1 /*
2 * Copyright (C) 2015 The Android Open Source Project
3 *
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
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
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.
15 */
16
17 #include <inttypes.h>
18 #include <signal.h>
19 #include <stdio.h>
20 #include <string.h>
21 #include <sys/prctl.h>
22
23 #include <algorithm>
24 #include <chrono>
25 #include <optional>
26 #include <set>
27 #include <string>
28 #include <string_view>
29 #include <vector>
30
31 #include <android-base/file.h>
32 #include <android-base/logging.h>
33 #include <android-base/strings.h>
34 #include <android-base/unique_fd.h>
35
36 #include "IOEventLoop.h"
37 #include "ProbeEvents.h"
38 #include "cmd_stat_impl.h"
39 #include "command.h"
40 #include "environment.h"
41 #include "event_attr.h"
42 #include "event_fd.h"
43 #include "event_selection_set.h"
44 #include "event_type.h"
45 #include "utils.h"
46 #include "workload.h"
47
48 namespace simpleperf {
49
50 using android::base::Split;
51
52 static std::vector<std::string> default_measured_event_types{
53 "cpu-cycles", "stalled-cycles-frontend", "stalled-cycles-backend",
54 "instructions", "branch-instructions", "branch-misses",
55 "task-clock", "context-switches", "page-faults",
56 };
57
58 static const std::unordered_map<std::string_view, std::pair<std::string_view, std::string_view>>
59 COMMON_EVENT_RATE_MAP = {
60 {"cache-misses", {"cache-references", "miss rate"}},
61 {"branch-misses", {"branch-instructions", "miss rate"}},
62 };
63
64 static const std::unordered_map<std::string_view, std::pair<std::string_view, std::string_view>>
65 ARM_EVENT_RATE_MAP = {
66 // Refer to "D6.10.5 Meaningful ratios between common microarchitectural events" in ARMv8
67 // specification.
68 {"raw-l1i-cache-refill", {"raw-l1i-cache", "level 1 instruction cache refill rate"}},
69 {"raw-l1i-tlb-refill", {"raw-l1i-tlb", "level 1 instruction TLB refill rate"}},
70 {"raw-l1d-cache-refill", {"raw-l1d-cache", "level 1 data or unified cache refill rate"}},
71 {"raw-l1d-tlb-refill", {"raw-l1d-tlb", "level 1 data or unified TLB refill rate"}},
72 {"raw-l2d-cache-refill", {"raw-l2d-cache", "level 2 data or unified cache refill rate"}},
73 {"raw-l2i-cache-refill", {"raw-l2i-cache", "level 2 instruction cache refill rate"}},
74 {"raw-l3d-cache-refill", {"raw-l3d-cache", "level 3 data or unified cache refill rate"}},
75 {"raw-l2d-tlb-refill", {"raw-l2d-tlb", "level 2 data or unified TLB refill rate"}},
76 {"raw-l2i-tlb-refill", {"raw-l2i-tlb", "level 2 instruction TLB refill rate"}},
77 {"raw-bus-access", {"raw-bus-cycles", "bus accesses per cycle"}},
78 {"raw-ll-cache-miss", {"raw-ll-cache", "last level data or unified cache refill rate"}},
79 {"raw-dtlb-walk", {"raw-l1d-tlb", "data TLB miss rate"}},
80 {"raw-itlb-walk", {"raw-l1i-tlb", "instruction TLB miss rate"}},
81 {"raw-ll-cache-miss-rd", {"raw-ll-cache-rd", "memory read operation miss rate"}},
82 {"raw-remote-access-rd",
83 {"raw-remote-access", "read accesses to another socket in a multi-socket system"}},
84 // Refer to "Table K3-2 Relationship between REFILL events and associated access events" in
85 // ARMv8 specification.
86 {"raw-l1d-cache-refill-rd", {"raw-l1d-cache-rd", "level 1 cache refill rate, read"}},
87 {"raw-l1d-cache-refill-wr", {"raw-l1d-cache-wr", "level 1 cache refill rate, write"}},
88 {"raw-l1d-tlb-refill-rd", {"raw-l1d-tlb-rd", "level 1 TLB refill rate, read"}},
89 {"raw-l1d-tlb-refill-wr", {"raw-l1d-tlb-wr", "level 1 TLB refill rate, write"}},
90 {"raw-l2d-cache-refill-rd", {"raw-l2d-cache-rd", "level 2 data cache refill rate, read"}},
91 {"raw-l2d-cache-refill-wr", {"raw-l2d-cache-wr", "level 2 data cache refill rate, write"}},
92 {"raw-l2d-tlb-refill-rd", {"raw-l2d-tlb-rd", "level 2 data TLB refill rate, read"}},
93 };
94
ReadableCountValue(bool csv)95 std::string CounterSummary::ReadableCountValue(bool csv) {
96 if (type_name == "cpu-clock" || type_name == "task-clock") {
97 // Convert nanoseconds to milliseconds.
98 double value = count / 1e6;
99 return android::base::StringPrintf("%lf(ms)", value);
100 }
101 if (csv) {
102 return android::base::StringPrintf("%" PRIu64, count);
103 }
104 return ReadableCount(count);
105 }
106
FindSummary(const std::string & type_name,const std::string & modifier,const ThreadInfo * thread,int cpu)107 const CounterSummary* CounterSummaries::FindSummary(const std::string& type_name,
108 const std::string& modifier,
109 const ThreadInfo* thread, int cpu) {
110 for (const auto& s : summaries_) {
111 if (s.type_name == type_name && s.modifier == modifier && s.thread == thread && s.cpu == cpu) {
112 return &s;
113 }
114 }
115 return nullptr;
116 }
117
AutoGenerateSummaries()118 void CounterSummaries::AutoGenerateSummaries() {
119 for (size_t i = 0; i < summaries_.size(); ++i) {
120 const CounterSummary& s = summaries_[i];
121 if (s.modifier == "u") {
122 const CounterSummary* other = FindSummary(s.type_name, "k", s.thread, s.cpu);
123 if (other != nullptr && other->IsMonitoredAtTheSameTime(s)) {
124 if (FindSummary(s.type_name, "", s.thread, s.cpu) == nullptr) {
125 summaries_.emplace_back(s.type_name, "", s.group_id, s.thread, s.cpu,
126 s.count + other->count, s.runtime_in_ns, s.scale, true, csv_);
127 }
128 }
129 }
130 }
131 }
132
GenerateComments(double duration_in_sec)133 void CounterSummaries::GenerateComments(double duration_in_sec) {
134 for (auto& s : summaries_) {
135 s.comment = GetCommentForSummary(s, duration_in_sec);
136 }
137 }
138
Show(FILE * fp)139 void CounterSummaries::Show(FILE* fp) {
140 bool show_thread = !summaries_.empty() && summaries_[0].thread != nullptr;
141 bool show_cpu = !summaries_.empty() && summaries_[0].cpu != -1;
142 if (csv_) {
143 ShowCSV(fp, show_thread, show_cpu);
144 } else {
145 ShowText(fp, show_thread, show_cpu);
146 }
147 }
148
ShowCSV(FILE * fp,bool show_thread,bool show_cpu)149 void CounterSummaries::ShowCSV(FILE* fp, bool show_thread, bool show_cpu) {
150 for (auto& s : summaries_) {
151 if (show_thread) {
152 fprintf(fp, "%s,%d,%d,", s.thread->name.c_str(), s.thread->pid, s.thread->tid);
153 }
154 if (show_cpu) {
155 fprintf(fp, "%d,", s.cpu);
156 }
157 fprintf(fp, "%s,%s,%s,%s\n", s.readable_count.c_str(), s.Name().c_str(), s.comment.c_str(),
158 (s.auto_generated ? "(generated)," : ""));
159 }
160 }
161
ShowText(FILE * fp,bool show_thread,bool show_cpu)162 void CounterSummaries::ShowText(FILE* fp, bool show_thread, bool show_cpu) {
163 std::vector<std::string> titles;
164
165 if (show_thread) {
166 titles = {"thread_name", "pid", "tid"};
167 }
168 if (show_cpu) {
169 titles.emplace_back("cpu");
170 }
171 titles.emplace_back("count");
172 titles.emplace_back("event_name");
173 titles.emplace_back(" # count / runtime");
174
175 std::vector<size_t> width(titles.size(), 0);
176
177 auto adjust_width = [](size_t& w, size_t size) { w = std::max(w, size); };
178
179 // The last title is too long. Don't include it for width adjustment.
180 for (size_t i = 0; i + 1 < titles.size(); i++) {
181 adjust_width(width[i], titles[i].size());
182 }
183
184 for (auto& s : summaries_) {
185 size_t i = 0;
186 if (show_thread) {
187 adjust_width(width[i++], s.thread->name.size());
188 adjust_width(width[i++], std::to_string(s.thread->pid).size());
189 adjust_width(width[i++], std::to_string(s.thread->tid).size());
190 }
191 if (show_cpu) {
192 adjust_width(width[i++], std::to_string(s.cpu).size());
193 }
194 adjust_width(width[i++], s.readable_count.size());
195 adjust_width(width[i++], s.Name().size());
196 adjust_width(width[i++], s.comment.size());
197 }
198
199 fprintf(fp, "# ");
200 for (size_t i = 0; i < titles.size(); i++) {
201 if (titles[i] == "count") {
202 fprintf(fp, "%*s", static_cast<int>(width[i]), titles[i].c_str());
203 } else {
204 fprintf(fp, "%-*s", static_cast<int>(width[i]), titles[i].c_str());
205 }
206 if (i + 1 < titles.size()) {
207 fprintf(fp, " ");
208 }
209 }
210 fprintf(fp, "\n");
211
212 for (auto& s : summaries_) {
213 size_t i = 0;
214 if (show_thread) {
215 fprintf(fp, " %-*s", static_cast<int>(width[i++]), s.thread->name.c_str());
216 fprintf(fp, " %-*d", static_cast<int>(width[i++]), s.thread->pid);
217 fprintf(fp, " %-*d", static_cast<int>(width[i++]), s.thread->tid);
218 }
219 if (show_cpu) {
220 fprintf(fp, " %-*d", static_cast<int>(width[i++]), s.cpu);
221 }
222 fprintf(fp, " %*s %-*s # %-*s%s\n", static_cast<int>(width[i]), s.readable_count.c_str(),
223 static_cast<int>(width[i + 1]), s.Name().c_str(), static_cast<int>(width[i + 2]),
224 s.comment.c_str(), (s.auto_generated ? " (generated)" : ""));
225 }
226 }
227
GetCommentForSummary(const CounterSummary & s,double duration_in_sec)228 std::string CounterSummaries::GetCommentForSummary(const CounterSummary& s,
229 double duration_in_sec) {
230 char sap_mid;
231 if (csv_) {
232 sap_mid = ',';
233 } else {
234 sap_mid = ' ';
235 }
236 if (s.type_name == "task-clock") {
237 double run_sec = s.count / 1e9;
238 double used_cpus = run_sec / duration_in_sec;
239 return android::base::StringPrintf("%f%ccpus used", used_cpus, sap_mid);
240 }
241 if (s.type_name == "cpu-clock") {
242 return "";
243 }
244 if (s.type_name == "cpu-cycles") {
245 if (s.runtime_in_ns == 0) {
246 return "";
247 }
248 double ghz = static_cast<double>(s.count) / s.runtime_in_ns;
249 return android::base::StringPrintf("%f%cGHz", ghz, sap_mid);
250 }
251 if (s.type_name == "instructions" && s.count != 0) {
252 const CounterSummary* other = FindSummary("cpu-cycles", s.modifier, s.thread, s.cpu);
253 if (other != nullptr && other->IsMonitoredAtTheSameTime(s)) {
254 double cpi = static_cast<double>(other->count) / s.count;
255 return android::base::StringPrintf("%f%ccycles per instruction", cpi, sap_mid);
256 }
257 }
258 std::string rate_comment = GetRateComment(s, sap_mid);
259 if (!rate_comment.empty()) {
260 return rate_comment;
261 }
262 if (s.runtime_in_ns == 0) {
263 return "";
264 }
265 double runtime_in_sec = static_cast<double>(s.runtime_in_ns) / 1e9;
266 double rate = s.count / runtime_in_sec;
267 if (rate >= 1e9 - 1e5) {
268 return android::base::StringPrintf("%.3f%cG/sec", rate / 1e9, sap_mid);
269 }
270 if (rate >= 1e6 - 1e2) {
271 return android::base::StringPrintf("%.3f%cM/sec", rate / 1e6, sap_mid);
272 }
273 if (rate >= 1e3) {
274 return android::base::StringPrintf("%.3f%cK/sec", rate / 1e3, sap_mid);
275 }
276 return android::base::StringPrintf("%.3f%c/sec", rate, sap_mid);
277 }
278
GetRateComment(const CounterSummary & s,char sep)279 std::string CounterSummaries::GetRateComment(const CounterSummary& s, char sep) {
280 std::string_view miss_event_name = s.type_name;
281 std::string event_name;
282 std::string rate_desc;
283 if (auto it = COMMON_EVENT_RATE_MAP.find(miss_event_name); it != COMMON_EVENT_RATE_MAP.end()) {
284 event_name = it->second.first;
285 rate_desc = it->second.second;
286 }
287 if (event_name.empty() && (GetTargetArch() == ARCH_ARM || GetTargetArch() == ARCH_ARM64)) {
288 if (auto it = ARM_EVENT_RATE_MAP.find(miss_event_name); it != ARM_EVENT_RATE_MAP.end()) {
289 event_name = it->second.first;
290 rate_desc = it->second.second;
291 }
292 }
293 if (event_name.empty() && android::base::ConsumeSuffix(&miss_event_name, "-misses")) {
294 event_name = std::string(miss_event_name) + "s";
295 rate_desc = "miss rate";
296 }
297 if (!event_name.empty()) {
298 const CounterSummary* other = FindSummary(event_name, s.modifier, s.thread, s.cpu);
299 if (other != nullptr && other->IsMonitoredAtTheSameTime(s) && other->count != 0) {
300 double miss_rate = static_cast<double>(s.count) / other->count;
301 return android::base::StringPrintf("%f%%%c%s", miss_rate * 100, sep, rate_desc.c_str());
302 }
303 }
304 return "";
305 }
306
307 namespace {
308
309 // devfreq may use performance counters to calculate memory latency (as in
310 // drivers/devfreq/arm-memlat-mon.c). Hopefully we can get more available counters by asking devfreq
311 // to not use the memory latency governor temporarily.
312 class DevfreqCounters {
313 public:
Use()314 bool Use() {
315 if (!IsRoot()) {
316 LOG(ERROR) << "--use-devfreq-counters needs root permission to set devfreq governors";
317 return false;
318 }
319 std::string devfreq_dir = "/sys/class/devfreq/";
320 for (auto& name : GetSubDirs(devfreq_dir)) {
321 std::string governor_path = devfreq_dir + name + "/governor";
322 if (IsRegularFile(governor_path)) {
323 std::string governor;
324 if (!android::base::ReadFileToString(governor_path, &governor)) {
325 LOG(ERROR) << "failed to read " << governor_path;
326 return false;
327 }
328 governor = android::base::Trim(governor);
329 if (governor == "mem_latency") {
330 if (!android::base::WriteStringToFile("performance", governor_path)) {
331 PLOG(ERROR) << "failed to write " << governor_path;
332 return false;
333 }
334 mem_latency_governor_paths_.emplace_back(std::move(governor_path));
335 }
336 }
337 }
338 return true;
339 }
340
~DevfreqCounters()341 ~DevfreqCounters() {
342 for (auto& path : mem_latency_governor_paths_) {
343 android::base::WriteStringToFile("mem_latency", path);
344 }
345 }
346
347 private:
348 std::vector<std::string> mem_latency_governor_paths_;
349 };
350
351 // Periodically scan /proc for new threads. If found, create new perf event files for the
352 // new threads.
353 class NewThreadMonitor {
354 private:
355 const int SCAN_INTERVAL_US = 1;
356
357 public:
NewThreadMonitor(EventSelectionSet & event_selection_set,bool monitor_all_processes,const std::set<pid_t> & monitored_processes,std::unordered_map<pid_t,ThreadInfo> & threads)358 NewThreadMonitor(EventSelectionSet& event_selection_set, bool monitor_all_processes,
359 const std::set<pid_t>& monitored_processes,
360 std::unordered_map<pid_t, ThreadInfo>& threads)
361 : event_selection_set_(event_selection_set),
362 monitor_all_processes_(monitor_all_processes),
363 monitored_processes_(monitored_processes),
364 threads_(threads) {}
365
Start()366 bool Start() {
367 IOEventLoop* loop = event_selection_set_.GetIOEventLoop();
368 timeval tv;
369 tv.tv_sec = 0;
370 tv.tv_usec = SCAN_INTERVAL_US;
371 if (!loop->AddPeriodicEvent(tv, std::bind(&NewThreadMonitor::Scan, this))) {
372 return false;
373 }
374 // Ensure perf event files opened for new threads are immediately enabled.
375 event_selection_set_.SetEnableCondition(true, false);
376 return true;
377 }
378
379 private:
Scan()380 bool Scan() {
381 std::unordered_set<pid_t> new_tids;
382 if (monitor_all_processes_) {
383 for (int pid : GetAllProcesses()) {
384 for (auto tid : GetThreadsInProcess(pid)) {
385 if (threads_.count(tid) == 0) {
386 new_tids.insert(tid);
387 }
388 }
389 }
390 } else {
391 for (auto tid : monitored_processes_) {
392 for (auto tid : GetThreadsInProcess(tid)) {
393 if (threads_.count(tid) == 0) {
394 new_tids.insert(tid);
395 }
396 }
397 }
398 }
399 std::set<pid_t> open_event_file_tids;
400 for (auto tid : new_tids) {
401 ThreadInfo info;
402 if (ReadThreadNameAndPid(tid, &info.name, &info.pid)) {
403 info.tid = tid;
404 threads_[tid] = std::move(info);
405 open_event_file_tids.insert(tid);
406 }
407 }
408 if (!open_event_file_tids.empty()) {
409 // It's okay for OpenEventFilesForThreads() to return false. It happens
410 // when the new threads exit before we can open event files for them.
411 event_selection_set_.OpenEventFilesForThreads(open_event_file_tids);
412 }
413 return true;
414 }
415
416 private:
417 EventSelectionSet& event_selection_set_;
418 bool monitor_all_processes_ = false;
419 std::set<pid_t> monitored_processes_;
420 std::unordered_map<pid_t, ThreadInfo>& threads_;
421 };
422
423 class StatCommand : public Command {
424 public:
StatCommand()425 StatCommand()
426 : Command(
427 "stat", "gather performance counter information",
428 // clang-format off
429 "Usage: simpleperf stat [options] [command [command-args]]\n"
430 " Gather performance counter information of running [command].\n"
431 " And -a/-p/-t option can be used to change target of counter information.\n"
432 "-a Collect system-wide information.\n"
433 #if defined(__ANDROID__)
434 "--app package_name Profile the process of an Android application.\n"
435 " On non-rooted devices, the app must be debuggable,\n"
436 " because we use run-as to switch to the app's context.\n"
437 #endif
438 "--cpu cpu_item1,cpu_item2,... Monitor events on selected cpus. cpu_item can be a number like\n"
439 " 1, or a range like 0-3. A --cpu option affects all event types\n"
440 " following it until meeting another --cpu option.\n"
441 "--csv Write report in comma separate form.\n"
442 "--duration time_in_sec Monitor for time_in_sec seconds instead of running\n"
443 " [command]. Here time_in_sec may be any positive\n"
444 " floating point number.\n"
445 "--interval time_in_ms Print stat for every time_in_ms milliseconds.\n"
446 " Here time_in_ms may be any positive floating point\n"
447 " number. Simpleperf prints total values from the\n"
448 " starting point. But this can be changed by\n"
449 " --interval-only-values.\n"
450 "--interval-only-values Print numbers of events happened in each interval.\n"
451 "-e event1[:modifier1],event2[:modifier2],...\n"
452 " Select a list of events to count. An event can be:\n"
453 " 1) an event name listed in `simpleperf list`;\n"
454 " 2) a raw PMU event in rN format. N is a hex number.\n"
455 " For example, r1b selects event number 0x1b.\n"
456 " Modifiers can be added to define how the event should be\n"
457 " monitored. Possible modifiers are:\n"
458 " u - monitor user space events only\n"
459 " k - monitor kernel space events only\n"
460 "--group event1[:modifier],event2[:modifier2],...\n"
461 " Similar to -e option. But events specified in the same --group\n"
462 " option are monitored as a group, and scheduled in and out at the\n"
463 " same time.\n"
464 "--kprobe kprobe_event1,kprobe_event2,...\n"
465 " Add kprobe events during stating. The kprobe_event format is in\n"
466 " Documentation/trace/kprobetrace.rst in the kernel. Examples:\n"
467 " 'p:myprobe do_sys_openat2 $arg2:string' - add event kprobes:myprobe\n"
468 " 'r:myretprobe do_sys_openat2 $retval:s64' - add event kprobes:myretprobe\n"
469 "--uprobe uprobe_event1,uprobe_event2,...\n"
470 " Add uprobe events during stating. The uprobe_event format is in\n"
471 " Documentation/trace/uprobetracer.rst in the kernel. Examples:\n"
472 " 'p:myprobe /system/lib64/libc.so:0x1000'\n"
473 " - add event uprobes:myprobe\n"
474 " 'r:myretprobe /system/lib64/libc.so:0x1000'\n"
475 " - add event uprobes:myretprobe\n"
476 "--no-inherit Don't stat created child threads/processes.\n"
477 "-o output_filename Write report to output_filename instead of standard output.\n"
478 "--per-core Print counters for each cpu core.\n"
479 "--per-thread Print counters for each thread.\n"
480 "--monitor-new-thread Print counters for new threads created after stating. It should be used\n"
481 " With --per-thread and --no-inherit.\n"
482 "-p pid_or_process_name_regex1,pid_or_process_name_regex2,...\n"
483 " Stat events on existing processes. Processes are searched either by pid\n"
484 " or process name regex. Mutually exclusive with -a.\n"
485 "-t tid1,tid2,... Stat events on existing threads. Mutually exclusive with -a.\n"
486 "--tp-filter filter_string Set filter_string for the previous tracepoint event.\n"
487 " Format is in Documentation/trace/events.rst in the kernel.\n"
488 " An example: 'prev_comm != \"simpleperf\" && (prev_pid > 1)'.\n"
489 "--print-hw-counter Test and print CPU PMU hardware counters available on the device.\n"
490 "--sort key1,key2,... Select keys used to sort the report, used when --per-thread\n"
491 " or --per-core appears. The appearance order of keys decides\n"
492 " the order of keys used to sort the report.\n"
493 " Possible keys include:\n"
494 " count -- event count for each entry\n"
495 " count_per_thread -- event count for a thread on all cpus\n"
496 " cpu -- cpu id\n"
497 " pid -- process id\n"
498 " tid -- thread id\n"
499 " comm -- thread name\n"
500 " The default sort keys are:\n"
501 " count_per_thread,tid,cpu,count\n"
502 #if defined(__ANDROID__)
503 "--use-devfreq-counters On devices with Qualcomm SOCs, some hardware counters may be used\n"
504 " to monitor memory latency (in drivers/devfreq/arm-memlat-mon.c),\n"
505 " making fewer counters available to users. This option asks devfreq\n"
506 " to temporarily release counters by replacing memory-latency governor\n"
507 " with performance governor. It affects memory latency during profiling,\n"
508 " and may cause wedged power if simpleperf is killed in between.\n"
509 #endif
510 "--verbose Show result in verbose mode.\n"
511 #if 0
512 // Below options are only used internally and shouldn't be visible to the public.
513 "--in-app We are already running in the app's context.\n"
514 "--tracepoint-events file_name Read tracepoint events from [file_name] instead of tracefs.\n"
515 "--out-fd <fd> Write output to a file descriptor.\n"
516 "--stop-signal-fd <fd> Stop stating when fd is readable.\n"
517 #endif
518 // clang-format on
519 ),
520 verbose_mode_(false),
521 system_wide_collection_(false),
522 child_inherit_(true),
523 duration_in_sec_(0),
524 interval_in_ms_(0),
525 interval_only_values_(false),
526 event_selection_set_(true),
527 csv_(false),
528 in_app_context_(false) {
529 // Die if parent exits.
530 prctl(PR_SET_PDEATHSIG, SIGHUP, 0, 0, 0);
531 // Set default sort keys. Full key list is in BuildSummaryComparator().
532 sort_keys_ = {"count_per_thread", "tid", "cpu", "count"};
533 }
534
535 bool Run(const std::vector<std::string>& args);
536
537 private:
538 bool ParseOptions(const std::vector<std::string>& args, std::vector<std::string>* non_option_args,
539 ProbeEvents& probe_events);
540 void PrintHardwareCounters();
541 bool AddDefaultMeasuredEventTypes();
542 void SetEventSelectionFlags();
543 void MonitorEachThread(std::unique_ptr<Workload>& workload);
544 void AdjustToIntervalOnlyValues(std::vector<CountersInfo>& counters);
545 bool ShowCounters(const std::vector<CountersInfo>& counters, double duration_in_sec, FILE* fp);
546 void CheckHardwareCounterMultiplexing();
547 void PrintWarningForInaccurateEvents();
548
549 bool verbose_mode_;
550 bool system_wide_collection_;
551 bool child_inherit_;
552 double duration_in_sec_;
553 double interval_in_ms_;
554 bool interval_only_values_;
555 std::vector<std::vector<CounterSum>> last_sum_values_;
556 EventSelectionSet event_selection_set_;
557 std::string output_filename_;
558 android::base::unique_fd out_fd_;
559 bool csv_;
560 std::string app_package_name_;
561 bool in_app_context_;
562 android::base::unique_fd stop_signal_fd_;
563 bool use_devfreq_counters_ = false;
564
565 bool report_per_core_ = false;
566 bool report_per_thread_ = false;
567 bool monitor_new_thread_ = false;
568 // used to report event count for each thread
569 std::unordered_map<pid_t, ThreadInfo> thread_info_;
570 // used to sort report
571 std::vector<std::string> sort_keys_;
572 std::optional<SummaryComparator> summary_comparator_;
573 bool print_hw_counter_ = false;
574 };
575
Run(const std::vector<std::string> & args)576 bool StatCommand::Run(const std::vector<std::string>& args) {
577 if (!CheckPerfEventLimit()) {
578 return false;
579 }
580 AllowMoreOpenedFiles();
581
582 // 1. Parse options, and use default measured event types if not given.
583 std::vector<std::string> workload_args;
584 ProbeEvents probe_events(event_selection_set_);
585 if (!ParseOptions(args, &workload_args, probe_events)) {
586 return false;
587 }
588 if (print_hw_counter_) {
589 PrintHardwareCounters();
590 return true;
591 }
592 if (!app_package_name_.empty() && !in_app_context_) {
593 if (!IsRoot()) {
594 return RunInAppContext(app_package_name_, "stat", args, workload_args.size(),
595 output_filename_, !event_selection_set_.GetTracepointEvents().empty());
596 }
597 }
598 DevfreqCounters devfreq_counters;
599 if (use_devfreq_counters_) {
600 if (!devfreq_counters.Use()) {
601 return false;
602 }
603 }
604 if (event_selection_set_.empty()) {
605 if (!AddDefaultMeasuredEventTypes()) {
606 return false;
607 }
608 }
609 SetEventSelectionFlags();
610
611 // 2. Create workload.
612 std::unique_ptr<Workload> workload;
613 if (!workload_args.empty()) {
614 workload = Workload::CreateWorkload(workload_args);
615 if (workload == nullptr) {
616 return false;
617 }
618 }
619 bool need_to_check_targets = false;
620 if (system_wide_collection_) {
621 if (report_per_thread_) {
622 event_selection_set_.AddMonitoredProcesses(GetAllProcesses());
623 } else {
624 event_selection_set_.AddMonitoredThreads({-1});
625 }
626 } else if (!event_selection_set_.HasMonitoredTarget()) {
627 if (workload != nullptr) {
628 event_selection_set_.AddMonitoredProcesses({workload->GetPid()});
629 event_selection_set_.SetEnableCondition(false, true);
630 } else if (!app_package_name_.empty()) {
631 std::set<pid_t> pids = WaitForAppProcesses(app_package_name_);
632 event_selection_set_.AddMonitoredProcesses(pids);
633 } else {
634 LOG(ERROR) << "No threads to monitor. Try `simpleperf help stat` for help\n";
635 return false;
636 }
637 } else {
638 need_to_check_targets = true;
639 }
640 std::unique_ptr<NewThreadMonitor> new_thread_monitor;
641 if (monitor_new_thread_) {
642 new_thread_monitor.reset(new NewThreadMonitor(event_selection_set_, system_wide_collection_,
643 event_selection_set_.GetMonitoredProcesses(),
644 thread_info_));
645 }
646 if (report_per_thread_) {
647 MonitorEachThread(workload);
648 }
649
650 // 3. Open perf_event_files and output file if defined.
651 if (!event_selection_set_.OpenEventFiles()) {
652 return false;
653 }
654 std::unique_ptr<FILE, decltype(&fclose)> fp_holder(nullptr, fclose);
655 if (!output_filename_.empty()) {
656 fp_holder.reset(fopen(output_filename_.c_str(), "we"));
657 if (fp_holder == nullptr) {
658 PLOG(ERROR) << "failed to open " << output_filename_;
659 return false;
660 }
661 } else if (out_fd_ != -1) {
662 fp_holder.reset(fdopen(out_fd_.release(), "we"));
663 if (fp_holder == nullptr) {
664 PLOG(ERROR) << "failed to write output.";
665 return false;
666 }
667 }
668 FILE* fp = fp_holder ? fp_holder.get() : stdout;
669
670 // 4. Add signal/periodic Events.
671 IOEventLoop* loop = event_selection_set_.GetIOEventLoop();
672 std::chrono::time_point<std::chrono::steady_clock> start_time;
673 std::vector<CountersInfo> counters;
674 if (need_to_check_targets && !event_selection_set_.StopWhenNoMoreTargets()) {
675 return false;
676 }
677 auto exit_loop_callback = [loop]() { return loop->ExitLoop(); };
678 if (!loop->AddSignalEvents({SIGCHLD, SIGINT, SIGTERM, SIGHUP}, exit_loop_callback)) {
679 return false;
680 }
681 if (stop_signal_fd_ != -1) {
682 if (!loop->AddReadEvent(stop_signal_fd_, exit_loop_callback)) {
683 return false;
684 }
685 }
686 if (duration_in_sec_ != 0) {
687 if (!loop->AddPeriodicEvent(SecondToTimeval(duration_in_sec_), exit_loop_callback)) {
688 return false;
689 }
690 }
691 auto print_counters = [&]() {
692 auto end_time = std::chrono::steady_clock::now();
693 if (!event_selection_set_.ReadCounters(&counters)) {
694 return false;
695 }
696 double duration_in_sec =
697 std::chrono::duration_cast<std::chrono::duration<double>>(end_time - start_time).count();
698 if (interval_only_values_) {
699 AdjustToIntervalOnlyValues(counters);
700 }
701 if (!ShowCounters(counters, duration_in_sec, fp)) {
702 return false;
703 }
704 return true;
705 };
706
707 if (interval_in_ms_ != 0) {
708 if (!loop->AddPeriodicEvent(SecondToTimeval(interval_in_ms_ / 1000.0), print_counters)) {
709 return false;
710 }
711 }
712 if (new_thread_monitor && !new_thread_monitor->Start()) {
713 return false;
714 }
715
716 // 5. Count events while workload running.
717 start_time = std::chrono::steady_clock::now();
718 if (workload != nullptr && !workload->Start()) {
719 return false;
720 }
721 if (!loop->RunLoop()) {
722 return false;
723 }
724
725 // 6. Read and print counters.
726 if (interval_in_ms_ == 0) {
727 if (!print_counters()) {
728 return false;
729 }
730 }
731
732 // 7. Print warnings when needed.
733 event_selection_set_.CloseEventFiles();
734 CheckHardwareCounterMultiplexing();
735 PrintWarningForInaccurateEvents();
736
737 return true;
738 }
739
ParseOptions(const std::vector<std::string> & args,std::vector<std::string> * non_option_args,ProbeEvents & probe_events)740 bool StatCommand::ParseOptions(const std::vector<std::string>& args,
741 std::vector<std::string>* non_option_args,
742 ProbeEvents& probe_events) {
743 OptionValueMap options;
744 std::vector<std::pair<OptionName, OptionValue>> ordered_options;
745
746 if (!PreprocessOptions(args, GetStatCmdOptionFormats(), &options, &ordered_options,
747 non_option_args)) {
748 return false;
749 }
750
751 // Process options.
752 system_wide_collection_ = options.PullBoolValue("-a");
753
754 if (auto value = options.PullValue("--app"); value) {
755 app_package_name_ = value->str_value;
756 }
757 csv_ = options.PullBoolValue("--csv");
758
759 if (!options.PullDoubleValue("--duration", &duration_in_sec_, 1e-9)) {
760 return false;
761 }
762 if (!options.PullDoubleValue("--interval", &interval_in_ms_, 1e-9)) {
763 return false;
764 }
765 interval_only_values_ = options.PullBoolValue("--interval-only-values");
766
767 in_app_context_ = options.PullBoolValue("--in-app");
768 for (const OptionValue& value : options.PullValues("--kprobe")) {
769 for (const auto& cmd : Split(value.str_value, ",")) {
770 if (!probe_events.AddProbe(ProbeEventType::kKprobe, cmd)) {
771 return false;
772 }
773 }
774 }
775 for (const OptionValue& value : options.PullValues("--uprobe")) {
776 for (const auto& cmd : Split(value.str_value, ",")) {
777 if (!probe_events.AddProbe(ProbeEventType::kUprobe, cmd)) {
778 return false;
779 }
780 }
781 }
782 monitor_new_thread_ = options.PullBoolValue("--monitor-new-thread");
783 child_inherit_ = !options.PullBoolValue("--no-inherit");
784
785 if (auto value = options.PullValue("-o"); value) {
786 output_filename_ = value->str_value;
787 }
788 if (auto value = options.PullValue("--out-fd"); value) {
789 out_fd_.reset(static_cast<int>(value->uint_value));
790 }
791
792 report_per_core_ = options.PullBoolValue("--per-core");
793 report_per_thread_ = options.PullBoolValue("--per-thread");
794
795 if (auto strs = options.PullStringValues("-p"); !strs.empty()) {
796 if (auto pids = GetPidsFromStrings(strs, true, true); pids) {
797 event_selection_set_.AddMonitoredProcesses(pids.value());
798 } else {
799 return false;
800 }
801 }
802 print_hw_counter_ = options.PullBoolValue("--print-hw-counter");
803
804 if (auto value = options.PullValue("--sort"); value) {
805 sort_keys_ = Split(value->str_value, ",");
806 }
807
808 if (auto value = options.PullValue("--stop-signal-fd"); value) {
809 stop_signal_fd_.reset(static_cast<int>(value->uint_value));
810 }
811
812 for (const OptionValue& value : options.PullValues("-t")) {
813 if (auto tids = GetTidsFromString(value.str_value, true); tids) {
814 event_selection_set_.AddMonitoredThreads(tids.value());
815 } else {
816 return false;
817 }
818 }
819
820 if (auto value = options.PullValue("--tracepoint-events"); value) {
821 if (!EventTypeManager::Instance().ReadTracepointsFromFile(value->str_value)) {
822 return false;
823 }
824 }
825
826 use_devfreq_counters_ = options.PullBoolValue("--use-devfreq-counters");
827 verbose_mode_ = options.PullBoolValue("--verbose");
828
829 CHECK(options.values.empty());
830
831 bool check_event_type = true;
832 if (!app_package_name_.empty() && !in_app_context_ && !IsRoot()) {
833 // Defer event type checking when RunInAppContext() is called.
834 check_event_type = false;
835 }
836
837 // Process ordered options.
838 for (const auto& pair : ordered_options) {
839 const OptionName& name = pair.first;
840 const OptionValue& value = pair.second;
841
842 if (name == "--cpu") {
843 if (auto v = GetCpusFromString(value.str_value); v) {
844 std::set<int>& cpus = v.value();
845 event_selection_set_.SetCpusForNewEvents(std::vector<int>(cpus.begin(), cpus.end()));
846 } else {
847 return false;
848 }
849 } else if (name == "-e") {
850 for (const auto& event_type : Split(value.str_value, ",")) {
851 if (!probe_events.CreateProbeEventIfNotExist(event_type)) {
852 return false;
853 }
854 if (!event_selection_set_.AddEventType(event_type, check_event_type)) {
855 return false;
856 }
857 }
858 } else if (name == "--group") {
859 std::vector<std::string> event_types = Split(value.str_value, ",");
860 for (const auto& event_type : event_types) {
861 if (!probe_events.CreateProbeEventIfNotExist(event_type)) {
862 return false;
863 }
864 }
865 if (!event_selection_set_.AddEventGroup(event_types, check_event_type)) {
866 return false;
867 }
868 } else if (name == "--tp-filter") {
869 if (!event_selection_set_.SetTracepointFilter(value.str_value)) {
870 return false;
871 }
872 } else {
873 LOG(ERROR) << "unprocessed option: " << name;
874 return false;
875 }
876 }
877
878 if (system_wide_collection_ && event_selection_set_.HasMonitoredTarget()) {
879 LOG(ERROR) << "Stat system wide and existing processes/threads can't be "
880 "used at the same time.";
881 return false;
882 }
883 if (system_wide_collection_ && !IsRoot()) {
884 LOG(ERROR) << "System wide profiling needs root privilege.";
885 return false;
886 }
887 if (monitor_new_thread_) {
888 if (!report_per_thread_ || child_inherit_) {
889 LOG(ERROR) << "--monitor-new-thread should be used with --per-thread and --no-inherit";
890 return false;
891 }
892 }
893
894 if (report_per_core_ || report_per_thread_) {
895 summary_comparator_ = BuildSummaryComparator(sort_keys_, report_per_thread_, report_per_core_);
896 if (!summary_comparator_) {
897 return false;
898 }
899 }
900 return true;
901 }
902
CheckHardwareCountersOnCpu(int cpu,size_t counters)903 std::optional<bool> CheckHardwareCountersOnCpu(int cpu, size_t counters) {
904 if (counters == 0) {
905 return true;
906 }
907 const EventType* event = FindEventTypeByName("cpu-cycles", true);
908 if (event == nullptr) {
909 return std::nullopt;
910 }
911 perf_event_attr attr = CreateDefaultPerfEventAttr(*event);
912 attr.exclude_kernel = true;
913 auto workload = Workload::CreateWorkload({"sleep", "0.1"});
914 if (!workload || !workload->SetCpuAffinity(cpu)) {
915 return std::nullopt;
916 }
917 std::vector<std::unique_ptr<EventFd>> event_fds;
918 for (size_t i = 0; i < counters; i++) {
919 EventFd* group_event_fd = event_fds.empty() ? nullptr : event_fds[0].get();
920 auto event_fd =
921 EventFd::OpenEventFile(attr, workload->GetPid(), cpu, group_event_fd, "cpu-cycles", false);
922 if (!event_fd) {
923 return false;
924 }
925 event_fds.emplace_back(std::move(event_fd));
926 }
927 if (!workload->Start() || !workload->WaitChildProcess(true, nullptr)) {
928 return std::nullopt;
929 }
930 for (auto& event_fd : event_fds) {
931 PerfCounter counter;
932 if (!event_fd->ReadCounter(&counter)) {
933 return std::nullopt;
934 }
935 if (counter.time_enabled == 0 || counter.time_enabled > counter.time_running) {
936 return false;
937 }
938 }
939 return true;
940 }
941
GetHardwareCountersOnCpu(int cpu)942 std::optional<size_t> GetHardwareCountersOnCpu(int cpu) {
943 size_t available_counters = 0;
944 while (true) {
945 std::optional<bool> result = CheckHardwareCountersOnCpu(cpu, available_counters + 1);
946 if (!result.has_value()) {
947 return std::nullopt;
948 }
949 if (!result.value()) {
950 break;
951 }
952 available_counters++;
953 }
954 return available_counters;
955 }
956
PrintHardwareCounters()957 void StatCommand::PrintHardwareCounters() {
958 for (int cpu : GetOnlineCpus()) {
959 std::optional<size_t> counters = GetHardwareCountersOnCpu(cpu);
960 if (!counters) {
961 // When built as a 32-bit program, we can't set sched_affinity to a 64-bit only CPU. So we
962 // may not be able to get hardware counters on that CPU.
963 LOG(WARNING) << "Failed to get CPU PMU hardware counters on cpu " << cpu;
964 continue;
965 }
966 printf("There are %zu CPU PMU hardware counters available on cpu %d.\n", counters.value(), cpu);
967 }
968 }
969
AddDefaultMeasuredEventTypes()970 bool StatCommand::AddDefaultMeasuredEventTypes() {
971 for (std::string name : default_measured_event_types) {
972 // It is not an error when some event types in the default list are not
973 // supported by the kernel.
974 const EventType* type = FindEventTypeByName(name);
975 if (type == nullptr) {
976 continue;
977 }
978 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
979 if (!IsKernelEventSupported()) {
980 attr.exclude_kernel = true;
981 if (name == "cpu-clock" || name == "task-clock") {
982 continue;
983 }
984 name += ":u";
985 }
986 if (IsEventAttrSupported(attr, name)) {
987 if (!event_selection_set_.AddEventType(name)) {
988 return false;
989 }
990 }
991 }
992 if (event_selection_set_.empty()) {
993 LOG(ERROR) << "Failed to add any supported default measured types";
994 return false;
995 }
996 return true;
997 }
998
SetEventSelectionFlags()999 void StatCommand::SetEventSelectionFlags() {
1000 event_selection_set_.SetInherit(child_inherit_);
1001 }
1002
MonitorEachThread(std::unique_ptr<Workload> & workload)1003 void StatCommand::MonitorEachThread(std::unique_ptr<Workload>& workload) {
1004 std::vector<pid_t> threads;
1005 for (auto pid : event_selection_set_.GetMonitoredProcesses()) {
1006 for (auto tid : GetThreadsInProcess(pid)) {
1007 ThreadInfo info;
1008 if (GetThreadName(tid, &info.name)) {
1009 if (tid == pid && workload && workload->GetPid() == pid) {
1010 info.name = workload->GetCommandName();
1011 }
1012 info.tid = tid;
1013 info.pid = pid;
1014 thread_info_[tid] = std::move(info);
1015 threads.push_back(tid);
1016 }
1017 }
1018 }
1019 for (auto tid : event_selection_set_.GetMonitoredThreads()) {
1020 ThreadInfo info;
1021 if (ReadThreadNameAndPid(tid, &info.name, &info.pid)) {
1022 info.tid = tid;
1023 thread_info_[tid] = std::move(info);
1024 threads.push_back(tid);
1025 }
1026 }
1027 event_selection_set_.ClearMonitoredTargets();
1028 event_selection_set_.AddMonitoredThreads(threads);
1029 }
1030
AdjustToIntervalOnlyValues(std::vector<CountersInfo> & counters)1031 void StatCommand::AdjustToIntervalOnlyValues(std::vector<CountersInfo>& counters) {
1032 if (last_sum_values_.size() < counters.size()) {
1033 last_sum_values_.resize(counters.size());
1034 }
1035 for (size_t i = 0; i < counters.size(); i++) {
1036 std::vector<CounterInfo>& counters_per_event = counters[i].counters;
1037 std::vector<CounterSum>& last_sum = last_sum_values_[i];
1038
1039 if (last_sum.size() < counters_per_event.size()) {
1040 last_sum.resize(counters_per_event.size());
1041 }
1042 for (size_t j = 0; j < counters_per_event.size(); j++) {
1043 PerfCounter& counter = counters_per_event[j].counter;
1044 CounterSum new_sum;
1045 new_sum.FromCounter(counter);
1046 CounterSum delta = new_sum - last_sum[j];
1047 delta.ToCounter(counter);
1048 last_sum[j] = new_sum;
1049 }
1050 }
1051 }
1052
ShowCounters(const std::vector<CountersInfo> & counters,double duration_in_sec,FILE * fp)1053 bool StatCommand::ShowCounters(const std::vector<CountersInfo>& counters, double duration_in_sec,
1054 FILE* fp) {
1055 if (csv_) {
1056 fprintf(fp, "Performance counter statistics,\n");
1057 } else {
1058 fprintf(fp, "Performance counter statistics:\n\n");
1059 }
1060
1061 if (verbose_mode_) {
1062 for (auto& counters_info : counters) {
1063 for (auto& counter_info : counters_info.counters) {
1064 if (csv_) {
1065 fprintf(fp,
1066 "%s,tid,%d,cpu,%d,count,%" PRIu64 ",time_enabled,%" PRIu64
1067 ",time running,%" PRIu64 ",id,%" PRIu64 ",\n",
1068 counters_info.event_name.c_str(), counter_info.tid, counter_info.cpu,
1069 counter_info.counter.value, counter_info.counter.time_enabled,
1070 counter_info.counter.time_running, counter_info.counter.id);
1071 } else {
1072 fprintf(fp,
1073 "%s(tid %d, cpu %d): count %" PRIu64 ", time_enabled %" PRIu64
1074 ", time running %" PRIu64 ", id %" PRIu64 "\n",
1075 counters_info.event_name.c_str(), counter_info.tid, counter_info.cpu,
1076 counter_info.counter.value, counter_info.counter.time_enabled,
1077 counter_info.counter.time_running, counter_info.counter.id);
1078 }
1079 }
1080 }
1081 }
1082
1083 CounterSummaryBuilder builder(report_per_thread_, report_per_core_, csv_, thread_info_,
1084 summary_comparator_);
1085 for (const auto& info : counters) {
1086 builder.AddCountersForOneEventType(info);
1087 }
1088 CounterSummaries summaries(builder.Build(), csv_);
1089 summaries.AutoGenerateSummaries();
1090 summaries.GenerateComments(duration_in_sec);
1091 summaries.Show(fp);
1092
1093 if (csv_) {
1094 fprintf(fp, "Total test time,%lf,seconds,\n", duration_in_sec);
1095 } else {
1096 fprintf(fp, "\nTotal test time: %lf seconds.\n", duration_in_sec);
1097 }
1098 return true;
1099 }
1100
CheckHardwareCounterMultiplexing()1101 void StatCommand::CheckHardwareCounterMultiplexing() {
1102 for (const auto& [cpu, hardware_events] : event_selection_set_.GetHardwareCountersForCpus()) {
1103 std::optional<bool> result = CheckHardwareCountersOnCpu(cpu, hardware_events);
1104 if (result.has_value() && !result.value()) {
1105 LOG(WARNING) << "It seems the number of hardware events are more than the number of\n"
1106 << "available CPU PMU hardware counters. That will trigger hardware counter\n"
1107 << "multiplexing. As a result, events are not counted all the time processes\n"
1108 << "running, and event counts are smaller than what really happen.\n"
1109 << "Use --print-hw-counter to show available hardware counters.\n"
1110 #if defined(__ANDROID__)
1111 << "If on a rooted device, try --use-devfreq-counters to get more counters.\n"
1112 #endif
1113 ;
1114 break;
1115 }
1116 }
1117 }
1118
PrintWarningForInaccurateEvents()1119 void StatCommand::PrintWarningForInaccurateEvents() {
1120 for (const EventType* event : event_selection_set_.GetEvents()) {
1121 if (event->name == "raw-l3d-cache-lmiss-rd") {
1122 LOG(WARNING) << "PMU event L3D_CACHE_LMISS_RD might undercount on A510. Please use "
1123 "L3D_CACHE_REFILL_RD instead.";
1124 break;
1125 }
1126 }
1127 }
1128
1129 } // namespace
1130
RegisterStatCommand()1131 void RegisterStatCommand() {
1132 RegisterCommand("stat", [] { return std::unique_ptr<Command>(new StatCommand); });
1133 }
1134
1135 } // namespace simpleperf
1136