vhost_get_vq_desc (correctly) uses smp_rmb to order
avail ring reads after index reads.
However, over time we added two more places that read the
index and do not bother with barriers.
Since vhost_get_vq_desc when it was written assumed it is the
only reader when it sees a new index value is cached
it does not bother with a barrier either, as a result,
on the nvidia-gracehopper platform (arm64) available ring
entry reads have been observed bypassing ring reads, causing
a ring corruption.
To fix, factor out the correct index access code from vhost_get_vq_desc.
As a side benefit, we also validate the index on all paths now, which
will hopefully help catch future errors earlier.
Note: current code is inconsistent in how it handles errors:
some places treat it as an empty ring, others - non empty.
This patch does not attempt to change the existing behaviour.
Cc: [email protected]
Reported-by: Gavin Shan <[email protected]>
Reported-by: Will Deacon <[email protected]>
Suggested-by: Will Deacon <[email protected]>
Fixes: 275bf960ac69 ("vhost: better detection of available buffers")
Cc: "Jason Wang" <[email protected]>
Fixes: d3bb267bbdcb ("vhost: cache avail index in vhost_enable_notify()")
Cc: "Stefano Garzarella" <[email protected]>
Signed-off-by: Michael S. Tsirkin <[email protected]>
---
I think it's better to bite the bullet and clean up the code.
Note: this is still only built, not tested.
Gavin could you help test please?
Especially on the arm platform you have?
Will thanks so much for finding this race!
drivers/vhost/vhost.c | 80 +++++++++++++++++++++++--------------------
1 file changed, 42 insertions(+), 38 deletions(-)
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index 045f666b4f12..26b70b1fd9ff 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -1290,10 +1290,38 @@ static void vhost_dev_unlock_vqs(struct vhost_dev *d)
mutex_unlock(&d->vqs[i]->mutex);
}
-static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
- __virtio16 *idx)
+static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq)
{
- return vhost_get_avail(vq, *idx, &vq->avail->idx);
+ __virtio16 idx;
+ u16 avail_idx;
+ int r = vhost_get_avail(vq, idx, &vq->avail->idx);
+
+ if (unlikely(r < 0)) {
+ vq_err(vq, "Failed to access avail idx at %p: %d\n",
+ &vq->avail->idx, r);
+ return -EFAULT;
+ }
+
+ avail_idx = vhost16_to_cpu(vq, idx);
+
+ /* Check it isn't doing very strange things with descriptor numbers. */
+ if (unlikely((u16)(avail_idx - vq->last_avail_idx) > vq->num)) {
+ vq_err(vq, "Guest moved used index from %u to %u",
+ vq->last_avail_idx, vq->avail_idx);
+ return -EFAULT;
+ }
+
+ /* Nothing new? We are done. */
+ if (avail_idx == vq->avail_idx)
+ return 0;
+
+ vq->avail_idx = avail_idx;
+
+ /* We updated vq->avail_idx so we need a memory barrier between
+ * the index read above and the caller reading avail ring entries.
+ */
+ smp_rmb();
+ return 1;
}
static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
@@ -2498,38 +2526,21 @@ int vhost_get_vq_desc(struct vhost_virtqueue *vq,
{
struct vring_desc desc;
unsigned int i, head, found = 0;
- u16 last_avail_idx;
- __virtio16 avail_idx;
+ u16 last_avail_idx = vq->last_avail_idx;
__virtio16 ring_head;
int ret, access;
- /* Check it isn't doing very strange things with descriptor numbers. */
- last_avail_idx = vq->last_avail_idx;
if (vq->avail_idx == vq->last_avail_idx) {
- if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
- vq_err(vq, "Failed to access avail idx at %p\n",
- &vq->avail->idx);
- return -EFAULT;
- }
- vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
-
- if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
- vq_err(vq, "Guest moved used index from %u to %u",
- last_avail_idx, vq->avail_idx);
- return -EFAULT;
- }
+ ret = vhost_get_avail_idx(vq);
+ if (unlikely(ret < 0))
+ return ret;
/* If there's nothing new since last we looked, return
* invalid.
*/
- if (vq->avail_idx == last_avail_idx)
+ if (!ret)
return vq->num;
-
- /* Only get avail ring entries after they have been
- * exposed by guest.
- */
- smp_rmb();
}
/* Grab the next descriptor number they're advertising, and increment
@@ -2790,25 +2801,21 @@ EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
/* return true if we're sure that avaiable ring is empty */
bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
{
- __virtio16 avail_idx;
int r;
if (vq->avail_idx != vq->last_avail_idx)
return false;
- r = vhost_get_avail_idx(vq, &avail_idx);
- if (unlikely(r))
- return false;
- vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
+ r = vhost_get_avail_idx(vq);
- return vq->avail_idx == vq->last_avail_idx;
+ /* Note: we treat error as non-empty here */
+ return r == 0;
}
EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
/* OK, now we need to know about added descriptors. */
bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
{
- __virtio16 avail_idx;
int r;
if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
@@ -2832,13 +2839,10 @@ bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
/* They could have slipped one in as we were doing that: make
* sure it's written, then check again. */
smp_mb();
- r = vhost_get_avail_idx(vq, &avail_idx);
- if (r) {
- vq_err(vq, "Failed to check avail idx at %p: %d\n",
- &vq->avail->idx, r);
+ r = vhost_get_avail_idx(vq);
+ /* Note: we treat error as empty here */
+ if (r < 0)
return false;
- }
- vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
return vq->avail_idx != vq->last_avail_idx;
}
--
MST
On Wed, Mar 27, 2024 at 01:26:23PM -0400, Michael S. Tsirkin wrote:
> vhost_get_vq_desc (correctly) uses smp_rmb to order
> avail ring reads after index reads.
> However, over time we added two more places that read the
> index and do not bother with barriers.
> Since vhost_get_vq_desc when it was written assumed it is the
> only reader when it sees a new index value is cached
> it does not bother with a barrier either, as a result,
> on the nvidia-gracehopper platform (arm64) available ring
> entry reads have been observed bypassing ring reads, causing
> a ring corruption.
>
> To fix, factor out the correct index access code from vhost_get_vq_desc.
> As a side benefit, we also validate the index on all paths now, which
> will hopefully help catch future errors earlier.
>
> Note: current code is inconsistent in how it handles errors:
> some places treat it as an empty ring, others - non empty.
> This patch does not attempt to change the existing behaviour.
>
> Cc: [email protected]
> Reported-by: Gavin Shan <[email protected]>
> Reported-by: Will Deacon <[email protected]>
> Suggested-by: Will Deacon <[email protected]>
> Fixes: 275bf960ac69 ("vhost: better detection of available buffers")
> Cc: "Jason Wang" <[email protected]>
> Fixes: d3bb267bbdcb ("vhost: cache avail index in vhost_enable_notify()")
> Cc: "Stefano Garzarella" <[email protected]>
> Signed-off-by: Michael S. Tsirkin <[email protected]>
> ---
>
> I think it's better to bite the bullet and clean up the code.
> Note: this is still only built, not tested.
> Gavin could you help test please?
> Especially on the arm platform you have?
>
> Will thanks so much for finding this race!
No problem, and I was also hoping that the smp_rmb() could be
consolidated into a single helper like you've done here.
One minor comment below:
> drivers/vhost/vhost.c | 80 +++++++++++++++++++++++--------------------
> 1 file changed, 42 insertions(+), 38 deletions(-)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 045f666b4f12..26b70b1fd9ff 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -1290,10 +1290,38 @@ static void vhost_dev_unlock_vqs(struct vhost_dev *d)
> mutex_unlock(&d->vqs[i]->mutex);
> }
>
> -static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
> - __virtio16 *idx)
> +static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq)
> {
> - return vhost_get_avail(vq, *idx, &vq->avail->idx);
> + __virtio16 idx;
> + u16 avail_idx;
> + int r = vhost_get_avail(vq, idx, &vq->avail->idx);
> +
> + if (unlikely(r < 0)) {
> + vq_err(vq, "Failed to access avail idx at %p: %d\n",
> + &vq->avail->idx, r);
> + return -EFAULT;
> + }
> +
> + avail_idx = vhost16_to_cpu(vq, idx);
> +
> + /* Check it isn't doing very strange things with descriptor numbers. */
> + if (unlikely((u16)(avail_idx - vq->last_avail_idx) > vq->num)) {
> + vq_err(vq, "Guest moved used index from %u to %u",
> + vq->last_avail_idx, vq->avail_idx);
> + return -EFAULT;
> + }
> +
> + /* Nothing new? We are done. */
> + if (avail_idx == vq->avail_idx)
> + return 0;
> +
> + vq->avail_idx = avail_idx;
> +
> + /* We updated vq->avail_idx so we need a memory barrier between
> + * the index read above and the caller reading avail ring entries.
> + */
> + smp_rmb();
I think you could use smp_acquire__after_ctrl_dep() if you're feeling
brave, but to be honest I'd prefer we went in the opposite direction
and used READ/WRITE_ONCE + smp_load_acquire()/smp_store_release() across
the board. It's just a thankless, error-prone task to get there :(
So, for the patch as-is:
Acked-by: Will Deacon <[email protected]>
(I've not tested it either though, so definitely wait for Gavin on that!)
Cheers,
Will
On Wed, Mar 27, 2024 at 07:52:02PM +0000, Will Deacon wrote:
> On Wed, Mar 27, 2024 at 01:26:23PM -0400, Michael S. Tsirkin wrote:
> > vhost_get_vq_desc (correctly) uses smp_rmb to order
> > avail ring reads after index reads.
> > However, over time we added two more places that read the
> > index and do not bother with barriers.
> > Since vhost_get_vq_desc when it was written assumed it is the
> > only reader when it sees a new index value is cached
> > it does not bother with a barrier either, as a result,
> > on the nvidia-gracehopper platform (arm64) available ring
> > entry reads have been observed bypassing ring reads, causing
> > a ring corruption.
> >
> > To fix, factor out the correct index access code from vhost_get_vq_desc.
> > As a side benefit, we also validate the index on all paths now, which
> > will hopefully help catch future errors earlier.
> >
> > Note: current code is inconsistent in how it handles errors:
> > some places treat it as an empty ring, others - non empty.
> > This patch does not attempt to change the existing behaviour.
> >
> > Cc: [email protected]
> > Reported-by: Gavin Shan <[email protected]>
> > Reported-by: Will Deacon <[email protected]>
> > Suggested-by: Will Deacon <[email protected]>
> > Fixes: 275bf960ac69 ("vhost: better detection of available buffers")
> > Cc: "Jason Wang" <[email protected]>
> > Fixes: d3bb267bbdcb ("vhost: cache avail index in vhost_enable_notify()")
> > Cc: "Stefano Garzarella" <[email protected]>
> > Signed-off-by: Michael S. Tsirkin <[email protected]>
> > ---
> >
> > I think it's better to bite the bullet and clean up the code.
> > Note: this is still only built, not tested.
> > Gavin could you help test please?
> > Especially on the arm platform you have?
> >
> > Will thanks so much for finding this race!
>
> No problem, and I was also hoping that the smp_rmb() could be
> consolidated into a single helper like you've done here.
>
> One minor comment below:
>
> > drivers/vhost/vhost.c | 80 +++++++++++++++++++++++--------------------
> > 1 file changed, 42 insertions(+), 38 deletions(-)
> >
> > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > index 045f666b4f12..26b70b1fd9ff 100644
> > --- a/drivers/vhost/vhost.c
> > +++ b/drivers/vhost/vhost.c
> > @@ -1290,10 +1290,38 @@ static void vhost_dev_unlock_vqs(struct vhost_dev *d)
> > mutex_unlock(&d->vqs[i]->mutex);
> > }
> >
> > -static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
> > - __virtio16 *idx)
> > +static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq)
> > {
> > - return vhost_get_avail(vq, *idx, &vq->avail->idx);
> > + __virtio16 idx;
> > + u16 avail_idx;
> > + int r = vhost_get_avail(vq, idx, &vq->avail->idx);
> > +
> > + if (unlikely(r < 0)) {
> > + vq_err(vq, "Failed to access avail idx at %p: %d\n",
> > + &vq->avail->idx, r);
> > + return -EFAULT;
> > + }
> > +
> > + avail_idx = vhost16_to_cpu(vq, idx);
> > +
> > + /* Check it isn't doing very strange things with descriptor numbers. */
> > + if (unlikely((u16)(avail_idx - vq->last_avail_idx) > vq->num)) {
> > + vq_err(vq, "Guest moved used index from %u to %u",
> > + vq->last_avail_idx, vq->avail_idx);
> > + return -EFAULT;
> > + }
> > +
> > + /* Nothing new? We are done. */
> > + if (avail_idx == vq->avail_idx)
> > + return 0;
> > +
> > + vq->avail_idx = avail_idx;
> > +
> > + /* We updated vq->avail_idx so we need a memory barrier between
> > + * the index read above and the caller reading avail ring entries.
> > + */
> > + smp_rmb();
>
> I think you could use smp_acquire__after_ctrl_dep() if you're feeling
> brave, but to be honest I'd prefer we went in the opposite direction
> and used READ/WRITE_ONCE + smp_load_acquire()/smp_store_release() across
> the board. It's just a thankless, error-prone task to get there :(
Let's just say that's a separate patch, I tried hard to make this one
a bugfix only, no other functional changes at all.
> So, for the patch as-is:
>
> Acked-by: Will Deacon <[email protected]>
>
> (I've not tested it either though, so definitely wait for Gavin on that!)
>
> Cheers,
>
> Will