Currently, under certain circumstances vhost_init_is_le does just a part
of the initialization job, and depends on vhost_reset_is_le being called
too. For this reason vhost_vq_init_access used to call vhost_reset_is_le
when vq->private_data is NULL. This is not only counter intuitive, but
also real a problem because it breaks vhost_net. The bug was introduced to
vhost_net with commit 2751c9882b94 ("vhost: cross-endian support for
legacy devices"). The symptom is corruption of the vq's used.idx field
(virtio) after VHOST_NET_SET_BACKEND was issued as a part of the vhost
shutdown on a vq with pending descriptors.
Let us make sure the outcome of vhost_init_is_le never depend on the state
it is actually supposed to initialize, and fix virtio_net by removing the
reset from vhost_vq_init_access.
With the above, there is no reason for vhost_reset_is_le to do just half
of the job. Let us make vhost_reset_is_le reinitialize is_le.
Signed-off-by: Halil Pasic <pasic at linux.vnet.ibm.com>
Reported-by: Michael A. Tebolt <miket at us.ibm.com>
Reported-by: Dr. David Alan Gilbert <dgilbert at redhat.com>
Fixes: commit 2751c9882b94 ("vhost: cross-endian support for legacy
devices")
---
The bug was already discussed here:
http://www.spinics.net/lists/kvm/msg144365.html
This is a follow up patch.
---
drivers/vhost/vhost.c | 10 ++++------
1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index d643260..8f99fe0 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -130,14 +130,14 @@ static long vhost_get_vring_endian(struct vhost_virtqueue
*vq, u32 idx,
static void vhost_init_is_le(struct vhost_virtqueue *vq)
{
- if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
- vq->is_le = true;
+ vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
+ || virtio_legacy_is_little_endian();
}
#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
static void vhost_reset_is_le(struct vhost_virtqueue *vq)
{
- vq->is_le = virtio_legacy_is_little_endian();
+ vhost_init_is_le(vq);
}
struct vhost_flush_struct {
@@ -1714,10 +1714,8 @@ int vhost_vq_init_access(struct vhost_virtqueue *vq)
int r;
bool is_le = vq->is_le;
- if (!vq->private_data) {
- vhost_reset_is_le(vq);
+ if (!vq->private_data)
return 0;
- }
vhost_init_is_le(vq);
--
2.8.4
On Mon, 30 Jan 2017 11:09:36 +0100 Halil Pasic <pasic at linux.vnet.ibm.com> wrote:> Currently, under certain circumstances vhost_init_is_le does just a part > of the initialization job, and depends on vhost_reset_is_le being called > too. For this reason vhost_vq_init_access used to call vhost_reset_is_le > when vq->private_data is NULL. This is not only counter intuitive, but > also real a problem because it breaks vhost_net. The bug was introduced to > vhost_net with commit 2751c9882b94 ("vhost: cross-endian support for > legacy devices"). The symptom is corruption of the vq's used.idx field > (virtio) after VHOST_NET_SET_BACKEND was issued as a part of the vhost > shutdown on a vq with pending descriptors. > > Let us make sure the outcome of vhost_init_is_le never depend on the state > it is actually supposed to initialize, and fix virtio_net by removing the > reset from vhost_vq_init_access. > > With the above, there is no reason for vhost_reset_is_le to do just half > of the job. Let us make vhost_reset_is_le reinitialize is_le. > > Signed-off-by: Halil Pasic <pasic at linux.vnet.ibm.com> > Reported-by: Michael A. Tebolt <miket at us.ibm.com> > Reported-by: Dr. David Alan Gilbert <dgilbert at redhat.com> > Fixes: commit 2751c9882b94 ("vhost: cross-endian support for legacy devices") > ---Reviewed-by: Greg Kurz <groug at kaod.org>> > The bug was already discussed here: > http://www.spinics.net/lists/kvm/msg144365.html > This is a follow up patch. > > --- > drivers/vhost/vhost.c | 10 ++++------ > 1 file changed, 4 insertions(+), 6 deletions(-) > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c > index d643260..8f99fe0 100644 > --- a/drivers/vhost/vhost.c > +++ b/drivers/vhost/vhost.c > @@ -130,14 +130,14 @@ static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx, > > static void vhost_init_is_le(struct vhost_virtqueue *vq) > { > - if (vhost_has_feature(vq, VIRTIO_F_VERSION_1)) > - vq->is_le = true; > + vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) > + || virtio_legacy_is_little_endian(); > } > #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */ > > static void vhost_reset_is_le(struct vhost_virtqueue *vq) > { > - vq->is_le = virtio_legacy_is_little_endian(); > + vhost_init_is_le(vq); > } > > struct vhost_flush_struct { > @@ -1714,10 +1714,8 @@ int vhost_vq_init_access(struct vhost_virtqueue *vq) > int r; > bool is_le = vq->is_le; > > - if (!vq->private_data) { > - vhost_reset_is_le(vq); > + if (!vq->private_data) > return 0; > - } > > vhost_init_is_le(vq); >
On 01/30/2017 08:06 PM, Greg Kurz wrote:>> Currently, under certain circumstances vhost_init_is_le does just a part >> of the initialization job, and depends on vhost_reset_is_le being called >> too. For this reason vhost_vq_init_access used to call vhost_reset_is_le >> when vq->private_data is NULL. This is not only counter intuitive, but >> also real a problem because it breaks vhost_net. The bug was introduced to >> vhost_net with commit 2751c9882b94 ("vhost: cross-endian support for >> legacy devices"). The symptom is corruption of the vq's used.idx field >> (virtio) after VHOST_NET_SET_BACKEND was issued as a part of the vhost >> shutdown on a vq with pending descriptors. >> >> Let us make sure the outcome of vhost_init_is_le never depend on the state >> it is actually supposed to initialize, and fix virtio_net by removing the >> reset from vhost_vq_init_access. >> >> With the above, there is no reason for vhost_reset_is_le to do just half >> of the job. Let us make vhost_reset_is_le reinitialize is_le. >> >> Signed-off-by: Halil Pasic <pasic at linux.vnet.ibm.com> >> Reported-by: Michael A. Tebolt <miket at us.ibm.com> >> Reported-by: Dr. David Alan Gilbert <dgilbert at redhat.com> >> Fixes: commit 2751c9882b94 ("vhost: cross-endian support for legacy devices") >> --- > Reviewed-by: Greg Kurz <groug at kaod.org> >Thanks! We have some tests on s390x (that is BE) running, but I won't be able to test the change with cross endian and legacy. What do you think, should I/we RFT or are we fine without? Regards, Halil
On 2017?01?30? 18:09, Halil Pasic wrote:> Currently, under certain circumstances vhost_init_is_le does just a part > of the initialization job, and depends on vhost_reset_is_le being called > too. For this reason vhost_vq_init_access used to call vhost_reset_is_le > when vq->private_data is NULL. This is not only counter intuitive, but > also real a problem because it breaks vhost_net. The bug was introduced to > vhost_net with commit 2751c9882b94 ("vhost: cross-endian support for > legacy devices"). The symptom is corruption of the vq's used.idx field > (virtio) after VHOST_NET_SET_BACKEND was issued as a part of the vhost > shutdown on a vq with pending descriptors. > > Let us make sure the outcome of vhost_init_is_le never depend on the state > it is actually supposed to initialize, and fix virtio_net by removing the > reset from vhost_vq_init_access. > > With the above, there is no reason for vhost_reset_is_le to do just half > of the job. Let us make vhost_reset_is_le reinitialize is_le. > > Signed-off-by: Halil Pasic <pasic at linux.vnet.ibm.com> > Reported-by: Michael A. Tebolt <miket at us.ibm.com> > Reported-by: Dr. David Alan Gilbert <dgilbert at redhat.com> > Fixes: commit 2751c9882b94 ("vhost: cross-endian support for legacy devices") > --- > > The bug was already discussed here: > http://www.spinics.net/lists/kvm/msg144365.html > This is a follow up patch. > > --- > drivers/vhost/vhost.c | 10 ++++------ > 1 file changed, 4 insertions(+), 6 deletions(-) > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c > index d643260..8f99fe0 100644 > --- a/drivers/vhost/vhost.c > +++ b/drivers/vhost/vhost.c > @@ -130,14 +130,14 @@ static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx, > > static void vhost_init_is_le(struct vhost_virtqueue *vq) > { > - if (vhost_has_feature(vq, VIRTIO_F_VERSION_1)) > - vq->is_le = true; > + vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) > + || virtio_legacy_is_little_endian(); > } > #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */ > > static void vhost_reset_is_le(struct vhost_virtqueue *vq) > { > - vq->is_le = virtio_legacy_is_little_endian(); > + vhost_init_is_le(vq); > } > > struct vhost_flush_struct { > @@ -1714,10 +1714,8 @@ int vhost_vq_init_access(struct vhost_virtqueue *vq) > int r; > bool is_le = vq->is_le; > > - if (!vq->private_data) { > - vhost_reset_is_le(vq); > + if (!vq->private_data) > return 0; > - } > > vhost_init_is_le(vq); >Acked-by: Jason Wang <jasowang at redhat.com> We can probably just drop vhost_reset_is_le() and just use vhost_init_is_le() instead. Thanks