Add ftrace events for user queue fence emission and completion.
Signed-off-by: Prike Liang <[email protected]>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h | 48 +++++++++++++++++++
.../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c | 8 ++++
2 files changed, 56 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
index 484fbb00068b..d421a44ef933 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
@@ -661,6 +661,54 @@ DEFINE_EVENT(amdgpu_userq_queue_result,
amdgpu_userq_map_end,
DEFINE_EVENT(amdgpu_userq_queue_result, amdgpu_userq_unmap_end,
TP_PROTO(struct amdgpu_usermode_queue *queue, int result),
TP_ARGS(queue, result));
+
+TRACE_EVENT(amdgpu_userq_fence_emit,
+ TP_PROTO(struct amdgpu_usermode_queue *queue, u64 seqno,
+ unsigned int num_read_handles,
+ unsigned int num_write_handles,
+ unsigned int num_syncobj_handles),
+ TP_ARGS(queue, seqno, num_read_handles, num_write_handles,
+ num_syncobj_handles),
+ TP_STRUCT__entry(
+ __field(struct amdgpu_usermode_queue *, queue)
+ __field(u64, doorbell_index)
+ __field(int, queue_type)
+ __field(u64, seqno)
+ __field(unsigned int, num_read_handles)
+ __field(unsigned int, num_write_handles)
+ __field(unsigned int, num_syncobj_handles)
+ ),
+ TP_fast_assign(
+ __entry->queue = queue;
+ __entry->doorbell_index = queue ?
queue->doorbell_index : 0;
+ __entry->queue_type = queue ? queue->queue_type : -1;
+ __entry->seqno = seqno;
+ __entry->num_read_handles = num_read_handles;
+ __entry->num_write_handles = num_write_handles;
+ __entry->num_syncobj_handles = num_syncobj_handles;
+ ),
+ TP_printk("queue=%p, doorbell=%llu, type=%d, seqno=%llu, read=%u,
write=%u, syncobj=%u",
+ __entry->queue, __entry->doorbell_index,
+ __entry->queue_type, __entry->seqno,
+ __entry->num_read_handles, __entry->num_write_handles,
+ __entry->num_syncobj_handles)
+);
+TRACE_EVENT(amdgpu_userq_fence_signal,
+ TP_PROTO(u64 context, u64 seqno, u64 rptr),
+ TP_ARGS(context, seqno, rptr),
+ TP_STRUCT__entry(
+ __field(u64, context)
+ __field(u64, seqno)
+ __field(u64, rptr)
+ ),
+ TP_fast_assign(
+ __entry->context = context;
+ __entry->seqno = seqno;
+ __entry->rptr = rptr;
+ ),
+ TP_printk("context=%llu, seqno=%llu, rptr=%llu",
+ __entry->context, __entry->seqno, __entry->rptr)
+);
#undef AMDGPU_JOB_GET_TIMELINE_NAME
#endif
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
index 369914cab555..50d1c3627279 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
@@ -31,6 +31,7 @@
#include "amdgpu.h"
#include "amdgpu_userq_fence.h"
+#include "amdgpu_trace.h"
#define AMDGPU_USERQ_MAX_HANDLES (1U << 16)
@@ -168,6 +169,9 @@ amdgpu_userq_fence_driver_process(struct
amdgpu_userq_fence_driver *fence_drv)
list_for_each_entry_safe(userq_fence, tmp, &to_be_signaled, link) {
fence = &userq_fence->base;
list_del_init(&userq_fence->link);
+ trace_amdgpu_userq_fence_signal(fence_drv->context,
+ userq_fence->base.seqno,
+ rptr);
dma_fence_signal(fence);
/* Drop fence_drv_array outside fence_list_lock
* to avoid the recursion lock.
@@ -530,6 +534,10 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, void
*data,
/* Create the new fence */
amdgpu_userq_fence_init(queue, fence, wptr);
+ trace_amdgpu_userq_fence_emit(queue, fence->base.seqno,
+ num_read_bo_handles,
+ num_write_bo_handles,
+ num_syncobj_handles);
mutex_unlock(&userq_mgr->userq_mutex);
--
2.34.1
