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 "event_selection_set.h"
18
19 #include <algorithm>
20 #include <atomic>
21 #include <thread>
22 #include <unordered_map>
23
24 #include <android-base/logging.h>
25 #include <android-base/stringprintf.h>
26 #include <android-base/strings.h>
27
28 #include "ETMRecorder.h"
29 #include "IOEventLoop.h"
30 #include "RecordReadThread.h"
31 #include "environment.h"
32 #include "event_attr.h"
33 #include "event_type.h"
34 #include "perf_regs.h"
35 #include "tracing.h"
36 #include "utils.h"
37
38 namespace simpleperf {
39
40 using android::base::StringPrintf;
41
IsBranchSamplingSupported()42 bool IsBranchSamplingSupported() {
43 const EventType* type = FindEventTypeByName("cpu-cycles");
44 if (type == nullptr) {
45 return false;
46 }
47 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
48 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
49 attr.branch_sample_type = PERF_SAMPLE_BRANCH_ANY;
50 return IsEventAttrSupported(attr, type->name);
51 }
52
IsDwarfCallChainSamplingSupported()53 bool IsDwarfCallChainSamplingSupported() {
54 if (auto version = GetKernelVersion(); version && version.value() >= std::make_pair(3, 18)) {
55 // Skip test on kernel >= 3.18, which has all patches needed to support dwarf callchain.
56 return true;
57 }
58 const EventType* type = FindEventTypeByName("cpu-clock");
59 if (type == nullptr) {
60 return false;
61 }
62 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
63 attr.sample_type |= PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER;
64 attr.exclude_callchain_user = 1;
65 attr.sample_regs_user = GetSupportedRegMask(GetTargetArch());
66 attr.sample_stack_user = 8192;
67 return IsEventAttrSupported(attr, type->name);
68 }
69
IsDumpingRegsForTracepointEventsSupported()70 bool IsDumpingRegsForTracepointEventsSupported() {
71 if (auto version = GetKernelVersion(); version && version.value() >= std::make_pair(4, 2)) {
72 // Kernel >= 4.2 has patch "5b09a094f2 arm64: perf: Fix callchain parse error with kernel
73 // tracepoint events". So no need to test.
74 return true;
75 }
76 const EventType* event_type = FindEventTypeByName("sched:sched_switch", false);
77 if (event_type == nullptr) {
78 return false;
79 }
80 std::atomic<bool> done(false);
81 std::atomic<pid_t> thread_id(0);
82 std::thread thread([&]() {
83 thread_id = gettid();
84 while (!done) {
85 usleep(1);
86 }
87 usleep(1); // Make a sched out to generate one sample.
88 });
89 while (thread_id == 0) {
90 usleep(1);
91 }
92 perf_event_attr attr = CreateDefaultPerfEventAttr(*event_type);
93 attr.freq = 0;
94 attr.sample_period = 1;
95 std::unique_ptr<EventFd> event_fd =
96 EventFd::OpenEventFile(attr, thread_id, -1, nullptr, event_type->name);
97 if (event_fd == nullptr || !event_fd->CreateMappedBuffer(4, true)) {
98 done = true;
99 thread.join();
100 return false;
101 }
102 done = true;
103 thread.join();
104
105 // There are small chances that we don't see samples immediately after joining the thread on
106 // cuttlefish, probably due to data synchronization between cpus. To avoid flaky tests, use a
107 // loop to wait for samples.
108 for (int timeout = 0; timeout < 1000; timeout++) {
109 std::vector<char> buffer = event_fd->GetAvailableMmapData();
110 std::vector<std::unique_ptr<Record>> records =
111 ReadRecordsFromBuffer(attr, buffer.data(), buffer.size());
112 for (auto& r : records) {
113 if (r->type() == PERF_RECORD_SAMPLE) {
114 auto& record = *static_cast<SampleRecord*>(r.get());
115 return record.ip_data.ip != 0;
116 }
117 }
118 usleep(1);
119 }
120 return false;
121 }
122
IsSettingClockIdSupported()123 bool IsSettingClockIdSupported() {
124 // Do the real check only once and keep the result in a static variable.
125 static int is_supported = -1;
126 if (is_supported == -1) {
127 is_supported = 0;
128 if (auto version = GetKernelVersion(); version && version.value() >= std::make_pair(4, 1)) {
129 // Kernel >= 4.1 has patch "34f43927 perf: Add per event clockid support". So no need to test.
130 is_supported = 1;
131 } else if (const EventType* type = FindEventTypeByName("cpu-clock"); type != nullptr) {
132 // Check if the kernel supports setting clockid, which was added in kernel 4.0. Just check
133 // with one clockid is enough. Because all needed clockids were supported before kernel 4.0.
134 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
135 attr.use_clockid = 1;
136 attr.clockid = CLOCK_MONOTONIC;
137 is_supported = IsEventAttrSupported(attr, type->name) ? 1 : 0;
138 }
139 }
140 return is_supported;
141 }
142
IsMmap2Supported()143 bool IsMmap2Supported() {
144 if (auto version = GetKernelVersion(); version && version.value() >= std::make_pair(3, 12)) {
145 // Kernel >= 3.12 has patch "13d7a2410 perf: Add attr->mmap2 attribute to an event". So no need
146 // to test.
147 return true;
148 }
149 const EventType* type = FindEventTypeByName("cpu-clock");
150 if (type == nullptr) {
151 return false;
152 }
153 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
154 attr.mmap2 = 1;
155 return IsEventAttrSupported(attr, type->name);
156 }
157
IsHardwareEventSupported()158 bool IsHardwareEventSupported() {
159 const EventType* type = FindEventTypeByName("cpu-cycles");
160 if (type == nullptr) {
161 return false;
162 }
163 perf_event_attr attr = CreateDefaultPerfEventAttr(*type);
164 return IsEventAttrSupported(attr, type->name);
165 }
166
IsSwitchRecordSupported()167 bool IsSwitchRecordSupported() {
168 // Kernel >= 4.3 has patch "45ac1403f perf: Add PERF_RECORD_SWITCH to indicate context switches".
169 auto version = GetKernelVersion();
170 return version && version.value() >= std::make_pair(4, 3);
171 }
172
ToString() const173 std::string AddrFilter::ToString() const {
174 switch (type) {
175 case FILE_RANGE:
176 return StringPrintf("filter 0x%" PRIx64 "/0x%" PRIx64 "@%s", addr, size, file_path.c_str());
177 case AddrFilter::FILE_START:
178 return StringPrintf("start 0x%" PRIx64 "@%s", addr, file_path.c_str());
179 case AddrFilter::FILE_STOP:
180 return StringPrintf("stop 0x%" PRIx64 "@%s", addr, file_path.c_str());
181 case AddrFilter::KERNEL_RANGE:
182 return StringPrintf("filter 0x%" PRIx64 "/0x%" PRIx64, addr, size);
183 case AddrFilter::KERNEL_START:
184 return StringPrintf("start 0x%" PRIx64, addr);
185 case AddrFilter::KERNEL_STOP:
186 return StringPrintf("stop 0x%" PRIx64, addr);
187 }
188 }
189
EventSelectionSet(bool for_stat_cmd)190 EventSelectionSet::EventSelectionSet(bool for_stat_cmd)
191 : for_stat_cmd_(for_stat_cmd), loop_(new IOEventLoop) {}
192
~EventSelectionSet()193 EventSelectionSet::~EventSelectionSet() {}
194
BuildAndCheckEventSelection(const std::string & event_name,bool first_event,EventSelection * selection)195 bool EventSelectionSet::BuildAndCheckEventSelection(const std::string& event_name, bool first_event,
196 EventSelection* selection) {
197 std::unique_ptr<EventTypeAndModifier> event_type = ParseEventType(event_name);
198 if (event_type == nullptr) {
199 return false;
200 }
201 if (for_stat_cmd_) {
202 if (event_type->event_type.name == "cpu-clock" || event_type->event_type.name == "task-clock") {
203 if (event_type->exclude_user || event_type->exclude_kernel) {
204 LOG(ERROR) << "Modifier u and modifier k used in event type " << event_type->event_type.name
205 << " are not supported by the kernel.";
206 return false;
207 }
208 }
209 }
210 selection->event_type_modifier = *event_type;
211 selection->event_attr = CreateDefaultPerfEventAttr(event_type->event_type);
212 selection->event_attr.exclude_user = event_type->exclude_user;
213 selection->event_attr.exclude_kernel = event_type->exclude_kernel;
214 selection->event_attr.exclude_hv = event_type->exclude_hv;
215 selection->event_attr.exclude_host = event_type->exclude_host;
216 selection->event_attr.exclude_guest = event_type->exclude_guest;
217 selection->event_attr.precise_ip = event_type->precise_ip;
218 if (IsEtmEventType(event_type->event_type.type)) {
219 auto& etm_recorder = ETMRecorder::GetInstance();
220 if (auto result = etm_recorder.CheckEtmSupport(); !result.ok()) {
221 LOG(ERROR) << result.error();
222 return false;
223 }
224 ETMRecorder::GetInstance().SetEtmPerfEventAttr(&selection->event_attr);
225 }
226 bool set_default_sample_freq = false;
227 if (!for_stat_cmd_) {
228 if (event_type->event_type.type == PERF_TYPE_TRACEPOINT) {
229 selection->event_attr.freq = 0;
230 selection->event_attr.sample_period = DEFAULT_SAMPLE_PERIOD_FOR_TRACEPOINT_EVENT;
231 } else if (IsEtmEventType(event_type->event_type.type)) {
232 // ETM recording has no sample frequency to adjust. Using sample frequency only wastes time
233 // enabling/disabling etm devices. So don't adjust frequency by default.
234 selection->event_attr.freq = 0;
235 selection->event_attr.sample_period = 1;
236 // An ETM event can't be enabled without mmap aux buffer. So disable it by default.
237 selection->event_attr.disabled = 1;
238 } else {
239 selection->event_attr.freq = 1;
240 // Set default sample freq here may print msg "Adjust sample freq to max allowed sample
241 // freq". But this is misleading. Because default sample freq may not be the final sample
242 // freq we use. So use minimum sample freq (1) here.
243 selection->event_attr.sample_freq = 1;
244 set_default_sample_freq = true;
245 }
246 // We only need to dump mmap and comm records for the first event type. Because all event types
247 // are monitoring the same processes.
248 if (first_event) {
249 selection->event_attr.mmap = 1;
250 selection->event_attr.comm = 1;
251 if (IsMmap2Supported()) {
252 selection->event_attr.mmap2 = 1;
253 }
254 }
255 }
256 // PMU events are provided by kernel, so they should be supported
257 if (!event_type->event_type.IsPmuEvent() &&
258 !IsEventAttrSupported(selection->event_attr, selection->event_type_modifier.name)) {
259 LOG(ERROR) << "Event type '" << event_type->name << "' is not supported on the device";
260 return false;
261 }
262 if (set_default_sample_freq) {
263 selection->event_attr.sample_freq = DEFAULT_SAMPLE_FREQ_FOR_NONTRACEPOINT_EVENT;
264 }
265
266 selection->event_fds.clear();
267
268 for (const auto& group : groups_) {
269 for (const auto& sel : group.selections) {
270 if (sel.event_type_modifier.name == selection->event_type_modifier.name) {
271 LOG(ERROR) << "Event type '" << sel.event_type_modifier.name << "' appears more than once";
272 return false;
273 }
274 }
275 }
276 return true;
277 }
278
AddEventType(const std::string & event_name)279 bool EventSelectionSet::AddEventType(const std::string& event_name) {
280 return AddEventGroup(std::vector<std::string>(1, event_name));
281 }
282
AddEventType(const std::string & event_name,const SampleRate & sample_rate)283 bool EventSelectionSet::AddEventType(const std::string& event_name, const SampleRate& sample_rate) {
284 if (!AddEventGroup(std::vector<std::string>(1, event_name))) {
285 return false;
286 }
287 SetSampleRateForGroup(groups_.back(), sample_rate);
288 return true;
289 }
290
AddEventGroup(const std::vector<std::string> & event_names)291 bool EventSelectionSet::AddEventGroup(const std::vector<std::string>& event_names) {
292 EventSelectionGroup group;
293 bool first_event = groups_.empty();
294 bool first_in_group = true;
295 for (const auto& event_name : event_names) {
296 EventSelection selection;
297 if (!BuildAndCheckEventSelection(event_name, first_event, &selection)) {
298 return false;
299 }
300 if (IsEtmEventType(selection.event_attr.type)) {
301 has_aux_trace_ = true;
302 }
303 if (first_in_group) {
304 auto& event_type = selection.event_type_modifier.event_type;
305 if (event_type.IsPmuEvent()) {
306 selection.allowed_cpus = event_type.GetPmuCpumask();
307 }
308 }
309 first_event = false;
310 first_in_group = false;
311 group.selections.emplace_back(std::move(selection));
312 }
313 if (sample_rate_) {
314 SetSampleRateForGroup(group, sample_rate_.value());
315 }
316 if (cpus_) {
317 group.cpus = cpus_.value();
318 }
319 groups_.emplace_back(std::move(group));
320 UnionSampleType();
321 return true;
322 }
323
AddCounters(const std::vector<std::string> & event_names)324 bool EventSelectionSet::AddCounters(const std::vector<std::string>& event_names) {
325 CHECK(!groups_.empty());
326 if (groups_.size() > 1) {
327 LOG(ERROR) << "Failed to add counters. Only one event group is allowed.";
328 return false;
329 }
330 for (const auto& event_name : event_names) {
331 EventSelection selection;
332 if (!BuildAndCheckEventSelection(event_name, false, &selection)) {
333 return false;
334 }
335 // Use a big sample_period to avoid getting samples for added counters.
336 selection.event_attr.freq = 0;
337 selection.event_attr.sample_period = INFINITE_SAMPLE_PERIOD;
338 selection.event_attr.inherit = 0;
339 groups_[0].selections.emplace_back(std::move(selection));
340 }
341 // Add counters in each sample.
342 for (auto& selection : groups_[0].selections) {
343 selection.event_attr.sample_type |= PERF_SAMPLE_READ;
344 selection.event_attr.read_format |= PERF_FORMAT_GROUP;
345 }
346 return true;
347 }
348
GetEvents() const349 std::vector<const EventType*> EventSelectionSet::GetEvents() const {
350 std::vector<const EventType*> result;
351 for (const auto& group : groups_) {
352 for (const auto& selection : group.selections) {
353 result.push_back(&selection.event_type_modifier.event_type);
354 }
355 }
356 return result;
357 }
358
GetTracepointEvents() const359 std::vector<const EventType*> EventSelectionSet::GetTracepointEvents() const {
360 std::vector<const EventType*> result;
361 for (const auto& group : groups_) {
362 for (const auto& selection : group.selections) {
363 if (selection.event_type_modifier.event_type.type == PERF_TYPE_TRACEPOINT) {
364 result.push_back(&selection.event_type_modifier.event_type);
365 }
366 }
367 }
368 return result;
369 }
370
ExcludeKernel() const371 bool EventSelectionSet::ExcludeKernel() const {
372 for (const auto& group : groups_) {
373 for (const auto& selection : group.selections) {
374 if (!selection.event_type_modifier.exclude_kernel) {
375 return false;
376 }
377 }
378 }
379 return true;
380 }
381
GetEventAttrWithId() const382 EventAttrIds EventSelectionSet::GetEventAttrWithId() const {
383 EventAttrIds result;
384 for (const auto& group : groups_) {
385 for (const auto& selection : group.selections) {
386 std::vector<uint64_t> ids;
387 for (const auto& fd : selection.event_fds) {
388 ids.push_back(fd->Id());
389 }
390 result.resize(result.size() + 1);
391 result.back().attr = selection.event_attr;
392 result.back().ids = std::move(ids);
393 }
394 }
395 return result;
396 }
397
GetEventNamesById() const398 std::unordered_map<uint64_t, std::string> EventSelectionSet::GetEventNamesById() const {
399 std::unordered_map<uint64_t, std::string> result;
400 for (const auto& group : groups_) {
401 for (const auto& selection : group.selections) {
402 for (const auto& fd : selection.event_fds) {
403 result[fd->Id()] = selection.event_type_modifier.name;
404 }
405 }
406 }
407 return result;
408 }
409
GetCpusById() const410 std::unordered_map<uint64_t, int> EventSelectionSet::GetCpusById() const {
411 std::unordered_map<uint64_t, int> result;
412 for (const auto& group : groups_) {
413 for (const auto& selection : group.selections) {
414 for (const auto& fd : selection.event_fds) {
415 result[fd->Id()] = fd->Cpu();
416 }
417 }
418 }
419 return result;
420 }
421
GetHardwareCountersForCpus() const422 std::map<int, size_t> EventSelectionSet::GetHardwareCountersForCpus() const {
423 std::map<int, size_t> cpu_map;
424 std::vector<int> online_cpus = GetOnlineCpus();
425
426 for (const auto& group : groups_) {
427 size_t hardware_events = 0;
428 for (const auto& selection : group.selections) {
429 if (selection.event_type_modifier.event_type.IsHardwareEvent()) {
430 hardware_events++;
431 }
432 }
433 const std::vector<int>* pcpus = group.cpus.empty() ? &online_cpus : &group.cpus;
434 for (int cpu : *pcpus) {
435 cpu_map[cpu] += hardware_events;
436 }
437 }
438 return cpu_map;
439 }
440
441 // Union the sample type of different event attrs can make reading sample
442 // records in perf.data easier.
UnionSampleType()443 void EventSelectionSet::UnionSampleType() {
444 uint64_t sample_type = 0;
445 for (const auto& group : groups_) {
446 for (const auto& selection : group.selections) {
447 sample_type |= selection.event_attr.sample_type;
448 }
449 }
450 for (auto& group : groups_) {
451 for (auto& selection : group.selections) {
452 selection.event_attr.sample_type = sample_type;
453 }
454 }
455 }
456
SetEnableCondition(bool enable_on_open,bool enable_on_exec)457 void EventSelectionSet::SetEnableCondition(bool enable_on_open, bool enable_on_exec) {
458 for (auto& group : groups_) {
459 for (auto& selection : group.selections) {
460 selection.event_attr.disabled = !enable_on_open;
461 selection.event_attr.enable_on_exec = enable_on_exec;
462 }
463 }
464 }
465
IsEnabledOnExec() const466 bool EventSelectionSet::IsEnabledOnExec() const {
467 for (const auto& group : groups_) {
468 for (const auto& selection : group.selections) {
469 if (!selection.event_attr.enable_on_exec) {
470 return false;
471 }
472 }
473 }
474 return true;
475 }
476
SampleIdAll()477 void EventSelectionSet::SampleIdAll() {
478 for (auto& group : groups_) {
479 for (auto& selection : group.selections) {
480 selection.event_attr.sample_id_all = 1;
481 }
482 }
483 }
484
SetSampleRateForNewEvents(const SampleRate & rate)485 void EventSelectionSet::SetSampleRateForNewEvents(const SampleRate& rate) {
486 sample_rate_ = rate;
487 for (auto& group : groups_) {
488 if (!group.set_sample_rate) {
489 SetSampleRateForGroup(group, rate);
490 }
491 }
492 }
493
SetCpusForNewEvents(const std::vector<int> & cpus)494 void EventSelectionSet::SetCpusForNewEvents(const std::vector<int>& cpus) {
495 cpus_ = cpus;
496 for (auto& group : groups_) {
497 if (group.cpus.empty()) {
498 group.cpus = cpus_.value();
499 }
500 }
501 }
502
SetSampleRateForGroup(EventSelectionSet::EventSelectionGroup & group,const SampleRate & rate)503 void EventSelectionSet::SetSampleRateForGroup(EventSelectionSet::EventSelectionGroup& group,
504 const SampleRate& rate) {
505 group.set_sample_rate = true;
506 for (auto& selection : group.selections) {
507 if (rate.UseFreq()) {
508 selection.event_attr.freq = 1;
509 selection.event_attr.sample_freq = rate.sample_freq;
510 } else {
511 selection.event_attr.freq = 0;
512 selection.event_attr.sample_period = rate.sample_period;
513 }
514 }
515 }
516
SetBranchSampling(uint64_t branch_sample_type)517 bool EventSelectionSet::SetBranchSampling(uint64_t branch_sample_type) {
518 if (branch_sample_type != 0 &&
519 (branch_sample_type & (PERF_SAMPLE_BRANCH_ANY | PERF_SAMPLE_BRANCH_ANY_CALL |
520 PERF_SAMPLE_BRANCH_ANY_RETURN | PERF_SAMPLE_BRANCH_IND_CALL)) == 0) {
521 LOG(ERROR) << "Invalid branch_sample_type: 0x" << std::hex << branch_sample_type;
522 return false;
523 }
524 if (branch_sample_type != 0 && !IsBranchSamplingSupported()) {
525 LOG(ERROR) << "branch stack sampling is not supported on this device.";
526 return false;
527 }
528 for (auto& group : groups_) {
529 for (auto& selection : group.selections) {
530 perf_event_attr& attr = selection.event_attr;
531 if (branch_sample_type != 0) {
532 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
533 } else {
534 attr.sample_type &= ~PERF_SAMPLE_BRANCH_STACK;
535 }
536 attr.branch_sample_type = branch_sample_type;
537 }
538 }
539 return true;
540 }
541
EnableFpCallChainSampling()542 void EventSelectionSet::EnableFpCallChainSampling() {
543 for (auto& group : groups_) {
544 for (auto& selection : group.selections) {
545 selection.event_attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
546 }
547 }
548 }
549
EnableDwarfCallChainSampling(uint32_t dump_stack_size)550 bool EventSelectionSet::EnableDwarfCallChainSampling(uint32_t dump_stack_size) {
551 if (!IsDwarfCallChainSamplingSupported()) {
552 LOG(ERROR) << "dwarf callchain sampling is not supported on this device.";
553 return false;
554 }
555 for (auto& group : groups_) {
556 for (auto& selection : group.selections) {
557 selection.event_attr.sample_type |=
558 PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER;
559 selection.event_attr.exclude_callchain_user = 1;
560 selection.event_attr.sample_regs_user = GetSupportedRegMask(GetMachineArch());
561 selection.event_attr.sample_stack_user = dump_stack_size;
562 }
563 }
564 return true;
565 }
566
SetInherit(bool enable)567 void EventSelectionSet::SetInherit(bool enable) {
568 for (auto& group : groups_) {
569 for (auto& selection : group.selections) {
570 selection.event_attr.inherit = (enable ? 1 : 0);
571 }
572 }
573 }
574
SetClockId(int clock_id)575 void EventSelectionSet::SetClockId(int clock_id) {
576 for (auto& group : groups_) {
577 for (auto& selection : group.selections) {
578 selection.event_attr.use_clockid = 1;
579 selection.event_attr.clockid = clock_id;
580 }
581 }
582 }
583
NeedKernelSymbol() const584 bool EventSelectionSet::NeedKernelSymbol() const {
585 return !ExcludeKernel();
586 }
587
SetRecordNotExecutableMaps(bool record)588 void EventSelectionSet::SetRecordNotExecutableMaps(bool record) {
589 // We only need to dump non-executable mmap records for the first event type.
590 groups_[0].selections[0].event_attr.mmap_data = record ? 1 : 0;
591 }
592
RecordNotExecutableMaps() const593 bool EventSelectionSet::RecordNotExecutableMaps() const {
594 return groups_[0].selections[0].event_attr.mmap_data == 1;
595 }
596
EnableSwitchRecord()597 void EventSelectionSet::EnableSwitchRecord() {
598 groups_[0].selections[0].event_attr.context_switch = 1;
599 }
600
WakeupPerSample()601 void EventSelectionSet::WakeupPerSample() {
602 for (auto& group : groups_) {
603 for (auto& selection : group.selections) {
604 selection.event_attr.watermark = 0;
605 selection.event_attr.wakeup_events = 1;
606 }
607 }
608 }
609
SetTracepointFilter(const std::string & filter)610 bool EventSelectionSet::SetTracepointFilter(const std::string& filter) {
611 // 1. Find the tracepoint event to set filter.
612 EventSelection* selection = nullptr;
613 if (!groups_.empty()) {
614 auto& group = groups_.back();
615 if (group.selections.size() == 1) {
616 if (group.selections[0].event_attr.type == PERF_TYPE_TRACEPOINT) {
617 selection = &group.selections[0];
618 }
619 }
620 }
621 if (selection == nullptr) {
622 LOG(ERROR) << "No tracepoint event before filter: " << filter;
623 return false;
624 }
625
626 // 2. Check the format of the filter.
627 bool use_quote = false;
628 // Quotes are needed for string operands in kernel >= 4.19, probably after patch "tracing: Rewrite
629 // filter logic to be simpler and faster".
630 if (auto version = GetKernelVersion(); version && version.value() >= std::make_pair(4, 19)) {
631 use_quote = true;
632 }
633
634 FieldNameSet used_fields;
635 auto adjusted_filter = AdjustTracepointFilter(filter, use_quote, &used_fields);
636 if (!adjusted_filter) {
637 return false;
638 }
639
640 // 3. Check if used fields are available in the tracepoint event.
641 auto& event_type = selection->event_type_modifier.event_type;
642 if (auto opt_fields = GetFieldNamesForTracepointEvent(event_type); opt_fields) {
643 FieldNameSet& fields = opt_fields.value();
644 for (const auto& field : used_fields) {
645 if (fields.find(field) == fields.end()) {
646 LOG(ERROR) << "field name " << field << " used in \"" << filter << "\" doesn't exist in "
647 << event_type.name << ". Available fields are "
648 << android::base::Join(fields, ",");
649 return false;
650 }
651 }
652 }
653
654 // 4. Connect the filter to the event.
655 selection->tracepoint_filter = adjusted_filter.value();
656 return true;
657 }
658
OpenEventFilesOnGroup(EventSelectionGroup & group,pid_t tid,int cpu,std::string * failed_event_type)659 bool EventSelectionSet::OpenEventFilesOnGroup(EventSelectionGroup& group, pid_t tid, int cpu,
660 std::string* failed_event_type) {
661 std::vector<std::unique_ptr<EventFd>> event_fds;
662 // Given a tid and cpu, events on the same group should be all opened
663 // successfully or all failed to open.
664 EventFd* group_fd = nullptr;
665 for (auto& selection : group.selections) {
666 std::unique_ptr<EventFd> event_fd = EventFd::OpenEventFile(
667 selection.event_attr, tid, cpu, group_fd, selection.event_type_modifier.name, false);
668 if (!event_fd) {
669 *failed_event_type = selection.event_type_modifier.name;
670 return false;
671 }
672 LOG(VERBOSE) << "OpenEventFile for " << event_fd->Name();
673 event_fds.emplace_back(std::move(event_fd));
674 if (group_fd == nullptr) {
675 group_fd = event_fds.back().get();
676 }
677 }
678 for (size_t i = 0; i < group.selections.size(); ++i) {
679 group.selections[i].event_fds.emplace_back(std::move(event_fds[i]));
680 }
681 return true;
682 }
683
PrepareThreads(const std::set<pid_t> & processes,const std::set<pid_t> & threads)684 static std::set<pid_t> PrepareThreads(const std::set<pid_t>& processes,
685 const std::set<pid_t>& threads) {
686 std::set<pid_t> result = threads;
687 for (auto& pid : processes) {
688 std::vector<pid_t> tids = GetThreadsInProcess(pid);
689 result.insert(tids.begin(), tids.end());
690 }
691 return result;
692 }
693
OpenEventFiles()694 bool EventSelectionSet::OpenEventFiles() {
695 std::vector<int> online_cpus = GetOnlineCpus();
696
697 auto check_if_cpus_online = [&](const std::vector<int>& cpus) {
698 if (cpus.size() == 1 && cpus[0] == -1) {
699 return true;
700 }
701 for (int cpu : cpus) {
702 if (std::find(online_cpus.begin(), online_cpus.end(), cpu) == online_cpus.end()) {
703 LOG(ERROR) << "cpu " << cpu << " is not online.";
704 return false;
705 }
706 }
707 return true;
708 };
709
710 std::set<pid_t> threads = PrepareThreads(processes_, threads_);
711 for (auto& group : groups_) {
712 const std::vector<int>* pcpus = &group.cpus;
713 if (!group.selections[0].allowed_cpus.empty()) {
714 // override cpu list if event's PMU has a cpumask as those PMUs are
715 // agnostic to cpu and it's meaningless to specify cpus for them.
716 pcpus = &group.selections[0].allowed_cpus;
717 }
718 if (pcpus->empty()) {
719 pcpus = &online_cpus;
720 } else if (!check_if_cpus_online(*pcpus)) {
721 return false;
722 }
723
724 size_t success_count = 0;
725 std::string failed_event_type;
726 for (const auto tid : threads) {
727 for (const auto& cpu : *pcpus) {
728 if (OpenEventFilesOnGroup(group, tid, cpu, &failed_event_type)) {
729 success_count++;
730 }
731 }
732 }
733 // We can't guarantee to open perf event file successfully for each thread on each cpu.
734 // Because threads may exit between PrepareThreads() and OpenEventFilesOnGroup(), and
735 // cpus may be offlined between GetOnlineCpus() and OpenEventFilesOnGroup().
736 // So we only check that we can at least monitor one thread for each event group.
737 if (success_count == 0) {
738 int error_number = errno;
739 PLOG(ERROR) << "failed to open perf event file for event_type " << failed_event_type;
740 if (error_number == EMFILE) {
741 LOG(ERROR) << "Please increase hard limit of open file numbers.";
742 }
743 return false;
744 }
745 }
746 return ApplyFilters();
747 }
748
ApplyFilters()749 bool EventSelectionSet::ApplyFilters() {
750 return ApplyAddrFilters() && ApplyTracepointFilters();
751 }
752
ApplyAddrFilters()753 bool EventSelectionSet::ApplyAddrFilters() {
754 if (addr_filters_.empty()) {
755 return true;
756 }
757 if (!has_aux_trace_) {
758 LOG(ERROR) << "addr filters only take effect in cs-etm instruction tracing";
759 return false;
760 }
761
762 // Check filter count limit.
763 size_t required_etm_filter_count = 0;
764 for (auto& filter : addr_filters_) {
765 // A range filter needs two etm filters.
766 required_etm_filter_count +=
767 (filter.type == AddrFilter::FILE_RANGE || filter.type == AddrFilter::KERNEL_RANGE) ? 2 : 1;
768 }
769 size_t etm_filter_count = ETMRecorder::GetInstance().GetAddrFilterPairs() * 2;
770 if (etm_filter_count < required_etm_filter_count) {
771 LOG(ERROR) << "needed " << required_etm_filter_count << " etm filters, but only "
772 << etm_filter_count << " filters are available.";
773 return false;
774 }
775
776 std::string filter_str;
777 for (auto& filter : addr_filters_) {
778 if (!filter_str.empty()) {
779 filter_str += ',';
780 }
781 filter_str += filter.ToString();
782 }
783
784 for (auto& group : groups_) {
785 for (auto& selection : group.selections) {
786 if (IsEtmEventType(selection.event_type_modifier.event_type.type)) {
787 for (auto& event_fd : selection.event_fds) {
788 if (!event_fd->SetFilter(filter_str)) {
789 return false;
790 }
791 }
792 }
793 }
794 }
795 return true;
796 }
797
ApplyTracepointFilters()798 bool EventSelectionSet::ApplyTracepointFilters() {
799 for (auto& group : groups_) {
800 for (auto& selection : group.selections) {
801 if (!selection.tracepoint_filter.empty()) {
802 for (auto& event_fd : selection.event_fds) {
803 if (!event_fd->SetFilter(selection.tracepoint_filter)) {
804 return false;
805 }
806 }
807 }
808 }
809 }
810 return true;
811 }
812
ReadCounter(EventFd * event_fd,CounterInfo * counter)813 static bool ReadCounter(EventFd* event_fd, CounterInfo* counter) {
814 if (!event_fd->ReadCounter(&counter->counter)) {
815 return false;
816 }
817 counter->tid = event_fd->ThreadId();
818 counter->cpu = event_fd->Cpu();
819 return true;
820 }
821
ReadCounters(std::vector<CountersInfo> * counters)822 bool EventSelectionSet::ReadCounters(std::vector<CountersInfo>* counters) {
823 counters->clear();
824 for (size_t i = 0; i < groups_.size(); ++i) {
825 for (auto& selection : groups_[i].selections) {
826 CountersInfo counters_info;
827 counters_info.group_id = i;
828 counters_info.event_name = selection.event_type_modifier.event_type.name;
829 counters_info.event_modifier = selection.event_type_modifier.modifier;
830 counters_info.counters = selection.hotplugged_counters;
831 for (auto& event_fd : selection.event_fds) {
832 CounterInfo counter;
833 if (!ReadCounter(event_fd.get(), &counter)) {
834 return false;
835 }
836 counters_info.counters.push_back(counter);
837 }
838 counters->push_back(counters_info);
839 }
840 }
841 return true;
842 }
843
MmapEventFiles(size_t min_mmap_pages,size_t max_mmap_pages,size_t aux_buffer_size,size_t record_buffer_size,bool allow_truncating_samples,bool exclude_perf)844 bool EventSelectionSet::MmapEventFiles(size_t min_mmap_pages, size_t max_mmap_pages,
845 size_t aux_buffer_size, size_t record_buffer_size,
846 bool allow_truncating_samples, bool exclude_perf) {
847 record_read_thread_.reset(new simpleperf::RecordReadThread(
848 record_buffer_size, groups_[0].selections[0].event_attr, min_mmap_pages, max_mmap_pages,
849 aux_buffer_size, allow_truncating_samples, exclude_perf));
850 return true;
851 }
852
PrepareToReadMmapEventData(const std::function<bool (Record *)> & callback)853 bool EventSelectionSet::PrepareToReadMmapEventData(const std::function<bool(Record*)>& callback) {
854 // Prepare record callback function.
855 record_callback_ = callback;
856 if (!record_read_thread_->RegisterDataCallback(*loop_,
857 [this]() { return ReadMmapEventData(true); })) {
858 return false;
859 }
860 std::vector<EventFd*> event_fds;
861 for (auto& group : groups_) {
862 for (auto& selection : group.selections) {
863 for (auto& event_fd : selection.event_fds) {
864 event_fds.push_back(event_fd.get());
865 }
866 }
867 }
868 return record_read_thread_->AddEventFds(event_fds);
869 }
870
SyncKernelBuffer()871 bool EventSelectionSet::SyncKernelBuffer() {
872 return record_read_thread_->SyncKernelBuffer();
873 }
874
875 // Read records from the RecordBuffer. If with_time_limit is false, read until the RecordBuffer is
876 // empty, otherwise stop after 100 ms or when the record buffer is empty.
ReadMmapEventData(bool with_time_limit)877 bool EventSelectionSet::ReadMmapEventData(bool with_time_limit) {
878 uint64_t start_time_in_ns;
879 if (with_time_limit) {
880 start_time_in_ns = GetSystemClock();
881 }
882 std::unique_ptr<Record> r;
883 while ((r = record_read_thread_->GetRecord()) != nullptr) {
884 if (!record_callback_(r.get())) {
885 return false;
886 }
887 if (with_time_limit && (GetSystemClock() - start_time_in_ns) >= 1e8) {
888 break;
889 }
890 }
891 return true;
892 }
893
FinishReadMmapEventData()894 bool EventSelectionSet::FinishReadMmapEventData() {
895 return ReadMmapEventData(false);
896 }
897
CloseEventFiles()898 void EventSelectionSet::CloseEventFiles() {
899 if (record_read_thread_) {
900 record_read_thread_->StopReadThread();
901 }
902 for (auto& group : groups_) {
903 for (auto& event : group.selections) {
904 event.event_fds.clear();
905 }
906 }
907 }
908
StopWhenNoMoreTargets(double check_interval_in_sec)909 bool EventSelectionSet::StopWhenNoMoreTargets(double check_interval_in_sec) {
910 return loop_->AddPeriodicEvent(SecondToTimeval(check_interval_in_sec),
911 [&]() { return CheckMonitoredTargets(); });
912 }
913
CheckMonitoredTargets()914 bool EventSelectionSet::CheckMonitoredTargets() {
915 if (!HasSampler()) {
916 return loop_->ExitLoop();
917 }
918 for (const auto& tid : threads_) {
919 if (IsThreadAlive(tid)) {
920 return true;
921 }
922 }
923 for (const auto& pid : processes_) {
924 if (IsThreadAlive(pid)) {
925 return true;
926 }
927 }
928 return loop_->ExitLoop();
929 }
930
HasSampler()931 bool EventSelectionSet::HasSampler() {
932 for (auto& group : groups_) {
933 for (auto& sel : group.selections) {
934 if (!sel.event_fds.empty()) {
935 return true;
936 }
937 }
938 }
939 return false;
940 }
941
SetEnableEvents(bool enable)942 bool EventSelectionSet::SetEnableEvents(bool enable) {
943 for (auto& group : groups_) {
944 for (auto& sel : group.selections) {
945 for (auto& fd : sel.event_fds) {
946 if (!fd->SetEnableEvent(enable)) {
947 return false;
948 }
949 }
950 }
951 }
952 return true;
953 }
954
EnableETMEvents()955 bool EventSelectionSet::EnableETMEvents() {
956 for (auto& group : groups_) {
957 for (auto& sel : group.selections) {
958 if (!sel.event_type_modifier.event_type.IsEtmEvent()) {
959 continue;
960 }
961 for (auto& fd : sel.event_fds) {
962 if (!fd->SetEnableEvent(true)) {
963 return false;
964 }
965 }
966 }
967 }
968 return true;
969 }
970
DisableETMEvents()971 bool EventSelectionSet::DisableETMEvents() {
972 for (auto& group : groups_) {
973 for (auto& sel : group.selections) {
974 if (!sel.event_type_modifier.event_type.IsEtmEvent()) {
975 continue;
976 }
977 // When using ETR, ETM data is flushed to the aux buffer of the last cpu disabling ETM events.
978 // To avoid overflowing the aux buffer for one cpu, rotate the last cpu disabling ETM events.
979 if (etm_event_cpus_.empty()) {
980 for (const auto& fd : sel.event_fds) {
981 etm_event_cpus_.insert(fd->Cpu());
982 }
983 if (etm_event_cpus_.empty()) {
984 continue;
985 }
986 etm_event_cpus_it_ = etm_event_cpus_.begin();
987 }
988 int last_disabled_cpu = *etm_event_cpus_it_;
989 if (++etm_event_cpus_it_ == etm_event_cpus_.end()) {
990 etm_event_cpus_it_ = etm_event_cpus_.begin();
991 }
992
993 for (auto& fd : sel.event_fds) {
994 if (fd->Cpu() != last_disabled_cpu) {
995 if (!fd->SetEnableEvent(false)) {
996 return false;
997 }
998 }
999 }
1000 for (auto& fd : sel.event_fds) {
1001 if (fd->Cpu() == last_disabled_cpu) {
1002 if (!fd->SetEnableEvent(false)) {
1003 return false;
1004 }
1005 }
1006 }
1007 }
1008 }
1009 return true;
1010 }
1011
1012 } // namespace simpleperf
1013