On 16.02.2017 02:00, Michel Dänzer wrote:
> On 16/02/17 04:10 AM, Nicolai Hähnle wrote:
>> From: Nicolai Hähnle <nicolai.haehnle at amd.com>
>>
>> Open-code the initial ttm_bo_validate call, so that we can properly
>> unlock the reservation lock when it fails. Also, properly destruct
>> the reservation object when the first part of TTM BO initialization
>> fails.
>>
>> Actual deadlocks caused by the missing unlock should have been fixed
>> by "drm/ttm: never add BO that failed to validate to the LRU list",
>> superseding the flawed fix in commit 38fc4856ad98 ("drm/amdgpu: fix
>> a potential deadlock in amdgpu_bo_create_restricted()").
>>
>> This change fixes remaining recursive locking errors that can be seen
>> with lock debugging enabled, and avoids the error of freeing a locked
>> mutex.
>>
>> v2: use ttm_bo_unref for error handling, and rebase on latest changes
>>
>> Fixes: 12a852219583 ("drm/amdgpu: improve AMDGPU_GEM_CREATE_VRAM_CLEARED handling (v2)")
>> Signed-off-by: Nicolai Hähnle <nicolai.haehnle at amd.com>
>> ---
>> drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 28 ++++++++++++++++++++++------
>> 1 file changed, 22 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> index c2e57f7..15944ea 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> @@ -404,15 +404,31 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
>> }
>>
>> initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
>> - r = ttm_bo_init(&adev->mman.bdev, &bo->tbo, size, type,
>> - &bo->placement, page_align, !kernel, NULL,
>> - acc_size, sg, resv ? resv : &bo->tbo.ttm_resv,
>> - &amdgpu_ttm_bo_destroy);
>> + r = ttm_bo_init_top(&adev->mman.bdev, &bo->tbo, size, type,
>> + page_align, NULL,
>> + acc_size, sg, resv ? resv : &bo->tbo.ttm_resv,
>> + &amdgpu_ttm_bo_destroy);
>> +
>> + if (likely(r == 0))
>> + r = ttm_bo_validate(&bo->tbo, &bo->placement, !kernel, false);
>> +
>> + if (unlikely(r != 0)) {
>> + struct ttm_buffer_object *tbo = &bo->tbo;
>> +
>> + if (!resv)
>> + ww_mutex_unlock(&bo->tbo.ttm_resv.lock);
>> + ttm_bo_unref(&tbo);
>> + return r;
>> + }
>
> I think this would be clearer as
>
> if (unlikely(r != 0)) {
> struct ttm_buffer_object *tbo = &bo->tbo;
>
> if (!resv)
> ww_mutex_unlock(&bo->tbo.ttm_resv.lock);
> ttm_bo_unref(&tbo);
> return r;
> }
>
> r = ttm_bo_validate(&bo->tbo, &bo->placement, !kernel, false);
> [...]
>
>
>> amdgpu_cs_report_moved_bytes(adev,
>> atomic64_read(&adev->num_bytes_moved) - initial_bytes_moved);
>>
>> - if (unlikely(r != 0))
>> - return r;
>> + if (!(bo->tbo.mem.placement & TTM_PL_FLAG_NO_EVICT)) {
>> + spin_lock(&bo->tbo.glob->lru_lock);
>> + ttm_bo_add_to_lru(&bo->tbo);
>> + spin_unlock(&bo->tbo.glob->lru_lock);
>> + }
>
> It's a bit ugly to open-code this logic in driver code. Maybe add
> another TTM helper which calls ttm_bo_validate and ttm_bo_add_to_lru?
On further reflection, maybe a better approach would be to have a
function ttm_bo_init_reserved, which returns with the reservation lock
held upon success. We can then just change amdgpu_bo_create_restricted
to ttm_bo_unreserve instead of ww_mutex_unlock, which does the add-to-LRU.
AFAICT, it doesn't make much sense to add the buffer to the LRU that
early anyway, if amdgpu_fill_buffer is used.
Cheers,
Nicolai
>
>
