Re: [PATCH 09/10] drm/amdgpu: add gang submit backend

["Christian König" <ckoenig.leichtzumerken@gmail.com>]

Am 04.03.22 um 18:10 schrieb Andrey Grodzovsky:
>
> On 2022-03-03 03:23, Christian König wrote:
>> Allows submitting jobs as gang which needs to run on multiple
>> engines at the same time.
>>
>> Basic idea is that we have a global gang submit fence representing 
>> when the
>> gang leader is finally pushed to run on the hardware last.
>>
>> Jobs submitted as gang are never re-submitted in case of a GPU reset 
>> since this
>> won't work and will just deadlock the hardware immediately again.
>>
>> Signed-off-by: Christian König <christian.koenig@amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  3 ++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 34 ++++++++++++++++++++++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c    | 28 ++++++++++++++++--
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.h    |  3 ++
>>   4 files changed, 66 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>> index 7f447ed7a67f..a664d43d7502 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>> @@ -852,6 +852,7 @@ struct amdgpu_device {
>>       u64                fence_context;
>>       unsigned            num_rings;
>>       struct amdgpu_ring        *rings[AMDGPU_MAX_RINGS];
>> +    struct dma_fence __rcu        *gang_submit;
>>       bool                ib_pool_ready;
>>       struct amdgpu_sa_manager    ib_pools[AMDGPU_IB_POOL_MAX];
>>       struct amdgpu_sched 
>> gpu_sched[AMDGPU_HW_IP_NUM][AMDGPU_RING_PRIO_MAX];
>> @@ -1233,6 +1234,8 @@ void amdgpu_device_invalidate_hdp(struct 
>> amdgpu_device *adev,
>>           struct amdgpu_ring *ring);
>>     void amdgpu_device_halt(struct amdgpu_device *adev);
>> +struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
>> +                        struct dma_fence *gang);
>>     /* atpx handler */
>>   #if defined(CONFIG_VGA_SWITCHEROO)
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index d78141e2c509..a116b8c08827 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -3512,6 +3512,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
>>       adev->gmc.gart_size = 512 * 1024 * 1024;
>>       adev->accel_working = false;
>>       adev->num_rings = 0;
>> +    RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
>>       adev->mman.buffer_funcs = NULL;
>>       adev->mman.buffer_funcs_ring = NULL;
>>       adev->vm_manager.vm_pte_funcs = NULL;
>> @@ -3989,6 +3990,7 @@ void amdgpu_device_fini_sw(struct amdgpu_device 
>> *adev)
>>       release_firmware(adev->firmware.gpu_info_fw);
>>       adev->firmware.gpu_info_fw = NULL;
>>       adev->accel_working = false;
>> + dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
>>         amdgpu_reset_fini(adev);
>>   @@ -5744,3 +5746,35 @@ void amdgpu_device_halt(struct amdgpu_device 
>> *adev)
>>       pci_disable_device(pdev);
>>       pci_wait_for_pending_transaction(pdev);
>>   }
>> +
>> +/**
>> + * amdgpu_device_switch_gang - switch to a new gang
>> + * @adev: amdgpu_device pointer
>> + * @gang: the gang to switch to
>> + *
>> + * Try to switch to a new gang or return a reference to the current 
>> gang if that
>> + * isn't possible.
>> + * Returns: Either NULL if we switched correctly or a reference to 
>> the existing
>> + * gang.
>> + */
>> +struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
>> +                        struct dma_fence *gang)
>> +{
>> +    struct dma_fence *old = NULL;
>> +
>> +    do {
>> +        dma_fence_put(old);
>> +        old = dma_fence_get_rcu_safe(&adev->gang_submit);
>> +
>> +        if (old == gang)
>> +            break;
>> +
>> +        if (!dma_fence_is_signaled(old))
>> +            return old;
>> +
>> +    } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
>> +             old, gang) != old);
>> +
>> +    dma_fence_put(old);
>> +    return NULL;
>> +}
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> index e07ceae36a5c..059e11c7898c 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
>> @@ -169,11 +169,29 @@ static void amdgpu_job_free_cb(struct 
>> drm_sched_job *s_job)
>>           kfree(job);
>>   }
>>   +void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
>> +                struct amdgpu_job *leader)
>> +{
>> +    struct dma_fence *fence = &leader->base.s_fence->scheduled;
>> +
>> +    WARN_ON(job->gang_submit);
>> +
>> +    /*
>> +     * Don't add a reference when we are the gang leader to avoid 
>> circle
>> +     * dependency.
>> +     */
>> +    if (job != leader)
>> +        dma_fence_get(fence);
>> +    job->gang_submit = fence;
>> +}
>> +
>>   void amdgpu_job_free(struct amdgpu_job *job)
>>   {
>>       amdgpu_job_free_resources(job);
>>       amdgpu_sync_free(&job->sync);
>>       amdgpu_sync_free(&job->sched_sync);
>> +    if (job->gang_submit != &job->base.s_fence->scheduled)
>> +        dma_fence_put(job->gang_submit);
>>         /* only put the hw fence if has embedded fence */
>>       if (job->hw_fence.ops != NULL)
>> @@ -247,12 +265,16 @@ static struct dma_fence 
>> *amdgpu_job_dependency(struct drm_sched_job *sched_job,
>>           fence = amdgpu_sync_get_fence(&job->sync);
>>       }
>>   +    if (!fence && !job->gang_submit)
>> +        fence = amdgpu_device_switch_gang(ring->adev, 
>> job->gang_submit);
>> +
>
>
> Why job->gang_submit should be NULL in the check above ? Don't you 
> want to switch to an actual new gang fence here ?
> Jobs that don't have gang_submit fence set are not gang jobs anyway 
> and we don't care for this dependency
> for them right ?

Well exactly that's the point. That a job is not part of a gang submit 
is signaled by setting the pointer to NULL.

If we don't check for NULL here we would just crash.

Christian.

>
> Andrey
>
>
>>       return fence;
>>   }
>>     static struct dma_fence *amdgpu_job_run(struct drm_sched_job 
>> *sched_job)
>>   {
>>       struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
>> +    struct amdgpu_device *adev = ring->adev;
>>       struct dma_fence *fence = NULL, *finished;
>>       struct amdgpu_job *job;
>>       int r = 0;
>> @@ -264,8 +286,10 @@ static struct dma_fence *amdgpu_job_run(struct 
>> drm_sched_job *sched_job)
>>         trace_amdgpu_sched_run_job(job);
>>   -    if (job->vram_lost_counter != 
>> atomic_read(&ring->adev->vram_lost_counter))
>> -        dma_fence_set_error(finished, -ECANCELED);/* skip IB as well 
>> if VRAM lost */
>> +    /* Skip job if VRAM is lost and never resubmit gangs */
>> +    if (job->vram_lost_counter != 
>> atomic_read(&adev->vram_lost_counter) ||
>> +        (job->job_run_counter && job->gang_submit))
>> +        dma_fence_set_error(finished, -ECANCELED);
>>         if (finished->error < 0) {
>>           DRM_INFO("Skip scheduling IBs!\n");
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>> index 0bab8fe0d419..615328130615 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
>> @@ -51,6 +51,7 @@ struct amdgpu_job {
>>       struct amdgpu_sync    sched_sync;
>>       struct dma_fence    hw_fence;
>>       struct dma_fence    *external_hw_fence;
>> +    struct dma_fence    *gang_submit;
>>       uint32_t        preamble_status;
>>       uint32_t                preemption_status;
>>       bool                    vm_needs_flush;
>> @@ -80,6 +81,8 @@ int amdgpu_job_alloc_with_ib(struct amdgpu_device 
>> *adev, unsigned size,
>>   void amdgpu_job_set_resources(struct amdgpu_job *job, struct 
>> amdgpu_bo *gds,
>>                     struct amdgpu_bo *gws, struct amdgpu_bo *oa);
>>   void amdgpu_job_free_resources(struct amdgpu_job *job);
>> +void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
>> +                struct amdgpu_job *leader);
>>   void amdgpu_job_free(struct amdgpu_job *job);
>>   int amdgpu_job_submit(struct amdgpu_job *job, struct 
>> drm_sched_entity *entity,
>>                 void *owner, struct dma_fence **f);