Re: [PATCH v2 1/2] vhost: Add smp_rmb() in vhost_vq_avail_empty()

[Gavin Shan]

On 3/27/24 12:44, Jason Wang wrote:
> On Wed, Mar 27, 2024 at 10:34 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > On Wed, Mar 27, 2024 at 7:39 AM Gavin Shan <gshan@xxxxxxxxxx> wrote:
> > > A smp_rmb() has been missed in vhost_vq_avail_empty(), spotted by
> > > Will Deacon <will@xxxxxxxxxx>. Otherwise, it's not ensured the
> > > available ring entries pushed by guest can be observed by vhost
> > > in time, leading to stale available ring entries fetched by vhost
> > > in vhost_get_vq_desc(), as reported by Yihuang Yu on NVidia's
> > > grace-hopper (ARM64) platform.
> > >
> > >    /home/gavin/sandbox/qemu.main/build/qemu-system-aarch64      \
> > >    -accel kvm -machine virt,gic-version=host -cpu host          \
> > >    -smp maxcpus=1,cpus=1,sockets=1,clusters=1,cores=1,threads=1 \
> > >    -m 4096M,slots=16,maxmem=64G                                 \
> > >    -object memory-backend-ram,id=mem0,size=4096M                \
> > >     :                                                           \
> > >    -netdev tap,id=vnet0,vhost=true                              \
> > >    -device virtio-net-pci,bus=pcie.8,netdev=vnet0,mac=52:54:00:f1:26:b0
> > >     :
> > >    guest# netperf -H 10.26.1.81 -l 60 -C -c -t UDP_STREAM
> > >    virtio_net virtio0: output.0:id 100 is not a head!
> > >
> > > Add the missed smp_rmb() in vhost_vq_avail_empty(). Note that it
> > > should be safe until vq->avail_idx is changed by commit 275bf960ac697
> > > ("vhost: better detection of available buffers").
> > >
> > > Fixes: 275bf960ac697 ("vhost: better detection of available buffers")
> > > Cc: <stable@xxxxxxxxxx> # v4.11+
> > > Reported-by: Yihuang Yu <yihyu@xxxxxxxxxx>
> > > Signed-off-by: Gavin Shan <gshan@xxxxxxxxxx>
> > > ---
> > >   drivers/vhost/vhost.c | 11 ++++++++++-
> > >   1 file changed, 10 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > > index 045f666b4f12..00445ab172b3 100644
> > > --- a/drivers/vhost/vhost.c
> > > +++ b/drivers/vhost/vhost.c
> > > @@ -2799,9 +2799,18 @@ bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
> > >          r = vhost_get_avail_idx(vq, &avail_idx);
> > >          if (unlikely(r))
> > >                  return false;
> > > +
> > >          vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
> > > +       if (vq->avail_idx != vq->last_avail_idx) {
> > > +               /* Similar to what's done in vhost_get_vq_desc(), we need
> > > +                * to ensure the available ring entries have been exposed
> > > +                * by guest.
> > > +                */
> >
> > We need to be more verbose here. For example, which load needs to be
> > ordered with which load.
> >
> > The rmb in vhost_get_vq_desc() is used to order the load of avail idx
> > and the load of head. It is paired with e.g virtio_wmb() in
> > virtqueue_add_split().
> >
> > vhost_vq_avail_empty() are mostly used as a hint in
> > vhost_net_busy_poll() which is under the protection of the vq mutex.
> >
> > An exception is the tx_can_batch(), but in that case it doesn't even
> > want to read the head.
>
> Ok, if it is needed only in that path, maybe we can move the barriers there.

[cc Will Deacon]

Jason, appreciate for your review and comments. I think PATCH[1/2] is
the fix for the hypothesis, meaning PATCH[2/2] is the real fix. However,
it would be nice to fix all of them in one shoot. I will try with PATCH[2/2]
only to see if our issue will disappear or not. However, the issue still
exists if PATCH[2/2] is missed.

Firstly, We were failing on the transmit queue and {tvq, rvq}->busyloop_timeout
== false if I remember correctly. So the added smp_rmb() in vhost_vq_avail_empty()
is only a concern to tx_can_batch(). A mutex isn't enough to ensure the order
for the available index and available ring entry (head). For example, vhost_vq_avail_empty()
called by tx_can_batch() can see next available index, but its corresponding
available ring entry (head) may not be seen by vhost yet if smp_rmb() is missed.
The next call to get_tx_bufs(), where the available ring entry (head) doesn't
arrived yet, leading to stale available ring entry (head) being fetched.

  handle_tx_copy
    get_tx_bufs                 // smp_rmb() won't be executed when vq->avail_idx != vq->last_avail_idx
    tx_can_batch
      vhost_vq_avail_empty      // vq->avail_idx is updated from vq->avail->idx

The reason why I added smp_rmb() to vhost_vq_avail_empty() is because the function
is a exposed API, even it's only used by drivers/vhost/net.c at present. It means
the API has been broken internally. So it seems more appropriate to fix it up in
vhost_vq_avail_empty() so that the API's users needn't worry about the memory access
order.

> > > +               smp_rmb();
> > > +               return false;
> > > +       }
> > >
> > > -       return vq->avail_idx == vq->last_avail_idx;
> > > +       return true;
> > >   }
> > >   EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);

Thanks,
Gavin