freebsd-dev/sys/dev/netmap/netmap.c

3339 lines
94 KiB
C
Raw Normal View History

Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Copyright (C) 2011-2014 Matteo Landi
* Copyright (C) 2011-2016 Luigi Rizzo
* Copyright (C) 2011-2016 Giuseppe Lettieri
* Copyright (C) 2011-2016 Vincenzo Maffione
* All rights reserved.
*
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* $FreeBSD$
*
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* This module supports memory mapped access to network devices,
* see netmap(4).
*
* The module uses a large, memory pool allocated by the kernel
* and accessible as mmapped memory by multiple userspace threads/processes.
* The memory pool contains packet buffers and "netmap rings",
* i.e. user-accessible copies of the interface's queues.
*
* Access to the network card works like this:
* 1. a process/thread issues one or more open() on /dev/netmap, to create
* select()able file descriptor on which events are reported.
* 2. on each descriptor, the process issues an ioctl() to identify
* the interface that should report events to the file descriptor.
* 3. on each descriptor, the process issues an mmap() request to
* map the shared memory region within the process' address space.
* The list of interesting queues is indicated by a location in
* the shared memory region.
* 4. using the functions in the netmap(4) userspace API, a process
* can look up the occupation state of a queue, access memory buffers,
* and retrieve received packets or enqueue packets to transmit.
* 5. using some ioctl()s the process can synchronize the userspace view
* of the queue with the actual status in the kernel. This includes both
* receiving the notification of new packets, and transmitting new
* packets on the output interface.
* 6. select() or poll() can be used to wait for events on individual
* transmit or receive queues (or all queues for a given interface).
*
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
SYNCHRONIZATION (USER)
The netmap rings and data structures may be shared among multiple
user threads or even independent processes.
Any synchronization among those threads/processes is delegated
to the threads themselves. Only one thread at a time can be in
a system call on the same netmap ring. The OS does not enforce
this and only guarantees against system crashes in case of
invalid usage.
LOCKING (INTERNAL)
Within the kernel, access to the netmap rings is protected as follows:
- a spinlock on each ring, to handle producer/consumer races on
RX rings attached to the host stack (against multiple host
threads writing from the host stack to the same ring),
and on 'destination' rings attached to a VALE switch
(i.e. RX rings in VALE ports, and TX rings in NIC/host ports)
protecting multiple active senders for the same destination)
- an atomic variable to guarantee that there is at most one
instance of *_*xsync() on the ring at any time.
For rings connected to user file
descriptors, an atomic_test_and_set() protects this, and the
lock on the ring is not actually used.
For NIC RX rings connected to a VALE switch, an atomic_test_and_set()
is also used to prevent multiple executions (the driver might indeed
already guarantee this).
For NIC TX rings connected to a VALE switch, the lock arbitrates
access to the queue (both when allocating buffers and when pushing
them out).
- *xsync() should be protected against initializations of the card.
On FreeBSD most devices have the reset routine protected by
a RING lock (ixgbe, igb, em) or core lock (re). lem is missing
the RING protection on rx_reset(), this should be added.
On linux there is an external lock on the tx path, which probably
also arbitrates access to the reset routine. XXX to be revised
- a per-interface core_lock protecting access from the host stack
while interfaces may be detached from netmap mode.
XXX there should be no need for this lock if we detach the interfaces
only while they are down.
--- VALE SWITCH ---
NMG_LOCK() serializes all modifications to switches and ports.
A switch cannot be deleted until all ports are gone.
For each switch, an SX lock (RWlock on linux) protects
deletion of ports. When configuring or deleting a new port, the
lock is acquired in exclusive mode (after holding NMG_LOCK).
When forwarding, the lock is acquired in shared mode (without NMG_LOCK).
The lock is held throughout the entire forwarding cycle,
during which the thread may incur in a page fault.
Hence it is important that sleepable shared locks are used.
On the rx ring, the per-port lock is grabbed initially to reserve
a number of slot in the ring, then the lock is released,
packets are copied from source to destination, and then
the lock is acquired again and the receive ring is updated.
(A similar thing is done on the tx ring for NIC and host stack
ports attached to the switch)
*/
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/* --- internals ----
*
* Roadmap to the code that implements the above.
*
* > 1. a process/thread issues one or more open() on /dev/netmap, to create
* > select()able file descriptor on which events are reported.
*
* Internally, we allocate a netmap_priv_d structure, that will be
* initialized on ioctl(NIOCREGIF). There is one netmap_priv_d
* structure for each open().
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
* os-specific:
* FreeBSD: see netmap_open() (netmap_freebsd.c)
* linux: see linux_netmap_open() (netmap_linux.c)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
* > 2. on each descriptor, the process issues an ioctl() to identify
* > the interface that should report events to the file descriptor.
*
* Implemented by netmap_ioctl(), NIOCREGIF case, with nmr->nr_cmd==0.
* Most important things happen in netmap_get_na() and
* netmap_do_regif(), called from there. Additional details can be
* found in the comments above those functions.
*
* In all cases, this action creates/takes-a-reference-to a
* netmap_*_adapter describing the port, and allocates a netmap_if
* and all necessary netmap rings, filling them with netmap buffers.
*
* In this phase, the sync callbacks for each ring are set (these are used
* in steps 5 and 6 below). The callbacks depend on the type of adapter.
* The adapter creation/initialization code puts them in the
* netmap_adapter (fields na->nm_txsync and na->nm_rxsync). Then, they
* are copied from there to the netmap_kring's during netmap_do_regif(), by
* the nm_krings_create() callback. All the nm_krings_create callbacks
* actually call netmap_krings_create() to perform this and the other
* common stuff. netmap_krings_create() also takes care of the host rings,
* if needed, by setting their sync callbacks appropriately.
*
* Additional actions depend on the kind of netmap_adapter that has been
* registered:
*
* - netmap_hw_adapter: [netmap.c]
* This is a system netdev/ifp with native netmap support.
* The ifp is detached from the host stack by redirecting:
* - transmissions (from the network stack) to netmap_transmit()
* - receive notifications to the nm_notify() callback for
* this adapter. The callback is normally netmap_notify(), unless
* the ifp is attached to a bridge using bwrap, in which case it
* is netmap_bwrap_intr_notify().
*
* - netmap_generic_adapter: [netmap_generic.c]
* A system netdev/ifp without native netmap support.
*
* (the decision about native/non native support is taken in
* netmap_get_hw_na(), called by netmap_get_na())
*
* - netmap_vp_adapter [netmap_vale.c]
* Returned by netmap_get_bdg_na().
* This is a persistent or ephemeral VALE port. Ephemeral ports
* are created on the fly if they don't already exist, and are
* always attached to a bridge.
* Persistent VALE ports must must be created separately, and i
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* then attached like normal NICs. The NIOCREGIF we are examining
* will find them only if they had previosly been created and
* attached (see VALE_CTL below).
*
* - netmap_pipe_adapter [netmap_pipe.c]
* Returned by netmap_get_pipe_na().
* Both pipe ends are created, if they didn't already exist.
*
* - netmap_monitor_adapter [netmap_monitor.c]
* Returned by netmap_get_monitor_na().
* If successful, the nm_sync callbacks of the monitored adapter
* will be intercepted by the returned monitor.
*
* - netmap_bwrap_adapter [netmap_vale.c]
* Cannot be obtained in this way, see VALE_CTL below
*
*
* os-specific:
* linux: we first go through linux_netmap_ioctl() to
* adapt the FreeBSD interface to the linux one.
*
*
* > 3. on each descriptor, the process issues an mmap() request to
* > map the shared memory region within the process' address space.
* > The list of interesting queues is indicated by a location in
* > the shared memory region.
*
* os-specific:
* FreeBSD: netmap_mmap_single (netmap_freebsd.c).
* linux: linux_netmap_mmap (netmap_linux.c).
*
* > 4. using the functions in the netmap(4) userspace API, a process
* > can look up the occupation state of a queue, access memory buffers,
* > and retrieve received packets or enqueue packets to transmit.
*
* these actions do not involve the kernel.
*
* > 5. using some ioctl()s the process can synchronize the userspace view
* > of the queue with the actual status in the kernel. This includes both
* > receiving the notification of new packets, and transmitting new
* > packets on the output interface.
*
* These are implemented in netmap_ioctl(), NIOCTXSYNC and NIOCRXSYNC
* cases. They invoke the nm_sync callbacks on the netmap_kring
* structures, as initialized in step 2 and maybe later modified
* by a monitor. Monitors, however, will always call the original
* callback before doing anything else.
*
*
* > 6. select() or poll() can be used to wait for events on individual
* > transmit or receive queues (or all queues for a given interface).
*
* Implemented in netmap_poll(). This will call the same nm_sync()
* callbacks as in step 5 above.
*
* os-specific:
* linux: we first go through linux_netmap_poll() to adapt
* the FreeBSD interface to the linux one.
*
*
* ---- VALE_CTL -----
*
* VALE switches are controlled by issuing a NIOCREGIF with a non-null
* nr_cmd in the nmreq structure. These subcommands are handled by
* netmap_bdg_ctl() in netmap_vale.c. Persistent VALE ports are created
* and destroyed by issuing the NETMAP_BDG_NEWIF and NETMAP_BDG_DELIF
* subcommands, respectively.
*
* Any network interface known to the system (including a persistent VALE
* port) can be attached to a VALE switch by issuing the
* NETMAP_BDG_ATTACH subcommand. After the attachment, persistent VALE ports
* look exactly like ephemeral VALE ports (as created in step 2 above). The
* attachment of other interfaces, instead, requires the creation of a
* netmap_bwrap_adapter. Moreover, the attached interface must be put in
* netmap mode. This may require the creation of a netmap_generic_adapter if
* we have no native support for the interface, or if generic adapters have
* been forced by sysctl.
*
* Both persistent VALE ports and bwraps are handled by netmap_get_bdg_na(),
* called by nm_bdg_ctl_attach(), and discriminated by the nm_bdg_attach()
* callback. In the case of the bwrap, the callback creates the
* netmap_bwrap_adapter. The initialization of the bwrap is then
* completed by calling netmap_do_regif() on it, in the nm_bdg_ctl()
* callback (netmap_bwrap_bdg_ctl in netmap_vale.c).
* A generic adapter for the wrapped ifp will be created if needed, when
* netmap_get_bdg_na() calls netmap_get_hw_na().
*
*
* ---- DATAPATHS -----
*
* -= SYSTEM DEVICE WITH NATIVE SUPPORT =-
*
* na == NA(ifp) == netmap_hw_adapter created in DEVICE_netmap_attach()
*
* - tx from netmap userspace:
* concurrently:
* 1) ioctl(NIOCTXSYNC)/netmap_poll() in process context
* kring->nm_sync() == DEVICE_netmap_txsync()
* 2) device interrupt handler
* na->nm_notify() == netmap_notify()
* - rx from netmap userspace:
* concurrently:
* 1) ioctl(NIOCRXSYNC)/netmap_poll() in process context
* kring->nm_sync() == DEVICE_netmap_rxsync()
* 2) device interrupt handler
* na->nm_notify() == netmap_notify()
2015-07-10 05:51:36 +00:00
* - rx from host stack
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* concurrently:
* 1) host stack
* netmap_transmit()
* na->nm_notify == netmap_notify()
* 2) ioctl(NIOCRXSYNC)/netmap_poll() in process context
* kring->nm_sync() == netmap_rxsync_from_host
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* netmap_rxsync_from_host(na, NULL, NULL)
* - tx to host stack
* ioctl(NIOCTXSYNC)/netmap_poll() in process context
* kring->nm_sync() == netmap_txsync_to_host
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* netmap_txsync_to_host(na)
* nm_os_send_up()
* FreeBSD: na->if_input() == ether_input()
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* linux: netif_rx() with NM_MAGIC_PRIORITY_RX
*
*
* -= SYSTEM DEVICE WITH GENERIC SUPPORT =-
*
2015-07-10 05:51:36 +00:00
* na == NA(ifp) == generic_netmap_adapter created in generic_netmap_attach()
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
2015-07-10 05:51:36 +00:00
* - tx from netmap userspace:
* concurrently:
* 1) ioctl(NIOCTXSYNC)/netmap_poll() in process context
* kring->nm_sync() == generic_netmap_txsync()
* nm_os_generic_xmit_frame()
* linux: dev_queue_xmit() with NM_MAGIC_PRIORITY_TX
* ifp->ndo_start_xmit == generic_ndo_start_xmit()
* gna->save_start_xmit == orig. dev. start_xmit
* FreeBSD: na->if_transmit() == orig. dev if_transmit
2015-07-10 05:51:36 +00:00
* 2) generic_mbuf_destructor()
* na->nm_notify() == netmap_notify()
* - rx from netmap userspace:
* 1) ioctl(NIOCRXSYNC)/netmap_poll() in process context
* kring->nm_sync() == generic_netmap_rxsync()
* mbq_safe_dequeue()
* 2) device driver
* generic_rx_handler()
* mbq_safe_enqueue()
* na->nm_notify() == netmap_notify()
* - rx from host stack
* FreeBSD: same as native
* Linux: same as native except:
2015-07-10 05:51:36 +00:00
* 1) host stack
* dev_queue_xmit() without NM_MAGIC_PRIORITY_TX
* ifp->ndo_start_xmit == generic_ndo_start_xmit()
* netmap_transmit()
* na->nm_notify() == netmap_notify()
* - tx to host stack (same as native):
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
*
2015-07-10 05:51:36 +00:00
* -= VALE =-
*
* INCOMING:
*
* - VALE ports:
* ioctl(NIOCTXSYNC)/netmap_poll() in process context
* kring->nm_sync() == netmap_vp_txsync()
*
* - system device with native support:
* from cable:
* interrupt
* na->nm_notify() == netmap_bwrap_intr_notify(ring_nr != host ring)
* kring->nm_sync() == DEVICE_netmap_rxsync()
* netmap_vp_txsync()
* kring->nm_sync() == DEVICE_netmap_rxsync()
* from host stack:
* netmap_transmit()
* na->nm_notify() == netmap_bwrap_intr_notify(ring_nr == host ring)
* kring->nm_sync() == netmap_rxsync_from_host()
2015-07-10 05:51:36 +00:00
* netmap_vp_txsync()
*
* - system device with generic support:
* from device driver:
* generic_rx_handler()
* na->nm_notify() == netmap_bwrap_intr_notify(ring_nr != host ring)
* kring->nm_sync() == generic_netmap_rxsync()
* netmap_vp_txsync()
* kring->nm_sync() == generic_netmap_rxsync()
* from host stack:
* netmap_transmit()
* na->nm_notify() == netmap_bwrap_intr_notify(ring_nr == host ring)
* kring->nm_sync() == netmap_rxsync_from_host()
2015-07-10 05:51:36 +00:00
* netmap_vp_txsync()
*
* (all cases) --> nm_bdg_flush()
* dest_na->nm_notify() == (see below)
*
* OUTGOING:
*
* - VALE ports:
* concurrently:
* 1) ioctlNIOCRXSYNC)/netmap_poll() in process context
* kring->nm_sync() == netmap_vp_rxsync()
* 2) from nm_bdg_flush()
* na->nm_notify() == netmap_notify()
*
* - system device with native support:
* to cable:
* na->nm_notify() == netmap_bwrap_notify()
* netmap_vp_rxsync()
* kring->nm_sync() == DEVICE_netmap_txsync()
* netmap_vp_rxsync()
* to host stack:
* netmap_vp_rxsync()
* kring->nm_sync() == netmap_txsync_to_host
2015-07-10 05:51:36 +00:00
* netmap_vp_rxsync_locked()
*
* - system device with generic adapter:
* to device driver:
* na->nm_notify() == netmap_bwrap_notify()
* netmap_vp_rxsync()
* kring->nm_sync() == generic_netmap_txsync()
* netmap_vp_rxsync()
* to host stack:
* netmap_vp_rxsync()
* kring->nm_sync() == netmap_txsync_to_host
2015-07-10 05:51:36 +00:00
* netmap_vp_rxsync()
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
*/
/*
* OS-specific code that is used only within this file.
* Other OS-specific code that must be accessed by drivers
* is present in netmap_kern.h
*/
#if defined(__FreeBSD__)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#include <sys/cdefs.h> /* prerequisite */
#include <sys/types.h>
#include <sys/errno.h>
#include <sys/param.h> /* defines used in kernel.h */
#include <sys/kernel.h> /* types used in module initialization */
#include <sys/conf.h> /* cdevsw struct, UID, GID */
#include <sys/filio.h> /* FIONBIO */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#include <sys/sockio.h>
#include <sys/socketvar.h> /* struct socket */
#include <sys/malloc.h>
#include <sys/poll.h>
Switch the vm_object mutex to be a rwlock. This will enable in the future further optimizations where the vm_object lock will be held in read mode most of the time the page cache resident pool of pages are accessed for reading purposes. The change is mostly mechanical but few notes are reported: * The KPI changes as follow: - VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK() - VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK() - VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK() - VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED() (in order to avoid visibility of implementation details) - The read-mode operations are added: VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(), VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED() * The vm/vm_pager.h namespace pollution avoidance (forcing requiring sys/mutex.h in consumers directly to cater its inlining functions using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h consumers now must include also sys/rwlock.h. * zfs requires a quite convoluted fix to include FreeBSD rwlocks into the compat layer because the name clash between FreeBSD and solaris versions must be avoided. At this purpose zfs redefines the vm_object locking functions directly, isolating the FreeBSD components in specific compat stubs. The KPI results heavilly broken by this commit. Thirdy part ports must be updated accordingly (I can think off-hand of VirtualBox, for example). Sponsored by: EMC / Isilon storage division Reviewed by: jeff Reviewed by: pjd (ZFS specific review) Discussed with: alc Tested by: pho
2013-03-09 02:32:23 +00:00
#include <sys/rwlock.h>
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#include <sys/socket.h> /* sockaddrs */
#include <sys/selinfo.h>
#include <sys/sysctl.h>
2014-01-09 00:59:03 +00:00
#include <sys/jail.h>
#include <net/vnet.h>
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#include <net/if.h>
#include <net/if_var.h>
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#include <net/bpf.h> /* BIOCIMMEDIATE */
#include <machine/bus.h> /* bus_dmamap_* */
#include <sys/endian.h>
#include <sys/refcount.h>
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#elif defined(linux)
#include "bsd_glue.h"
#elif defined(__APPLE__)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#warning OSX support is only partial
#include "osx_glue.h"
#elif defined (_WIN32)
#include "win_glue.h"
#else
#error Unsupported platform
#endif /* unsupported */
/*
* common headers
*/
#include <net/netmap.h>
#include <dev/netmap/netmap_kern.h>
#include <dev/netmap/netmap_mem2.h>
/* user-controlled variables */
int netmap_verbose;
static int netmap_no_timestamp; /* don't timestamp on rxsync */
int netmap_mitigate = 1;
int netmap_no_pendintr = 1;
int netmap_txsync_retry = 2;
int netmap_flags = 0; /* debug flags */
static int netmap_fwd = 0; /* force transparent mode */
/*
* netmap_admode selects the netmap mode to use.
* Invalid values are reset to NETMAP_ADMODE_BEST
*/
enum { NETMAP_ADMODE_BEST = 0, /* use native, fallback to generic */
NETMAP_ADMODE_NATIVE, /* either native or none */
NETMAP_ADMODE_GENERIC, /* force generic */
NETMAP_ADMODE_LAST };
static int netmap_admode = NETMAP_ADMODE_BEST;
/* netmap_generic_mit controls mitigation of RX notifications for
* the generic netmap adapter. The value is a time interval in
* nanoseconds. */
int netmap_generic_mit = 100*1000;
/* We use by default netmap-aware qdiscs with generic netmap adapters,
* even if there can be a little performance hit with hardware NICs.
* However, using the qdisc is the safer approach, for two reasons:
* 1) it prevents non-fifo qdiscs to break the TX notification
* scheme, which is based on mbuf destructors when txqdisc is
* not used.
* 2) it makes it possible to transmit over software devices that
* change skb->dev, like bridge, veth, ...
*
* Anyway users looking for the best performance should
* use native adapters.
*/
int netmap_generic_txqdisc = 1;
/* Default number of slots and queues for generic adapters. */
int netmap_generic_ringsize = 1024;
int netmap_generic_rings = 1;
/* Non-zero if ptnet devices are allowed to use virtio-net headers. */
int ptnet_vnet_hdr = 1;
/*
* SYSCTL calls are grouped between SYSBEGIN and SYSEND to be emulated
* in some other operating systems
*/
SYSBEGIN(main_init);
SYSCTL_DECL(_dev_netmap);
SYSCTL_NODE(_dev, OID_AUTO, netmap, CTLFLAG_RW, 0, "Netmap args");
SYSCTL_INT(_dev_netmap, OID_AUTO, verbose,
CTLFLAG_RW, &netmap_verbose, 0, "Verbose mode");
SYSCTL_INT(_dev_netmap, OID_AUTO, no_timestamp,
CTLFLAG_RW, &netmap_no_timestamp, 0, "no_timestamp");
SYSCTL_INT(_dev_netmap, OID_AUTO, mitigate, CTLFLAG_RW, &netmap_mitigate, 0, "");
SYSCTL_INT(_dev_netmap, OID_AUTO, no_pendintr,
CTLFLAG_RW, &netmap_no_pendintr, 0, "Always look for new received packets.");
SYSCTL_INT(_dev_netmap, OID_AUTO, txsync_retry, CTLFLAG_RW,
&netmap_txsync_retry, 0 , "Number of txsync loops in bridge's flush.");
SYSCTL_INT(_dev_netmap, OID_AUTO, flags, CTLFLAG_RW, &netmap_flags, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, fwd, CTLFLAG_RW, &netmap_fwd, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, admode, CTLFLAG_RW, &netmap_admode, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, generic_mit, CTLFLAG_RW, &netmap_generic_mit, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, generic_ringsize, CTLFLAG_RW, &netmap_generic_ringsize, 0 , "");
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
SYSCTL_INT(_dev_netmap, OID_AUTO, generic_rings, CTLFLAG_RW, &netmap_generic_rings, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, generic_txqdisc, CTLFLAG_RW, &netmap_generic_txqdisc, 0 , "");
SYSCTL_INT(_dev_netmap, OID_AUTO, ptnet_vnet_hdr, CTLFLAG_RW, &ptnet_vnet_hdr, 0 , "");
SYSEND;
NMG_LOCK_T netmap_global_lock;
/*
* mark the ring as stopped, and run through the locks
* to make sure other users get to see it.
* stopped must be either NR_KR_STOPPED (for unbounded stop)
* of NR_KR_LOCKED (brief stop for mutual exclusion purposes)
*/
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
static void
netmap_disable_ring(struct netmap_kring *kr, int stopped)
{
nm_kr_stop(kr, stopped);
// XXX check if nm_kr_stop is sufficient
mtx_lock(&kr->q_lock);
mtx_unlock(&kr->q_lock);
nm_kr_put(kr);
}
2015-07-10 05:51:36 +00:00
/* stop or enable a single ring */
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
void
2015-07-10 05:51:36 +00:00
netmap_set_ring(struct netmap_adapter *na, u_int ring_id, enum txrx t, int stopped)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
{
if (stopped)
netmap_disable_ring(NMR(na, t) + ring_id, stopped);
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
else
2015-07-10 05:51:36 +00:00
NMR(na, t)[ring_id].nkr_stopped = 0;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
}
/* stop or enable all the rings of na */
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
void
netmap_set_all_rings(struct netmap_adapter *na, int stopped)
{
int i;
2015-07-10 05:51:36 +00:00
enum txrx t;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (!nm_netmap_on(na))
return;
2015-07-10 05:51:36 +00:00
for_rx_tx(t) {
for (i = 0; i < netmap_real_rings(na, t); i++) {
netmap_set_ring(na, i, t, stopped);
}
}
}
/*
* Convenience function used in drivers. Waits for current txsync()s/rxsync()s
* to finish and prevents any new one from starting. Call this before turning
2015-12-25 14:51:36 +00:00
* netmap mode off, or before removing the hardware rings (e.g., on module
* onload).
*/
void
netmap_disable_all_rings(struct ifnet *ifp)
{
if (NM_NA_VALID(ifp)) {
netmap_set_all_rings(NA(ifp), NM_KR_STOPPED);
}
}
/*
* Convenience function used in drivers. Re-enables rxsync and txsync on the
* adapter's rings In linux drivers, this should be placed near each
* napi_enable().
*/
void
netmap_enable_all_rings(struct ifnet *ifp)
{
if (NM_NA_VALID(ifp)) {
netmap_set_all_rings(NA(ifp), 0 /* enabled */);
}
}
void
netmap_make_zombie(struct ifnet *ifp)
{
if (NM_NA_VALID(ifp)) {
struct netmap_adapter *na = NA(ifp);
netmap_set_all_rings(na, NM_KR_LOCKED);
na->na_flags |= NAF_ZOMBIE;
netmap_set_all_rings(na, 0);
}
}
void
netmap_undo_zombie(struct ifnet *ifp)
{
if (NM_NA_VALID(ifp)) {
struct netmap_adapter *na = NA(ifp);
if (na->na_flags & NAF_ZOMBIE) {
netmap_set_all_rings(na, NM_KR_LOCKED);
na->na_flags &= ~NAF_ZOMBIE;
netmap_set_all_rings(na, 0);
}
}
}
/*
* generic bound_checking function
*/
u_int
nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg)
{
u_int oldv = *v;
const char *op = NULL;
if (dflt < lo)
dflt = lo;
if (dflt > hi)
dflt = hi;
if (oldv < lo) {
*v = dflt;
op = "Bump";
} else if (oldv > hi) {
*v = hi;
op = "Clamp";
}
if (op && msg)
printf("%s %s to %d (was %d)\n", op, msg, *v, oldv);
return *v;
}
/*
* packet-dump function, user-supplied or static buffer.
* The destination buffer must be at least 30+4*len
*/
const char *
nm_dump_buf(char *p, int len, int lim, char *dst)
{
static char _dst[8192];
int i, j, i0;
static char hex[] ="0123456789abcdef";
char *o; /* output position */
#define P_HI(x) hex[((x) & 0xf0)>>4]
#define P_LO(x) hex[((x) & 0xf)]
#define P_C(x) ((x) >= 0x20 && (x) <= 0x7e ? (x) : '.')
if (!dst)
dst = _dst;
if (lim <= 0 || lim > len)
lim = len;
o = dst;
sprintf(o, "buf 0x%p len %d lim %d\n", p, len, lim);
o += strlen(o);
/* hexdump routine */
for (i = 0; i < lim; ) {
sprintf(o, "%5d: ", i);
o += strlen(o);
memset(o, ' ', 48);
i0 = i;
for (j=0; j < 16 && i < lim; i++, j++) {
o[j*3] = P_HI(p[i]);
o[j*3+1] = P_LO(p[i]);
}
i = i0;
for (j=0; j < 16 && i < lim; i++, j++)
o[j + 48] = P_C(p[i]);
o[j+48] = '\n';
o += j+49;
}
*o = '\0';
#undef P_HI
#undef P_LO
#undef P_C
return dst;
}
/*
* Fetch configuration from the device, to cope with dynamic
* reconfigurations after loading the module.
*/
/* call with NMG_LOCK held */
int
netmap_update_config(struct netmap_adapter *na)
{
u_int txr, txd, rxr, rxd;
txr = txd = rxr = rxd = 0;
if (na->nm_config == NULL ||
2015-07-10 05:51:36 +00:00
na->nm_config(na, &txr, &txd, &rxr, &rxd))
{
/* take whatever we had at init time */
txr = na->num_tx_rings;
txd = na->num_tx_desc;
rxr = na->num_rx_rings;
rxd = na->num_rx_desc;
2013-05-02 16:01:04 +00:00
}
if (na->num_tx_rings == txr && na->num_tx_desc == txd &&
na->num_rx_rings == rxr && na->num_rx_desc == rxd)
return 0; /* nothing changed */
if (netmap_verbose || na->active_fds > 0) {
D("stored config %s: txring %d x %d, rxring %d x %d",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name,
na->num_tx_rings, na->num_tx_desc,
na->num_rx_rings, na->num_rx_desc);
D("new config %s: txring %d x %d, rxring %d x %d",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name, txr, txd, rxr, rxd);
}
if (na->active_fds == 0) {
D("configuration changed (but fine)");
na->num_tx_rings = txr;
na->num_tx_desc = txd;
na->num_rx_rings = rxr;
na->num_rx_desc = rxd;
return 0;
}
D("configuration changed while active, this is bad...");
return 1;
}
/* nm_sync callbacks for the host rings */
static int netmap_txsync_to_host(struct netmap_kring *kring, int flags);
static int netmap_rxsync_from_host(struct netmap_kring *kring, int flags);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
/* create the krings array and initialize the fields common to all adapters.
* The array layout is this:
*
* +----------+
* na->tx_rings ----->| | \
* | | } na->num_tx_ring
* | | /
* +----------+
* | | host tx kring
* na->rx_rings ----> +----------+
* | | \
* | | } na->num_rx_rings
* | | /
* +----------+
* | | host rx kring
* +----------+
* na->tailroom ----->| | \
* | | } tailroom bytes
* | | /
* +----------+
*
* Note: for compatibility, host krings are created even when not needed.
* The tailroom space is currently used by vale ports for allocating leases.
*/
/* call with NMG_LOCK held */
int
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
netmap_krings_create(struct netmap_adapter *na, u_int tailroom)
{
u_int i, len, ndesc;
struct netmap_kring *kring;
2015-07-10 05:51:36 +00:00
u_int n[NR_TXRX];
enum txrx t;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
/* account for the (possibly fake) host rings */
2015-07-10 05:51:36 +00:00
n[NR_TX] = na->num_tx_rings + 1;
n[NR_RX] = na->num_rx_rings + 1;
2015-07-10 05:51:36 +00:00
len = (n[NR_TX] + n[NR_RX]) * sizeof(struct netmap_kring) + tailroom;
na->tx_rings = malloc((size_t)len, M_DEVBUF, M_NOWAIT | M_ZERO);
if (na->tx_rings == NULL) {
D("Cannot allocate krings");
return ENOMEM;
}
2015-07-10 05:51:36 +00:00
na->rx_rings = na->tx_rings + n[NR_TX];
/*
* All fields in krings are 0 except the one initialized below.
* but better be explicit on important kring fields.
*/
2015-07-10 05:51:36 +00:00
for_rx_tx(t) {
ndesc = nma_get_ndesc(na, t);
for (i = 0; i < n[t]; i++) {
kring = &NMR(na, t)[i];
bzero(kring, sizeof(*kring));
kring->na = na;
kring->ring_id = i;
kring->tx = t;
kring->nkr_num_slots = ndesc;
kring->nr_mode = NKR_NETMAP_OFF;
kring->nr_pending_mode = NKR_NETMAP_OFF;
2015-07-10 05:51:36 +00:00
if (i < nma_get_nrings(na, t)) {
kring->nm_sync = (t == NR_TX ? na->nm_txsync : na->nm_rxsync);
} else {
2015-07-10 05:51:36 +00:00
kring->nm_sync = (t == NR_TX ?
netmap_txsync_to_host:
netmap_rxsync_from_host);
2015-07-10 05:51:36 +00:00
}
kring->nm_notify = na->nm_notify;
kring->rhead = kring->rcur = kring->nr_hwcur = 0;
/*
* IMPORTANT: Always keep one slot empty.
*/
kring->rtail = kring->nr_hwtail = (t == NR_TX ? ndesc - 1 : 0);
snprintf(kring->name, sizeof(kring->name) - 1, "%s %s%d", na->name,
2015-07-10 05:51:36 +00:00
nm_txrx2str(t), i);
ND("ktx %s h %d c %d t %d",
kring->name, kring->rhead, kring->rcur, kring->rtail);
mtx_init(&kring->q_lock, (t == NR_TX ? "nm_txq_lock" : "nm_rxq_lock"), NULL, MTX_DEF);
nm_os_selinfo_init(&kring->si);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
}
nm_os_selinfo_init(&na->si[t]);
}
2015-07-10 05:51:36 +00:00
na->tailroom = na->rx_rings + n[NR_RX];
return 0;
}
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
/* undo the actions performed by netmap_krings_create */
/* call with NMG_LOCK held */
void
netmap_krings_delete(struct netmap_adapter *na)
{
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
struct netmap_kring *kring = na->tx_rings;
2015-07-10 05:51:36 +00:00
enum txrx t;
for_rx_tx(t)
nm_os_selinfo_uninit(&na->si[t]);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
/* we rely on the krings layout described above */
for ( ; kring != na->tailroom; kring++) {
mtx_destroy(&kring->q_lock);
nm_os_selinfo_uninit(&kring->si);
}
free(na->tx_rings, M_DEVBUF);
na->tx_rings = na->rx_rings = na->tailroom = NULL;
}
/*
* Destructor for NIC ports. They also have an mbuf queue
* on the rings connected to the host so we need to purge
* them first.
*/
/* call with NMG_LOCK held */
void
netmap_hw_krings_delete(struct netmap_adapter *na)
{
struct mbq *q = &na->rx_rings[na->num_rx_rings].rx_queue;
ND("destroy sw mbq with len %d", mbq_len(q));
mbq_purge(q);
mbq_safe_fini(q);
netmap_krings_delete(na);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
2015-07-10 05:51:36 +00:00
* Undo everything that was done in netmap_do_regif(). In particular,
* call nm_register(ifp,0) to stop netmap mode on the interface and
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* revert to normal operation.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
/* call with NMG_LOCK held */
2015-07-10 05:51:36 +00:00
static void netmap_unset_ringid(struct netmap_priv_d *);
static void netmap_krings_put(struct netmap_priv_d *);
void
2015-07-10 05:51:36 +00:00
netmap_do_unregif(struct netmap_priv_d *priv)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_adapter *na = priv->np_na;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
NMG_LOCK_ASSERT();
na->active_fds--;
/* unset nr_pending_mode and possibly release exclusive mode */
netmap_krings_put(priv);
2015-07-10 05:51:36 +00:00
#ifdef WITH_MONITOR
/* XXX check whether we have to do something with monitor
* when rings change nr_mode. */
if (na->active_fds <= 0) {
2015-07-10 05:51:36 +00:00
/* walk through all the rings and tell any monitor
* that the port is going to exit netmap mode
*/
netmap_monitor_stop(na);
}
2015-07-10 05:51:36 +00:00
#endif
if (na->active_fds <= 0 || nm_kring_pending(priv)) {
na->nm_register(na, 0);
}
/* delete rings and buffers that are no longer needed */
netmap_mem_rings_delete(na);
if (na->active_fds <= 0) { /* last instance */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* (TO CHECK) We enter here
* when the last reference to this file descriptor goes
* away. This means we cannot have any pending poll()
* or interrupt routine operating on the structure.
* XXX The file may be closed in a thread while
* another thread is using it.
* Linux keeps the file opened until the last reference
* by any outstanding ioctl/poll or mmap is gone.
* FreeBSD does not track mmap()s (but we do) and
* wakes up any sleeping poll(). Need to check what
* happens if the close() occurs while a concurrent
* syscall is running.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
if (netmap_verbose)
D("deleting last instance for %s", na->name);
if (nm_netmap_on(na)) {
D("BUG: netmap on while going to delete the krings");
}
na->nm_krings_delete(na);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
2015-07-10 05:51:36 +00:00
/* possibily decrement counter of tx_si/rx_si users */
netmap_unset_ringid(priv);
/* delete the nifp */
2015-07-10 05:51:36 +00:00
netmap_mem_if_delete(na, priv->np_nifp);
/* drop the allocator */
netmap_mem_deref(na->nm_mem, na);
/* mark the priv as unregistered */
priv->np_na = NULL;
priv->np_nifp = NULL;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
}
/* call with NMG_LOCK held */
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
static __inline int
2015-07-10 05:51:36 +00:00
nm_si_user(struct netmap_priv_d *priv, enum txrx t)
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
{
return (priv->np_na != NULL &&
2015-07-10 05:51:36 +00:00
(priv->np_qlast[t] - priv->np_qfirst[t] > 1));
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
}
struct netmap_priv_d*
netmap_priv_new(void)
{
struct netmap_priv_d *priv;
priv = malloc(sizeof(struct netmap_priv_d), M_DEVBUF,
M_NOWAIT | M_ZERO);
if (priv == NULL)
return NULL;
priv->np_refs = 1;
nm_os_get_module();
return priv;
}
/*
* Destructor of the netmap_priv_d, called when the fd is closed
* Action: undo all the things done by NIOCREGIF,
* On FreeBSD we need to track whether there are active mmap()s,
* and we use np_active_mmaps for that. On linux, the field is always 0.
* Return: 1 if we can free priv, 0 otherwise.
*
*/
/* call with NMG_LOCK held */
void
netmap_priv_delete(struct netmap_priv_d *priv)
{
struct netmap_adapter *na = priv->np_na;
/* number of active references to this fd */
if (--priv->np_refs > 0) {
return;
}
nm_os_put_module();
if (na) {
netmap_do_unregif(priv);
}
netmap_unget_na(na, priv->np_ifp);
bzero(priv, sizeof(*priv)); /* for safety */
free(priv, M_DEVBUF);
}
/* call with NMG_LOCK *not* held */
void
netmap_dtor(void *data)
{
struct netmap_priv_d *priv = data;
NMG_LOCK();
netmap_priv_delete(priv);
NMG_UNLOCK();
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Handlers for synchronization of the queues from/to the host.
* Netmap has two operating modes:
* - in the default mode, the rings connected to the host stack are
* just another ring pair managed by userspace;
* - in transparent mode (XXX to be defined) incoming packets
* (from the host or the NIC) are marked as NS_FORWARD upon
* arrival, and the user application has a chance to reset the
* flag for packets that should be dropped.
* On the RXSYNC or poll(), packets in RX rings between
* kring->nr_kcur and ring->cur with NS_FORWARD still set are moved
* to the other side.
* The transfer NIC --> host is relatively easy, just encapsulate
* into mbufs and we are done. The host --> NIC side is slightly
* harder because there might not be room in the tx ring so it
* might take a while before releasing the buffer.
*/
/*
* pass a chain of buffers to the host stack as coming from 'dst'
* We do not need to lock because the queue is private.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
static void
netmap_send_up(struct ifnet *dst, struct mbq *q)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct mbuf *m;
struct mbuf *head = NULL, *prev = NULL;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/* send packets up, outside the lock */
while ((m = mbq_dequeue(q)) != NULL) {
if (netmap_verbose & NM_VERB_HOST)
D("sending up pkt %p size %d", m, MBUF_LEN(m));
prev = nm_os_send_up(dst, m, prev);
if (head == NULL)
head = prev;
}
if (head)
nm_os_send_up(dst, NULL, head);
mbq_fini(q);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* put a copy of the buffers marked NS_FORWARD into an mbuf chain.
* Take packets from hwcur to ring->head marked NS_FORWARD (or forced)
* and pass them up. Drop remaining packets in the unlikely event
* of an mbuf shortage.
*/
static void
netmap_grab_packets(struct netmap_kring *kring, struct mbq *q, int force)
{
u_int const lim = kring->nkr_num_slots - 1;
2015-07-10 05:51:36 +00:00
u_int const head = kring->rhead;
u_int n;
struct netmap_adapter *na = kring->na;
for (n = kring->nr_hwcur; n != head; n = nm_next(n, lim)) {
struct mbuf *m;
struct netmap_slot *slot = &kring->ring->slot[n];
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
if ((slot->flags & NS_FORWARD) == 0 && !force)
continue;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (slot->len < 14 || slot->len > NETMAP_BUF_SIZE(na)) {
RD(5, "bad pkt at %d len %d", n, slot->len);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
continue;
}
slot->flags &= ~NS_FORWARD; // XXX needed ?
/* XXX TODO: adapt to the case of a multisegment packet */
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
m = m_devget(NMB(na, slot), slot->len, 0, na->ifp, NULL);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
if (m == NULL)
break;
mbq_enqueue(q, m);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
}
static inline int
_nm_may_forward(struct netmap_kring *kring)
{
return ((netmap_fwd || kring->ring->flags & NR_FORWARD) &&
kring->na->na_flags & NAF_HOST_RINGS &&
kring->tx == NR_RX);
}
static inline int
nm_may_forward_up(struct netmap_kring *kring)
{
return _nm_may_forward(kring) &&
kring->ring_id != kring->na->num_rx_rings;
}
static inline int
nm_may_forward_down(struct netmap_kring *kring)
{
return _nm_may_forward(kring) &&
kring->ring_id == kring->na->num_rx_rings;
}
/*
* Send to the NIC rings packets marked NS_FORWARD between
* kring->nr_hwcur and kring->rhead
* Called under kring->rx_queue.lock on the sw rx ring,
*/
static u_int
netmap_sw_to_nic(struct netmap_adapter *na)
{
struct netmap_kring *kring = &na->rx_rings[na->num_rx_rings];
struct netmap_slot *rxslot = kring->ring->slot;
u_int i, rxcur = kring->nr_hwcur;
u_int const head = kring->rhead;
u_int const src_lim = kring->nkr_num_slots - 1;
u_int sent = 0;
/* scan rings to find space, then fill as much as possible */
for (i = 0; i < na->num_tx_rings; i++) {
struct netmap_kring *kdst = &na->tx_rings[i];
struct netmap_ring *rdst = kdst->ring;
u_int const dst_lim = kdst->nkr_num_slots - 1;
/* XXX do we trust ring or kring->rcur,rtail ? */
for (; rxcur != head && !nm_ring_empty(rdst);
rxcur = nm_next(rxcur, src_lim) ) {
struct netmap_slot *src, *dst, tmp;
u_int dst_head = rdst->head;
src = &rxslot[rxcur];
if ((src->flags & NS_FORWARD) == 0 && !netmap_fwd)
continue;
sent++;
dst = &rdst->slot[dst_head];
tmp = *src;
src->buf_idx = dst->buf_idx;
src->flags = NS_BUF_CHANGED;
dst->buf_idx = tmp.buf_idx;
dst->len = tmp.len;
dst->flags = NS_BUF_CHANGED;
rdst->head = rdst->cur = nm_next(dst_head, dst_lim);
}
/* if (sent) XXX txsync ? */
}
return sent;
}
/*
* netmap_txsync_to_host() passes packets up. We are called from a
* system call in user process context, and the only contention
* can be among multiple user threads erroneously calling
* this routine concurrently.
*/
static int
netmap_txsync_to_host(struct netmap_kring *kring, int flags)
{
struct netmap_adapter *na = kring->na;
u_int const lim = kring->nkr_num_slots - 1;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
u_int const head = kring->rhead;
struct mbq q;
/* Take packets from hwcur to head and pass them up.
* force head = cur since netmap_grab_packets() stops at head
* In case of no buffers we give up. At the end of the loop,
* the queue is drained in all cases.
*/
mbq_init(&q);
netmap_grab_packets(kring, &q, 1 /* force */);
ND("have %d pkts in queue", mbq_len(&q));
kring->nr_hwcur = head;
kring->nr_hwtail = head + lim;
if (kring->nr_hwtail > lim)
kring->nr_hwtail -= lim + 1;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
netmap_send_up(na->ifp, &q);
return 0;
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* rxsync backend for packets coming from the host stack.
* They have been put in kring->rx_queue by netmap_transmit().
* We protect access to the kring using kring->rx_queue.lock
*
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* This routine also does the selrecord if called from the poll handler
* (we know because sr != NULL).
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
* returns the number of packets delivered to tx queues in
* transparent mode, or a negative value if error
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
static int
netmap_rxsync_from_host(struct netmap_kring *kring, int flags)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_adapter *na = kring->na;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
struct netmap_ring *ring = kring->ring;
u_int nm_i, n;
u_int const lim = kring->nkr_num_slots - 1;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
u_int const head = kring->rhead;
int ret = 0;
2015-07-10 05:51:36 +00:00
struct mbq *q = &kring->rx_queue, fq;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
2015-07-10 05:51:36 +00:00
mbq_init(&fq); /* fq holds packets to be freed */
mbq_lock(q);
/* First part: import newly received packets */
n = mbq_len(q);
if (n) { /* grab packets from the queue */
struct mbuf *m;
uint32_t stop_i;
nm_i = kring->nr_hwtail;
stop_i = nm_prev(nm_i, lim);
while ( nm_i != stop_i && (m = mbq_dequeue(q)) != NULL ) {
int len = MBUF_LEN(m);
struct netmap_slot *slot = &ring->slot[nm_i];
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
m_copydata(m, 0, len, NMB(na, slot));
ND("nm %d len %d", nm_i, len);
if (netmap_verbose)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
D("%s", nm_dump_buf(NMB(na, slot),len, 128, NULL));
slot->len = len;
slot->flags = kring->nkr_slot_flags;
nm_i = nm_next(nm_i, lim);
2015-07-10 05:51:36 +00:00
mbq_enqueue(&fq, m);
}
kring->nr_hwtail = nm_i;
}
/*
* Second part: skip past packets that userspace has released.
*/
nm_i = kring->nr_hwcur;
if (nm_i != head) { /* something was released */
if (nm_may_forward_down(kring)) {
ret = netmap_sw_to_nic(na);
if (ret > 0) {
kring->nr_kflags |= NR_FORWARD;
ret = 0;
}
}
kring->nr_hwcur = head;
}
mbq_unlock(q);
2015-07-10 05:51:36 +00:00
mbq_purge(&fq);
mbq_fini(&fq);
2015-07-10 05:51:36 +00:00
return ret;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
/* Get a netmap adapter for the port.
*
* If it is possible to satisfy the request, return 0
* with *na containing the netmap adapter found.
* Otherwise return an error code, with *na containing NULL.
*
* When the port is attached to a bridge, we always return
* EBUSY.
* Otherwise, if the port is already bound to a file descriptor,
* then we unconditionally return the existing adapter into *na.
* In all the other cases, we return (into *na) either native,
* generic or NULL, according to the following table:
*
* native_support
* active_fds dev.netmap.admode YES NO
* -------------------------------------------------------
* >0 * NA(ifp) NA(ifp)
*
* 0 NETMAP_ADMODE_BEST NATIVE GENERIC
* 0 NETMAP_ADMODE_NATIVE NATIVE NULL
* 0 NETMAP_ADMODE_GENERIC GENERIC GENERIC
*
*/
static void netmap_hw_dtor(struct netmap_adapter *); /* needed by NM_IS_NATIVE() */
int
netmap_get_hw_na(struct ifnet *ifp, struct netmap_adapter **na)
{
/* generic support */
int i = netmap_admode; /* Take a snapshot. */
struct netmap_adapter *prev_na;
2015-07-10 05:51:36 +00:00
int error = 0;
*na = NULL; /* default */
/* reset in case of invalid value */
if (i < NETMAP_ADMODE_BEST || i >= NETMAP_ADMODE_LAST)
i = netmap_admode = NETMAP_ADMODE_BEST;
if (NM_NA_VALID(ifp)) {
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
prev_na = NA(ifp);
/* If an adapter already exists, return it if
* there are active file descriptors or if
* netmap is not forced to use generic
* adapters.
*/
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (NETMAP_OWNED_BY_ANY(prev_na)
|| i != NETMAP_ADMODE_GENERIC
|| prev_na->na_flags & NAF_FORCE_NATIVE
#ifdef WITH_PIPES
/* ugly, but we cannot allow an adapter switch
* if some pipe is referring to this one
*/
|| prev_na->na_next_pipe > 0
#endif
) {
*na = prev_na;
return 0;
}
}
/* If there isn't native support and netmap is not allowed
* to use generic adapters, we cannot satisfy the request.
*/
if (!NM_IS_NATIVE(ifp) && i == NETMAP_ADMODE_NATIVE)
return EOPNOTSUPP;
/* Otherwise, create a generic adapter and return it,
* saving the previously used netmap adapter, if any.
*
* Note that here 'prev_na', if not NULL, MUST be a
* native adapter, and CANNOT be a generic one. This is
* true because generic adapters are created on demand, and
* destroyed when not used anymore. Therefore, if the adapter
* currently attached to an interface 'ifp' is generic, it
* must be that
* (NA(ifp)->active_fds > 0 || NETMAP_OWNED_BY_KERN(NA(ifp))).
* Consequently, if NA(ifp) is generic, we will enter one of
* the branches above. This ensures that we never override
* a generic adapter with another generic adapter.
*/
error = generic_netmap_attach(ifp);
if (error)
return error;
*na = NA(ifp);
return 0;
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* MUST BE CALLED UNDER NMG_LOCK()
*
* Get a refcounted reference to a netmap adapter attached
* to the interface specified by nmr.
* This is always called in the execution of an ioctl().
*
* Return ENXIO if the interface specified by the request does
* not exist, ENOTSUP if netmap is not supported by the interface,
* EBUSY if the interface is already attached to a bridge,
* EINVAL if parameters are invalid, ENOMEM if needed resources
* could not be allocated.
* If successful, hold a reference to the netmap adapter.
*
* If the interface specified by nmr is a system one, also keep
* a reference to it and return a valid *ifp.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
int
netmap_get_na(struct nmreq *nmr, struct netmap_adapter **na,
struct ifnet **ifp, int create)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
int error = 0;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
struct netmap_adapter *ret = NULL;
*na = NULL; /* default return value */
*ifp = NULL;
NMG_LOCK_ASSERT();
/* We cascade through all possible types of netmap adapter.
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* All netmap_get_*_na() functions return an error and an na,
* with the following combinations:
*
* error na
* 0 NULL type doesn't match
* !0 NULL type matches, but na creation/lookup failed
* 0 !NULL type matches and na created/found
* !0 !NULL impossible
*/
/* try to see if this is a ptnetmap port */
error = netmap_get_pt_host_na(nmr, na, create);
if (error || *na != NULL)
return error;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/* try to see if this is a monitor port */
error = netmap_get_monitor_na(nmr, na, create);
if (error || *na != NULL)
return error;
/* try to see if this is a pipe port */
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
error = netmap_get_pipe_na(nmr, na, create);
if (error || *na != NULL)
return error;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/* try to see if this is a bridge port */
error = netmap_get_bdg_na(nmr, na, create);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (error)
return error;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (*na != NULL) /* valid match in netmap_get_bdg_na() */
2015-07-10 05:51:36 +00:00
goto out;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
/*
* This must be a hardware na, lookup the name in the system.
* Note that by hardware we actually mean "it shows up in ifconfig".
* This may still be a tap, a veth/epair, or even a
* persistent VALE port.
*/
*ifp = ifunit_ref(nmr->nr_name);
if (*ifp == NULL) {
return ENXIO;
}
error = netmap_get_hw_na(*ifp, &ret);
if (error)
goto out;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
*na = ret;
netmap_adapter_get(ret);
out:
if (error) {
if (ret)
netmap_adapter_put(ret);
if (*ifp) {
if_rele(*ifp);
*ifp = NULL;
}
}
return error;
}
/* undo netmap_get_na() */
void
netmap_unget_na(struct netmap_adapter *na, struct ifnet *ifp)
{
if (ifp)
if_rele(ifp);
if (na)
netmap_adapter_put(na);
}
#define NM_FAIL_ON(t) do { \
if (unlikely(t)) { \
RD(5, "%s: fail '" #t "' " \
"h %d c %d t %d " \
"rh %d rc %d rt %d " \
"hc %d ht %d", \
kring->name, \
head, cur, ring->tail, \
kring->rhead, kring->rcur, kring->rtail, \
kring->nr_hwcur, kring->nr_hwtail); \
return kring->nkr_num_slots; \
} \
} while (0)
/*
* validate parameters on entry for *_txsync()
* Returns ring->cur if ok, or something >= kring->nkr_num_slots
* in case of error.
*
* rhead, rcur and rtail=hwtail are stored from previous round.
* hwcur is the next packet to send to the ring.
*
* We want
* hwcur <= *rhead <= head <= cur <= tail = *rtail <= hwtail
*
* hwcur, rhead, rtail and hwtail are reliable
*/
u_int
nm_txsync_prologue(struct netmap_kring *kring, struct netmap_ring *ring)
{
u_int head = ring->head; /* read only once */
u_int cur = ring->cur; /* read only once */
u_int n = kring->nkr_num_slots;
ND(5, "%s kcur %d ktail %d head %d cur %d tail %d",
kring->name,
kring->nr_hwcur, kring->nr_hwtail,
ring->head, ring->cur, ring->tail);
#if 1 /* kernel sanity checks; but we can trust the kring. */
NM_FAIL_ON(kring->nr_hwcur >= n || kring->rhead >= n ||
kring->rtail >= n || kring->nr_hwtail >= n);
#endif /* kernel sanity checks */
/*
* user sanity checks. We only use head,
* A, B, ... are possible positions for head:
*
* 0 A rhead B rtail C n-1
* 0 D rtail E rhead F n-1
*
* B, F, D are valid. A, C, E are wrong
*/
if (kring->rtail >= kring->rhead) {
/* want rhead <= head <= rtail */
NM_FAIL_ON(head < kring->rhead || head > kring->rtail);
/* and also head <= cur <= rtail */
NM_FAIL_ON(cur < head || cur > kring->rtail);
} else { /* here rtail < rhead */
/* we need head outside rtail .. rhead */
NM_FAIL_ON(head > kring->rtail && head < kring->rhead);
/* two cases now: head <= rtail or head >= rhead */
if (head <= kring->rtail) {
/* want head <= cur <= rtail */
NM_FAIL_ON(cur < head || cur > kring->rtail);
} else { /* head >= rhead */
/* cur must be outside rtail..head */
NM_FAIL_ON(cur > kring->rtail && cur < head);
}
}
if (ring->tail != kring->rtail) {
RD(5, "%s tail overwritten was %d need %d", kring->name,
ring->tail, kring->rtail);
ring->tail = kring->rtail;
}
kring->rhead = head;
kring->rcur = cur;
return head;
}
/*
* validate parameters on entry for *_rxsync()
* Returns ring->head if ok, kring->nkr_num_slots on error.
*
* For a valid configuration,
* hwcur <= head <= cur <= tail <= hwtail
*
* We only consider head and cur.
* hwcur and hwtail are reliable.
*
*/
u_int
nm_rxsync_prologue(struct netmap_kring *kring, struct netmap_ring *ring)
{
uint32_t const n = kring->nkr_num_slots;
uint32_t head, cur;
2015-07-10 05:51:36 +00:00
ND(5,"%s kc %d kt %d h %d c %d t %d",
kring->name,
kring->nr_hwcur, kring->nr_hwtail,
ring->head, ring->cur, ring->tail);
/*
* Before storing the new values, we should check they do not
* move backwards. However:
* - head is not an issue because the previous value is hwcur;
* - cur could in principle go back, however it does not matter
* because we are processing a brand new rxsync()
*/
cur = kring->rcur = ring->cur; /* read only once */
head = kring->rhead = ring->head; /* read only once */
#if 1 /* kernel sanity checks */
NM_FAIL_ON(kring->nr_hwcur >= n || kring->nr_hwtail >= n);
#endif /* kernel sanity checks */
/* user sanity checks */
if (kring->nr_hwtail >= kring->nr_hwcur) {
/* want hwcur <= rhead <= hwtail */
NM_FAIL_ON(head < kring->nr_hwcur || head > kring->nr_hwtail);
/* and also rhead <= rcur <= hwtail */
NM_FAIL_ON(cur < head || cur > kring->nr_hwtail);
} else {
/* we need rhead outside hwtail..hwcur */
NM_FAIL_ON(head < kring->nr_hwcur && head > kring->nr_hwtail);
/* two cases now: head <= hwtail or head >= hwcur */
if (head <= kring->nr_hwtail) {
/* want head <= cur <= hwtail */
NM_FAIL_ON(cur < head || cur > kring->nr_hwtail);
} else {
/* cur must be outside hwtail..head */
NM_FAIL_ON(cur < head && cur > kring->nr_hwtail);
}
}
if (ring->tail != kring->rtail) {
RD(5, "%s tail overwritten was %d need %d",
kring->name,
ring->tail, kring->rtail);
ring->tail = kring->rtail;
}
return head;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Error routine called when txsync/rxsync detects an error.
* Can't do much more than resetting head =cur = hwcur, tail = hwtail
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* Return 1 on reinit.
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
*
* This routine is only called by the upper half of the kernel.
* It only reads hwcur (which is changed only by the upper half, too)
* and hwtail (which may be changed by the lower half, but only on
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
* a tx ring and only to increase it, so any error will be recovered
* on the next call). For the above, we don't strictly need to call
* it under lock.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
int
netmap_ring_reinit(struct netmap_kring *kring)
{
struct netmap_ring *ring = kring->ring;
u_int i, lim = kring->nkr_num_slots - 1;
int errors = 0;
// XXX KASSERT nm_kr_tryget
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
RD(10, "called for %s", kring->name);
// XXX probably wrong to trust userspace
kring->rhead = ring->head;
kring->rcur = ring->cur;
kring->rtail = ring->tail;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
if (ring->cur > lim)
errors++;
if (ring->head > lim)
errors++;
if (ring->tail > lim)
errors++;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
for (i = 0; i <= lim; i++) {
u_int idx = ring->slot[i].buf_idx;
u_int len = ring->slot[i].len;
2015-07-10 05:51:36 +00:00
if (idx < 2 || idx >= kring->na->na_lut.objtotal) {
RD(5, "bad index at slot %d idx %d len %d ", i, idx, len);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
ring->slot[i].buf_idx = 0;
ring->slot[i].len = 0;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
} else if (len > NETMAP_BUF_SIZE(kring->na)) {
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
ring->slot[i].len = 0;
RD(5, "bad len at slot %d idx %d len %d", i, idx, len);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
}
if (errors) {
RD(10, "total %d errors", errors);
RD(10, "%s reinit, cur %d -> %d tail %d -> %d",
kring->name,
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
ring->cur, kring->nr_hwcur,
ring->tail, kring->nr_hwtail);
ring->head = kring->rhead = kring->nr_hwcur;
ring->cur = kring->rcur = kring->nr_hwcur;
ring->tail = kring->rtail = kring->nr_hwtail;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
return (errors ? 1 : 0);
}
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/* interpret the ringid and flags fields of an nmreq, by translating them
* into a pair of intervals of ring indices:
*
* [priv->np_txqfirst, priv->np_txqlast) and
* [priv->np_rxqfirst, priv->np_rxqlast)
*
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
int
netmap_interp_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_adapter *na = priv->np_na;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
u_int j, i = ringid & NETMAP_RING_MASK;
u_int reg = flags & NR_REG_MASK;
int excluded_direction[] = { NR_TX_RINGS_ONLY, NR_RX_RINGS_ONLY };
2015-07-10 05:51:36 +00:00
enum txrx t;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (reg == NR_REG_DEFAULT) {
/* convert from old ringid to flags */
if (ringid & NETMAP_SW_RING) {
reg = NR_REG_SW;
} else if (ringid & NETMAP_HW_RING) {
reg = NR_REG_ONE_NIC;
} else {
reg = NR_REG_ALL_NIC;
}
D("deprecated API, old ringid 0x%x -> ringid %x reg %d", ringid, i, reg);
}
if ((flags & NR_PTNETMAP_HOST) && (reg != NR_REG_ALL_NIC ||
flags & (NR_RX_RINGS_ONLY|NR_TX_RINGS_ONLY))) {
D("Error: only NR_REG_ALL_NIC supported with netmap passthrough");
return EINVAL;
}
for_rx_tx(t) {
if (flags & excluded_direction[t]) {
priv->np_qfirst[t] = priv->np_qlast[t] = 0;
continue;
}
switch (reg) {
case NR_REG_ALL_NIC:
case NR_REG_PIPE_MASTER:
case NR_REG_PIPE_SLAVE:
2015-07-10 05:51:36 +00:00
priv->np_qfirst[t] = 0;
priv->np_qlast[t] = nma_get_nrings(na, t);
ND("ALL/PIPE: %s %d %d", nm_txrx2str(t),
priv->np_qfirst[t], priv->np_qlast[t]);
break;
case NR_REG_SW:
case NR_REG_NIC_SW:
if (!(na->na_flags & NAF_HOST_RINGS)) {
D("host rings not supported");
return EINVAL;
}
2015-07-10 05:51:36 +00:00
priv->np_qfirst[t] = (reg == NR_REG_SW ?
nma_get_nrings(na, t) : 0);
priv->np_qlast[t] = nma_get_nrings(na, t) + 1;
ND("%s: %s %d %d", reg == NR_REG_SW ? "SW" : "NIC+SW",
nm_txrx2str(t),
priv->np_qfirst[t], priv->np_qlast[t]);
break;
case NR_REG_ONE_NIC:
if (i >= na->num_tx_rings && i >= na->num_rx_rings) {
D("invalid ring id %d", i);
return EINVAL;
}
2015-07-10 05:51:36 +00:00
/* if not enough rings, use the first one */
j = i;
if (j >= nma_get_nrings(na, t))
j = 0;
priv->np_qfirst[t] = j;
priv->np_qlast[t] = j + 1;
ND("ONE_NIC: %s %d %d", nm_txrx2str(t),
priv->np_qfirst[t], priv->np_qlast[t]);
break;
default:
D("invalid regif type %d", reg);
return EINVAL;
2015-07-10 05:51:36 +00:00
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
priv->np_flags = (flags & ~NR_REG_MASK) | reg;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (netmap_verbose) {
D("%s: tx [%d,%d) rx [%d,%d) id %d",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name,
2015-07-10 05:51:36 +00:00
priv->np_qfirst[NR_TX],
priv->np_qlast[NR_TX],
priv->np_qfirst[NR_RX],
priv->np_qlast[NR_RX],
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
i);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return 0;
}
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/*
* Set the ring ID. For devices with a single queue, a request
* for all rings is the same as a single ring.
*/
static int
netmap_set_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags)
{
struct netmap_adapter *na = priv->np_na;
int error;
2015-07-10 05:51:36 +00:00
enum txrx t;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
error = netmap_interp_ringid(priv, ringid, flags);
if (error) {
return error;
}
priv->np_txpoll = (ringid & NETMAP_NO_TX_POLL) ? 0 : 1;
/* optimization: count the users registered for more than
* one ring, which are the ones sleeping on the global queue.
* The default netmap_notify() callback will then
* avoid signaling the global queue if nobody is using it
*/
2015-07-10 05:51:36 +00:00
for_rx_tx(t) {
if (nm_si_user(priv, t))
na->si_users[t]++;
}
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
return 0;
}
2015-07-10 05:51:36 +00:00
static void
netmap_unset_ringid(struct netmap_priv_d *priv)
{
struct netmap_adapter *na = priv->np_na;
enum txrx t;
for_rx_tx(t) {
if (nm_si_user(priv, t))
na->si_users[t]--;
priv->np_qfirst[t] = priv->np_qlast[t] = 0;
}
priv->np_flags = 0;
priv->np_txpoll = 0;
}
/* Set the nr_pending_mode for the requested rings.
* If requested, also try to get exclusive access to the rings, provided
* the rings we want to bind are not exclusively owned by a previous bind.
2015-07-10 05:51:36 +00:00
*/
static int
netmap_krings_get(struct netmap_priv_d *priv)
2015-07-10 05:51:36 +00:00
{
struct netmap_adapter *na = priv->np_na;
u_int i;
struct netmap_kring *kring;
int excl = (priv->np_flags & NR_EXCLUSIVE);
enum txrx t;
ND("%s: grabbing tx [%d, %d) rx [%d, %d)",
na->name,
priv->np_qfirst[NR_TX],
priv->np_qlast[NR_TX],
priv->np_qfirst[NR_RX],
priv->np_qlast[NR_RX]);
/* first round: check that all the requested rings
* are neither alread exclusively owned, nor we
* want exclusive ownership when they are already in use
*/
for_rx_tx(t) {
for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
if ((kring->nr_kflags & NKR_EXCLUSIVE) ||
(kring->users && excl))
{
ND("ring %s busy", kring->name);
return EBUSY;
}
}
}
/* second round: increment usage count (possibly marking them
* as exclusive) and set the nr_pending_mode
2015-07-10 05:51:36 +00:00
*/
for_rx_tx(t) {
for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
kring->users++;
if (excl)
kring->nr_kflags |= NKR_EXCLUSIVE;
kring->nr_pending_mode = NKR_NETMAP_ON;
2015-07-10 05:51:36 +00:00
}
}
return 0;
}
/* Undo netmap_krings_get(). This is done by clearing the exclusive mode
* if was asked on regif, and unset the nr_pending_mode if we are the
* last users of the involved rings. */
2015-07-10 05:51:36 +00:00
static void
netmap_krings_put(struct netmap_priv_d *priv)
2015-07-10 05:51:36 +00:00
{
struct netmap_adapter *na = priv->np_na;
u_int i;
struct netmap_kring *kring;
int excl = (priv->np_flags & NR_EXCLUSIVE);
enum txrx t;
ND("%s: releasing tx [%d, %d) rx [%d, %d)",
na->name,
priv->np_qfirst[NR_TX],
priv->np_qlast[NR_TX],
priv->np_qfirst[NR_RX],
priv->np_qlast[MR_RX]);
for_rx_tx(t) {
for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
if (excl)
kring->nr_kflags &= ~NKR_EXCLUSIVE;
kring->users--;
if (kring->users == 0)
kring->nr_pending_mode = NKR_NETMAP_OFF;
2015-07-10 05:51:36 +00:00
}
}
}
/*
* possibly move the interface to netmap-mode.
* If success it returns a pointer to netmap_if, otherwise NULL.
* This must be called with NMG_LOCK held.
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*
* The following na callbacks are called in the process:
*
* na->nm_config() [by netmap_update_config]
* (get current number and size of rings)
*
* We have a generic one for linux (netmap_linux_config).
* The bwrap has to override this, since it has to forward
* the request to the wrapped adapter (netmap_bwrap_config).
*
*
2015-07-10 05:51:36 +00:00
* na->nm_krings_create()
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* (create and init the krings array)
*
* One of the following:
*
* * netmap_hw_krings_create, (hw ports)
* creates the standard layout for the krings
* and adds the mbq (used for the host rings).
*
* * netmap_vp_krings_create (VALE ports)
* add leases and scratchpads
*
* * netmap_pipe_krings_create (pipes)
* create the krings and rings of both ends and
* cross-link them
*
* * netmap_monitor_krings_create (monitors)
* avoid allocating the mbq
*
* * netmap_bwrap_krings_create (bwraps)
* create both the brap krings array,
* the krings array of the wrapped adapter, and
* (if needed) the fake array for the host adapter
*
* na->nm_register(, 1)
* (put the adapter in netmap mode)
*
* This may be one of the following:
*
* * netmap_hw_reg (hw ports)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* checks that the ifp is still there, then calls
* the hardware specific callback;
*
* * netmap_vp_reg (VALE ports)
* If the port is connected to a bridge,
* set the NAF_NETMAP_ON flag under the
* bridge write lock.
*
* * netmap_pipe_reg (pipes)
* inform the other pipe end that it is no
* longer responsible for the lifetime of this
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* pipe end
*
* * netmap_monitor_reg (monitors)
* intercept the sync callbacks of the monitored
* rings
*
* * netmap_bwrap_reg (bwraps)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* cross-link the bwrap and hwna rings,
* forward the request to the hwna, override
* the hwna notify callback (to get the frames
* coming from outside go through the bridge).
*
*
*/
2015-07-10 05:51:36 +00:00
int
netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na,
2015-07-10 05:51:36 +00:00
uint16_t ringid, uint32_t flags)
{
struct netmap_if *nifp = NULL;
2015-07-10 05:51:36 +00:00
int error;
NMG_LOCK_ASSERT();
/* ring configuration may have changed, fetch from the card */
netmap_update_config(na);
priv->np_na = na; /* store the reference */
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
error = netmap_set_ringid(priv, ringid, flags);
if (error)
2015-07-10 05:51:36 +00:00
goto err;
error = netmap_mem_finalize(na->nm_mem, na);
if (error)
goto err;
if (na->active_fds == 0) {
/*
* If this is the first registration of the adapter,
* create the in-kernel view of the netmap rings,
2015-07-10 05:51:36 +00:00
* the netmap krings.
*/
/*
* Depending on the adapter, this may also create
* the netmap rings themselves
*/
error = na->nm_krings_create(na);
if (error)
goto err_drop_mem;
}
2015-07-10 05:51:36 +00:00
/* now the krings must exist and we can check whether some
* previous bind has exclusive ownership on them, and set
* nr_pending_mode
2015-07-10 05:51:36 +00:00
*/
error = netmap_krings_get(priv);
2015-07-10 05:51:36 +00:00
if (error)
goto err_del_krings;
/* create all needed missing netmap rings */
error = netmap_mem_rings_create(na);
if (error)
goto err_rel_excl;
2015-07-10 05:51:36 +00:00
/* in all cases, create a new netmap if */
nifp = netmap_mem_if_new(na);
if (nifp == NULL) {
error = ENOMEM;
goto err_del_rings;
}
2015-07-10 05:51:36 +00:00
if (na->active_fds == 0) {
/* cache the allocator info in the na */
error = netmap_mem_get_lut(na->nm_mem, &na->na_lut);
if (error)
2015-07-10 05:51:36 +00:00
goto err_del_if;
ND("lut %p bufs %u size %u", na->na_lut.lut, na->na_lut.objtotal,
na->na_lut.objsize);
}
if (nm_kring_pending(priv)) {
/* Some kring is switching mode, tell the adapter to
* react on this. */
error = na->nm_register(na, 1);
if (error)
goto err_put_lut;
}
2015-07-10 05:51:36 +00:00
/* Commit the reference. */
na->active_fds++;
2015-07-10 05:51:36 +00:00
/*
* advertise that the interface is ready by setting np_nifp.
* The barrier is needed because readers (poll, *SYNC and mmap)
* check for priv->np_nifp != NULL without locking
*/
mb(); /* make sure previous writes are visible to all CPUs */
priv->np_nifp = nifp;
return 0;
err_put_lut:
if (na->active_fds == 0)
memset(&na->na_lut, 0, sizeof(na->na_lut));
2015-07-10 05:51:36 +00:00
err_del_if:
netmap_mem_if_delete(na, nifp);
err_rel_excl:
netmap_krings_put(priv);
2015-07-10 05:51:36 +00:00
err_del_rings:
netmap_mem_rings_delete(na);
2015-07-10 05:51:36 +00:00
err_del_krings:
if (na->active_fds == 0)
na->nm_krings_delete(na);
err_drop_mem:
netmap_mem_deref(na->nm_mem, na);
err:
priv->np_na = NULL;
return error;
}
/*
* update kring and ring at the end of rxsync/txsync.
2015-07-10 05:51:36 +00:00
*/
static inline void
nm_sync_finalize(struct netmap_kring *kring)
2015-07-10 05:51:36 +00:00
{
/*
* Update ring tail to what the kernel knows
* After txsync: head/rhead/hwcur might be behind cur/rcur
* if no carrier.
*/
2015-07-10 05:51:36 +00:00
kring->ring->tail = kring->rtail = kring->nr_hwtail;
ND(5, "%s now hwcur %d hwtail %d head %d cur %d tail %d",
kring->name, kring->nr_hwcur, kring->nr_hwtail,
kring->rhead, kring->rcur, kring->rtail);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* ioctl(2) support for the "netmap" device.
*
* Following a list of accepted commands:
* - NIOCGINFO
* - SIOCGIFADDR just for convenience
* - NIOCREGIF
* - NIOCTXSYNC
* - NIOCRXSYNC
*
* Return 0 on success, errno otherwise.
*/
int
netmap_ioctl(struct netmap_priv_d *priv, u_long cmd, caddr_t data, struct thread *td)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct nmreq *nmr = (struct nmreq *) data;
struct netmap_adapter *na = NULL;
struct ifnet *ifp = NULL;
int error = 0;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
u_int i, qfirst, qlast;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
struct netmap_if *nifp;
struct netmap_kring *krings;
2015-07-10 05:51:36 +00:00
enum txrx t;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
if (cmd == NIOCGINFO || cmd == NIOCREGIF) {
/* truncate name */
nmr->nr_name[sizeof(nmr->nr_name) - 1] = '\0';
if (nmr->nr_version != NETMAP_API) {
D("API mismatch for %s got %d need %d",
nmr->nr_name,
nmr->nr_version, NETMAP_API);
nmr->nr_version = NETMAP_API;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
}
if (nmr->nr_version < NETMAP_MIN_API ||
nmr->nr_version > NETMAP_MAX_API) {
return EINVAL;
}
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
switch (cmd) {
case NIOCGINFO: /* return capabilities etc */
if (nmr->nr_cmd == NETMAP_BDG_LIST) {
error = netmap_bdg_ctl(nmr, NULL);
break;
}
NMG_LOCK();
do {
/* memsize is always valid */
struct netmap_mem_d *nmd = &nm_mem;
u_int memflags;
if (nmr->nr_name[0] != '\0') {
/* get a refcount */
error = netmap_get_na(nmr, &na, &ifp, 1 /* create */);
if (error) {
na = NULL;
ifp = NULL;
break;
}
nmd = na->nm_mem; /* get memory allocator */
}
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
error = netmap_mem_get_info(nmd, &nmr->nr_memsize, &memflags,
&nmr->nr_arg2);
if (error)
break;
if (na == NULL) /* only memory info */
break;
nmr->nr_offset = 0;
nmr->nr_rx_slots = nmr->nr_tx_slots = 0;
netmap_update_config(na);
nmr->nr_rx_rings = na->num_rx_rings;
nmr->nr_tx_rings = na->num_tx_rings;
nmr->nr_rx_slots = na->num_rx_desc;
nmr->nr_tx_slots = na->num_tx_desc;
} while (0);
netmap_unget_na(na, ifp);
NMG_UNLOCK();
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
break;
case NIOCREGIF:
/*
* If nmr->nr_cmd is not zero, this NIOCREGIF is not really
* a regif operation, but a different one, specified by the
* value of nmr->nr_cmd.
*/
i = nmr->nr_cmd;
if (i == NETMAP_BDG_ATTACH || i == NETMAP_BDG_DETACH
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
|| i == NETMAP_BDG_VNET_HDR
|| i == NETMAP_BDG_NEWIF
|| i == NETMAP_BDG_DELIF
|| i == NETMAP_BDG_POLLING_ON
|| i == NETMAP_BDG_POLLING_OFF) {
/* possibly attach/detach NIC and VALE switch */
error = netmap_bdg_ctl(nmr, NULL);
break;
} else if (i == NETMAP_PT_HOST_CREATE || i == NETMAP_PT_HOST_DELETE) {
/* forward the command to the ptnetmap subsystem */
error = ptnetmap_ctl(nmr, priv->np_na);
break;
} else if (i == NETMAP_VNET_HDR_GET) {
/* get vnet-header length for this netmap port */
struct ifnet *ifp;
NMG_LOCK();
error = netmap_get_na(nmr, &na, &ifp, 0);
if (na && !error) {
nmr->nr_arg1 = na->virt_hdr_len;
}
netmap_unget_na(na, ifp);
NMG_UNLOCK();
break;
} else if (i == NETMAP_POOLS_INFO_GET) {
/* get information from the memory allocator */
error = netmap_mem_pools_info_get(nmr, priv->np_na);
break;
} else if (i != 0) {
D("nr_cmd must be 0 not %d", i);
error = EINVAL;
break;
}
/* protect access to priv from concurrent NIOCREGIF */
NMG_LOCK();
do {
u_int memflags;
struct ifnet *ifp;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
2015-07-10 05:51:36 +00:00
if (priv->np_nifp != NULL) { /* thread already registered */
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
error = EBUSY;
break;
}
/* find the interface and a reference */
error = netmap_get_na(nmr, &na, &ifp,
1 /* create */); /* keep reference */
if (error)
break;
if (NETMAP_OWNED_BY_KERN(na)) {
netmap_unget_na(na, ifp);
error = EBUSY;
break;
}
if (na->virt_hdr_len && !(nmr->nr_flags & NR_ACCEPT_VNET_HDR)) {
netmap_unget_na(na, ifp);
error = EIO;
break;
}
2015-07-10 05:51:36 +00:00
error = netmap_do_regif(priv, na, nmr->nr_ringid, nmr->nr_flags);
if (error) { /* reg. failed, release priv and ref */
netmap_unget_na(na, ifp);
break;
}
2015-07-10 05:51:36 +00:00
nifp = priv->np_nifp;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
priv->np_td = td; // XXX kqueue, debugging only
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/* return the offset of the netmap_if object */
nmr->nr_rx_rings = na->num_rx_rings;
nmr->nr_tx_rings = na->num_tx_rings;
nmr->nr_rx_slots = na->num_rx_desc;
nmr->nr_tx_slots = na->num_tx_desc;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
error = netmap_mem_get_info(na->nm_mem, &nmr->nr_memsize, &memflags,
&nmr->nr_arg2);
if (error) {
2015-07-10 05:51:36 +00:00
netmap_do_unregif(priv);
netmap_unget_na(na, ifp);
break;
}
if (memflags & NETMAP_MEM_PRIVATE) {
*(uint32_t *)(uintptr_t)&nifp->ni_flags |= NI_PRIV_MEM;
}
2015-07-10 05:51:36 +00:00
for_rx_tx(t) {
priv->np_si[t] = nm_si_user(priv, t) ?
&na->si[t] : &NMR(na, t)[priv->np_qfirst[t]].si;
}
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (nmr->nr_arg3) {
if (netmap_verbose)
D("requested %d extra buffers", nmr->nr_arg3);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
nmr->nr_arg3 = netmap_extra_alloc(na,
&nifp->ni_bufs_head, nmr->nr_arg3);
if (netmap_verbose)
D("got %d extra buffers", nmr->nr_arg3);
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
}
nmr->nr_offset = netmap_mem_if_offset(na->nm_mem, nifp);
/* store ifp reference so that priv destructor may release it */
priv->np_ifp = ifp;
} while (0);
NMG_UNLOCK();
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
break;
case NIOCTXSYNC:
case NIOCRXSYNC:
nifp = priv->np_nifp;
if (nifp == NULL) {
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
error = ENXIO;
break;
}
mb(); /* make sure following reads are not from cache */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
na = priv->np_na; /* we have a reference */
if (na == NULL) {
D("Internal error: nifp != NULL && na == NULL");
error = ENXIO;
break;
}
2015-07-10 05:51:36 +00:00
t = (cmd == NIOCTXSYNC ? NR_TX : NR_RX);
krings = NMR(na, t);
qfirst = priv->np_qfirst[t];
qlast = priv->np_qlast[t];
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
for (i = qfirst; i < qlast; i++) {
struct netmap_kring *kring = krings + i;
struct netmap_ring *ring = kring->ring;
if (unlikely(nm_kr_tryget(kring, 1, &error))) {
error = (error ? EIO : 0);
continue;
}
if (cmd == NIOCTXSYNC) {
if (netmap_verbose & NM_VERB_TXSYNC)
D("pre txsync ring %d cur %d hwcur %d",
i, ring->cur,
kring->nr_hwcur);
if (nm_txsync_prologue(kring, ring) >= kring->nkr_num_slots) {
netmap_ring_reinit(kring);
2015-07-10 05:51:36 +00:00
} else if (kring->nm_sync(kring, NAF_FORCE_RECLAIM) == 0) {
nm_sync_finalize(kring);
}
if (netmap_verbose & NM_VERB_TXSYNC)
D("post txsync ring %d cur %d hwcur %d",
i, ring->cur,
kring->nr_hwcur);
} else {
if (nm_rxsync_prologue(kring, ring) >= kring->nkr_num_slots) {
2015-07-10 05:51:36 +00:00
netmap_ring_reinit(kring);
} else if (kring->nm_sync(kring, NAF_FORCE_READ) == 0) {
nm_sync_finalize(kring);
2015-07-10 05:51:36 +00:00
}
microtime(&ring->ts);
}
nm_kr_put(kring);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
break;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
2015-07-10 05:51:36 +00:00
#ifdef WITH_VALE
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
case NIOCCONFIG:
error = netmap_bdg_config(nmr);
break;
2015-07-10 05:51:36 +00:00
#endif
#ifdef __FreeBSD__
case FIONBIO:
case FIOASYNC:
ND("FIONBIO/FIOASYNC are no-ops");
break;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
case BIOCIMMEDIATE:
case BIOCGHDRCMPLT:
case BIOCSHDRCMPLT:
case BIOCSSEESENT:
D("ignore BIOCIMMEDIATE/BIOCSHDRCMPLT/BIOCSHDRCMPLT/BIOCSSEESENT");
break;
default: /* allow device-specific ioctls */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct ifnet *ifp = ifunit_ref(nmr->nr_name);
if (ifp == NULL) {
error = ENXIO;
} else {
struct socket so;
bzero(&so, sizeof(so));
so.so_vnet = ifp->if_vnet;
// so->so_proto not null.
error = ifioctl(&so, cmd, data, td);
if_rele(ifp);
}
break;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
#else /* linux */
default:
error = EOPNOTSUPP;
#endif /* linux */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
return (error);
}
/*
* select(2) and poll(2) handlers for the "netmap" device.
*
* Can be called for one or more queues.
* Return true the event mask corresponding to ready events.
* If there are no ready events, do a selrecord on either individual
* selinfo or on the global one.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* Device-dependent parts (locking and sync of tx/rx rings)
* are done through callbacks.
*
* On linux, arguments are really pwait, the poll table, and 'td' is struct file *
* The first one is remapped to pwait as selrecord() uses the name as an
* hidden argument.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
int
netmap_poll(struct netmap_priv_d *priv, int events, NM_SELRECORD_T *sr)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_adapter *na;
struct netmap_kring *kring;
struct netmap_ring *ring;
2015-07-10 05:51:36 +00:00
u_int i, check_all_tx, check_all_rx, want[NR_TXRX], revents = 0;
#define want_tx want[NR_TX]
#define want_rx want[NR_RX]
struct mbq q; /* packets from hw queues to host stack */
2015-07-10 05:51:36 +00:00
enum txrx t;
/*
* In order to avoid nested locks, we need to "double check"
* txsync and rxsync if we decide to do a selrecord().
* retry_tx (and retry_rx, later) prevent looping forever.
*/
int retry_tx = 1, retry_rx = 1;
/* transparent mode: send_down is 1 if we have found some
* packets to forward during the rx scan and we have not
* sent them down to the nic yet
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
*/
int send_down = 0;
mbq_init(&q);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
if (priv->np_nifp == NULL) {
D("No if registered");
return POLLERR;
}
2015-07-10 05:51:36 +00:00
mb(); /* make sure following reads are not from cache */
na = priv->np_na;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (!nm_netmap_on(na))
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return POLLERR;
if (netmap_verbose & 0x8000)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
D("device %s events 0x%x", na->name, events);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
want_tx = events & (POLLOUT | POLLWRNORM);
want_rx = events & (POLLIN | POLLRDNORM);
/*
* check_all_{tx|rx} are set if the card has more than one queue AND
* the file descriptor is bound to all of them. If so, we sleep on
* the "global" selinfo, otherwise we sleep on individual selinfo
* (FreeBSD only allows two selinfo's per file descriptor).
* The interrupt routine in the driver wake one or the other
* (or both) depending on which clients are active.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*
* rxsync() is only called if we run out of buffers on a POLLIN.
* txsync() is called if we run out of buffers on POLLOUT, or
* there are pending packets to send. The latter can be disabled
* passing NETMAP_NO_TX_POLL in the NIOCREG call.
*/
2015-07-10 05:51:36 +00:00
check_all_tx = nm_si_user(priv, NR_TX);
check_all_rx = nm_si_user(priv, NR_RX);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* We start with a lock free round which is cheap if we have
* slots available. If this fails, then lock and call the sync
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* routines.
*/
#if 1 /* new code- call rx if any of the ring needs to release or read buffers */
if (want_tx) {
t = NR_TX;
for (i = priv->np_qfirst[t]; want[t] && i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
/* XXX compare ring->cur and kring->tail */
if (!nm_ring_empty(kring->ring)) {
revents |= want[t];
want[t] = 0; /* also breaks the loop */
}
}
}
if (want_rx) {
want_rx = 0; /* look for a reason to run the handlers */
t = NR_RX;
for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
if (kring->ring->cur == kring->ring->tail /* try fetch new buffers */
|| kring->rhead != kring->ring->head /* release buffers */) {
want_rx = 1;
}
}
if (!want_rx)
revents |= events & (POLLIN | POLLRDNORM); /* we have data */
}
#else /* old code */
2015-07-10 05:51:36 +00:00
for_rx_tx(t) {
for (i = priv->np_qfirst[t]; want[t] && i < priv->np_qlast[t]; i++) {
kring = &NMR(na, t)[i];
/* XXX compare ring->cur and kring->tail */
if (!nm_ring_empty(kring->ring)) {
revents |= want[t];
want[t] = 0; /* also breaks the loop */
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
}
#endif /* old code */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* If we want to push packets out (priv->np_txpoll) or
* want_tx is still set, we must issue txsync calls
* (on all rings, to avoid that the tx rings stall).
* XXX should also check cur != hwcur on the tx rings.
* Fortunately, normal tx mode has np_txpoll set.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
if (priv->np_txpoll || want_tx) {
/*
* The first round checks if anyone is ready, if not
* do a selrecord and another round to handle races.
* want_tx goes to 0 if any space is found, and is
* used to skip rings with no pending transmissions.
*/
flush_tx:
for (i = priv->np_qfirst[NR_TX]; i < priv->np_qlast[NR_TX]; i++) {
int found = 0;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
kring = &na->tx_rings[i];
ring = kring->ring;
if (!send_down && !want_tx && ring->cur == kring->nr_hwcur)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
continue;
if (nm_kr_tryget(kring, 1, &revents))
continue;
if (nm_txsync_prologue(kring, ring) >= kring->nkr_num_slots) {
netmap_ring_reinit(kring);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
revents |= POLLERR;
} else {
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (kring->nm_sync(kring, 0))
revents |= POLLERR;
2015-07-10 05:51:36 +00:00
else
nm_sync_finalize(kring);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* If we found new slots, notify potential
* listeners on the same ring.
* Since we just did a txsync, look at the copies
* of cur,tail in the kring.
*/
found = kring->rcur != kring->rtail;
nm_kr_put(kring);
if (found) { /* notify other listeners */
revents |= want_tx;
want_tx = 0;
2015-07-10 05:51:36 +00:00
kring->nm_notify(kring, 0);
}
}
/* if there were any packet to forward we must have handled them by now */
send_down = 0;
if (want_tx && retry_tx && sr) {
nm_os_selrecord(sr, check_all_tx ?
2015-07-10 05:51:36 +00:00
&na->si[NR_TX] : &na->tx_rings[priv->np_qfirst[NR_TX]].si);
retry_tx = 0;
goto flush_tx;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
}
/*
* If want_rx is still set scan receive rings.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* Do it on all rings because otherwise we starve.
*/
if (want_rx) {
/* two rounds here for race avoidance */
do_retry_rx:
2015-07-10 05:51:36 +00:00
for (i = priv->np_qfirst[NR_RX]; i < priv->np_qlast[NR_RX]; i++) {
int found = 0;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
kring = &na->rx_rings[i];
ring = kring->ring;
if (unlikely(nm_kr_tryget(kring, 1, &revents)))
continue;
if (nm_rxsync_prologue(kring, ring) >= kring->nkr_num_slots) {
2015-07-10 05:51:36 +00:00
netmap_ring_reinit(kring);
revents |= POLLERR;
}
/* now we can use kring->rcur, rtail */
/*
* transparent mode support: collect packets
* from the rxring(s).
*/
if (nm_may_forward_up(kring)) {
ND(10, "forwarding some buffers up %d to %d",
kring->nr_hwcur, ring->cur);
netmap_grab_packets(kring, &q, netmap_fwd);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
kring->nr_kflags &= ~NR_FORWARD;
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
if (kring->nm_sync(kring, 0))
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
revents |= POLLERR;
2015-07-10 05:51:36 +00:00
else
nm_sync_finalize(kring);
send_down |= (kring->nr_kflags & NR_FORWARD); /* host ring only */
if (netmap_no_timestamp == 0 ||
ring->flags & NR_TIMESTAMP) {
microtime(&ring->ts);
}
found = kring->rcur != kring->rtail;
nm_kr_put(kring);
if (found) {
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
revents |= want_rx;
retry_rx = 0;
2015-07-10 05:51:36 +00:00
kring->nm_notify(kring, 0);
}
}
if (retry_rx && sr) {
nm_os_selrecord(sr, check_all_rx ?
2015-07-10 05:51:36 +00:00
&na->si[NR_RX] : &na->rx_rings[priv->np_qfirst[NR_RX]].si);
}
if (send_down > 0 || retry_rx) {
retry_rx = 0;
if (send_down)
goto flush_tx; /* and retry_rx */
else
goto do_retry_rx;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
}
/*
* Transparent mode: marked bufs on rx rings between
* kring->nr_hwcur and ring->head
* are passed to the other endpoint.
*
* Transparent mode requires to bind all
* rings to a single file descriptor.
*/
if (q.head && !nm_kr_tryget(&na->tx_rings[na->num_tx_rings], 1, &revents)) {
netmap_send_up(na->ifp, &q);
nm_kr_put(&na->tx_rings[na->num_tx_rings]);
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return (revents);
2015-07-10 05:51:36 +00:00
#undef want_tx
#undef want_rx
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
/*-------------------- driver support routines -------------------*/
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/* default notify callback */
static int
2015-07-10 05:51:36 +00:00
netmap_notify(struct netmap_kring *kring, int flags)
{
2015-07-10 05:51:36 +00:00
struct netmap_adapter *na = kring->na;
enum txrx t = kring->tx;
nm_os_selwakeup(&kring->si);
2015-07-10 05:51:36 +00:00
/* optimization: avoid a wake up on the global
* queue if nobody has registered for more
* than one ring
*/
if (na->si_users[t] > 0)
nm_os_selwakeup(&na->si[t]);
return NM_IRQ_COMPLETED;
}
#if 0
static int
netmap_notify(struct netmap_adapter *na, u_int n_ring,
enum txrx tx, int flags)
{
if (tx == NR_TX) {
KeSetEvent(notes->TX_EVENT, 0, FALSE);
}
else
{
KeSetEvent(notes->RX_EVENT, 0, FALSE);
}
return 0;
}
#endif
/* called by all routines that create netmap_adapters.
* provide some defaults and get a reference to the
* memory allocator
*/
int
netmap_attach_common(struct netmap_adapter *na)
{
if (na->num_tx_rings == 0 || na->num_rx_rings == 0) {
D("%s: invalid rings tx %d rx %d",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name, na->num_tx_rings, na->num_rx_rings);
return EINVAL;
}
#ifdef __FreeBSD__
if (na->na_flags & NAF_HOST_RINGS && na->ifp) {
na->if_input = na->ifp->if_input; /* for netmap_send_up */
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
}
#endif /* __FreeBSD__ */
if (na->nm_krings_create == NULL) {
/* we assume that we have been called by a driver,
* since other port types all provide their own
* nm_krings_create
*/
na->nm_krings_create = netmap_hw_krings_create;
na->nm_krings_delete = netmap_hw_krings_delete;
}
if (na->nm_notify == NULL)
na->nm_notify = netmap_notify;
na->active_fds = 0;
if (na->nm_mem == NULL)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
/* use the global allocator */
na->nm_mem = &nm_mem;
2015-07-10 05:51:36 +00:00
netmap_mem_get(na->nm_mem);
#ifdef WITH_VALE
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (na->nm_bdg_attach == NULL)
/* no special nm_bdg_attach callback. On VALE
* attach, we need to interpose a bwrap
*/
na->nm_bdg_attach = netmap_bwrap_attach;
2015-07-10 05:51:36 +00:00
#endif
return 0;
}
/* standard cleanup, called by all destructors */
void
netmap_detach_common(struct netmap_adapter *na)
{
if (na->tx_rings) { /* XXX should not happen */
D("freeing leftover tx_rings");
na->nm_krings_delete(na);
}
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
netmap_pipe_dealloc(na);
2015-07-10 05:51:36 +00:00
if (na->nm_mem)
netmap_mem_put(na->nm_mem);
bzero(na, sizeof(*na));
free(na, M_DEVBUF);
}
/* Wrapper for the register callback provided netmap-enabled
* hardware drivers.
* nm_iszombie(na) means that the driver module has been
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* unloaded, so we cannot call into it.
* nm_os_ifnet_lock() must guarantee mutual exclusion with
* module unloading.
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
*/
static int
netmap_hw_reg(struct netmap_adapter *na, int onoff)
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
{
struct netmap_hw_adapter *hwna =
(struct netmap_hw_adapter*)na;
int error = 0;
nm_os_ifnet_lock();
if (nm_iszombie(na)) {
if (onoff) {
error = ENXIO;
} else if (na != NULL) {
na->na_flags &= ~NAF_NETMAP_ON;
}
goto out;
}
error = hwna->nm_hw_register(na, onoff);
out:
nm_os_ifnet_unlock();
return error;
}
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
static void
netmap_hw_dtor(struct netmap_adapter *na)
{
if (nm_iszombie(na) || na->ifp == NULL)
return;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
WNA(na->ifp) = NULL;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Allocate a ``netmap_adapter`` object, and initialize it from the
* 'arg' passed by the driver on attach.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* We allocate a block of memory with room for a struct netmap_adapter
* plus two sets of N+2 struct netmap_kring (where N is the number
* of hardware rings):
* krings 0..N-1 are for the hardware queues.
* kring N is for the host stack queue
* kring N+1 is only used for the selinfo for all queues. // XXX still true ?
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
* Return 0 on success, ENOMEM otherwise.
*/
static int
_netmap_attach(struct netmap_adapter *arg, size_t size)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_hw_adapter *hwna = NULL;
struct ifnet *ifp = NULL;
if (arg == NULL || arg->ifp == NULL)
goto fail;
ifp = arg->ifp;
hwna = malloc(size, M_DEVBUF, M_NOWAIT | M_ZERO);
if (hwna == NULL)
goto fail;
hwna->up = *arg;
2015-07-10 05:51:36 +00:00
hwna->up.na_flags |= NAF_HOST_RINGS | NAF_NATIVE;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
strncpy(hwna->up.name, ifp->if_xname, sizeof(hwna->up.name));
hwna->nm_hw_register = hwna->up.nm_register;
hwna->up.nm_register = netmap_hw_reg;
if (netmap_attach_common(&hwna->up)) {
free(hwna, M_DEVBUF);
goto fail;
}
netmap_adapter_get(&hwna->up);
NM_ATTACH_NA(ifp, &hwna->up);
#ifdef linux
if (ifp->netdev_ops) {
/* prepare a clone of the netdev ops */
2015-07-10 05:51:36 +00:00
#ifndef NETMAP_LINUX_HAVE_NETDEV_OPS
hwna->nm_ndo.ndo_start_xmit = ifp->netdev_ops;
#else
hwna->nm_ndo = *ifp->netdev_ops;
#endif /* NETMAP_LINUX_HAVE_NETDEV_OPS */
}
hwna->nm_ndo.ndo_start_xmit = linux_netmap_start_xmit;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (ifp->ethtool_ops) {
hwna->nm_eto = *ifp->ethtool_ops;
}
hwna->nm_eto.set_ringparam = linux_netmap_set_ringparam;
2015-07-10 05:51:36 +00:00
#ifdef NETMAP_LINUX_HAVE_SET_CHANNELS
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
hwna->nm_eto.set_channels = linux_netmap_set_channels;
#endif /* NETMAP_LINUX_HAVE_SET_CHANNELS */
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (arg->nm_config == NULL) {
hwna->up.nm_config = netmap_linux_config;
}
#endif /* linux */
if (arg->nm_dtor == NULL) {
hwna->up.nm_dtor = netmap_hw_dtor;
}
if_printf(ifp, "netmap queues/slots: TX %d/%d, RX %d/%d\n",
hwna->up.num_tx_rings, hwna->up.num_tx_desc,
hwna->up.num_rx_rings, hwna->up.num_rx_desc);
return 0;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
fail:
D("fail, arg %p ifp %p na %p", arg, ifp, hwna);
return (hwna ? EINVAL : ENOMEM);
}
int
netmap_attach(struct netmap_adapter *arg)
{
return _netmap_attach(arg, sizeof(struct netmap_hw_adapter));
}
#ifdef WITH_PTNETMAP_GUEST
int
netmap_pt_guest_attach(struct netmap_adapter *arg, void *csb,
unsigned int nifp_offset, unsigned int memid)
{
struct netmap_pt_guest_adapter *ptna;
struct ifnet *ifp = arg ? arg->ifp : NULL;
int error;
/* get allocator */
arg->nm_mem = netmap_mem_pt_guest_new(ifp, nifp_offset, memid);
if (arg->nm_mem == NULL)
return ENOMEM;
arg->na_flags |= NAF_MEM_OWNER;
error = _netmap_attach(arg, sizeof(struct netmap_pt_guest_adapter));
if (error)
return error;
/* get the netmap_pt_guest_adapter */
ptna = (struct netmap_pt_guest_adapter *) NA(ifp);
ptna->csb = csb;
/* Initialize a separate pass-through netmap adapter that is going to
* be used by the ptnet driver only, and so never exposed to netmap
* applications. We only need a subset of the available fields. */
memset(&ptna->dr, 0, sizeof(ptna->dr));
ptna->dr.up.ifp = ifp;
ptna->dr.up.nm_mem = ptna->hwup.up.nm_mem;
netmap_mem_get(ptna->dr.up.nm_mem);
ptna->dr.up.nm_config = ptna->hwup.up.nm_config;
ptna->backend_regifs = 0;
return 0;
}
#endif /* WITH_PTNETMAP_GUEST */
void
NM_DBG(netmap_adapter_get)(struct netmap_adapter *na)
{
if (!na) {
return;
}
refcount_acquire(&na->na_refcount);
}
/* returns 1 iff the netmap_adapter is destroyed */
int
NM_DBG(netmap_adapter_put)(struct netmap_adapter *na)
{
if (!na)
return 1;
if (!refcount_release(&na->na_refcount))
return 0;
if (na->nm_dtor)
na->nm_dtor(na);
netmap_detach_common(na);
return 1;
}
/* nm_krings_create callback for all hardware native adapters */
int
netmap_hw_krings_create(struct netmap_adapter *na)
{
This new version of netmap brings you the following: - netmap pipes, providing bidirectional blocking I/O while moving 100+ Mpps between processes using shared memory channels (no mistake: over one hundred million. But mind you, i said *moving* not *processing*); - kqueue support (BHyVe needs it); - improved user library. Just the interface name lets you select a NIC, host port, VALE switch port, netmap pipe, and individual queues. The upcoming netmap-enabled libpcap will use this feature. - optional extra buffers associated to netmap ports, for applications that need to buffer data yet don't want to make copies. - segmentation offloading for the VALE switch, useful between VMs. and a number of bug fixes and performance improvements. My colleagues Giuseppe Lettieri and Vincenzo Maffione did a substantial amount of work on these features so we owe them a big thanks. There are some external repositories that can be of interest: https://code.google.com/p/netmap our public repository for netmap/VALE code, including linux versions and other stuff that does not belong here, such as python bindings. https://code.google.com/p/netmap-libpcap a clone of the libpcap repository with netmap support. With this any libpcap client has access to most netmap feature with no recompilation. E.g. tcpdump can filter packets at 10-15 Mpps. https://code.google.com/p/netmap-ipfw a userspace version of ipfw+dummynet which uses netmap to send/receive packets. Speed is up in the 7-10 Mpps range per core for simple rulesets. Both netmap-libpcap and netmap-ipfw will be merged upstream at some point, but while this happens it is useful to have access to them. And yes, this code will be merged soon. It is infinitely better than the version currently in 10 and 9. MFC after: 3 days
2014-02-15 04:53:04 +00:00
int ret = netmap_krings_create(na, 0);
if (ret == 0) {
/* initialize the mbq for the sw rx ring */
mbq_safe_init(&na->rx_rings[na->num_rx_rings].rx_queue);
ND("initialized sw rx queue %d", na->num_rx_rings);
}
return ret;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Called on module unload by the netmap-enabled drivers
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
void
netmap_detach(struct ifnet *ifp)
{
struct netmap_adapter *na = NA(ifp);
if (!na)
return;
NMG_LOCK();
netmap_set_all_rings(na, NM_KR_LOCKED);
na->na_flags |= NAF_ZOMBIE;
2015-07-10 05:51:36 +00:00
/*
* if the netmap adapter is not native, somebody
* changed it, so we can not release it here.
* The NAF_ZOMBIE flag will notify the new owner that
2015-07-10 05:51:36 +00:00
* the driver is gone.
*/
if (na->na_flags & NAF_NATIVE) {
netmap_adapter_put(na);
}
/* give active users a chance to notice that NAF_ZOMBIE has been
* turned on, so that they can stop and return an error to userspace.
* Note that this becomes a NOP if there are no active users and,
* therefore, the put() above has deleted the na, since now NA(ifp) is
* NULL.
*/
netmap_enable_all_rings(ifp);
NMG_UNLOCK();
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}
/*
* Intercept packets from the network stack and pass them
* to netmap as incoming packets on the 'software' ring.
*
* We only store packets in a bounded mbq and then copy them
* in the relevant rxsync routine.
*
* We rely on the OS to make sure that the ifp and na do not go
* away (typically the caller checks for IFF_DRV_RUNNING or the like).
* In nm_register() or whenever there is a reinitialization,
* we make sure to make the mode change visible here.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
int
netmap_transmit(struct ifnet *ifp, struct mbuf *m)
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
{
struct netmap_adapter *na = NA(ifp);
struct netmap_kring *kring, *tx_kring;
u_int len = MBUF_LEN(m);
u_int error = ENOBUFS;
unsigned int txr;
struct mbq *q;
int space;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
2015-07-10 05:51:36 +00:00
kring = &na->rx_rings[na->num_rx_rings];
// XXX [Linux] we do not need this lock
// if we follow the down/configure/up protocol -gl
// mtx_lock(&na->core_lock);
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (!nm_netmap_on(na)) {
D("%s not in netmap mode anymore", na->name);
error = ENXIO;
goto done;
}
txr = MBUF_TXQ(m);
if (txr >= na->num_tx_rings) {
txr %= na->num_tx_rings;
}
tx_kring = &NMR(na, NR_TX)[txr];
if (tx_kring->nr_mode == NKR_NETMAP_OFF) {
return MBUF_TRANSMIT(na, ifp, m);
}
q = &kring->rx_queue;
// XXX reconsider long packets if we handle fragments
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (len > NETMAP_BUF_SIZE(na)) { /* too long for us */
D("%s from_host, drop packet size %d > %d", na->name,
len, NETMAP_BUF_SIZE(na));
goto done;
}
if (nm_os_mbuf_has_offld(m)) {
RD(1, "%s drop mbuf requiring offloadings", na->name);
goto done;
}
/* protect against rxsync_from_host(), netmap_sw_to_nic()
* and maybe other instances of netmap_transmit (the latter
* not possible on Linux).
* Also avoid overflowing the queue.
*/
mbq_lock(q);
space = kring->nr_hwtail - kring->nr_hwcur;
if (space < 0)
space += kring->nkr_num_slots;
if (space + mbq_len(q) >= kring->nkr_num_slots - 1) { // XXX
RD(10, "%s full hwcur %d hwtail %d qlen %d len %d m %p",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name, kring->nr_hwcur, kring->nr_hwtail, mbq_len(q),
len, m);
} else {
mbq_enqueue(q, m);
ND(10, "%s %d bufs in queue len %d m %p",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name, mbq_len(q), len, m);
/* notify outside the lock */
m = NULL;
error = 0;
}
mbq_unlock(q);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
done:
if (m)
m_freem(m);
/* unconditionally wake up listeners */
2015-07-10 05:51:36 +00:00
kring->nm_notify(kring, 0);
/* this is normally netmap_notify(), but for nics
* connected to a bridge it is netmap_bwrap_intr_notify(),
* that possibly forwards the frames through the switch
*/
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return (error);
}
/*
* netmap_reset() is called by the driver routines when reinitializing
* a ring. The driver is in charge of locking to protect the kring.
* If native netmap mode is not set just return NULL.
* If native netmap mode is set, in particular, we have to set nr_mode to
* NKR_NETMAP_ON.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
struct netmap_slot *
netmap_reset(struct netmap_adapter *na, enum txrx tx, u_int n,
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
u_int new_cur)
{
struct netmap_kring *kring;
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
int new_hwofs, lim;
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (!nm_native_on(na)) {
ND("interface not in native netmap mode");
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return NULL; /* nothing to reinitialize */
}
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
/* XXX note- in the new scheme, we are not guaranteed to be
* under lock (e.g. when called on a device reset).
* In this case, we should set a flag and do not trust too
* much the values. In practice: TODO
* - set a RESET flag somewhere in the kring
* - do the processing in a conservative way
* - let the *sync() fixup at the end.
*/
if (tx == NR_TX) {
if (n >= na->num_tx_rings)
return NULL;
kring = na->tx_rings + n;
if (kring->nr_pending_mode == NKR_NETMAP_OFF) {
kring->nr_mode = NKR_NETMAP_OFF;
return NULL;
}
// XXX check whether we should use hwcur or rcur
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
new_hwofs = kring->nr_hwcur - new_cur;
} else {
if (n >= na->num_rx_rings)
return NULL;
kring = na->rx_rings + n;
if (kring->nr_pending_mode == NKR_NETMAP_OFF) {
kring->nr_mode = NKR_NETMAP_OFF;
return NULL;
}
new_hwofs = kring->nr_hwtail - new_cur;
}
lim = kring->nkr_num_slots - 1;
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
if (new_hwofs > lim)
new_hwofs -= lim + 1;
/* Always set the new offset value and realign the ring. */
if (netmap_verbose)
D("%s %s%d hwofs %d -> %d, hwtail %d -> %d",
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
na->name,
tx == NR_TX ? "TX" : "RX", n,
kring->nkr_hwofs, new_hwofs,
kring->nr_hwtail,
tx == NR_TX ? lim : kring->nr_hwtail);
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
kring->nkr_hwofs = new_hwofs;
if (tx == NR_TX) {
kring->nr_hwtail = kring->nr_hwcur + lim;
if (kring->nr_hwtail > lim)
kring->nr_hwtail -= lim + 1;
}
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
#if 0 // def linux
/* XXX check that the mappings are correct */
/* need ring_nr, adapter->pdev, direction */
buffer_info->dma = dma_map_single(&pdev->dev, addr, adapter->rx_buffer_len, DMA_FROM_DEVICE);
if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) {
D("error mapping rx netmap buffer %d", i);
// XXX fix error handling
}
#endif /* linux */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
/*
* Wakeup on the individual and global selwait
1. Fix the handling of link reset while in netmap more. A link reset now is completely transparent for the netmap client: even if the NIC resets its own ring (e.g. restarting from 0), the client will not see any change in the current rx/tx positions, because the driver will keep track of the offset between the two. 2. make the device-specific code more uniform across different drivers There were some inconsistencies in the implementation of the netmap support routines, now drivers have been aligned to a common code structure. 3. import netmap support for ixgbe . This is implemented as a very small patch for ixgbe.c (233 lines, 11 chunks, mostly comments: in total the patch has only 54 lines of new code) , as most of the code is in an external file sys/dev/netmap/ixgbe_netmap.h , following some initial comments from Jack Vogel about making changes less intrusive. (Note, i have emailed Jack multiple times asking if he had comments on this structure of the code; i got no reply so i assume he is fine with it). Support for other drivers (em, lem, re, igb) will come later. "ixgbe" is now the reference driver for netmap support. Both the external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should serve as a reference for other device drivers. Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap, the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz on an i7-860 with 4 cores and 82599 card. Haven't tried yet more aggressive optimizations such as adding 'prefetch' instructions in the time-critical parts of the code.
2011-12-05 12:06:53 +00:00
* We do the wakeup here, but the ring is not yet reconfigured.
* However, we are under lock so there are no races.
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
*/
kring->nr_mode = NKR_NETMAP_ON;
2015-07-10 05:51:36 +00:00
kring->nm_notify(kring, 0);
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
return kring->ring->slot;
}
/*
* Dispatch rx/tx interrupts to the netmap rings.
*
* "work_done" is non-null on the RX path, NULL for the TX path.
* We rely on the OS to make sure that there is only one active
* instance per queue, and that there is appropriate locking.
*
* The 'notify' routine depends on what the ring is attached to.
* - for a netmap file descriptor, do a selwakeup on the individual
* waitqueue, plus one on the global one if needed
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
* (see netmap_notify)
* - for a nic connected to a switch, call the proper forwarding routine
* (see netmap_bwrap_intr_notify)
*/
int
netmap_common_irq(struct netmap_adapter *na, u_int q, u_int *work_done)
{
struct netmap_kring *kring;
2015-07-10 05:51:36 +00:00
enum txrx t = (work_done ? NR_RX : NR_TX);
q &= NETMAP_RING_MASK;
if (netmap_verbose) {
RD(5, "received %s queue %d", work_done ? "RX" : "TX" , q);
}
2015-07-10 05:51:36 +00:00
if (q >= nma_get_nrings(na, t))
return NM_IRQ_PASS; // not a physical queue
2015-07-10 05:51:36 +00:00
kring = NMR(na, t) + q;
if (kring->nr_mode == NKR_NETMAP_OFF) {
return NM_IRQ_PASS;
}
2015-07-10 05:51:36 +00:00
if (t == NR_RX) {
kring->nr_kflags |= NKR_PENDINTR; // XXX atomic ?
*work_done = 1; /* do not fire napi again */
}
return kring->nm_notify(kring, 0);
}
/*
* Default functions to handle rx/tx interrupts from a physical device.
* "work_done" is non-null on the RX path, NULL for the TX path.
*
* If the card is not in netmap mode, simply return NM_IRQ_PASS,
* so that the caller proceeds with regular processing.
* Otherwise call netmap_common_irq().
*
* If the card is connected to a netmap file descriptor,
* do a selwakeup on the individual queue, plus one on the global one
* if needed (multiqueue card _and_ there are multiqueue listeners),
* and return NR_IRQ_COMPLETED.
*
* Finally, if called on rx from an interface connected to a switch,
* calls the proper forwarding routine.
*/
int
netmap_rx_irq(struct ifnet *ifp, u_int q, u_int *work_done)
{
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
struct netmap_adapter *na = NA(ifp);
/*
* XXX emulated netmap mode sets NAF_SKIP_INTR so
* we still use the regular driver even though the previous
* check fails. It is unclear whether we should use
* nm_native_on() here.
*/
if (!nm_netmap_on(na))
return NM_IRQ_PASS;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
if (na->na_flags & NAF_SKIP_INTR) {
ND("use regular interrupt");
return NM_IRQ_PASS;
}
return netmap_common_irq(na, q, work_done);
}
/*
* Module loader and unloader
*
* netmap_init() creates the /dev/netmap device and initializes
* all global variables. Returns 0 on success, errno on failure
* (but there is no chance)
*
* netmap_fini() destroys everything.
*/
static struct cdev *netmap_dev; /* /dev/netmap character device. */
extern struct cdevsw netmap_cdevsw;
void
netmap_fini(void)
{
if (netmap_dev)
destroy_dev(netmap_dev);
/* we assume that there are no longer netmap users */
nm_os_ifnet_fini();
netmap_uninit_bridges();
netmap_mem_fini();
NMG_LOCK_DESTROY();
printf("netmap: unloaded module.\n");
}
int
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
netmap_init(void)
{
int error;
NMG_LOCK_INIT();
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
error = netmap_mem_init();
if (error != 0)
goto fail;
/*
* MAKEDEV_ETERNAL_KLD avoids an expensive check on syscalls
* when the module is compiled in.
* XXX could use make_dev_credv() to get error number
*/
netmap_dev = make_dev_credf(MAKEDEV_ETERNAL_KLD,
&netmap_cdevsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600,
"netmap");
if (!netmap_dev)
goto fail;
2015-07-10 05:51:36 +00:00
error = netmap_init_bridges();
if (error)
goto fail;
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
#ifdef __FreeBSD__
nm_os_vi_init_index();
Update to the current version of netmap. Mostly bugfixes or features developed in the past 6 months, so this is a 10.1 candidate. Basically no user API changes (some bugfixes in sys/net/netmap_user.h). In detail: 1. netmap support for virtio-net, including in netmap mode. Under bhyve and with a netmap backend [2] we reach over 1Mpps with standard APIs (e.g. libpcap), and 5-8 Mpps in netmap mode. 2. (kernel) add support for multiple memory allocators, so we can better partition physical and virtual interfaces giving access to separate users. The most visible effect is one additional argument to the various kernel functions to compute buffer addresses. All netmap-supported drivers are affected, but changes are mechanical and trivial 3. (kernel) simplify the prototype for *txsync() and *rxsync() driver methods. All netmap drivers affected, changes mostly mechanical. 4. add support for netmap-monitor ports. Think of it as a mirroring port on a physical switch: a netmap monitor port replicates traffic present on the main port. Restrictions apply. Drive carefully. 5. if_lem.c: support for various paravirtualization features, experimental and disabled by default. Most of these are described in our ANCS'13 paper [1]. Paravirtualized support in netmap mode is new, and beats the numbers in the paper by a large factor (under qemu-kvm, we measured gues-host throughput up to 10-12 Mpps). A lot of refactoring and additional documentation in the files in sys/dev/netmap, but apart from #2 and #3 above, almost nothing of this stuff is visible to other kernel parts. Example programs in tools/tools/netmap have been updated with bugfixes and to support more of the existing features. This is meant to go into 10.1 so we plan an MFC before the Aug.22 deadline. A lot of this code has been contributed by my colleagues at UNIPI, including Giuseppe Lettieri, Vincenzo Maffione, Stefano Garzarella. MFC after: 3 days.
2014-08-16 15:00:01 +00:00
#endif
2015-07-10 05:51:36 +00:00
error = nm_os_ifnet_init();
if (error)
goto fail;
printf("netmap: loaded module\n");
return (0);
fail:
netmap_fini();
return (EINVAL); /* may be incorrect */
Bring in support for netmap, a framework for very efficient packet I/O from userspace, capable of line rate at 10G, see http://info.iet.unipi.it/~luigi/netmap/ At this time I am bringing in only the generic code (sys/dev/netmap/ plus two headers under sys/net/), and some sample applications in tools/tools/netmap. There is also a manpage in share/man/man4 [1] In order to make use of the framework you need to build a kernel with "device netmap", and patch individual drivers with the code that you can find in sys/dev/netmap/head.diff The file will go away as the relevant pieces are committed to the various device drivers, which should happen in a few days after talking to the driver maintainers. Netmap support is available at the moment for Intel 10G and 1G cards (ixgbe, em/lem/igb), and for the Realtek 1G card ("re"). I have partial patches for "bge" and am starting to work on "cxgbe". Hopefully changes are trivial enough so interested third parties can submit their patches. Interested people can contact me for advice on how to add netmap support to specific devices. CREDITS: Netmap has been developed by Luigi Rizzo and other collaborators at the Universita` di Pisa, and supported by EU project CHANGE (http://www.change-project.eu/) The code is distributed under a BSD Copyright. [1] In my opinion is a bad idea to have all manpage in one directory. We should place kernel documentation in the same dir that contains the code, which would make it much simpler to keep doc and code in sync, reduce the clutter in share/man/ and incidentally is the policy used for all of userspace code. Makefiles and doc tools can be trivially adjusted to find the manpages in the relevant subdirs.
2011-11-17 12:17:39 +00:00
}