drm/radeon: fix race between GPU reset and TTM delayed delete thread.
Dave Airlie [Thu, 10 Feb 2011 04:46:06 +0000 (14:46 +1000)]
My evergreen has been in a remote PC for week and reset has never once
saved me from certain doom, I finally relocated to the box with a
serial cable and noticed an oops when the GPU resets, and the TTM
delayed delete thread tries to remove something from the GTT.

This stops the delayed delete thread from executing across the GPU
reset handler, and woot I can GPU reset now.

Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/radeon/radeon_device.c

index 0d47893..4954e2d 100644 (file)
@@ -936,8 +936,11 @@ int radeon_resume_kms(struct drm_device *dev)
 int radeon_gpu_reset(struct radeon_device *rdev)
 {
        int r;
+       int resched;
 
        radeon_save_bios_scratch_regs(rdev);
+       /* block TTM */
+       resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
        radeon_suspend(rdev);
 
        r = radeon_asic_reset(rdev);
@@ -946,6 +949,7 @@ int radeon_gpu_reset(struct radeon_device *rdev)
                radeon_resume(rdev);
                radeon_restore_bios_scratch_regs(rdev);
                drm_helper_resume_force_mode(rdev->ddev);
+               ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
                return 0;
        }
        /* bad news, how to tell it to userspace ? */