drm/ttm: Remove pinned bos from LRU in ttm_bo_move_to_lru_tail() v2
authorLyude Paul <lyude@redhat.com>
Tue, 5 Jan 2021 11:45:05 +0000 (12:45 +0100)
committerLyude Paul <lyude@redhat.com>
Wed, 6 Jan 2021 17:50:57 +0000 (12:50 -0500)
Recently a regression was introduced which caused TTM's buffer eviction to
attempt to evict already-pinned BOs, causing issues with buffer eviction
under memory pressure along with suspend/resume:

  nouveau 0000:1f:00.0: DRM: evicting buffers...
  nouveau 0000:1f:00.0: DRM: Moving pinned object 00000000c428c3ff!
  nouveau 0000:1f:00.0: fifo: fault 00 [READ] at 0000000000200000 engine 04
  [BAR1] client 07 [HUB/HOST_CPU] reason 02 [PTE] on channel -1 [00ffeaa000
  unknown]
  nouveau 0000:1f:00.0: fifo: DROPPED_MMU_FAULT 00001000
  nouveau 0000:1f:00.0: fifo: fault 01 [WRITE] at 0000000000020000 engine
  0c [HOST6] client 07 [HUB/HOST_CPU] reason 02 [PTE] on channel 1
  [00ffb28000 DRM]
  nouveau 0000:1f:00.0: fifo: channel 1: killed
  nouveau 0000:1f:00.0: fifo: runlist 0: scheduled for recovery
  [TTM] Buffer eviction failed
  nouveau 0000:1f:00.0: DRM: waiting for kernel channels to go idle...
  nouveau 0000:1f:00.0: DRM: failed to idle channel 1 [DRM]
  nouveau 0000:1f:00.0: DRM: resuming display...

After some bisection and investigation, it appears this resulted from the
recent changes to ttm_bo_move_to_lru_tail(). Previously when a buffer was
pinned, the buffer would be removed from the LRU once ttm_bo_unreserve
to maintain the LRU list when pinning or unpinning BOs. However, since:

commit 3d1a88e1051f ("drm/ttm: cleanup LRU handling further")

We've been exiting from ttm_bo_move_to_lru_tail() at the very beginning of
the function if the bo we're looking at is pinned, resulting in the pinned
BO never getting removed from the lru and as a result - causing issues when
it eventually becomes time for eviction.

So, let's fix this by calling ttm_bo_del_from_lru() from
ttm_bo_move_to_lru_tail() in the event that we're dealing with a pinned
buffer.

v2 (chk): reduce to only the fixing one liner since we always want to
call the callback whenever we would move on the LRU.

Fixes: 3d1a88e1051f ("drm/ttm: cleanup LRU handling further")
Cc: Dave Airlie <airlied@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210105114505.38210-1-christian.koenig@amd.com
drivers/gpu/drm/ttm/ttm_bo.c

index 31e8b3d..b65f4b1 100644 (file)
@@ -138,8 +138,10 @@ void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
 
        dma_resv_assert_held(bo->base.resv);
 
-       if (bo->pin_count)
+       if (bo->pin_count) {
+               ttm_bo_del_from_lru(bo);
                return;
+       }
 
        man = ttm_manager_type(bdev, mem->mem_type);
        list_move_tail(&bo->lru, &man->lru[bo->priority]);