From b3f8de4d0505795817029d7a77b1417bd440789f Mon Sep 17 00:00:00 2001 From: Jamie Gennis Date: Thu, 21 Feb 2013 17:55:28 -0800 Subject: [PATCH] trace/events: add gpu trace events Change-Id: I0607b9c776acf61cb796b8572cf8cfb8b2dc1377 Signed-off-by: Jamie Gennis --- include/trace/events/gpu.h | 142 +++++++++++++++++++++++++++++++++++++++++++++ kernel/trace/Kconfig | 3 + kernel/trace/Makefile | 1 + kernel/trace/gpu-traces.c | 23 ++++++++ 4 files changed, 169 insertions(+) create mode 100644 include/trace/events/gpu.h create mode 100644 kernel/trace/gpu-traces.c diff --git a/include/trace/events/gpu.h b/include/trace/events/gpu.h new file mode 100644 index 000000000000..09efa71d66c6 --- /dev/null +++ b/include/trace/events/gpu.h @@ -0,0 +1,142 @@ +#undef TRACE_SYSTEM +#define TRACE_SYSTEM gpu + +#if !defined(_TRACE_GPU_H) || defined(TRACE_HEADER_MULTI_READ) +#define _TRACE_GPU_H + +#include +#include + +#define show_secs_from_ns(ns) \ + ({ \ + u64 t = ns + (NSEC_PER_USEC / 2); \ + do_div(t, NSEC_PER_SEC); \ + }) + +#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=%5llu.%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 diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig index a5da09c899dd..37e1fb831d95 100644 --- a/kernel/trace/Kconfig +++ b/kernel/trace/Kconfig @@ -77,6 +77,9 @@ config EVENT_TRACING select CONTEXT_SWITCH_TRACER bool +config GPU_TRACEPOINTS + bool + config CONTEXT_SWITCH_TRACER bool diff --git a/kernel/trace/Makefile b/kernel/trace/Makefile index 67d6369ddf83..5da3cfb1bb84 100644 --- a/kernel/trace/Makefile +++ b/kernel/trace/Makefile @@ -63,6 +63,7 @@ obj-$(CONFIG_KGDB_KDB) += trace_kdb.o endif obj-$(CONFIG_PROBE_EVENTS) += trace_probe.o obj-$(CONFIG_UPROBE_EVENT) += trace_uprobe.o +obj-$(CONFIG_GPU_TRACEPOINTS) += gpu-traces.o obj-$(CONFIG_TRACEPOINT_BENCHMARK) += trace_benchmark.o diff --git a/kernel/trace/gpu-traces.c b/kernel/trace/gpu-traces.c new file mode 100644 index 000000000000..a4b3f00faee3 --- /dev/null +++ b/kernel/trace/gpu-traces.c @@ -0,0 +1,23 @@ +/* + * GPU tracepoints + * + * Copyright (C) 2013 Google, Inc. + * + * This software is licensed under the terms of the GNU General Public + * License version 2, as published by the Free Software Foundation, and + * may be copied, distributed, and modified under those terms. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + */ + +#include + +#define CREATE_TRACE_POINTS +#include + +EXPORT_TRACEPOINT_SYMBOL(gpu_sched_switch); +EXPORT_TRACEPOINT_SYMBOL(gpu_job_enqueue); -- 2.11.0