aboutsummaryrefslogtreecommitdiff
path: root/include/trace/events/gpu.h
diff options
context:
space:
mode:
Diffstat (limited to 'include/trace/events/gpu.h')
-rw-r--r--include/trace/events/gpu.h143
1 files changed, 143 insertions, 0 deletions
diff --git a/include/trace/events/gpu.h b/include/trace/events/gpu.h
new file mode 100644
index 000000000000..7e15cdfafe5a
--- /dev/null
+++ b/include/trace/events/gpu.h
@@ -0,0 +1,143 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM gpu
+
+#if !defined(_TRACE_GPU_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_GPU_H
+
+#include <linux/tracepoint.h>
+#include <linux/time.h>
+
+#define show_secs_from_ns(ns) \
+ ({ \
+ u64 t = ns + (NSEC_PER_USEC / 2); \
+ do_div(t, NSEC_PER_SEC); \
+ t; \
+ })
+
+#define show_usecs_from_ns(ns) \
+ ({ \
+ u64 t = ns + (NSEC_PER_USEC / 2) ; \
+ u32 rem; \
+ do_div(t, NSEC_PER_USEC); \
+ rem = do_div(t, USEC_PER_SEC); \
+ })
+
+/*
+ * The gpu_sched_switch event indicates that a switch from one GPU context to
+ * another occurred on one of the GPU hardware blocks.
+ *
+ * The gpu_name argument identifies the GPU hardware block. Each independently
+ * scheduled GPU hardware block should have a different name. This may be used
+ * in different ways for different GPUs. For example, if a GPU includes
+ * multiple processing cores it may use names "GPU 0", "GPU 1", etc. If a GPU
+ * includes a separately scheduled 2D and 3D hardware block, it might use the
+ * names "2D" and "3D".
+ *
+ * The timestamp argument is the timestamp at which the switch occurred on the
+ * GPU. These timestamps are in units of nanoseconds and must use
+ * approximately the same time as sched_clock, though they need not come from
+ * any CPU clock. The timestamps for a single hardware block must be
+ * monotonically nondecreasing. This means that if a variable compensation
+ * offset is used to translate from some other clock to the sched_clock, then
+ * care must be taken when increasing that offset, and doing so may result in
+ * multiple events with the same timestamp.
+ *
+ * The next_ctx_id argument identifies the next context that was running on
+ * the GPU hardware block. A value of 0 indicates that the hardware block
+ * will be idle.
+ *
+ * The next_prio argument indicates the priority of the next context at the
+ * time of the event. The exact numeric values may mean different things for
+ * different GPUs, but they should follow the rule that lower values indicate a
+ * higher priority.
+ *
+ * The next_job_id argument identifies the batch of work that the GPU will be
+ * working on. This should correspond to a job_id that was previously traced
+ * as a gpu_job_enqueue event when the batch of work was created.
+ */
+TRACE_EVENT(gpu_sched_switch,
+
+ TP_PROTO(const char *gpu_name, u64 timestamp,
+ u32 next_ctx_id, s32 next_prio, u32 next_job_id),
+
+ TP_ARGS(gpu_name, timestamp, next_ctx_id, next_prio, next_job_id),
+
+ TP_STRUCT__entry(
+ __string( gpu_name, gpu_name )
+ __field( u64, timestamp )
+ __field( u32, next_ctx_id )
+ __field( s32, next_prio )
+ __field( u32, next_job_id )
+ ),
+
+ TP_fast_assign(
+ __assign_str(gpu_name, gpu_name);
+ __entry->timestamp = timestamp;
+ __entry->next_ctx_id = next_ctx_id;
+ __entry->next_prio = next_prio;
+ __entry->next_job_id = next_job_id;
+ ),
+
+ TP_printk("gpu_name=%s ts=%llu.%06lu next_ctx_id=%lu next_prio=%ld "
+ "next_job_id=%lu",
+ __get_str(gpu_name),
+ (unsigned long long)show_secs_from_ns(__entry->timestamp),
+ (unsigned long)show_usecs_from_ns(__entry->timestamp),
+ (unsigned long)__entry->next_ctx_id,
+ (long)__entry->next_prio,
+ (unsigned long)__entry->next_job_id)
+);
+
+/*
+ * The gpu_job_enqueue event indicates that a batch of work has been queued up
+ * to be processed by the GPU. This event is not intended to indicate that
+ * the batch of work has been submitted to the GPU hardware, but rather that
+ * it has been submitted to the GPU kernel driver.
+ *
+ * This event should be traced on the thread that initiated the work being
+ * queued. For example, if a batch of work is submitted to the kernel by a
+ * userland thread, the event should be traced on that thread.
+ *
+ * The ctx_id field identifies the GPU context in which the batch of work
+ * being queued is to be run.
+ *
+ * The job_id field identifies the batch of work being queued within the given
+ * GPU context. The first batch of work submitted for a given GPU context
+ * should have a job_id of 0, and each subsequent batch of work should
+ * increment the job_id by 1.
+ *
+ * The type field identifies the type of the job being enqueued. The job
+ * types may be different for different GPU hardware. For example, a GPU may
+ * differentiate between "2D", "3D", and "compute" jobs.
+ */
+TRACE_EVENT(gpu_job_enqueue,
+
+ TP_PROTO(u32 ctx_id, u32 job_id, const char *type),
+
+ TP_ARGS(ctx_id, job_id, type),
+
+ TP_STRUCT__entry(
+ __field( u32, ctx_id )
+ __field( u32, job_id )
+ __string( type, type )
+ ),
+
+ TP_fast_assign(
+ __entry->ctx_id = ctx_id;
+ __entry->job_id = job_id;
+ __assign_str(type, type);
+ ),
+
+ TP_printk("ctx_id=%lu job_id=%lu type=%s",
+ (unsigned long)__entry->ctx_id,
+ (unsigned long)__entry->job_id,
+ __get_str(type))
+);
+
+#undef show_secs_from_ns
+#undef show_usecs_from_ns
+
+#endif /* _TRACE_GPU_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>