Received-SPF: pass (google.com: domain of linux-kernel+bounces-120443-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Precedence: bulk
MIME-Version: 1.0
References: <20240326233846.1086253-1-gshan@redhat.com> <20240326233846.1086253-2-gshan@redhat.com>
 <CACGkMEtyEo2QQSEh0ZnZDUJSXMMhzeO97Jp0wF4_rhzUBGk4Zg@mail.gmail.com>
 <CACGkMEtUdgDr_M-F8-gdFkJp+29Xuw9DCib2-diFmJxFxDN2Bw@mail.gmail.com>
 <b6f4e8c1-bdd1-496e-aa1a-68349674671c@redhat.com> <92ac171a-0004-45ef-a6c8-01b1c135a8ca@redhat.com>
In-Reply-To: <92ac171a-0004-45ef-a6c8-01b1c135a8ca@redhat.com>
From: Jason Wang <jasowang@redhat.com>
Date: Wed, 27 Mar 2024 15:42:45 +0800
Message-ID: <CACGkMEti3xtqHxhM-DGcquP6UncELUzrNeVor45wfGRCBkkZrg@mail.gmail.com>
Subject: Re: [PATCH v2 1/2] vhost: Add smp_rmb() in vhost_vq_avail_empty()
To: Gavin Shan <gshan@redhat.com>
Cc: virtualization@lists.linux.dev, linux-kernel@vger.kernel.org, 
	mst@redhat.com, davem@davemloft.net, stefanha@redhat.com, sgarzare@redhat.com, 
	keirf@google.com, yihyu@redhat.com, shan.gavin@gmail.com, 
	Will Deacon <will@kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Wed, Mar 27, 2024 at 3:35=E2=80=AFPM Gavin Shan <gshan@redhat.com> wrote=
:
>
> On 3/27/24 14:08, Gavin Shan wrote:
> > On 3/27/24 12:44, Jason Wang wrote:
> >> On Wed, Mar 27, 2024 at 10:34=E2=80=AFAM Jason Wang <jasowang@redhat.c=
om> wrote:
> >>> On Wed, Mar 27, 2024 at 7:39=E2=80=AFAM Gavin Shan <gshan@redhat.com>=
 wrote:
> >>>>
> >>>> A smp_rmb() has been missed in vhost_vq_avail_empty(), spotted by
> >>>> Will Deacon <will@kernel.org>. 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=3Dhost -cpu host          \
> >>>>    -smp maxcpus=3D1,cpus=3D1,sockets=3D1,clusters=3D1,cores=3D1,thre=
ads=3D1 \
> >>>>    -m 4096M,slots=3D16,maxmem=3D64G                                 =
\
> >>>>    -object memory-backend-ram,id=3Dmem0,size=3D4096M                =
\
> >>>>     :                                                           \
> >>>>    -netdev tap,id=3Dvnet0,vhost=3Dtrue                              =
\
> >>>>    -device virtio-net-pci,bus=3Dpcie.8,netdev=3Dvnet0,mac=3D52:54:00=
: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 275bf960ac69=
7
> >>>> ("vhost: better detection of available buffers").
> >>>>
> >>>> Fixes: 275bf960ac697 ("vhost: better detection of available buffers"=
)
> >>>> Cc: <stable@kernel.org> # v4.11+
> >>>> Reported-by: Yihuang Yu <yihyu@redhat.com>
> >>>> Signed-off-by: Gavin Shan <gshan@redhat.com>
> >>>> ---
> >>>>   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 *d=
ev, struct vhost_virtqueue *vq)
> >>>>          r =3D vhost_get_avail_idx(vq, &avail_idx);
> >>>>          if (unlikely(r))
> >>>>                  return false;
> >>>> +
> >>>>          vq->avail_idx =3D vhost16_to_cpu(vq, avail_idx);
> >>>> +       if (vq->avail_idx !=3D vq->last_avail_idx) {
> >>>> +               /* Similar to what's done in vhost_get_vq_desc(), we=
 need
> >>>> +                * to ensure the available ring entries have been ex=
posed
> >>>> +                * 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 stil=
l
> > exists if PATCH[2/2] is missed.
> >
>
> Jason, PATCH[2/2] is sufficient to fix our current issue. I tried with PA=
TCH[2/2]
> only and unable to hit the issue. However, PATCH[1/2] may be needed by ot=
her scenarios.
> So it would be nice to fix them in one shoot.

Yes, see below.

>
>
> > Firstly, We were failing on the transmit queue and {tvq, rvq}->busyloop=
_timeout
> > =3D=3D false if I remember correctly. So the added smp_rmb() in vhost_v=
q_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, v=
host_vq_avail_empty()
> > called by tx_can_batch() can see next available index, but its correspo=
nding
> > available ring entry (head) may not be seen by vhost yet if smp_rmb() i=
s missed.
> > The next call to get_tx_bufs(), where the available ring entry (head) d=
oesn'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 !=3D vq->last_avail_idx
> >      tx_can_batch
> >        vhost_vq_avail_empty      // vq->avail_idx is updated from vq->a=
vail->idx
> >
> > The reason why I added smp_rmb() to vhost_vq_avail_empty() is because t=
he 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.

When tx_can_batch returns true it means there's still pending tx
buffers. Since it might read indices so it still can bypass the
smp_rmb() in the vhost_get_vq_desc().

I'd suggest adding those above to change log.

With this,

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

> >
> >>>
> >>>
> >>>> +               smp_rmb();
> >>>> +               return false;
> >>>> +       }
> >>>>
> >>>> -       return vq->avail_idx =3D=3D vq->last_avail_idx;
> >>>> +       return true;
> >>>>   }
> >>>>   EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
>
> Thanks,
> Gavin
>