This patch enables the handling of buffers non-contiguous in
process virtual address space in the enqueue path when mergeable
buffers are used.
When virtio-net header doesn't fit in a single chunck, it is
computed in a local variable and copied to the buffer chuncks
afterwards.
For packet content, the copy length is limited to the chunck
size, next chuncks VAs being fetched afterward.
This issue has been assigned CVE-2018-1059.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch enables the handling of buffers non-contiguous in
process virtual address space in the enqueue path when mergeable
buffers aren't used.
When virtio-net header doesn't fit in a single chunck, it is
computed in a local variable and copied to the buffer chuncks
afterwards.
For packet content, the copy length is limited to the chunck
size, next chuncks VAs being fetched afterward.
This issue has been assigned CVE-2018-1059.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch enables the handling of buffers non-contiguous in
process virtual address space in the dequeue path.
When virtio-net header doesn't fit in a single chunck, it is
copied into a local variablei before being processed.
For packet content, the copy length is limited to the chunck
size, next chuncks VAs being fetched afterward.
This issue has been assigned CVE-2018-1059.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch adds support for non-contiguous indirect descriptor
tables in VA space.
When it happens, which is unlikely, a table is allocated and the
non-contiguous content is copied into it.
This issue has been assigned CVE-2018-1059.
Reported-by: Yongji Xie <xieyongji@baidu.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
There is currently no check done on the length when translating
guest addresses into host virtual addresses. Also, there is no
guanrantee that the guest addresses range is contiguous in
the host virtual address space.
This patch prepares vhost_iova_to_vva() and its callers to
return and check the mapped size. If the mapped size is smaller
than the requested size, the caller handle it as an error.
This issue has been assigned CVE-2018-1059.
Reported-by: Yongji Xie <xieyongji@baidu.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch fixes the size passed at the indirect descriptor
table translation time, which is the len field of the descriptor,
and not a single descriptor.
This issue has been assigned CVE-2018-1059.
Fixes: 62fdb8255a ("vhost: use the guest IOVA to host VA helper")
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
LOG_DEBUG is a symbol defined by POSIX, so if sys/log.h is
included the symbols conflict.
This patch changes LOG_DEBUG to VHOST_LOG_DEBUG.
Fixes: 1c01d52392 ("vhost: add debug print")
Cc: stable@dpdk.org
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Signed-off-by: Tomasz Kulasek <tomaszx.kulasek@intel.com>
Reviewed-by: Jianfeng Tan <jianfeng.tan@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The librte_vhost API is used in two ways:
1. As a vhost net device backend via rte_vhost_enqueue/dequeue_burst().
2. As a library for implementing vhost device backends.
There is no distinction between the two at the API level or in the
librte_vhost implementation. For example, device state is kept in
"struct virtio_net" regardless of whether this is actually a net device
backend or whether the built-in virtio_net.c driver is in use.
The virtio_net.c driver should be a librte_vhost API client just like
the vhost-scsi code and have no special access to vhost.h internals.
Unfortunately, fixing this requires significant librte_vhost API
changes.
This patch takes a different approach: keep the librte_vhost API
unchanged but track whether the built-in virtio_net.c driver is in use.
See the next patch for a bug fix that requires knowledge of whether
virtio_net.c is in use.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
When performing live migration or memory hot-plugging,
the changes to the device and vrings made by message handler
done independently from vring usage by PMD threads.
This causes for example segfaults during live-migration
with MQ enable, but in general virtually any request
sent by qemu changing the state of device can cause
problems.
These patches fixes all above issues by adding a spinlock
to every vring and requiring message handler to start operation
only after ensuring that all PMD threads related to the device
are out of critical section accessing the vring data.
Each vring has its own lock in order to not create contention
between PMD threads of different vrings and to prevent
performance degradation by scaling queue pair number.
See https://bugzilla.redhat.com/show_bug.cgi?id=1450680
Cc: stable@dpdk.org
Signed-off-by: Victor Kaplansky <victork@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
dequeue zero copy change buf_addr and buf_iova of mbuf, and return
to mbuf pool without restore them, it breaks vm memory if others allocate
mbuf from same pool since mbuf reset doesn't reset buf_addr and buf_iova.
Fixes: b0a985d1f3 ("vhost: add dequeue zero copy")
Cc: stable@dpdk.org
Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Due to a mistake operation from me, older version (v10) was merged to
master branch. It's the v11 should be applied. However, the master branch
is not rebase-able. Thus, this patch is made, from the diff between v10
and v11.
The diffs are:
- Add check for parameter and tailroom in rte_net_make_rarp_packet
- Allocate mbuf in rte_net_make_rarp_packet
Besides that, a link error is fixed when shared lib is enabled.
Fixes: 45ae05df82 ("net: add a helper for making RARP packet")
Fixes: c3ffdba0e8 ("vhost: use API to make RARP packet")
Signed-off-by: Xiao Wang <xiao.w.wang@intel.com>
Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
The driver can suppress interrupt when VIRTIO_F_EVENT_IDX feature bit is
negotiated. The driver set vring flags to 0, and MAY use used_event in
available ring to advise device interrupt util reach an index specified
by used_event. The device ignore the lower bit of vring flags, and send
an interrupt when index reach used_event.
The device can suppress notification in a manner analogous to the ways
driver suppress interrupt. The device manipulates flags or avail_event in
the used ring in the same way the driver manipulates flags or used_event in
available ring.
Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
In virtio, UDP Fragmentation Offload (UFO) includes two parts: host UFO
and guest UFO. Guest UFO means the frontend can receive large UDP
packets, and host UFO means the backend can receive large UDP packets.
This patch supports host UFO and guest UFO for vhost-user.
Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Extract the callfd eventfd signal operation so virtio_net.c does not
have to repeat it multiple times.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
This fix dequeue zero copy can not work with Qemu
version >= 2.7. Since from Qemu 2.7 virtio device
use virtio-1 protocol, the zero copy code path
forget to add offset to buffer address.
Fixes: b0a985d1f3 ("vhost: add dequeue zero copy")
Cc: stable@dpdk.org
Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Replace the BSD license header with the SPDX tag for files
with only an Intel copyright on them.
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Rename buf_physaddr to buf_iova.
Keep the deprecated name in an anonymous union to avoid breaking
the API.
Signed-off-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
When offload is enabled, vhost needs to access the first mbuf
to get the packet info, e.g. TCP header. So we couldn't delay
the data copy in this case.
Fixes: e5c494a7a2 ("vhost: batch small guest memory copies")
Reported-by: Lei Yao <lei.a.yao@intel.com>
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Translating the start addresses of the rings is not enough, we need to
be sure all the ring is made available by the guest.
It depends on the size of the rings, which is not known on SET_VRING_ADDR
reception. Furthermore, we need to be be safe against vring pages
invalidates.
This patch introduces a new access_ok flag per virtqueue, which is set
when all the rings are mapped, and cleared as soon as a page used by a
ring is invalidated. The invalidation part is implemented in a following
patch.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Replace rte_vhost_gpa_to_vva() calls with vhost_iova_to_vva(), which
requires to also pass the mapped len and the access permissions needed.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
In the non-mergeable receive case, when copy_mbuf_to_desc()
call fails the packet is skipped, the corresponding used element
len field is set to vnet header size, and it continues with next
packet/desc. It could be a problem because it does not know why
it failed, and assume the desc buffer is large enough.
In mergeable receive case, when copy_mbuf_to_desc_mergeable()
fails, packets burst is simply stopped.
This patch makes the non-mergeable error path to behave as the
mergeable one, as it seems the safest way. Also, doing this way
will simplify pending IOTLB miss requests handling.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
This patch adaptively batches the small guest memory copies.
By batching the small copies, the efficiency of executing the
memory LOAD instructions can be improved greatly, because the
memory LOAD latency can be effectively hidden by the pipeline.
We saw great performance boosts for small packets PVP test.
This patch improves the performance for small packets, and has
distinguished the packets by size. So although the performance
for big packets doesn't change, it makes it relatively easy to
do some special optimizations for the big packets too.
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Zhiyong Yang <zhiyong.yang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
To compare enabled features in current device we must use bit
mask instead of bit position.
Fixes: c843af3aa1 ("vhost: access header only if offloading is supported")
Cc: stable@dpdk.org
Signed-off-by: Ivan Dyukov <i.dyukov@samsung.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
There is no way to bypass IP checksum verification in Linux
kernel, no matter skb->ip_summed is assigned as CHECKSUM_UNNECESSARY
or CHECKSUM_PARTIAL.
So any packets with bad IP checksum will be dropped at VM IP layer.
To correct, we check this flag PKT_TX_IP_CKSUM to calculate IP csum.
Fixes: 859b480d5a ("vhost: add guest offload setting")
Cc: stable@dpdk.org
Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
As PKT_TX_TCP_SEG flag in mbuf->ol_flags implies PKT_TX_TCP_CKSUM,
applications, e.g., testpmd, don't set PKT_TX_TCP_CKSUM when TSO
is set.
This leads to that packets get dropped in VM tcp stack layer because
of bad TCP csum.
To fix this, we make sure TCP NEEDS_CSUM info is set into virtio net
header when PKT_TX_TCP_SEG is set, so that VM tcp stack will not
check the TCP csum.
Fixes: 859b480d5a ("vhost: add guest offload setting")
Cc: stable@dpdk.org
Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Fixing typos across dpdk source code using codespell utility.
Skipped the ethdev driver's base code fixes to keep the base
code intact.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Different drivers use internal macros like force_inline for compiler
always inline feature.
Standardizing it through __rte_always_inline macro.
Verified the change by comparing the output binary file.
No difference found in the output binary file with this change.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
For zero copy mode, we need pin the mbuf to not let the underlaying PMD
driver (or the app) free the mbuf. Currently, only the heading mbuf is
pinned. However, the mbuf free function would try to free all mbufs
in the mbuf chain (-1 to the refcnt). This may lead the head mbuf being
still pinned, while the other subsequent mbufs are actually freed. Which
is wrong.
It becomes more fatal after the mbuf refactor, more specificly, after
the commit 8f094a9ac5 ("mbuf: set mbuf fields while in pool"). The
refcnt resets to 1 after the last real reference. OTOH, it leads to a
situtation that we never know one mbuf is actually freed or not. This
would result the mbuf __just__ after the heading mbuf being freed twice:
it's firstly freed (and put back to mempool) when the underlaying PMD
finishes the DMA. Later, it will then be freed again when vhost unpins
it. Meaning, one mbuf may be returned to the mempool twice, while in
turn, being allocated twice later. Something uncertain may happen then.
For example, the VM2VM case becomes broken.
Fixes: b0a985d1f3 ("vhost: add dequeue zero copy")
Cc: stable@dpdk.org
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Like what we did for virtio PMD driver [0][1], we could also apply such
trick to vhost, to avoid the memory write on net header when necessary.
[0]: c9ea670c1d ("net/virtio: fix performance regression due to TSO")
[1]: 16994abee2 ("net/virtio: optimize header reset on any layout")
With this, the cache issue of the mergeable path is again greatly reduced:
even the write of "num_buffers" could be avoided. A quick PVP test shows
the gap between the mergeable Rx and non-mergeable Rx is pretty small now:
they are basically the same in my test.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Rename "rte_virtio_net.h" to "rte_vhost.h", to not let it be virtio
net specific.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Export few APIs for the vhost-user driver to log the guest memory writes,
which is a must for live migration support.
This patch basically moves vhost_log_write() and vhost_log_used_vring()
into vhost.h and then add an wrapper (the public API) to them.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The queue pair is very virtio-net specific, other devices don't have
such concept. To make it generic, we should log the number of vrings
instead of the number of queue pairs.
This patch just does a simple convert, a later patch would export the
number of vrings to applications.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The broadcast_rarp field in the virtio_net struct is checked in the
dequeue datapath regardless of whether descriptors are available or not.
As it is checked with cmpset leading to a write, false sharing on the
virtio_net struct can happen between enqueue and dequeue datapaths
regardless of whether a RARP is requested. In OVS, the issue can cause
a uni-directional performance drop of up to 15%.
Fix that by only performing the cmpset if a read of broadcast_rarp
indicates that the cmpset is likely to succeed.
Fixes: a66bcad322 ("vhost: arrange struct fields for better cache sharing")
Cc: stable@dpdk.org
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Found with clang static analysis:
lib/librte_vhost/virtio_net.c:723:17: warning:
Access to field 'data_off' results in a dereference of a null pointer
(loaded from variable 'tcp_hdr')
m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
^~~~~~~~~~~~~~~~~
Fixes: d0cf91303d ("vhost: add Tx offload capabilities")
Cc: stable@dpdk.org
Signed-off-by: Emmanuel Roullit <emmanuel.roullit@gmail.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
If a malicious guest forges a dead loop desc chain (let desc->next point
to itself) and desc->len is zero, this could lead to a dead loop in
copy_mbuf_to_desc(following is a simplified code to show this issue
clearly):
while (mbuf_is_not_totally_consumed) {
if (desc_avail == 0) {
desc = &descs[desc->next];
desc_avail = desc->len;
}
COPY(desc, mbuf, desc_avail);
}
I have actually fixed a same issue before: commit a436f53ebf ("vhost:
avoid dead loop chain"); it fixes the dequeue path though, leaving the
enqueue path still vulnerable.
The fix is the same. Add a var nr_desc to avoid the dead loop.
Fixes: f1a519ad98 ("vhost: fix enqueue/dequeue to handle chained vring descriptors")
Cc: stable@dpdk.org
Reported-by: Xieming Katty <katty.xieming@huawei.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Linux virtio-net kernel driver uses indirect descriptors when
mergeable buffers are not used.
This patch adds its support, fixing the use of indirect
descriptors with these guests.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Windows virtio-net driver uses indirect descriptors with
mergeable buffers.
This patch adds its support, fixing the use of indirect
descriptors with these guests.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
There is no need to retrieve the latest avail head every time we enqueue
a packet in the mereable Rx path by
avail_idx = *((volatile uint16_t *)&vq->avail->idx);
Instead, we could just retrieve it once at the beginning of the enqueue
path. This could diminish the cache penalty slightly, because the virtio
driver could be updating it while vhost is reading it (for each packet).
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The basic idea is to shadow the used ring update: update them into a
local buffer first, and then flush them all to the virtio used vring
at once in the end.
And since we do avail ring reservation before enqueuing data, we would
know which and how many descs will be used. Which means we could update
the shadow used ring at the reservation time. It also introduce another
slight advantage: we don't need access the desc->flag any more inside
copy_mbuf_to_desc_mergeable().
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
shadow_used_ring will be introduced later. Since then last avail
idx will not be updated together with last used idx.
So, here we use last_avail_idx for avail ring reservation.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Let it return "num_buffers" we reserved, so that we could re-use it
with copy_mbuf_to_desc_mergeable() directly, instead of calculating
it again there.
Meanwhile, the return type of copy_mbuf_to_desc_mergeable is changed
to "int". -1 will be return on error.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch reorders the code to delay virtio header write to improve
cache access efficiency for cases where the mrg_rxbuf feature is turned
on. CPU pipeline stall cycles can be significantly reduced.
Virtio header write and mbuf data copy are all remote store operations
which takes a long time to finish. It's a good idea to put them together
to remove bubbles in between, to let as many remote store instructions
as possible go into store buffer at the same time to hide latency, and
to let the H/W prefetcher goes to work as early as possible.
On a Haswell machine, about 100 cycles can be saved per packet by this
patch alone. Taking 64B packets traffic for example, this means about 60%
efficiency improvement for the enqueue operation.
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
If offloading features are not negotiated, parsing the virtio header
is not needed.
Micro-benchmark with testpmd shows that the gain is +4% with indirect
descriptors, +1% when using direct descriptors.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
This patch fixes a Windows VM compatibility issue in DPDK 16.07 vhost code
which causes the guest to hang once any packets are enqueued when mrg_rxbuf
is turned on by setting the right id and len in the used ring.
As defined in virtio spec 0.95 and 1.0, in each used ring element, id means
index of start of used descriptor chain, and len means total length of the
descriptor chain which was written to. While in 16.07 code, index of the
last descriptor is assigned to id, and the length of the last descriptor is
assigned to len.
How to test?
1. Start testpmd in the host with a vhost port.
2. Start a Windows VM image with qemu and connect to the vhost port.
3. Start io forwarding with tx_first in host testpmd.
For 16.07 code, the Windows VM will hang once any packets are enqueued.
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
The basic idea of dequeue zero copy is, instead of copying data from
the desc buf, here we let the mbuf reference the desc buf addr directly.
Doing so, however, has one major issue: we can't update the used ring
at the end of rte_vhost_dequeue_burst. Because we don't do the copy
here, an update of the used ring would let the driver to reclaim the
desc buf. As a result, DPDK might reference a stale memory region.
To update the used ring properly, this patch does several tricks:
- when mbuf references a desc buf, refcnt is added by 1.
This is to pin lock the mbuf, so that a mbuf free from the DPDK
won't actually free it, instead, refcnt is subtracted by 1.
- We chain all those mbuf together (by tailq)
And we check it every time on the rte_vhost_dequeue_burst entrance,
to see if the mbuf is freed (when refcnt equals to 1). If that
happens, it means we are the last user of this mbuf and we are
safe to update the used ring.
- "struct zcopy_mbuf" is introduced, to associate an mbuf with the
right desc idx.
Dequeue zero copy is introduced for performance reason, and some rough
tests show about 50% perfomance boost for packet size 1500B. For small
packets, (e.g. 64B), it actually slows a bit down (well, it could up to
15%). That is expected because this patch introduces some extra works,
and it outweighs the benefit from saving few bytes copy.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
So far, we retrieve both the used ring and avail ring idx by the var
last_used_idx; it won't be a problem because the used ring is updated
immediately after those avail entries are consumed.
But that's not true when dequeue zero copy is enabled, that used ring is
updated only when the mbuf is consumed. Thus, we need use another var to
note the last avail ring idx we have consumed.
Therefore, last_avail_idx is introduced.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
