[PATCH v4] drm: do not sleep on vblank while holding a mutex

[PATCH v4] drm: do not sleep on vblank while holding a mutex

[Ilija Hadzic]

drm_wait_vblank must be DRM_UNLOCKED because otherwise it
will grab the drm_global_mutex and then go to sleep until the vblank
event it is waiting for. That can wreck havoc in the windowing system
because if one process issues this ioctl, it will block all other
processes for the duration of all vblanks between the current and the
one it is waiting for. In some cases it can block the entire windowing
system.

v2: incorporate comments received from Daniel Vetter and
    Michel Daenzer.

v3/v4: after a lengty discussion with Daniel Vetter, it was concluded
       that the only thing not yet protected with locks and atomic
       ops is the write to dev->last_vblank_wait. It's only used in a
       debug file in proc, and the current code already employs no
       correct locking: the proc file only takes dev->struct_mutex,
       whereas drm_wait_vblank implicitly took the drm_global_mutex.
       Given all this, it's not worth bothering to try to fix
       the locks at this time.

Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
---
 drivers/gpu/drm/drm_drv.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/drivers/gpu/drm/drm_drv.c b/drivers/gpu/drm/drm_drv.c
index dbabcb0..dc0eb0b 100644
--- a/drivers/gpu/drm/drm_drv.c
+++ b/drivers/gpu/drm/drm_drv.c
@@ -124,7 +124,7 @@ static struct drm_ioctl_desc drm_ioctls[] = {
 	DRM_IOCTL_DEF(DRM_IOCTL_SG_ALLOC, drm_sg_alloc_ioctl, DRM_AUTH|DRM_MASTER|DRM_ROOT_ONLY),
 	DRM_IOCTL_DEF(DRM_IOCTL_SG_FREE, drm_sg_free, DRM_AUTH|DRM_MASTER|DRM_ROOT_ONLY),
 
-	DRM_IOCTL_DEF(DRM_IOCTL_WAIT_VBLANK, drm_wait_vblank, 0),
+	DRM_IOCTL_DEF(DRM_IOCTL_WAIT_VBLANK, drm_wait_vblank, DRM_UNLOCKED),
 
 	DRM_IOCTL_DEF(DRM_IOCTL_MODESET_CTL, drm_modeset_ctl, 0),
 
-- 
1.7.7

Re: [PATCH v4] drm: do not sleep on vblank while holding a mutex

[Daniel Vetter]

On Mon, Oct 31, 2011 at 05:46:18PM -0400, Ilija Hadzic wrote:
> drm_wait_vblank must be DRM_UNLOCKED because otherwise it
> will grab the drm_global_mutex and then go to sleep until the vblank
> event it is waiting for. That can wreck havoc in the windowing system
> because if one process issues this ioctl, it will block all other
> processes for the duration of all vblanks between the current and the
> one it is waiting for. In some cases it can block the entire windowing
> system.
> 
> v2: incorporate comments received from Daniel Vetter and
>     Michel Daenzer.
> 
> v3/v4: after a lengty discussion with Daniel Vetter, it was concluded
>        that the only thing not yet protected with locks and atomic
>        ops is the write to dev->last_vblank_wait. It's only used in a
>        debug file in proc, and the current code already employs no
>        correct locking: the proc file only takes dev->struct_mutex,
>        whereas drm_wait_vblank implicitly took the drm_global_mutex.
>        Given all this, it's not worth bothering to try to fix
>        the locks at this time.
> 
> Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
-- 
Daniel Vetter
Mail: daniel@ffwll.ch
Mobile: +41 (0)79 365 57 48