On 03.07.2015 01:54, Grigori Goronzy wrote:
> We don't need to call the (expensive) radeon_bo_wait, checking the
> fences via RCU is much faster. The reservation done by radeon_bo_wait
> does not save us from any race conditions.
>
> Signed-off-by: Grigori Goronzy <greg at chown.ath.cx>
Patche #1-#3 are Reviewed-by: Christian König <christian.koenig at amd.com>
Patch #4 is a nice catch, but as Michel already noted needs more
thoughts for a complete fix.
Regards,
Christian.
> ---
> drivers/gpu/drm/radeon/radeon_gem.c | 11 ++++++++---
> 1 file changed, 8 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/radeon/radeon_gem.c
> b/drivers/gpu/drm/radeon/radeon_gem.c
> index ac3c131..7199e19 100644
> --- a/drivers/gpu/drm/radeon/radeon_gem.c
> +++ b/drivers/gpu/drm/radeon/radeon_gem.c
> @@ -428,7 +428,6 @@ int radeon_gem_mmap_ioctl(struct drm_device *dev, void
> *data,
> int radeon_gem_busy_ioctl(struct drm_device *dev, void *data,
> struct drm_file *filp)
> {
> - struct radeon_device *rdev = dev->dev_private;
> struct drm_radeon_gem_busy *args = data;
> struct drm_gem_object *gobj;
> struct radeon_bo *robj;
> @@ -440,10 +439,16 @@ int radeon_gem_busy_ioctl(struct drm_device *dev, void
> *data,
> return -ENOENT;
> }
> robj = gem_to_radeon_bo(gobj);
> - r = radeon_bo_wait(robj, &cur_placement, true);
> +
> + r = reservation_object_test_signaled_rcu(robj->tbo.resv, true);
> + if (r == 0)
> + r = -EBUSY;
> + else
> + r = 0;
> +
> + cur_placement = ACCESS_ONCE(robj->tbo.mem.mem_type);
> args->domain = radeon_mem_type_to_domain(cur_placement);
> drm_gem_object_unreference_unlocked(gobj);
> - r = radeon_gem_handle_lockup(rdev, r);
> return r;
> }
>
