linking NIC Receive Side Scaling (RSS) to the network stack's
connection-group implementation. This prototype (and derived patches)
are in use at Juniper and several other FreeBSD-using companies, so
despite some reservations about its maturity, merge the patch to the
base tree so that it can be iteratively refined in collaboration rather
than maintained as a set of gradually diverging patch sets.
(1) Merge a software implementation of the Toeplitz hash specified in
RSS implemented by David Malone. This is used to allow suitable
pcbgroup placement of connections before the first packet is
received from the NIC. Software hashing is generally avoided,
however, due to high cost of the hash on general-purpose CPUs.
(2) In in_rss.c, maintain authoritative versions of RSS state intended
to be pushed to each NIC, including keying material, hash
algorithm/ configuration, and buckets. Provide software-facing
interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both
the RSS standardised Toeplitz and a 'naive' variation with a hash
efficient in software but with poor distribution properties.
Implement rss_m2cpuid()to be used by netisr and other load
balancing code to look up the CPU on which an mbuf should be
processed.
(3) In the Ethernet link layer, allow netisr distribution using RSS as
a source of policy as an alternative to source ordering; continue
to default to direct dispatch (i.e., don't try and requeue packets
for processing on the 'right' CPU if they arrive in a directly
dispatchable context).
(4) Allow RSS to control tuning of connection groups in order to align
groups with RSS buckets. If a packet arrives on a protocol using
connection groups, and contains a suitable hardware-generated
hash, use that hash value to select the connection group for pcb
lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz
hash is available, we fall back on regular PCB lookup risking
contention rather than pay the cost of Toeplitz in software --
this is a less scalable but, at my last measurement, faster
approach. As core counts go up, we may want to revise this
strategy despite CPU overhead.
Where device drivers suitably configure NICs, and connection groups /
RSS are enabled, this should avoid both lock and line contention during
connection lookup for TCP. This commit does not modify any device
drivers to tune device RSS configuration to the global RSS
configuration; patches are in circulation to do this for at least
Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device
drivers is not particularly robust, nor aware of more advanced features
such as runtime reconfiguration/rebalancing. This will hopefully prove
a useful starting point for refinement.
No MFC is scheduled as we will first want to nail down a more mature
and maintainable KPI/KBI for device drivers.
Sponsored by: Juniper Networks (original work)
Sponsored by: EMC/Isilon (patch update and merge)
AppleTalk was a network transport protocol for Apple Macintosh devices
in 80s and then 90s. Starting with Mac OS X in 2000 the AppleTalk was
a legacy protocol and primary networking protocol is TCP/IP. The last
Mac OS X release to support AppleTalk happened in 2009. The same year
routing equipment vendors (namely Cisco) end their support.
Thus, AppleTalk won't be supported in FreeBSD 11.0-RELEASE.
IPX was a network transport protocol in Novell's NetWare network operating
system from late 80s and then 90s. The NetWare itself switched to TCP/IP
as default transport in 1998. Later, in this century the Novell Open
Enterprise Server became successor of Novell NetWare. The last release
that claimed to still support IPX was OES 2 in 2007. Routing equipment
vendors (e.g. Cisco) discontinued support for IPX in 2011.
Thus, IPX won't be supported in FreeBSD 11.0-RELEASE.
callback providers. link_init_sdl() function can be used to
fill most of the parameters. Use caller stack instead of
allocation / freing memory for each request. Do not drop support
for extra-long (probably non-existing) link-layer protocols by
introducing link_alloc_sdl() (used by if_resolvemulti() callback)
and link_free_sdl() (used by caller).
Since this change breaks KBI, MFC requires slightly different approach
(link_init_sdl() auto-allocating buffer if necessary to handle cases
with unmodified if_resolvemulti() callers).
MFC after: 2 weeks
rather than just void *.
Then, as part of this, convert a couple of mbuf m->m_data accesses
to mtod(m, const void *).
Reviewed by: markm
Approved by: security-officer (delphij)
Sponsored by: Netflix, Inc.
in net, to avoid compatibility breakage for no sake.
The future plan is to split most of non-kernel parts of
pfvar.h into pf.h, and then make pfvar.h a kernel only
include breaking compatibility.
Discussed with: bz
to this event, adding if_var.h to files that do need it. Also, include
all includes that now are included due to implicit pollution via if_var.h
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
and there are ifnets, that do that via counter(9). Provide a flag that
would skip cache line trashing '+=' operation in ether_input().
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
Reviewed by: melifaro, adrian
Approved by: re (marius)
been tested. With all sources turned on, this unlocks itself in
a couple of seconds! That is no my box, and there is no guarantee
that this will be the case everywhere.
* Cut debug prints.
* Use the same locks/mutexes all the way through.
* Be a tad more conservative about entropy estimates.
Contains:
* Refactor the hardware RNG CPU instruction sources to feed into
the software mixer. This is unfinished. The actual harvesting needs
to be sorted out. Modified by me (see below).
* Remove 'frac' parameter from random_harvest(). This was never
used and adds extra code for no good reason.
* Remove device write entropy harvesting. This provided a weak
attack vector, was not very good at bootstrapping the device. To
follow will be a replacement explicit reseed knob.
* Separate out all the RANDOM_PURE sources into separate harvest
entities. This adds some secuity in the case where more than one
is present.
* Review all the code and fix anything obviously messy or inconsistent.
Address som review concerns while I'm here, like rename the pseudo-rng
to 'dummy'.
Submitted by: Arthur Mesh <arthurmesh@gmail.com> (the first item)
addresses added to the UUID generator using uuid_ether_add(). The
UUID generator keeps an arbitrary number of MAC addresses, under
the assumption that they are rarely removed (= uuid_ether_del()).
This achieves the following:
1. It brings up closer to having the network stack as a loadable
module.
2. It allows the UUID generator to filter MAC addresses for best
results (= highest chance of uniqeness).
3. MAC addresses can come from anywhere, irrespactive of whether
it's used for an interface or not.
A side-effect of the change is that when no MAC addresses have been
added, a random multicast MAC address is created once and re-used if
needed. Previusly, when a random MAC address was needed, it was
created for every call. Thus, a change in behaviour is introduced
for when no MAC addresses exist.
Obtained from: Juniper Networks, Inc.
Merge ether_ipfw_chk() and part of bridge_pfil() into
unified ipfw_check_frame() function called by PFIL.
This change was suggested by rwatson? @ DevSummit.
Remove ipfw headers from ether/bridge code since they are unneeded now.
Note this thange introduce some (temporary) performance penalty since
PFIL read lock has to be acquired for every link-level packet.
MFC after: 3 weeks
Cleaner solution (e.g. adding another header) should be done here.
Original log:
Move several enums and structures required for L2 filtering from ip_fw_private.h to ip_fw.h.
Remove ipfw/ip_fw_private.h header from non-ipfw code.
Requested by: luigi
Approved by: kib(mentor)
from scratch, copying needed functionality from the old implemenation
on demand, with a thorough review of all code. The main change is that
interface layer has been removed from the CARP. Now redundant addresses
are configured exactly on the interfaces, they run on.
The CARP configuration itself is, as before, configured and read via
SIOCSVH/SIOCGVH ioctls. A new prefix created with SIOCAIFADDR or
SIOCAIFADDR_IN6 may now be configured to a particular virtual host id,
which makes the prefix redundant.
ifconfig(8) semantics has been changed too: now one doesn't need
to clone carpXX interface, he/she should directly configure a vhid
on a Ethernet interface.
To supply vhid data from the kernel to an application the getifaddrs(8)
function had been changed to pass ifam_data with each address. [1]
The new implementation definitely closes all PRs related to carp(4)
being an interface, and may close several others. It also allows
to run a single redundant IP per interface.
Big thanks to Bjoern Zeeb for his help with inet6 part of patch, for
idea on using ifam_data and for several rounds of reviewing!
PR: kern/117000, kern/126945, kern/126714, kern/120130, kern/117448
Reviewed by: bz
Submitted by: bz [1]
setting (either default or if supported as set by SIOCSIFFIB, e.g.
from ifconfig).
Submitted by: Alexander V. Chernikov (melifaro ipfw.ru)
Reviewed by: julian
MFC after: 2 weeks
default dispatch method to NETISR_DISPATCH_DIRECT in order to force
direct dispatch. This adds a fairly negligble overhead without
changing default behavior, but in the future will allow deferred or
hybrid dispatch to other worker threads before link layer processing
has taken place.
For example, this could allow redistribution using RSS hashes
without ethernet header cache line hits, if the NIC was unable to
adequately implement load balancing to too small a number of input
queues -- perhaps due to hard queueset counts of 1, 3, or 8, but in
a modern system with 16-128 threads. This can happen on highly
threaded systems, where you want want an ithread per core,
redistributing work to other queues, but also on virtualised systems
where hardware hashing is (or is not) available, but only a single
queue has been directed to one VCPU on a VM.
Note: this adds a previously non-present assertion about the
equivalence of the ifnet from which the packet is received, and the
ifnet stamped in the mbuf header. I believe this assertion to
generally be true, but we'll find out soon -- if it's not, we might
have to add additional overhead in some cases to add an m_tag with
the originating ifnet pointer stored in it.
Reviewed by: bz
MFC after: 3 weeks
Sponsored by: Juniper Networks, Inc.
DPCPU_DEFINE and VNET_DEFINE macros, as these cause problems for various
people working on the affected files. A better long-term solution is
still being considered. This reversal may give some modules empty
set_pcpu or set_vnet sections, but these are harmless.
Changes reverted:
------------------------------------------------------------------------
r215318 | dim | 2010-11-14 21:40:55 +0100 (Sun, 14 Nov 2010) | 4 lines
Instead of unconditionally emitting .globl's for the __start_set_xxx and
__stop_set_xxx symbols, only emit them when the set_vnet or set_pcpu
sections are actually defined.
------------------------------------------------------------------------
r215317 | dim | 2010-11-14 21:38:11 +0100 (Sun, 14 Nov 2010) | 3 lines
Apply the STATIC_VNET_DEFINE and STATIC_DPCPU_DEFINE macros throughout
the tree.
------------------------------------------------------------------------
r215316 | dim | 2010-11-14 21:23:02 +0100 (Sun, 14 Nov 2010) | 2 lines
Add macros to define static instances of VNET_DEFINE and DPCPU_DEFINE.
bridge(4), lagg(4) etc. and make use of function pointers and
pf_proto_register() to hook carp into the network stack.
Currently, because of the uncertainty about whether the unload path is free
of race condition panics, unloads are disallowed by default. Compiling with
CARPMOD_CAN_UNLOAD in CFLAGS removes this anti foot shooting measure.
This commit requires IP6PROTOSPACER, introduced in r211115.
Reviewed by: bz, simon
Approved by: ken (mentor)
MFC after: 2 weeks
and tested over the past two months in the ipfw3-head branch. This
also happens to be the same code available in the Linux and Windows
ports of ipfw and dummynet.
The major enhancement is a completely restructured version of
dummynet, with support for different packet scheduling algorithms
(loadable at runtime), faster queue/pipe lookup, and a much cleaner
internal architecture and kernel/userland ABI which simplifies
future extensions.
In addition to the existing schedulers (FIFO and WF2Q+), we include
a Deficit Round Robin (DRR or RR for brevity) scheduler, and a new,
very fast version of WF2Q+ called QFQ.
Some test code is also present (in sys/netinet/ipfw/test) that
lets you build and test schedulers in userland.
Also, we have added a compatibility layer that understands requests
from the RELENG_7 and RELENG_8 versions of the /sbin/ipfw binaries,
and replies correctly (at least, it does its best; sometimes you
just cannot tell who sent the request and how to answer).
The compatibility layer should make it possible to MFC this code in a
relatively short time.
Some minor glitches (e.g. handling of ipfw set enable/disable,
and a workaround for a bug in RELENG_7's /sbin/ipfw) will be
fixed with separate commits.
CREDITS:
This work has been partly supported by the ONELAB2 project, and
mostly developed by Riccardo Panicucci and myself.
The code for the qfq scheduler is mostly from Fabio Checconi,
and Marta Carbone and Francesco Magno have helped with testing,
debugging and some bug fixes.
- use a uniform mtag format for all packets that exit and re-enter
the firewall in the middle of a rulechain. On reentry, all tags
containing reinject info are renamed to MTAG_IPFW_RULE so the
processing is simpler.
- make ipfw and dummynet use ip_len and ip_off in network format
everywhere. Conversion is done only once instead of tracking
the format in every place.
- use a macro FREE_PKT to dispose of mbufs. This eases portability.
On passing i also removed a few typos, staticise or localise variables,
remove useless declarations and other minor things.
Overall the code shrinks a bit and is hopefully more readable.
I have tested functionality for all but ng_ipfw and if_bridge/if_ethersubr.
For ng_ipfw i am actually waiting for feedback from glebius@ because
we might have some small changes to make.
For if_bridge and if_ethersubr feedback would be welcome
(there are still some redundant parts in these two modules that
I would like to remove, but first i need to check functionality).
r201011
- move most of ng_ipfw.h into ip_fw_private.h, as this code is
ipfw-specific. This removes a dependency on ng_ipfw.h from some files.
- move many equivalent definitions of direction (IN, OUT) for
reinjected packets into ip_fw_private.h
- document the structure of the packet tags used for dummynet
and netgraph;
r201049
- merge some common code to attach/detach hooks into
a single function.
r201055
- remove some duplicated code in ip_fw_pfil. The input
and output processing uses almost exactly the same code so
there is no need to use two separate hooks.
ip_fw_pfil.o goes from 2096 to 1382 bytes of .text
r201057 (see the svn log for full details)
- macros to make the conversion of ip_len and ip_off
between host and network format more explicit
r201113 (the remaining parts)
- readability fixes -- put braces around some large for() blocks,
localize variables so the compiler does not think they are uninitialized,
do not insist on precise allocation size if we have more than we need.
r201119
- when doing a lookup, keys must be in big endian format because
this is what the radix code expects (this fixes a bug in the
recently-introduced 'lookup' option)
No ABI changes in this commit.
MFC after: 1 week
and remove all O(N) sequences from kernel critical sections in ipfw.
In detail:
1. introduce a IPFW_UH_LOCK to arbitrate requests from
the upper half of the kernel. Some things, such as 'ipfw show',
can be done holding this lock in read mode, whereas insert and
delete require IPFW_UH_WLOCK.
2. introduce a mapping structure to keep rules together. This replaces
the 'next' chain currently used in ipfw rules. At the moment
the map is a simple array (sorted by rule number and then rule_id),
so we can find a rule quickly instead of having to scan the list.
This reduces many expensive lookups from O(N) to O(log N).
3. when an expensive operation (such as insert or delete) is done
by userland, we grab IPFW_UH_WLOCK, create a new copy of the map
without blocking the bottom half of the kernel, then acquire
IPFW_WLOCK and quickly update pointers to the map and related info.
After dropping IPFW_LOCK we can then continue the cleanup protected
by IPFW_UH_LOCK. So userland still costs O(N) but the kernel side
is only blocked for O(1).
4. do not pass pointers to rules through dummynet, netgraph, divert etc,
but rather pass a <slot, chain_id, rulenum, rule_id> tuple.
We validate the slot index (in the array of #2) with chain_id,
and if successful do a O(1) dereference; otherwise, we can find
the rule in O(log N) through <rulenum, rule_id>
All the above does not change the userland/kernel ABI, though there
are some disgusting casts between pointers and uint32_t
Operation costs now are as follows:
Function Old Now Planned
-------------------------------------------------------------------
+ skipto X, non cached O(N) O(log N)
+ skipto X, cached O(1) O(1)
XXX dynamic rule lookup O(1) O(log N) O(1)
+ skipto tablearg O(N) O(1)
+ reinject, non cached O(N) O(log N)
+ reinject, cached O(1) O(1)
+ kernel blocked during setsockopt() O(N) O(1)
-------------------------------------------------------------------
The only (very small) regression is on dynamic rule lookup and this will
be fixed in a day or two, without changing the userland/kernel ABI
Supported by: Valeria Paoli
MFC after: 1 month
At this time we pull out from ip_fw2.c the logging functions, and
support for dynamic rules, and move kernel-only stuff into
netinet/ipfw/ip_fw_private.h
No ABI change involved in this commit, unless I made some mistake.
ip_fw.h has changed, though not in the userland-visible part.
Files touched by this commit:
conf/files
now references the two new source files
netinet/ip_fw.h
remove kernel-only definitions gone into netinet/ipfw/ip_fw_private.h.
netinet/ipfw/ip_fw_private.h
new file with kernel-specific ipfw definitions
netinet/ipfw/ip_fw_log.c
ipfw_log and related functions
netinet/ipfw/ip_fw_dynamic.c
code related to dynamic rules
netinet/ipfw/ip_fw2.c
removed the pieces that goes in the new files
netinet/ipfw/ip_fw_nat.c
minor rearrangement to remove LOOKUP_NAT from the
main headers. This require a new function pointer.
A bunch of other kernel files that included netinet/ip_fw.h now
require netinet/ipfw/ip_fw_private.h as well.
Not 100% sure i caught all of them.
MFC after: 1 month
packet filters. ALso allows ipfw to be enabled on on ejail and disabled
on another. In 8.0 it's a global setting.
Sitting aroung in tree waiting to commit for: 2 months
MFC after: 2 months
vnet.h, we now use jails (rather than vimages) as the abstraction
for virtualization management, and what remained was specific to
virtual network stacks. Minor cleanups are done in the process,
and comments updated to reflect these changes.
Reviewed by: bz
Approved by: re (vimage blanket)
L2 information. For an indirect route the cached L2 entry contains the
MAC address of the gateway. Typically the default route is used to
transmit multicast packets when explicit multicast routes are not
available. The ether_output() function bypasses L2 resolution function
if it verifies the L2 cache is valid, because the cached L2 address
(a unicast MAC address) is copied into the packets as the destination
MAC address. This validation, however, does not apply to broadcast and
multicast packets because the destination MAC address is mapped
according to a standard method instead.
Submitted by: Xin Li
Reviewed by: bz
Approved by: re
(DPCPU), as suggested by Peter Wemm, and implement a new per-virtual
network stack memory allocator. Modify vnet to use the allocator
instead of monolithic global container structures (vinet, ...). This
change solves many binary compatibility problems associated with
VIMAGE, and restores ELF symbols for virtualized global variables.
Each virtualized global variable exists as a "reference copy", and also
once per virtual network stack. Virtualized global variables are
tagged at compile-time, placing the in a special linker set, which is
loaded into a contiguous region of kernel memory. Virtualized global
variables in the base kernel are linked as normal, but those in modules
are copied and relocated to a reserved portion of the kernel's vnet
region with the help of a the kernel linker.
Virtualized global variables exist in per-vnet memory set up when the
network stack instance is created, and are initialized statically from
the reference copy. Run-time access occurs via an accessor macro, which
converts from the current vnet and requested symbol to a per-vnet
address. When "options VIMAGE" is not compiled into the kernel, normal
global ELF symbols will be used instead and indirection is avoided.
This change restores static initialization for network stack global
variables, restores support for non-global symbols and types, eliminates
the need for many subsystem constructors, eliminates large per-subsystem
structures that caused many binary compatibility issues both for
monitoring applications (netstat) and kernel modules, removes the
per-function INIT_VNET_*() macros throughout the stack, eliminates the
need for vnet_symmap ksym(2) munging, and eliminates duplicate
definitions of virtualized globals under VIMAGE_GLOBALS.
Bump __FreeBSD_version and update UPDATING.
Portions submitted by: bz
Reviewed by: bz, zec
Discussed with: gnn, jamie, jeff, jhb, julian, sam
Suggested by: peter
Approved by: re (kensmith)
- at_ifawithnet(), which acquires an locks it needs and returns an
at_ifaddr reference.
- at_ifawithnet_locked(), which relies on the caller locking
at_ifaddr_list, and returns a pointer rather than a reference.
Update various consumers to prefer one or the other, including ether
and fddi output, to properly release at_ifaddr references.
Rework at_control() to manage locking and references in a manner
identical to in_control().
MFC after: 6 weeks
use IPv4/v6 for inter-node communication (according to my reading).
Properly wrap the carp callouts in INET || INET6 and refelect this
in sys/conf/files as well. While in theory this should be ok,
it might be a bit optimistic to think that carp could build with
inet6 only[1].
Discussed with: mlaier [1]
