[Why]
In multi-monitor scenario, if first crtc's flip done event occurs delayed
(but within timeout), due to non-blocking design of commit_tail(), there
are more than one commit's scheduled by the time the second crtc's
wait_for_completion_timeout() is called in drm_atomic_helper_wait_for_flip_done.

Due to these in-between additions and deletions in the atomic state, it is
found that the system fails while accessing common data structures of the
second crtc in drm_atomic_helper_wait_for_flip_done(), leading to crash as
below:

        BUG: unable to handle kernel paging request at 000000010000001c
        IP: do_raw_spin_lock+0xf/0x94
        PGD 0 P4D 0
        Oops: 0000 [#1] PREEMPT SMP NOPTI
        Call Trace:
         __wait_for_common+0x36/0x60
         drm_atomic_helper_wait_for_flip_done+0x47/0x89
         amdgpu_dm_atomic_commit_tail+0xf4b/0xf84
         ? drm_atomic_helper_wait_for_dependencies+0x1cd/0x217
         commit_tail+0x41/0x5f

[How]
Move drm_atomic_helper_commit_hw_done() post wait_for_flip_done(),
which cleans up the atomic state's commit and completes pending hw_done and
flip_done works as a result there wont be dangling flip waits on random commits.

Signed-off-by: Shirish S <shiris...@amd.com>
---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c 
b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
index 0f10d92..41a1958 100644
--- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
+++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
@@ -4626,12 +4626,17 @@ static void amdgpu_dm_atomic_commit_tail(struct 
drm_atomic_state *state)
        }
        spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
 
-       /* Signal HW programming completion */
-       drm_atomic_helper_commit_hw_done(state);
 
        if (wait_for_vblank)
                drm_atomic_helper_wait_for_flip_done(dev, state);
 
+       /* Atomic state pointer gets corrupted in case of frequent
+        * modesets operations like changing resolutions.
+        * Hence discard state->commit before signalling to user
+        * space.
+        */
+       drm_atomic_helper_commit_hw_done(state);
+
        drm_atomic_helper_cleanup_planes(dev, state);
 
        /*
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Reply via email to