RE: [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
[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, >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_X(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
Sent: Friday, July 23, 2021 12:42 PM
To: amd-gfx@lists.freedesktop.org
Cc: Liu, Monk ; Grodzovsky, Andrey
; ckoenig.leichtzumer...@gmail.com; Zhang, Jack
(Jian) ; Chen, JingWen
Subject: [PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
From: Jack Zhang
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
Signed-off-by: Jack Zhang
---
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 && (>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 = _fence->base;
+ am_fence->ring = ring;
+ } else {
+ /* take use of job-embedded fence */
+ fence = >hw_fence;
+ job->ring = ring;
+ }
seq = ++ring->fence_drv.sync_seq;
- fence->ring = ring;
- dma_fence_init(>base,
[PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
From: Jack Zhang
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
Signed-off-by: Jack Zhang
---
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 && (>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 = _fence->base;
+ am_fence->ring = ring;
+ } else {
+ /* take use of job-embedded fence */
+ fence = >hw_fence;
+ job->ring = ring;
+ }
seq = ++ring->fence_drv.sync_seq;
- fence->ring = ring;
- dma_fence_init(>base, _fence_ops,
+ dma_fence_init(fence, _fence_ops,
>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, >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(>base));
+ rcu_assign_pointer(*ptr, dma_fence_get(fence));
- *f = >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;
+
+
[PATCH 1/2] drm/amd/amdgpu embed hw_fence into amdgpu_job
From: Jack Zhang
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
Signed-off-by: Jack Zhang
---
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 && (>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 = _fence->base;
+ am_fence->ring = ring;
+ } else {
+ /* take use of job-embedded fence */
+ fence = >hw_fence;
+ job->ring = ring;
+ }
seq = ++ring->fence_drv.sync_seq;
- fence->ring = ring;
- dma_fence_init(>base, _fence_ops,
+ dma_fence_init(fence, _fence_ops,
>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, >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(>base));
+ rcu_assign_pointer(*ptr, dma_fence_get(fence));
- *f = >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;
+
+
