2 * Copyright (C) 2015 The Android Open Source Project
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
28 #include <android-base/logging.h>
29 #include <android-base/strings.h>
32 #include "environment.h"
33 #include "event_attr.h"
35 #include "event_selection_set.h"
36 #include "event_type.h"
40 static std::vector<std::string> default_measured_event_types{
41 "cpu-cycles", "stalled-cycles-frontend", "stalled-cycles-backend",
42 "instructions", "branch-instructions", "branch-misses",
43 "task-clock", "context-switches", "page-faults",
46 static volatile bool signaled;
47 static void signal_handler(int) {
51 class StatCommand : public Command {
54 : Command("stat", "gather performance counter information",
55 "Usage: simpleperf stat [options] [command [command-args]]\n"
56 " Gather performance counter information of running [command].\n"
57 " -a Collect system-wide information.\n"
58 " --cpu cpu_item1,cpu_item2,...\n"
59 " Collect information only on the selected cpus. cpu_item can\n"
60 " be a cpu number like 1, or a cpu range like 0-3.\n"
61 " -e event1[:modifier1],event2[:modifier2],...\n"
62 " Select the event list to count. Use `simpleperf list` to find\n"
63 " all possible event names. Modifiers can be added to define\n"
64 " how the event should be monitored. Possible modifiers are:\n"
65 " u - monitor user space events only\n"
66 " k - monitor kernel space events only\n"
68 " Don't stat created child threads/processes.\n"
70 " Stat events on existing processes. Mutually exclusive with -a.\n"
72 " Stat events on existing threads. Mutually exclusive with -a.\n"
73 " --verbose Show result in verbose mode.\n"),
75 system_wide_collection_(false),
76 child_inherit_(true) {
78 signal_handler_register_.reset(
79 new SignalHandlerRegister({SIGCHLD, SIGINT, SIGTERM}, signal_handler));
82 bool Run(const std::vector<std::string>& args);
85 bool ParseOptions(const std::vector<std::string>& args, std::vector<std::string>* non_option_args);
86 bool AddMeasuredEventType(const std::string& event_type_name);
87 bool AddDefaultMeasuredEventTypes();
88 bool SetEventSelection();
89 bool ShowCounters(const std::vector<CountersInfo>& counters, double duration_in_sec);
92 bool system_wide_collection_;
94 std::vector<pid_t> monitored_threads_;
95 std::vector<int> cpus_;
96 std::vector<EventTypeAndModifier> measured_event_types_;
97 EventSelectionSet event_selection_set_;
99 std::unique_ptr<SignalHandlerRegister> signal_handler_register_;
102 bool StatCommand::Run(const std::vector<std::string>& args) {
103 // 1. Parse options, and use default measured event types if not given.
104 std::vector<std::string> workload_args;
105 if (!ParseOptions(args, &workload_args)) {
108 if (measured_event_types_.empty()) {
109 if (!AddDefaultMeasuredEventTypes()) {
113 if (!SetEventSelection()) {
117 // 2. Create workload.
118 std::unique_ptr<Workload> workload;
119 if (!workload_args.empty()) {
120 workload = Workload::CreateWorkload(workload_args);
121 if (workload == nullptr) {
125 if (!system_wide_collection_ && monitored_threads_.empty()) {
126 if (workload != nullptr) {
127 monitored_threads_.push_back(workload->GetPid());
128 event_selection_set_.SetEnableOnExec(true);
130 LOG(ERROR) << "No threads to monitor. Try `simpleperf help stat` for help\n";
135 // 3. Open perf_event_files.
136 if (system_wide_collection_) {
137 if (!event_selection_set_.OpenEventFilesForCpus(cpus_)) {
141 if (!event_selection_set_.OpenEventFilesForThreadsOnCpus(monitored_threads_, cpus_)) {
146 // 4. Count events while workload running.
147 auto start_time = std::chrono::steady_clock::now();
148 if (workload != nullptr && !workload->Start()) {
154 auto end_time = std::chrono::steady_clock::now();
156 // 5. Read and print counters.
157 std::vector<CountersInfo> counters;
158 if (!event_selection_set_.ReadCounters(&counters)) {
161 double duration_in_sec =
162 std::chrono::duration_cast<std::chrono::duration<double>>(end_time - start_time).count();
163 if (!ShowCounters(counters, duration_in_sec)) {
169 bool StatCommand::ParseOptions(const std::vector<std::string>& args,
170 std::vector<std::string>* non_option_args) {
171 std::set<pid_t> tid_set;
173 for (i = 0; i < args.size() && args[i].size() > 0 && args[i][0] == '-'; ++i) {
174 if (args[i] == "-a") {
175 system_wide_collection_ = true;
176 } else if (args[i] == "--cpu") {
177 if (!NextArgumentOrError(args, &i)) {
180 cpus_ = GetCpusFromString(args[i]);
181 } else if (args[i] == "-e") {
182 if (!NextArgumentOrError(args, &i)) {
185 std::vector<std::string> event_types = android::base::Split(args[i], ",");
186 for (auto& event_type : event_types) {
187 if (!AddMeasuredEventType(event_type)) {
191 } else if (args[i] == "--no-inherit") {
192 child_inherit_ = false;
193 } else if (args[i] == "-p") {
194 if (!NextArgumentOrError(args, &i)) {
197 if (!GetValidThreadsFromProcessString(args[i], &tid_set)) {
200 } else if (args[i] == "-t") {
201 if (!NextArgumentOrError(args, &i)) {
204 if (!GetValidThreadsFromThreadString(args[i], &tid_set)) {
207 } else if (args[i] == "--verbose") {
208 verbose_mode_ = true;
210 ReportUnknownOption(args, i);
215 monitored_threads_.insert(monitored_threads_.end(), tid_set.begin(), tid_set.end());
216 if (system_wide_collection_ && !monitored_threads_.empty()) {
217 LOG(ERROR) << "Stat system wide and existing processes/threads can't be used at the same time.";
221 if (non_option_args != nullptr) {
222 non_option_args->clear();
223 for (; i < args.size(); ++i) {
224 non_option_args->push_back(args[i]);
230 bool StatCommand::AddMeasuredEventType(const std::string& event_type_name) {
231 std::unique_ptr<EventTypeAndModifier> event_type_modifier = ParseEventType(event_type_name);
232 if (event_type_modifier == nullptr) {
235 measured_event_types_.push_back(*event_type_modifier);
239 bool StatCommand::AddDefaultMeasuredEventTypes() {
240 for (auto& name : default_measured_event_types) {
241 // It is not an error when some event types in the default list are not supported by the kernel.
242 const EventType* type = FindEventTypeByName(name);
243 if (type != nullptr && IsEventAttrSupportedByKernel(CreateDefaultPerfEventAttr(*type))) {
244 AddMeasuredEventType(name);
247 if (measured_event_types_.empty()) {
248 LOG(ERROR) << "Failed to add any supported default measured types";
254 bool StatCommand::SetEventSelection() {
255 for (auto& event_type : measured_event_types_) {
256 if (!event_selection_set_.AddEventType(event_type)) {
260 event_selection_set_.SetInherit(child_inherit_);
264 static std::string ReadableCountValue(uint64_t count,
265 const EventTypeAndModifier& event_type_modifier) {
266 if (event_type_modifier.event_type.name == "cpu-clock" ||
267 event_type_modifier.event_type.name == "task-clock") {
268 double value = count / 1e6;
269 return android::base::StringPrintf("%lf(ms)", value);
271 std::string s = android::base::StringPrintf("%" PRIu64, count);
272 for (size_t i = s.size() - 1, j = 1; i > 0; --i, ++j) {
274 s.insert(s.begin() + i, ',');
282 struct CounterSummary {
283 const EventTypeAndModifier* event_type;
286 std::string readable_count_str;
290 static std::string GetCommentForSummary(const CounterSummary& summary,
291 const std::vector<CounterSummary>& summaries,
292 double duration_in_sec) {
293 const std::string& type_name = summary.event_type->event_type.name;
294 const std::string& modifier = summary.event_type->modifier;
295 if (type_name == "task-clock") {
296 double run_sec = summary.count / 1e9;
297 double cpu_usage = run_sec / duration_in_sec;
298 return android::base::StringPrintf("%lf%% cpu usage", cpu_usage * 100);
300 if (type_name == "cpu-clock") {
303 if (type_name == "cpu-cycles") {
304 double hz = summary.count / duration_in_sec;
305 return android::base::StringPrintf("%lf GHz", hz / 1e9);
307 if (type_name == "instructions" && summary.count != 0) {
308 for (auto& t : summaries) {
309 if (t.event_type->event_type.name == "cpu-cycles" && t.event_type->modifier == modifier) {
310 double cycles_per_instruction = t.count * 1.0 / summary.count;
311 return android::base::StringPrintf("%lf cycles per instruction", cycles_per_instruction);
315 if (android::base::EndsWith(type_name, "-misses")) {
317 if (type_name == "cache-misses") {
318 s = "cache-references";
319 } else if (type_name == "branch-misses") {
320 s = "branch-instructions";
322 s = type_name.substr(0, type_name.size() - strlen("-misses")) + "s";
324 for (auto& t : summaries) {
325 if (t.event_type->event_type.name == s && t.event_type->modifier == modifier && t.count != 0) {
326 double miss_rate = summary.count * 1.0 / t.count;
327 return android::base::StringPrintf("%lf%% miss rate", miss_rate * 100);
331 double rate = summary.count / duration_in_sec;
333 return android::base::StringPrintf("%.3lf G/sec", rate / 1e9);
336 return android::base::StringPrintf("%.3lf M/sec", rate / 1e6);
339 return android::base::StringPrintf("%.3lf K/sec", rate / 1e3);
341 return android::base::StringPrintf("%.3lf /sec", rate);
344 bool StatCommand::ShowCounters(const std::vector<CountersInfo>& counters, double duration_in_sec) {
345 printf("Performance counter statistics:\n\n");
348 for (auto& counters_info : counters) {
349 const EventTypeAndModifier* event_type = counters_info.event_type;
350 for (auto& counter_info : counters_info.counters) {
351 printf("%s(tid %d, cpu %d): count %s, time_enabled %" PRIu64 ", time running %" PRIu64
352 ", id %" PRIu64 "\n",
353 event_type->name.c_str(), counter_info.tid, counter_info.cpu,
354 ReadableCountValue(counter_info.counter.value, *event_type).c_str(),
355 counter_info.counter.time_enabled, counter_info.counter.time_running,
356 counter_info.counter.id);
361 std::vector<CounterSummary> summaries;
362 for (auto& counters_info : counters) {
363 uint64_t value_sum = 0;
364 uint64_t time_enabled_sum = 0;
365 uint64_t time_running_sum = 0;
366 for (auto& counter_info : counters_info.counters) {
367 value_sum += counter_info.counter.value;
368 time_enabled_sum += counter_info.counter.time_enabled;
369 time_running_sum += counter_info.counter.time_running;
372 uint64_t scaled_count = value_sum;
373 if (time_running_sum < time_enabled_sum) {
374 if (time_running_sum == 0) {
377 scale = static_cast<double>(time_enabled_sum) / time_running_sum;
378 scaled_count = static_cast<uint64_t>(scale * value_sum);
381 CounterSummary summary;
382 summary.event_type = counters_info.event_type;
383 summary.count = scaled_count;
384 summary.scale = scale;
385 summary.readable_count_str = ReadableCountValue(summary.count, *summary.event_type);
386 summaries.push_back(summary);
389 for (auto& summary : summaries) {
390 summary.comment = GetCommentForSummary(summary, summaries, duration_in_sec);
393 size_t count_column_width = 0;
394 size_t name_column_width = 0;
395 size_t comment_column_width = 0;
396 for (auto& summary : summaries) {
397 count_column_width = std::max(count_column_width, summary.readable_count_str.size());
398 name_column_width = std::max(name_column_width, summary.event_type->name.size());
399 comment_column_width = std::max(comment_column_width, summary.comment.size());
402 for (auto& summary : summaries) {
403 printf(" %*s %-*s # %-*s (%.0lf%%)\n", static_cast<int>(count_column_width),
404 summary.readable_count_str.c_str(), static_cast<int>(name_column_width),
405 summary.event_type->name.c_str(), static_cast<int>(comment_column_width),
406 summary.comment.c_str(), 1.0 / summary.scale * 100);
409 printf("\nTotal test time: %lf seconds.\n", duration_in_sec);
413 __attribute__((constructor)) static void RegisterStatCommand() {
414 RegisterCommand("stat", [] { return std::unique_ptr<Command>(new StatCommand); });