Re: [PATCH] drm/ttm: Remove pinned bos from LRU in ttm_bo_move_to_lru_tail()

[Christian König]

Am 05.01.21 um 00:13 schrieb Lyude Paul:
> 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. As well, add back the hunks in ttm_bo_del_from_lru() that were
> removed which checked whether we want to call
> bdev->driver->del_from_lru_notify() or not. We do this last part to avoid
> calling the hook when the bo in question was already removed from the LRU.
>
> Signed-off-by: Lyude Paul <lyude@xxxxxxxxxx>
> Fixes: 3d1a88e1051f ("drm/ttm: cleanup LRU handling further")
> Cc: Christian König <christian.koenig@xxxxxxx>
> Cc: Dave Airlie <airlied@xxxxxxxxxx>

I was already working on a fix as well, but you have been faster than me :)

Reviewed-by: Christian König <christian.koenig@xxxxxxx>

Going to pick this up for drm-misc-next or drm-misc-fixes in a few minutes.

Thanks,
Christian.

> ---
>   drivers/gpu/drm/ttm/ttm_bo.c | 20 ++++++++++++++++----
>   1 file changed, 16 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index 31e8b3da5563..0f373b78e7fa 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -113,11 +113,18 @@ static struct kobj_type ttm_bo_glob_kobj_type  = {
>   static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
>   {
>   	struct ttm_bo_device *bdev = bo->bdev;
> +	bool notify = false;
>   
> -	list_del_init(&bo->swap);
> -	list_del_init(&bo->lru);
> +	if (!list_empty(&bo->swap)) {
> +		notify = true;
> +		list_del_init(&bo->swap);
> +	}
> +	if (!list_empty(&bo->lru)) {
> +		notify = true;
> +		list_del_init(&bo->lru);
> +	}
>   
> -	if (bdev->driver->del_from_lru_notify)
> +	if (notify && bdev->driver->del_from_lru_notify)
>   		bdev->driver->del_from_lru_notify(bo);
>   }
>   
> @@ -138,8 +145,13 @@ void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
>   
>   	dma_resv_assert_held(bo->base.resv);
>   
> -	if (bo->pin_count)
> +	/* Pinned bos will have been added to the LRU before they were pinned, so make sure we
> +	 * always remove them here
> +	 */
> +	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]);