[PATCH] drm/amdgpu: release parent fence reference

[Christian König]

Am 28.10.2016 um 13:47 schrieb Christian König:
> Am 24.10.2016 um 11:43 schrieb Grazvydas Ignotas:
>> On Mon, Oct 24, 2016 at 12:13 PM, Christian König
>> <deathsimple at vodafone.de> wrote:
>>> Am 24.10.2016 um 04:25 schrieb zhoucm1:
>>>>
>>>>
>>>> On 2016年10月24日 02:31, Grazvydas Ignotas wrote:
>>>>> It's done in amd_sched_hw_job_reset(), but not in normal job 
>>>>> processing.
>>>>> Leak reported by CONFIG_SLUB_DEBUG.
>>>>>
>>>>> Signed-off-by: Grazvydas Ignotas <notasas at gmail.com>
>>>>> ---
>>>>>    CONFIG_SLUB_DEBUG reports more leaks related to ioctls,
>>>>>    but I was unable to track them down...
>>>>>
>>>>>    drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 2 ++
>>>>>    1 file changed, 2 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>> b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>> index 910b8d5..cfb686e 100644
>>>>> --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>> +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>> @@ -522,6 +522,8 @@ static void amd_sched_process_job(struct fence 
>>>>> *f,
>>>>> struct fence_cb *cb)
>>>>>          trace_amd_sched_process_job(s_fence);
>>>>>        fence_put(&s_fence->finished);
>>>>> +    fence_put(s_fence->parent);
>>>> If I remember correctly, parent was put in sched fence release.
>>>
>>> Yes, exactly. It's only released in amd_sched_hw_job_reset() because 
>>> we want
>>> to assign a new parent fence.
>> Well then it doesn't work, SLUB_DEBUG detects leaks on teardown, and
>> with my patch they're gone.
>> Perhaps the problem is that when new parent fence is assigned, the old
>> one is not put? I won't be able to check until the weekend, so if
>> anybody else can do it, please go ahead.
>
> Mhm, what steps do you do to reproduce this?

Never mind, I was able to figure out how to trigger it. You actually 
have to create a fence to leak it :)

Regards,
Christian.

>
> I'm looking into this a bit and it sounds like we just doesn't 
> correctly tear down the scheduler on driver unload.
>
> Regards,
> Christian.
>
>>
>> Gražvydas
>
>