[PATCH] drm/ttm: fix pipelined gutting for evictions

Felix Kuehling felix.kuehling at amd.com
Fri Jul 24 02:56:25 UTC 2020


Am 2020-07-23 um 9:58 p.m. schrieb philip yang:
>
> On 2020-07-23 7:02 p.m., Felix Kuehling wrote:
>> Am 2020-07-23 um 5:00 a.m. schrieb Christian König:
>>> We can't pipeline that during eviction because the memory needs
>>> to be available immediately.
>>>
>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>> ---
>>>   drivers/gpu/drm/ttm/ttm_bo.c | 12 ++++++++++--
>>>   1 file changed, 10 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
>>> b/drivers/gpu/drm/ttm/ttm_bo.c
>>> index bc2230ecb7e3..122040056a07 100644
>>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>>> @@ -651,8 +651,16 @@ static int ttm_bo_evict(struct
>>> ttm_buffer_object *bo,
>>>       placement.num_busy_placement = 0;
>>>       bdev->driver->evict_flags(bo, &placement);
>>>   -    if (!placement.num_placement && !placement.num_busy_placement)
>>> -        return ttm_bo_pipeline_gutting(bo);
>>> +    if (!placement.num_placement && !placement.num_busy_placement) {
>>> +        ttm_bo_wait(bo, false, false);
>>> +
>>> +        ttm_tt_destroy(bo->ttm);
>>> +
>>> +        memset(&bo->mem, 0, sizeof(bo->mem));
>> Where does the memory in the bo->mem (ttm_mem_reg) get destroyed? It
>> doesn't get attached to a ghost BO in this case, so someone will have to
>> call ttm_bo_mem_put explicitly before you wipe out bo->mem.
>
> After migrating to ram,
> svm_range_bo_unref-->amdgpu_unref_bo->ttm_bo_put->ttm_bo_release calls
> ttm_bo_mem_put.

amdgpu_bo_unref won't free anything if the reference count doesn't go to
0. And TTM is still holding a reference here. The BO won't be freed
until ttm_mem_evict_first calls ttm_bo_put.

The memset above overwrites the ttm_mem_reg structure, which means it
will forget about the VRAM nodes held by the BO. They need to be
released first. That's what frees the space that ttm_bo_evict was trying
to make available in the first place.

Regards,
  Felix


> vram is already freed before we signal fence, right?
>
> Regards,
>
> Philip
>
>> Regards,
>>    Felix
>>
>>
>>> +        bo->mem.mem_type = TTM_PL_SYSTEM;
>>> +        bo->ttm = NULL;
>>> +        return 0;
>>> +    }
>>>         evict_mem = bo->mem;
>>>       evict_mem.mm_node = NULL;


More information about the dri-devel mailing list