Hi Christian
For scenario: Bad Job (hang, vmid1) -->Job A (context 10, explicit dep for Job B, vmid2) --> Job B(context 10, vmid2) --> Job C (context 11, vmid3)
Assume "job_hang_limit" is 0, and assume "sched_hw_submission" is 4, I give a second thought on the logic after GPU reset:
1) the bad Job would be set guilty and skipped by scheduler,
2) the first re-submitted job (Job A) would be forced with a pipeline-sync,
3) the first re-submitted job (Job A) would be forced with a vm-flush, and later its VMID's "current_gpu_reset_count" is updated to "adev->gpu_reset_count"
4) the second re-submitted job (Job B, assume it was from the same context of Job A, share the same page table/process, and no vm_update needed ) would not be forced with a pipeline-sync, and neither a vm-flush ...
Thus for Job B if it has an explicit dep on Job A, this explicit dep would get lost and there will be no pipeline sync inserted prior to Job B ...
Do you think that's a possible corner case ?
/Monk
-----Original Message-----
From: Koenig, Christian
Sent: Monday, November 5, 2018 3:48 PM
To: Liu, Monk <Monk.Liu@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx; Zhou, David(ChunMing) <David1.Zhou@xxxxxxx>
Subject: Re: [PATCH 2/3] drm/amdgpu: drop the sched_sync
Am 05.11.18 um 08:24 schrieb Liu, Monk:
>> David Zhou had an use case which saw a >10% performance drop the last time he tried it.
> I really don't believe that, because if you insert a WAIT_MEM on an already signaled fence, it only cost GPU couple clocks to move on, right ? no reason to slow down up to 10% ... with 3dmark vulkan version test, the performance is barely different ... with my patch applied ...
Why do you think that we insert a WAIT_MEM on an already signaled fence?
The pipeline sync always wait for the last fence value (because we can't handle wraparounds in PM4).
So you have a pipeline sync when you don't need one and that is really really bad for things shared between processes, e.g. X/Wayland and it's clients.
I also expects that this doesn't effect 3dmark at all, but everything which runs in a window which is composed by X could be slowed down massively.
David do you remember which use case was affected when you tried to drop this optimization?
>> When a reset happens we flush the VMIDs when re-submitting the jobs to the rings and while doing so we also always do a pipeline sync.
> I will check that point in my branch, I didn't use drm-next, maybe
> there is gap in this part
We had that logic for a very long time now, but we recently simplified it. Could be that there was a bug introduced doing so.
Maybe we should add a specific flag to run_job() to note that we are re-running a job and then always add VM flushes/pipeline syncs?
But my main question is why do you see any issues with quark? That is a workaround for an issue for Vulkan sync handling and should only surface when a specific test is run many many times.
Regards,
Christian.
>
> /Monk
> -----Original Message-----
> From: Koenig, Christian
> Sent: Monday, November 5, 2018 3:02 AM
> To: Liu, Monk <Monk.Liu@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH 2/3] drm/amdgpu: drop the sched_sync
>
>> Can you tell me which game/benchmark will have performance drop with this fix by your understanding ?
> When you sync between submission things like composing X windows are slowed down massively.
>
> David Zhou had an use case which saw a >10% performance drop the last time he tried it.
>
>> The problem I hit is during the massive stress test against
>> multi-process + quark , if the quark process hang the engine while there is another two job following the bad job, After the TDR these two job will lose the explicit and the pipeline-sync was also lost.
> Well that is really strange. This workaround is only for a very specific Vulkan CTS test which we are still not 100% sure is actually valid.
>
> When a reset happens we flush the VMIDs when re-submitting the jobs to the rings and while doing so we also always do a pipeline sync.
>
> So you should never ever run into any issues in quark with that, even when we completely disable this workaround.
>
> Regards,
> Christian.
>
> Am 04.11.18 um 01:48 schrieb Liu, Monk:
>>> NAK, that would result in a severe performance drop.
>>> We need the fence here to determine if we actually need to do the pipeline sync or not.
>>> E.g. the explicit requested fence could already be signaled.
>> For the performance issue, only insert a WAIT_REG_MEM on GFX/compute ring *doesn't* give the "severe" drop (it's mimic in fact) ... At least I didn't observe any performance drop with 3dmark benchmark (also tested vulkan CTS), Can you tell me which game/benchmark will have performance drop with this fix by your understanding ? let me check it .
>>
>> The problem I hit is during the massive stress test against
>> multi-process + quark , if the quark process hang the engine while there is another two job following the bad job, After the TDR these two job will lose the explicit and the pipeline-sync was also lost.
>>
>>
>> BTW: for original logic, the pipeline sync have another corner case:
>> Assume JobC depend on JobA with explicit flag, and there is jobB inserted in ring:
>>
>> jobA -> jobB -> (pipe sync)JobC
>>
>> if JobA really cost a lot of time to finish, in the
>> amdgpu_ib_schedule() stage you will insert a pipeline sync for JobC against its explicit dependency which is JobA, but there is a JobB between A and C and the pipeline sync of before JobC will wrongly wait on the JobB ...
>>
>> while it is not a big issue but obviously not necessary: C have no
>> relation with B
>>
>> /Monk
>>
>>
>>
>> -----Original Message-----
>> From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
>> Sent: Sunday, November 4, 2018 3:50 AM
>> To: Liu, Monk <Monk.Liu@xxxxxxx>; amd-gfx@xxxxxxxxxxxxxxxxxxxxx
>> Subject: Re: [PATCH 2/3] drm/amdgpu: drop the sched_sync
>>
>> Am 03.11.18 um 06:33 schrieb Monk Liu:
>>> Reasons to drop it:
>>>
>>> 1) simplify the code: just introduce field member "need_pipe_sync"
>>> for job is good enough to tell if the explicit dependency fence need
>>> followed by a pipeline sync.
>>>
>>> 2) after GPU_recover the explicit fence from sched_syn will not come
>>> back so the required pipeline_sync following it is missed, consider
>>> scenario below:
>>>> now on ring buffer:
>>> Job-A -> pipe_sync -> Job-B
>>>> TDR occured on Job-A, and after GPU recover:
>>>> now on ring buffer:
>>> Job-A -> Job-B
>>>
>>> because the fence from sched_sync is used and freed after
>>> ib_schedule in first time, it will never come back, with this patch
>>> this issue could be avoided.
>> NAK, that would result in a severe performance drop.
>>
>> We need the fence here to determine if we actually need to do the pipeline sync or not.
>>
>> E.g. the explicit requested fence could already be signaled.
>>
>> Christian.
>>
>>> Signed-off-by: Monk Liu <Monk.Liu@xxxxxxx>
>>> ---
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 16 ++++++----------
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 14 +++-----------
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 3 +--
>>> 3 files changed, 10 insertions(+), 23 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> index c48207b3..ac7d2da 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> @@ -122,7 +122,6 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>>> {
>>> struct amdgpu_device *adev = ring->adev;
>>> struct amdgpu_ib *ib = &ibs[0];
>>> - struct dma_fence *tmp = NULL;
>>> bool skip_preamble, need_ctx_switch;
>>> unsigned patch_offset = ~0;
>>> struct amdgpu_vm *vm;
>>> @@ -166,16 +165,13 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>>> }
>>>
>>> need_ctx_switch = ring->current_ctx != fence_ctx;
>>> - if (ring->funcs->emit_pipeline_sync && job &&
>>> - ((tmp = amdgpu_sync_get_fence(&job->sched_sync, NULL)) ||
>>> - (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
>>> - amdgpu_vm_need_pipeline_sync(ring, job))) {
>>> - need_pipe_sync = true;
>>>
>>> - if (tmp)
>>> - trace_amdgpu_ib_pipe_sync(job, tmp);
>>> -
>>> - dma_fence_put(tmp);
>>> + if (ring->funcs->emit_pipeline_sync && job) {
>>> + if ((need_ctx_switch && amdgpu_sriov_vf(adev)) ||
>>> + amdgpu_vm_need_pipeline_sync(ring, job))
>>> + need_pipe_sync = true;
>>> + else if (job->need_pipe_sync)
>>> + need_pipe_sync = true;
>>> }
>>>
>>> if (ring->funcs->insert_start)
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>> index 1d71f8c..dae997d 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>>> @@ -71,7 +71,6 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
>>> (*job)->num_ibs = num_ibs;
>>>
>>> amdgpu_sync_create(&(*job)->sync);
>>> - amdgpu_sync_create(&(*job)->sched_sync);
>>> (*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
>>> (*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
>>>
>>> @@ -117,7 +116,6 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
>>> amdgpu_ring_priority_put(ring, s_job->s_priority);
>>> dma_fence_put(job->fence);
>>> amdgpu_sync_free(&job->sync);
>>> - amdgpu_sync_free(&job->sched_sync);
>>> kfree(job);
>>> }
>>>
>>> @@ -127,7 +125,6 @@ void amdgpu_job_free(struct amdgpu_job *job)
>>>
>>> dma_fence_put(job->fence);
>>> amdgpu_sync_free(&job->sync);
>>> - amdgpu_sync_free(&job->sched_sync);
>>> kfree(job);
>>> }
>>>
>>> @@ -182,14 +179,9 @@ static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
>>> bool need_pipe_sync = false;
>>> int r;
>>>
>>> - fence = amdgpu_sync_get_fence(&job->sync, &need_pipe_sync);
>>> - if (fence && need_pipe_sync) {
>>> - if (drm_sched_dependency_optimized(fence, s_entity)) {
>>> - r = amdgpu_sync_fence(ring->adev, &job->sched_sync,
>>> - fence, false);
>>> - if (r)
>>> - DRM_ERROR("Error adding fence (%d)\n", r);
>>> - }
>>> + if (fence && need_pipe_sync && drm_sched_dependency_optimized(fence, s_entity)) {
>>> + trace_amdgpu_ib_pipe_sync(job, fence);
>>> + job->need_pipe_sync = true;
>>> }
>>>
>>> while (fence == NULL && vm && !job->vmid) { diff --git
>>> a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>>> index e1b46a6..c1d00f0 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>>> @@ -41,7 +41,6 @@ struct amdgpu_job {
>>> struct drm_sched_job base;
>>> struct amdgpu_vm *vm;
>>> struct amdgpu_sync sync;
>>> - struct amdgpu_sync sched_sync;
>>> struct amdgpu_ib *ibs;
>>> struct dma_fence *fence; /* the hw fence */
>>> uint32_t preamble_status;
>>> @@ -59,7 +58,7 @@ struct amdgpu_job {
>>> /* user fence handling */
>>> uint64_t uf_addr;
>>> uint64_t uf_sequence;
>>> -
>>> + bool need_pipe_sync; /* require a pipeline sync for this job */
>>> };
>>>
>>> int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned
>>> num_ibs,
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
