Skip to contentSkip to conversation

drm/msm: add trace_dma_fence_emit to msm_gpu_submit

PATCH Superseded Chia-I Wu 4 participants 6 messages Apr 8, 2022 – Apr 26, 2022

Subsystems (4)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Chia-I Wu

author
#0 lore raw

PATCH 1/1 drm/msm: add trace_dma_fence_emit to msm_gpu_submit

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

---
 drivers/gpu/drm/msm/msm_gpu.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -0
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
 #include <linux/string_helpers.h>
 #include <linux/devcoredump.h>
 #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
 /*
  * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 	gpu->active_submits++;
 	mutex_unlock(&gpu->active_lock);
 
+	trace_dma_fence_emit(submit->hw_fence);
 	gpu->funcs->submit(gpu, submit);
 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 
Signature · 1 line
2.35.1.1178.g4f1659d476-goog
Reviewed-by on v1 1/1
#1 lore raw
Quoted 3 lines · Chia-I Wu

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

Please mention in the commit message that the trace_dma_fence_init() is called from dma_fence_init().
With that in place:

Quoted 30 lines
---
 drivers/gpu/drm/msm/msm_gpu.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -1
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
 #include <linux/string_helpers.h>
 #include <linux/devcoredump.h>
 #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
 /*
  * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
        gpu->active_submits++;
        mutex_unlock(&gpu->active_lock);
 
+       trace_dma_fence_emit(submit->hw_fence);
        gpu->funcs->submit(gpu, submit);
        gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 
--

2.35.1.1178.g4f1659d476-goog

Signature · 2 lines
With best wishes
Dmitry
#2 lore raw
In reply to Chia-I Wu

Am 08.04.22 um 23:12 schrieb Chia-I Wu:

Quoted 2 lines

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

Well, the reason why basically no driver is calling this is because it is pretty much deprecated.

We do have a case in the GPU scheduler where it makes sense to distinct between init and emit, but it doesn't really matter for drivers.

So I'm not sure if it's a good idea to add that here.

Regards,
Christian.

Quoted 27 lines
---
  drivers/gpu/drm/msm/msm_gpu.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -0
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
  #include <linux/string_helpers.h>
  #include <linux/devcoredump.h>
  #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
  /*
   * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
  	gpu->active_submits++;
  	mutex_unlock(&gpu->active_lock);
 
+	trace_dma_fence_emit(submit->hw_fence);
  	gpu->funcs->submit(gpu, submit);
  	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 
author
#3 lore raw
Quoted 7 lines · Christian König

Am 08.04.22 um 23:12 schrieb Chia-I Wu:

Quoted 2 lines

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

Well, the reason why basically no driver is calling this is because it is pretty much deprecated.

Why is it considered deprecated? trace_dma_fence_{emit,signaled} are useful to visualize fence timelines. I am actually less sure about how trace_dma_fence_{init,destroy} are used.

Is it because trace_dma_fence_init is called automatically, and is good enough most of the time?

Quoted 3 lines

We do have a case in the GPU scheduler where it makes sense to distinct between init and emit, but it doesn't really matter for drivers.

virtio also has a case where init and emit can be far apart, when the host cannot process commands fast enough and there is no space in virtqueue.

Quoted 34 lines

So I'm not sure if it's a good idea to add that here.

Regards,
Christian.

Quoted 27 lines
---
  drivers/gpu/drm/msm/msm_gpu.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -0
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
  #include <linux/string_helpers.h>
  #include <linux/devcoredump.h>
  #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
  /*
   * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
      gpu->active_submits++;
      mutex_unlock(&gpu->active_lock);
 
+     trace_dma_fence_emit(submit->hw_fence);
      gpu->funcs->submit(gpu, submit);
      gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 
#4 lore raw
In reply to Christian König
Quoted 5 lines · Christian König

Am 08.04.22 um 23:12 schrieb Chia-I Wu:

Quoted 2 lines

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

Quoted 7 lines

Well, the reason why basically no driver is calling this is because it is pretty much deprecated.

We do have a case in the GPU scheduler where it makes sense to distinct between init and emit, but it doesn't really matter for drivers.

So I'm not sure if it's a good idea to add that here.

visualization can't easily differentiate between drivers/timelines where the split matters and ones where it doesn't.. IMO it is better to just have the extra trace even in the cases where it comes at the same time as the init trace

BR,
-R

Quoted 31 lines

Regards,
Christian.

Quoted 27 lines
---
  drivers/gpu/drm/msm/msm_gpu.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -0
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
  #include <linux/string_helpers.h>
  #include <linux/devcoredump.h>
  #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
  /*
   * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
      gpu->active_submits++;
      mutex_unlock(&gpu->active_lock);
 
+     trace_dma_fence_emit(submit->hw_fence);
      gpu->funcs->submit(gpu, submit);
      gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 
#5 lore raw

Sorry for the delayed reply.

Am 12.04.22 um 21:41 schrieb Rob Clark:

Quoted 7 lines · Christian König
Quoted 6 lines

Am 08.04.22 um 23:12 schrieb Chia-I Wu:

Quoted 2 lines

In practice, trace_dma_fence_init is good enough and almost no driver calls trace_dma_fence_emit. But this is still more correct in theory.

Well, the reason why basically no driver is calling this is because it is pretty much deprecated.

Quoted 8 lines
Quoted 4 lines

We do have a case in the GPU scheduler where it makes sense to distinct between init and emit, but it doesn't really matter for drivers.

So I'm not sure if it's a good idea to add that here.

visualization can't easily differentiate between drivers/timelines where the split matters and ones where it doesn't.. IMO it is better to just have the extra trace even in the cases where it comes at the same time as the init trace

That's exactly the reason why I want to remove the extra trace.

To make it clear this is only useful for debugging and *NOT* for actually visualizing things.

So by adding that here you add more confusion than solving anything.

Regards,
Christian.

Quoted 33 lines

BR,
-R

Quoted 29 lines

Regards,
Christian.

Quoted 26 lines
---
   drivers/gpu/drm/msm/msm_gpu.c | 2 ++
   1 file changed, 2 insertions(+)
drivers/gpu/drm/msm/msm_gpu.c +2 -0
index faf0c242874e..a82193f41ea2 100644
@@ -15,6 +15,7 @@
   #include <linux/string_helpers.h>
   #include <linux/devcoredump.h>
   #include <linux/sched/task.h>
+#include <trace/events/dma_fence.h>
 
   /*
    * Power Management:
@@ -769,6 +770,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
       gpu->active_submits++;
       mutex_unlock(&gpu->active_lock);
 
+     trace_dma_fence_emit(submit->hw_fence);
       gpu->funcs->submit(gpu, submit);
       gpu->cur_ctx_seqno = submit->queue->ctx->seqno;