[PATCH 1/4] drm/amdgpu: add sched sync for amdgpu job

[Christian König]

Am 10.05.2017 um 11:00 schrieb zhoucm1:
>
>
> On 2017年05月10日 16:50, Christian König wrote:
>> Am 10.05.2017 um 10:38 schrieb zhoucm1:
>>>
>>>
>>> On 2017年05月10日 16:26, Christian König wrote:
>>>> Am 10.05.2017 um 09:31 schrieb Chunming Zhou:
>>>>> this is an improvement for previous patch, the sched_sync is to 
>>>>> store fence
>>>>> that could be skipped as scheduled, when job is executed, we 
>>>>> didn't need
>>>>> pipeline_sync if all fences in sched_sync are signalled, otherwise 
>>>>> insert
>>>>> pipeline_sync still.
>>>>>
>>>>> Change-Id: I26d3a2794272ba94b25753d4bf367326d12f6939
>>>>> Signed-off-by: Chunming Zhou <David1.Zhou at amd.com>
>>>>> ---
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu.h     | 1 +
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c  | 7 ++++++-
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++++-
>>>>>   3 files changed, 11 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h 
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> index 787acd7..ef018bf 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> @@ -1162,6 +1162,7 @@ struct amdgpu_job {
>>>>>       struct amdgpu_vm    *vm;
>>>>>       struct amdgpu_ring    *ring;
>>>>>       struct amdgpu_sync    sync;
>>>>> +    struct amdgpu_sync    sched_sync;
>>>>>       struct amdgpu_ib    *ibs;
>>>>>       struct fence        *fence; /* the hw fence */
>>>>>       uint32_t        preamble_status;
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c 
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>>>> index 2c6624d..86ad507 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>>>> @@ -121,6 +121,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring 
>>>>> *ring, unsigned num_ibs,
>>>>>   {
>>>>>       struct amdgpu_device *adev = ring->adev;
>>>>>       struct amdgpu_ib *ib = &ibs[0];
>>>>> +    struct fence *tmp;
>>>>>       bool skip_preamble, need_ctx_switch;
>>>>>       unsigned patch_offset = ~0;
>>>>>       struct amdgpu_vm *vm;
>>>>> @@ -167,8 +168,12 @@ int amdgpu_ib_schedule(struct amdgpu_ring 
>>>>> *ring, unsigned num_ibs,
>>>>>           return r;
>>>>>       }
>>>>>   -    if (ring->funcs->emit_pipeline_sync && job && 
>>>>> job->need_pipeline_sync)
>>>>> +    if (ring->funcs->emit_pipeline_sync && job &&
>>>>> +        (tmp = amdgpu_sync_get_fence(&job->sched_sync))) {
>>>>> +        job->need_pipeline_sync = true;
>>>>>           amdgpu_ring_emit_pipeline_sync(ring);
>>>>> +        fence_put(tmp);
>>>>> +    }
>>>>>       if (vm) {
>>>>>           amdgpu_ring_insert_nop(ring, extra_nop); /* prevent CE 
>>>>> go too fast than DE */
>>>>>   diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>>>> index cfa97ab..fa0c8b1 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>>>> @@ -60,6 +60,7 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, 
>>>>> unsigned num_ibs,
>>>>>       (*job)->need_pipeline_sync = false;
>>>>>         amdgpu_sync_create(&(*job)->sync);
>>>>> +    amdgpu_sync_create(&(*job)->sched_sync);
>>>>>         return 0;
>>>>>   }
>>>>> @@ -98,6 +99,7 @@ static void amdgpu_job_free_cb(struct 
>>>>> amd_sched_job *s_job)
>>>>>         fence_put(job->fence);
>>>>>       amdgpu_sync_free(&job->sync);
>>>>> +    amdgpu_sync_free(&job->sched_sync);
>>>>>       kfree(job);
>>>>>   }
>>>>>   @@ -107,6 +109,7 @@ void amdgpu_job_free(struct amdgpu_job *job)
>>>>>         fence_put(job->fence);
>>>>>       amdgpu_sync_free(&job->sync);
>>>>> +    amdgpu_sync_free(&job->sched_sync);
>>>>>       kfree(job);
>>>>>   }
>>>>>   @@ -154,7 +157,7 @@ static struct fence 
>>>>> *amdgpu_job_dependency(struct amd_sched_job *sched_job)
>>>>>       }
>>>>>         if (amd_sched_dependency_optimized(fence, 
>>>>> sched_job->s_entity))
>>>>> -        job->need_pipeline_sync = true;
>>>>> +        amdgpu_sync_fence(job->adev, &job->sched_sync, fence);
>>>>
>>>> This can result in an -ENOMEM 
>>> will handle it.
>>>> and additional to that we only need to remember the last fence 
>>>> optimized like this, not all of them.
>>>>
>>>> So just keep the last one found here in job->sched_fence instead.
>>> I guess this isn't enough.
>>> The dependency is not in order when calling, so the last one is not 
>>> always the last scheduled fence.
>>> And they could be sched fence not hw fence, although they are 
>>> handled by same hw ring, but the sched fence context isn't same.
>>> so we still need sched_sync here, right?
>>
>> No, amdgpu_job_dependency is only called again when the returned 
>> fence is signaled (or scheduled on the same ring).
> Let use give an example for it:
> Assume job->sync has two fences(fenceA and fenceB) which could be 
> scheduled. fenceA is from entity1, fenceB is from entity2, but both 
> for gfx engine, but fenceA could be submitted to hw ring behind fenceB.
> the order in job->sync list is: others---->fenceA---->fenceB--->others.
> when calling amdgpu_job_dependency, fenceA will be checked first, and 
> then fenceB.
>
> If following your proposal, we only store fenceB, but fenceA is the 
> later. Which isn't  expected.

Ah! Indeed, I didn't realized that the dependent fence could have 
already been scheduled.

Mhm, how are we going to handle the out of memory situation then? Sine 
we are inside a kernel thread we are not supposed to fail at this point.

Regards,
Christian.

>
>
> Regards,
> David Zhou
>>
>> So when this is called and you find that you need to wait for another 
>> fence the order is guaranteed.
>>
>> Regards,
>> Christian.
>>
>>>
>>> Regards,
>>> David zhou
>>>>
>>>> Regards,
>>>> Christian.
>>>>
>>>>>         return fence;
>>>>>   }
>>>>
>>>>
>>>
>>
>