Indirect descriptors are usually supported by virtio-net devices,
allowing to dispatch a larger number of requests.
When the virtio device sends a packet using indirect descriptors,
only one slot is used in the ring, even for large packets.
The main effect is to improve the 0% packet loss benchmark.
A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
(fwd io for host, macswap for VM) on DUT shows a +50% gain for
zero loss.
On the downside, micro-benchmark using testpmd txonly in VM and
rxonly on host shows a loss between 1 and 4%. But depending on
the needs, feature can be disabled at VM boot time by passing
indirect_desc=off argument to vhost-user device in Qemu.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
As new_device and destroy_device use an int instead of a
"struct virtio_net *", The comment about setting VIRTIO_DEV_RUNNING
doesn't make sense anymore, plus If I've correctly understand the
code, the drivers take care of setting the flag before calling the
callbacks, so I guess that this comment is obsolet and I've remove it.
Signed-off-by: Matthias Gatto <matthias.gatto@outscale.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
No need to use a pointer to store/retrieve features.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Invoke get_device() at the beginning of vhost_user_msg_handler, so that
we could check the return value once. Which could save tons of duplicate
get-and-check device.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Some functions are with prefix "user_", while others with "vhost_".
Making them all starting with "vhost_user_" to unify the function names.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Due to history reason (that we have 2 vhost implementations), some
messages are handled in two calls: vhost specific implementation
handles it first and then invoke the common one to do another handling.
We have one implementation only now, we could write one method for
each message. Here fold those common handles to corresponding vhost
user handler.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The code structure is a bit messy now. For example, vhost-user message
handling is spread to three different files:
vhost-net-user.c virtio-net.c virtio-net-user.c
Where, vhost-net-user.c is the entrance to handle all those messages
and then invoke the right method for a specific message. Some of them
are stored at virtio-net.c, while others are stored at virtio-net-user.c.
The truth is all of them should be in one file, vhost_user.c.
So this patch refactors the source code structure: mainly on renaming
files and moving code from one file to another file that is more suitable
for storing it. Thus, no functional changes are made.
After the refactor, the code structure becomes to:
- socket.c handles all vhost-user socket file related stuff, such
as, socket file creation for server mode, reconnection
for client mode.
- vhost.c mainly on stuff like vhost device creation/destroy/reset.
Most of the vhost API implementation are there, too.
- vhost_user.c all stuff about vhost-user messages handling goes there.
- virtio_net.c all stuff about virtio-net should go there. It has virtio
net Rx/Tx implementation only so far: it's just a rename
from vhost_rxtx.c
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
We now have one vhost implementation; no sub source dir is needed.
Remove it by move them to upper dir.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
remove vhost-cuse code, including the eventfd_link kernel module that
is for vhost-cuse only.
The lib/virt/qemu-wrap.py is also removed, as it's mainly for vhost-cuse
usage.
As we have one vhost implementation now, one vhost config option is
needed only. Thus, CONFIG_RTE_LIBRTE_VHOST_USER is removed.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
nr_desc is not an index but the number of descriptors,
so can be equal to the virtqueue size.
Fixes: a436f53ebfeb ("vhost: avoid dead loop chain")
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Currently while calling of 'rte_vhost_driver_unregister()' connection
to QEMU will not be closed. This leads to inability to register driver
again and reconnect to same virtual machine.
This scenario is reproducible with OVS. While executing of the following
command vhost port will be re-created (will be executed
'rte_vhost_driver_register()' followed by 'rte_vhost_driver_unregister()')
network will be broken and QEMU possibly will crash:
ovs-vsctl set Interface vhost1 ofport_request=15
Fix this by closing all established connections on driver unregister and
removing of pending connections from reconnection list.
Fixes: 64ab701c3d1e ("vhost: add vhost-user client mode")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
If something abnormal happened to QEMU, 'connect()' can block calling
thread (e.g. main thread of OVS) forever or for a really long time.
This can break whole application or block the reconnection thread.
Example with OVS:
ovs_rcu(urcu2)|WARN|blocked 512000 ms waiting for main to quiesce
(gdb) bt
#0 connect () from /lib64/libpthread.so.0
#1 vhost_user_create_client (vsocket=0xa816e0)
#2 rte_vhost_driver_register
#3 netdev_dpdk_vhost_user_construct
#4 netdev_open (name=0xa664b0 "vhost1")
[...]
#11 main
Fix that by setting non-blocking mode for client sockets for connection.
Fixes: 64ab701c3d1e ("vhost: add vhost-user client mode")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Protect against DPDK crash when allocation of listen fd >= 1023.
For events on fd:s >1023, the current implementation will trigger
an abort due to access outside of allocated bit mask.
Corrections would include:
* Match fdset_add() signature in fd_man.c to fd_man.h
* Handling of return codes from fdset_add()
* Addition of check of fd number in fdset_add_fd()
The rationale behind the suggested code change is that,
fdset_event_dispatch() could attempt access outside of the FD_SET
bitmask if there is an event on a file descriptor that in turn
looks up a virtio file descriptor with a value > 1023.
Such an attempt will lead to an abort() and a restart of any
vswitch using DPDK.
A discussion topic exist in the ovs-discuss mailing list that can
provide a little more background:
http://openvswitch.org/pipermail/discuss/2016-February/020243.html
Fixes: 8f972312 ("vhost: support vhost-user")
Signed-off-by: Patrik Andersson <patrik.r.andersson@ericsson.com>
Signed-off-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
In current implementation vhost will crash with segmentation fault
if malicious or buggy virtio application breaks addresses of descriptors.
Before commit 0823c1cb0a73 ("vhost: workaround stale vring base")
this crash was reproducible even with normal DPDK application that tries
to change number of virtqueues dynamically inside VM.
Fix that by checking addresses of descriptors before using.
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Return value on error changed from '-1' to '0' because it returns
unsigned value and it means number of used descriptors.
Also fixed updating of 'last_used_idx' by using actual number of
used descriptors.
Fixes: 623bc47054d0 ("vhost: do sanity check for ring descriptor length")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Fix potential memory leak raised by Coverity.
>>> Variable "vsocket" going out of scope leaks the storage it
>>> points to.
Coverity issue: 127483
Fixes: e623e0c6d8a5 ("vhost: add reconnect ability")
Reported-by: John McNamara <john.mcnamara@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Commit 550c9d27d143 ("vhost: set/reset device flags internally") moves
the VIRTIO_DEV_RUNNING set/reset to vhost lib. But I missed one reset
on stop; here fixes it.
Fixes: 550c9d27d143 ("vhost: set/reset device flags internally")
Reported-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Ciara Loftus <ciara.loftus@intel.com>
Some libraries were missing their dependency on eal, mbuf, mempool,
ring and kvargs.
It is revealed by the linker option "-z defs".
Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
Value returned from fstat is not checked for errors before being used.
This patch fixes following coverity issue.
static uint64_t
get_blk_size(int fd)
{
struct stat stat;
fstat(fd, &stat);
return (uint64_t)stat.st_blksize;
>>> CID 107103 (#1 of 1): Unchecked return value from library
(CHECKED_RETURN)
>>> check_return: Calling fstat(fd, &stat) without checking
return value.
>>> This library function may fail and return an error code.
Fixes: 8f972312b8f4 ("vhost: support vhost-user")
Signed-off-by: Huawei Xie <huawei.xie@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Return value of function get_device() is not checking before
dereference. Fix this problem by adding checking condition.
Coverity issue: 119262
Fixes: 77d20126b4c2 ("vhost-user: handle message to enable vring")
Signed-off-by: Marcin Kerlin <marcinx.kerlin@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
All other DPDK PMDs doesn't support concurrent receiving or sending
packets to the same queue. The upper application should deal with
this, normally through queue and core bindings.
Due to historical reason, vhost internally supports concurrent lockless
enqueuing packets to the same virtio queue through costly cmpset operation.
This patch removes this internal lockless implementation and should improve
performance a bit.
Luckily DPDK OVS doesn't rely on this behavior.
Signed-off-by: Huawei Xie <huawei.xie@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
The ifname[] field takes so much space, that it seperates some frequently
used fields into different caches, say, features and broadcast_rarp.
This patch moves all those fields that will be accessed frequently in Rx/Tx
together (before the ifname[] field) to let them share one cache line.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
A virtio driver normally uses at least 2 desc buffers for Tx: the
first for storing the header, and the others for storing the data.
Therefore, we could fetch the first data desc buf before the main
loop, and do the copy first before the check of "are we done yet?".
This could save one check for small packets that just have one data
desc buffer and need one mbuf to store it.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Pre update and update used ring in batch for Tx and Rx at the stage
while fetching all avail desc idx. This would reduce some cache misses
and hence, increase the performance a bit.
Pre update would be feasible as guest driver will not start processing
those entries as far as we don't update "used->idx". (I'm not 100%
certain I don't miss anything, though).
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
When DPDK app crashes (or quits, or gets killed), a restart of DPDK
app would get stale vring base from QEMU. That would break the kernel
virtio net completely, making it non-work any more, unless a driver
reset is done.
So, instead of getting the stale vring base from QEMU, Huawei suggested
we could get a much saner (and may not the most accurate) vring base
from used->idx. That would work because:
- there is a memory barrier between updating used ring entries and
used->idx. So, even though we crashed at updating the used ring
entries, it will not cause any issue, as the guest driver will not
process those stale used entries, for used-idx is not updated yet.
- DPDK process vring in order, that means a crash may just lead some
packet retransmission for Tx and drop for Rx.
Suggested-by: Huawei Xie <huawei.xie@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
Allow reconnecting on failure by default when:
- DPDK app starts first and QEMU (as the server) is not started yet.
Without reconnecting, DPDK app would simply fail on vhost-user
registration.
- QEMU restarts, say due to OS reboot.
Without reconnecting, you can't re-establish the connection without
restarting DPDK app.
This patch make it work well for both above cases. It simply creates
a new thread, and keep trying calling "connect()", until it succeeds.
The reconnect could be disabled when RTE_VHOST_USER_NO_RECONNECT flag
is set.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Add a new paramter (flags) to rte_vhost_driver_register(). DPDK
vhost-user acts as client mode when RTE_VHOST_USER_CLIENT flag
is set. The flags would also allow future extensions without
breaking the API (again).
The rest is straingfoward then: allocate a unix socket, and
bind/listen for server, connect for client.
This extension is for vhost-user only, therefore we simply quit
and report error when any flags are given for vhost-cuse.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
DPDK vhost-user just acts as server so far, so, using a struct named
as "vhost_server" is okay. However, if we add client mode, it doesn't
make sense any more. Here renames it to "vhost_user_socket".
There was no obvious wrong about "connfd_ctx", but I think it's obviously
better to rename it to "vhost_user_connection", as it does represent
a connection, a connection between the backend (DPDK) and the frontend
(QEMU).
Similarly, few more renames are taken, such as "vserver_new_vq_conn"
to "vhost_user_new_connection".
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Array of buf_vector's is just an array for temporary storing information
about available descriptors. It used only locally in virtio_dev_merge_rx()
and there is no reason for that array to be shared.
Fix that by allocating local buf_vec inside virtio_dev_merge_rx().
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
Virtio net header length is set per device, but not per queue. So, there
is no reason to store it in vhost_virtqueue struct, instead, we should
store it in virtio_net struct, to make one copy only.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
"virtio_net_device_ops" is the only left open struct that an application
can access, therefore, it's the only place that might introduce potential
ABI break in future for extension.
So, do some reservation for it. 5 should be pretty enough, considering
that we have barely touched it for a long while. Another reason to
choose 5 is for cache alignment: 5 makes the struct 64 bytes for 64 bit
machine.
With this, it's confidence to say that we might be able to be free from
the ABI violation forever.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
It barely has anything useful there, just 2 functions prototype. Here
move them to vhost-net.h, and delete it.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
The "reserved" field in virtio_net and vhost_virtqueue struct is not
necessary any more. We now expose virtio_net device with a number "vid".
This patch also removes the "priv" field: all fields are priviate now:
application can't access it now. The only way that we could still access
it is to expose it by a function, but I doubt that's needed or worthwhile.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
We are now safe to move all those internal structs/macros/functions to
vhost-net.h, to hide them from external access.
This patch also breaks long lines and removes some redundant comments.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
With all the previous prepare works, we are just one step away from
the final ABI refactoring. That is, to change current API to let them
stick to vid instead of the old virtio_net dev.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
The new API rte_vhost_avail_entries() is actually a rename of
rte_vring_available_entries(), with the "vring" to "vhost" name
change to keep the consistency of other vhost exported APIs.
This change could let us avoid the dependency of "virtio_net"
struct, to prepare for the ABI refactoring.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
Introduce a new API rte_vhost_get_ifname() to export the ifname to
application.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
Introduce a new API rte_vhost_get_queue_num() to export the number of
queues.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
Introduce a new API rte_vhost_get_numa_node() to get the numa node
from which the virtio_net struct is allocated.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
vhost cuse is now the last reference of the vhost_device_ctx struct;
move it there, and do a rename to "vhost_cuse_device_ctx", to make
it clear that it's "cuse only".
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
get_device() just needs vid, so pass vid as the parameter only.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
I failed to figure out what does "fh" mean here for a long while.
The only guess I could have had is "file handle". So, you get the
point that it's not well named.
I then figured it out that "fh" is derived from the fuse lib, and
my above guess is right. However, device_fh represents a virtio
net device ID. Therefore, here I rename it to vid (Virtio-net device
ID, or Vhost device ID; choose one you prefer) to make it easier for
understanding.
This name (vid) then will be considered to the only interface to
applications. That's another reason to do the rename: it's our
interface, make it more understandable.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
device_fh repsents the device id for a specific virtio net device.
Firstly, "int" would be big enough: we don't need 64 bit. Secondly,
this could let us avoid the ugly "%" PRIu64 ".." stuff.
And since ctx.fh is derived from device_fh, declare it as int, too.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
It does not make sense to ask the application to set/unset the flag
VIRTIO_DEV_RUNNING (that used internal only) at new_device()/
destroy_device() callback.
Instead, it should be set after new_device() succeeds and reset before
destroy_device() is invoked inside vhost lib. This patch fixes it.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
It's an fd; so define it as "int", which could also save the unncessary
(int) case.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Rich Lane <rich.lane@bigswitch.com>
Acked-by: Rich Lane <rich.lane@bigswitch.com>
Fix issue reported by Coverity.
Coverity ID 124556
If the buffer is treated as a null terminated string in later operations,
a buffer overflow or over-read may occur.
In vhost_set_ifname: The string buffer may not have a null terminator if
the source string's length is equal to the buffer size
Fixes: 54292e9520e0 ("vhost: support ifname for vhost-user")
Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
