Linux Virtualization - Sep 2019 - [PATCH] drm/virtio: fix command submission with objects but without fence.

Only call virtio_gpu_array_add_fence if we actually have a fence.

Fixes: da758d51968a ("drm/virtio: rework virtio_gpu_execbuffer_ioctl
fencing")
Signed-off-by: Gerd Hoffmann <kraxel at redhat.com>
---
 drivers/gpu/drm/virtio/virtgpu_vq.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/virtio/virtgpu_vq.c
b/drivers/gpu/drm/virtio/virtgpu_vq.c
index 595fa6ec2d58..7fd2851f7b97 100644
--- a/drivers/gpu/drm/virtio/virtgpu_vq.c
+++ b/drivers/gpu/drm/virtio/virtgpu_vq.c
@@ -339,11 +339,12 @@ static void virtio_gpu_queue_fenced_ctrl_buffer(struct
virtio_gpu_device *vgdev,
 		goto again;
 	}
 
-	if (fence)
+	if (fence) {
 		virtio_gpu_fence_emit(vgdev, hdr, fence);
-	if (vbuf->objs) {
-		virtio_gpu_array_add_fence(vbuf->objs, &fence->f);
-		virtio_gpu_array_unlock_resv(vbuf->objs);
+		if (vbuf->objs) {
+			virtio_gpu_array_add_fence(vbuf->objs, &fence->f);
+			virtio_gpu_array_unlock_resv(vbuf->objs);
+		}
 	}
 	notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
 	spin_unlock(&vgdev->ctrlq.qlock);
-- 
2.18.1

On Wed, Sep 4, 2019 at 12:48 AM Gerd Hoffmann <kraxel at redhat.com>
wrote:
>
> Only call virtio_gpu_array_add_fence if we actually have a fence.
>
> Fixes: da758d51968a ("drm/virtio: rework virtio_gpu_execbuffer_ioctl
fencing")
> Signed-off-by: Gerd Hoffmann <kraxel at redhat.com>
> ---
>  drivers/gpu/drm/virtio/virtgpu_vq.c | 9 +++++----
>  1 file changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/virtio/virtgpu_vq.c
b/drivers/gpu/drm/virtio/virtgpu_vq.c
> index 595fa6ec2d58..7fd2851f7b97 100644
> --- a/drivers/gpu/drm/virtio/virtgpu_vq.c
> +++ b/drivers/gpu/drm/virtio/virtgpu_vq.c
> @@ -339,11 +339,12 @@ static void
virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
>                 goto again;
>         }
>
> -       if (fence)
> +       if (fence) {
>                 virtio_gpu_fence_emit(vgdev, hdr, fence);
> -       if (vbuf->objs) {
> -               virtio_gpu_array_add_fence(vbuf->objs,
&fence->f);
> -               virtio_gpu_array_unlock_resv(vbuf->objs);
> +               if (vbuf->objs) {
> +                       virtio_gpu_array_add_fence(vbuf->objs,
&fence->f);
> +                       virtio_gpu_array_unlock_resv(vbuf->objs);
> +               }
This leaks when fence == NULL and vbuf->objs != NULL (which can really
happen IIRC... not at my desk to check).

>         }
>         notify = virtio_gpu_queue_ctrl_buffer_locked(vgdev, vbuf);
>         spin_unlock(&vgdev->ctrlq.qlock);
> --
> 2.18.1
>

On Wed, Sep 04, 2019 at 04:10:30PM -0700, Chia-I Wu
wrote:
> On Wed, Sep 4, 2019 at 12:48 AM Gerd Hoffmann <kraxel at redhat.com>
wrote:
> >
> > Only call virtio_gpu_array_add_fence if we actually have a fence.
> >
> > Fixes: da758d51968a ("drm/virtio: rework
virtio_gpu_execbuffer_ioctl fencing")
> > Signed-off-by: Gerd Hoffmann <kraxel at redhat.com>
> > ---
> >  drivers/gpu/drm/virtio/virtgpu_vq.c | 9 +++++----
> >  1 file changed, 5 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/virtio/virtgpu_vq.c
b/drivers/gpu/drm/virtio/virtgpu_vq.c
> > index 595fa6ec2d58..7fd2851f7b97 100644
> > --- a/drivers/gpu/drm/virtio/virtgpu_vq.c
> > +++ b/drivers/gpu/drm/virtio/virtgpu_vq.c
> > @@ -339,11 +339,12 @@ static void
virtio_gpu_queue_fenced_ctrl_buffer(struct virtio_gpu_device *vgdev,
> >                 goto again;
> >         }
> >
> > -       if (fence)
> > +       if (fence) {
> >                 virtio_gpu_fence_emit(vgdev, hdr, fence);
> > -       if (vbuf->objs) {
> > -               virtio_gpu_array_add_fence(vbuf->objs,
&fence->f);
> > -               virtio_gpu_array_unlock_resv(vbuf->objs);
> > +               if (vbuf->objs) {
> > +                       virtio_gpu_array_add_fence(vbuf->objs,
&fence->f);
> > +                       virtio_gpu_array_unlock_resv(vbuf->objs);
> > +               }
> This leaks when fence == NULL and vbuf->objs != NULL (which can really
> happen IIRC... not at my desk to check).
Yes, it can happen, for example when flushing dumb buffers.

But I don't think we leak in this case.  The code paths which don't need
a fence also do not call virtio_gpu_array_lock_resv(), so things are
balanced.  The actual release of the objs happens in
virtio_gpu_dequeue_ctrl_func() via virtio_gpu_array_put_free_delayed().

cheers,
  Gerd