Re: [PATCH] drm/ttm: fix swapped-out resources never leaving their bulk_move range

[Christian König]

On 9/10/26 09:14, Thomas Hellström wrote:
> On Wed, 2026-09-09 at 23:50 +0300, Vadim Nikitushkin wrote:
>> ttm_tt_swapout() returns the number of pages swapped out on success
>> and
>> a negative error code on failure; for a populated ttm it never
>> returns
>> zero. Commit b2ed01e7ad3d ("drm/ttm: Fix ttm_bo_swapout() infinite
>> LRU
>> walk on swapout failure") moved the bulk_move bookkeeping in
>> ttm_bo_swapout_cb() under "if (!ret)", so the
>> ttm_resource_del_bulk_move_unevictable() /
>> ttm_resource_move_to_lru_tail()
>> pair is now skipped on every successful swapout. The equivalent
>> change
>> for the shrinker in commit 1d59f36e95f7 ("drm/ttm: Fix
>> ttm_bo_shrink()
>> infinite LRU walk on backup failure") tests "lret > 0", which is what
>> was intended here as well.
>>
>> Before b2ed01e7ad3d the resource was taken off the bulk_move before
>> the
>> swapout; since then a swapped-out resource stays inside its BO's
>> bulk_move range (and on the manager LRU) although it is unevictable.
>> When it is later freed or the BO leaves the bulk_move
>> (ttm_resource_free(), ttm_bo_set_bulk_move() via amdgpu_vm_bo_del()),
>> ttm_resource_del_bulk_move() skips it because of its
>> !ttm_resource_unevictable() guard, so a range endpoint in pos->first
>> /
>> pos->last is left pointing at freed memory. The next
>> ttm_lru_bulk_move_tail() or ttm_resource_add_bulk_move() on that
>> cursor
>> is a use-after-free, seen as the resv WARN in
>> ttm_lru_bulk_move_add(),
>> "list_del corruption" in ttm_resource_move_to_lru_tail() or a NULL
>> dereference in ttm_resource_manager_next() -- minutes to hours after
>> a
>> hibernation, or at process exit / reboot following one. Samuel
>> Ainsworth's analysis of drm/amd issue 5387 (see Link) identified the
>> dangling cursor; the missing removal at swapout time is the reason it
>> dangles.
>>
>> Testing the condition for success restores the removal. On an AMD
>> Phoenix APU (ASUS UM3406GA, gfx1103) running suspend-then-hibernate
>> on
>> a 7.0.y stable kernel carrying the backport (Ubuntu 7.0.0-31) the bug
>> crashed 5 of 18 hibernation cycles; a function profile of one
>> hibernation showed 336 ttm_tt_swapout() calls and zero
>> ttm_resource_del_bulk_move_unevictable() calls. With this change the
>> removal happens for every swapped-out resource and 12 further cycles
>> were clean.
>>
>> Fixes: b2ed01e7ad3d ("drm/ttm: Fix ttm_bo_swapout() infinite LRU walk
>> on swapout failure")
>> Cc: stable@xxxxxxxxxxxxxxx # v7.1+
>> Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/5387
>> Link:
>> https://lore.kernel.org/dri-devel/CAHYiNPa6aVacJoLOje-qZ1GyYx-9p0tN4NuP8D_eSL+UJeevXw@xxxxxxxxxxxxxx/
>> Signed-off-by: Vadim Nikitushkin <bub4z0r@xxxxxxxxx>
> 
> Nice catch.

Agreed, that is a really good one. We had tons of people staring at the code without seeing that.

> 
> This also explains why https://patchwork.freedesktop.org/series/170311/
> appeared to fix the issue. But that series actually kept the resource
> on the bulk sublist until someone bumped the LRU or removed it.
> 
> Reviewed-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>

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

If nobody comes up with some last second objections I'm going to push that to drm-misc-fixes ASAP.

Thanks,
Christian.

> 
>> ---
>>  drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
>> b/drivers/gpu/drm/ttm/ttm_bo.c
>> index ef56c18..9b85b5f 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -1434,7 +1434,7 @@ ttm_bo_swapout_cb(struct ttm_lru_walk *walk,
>> struct ttm_buffer_object *bo)
>>  
>>  	if (ttm_tt_is_populated(tt)) {
>>  		ret = ttm_tt_swapout(bdev, tt, swapout_walk-
>>> gfp_flags);
>> -		if (!ret) {
>> +		if (ret > 0) {
>>  			spin_lock(&bdev->lru_lock);
>>  			ttm_resource_del_bulk_move_unevictable(bo-
>>> resource, bo);
>>  			ttm_resource_move_to_lru_tail(bo->resource);