All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
@ 2021-07-23  4:41 Jingwen Chen
  2021-07-23  4:41 ` [PATCH] drm: add tdr support for embeded hw_fence Jingwen Chen
  2021-07-23  8:29 ` [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Liu, Monk
  0 siblings, 2 replies; 3+ messages in thread
From: Jingwen Chen @ 2021-07-23  4:41 UTC (permalink / raw)
  To: amd-gfx
  Cc: andrey.grodzovsky, Jack Zhang, Jingwen Chen, monk.liu,
	ckoenig.leichtzumerken

From: Jack Zhang <Jack.Zhang1@amd.com>

Why: Previously hw fence is alloced separately with job.
It caused historical lifetime issues and corner cases.
The ideal situation is to take fence to manage both job
and fence's lifetime, and simplify the design of gpu-scheduler.

How:
We propose to embed hw_fence into amdgpu_job.
1. We cover the normal job submission by this method.
2. For ib_test, and submit without a parent job keep the
legacy way to create a hw fence separately.

Signed-off-by: Jingwen Chen <Jingwen.Chen2@amd.com>
Signed-off-by: Jack Zhang <Jack.Zhang1@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c  |  1 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c   | 62 ++++++++++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c      |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c     | 35 ++++++++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.h     |  4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h    |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c      |  2 +-
 8 files changed, 80 insertions(+), 30 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index b6d33f13b476..bad403978bac 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -714,7 +714,6 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
 	ret = dma_fence_wait(f, false);
 
 err_ib_sched:
-	dma_fence_put(f);
 	amdgpu_job_free(job);
 err:
 	return ret;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
index 536005bff24a..277128846dd1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
@@ -1414,7 +1414,7 @@ static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
 			continue;
 		}
 		job = to_amdgpu_job(s_job);
-		if (preempted && job->fence == fence)
+		if (preempted && (&job->hw_fence) == fence)
 			/* mark the job as preempted */
 			job->preemption_status |= AMDGPU_IB_PREEMPTED;
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 30772608eac6..eecf21d8ec33 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -133,25 +133,40 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
  * Emits a fence command on the requested ring (all asics).
  * Returns 0 on success, -ENOMEM on failure.
  */
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amdgpu_job *job,
 		      unsigned flags)
 {
 	struct amdgpu_device *adev = ring->adev;
-	struct amdgpu_fence *fence;
+	struct dma_fence *fence;
+	struct amdgpu_fence *am_fence;
 	struct dma_fence __rcu **ptr;
 	uint32_t seq;
 	int r;
 
-	fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
-	if (fence == NULL)
-		return -ENOMEM;
+	if (job == NULL) {
+		/* create a sperate hw fence */
+		am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
+		if (am_fence == NULL)
+			return -ENOMEM;
+		fence = &am_fence->base;
+		am_fence->ring = ring;
+	} else {
+		/* take use of job-embedded fence */
+		fence = &job->hw_fence;
+		job->ring = ring;
+	}
 
 	seq = ++ring->fence_drv.sync_seq;
-	fence->ring = ring;
-	dma_fence_init(&fence->base, &amdgpu_fence_ops,
+	dma_fence_init(fence, &amdgpu_fence_ops,
 		       &ring->fence_drv.lock,
 		       adev->fence_context + ring->idx,
 		       seq);
+
+	if (job != NULL) {
+		/* mark this fence has a parent job */
+		set_bit(DMA_FENCE_FLAG_USER_BITS, &fence->flags);
+	}
+
 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
 			       seq, flags | AMDGPU_FENCE_FLAG_INT);
 	pm_runtime_get_noresume(adev_to_drm(adev)->dev);
@@ -174,9 +189,9 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
 	/* This function can't be called concurrently anyway, otherwise
 	 * emitting the fence would mess up the hardware ring buffer.
 	 */
-	rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
+	rcu_assign_pointer(*ptr, dma_fence_get(fence));
 
-	*f = &fence->base;
+	*f = fence;
 
 	return 0;
 }
@@ -636,8 +651,16 @@ static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
 
 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
 {
-	struct amdgpu_fence *fence = to_amdgpu_fence(f);
-	return (const char *)fence->ring->name;
+	struct amdgpu_ring *ring;
+
+	if (test_bit(DMA_FENCE_FLAG_USER_BITS, &f->flags)) {
+		struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
+
+		ring = job->ring;
+	} else {
+		ring = to_amdgpu_fence(f)->ring;
+	}
+	return (const char *)ring->name;
 }
 
 /**
@@ -671,8 +694,20 @@ static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
 static void amdgpu_fence_free(struct rcu_head *rcu)
 {
 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
-	struct amdgpu_fence *fence = to_amdgpu_fence(f);
-	kmem_cache_free(amdgpu_fence_slab, fence);
+
+	if (test_bit(DMA_FENCE_FLAG_USER_BITS, &f->flags)) {
+	/* free job if fence has a parent job */
+		struct amdgpu_job *job;
+
+		job = container_of(f, struct amdgpu_job, hw_fence);
+		kfree(job);
+	} else {
+	/* free fence_slab if it's separated fence*/
+		struct amdgpu_fence *fence;
+
+		fence = to_amdgpu_fence(f);
+		kmem_cache_free(amdgpu_fence_slab, fence);
+	}
 }
 
 /**
@@ -695,6 +730,7 @@ static const struct dma_fence_ops amdgpu_fence_ops = {
 	.release = amdgpu_fence_release,
 };
 
+
 /*
  * Fence debugfs
  */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
index ec65ab0ddf89..c076a6b9a5a2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -262,7 +262,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
 	}
 
-	r = amdgpu_fence_emit(ring, f, fence_flags);
+	r = amdgpu_fence_emit(ring, f, job, fence_flags);
 	if (r) {
 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
 		if (job && job->vmid)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..7c426e225b24 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -114,11 +114,16 @@ void amdgpu_job_free_resources(struct amdgpu_job *job)
 {
 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
 	struct dma_fence *f;
+	struct dma_fence *hw_fence;
 	unsigned i;
 
-	/* use sched fence if available */
-	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
+	if (job->hw_fence.ops == NULL)
+		hw_fence = job->external_hw_fence;
+	else
+		hw_fence = &job->hw_fence;
 
+	/* use sched fence if available */
+	f = job->base.s_fence ? &job->base.s_fence->finished : hw_fence;
 	for (i = 0; i < job->num_ibs; ++i)
 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);
 }
@@ -129,20 +134,27 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
 
 	drm_sched_job_cleanup(s_job);
 
-	dma_fence_put(job->fence);
 	amdgpu_sync_free(&job->sync);
 	amdgpu_sync_free(&job->sched_sync);
-	kfree(job);
+
+    /* only put the hw fence if has embedded fence */
+	if (job->hw_fence.ops != NULL)
+		dma_fence_put(&job->hw_fence);
+	else
+		kfree(job);
 }
 
 void amdgpu_job_free(struct amdgpu_job *job)
 {
 	amdgpu_job_free_resources(job);
-
-	dma_fence_put(job->fence);
 	amdgpu_sync_free(&job->sync);
 	amdgpu_sync_free(&job->sched_sync);
-	kfree(job);
+
+	/* only put the hw fence if has embedded fence */
+	if (job->hw_fence.ops != NULL)
+		dma_fence_put(&job->hw_fence);
+	else
+		kfree(job);
 }
 
 int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
@@ -171,11 +183,14 @@ int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
 
 	job->base.sched = &ring->sched;
 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, NULL, fence);
-	job->fence = dma_fence_get(*fence);
+	/* record external_hw_fence for direct submit */
+	job->external_hw_fence = dma_fence_get(*fence);
 	if (r)
 		return r;
 
 	amdgpu_job_free(job);
+	dma_fence_put(*fence);
+
 	return 0;
 }
 
@@ -233,10 +248,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 		if (r)
 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
 	}
-	/* if gpu reset, hw fence will be replaced here */
-	dma_fence_put(job->fence);
-	job->fence = dma_fence_get(fence);
 
+	dma_fence_get(fence);
 	amdgpu_job_free_resources(job);
 
 	fence = r ? ERR_PTR(r) : fence;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
index 81caac9b958a..92324c978534 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
@@ -46,7 +46,9 @@ struct amdgpu_job {
 	struct amdgpu_sync	sync;
 	struct amdgpu_sync	sched_sync;
 	struct amdgpu_ib	*ibs;
-	struct dma_fence	*fence; /* the hw fence */
+	struct dma_fence	hw_fence;
+	struct amdgpu_ring *ring;
+	struct dma_fence	*external_hw_fence;
 	uint32_t		preamble_status;
 	uint32_t                preemption_status;
 	uint32_t		num_ibs;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
index ca1622835296..2306424cbcb4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
@@ -118,7 +118,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
 				   unsigned irq_type);
 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev);
 void amdgpu_fence_driver_resume(struct amdgpu_device *adev);
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence,
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence, struct amdgpu_job *job,
 		      unsigned flags);
 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
 			      uint32_t timeout);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 358316d6a38c..0670f43ef22a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1216,7 +1216,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
 
 	if (vm_flush_needed || pasid_mapping_needed) {
-		r = amdgpu_fence_emit(ring, &fence, 0);
+		r = amdgpu_fence_emit(ring, &fence, NULL, 0);
 		if (r)
 			return r;
 	}
-- 
2.25.1

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

^ permalink raw reply related	[flat|nested] 3+ messages in thread
* [PATCH] drm: add tdr support for embeded hw_fence
  2021-07-23  4:41 [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Jingwen Chen
@ 2021-07-23  4:41 ` Jingwen Chen
  2021-07-23  8:29 ` [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Liu, Monk
  1 sibling, 0 replies; 3+ messages in thread
From: Jingwen Chen @ 2021-07-23  4:41 UTC (permalink / raw)
  To: amd-gfx
  Cc: andrey.grodzovsky, Jingwen Chen, Jack Zhang, monk.liu,
	ckoenig.leichtzumerken

[Why]
After embeded hw_fence to amdgpu_job, we need to add tdr support
for this feature.

[How]
1. Add a resubmit_flag for resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences in
   pre_asic_reset
3. skip dma_fence_get for resubmit jobs and add a dma_fence_put
   for guilty jobs.

Signed-off-by: Jack Zhang <Jack.Zhang1@amd.com>
Signed-off-by: Jingwen Chen <Jingwen.Chen2@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c  | 13 +++++++++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c    |  4 +++-
 drivers/gpu/drm/scheduler/sched_main.c     |  1 +
 include/drm/gpu_scheduler.h                |  1 +
 5 files changed, 25 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 280b1940e892..df73fe666e87 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
 				 struct amdgpu_reset_context *reset_context)
 {
-	int i, r = 0;
+	int i, j, r = 0;
 	struct amdgpu_job *job = NULL;
 	bool need_full_reset =
 		test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
 		if (!ring || !ring->sched.thread)
 			continue;
 
+		/*clear job fence from fence drv to avoid force_completion
+		 *leave NULL and vm flush fence in fence drv */
+		for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) {
+			struct dma_fence *old,**ptr;
+			ptr = &ring->fence_drv.fences[j];
+			old = rcu_dereference_protected(*ptr, 1);
+			if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) {
+				RCU_INIT_POINTER(*ptr, NULL);
+			}
+		}
 		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
 		amdgpu_fence_driver_force_completion(ring);
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index eecf21d8ec33..d5b3d5f8f951 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -157,10 +157,15 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
 	}
 
 	seq = ++ring->fence_drv.sync_seq;
-	dma_fence_init(fence, &amdgpu_fence_ops,
-		       &ring->fence_drv.lock,
-		       adev->fence_context + ring->idx,
-		       seq);
+	if (job != NULL && job->base.resubmit_flag == 1) {
+		/* reset seq for resubmitted jobs */
+		fence->seqno = seq;
+	} else {
+		dma_fence_init(fence, &amdgpu_fence_ops,
+				&ring->fence_drv.lock,
+				adev->fence_context + ring->idx,
+				seq);
+	}
 
 	if (job != NULL) {
 		/* mark this fence has a parent job */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 7c426e225b24..d6f848adc3f4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
 
 	if (finished->error < 0) {
+		dma_fence_put(&job->hw_fence);
 		DRM_INFO("Skip scheduling IBs!\n");
 	} else {
 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
 	}
 
-	dma_fence_get(fence);
+	if (!job->base.resubmit_flag)
+		dma_fence_get(fence);
 	amdgpu_job_free_resources(job);
 
 	fence = r ? ERR_PTR(r) : fence;
diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index f4f474944169..5a36ab5aea2d 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
 
 		dma_fence_put(s_job->s_fence->parent);
+		s_job->resubmit_flag = 1;
 		fence = sched->ops->run_job(s_job);
 		i++;
 
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 4ea8606d91fe..06c101af1f71 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -198,6 +198,7 @@ struct drm_sched_job {
 	enum drm_sched_priority		s_priority;
 	struct drm_sched_entity         *entity;
 	struct dma_fence_cb		cb;
+	int resubmit_flag;
 };
 
 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
-- 
2.25.1

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

^ permalink raw reply related	[flat|nested] 3+ messages in thread
* RE: [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
  2021-07-23  4:41 [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Jingwen Chen
  2021-07-23  4:41 ` [PATCH] drm: add tdr support for embeded hw_fence Jingwen Chen
@ 2021-07-23  8:29 ` Liu, Monk
  1 sibling, 0 replies; 3+ messages in thread
From: Liu, Monk @ 2021-07-23  8:29 UTC (permalink / raw)
  To: Chen, JingWen, amd-gfx
  Cc: Grodzovsky, Andrey, Zhang, Jack (Jian),
	Chen, JingWen, ckoenig.leichtzumerken

[AMD Official Use Only]

Hi jingwen

In the lines like  below:
+	if (job != NULL) {
+		/* mark this fence has a parent job */
+		set_bit(DMA_FENCE_FLAG_USER_BITS, &fence->flags);
+	}

The DMA_FENCE_FLAG_USER_BITS usage looks not good enough

Please try to improve it, e.g.:

#define AMDGPU_DMA_FENCE_FLAG_INSIDE_JOB  DMA_FENCE_FLAG_USER_BITS

This way in the future USER BITS could be extended to other purpose:
#define AMDGPU_DMA_FENCE_FLAG_XXXXX(AMDGPU_DMA_FENCE_FLAG_INSIDE_JOB  + 1)

Please try to use AMDGPU_DMA_FENCE_FLAG_INSIDE_JOB instead of DMA_FENCE_FLAG_USER_BITS everywhere 

For other logic parts I do not have suggestions since I already know those history and the corresponding changes much enough 

Thanks 

------------------------------------------
Monk Liu | Cloud-GPU Core team
------------------------------------------

-----Original Message-----
From: Jingwen Chen <Jingwen.Chen2@amd.com> 
Sent: Friday, July 23, 2021 12:42 PM
To: amd-gfx@lists.freedesktop.org
Cc: Liu, Monk <Monk.Liu@amd.com>; Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>; ckoenig.leichtzumerken@gmail.com; Zhang, Jack (Jian) <Jack.Zhang1@amd.com>; Chen, JingWen <JingWen.Chen2@amd.com>
Subject: [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job

From: Jack Zhang <Jack.Zhang1@amd.com>

Why: Previously hw fence is alloced separately with job.
It caused historical lifetime issues and corner cases.
The ideal situation is to take fence to manage both job and fence's lifetime, and simplify the design of gpu-scheduler.

How:
We propose to embed hw_fence into amdgpu_job.
1. We cover the normal job submission by this method.
2. For ib_test, and submit without a parent job keep the legacy way to create a hw fence separately.

Signed-off-by: Jingwen Chen <Jingwen.Chen2@amd.com>
Signed-off-by: Jack Zhang <Jack.Zhang1@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c  |  1 -  drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c   | 62 ++++++++++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c      |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c     | 35 ++++++++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.h     |  4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h    |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c      |  2 +-
 8 files changed, 80 insertions(+), 30 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index b6d33f13b476..bad403978bac 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -714,7 +714,6 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
 	ret = dma_fence_wait(f, false);
 
 err_ib_sched:
-	dma_fence_put(f);
 	amdgpu_job_free(job);
 err:
 	return ret;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
index 536005bff24a..277128846dd1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
@@ -1414,7 +1414,7 @@ static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
 			continue;
 		}
 		job = to_amdgpu_job(s_job);
-		if (preempted && job->fence == fence)
+		if (preempted && (&job->hw_fence) == fence)
 			/* mark the job as preempted */
 			job->preemption_status |= AMDGPU_IB_PREEMPTED;
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 30772608eac6..eecf21d8ec33 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -133,25 +133,40 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
  * Emits a fence command on the requested ring (all asics).
  * Returns 0 on success, -ENOMEM on failure.
  */
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, 
+struct amdgpu_job *job,
 		      unsigned flags)
 {
 	struct amdgpu_device *adev = ring->adev;
-	struct amdgpu_fence *fence;
+	struct dma_fence *fence;
+	struct amdgpu_fence *am_fence;
 	struct dma_fence __rcu **ptr;
 	uint32_t seq;
 	int r;
 
-	fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
-	if (fence == NULL)
-		return -ENOMEM;
+	if (job == NULL) {
+		/* create a sperate hw fence */
+		am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
+		if (am_fence == NULL)
+			return -ENOMEM;
+		fence = &am_fence->base;
+		am_fence->ring = ring;
+	} else {
+		/* take use of job-embedded fence */
+		fence = &job->hw_fence;
+		job->ring = ring;
+	}
 
 	seq = ++ring->fence_drv.sync_seq;
-	fence->ring = ring;
-	dma_fence_init(&fence->base, &amdgpu_fence_ops,
+	dma_fence_init(fence, &amdgpu_fence_ops,
 		       &ring->fence_drv.lock,
 		       adev->fence_context + ring->idx,
 		       seq);
+
+	if (job != NULL) {
+		/* mark this fence has a parent job */
+		set_bit(DMA_FENCE_FLAG_USER_BITS, &fence->flags);
+	}
+
 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
 			       seq, flags | AMDGPU_FENCE_FLAG_INT);
 	pm_runtime_get_noresume(adev_to_drm(adev)->dev);
@@ -174,9 +189,9 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
 	/* This function can't be called concurrently anyway, otherwise
 	 * emitting the fence would mess up the hardware ring buffer.
 	 */
-	rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
+	rcu_assign_pointer(*ptr, dma_fence_get(fence));
 
-	*f = &fence->base;
+	*f = fence;
 
 	return 0;
 }
@@ -636,8 +651,16 @@ static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
 
 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)  {
-	struct amdgpu_fence *fence = to_amdgpu_fence(f);
-	return (const char *)fence->ring->name;
+	struct amdgpu_ring *ring;
+
+	if (test_bit(DMA_FENCE_FLAG_USER_BITS, &f->flags)) {
+		struct amdgpu_job *job = container_of(f, struct amdgpu_job, 
+hw_fence);
+
+		ring = job->ring;
+	} else {
+		ring = to_amdgpu_fence(f)->ring;
+	}
+	return (const char *)ring->name;
 }
 
 /**
@@ -671,8 +694,20 @@ static bool amdgpu_fence_enable_signaling(struct dma_fence *f)  static void amdgpu_fence_free(struct rcu_head *rcu)  {
 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
-	struct amdgpu_fence *fence = to_amdgpu_fence(f);
-	kmem_cache_free(amdgpu_fence_slab, fence);
+
+	if (test_bit(DMA_FENCE_FLAG_USER_BITS, &f->flags)) {
+	/* free job if fence has a parent job */
+		struct amdgpu_job *job;
+
+		job = container_of(f, struct amdgpu_job, hw_fence);
+		kfree(job);
+	} else {
+	/* free fence_slab if it's separated fence*/
+		struct amdgpu_fence *fence;
+
+		fence = to_amdgpu_fence(f);
+		kmem_cache_free(amdgpu_fence_slab, fence);
+	}
 }
 
 /**
@@ -695,6 +730,7 @@ static const struct dma_fence_ops amdgpu_fence_ops = {
 	.release = amdgpu_fence_release,
 };
 
+
 /*
  * Fence debugfs
  */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
index ec65ab0ddf89..c076a6b9a5a2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -262,7 +262,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
 	}
 
-	r = amdgpu_fence_emit(ring, f, fence_flags);
+	r = amdgpu_fence_emit(ring, f, job, fence_flags);
 	if (r) {
 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
 		if (job && job->vmid)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 759b34799221..7c426e225b24 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -114,11 +114,16 @@ void amdgpu_job_free_resources(struct amdgpu_job *job)  {
 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
 	struct dma_fence *f;
+	struct dma_fence *hw_fence;
 	unsigned i;
 
-	/* use sched fence if available */
-	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
+	if (job->hw_fence.ops == NULL)
+		hw_fence = job->external_hw_fence;
+	else
+		hw_fence = &job->hw_fence;
 
+	/* use sched fence if available */
+	f = job->base.s_fence ? &job->base.s_fence->finished : hw_fence;
 	for (i = 0; i < job->num_ibs; ++i)
 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);  } @@ -129,20 +134,27 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
 
 	drm_sched_job_cleanup(s_job);
 
-	dma_fence_put(job->fence);
 	amdgpu_sync_free(&job->sync);
 	amdgpu_sync_free(&job->sched_sync);
-	kfree(job);
+
+    /* only put the hw fence if has embedded fence */
+	if (job->hw_fence.ops != NULL)
+		dma_fence_put(&job->hw_fence);
+	else
+		kfree(job);
 }
 
 void amdgpu_job_free(struct amdgpu_job *job)  {
 	amdgpu_job_free_resources(job);
-
-	dma_fence_put(job->fence);
 	amdgpu_sync_free(&job->sync);
 	amdgpu_sync_free(&job->sched_sync);
-	kfree(job);
+
+	/* only put the hw fence if has embedded fence */
+	if (job->hw_fence.ops != NULL)
+		dma_fence_put(&job->hw_fence);
+	else
+		kfree(job);
 }
 
 int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity, @@ -171,11 +183,14 @@ int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
 
 	job->base.sched = &ring->sched;
 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, NULL, fence);
-	job->fence = dma_fence_get(*fence);
+	/* record external_hw_fence for direct submit */
+	job->external_hw_fence = dma_fence_get(*fence);
 	if (r)
 		return r;
 
 	amdgpu_job_free(job);
+	dma_fence_put(*fence);
+
 	return 0;
 }
 
@@ -233,10 +248,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 		if (r)
 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
 	}
-	/* if gpu reset, hw fence will be replaced here */
-	dma_fence_put(job->fence);
-	job->fence = dma_fence_get(fence);
 
+	dma_fence_get(fence);
 	amdgpu_job_free_resources(job);
 
 	fence = r ? ERR_PTR(r) : fence;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
index 81caac9b958a..92324c978534 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
@@ -46,7 +46,9 @@ struct amdgpu_job {
 	struct amdgpu_sync	sync;
 	struct amdgpu_sync	sched_sync;
 	struct amdgpu_ib	*ibs;
-	struct dma_fence	*fence; /* the hw fence */
+	struct dma_fence	hw_fence;
+	struct amdgpu_ring *ring;
+	struct dma_fence	*external_hw_fence;
 	uint32_t		preamble_status;
 	uint32_t                preemption_status;
 	uint32_t		num_ibs;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
index ca1622835296..2306424cbcb4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
@@ -118,7 +118,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
 				   unsigned irq_type);
 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev);  void amdgpu_fence_driver_resume(struct amdgpu_device *adev); -int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence,
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence 
+**fence, struct amdgpu_job *job,
 		      unsigned flags);
 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
 			      uint32_t timeout);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 358316d6a38c..0670f43ef22a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1216,7 +1216,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
 
 	if (vm_flush_needed || pasid_mapping_needed) {
-		r = amdgpu_fence_emit(ring, &fence, 0);
+		r = amdgpu_fence_emit(ring, &fence, NULL, 0);
 		if (r)
 			return r;
 	}
--
2.25.1
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

^ permalink raw reply related	[flat|nested] 3+ messages in thread
end of thread, other threads:[~2021-07-23  8:29 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-07-23  4:41 [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Jingwen Chen
2021-07-23  4:41 ` [PATCH] drm: add tdr support for embeded hw_fence Jingwen Chen
2021-07-23  8:29 ` [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job Liu, Monk

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.