Re: [PATCH] drm/virtio: exbuf->fence_fd unmodified on interrupted wait

[Ryan Neph]

On Thu, Feb 02, 2023 at 05:24:34AM +0300, Dmitry Osipenko wrote:
> On 2/2/23 05:17, Dmitry Osipenko wrote:
> > On 2/1/23 18:48, Rob Clark wrote:
> >> On Wed, Feb 1, 2023 at 5:28 AM Dmitry Osipenko
> >> <dmitry.osipenko@xxxxxxxxxxxxx> wrote:
> >>>
> >>> On 1/27/23 01:58, Ryan Neph wrote:
> >>>> An interrupted dma_fence_wait() becomes an -ERESTARTSYS returned
> >>>> to userspace ioctl(DRM_IOCTL_VIRTGPU_EXECBUFFER) calls, prompting to
> >>>> retry the ioctl(), but the passed exbuf->fence_fd has been reset to -1,
> >>>> making the retry attempt fail at sync_file_get_fence().
> >>>>
> >>>> The uapi for DRM_IOCTL_VIRTGPU_EXECBUFFER is changed to retain the
> >>>> passed value for exbuf->fence_fd when returning ERESTARTSYS or EINTR.
> >>>>
> >>>> Fixes: 2cd7b6f08bc4 ("drm/virtio: add in/out fence support for explicit synchronization")
> >>>> Signed-off-by: Ryan Neph <ryanneph@xxxxxxxxxxxx>
> >>>> ---
> >>>>
> >>>>  drivers/gpu/drm/virtio/virtgpu_ioctl.c | 9 ++++++---
> >>>>  include/uapi/drm/virtgpu_drm.h         | 3 +++
> >>>>  2 files changed, 9 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/virtio/virtgpu_ioctl.c b/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> >>>> index 9f4a90493aea..ffce4e2a409a 100644
> >>>> --- a/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> >>>> +++ b/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> >>>> @@ -132,6 +132,8 @@ static int virtio_gpu_execbuffer_ioctl(struct drm_device *dev, void *data,
> >>>>       uint64_t fence_ctx;
> >>>>       uint32_t ring_idx;
> >>>>
> >>>> +     exbuf->fence_fd = -1;
> >>>> +
> >>>>       fence_ctx = vgdev->fence_drv.context;
> >>>>       ring_idx = 0;
> >>>>
> >>>> @@ -152,8 +154,6 @@ static int virtio_gpu_execbuffer_ioctl(struct drm_device *dev, void *data,
> >>>>               ring_idx = exbuf->ring_idx;
> >>>>       }
> >>>>
> >>>> -     exbuf->fence_fd = -1;
> >>>
> >>> Is there any userspace relying on this -1 behaviour? Wouldn't be better
> >>> to remove this offending assignment?
> >>
> >> Looking at current mesa, removing the assignment should be ok (and
> >> more consistent with other drivers).  But I can't say if this was
> >> always true, or that there aren't other non-mesa users, so I can see
> >> the argument for the more conservative uabi change that this patch
> >> went with.
> > 
> > Realistically, Mesa is the only user of this IOCTL. In general, in a
> > such case of doubt, I'll do the UABI change and then wait for complains.
> > If there is a complaint, then the change is reverted. Also will be good
> > to know about existence of other users :)
> > 
> > Given that -1 already wasn't consistently set for all error code paths,
> > it's tempting to see it removed.
> > 
> > The code change of this patch is trivial, hence should fine to keep the
> > -1 if you prefer that, but the patch won't apply cleanly to the stable
> > kernels because of the "exbuf->fence_fd = -1" movement. If stable
> > maintainers won't put effort into rebasing the patch, then better to do
> > the removal and live with a cleaner driver code, IMO.
> 
> Although, there will be a merge conflict either way. I'll give the r-b,
> still removing -1 feels more attractive to me.

I'm not opposed to removing the "exbuf->fence_fd = -1" on error. I made the
v1 patch with extra care to fix the known issue while minimizing the uabi
change, but I'd prefer to see it changed too; thanks for the comments.

I'll follow up with a v2 that removes the modification of "exbuf->fence_fd"
unless the IOCTL succeeds.