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 "profile_saver.h"
18
19 #include <fcntl.h>
20 #include <sys/resource.h>
21 #include <sys/stat.h>
22 #include <sys/types.h>
23 #include <unistd.h>
24
25 #include "android-base/strings.h"
26 #include "art_method-inl.h"
27 #include "base/compiler_filter.h"
28 #include "base/enums.h"
29 #include "base/logging.h" // For VLOG.
30 #include "base/scoped_arena_containers.h"
31 #include "base/stl_util.h"
32 #include "base/systrace.h"
33 #include "base/time_utils.h"
34 #include "base/unix_file/fd_file.h"
35 #include "class_table-inl.h"
36 #include "dex/dex_file_loader.h"
37 #include "dex_reference_collection.h"
38 #include "gc/collector_type.h"
39 #include "gc/gc_cause.h"
40 #include "jit/jit.h"
41 #include "jit/profiling_info.h"
42 #include "oat_file_manager.h"
43 #include "profile/profile_compilation_info.h"
44 #include "scoped_thread_state_change-inl.h"
45
46 namespace art {
47
48 using Hotness = ProfileCompilationInfo::MethodHotness;
49
50 ProfileSaver* ProfileSaver::instance_ = nullptr;
51 pthread_t ProfileSaver::profiler_pthread_ = 0U;
52
53 static_assert(ProfileCompilationInfo::kIndividualInlineCacheSize ==
54 InlineCache::kIndividualCacheSize,
55 "InlineCache and ProfileCompilationInfo do not agree on kIndividualCacheSize");
56
57 // At what priority to schedule the saver threads. 9 is the lowest foreground priority on device.
58 static constexpr int kProfileSaverPthreadPriority = 9;
59
SetProfileSaverThreadPriority(pthread_t thread,int priority)60 static void SetProfileSaverThreadPriority(pthread_t thread, int priority) {
61 #if defined(ART_TARGET_ANDROID)
62 int result = setpriority(PRIO_PROCESS, pthread_gettid_np(thread), priority);
63 if (result != 0) {
64 LOG(ERROR) << "Failed to setpriority to :" << priority;
65 }
66 #else
67 UNUSED(thread);
68 UNUSED(priority);
69 #endif
70 }
71
GetDefaultThreadPriority()72 static int GetDefaultThreadPriority() {
73 #if defined(ART_TARGET_ANDROID)
74 pthread_attr_t attr;
75 sched_param param;
76 pthread_attr_init(&attr);
77 pthread_attr_getschedparam(&attr, ¶m);
78 return param.sched_priority;
79 #else
80 return 0;
81 #endif
82 }
83
ProfileSaver(const ProfileSaverOptions & options,jit::JitCodeCache * jit_code_cache)84 ProfileSaver::ProfileSaver(const ProfileSaverOptions& options, jit::JitCodeCache* jit_code_cache)
85 : jit_code_cache_(jit_code_cache),
86 shutting_down_(false),
87 last_time_ns_saver_woke_up_(0),
88 jit_activity_notifications_(0),
89 wait_lock_("ProfileSaver wait lock"),
90 period_condition_("ProfileSaver period condition", wait_lock_),
91 total_bytes_written_(0),
92 total_number_of_writes_(0),
93 total_number_of_code_cache_queries_(0),
94 total_number_of_skipped_writes_(0),
95 total_number_of_failed_writes_(0),
96 total_ms_of_sleep_(0),
97 total_ns_of_work_(0),
98 total_number_of_hot_spikes_(0),
99 total_number_of_wake_ups_(0),
100 options_(options) {
101 DCHECK(options_.IsEnabled());
102 }
103
~ProfileSaver()104 ProfileSaver::~ProfileSaver() {
105 for (auto& it : profile_cache_) {
106 delete it.second;
107 }
108 }
109
NotifyStartupCompleted()110 void ProfileSaver::NotifyStartupCompleted() {
111 Thread* self = Thread::Current();
112 MutexLock mu(self, *Locks::profiler_lock_);
113 if (instance_ == nullptr || instance_->shutting_down_) {
114 return;
115 }
116 MutexLock mu2(self, instance_->wait_lock_);
117 instance_->period_condition_.Signal(self);
118 }
119
Run()120 void ProfileSaver::Run() {
121 Thread* self = Thread::Current();
122
123 // For thread annotalysis, the setup is more complicated than it should be. Run needs to start
124 // under mutex, but should drop it.
125 Locks::profiler_lock_->ExclusiveUnlock(self);
126
127 bool check_for_first_save =
128 options_.GetMinFirstSaveMs() != ProfileSaverOptions::kMinFirstSaveMsNotSet;
129 bool force_early_first_save = check_for_first_save && IsFirstSave();
130
131 // Fetch the resolved classes for the app images after sleeping for
132 // options_.GetSaveResolvedClassesDelayMs().
133 // TODO(calin) This only considers the case of the primary profile file.
134 // Anything that gets loaded in the same VM will not have their resolved
135 // classes save (unless they started before the initial saving was done).
136 {
137 MutexLock mu(self, wait_lock_);
138
139 const uint64_t sleep_time = MsToNs(force_early_first_save
140 ? options_.GetMinFirstSaveMs()
141 : options_.GetSaveResolvedClassesDelayMs());
142 const uint64_t start_time = NanoTime();
143 const uint64_t end_time = start_time + sleep_time;
144 while (!Runtime::Current()->GetStartupCompleted() || force_early_first_save) {
145 const uint64_t current_time = NanoTime();
146 if (current_time >= end_time) {
147 break;
148 }
149 period_condition_.TimedWait(self, NsToMs(end_time - current_time), 0);
150 }
151 total_ms_of_sleep_ += NsToMs(NanoTime() - start_time);
152 }
153
154 FetchAndCacheResolvedClassesAndMethods(/*startup=*/ true);
155
156 // When we save without waiting for JIT notifications we use a simple
157 // exponential back off policy bounded by max_wait_without_jit.
158 uint32_t max_wait_without_jit = options_.GetMinSavePeriodMs() * 16;
159 uint64_t cur_wait_without_jit = options_.GetMinSavePeriodMs();
160
161 // Loop for the profiled methods.
162 while (!ShuttingDown(self)) {
163 // Sleep only if we don't have to force an early first save configured
164 // with GetMinFirstSaveMs().
165 // If we do have to save early, move directly to the processing part
166 // since we already slept before fetching and resolving the startup
167 // classes.
168 if (!force_early_first_save) {
169 uint64_t sleep_start = NanoTime();
170 uint64_t sleep_time = 0;
171 {
172 MutexLock mu(self, wait_lock_);
173 if (options_.GetWaitForJitNotificationsToSave()) {
174 period_condition_.Wait(self);
175 } else {
176 period_condition_.TimedWait(self, cur_wait_without_jit, 0);
177 if (cur_wait_without_jit < max_wait_without_jit) {
178 cur_wait_without_jit *= 2;
179 }
180 }
181 sleep_time = NanoTime() - sleep_start;
182 }
183 // Check if the thread was woken up for shutdown.
184 if (ShuttingDown(self)) {
185 break;
186 }
187 total_number_of_wake_ups_++;
188 // We might have been woken up by a huge number of notifications to guarantee saving.
189 // If we didn't meet the minimum saving period go back to sleep (only if missed by
190 // a reasonable margin).
191 uint64_t min_save_period_ns = MsToNs(options_.GetMinSavePeriodMs());
192 while (min_save_period_ns * 0.9 > sleep_time) {
193 {
194 MutexLock mu(self, wait_lock_);
195 period_condition_.TimedWait(self, NsToMs(min_save_period_ns - sleep_time), 0);
196 sleep_time = NanoTime() - sleep_start;
197 }
198 // Check if the thread was woken up for shutdown.
199 if (ShuttingDown(self)) {
200 break;
201 }
202 total_number_of_wake_ups_++;
203 }
204 total_ms_of_sleep_ += NsToMs(NanoTime() - sleep_start);
205 }
206
207 if (ShuttingDown(self)) {
208 break;
209 }
210
211 uint16_t number_of_new_methods = 0;
212 uint64_t start_work = NanoTime();
213 // If we force an early_first_save do not run FetchAndCacheResolvedClassesAndMethods
214 // again. We just did it. So pass true to skip_class_and_method_fetching.
215 bool profile_saved_to_disk = ProcessProfilingInfo(
216 /*force_save=*/ false,
217 /*skip_class_and_method_fetching=*/ force_early_first_save,
218 &number_of_new_methods);
219
220 // Reset the flag, so we can continue on the normal schedule.
221 force_early_first_save = false;
222
223 // Update the notification counter based on result. Note that there might be contention on this
224 // but we don't care about to be 100% precise.
225 if (!profile_saved_to_disk) {
226 // If we didn't save to disk it may be because we didn't have enough new methods.
227 // Set the jit activity notifications to number_of_new_methods so we can wake up earlier
228 // if needed.
229 jit_activity_notifications_ = number_of_new_methods;
230 }
231 total_ns_of_work_ += NanoTime() - start_work;
232 }
233 }
234
235 // Checks if the profile file is empty.
236 // Return true if the size of the profile file is 0 or if there were errors when
237 // trying to open the file.
IsProfileEmpty(const std::string & location)238 static bool IsProfileEmpty(const std::string& location) {
239 if (location.empty()) {
240 return true;
241 }
242
243 struct stat stat_buffer;
244 if (stat(location.c_str(), &stat_buffer) != 0) {
245 if (VLOG_IS_ON(profiler)) {
246 PLOG(WARNING) << "Failed to stat profile location for IsFirstUse: " << location;
247 }
248 return true;
249 }
250
251 VLOG(profiler) << "Profile " << location << " size=" << stat_buffer.st_size;
252 return stat_buffer.st_size == 0;
253 }
254
IsFirstSave()255 bool ProfileSaver::IsFirstSave() {
256 Thread* self = Thread::Current();
257 SafeMap<std::string, std::string> tracked_locations;
258 {
259 // Make a copy so that we don't hold the lock while doing I/O.
260 MutexLock mu(self, *Locks::profiler_lock_);
261 tracked_locations = tracked_profiles_;
262 }
263
264 for (const auto& it : tracked_locations) {
265 if (ShuttingDown(self)) {
266 return false;
267 }
268 const std::string& cur_profile = it.first;
269 const std::string& ref_profile = it.second;
270
271 // Check if any profile is non empty. If so, then this is not the first save.
272 if (!IsProfileEmpty(cur_profile) || !IsProfileEmpty(ref_profile)) {
273 return false;
274 }
275 }
276
277 // All locations are empty. Assume this is the first use.
278 VLOG(profiler) << "All profile locations are empty. This is considered to be first save";
279 return true;
280 }
281
NotifyJitActivity()282 void ProfileSaver::NotifyJitActivity() {
283 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
284 if (instance_ == nullptr || instance_->shutting_down_) {
285 return;
286 }
287 instance_->NotifyJitActivityInternal();
288 }
289
WakeUpSaver()290 void ProfileSaver::WakeUpSaver() {
291 jit_activity_notifications_ = 0;
292 last_time_ns_saver_woke_up_ = NanoTime();
293 period_condition_.Signal(Thread::Current());
294 }
295
NotifyJitActivityInternal()296 void ProfileSaver::NotifyJitActivityInternal() {
297 // Unlikely to overflow but if it happens,
298 // we would have waken up the saver long before that.
299 jit_activity_notifications_++;
300 // Note that we are not as precise as we could be here but we don't want to wake the saver
301 // every time we see a hot method.
302 if (jit_activity_notifications_ > options_.GetMinNotificationBeforeWake()) {
303 MutexLock wait_mutex(Thread::Current(), wait_lock_);
304 if ((NanoTime() - last_time_ns_saver_woke_up_) > MsToNs(options_.GetMinSavePeriodMs())) {
305 WakeUpSaver();
306 } else if (jit_activity_notifications_ > options_.GetMaxNotificationBeforeWake()) {
307 // Make sure to wake up the saver if we see a spike in the number of notifications.
308 // This is a precaution to avoid losing a big number of methods in case
309 // this is a spike with no jit after.
310 total_number_of_hot_spikes_++;
311 WakeUpSaver();
312 }
313 }
314 }
315
316 class ProfileSaver::ScopedDefaultPriority {
317 public:
ScopedDefaultPriority(pthread_t thread)318 explicit ScopedDefaultPriority(pthread_t thread) : thread_(thread) {
319 SetProfileSaverThreadPriority(thread_, GetDefaultThreadPriority());
320 }
321
~ScopedDefaultPriority()322 ~ScopedDefaultPriority() {
323 SetProfileSaverThreadPriority(thread_, kProfileSaverPthreadPriority);
324 }
325
326 private:
327 const pthread_t thread_;
328 };
329
330 class ProfileSaver::GetClassesAndMethodsHelper {
331 public:
GetClassesAndMethodsHelper(bool startup,const ProfileSaverOptions & options,const ProfileCompilationInfo::ProfileSampleAnnotation & annotation)332 GetClassesAndMethodsHelper(bool startup,
333 const ProfileSaverOptions& options,
334 const ProfileCompilationInfo::ProfileSampleAnnotation& annotation)
335 REQUIRES_SHARED(Locks::mutator_lock_)
336 : startup_(startup),
337 profile_boot_class_path_(options.GetProfileBootClassPath()),
338 hot_method_sample_threshold_(CalculateHotMethodSampleThreshold(startup, options)),
339 extra_flags_(GetExtraMethodHotnessFlags(options)),
340 annotation_(annotation),
341 arena_stack_(Runtime::Current()->GetArenaPool()),
342 allocator_(&arena_stack_),
343 class_loaders_(std::nullopt),
344 dex_file_records_map_(allocator_.Adapter(kArenaAllocProfile)),
345 number_of_hot_methods_(0u),
346 number_of_sampled_methods_(0u) {
347 std::fill_n(max_primitive_array_dimensions_.data(), max_primitive_array_dimensions_.size(), 0u);
348 }
349
REQUIRES_SHARED(Locks::mutator_lock_)350 ~GetClassesAndMethodsHelper() REQUIRES_SHARED(Locks::mutator_lock_) {
351 // The `class_loaders_` member destructor needs the mutator lock.
352 // We need to destroy arena-allocated dex file records.
353 for (const auto& entry : dex_file_records_map_) {
354 delete entry.second;
355 }
356 }
357
358 void CollectClasses(Thread* self) REQUIRES_SHARED(Locks::mutator_lock_);
359 void UpdateProfile(const std::set<std::string>& locations, ProfileCompilationInfo* profile_info);
360
GetHotMethodSampleThreshold() const361 uint32_t GetHotMethodSampleThreshold() const {
362 return hot_method_sample_threshold_;
363 }
364
GetNumberOfHotMethods() const365 size_t GetNumberOfHotMethods() const {
366 return number_of_hot_methods_;
367 }
368
GetNumberOfSampledMethods() const369 size_t GetNumberOfSampledMethods() const {
370 return number_of_sampled_methods_;
371 }
372
373 private:
374 // GetClassLoadersVisitor collects visited class loaders.
375 class GetClassLoadersVisitor : public ClassLoaderVisitor {
376 public:
GetClassLoadersVisitor(VariableSizedHandleScope * class_loaders)377 explicit GetClassLoadersVisitor(VariableSizedHandleScope* class_loaders)
378 : class_loaders_(class_loaders) {}
379
Visit(ObjPtr<mirror::ClassLoader> class_loader)380 void Visit(ObjPtr<mirror::ClassLoader> class_loader)
381 REQUIRES_SHARED(Locks::classlinker_classes_lock_, Locks::mutator_lock_) override {
382 DCHECK(class_loader != nullptr);
383 class_loaders_->NewHandle(class_loader);
384 }
385
386 private:
387 VariableSizedHandleScope* const class_loaders_;
388 };
389
390 class CollectInternalVisitor {
391 public:
CollectInternalVisitor(GetClassesAndMethodsHelper * helper)392 explicit CollectInternalVisitor(GetClassesAndMethodsHelper* helper)
393 : helper_(helper) {}
394
VisitRootIfNonNull(StackReference<mirror::Object> * ref)395 void VisitRootIfNonNull(StackReference<mirror::Object>* ref)
396 REQUIRES_SHARED(Locks::mutator_lock_) {
397 if (!ref->IsNull()) {
398 helper_->CollectInternal</*kBootClassLoader=*/ false>(ref->AsMirrorPtr()->AsClassLoader());
399 }
400 }
401
402 private:
403 GetClassesAndMethodsHelper* helper_;
404 };
405
406 struct ClassRecord {
407 dex::TypeIndex type_index;
408 uint16_t array_dimension;
409 uint32_t copied_methods_start;
410 LengthPrefixedArray<ArtMethod>* methods;
411 };
412
413 struct DexFileRecords : public DeletableArenaObject<kArenaAllocProfile> {
DexFileRecordsart::ProfileSaver::GetClassesAndMethodsHelper::DexFileRecords414 explicit DexFileRecords(ScopedArenaAllocator* allocator)
415 : class_records(allocator->Adapter(kArenaAllocProfile)),
416 copied_methods(allocator->Adapter(kArenaAllocProfile)) {
417 class_records.reserve(kInitialClassRecordsReservation);
418 }
419
420 static constexpr size_t kInitialClassRecordsReservation = 512;
421
422 ScopedArenaVector<ClassRecord> class_records;
423 ScopedArenaVector<ArtMethod*> copied_methods;
424 };
425
426 using DexFileRecordsMap = ScopedArenaHashMap<const DexFile*, DexFileRecords*>;
427
CalculateHotMethodSampleThreshold(bool startup,const ProfileSaverOptions & options)428 static uint32_t CalculateHotMethodSampleThreshold(bool startup,
429 const ProfileSaverOptions& options) {
430 Runtime* runtime = Runtime::Current();
431 if (startup) {
432 const bool is_low_ram = runtime->GetHeap()->IsLowMemoryMode();
433 return options.GetHotStartupMethodSamples(is_low_ram);
434 } else if (runtime->GetJit() != nullptr) {
435 return runtime->GetJit()->WarmMethodThreshold();
436 } else {
437 return std::numeric_limits<uint32_t>::max();
438 }
439 }
440
ShouldCollectClasses(bool startup)441 ALWAYS_INLINE static bool ShouldCollectClasses(bool startup) {
442 // We only record classes for the startup case. This may change in the future.
443 return startup;
444 }
445
446 // Collect classes and methods from one class loader.
447 template <bool kBootClassLoader>
448 void CollectInternal(ObjPtr<mirror::ClassLoader> class_loader) NO_INLINE
449 REQUIRES_SHARED(Locks::mutator_lock_);
450
451 const bool startup_;
452 const bool profile_boot_class_path_;
453 const uint32_t hot_method_sample_threshold_;
454 const uint32_t extra_flags_;
455 const ProfileCompilationInfo::ProfileSampleAnnotation annotation_;
456 ArenaStack arena_stack_;
457 ScopedArenaAllocator allocator_;
458 std::optional<VariableSizedHandleScope> class_loaders_;
459 DexFileRecordsMap dex_file_records_map_;
460
461 static_assert(Primitive::kPrimLast == Primitive::kPrimVoid); // There are no arrays of void.
462 std::array<uint8_t, static_cast<size_t>(Primitive::kPrimLast)> max_primitive_array_dimensions_;
463
464 size_t number_of_hot_methods_;
465 size_t number_of_sampled_methods_;
466 };
467
468 template <bool kBootClassLoader>
CollectInternal(ObjPtr<mirror::ClassLoader> class_loader)469 void ProfileSaver::GetClassesAndMethodsHelper::CollectInternal(
470 ObjPtr<mirror::ClassLoader> class_loader) {
471 ScopedTrace trace(__PRETTY_FUNCTION__);
472 DCHECK_EQ(kBootClassLoader, class_loader == nullptr);
473
474 // If the class loader has not loaded any classes, it may have a null table.
475 ClassLinker* const class_linker = Runtime::Current()->GetClassLinker();
476 ClassTable* const table =
477 class_linker->ClassTableForClassLoader(kBootClassLoader ? nullptr : class_loader);
478 if (table == nullptr) {
479 return;
480 }
481
482 // Move members to local variables to allow the compiler to optimize this properly.
483 const bool startup = startup_;
484 table->Visit([&](ObjPtr<mirror::Class> klass) REQUIRES_SHARED(Locks::mutator_lock_) {
485 if (kBootClassLoader ? (!klass->IsBootStrapClassLoaded())
486 : (klass->GetClassLoader() != class_loader)) {
487 // To avoid processing a class more than once, we process each class only
488 // when we encounter it in the defining class loader's class table.
489 // This class has a different defining class loader, skip it.
490 return true;
491 }
492
493 uint16_t dim = 0u;
494 ObjPtr<mirror::Class> k = klass;
495 if (klass->IsArrayClass()) {
496 DCHECK_EQ(klass->NumMethods(), 0u); // No methods to collect.
497 if (!ShouldCollectClasses(startup)) {
498 return true;
499 }
500 do {
501 DCHECK(k->IsResolved()); // Array classes are always resolved.
502 ++dim;
503 // At the time of array class creation, the element type is already either
504 // resolved or erroneous unresoved and either shall remain an invariant.
505 // Similarly, the access flag indicating a proxy class is an invariant.
506 // Read barrier is unnecessary for reading a chain of constant references
507 // in order to read primitive fields to check such invariants, or to read
508 // other constant primitive fields (dex file, primitive type) below.
509 k = k->GetComponentType<kDefaultVerifyFlags, kWithoutReadBarrier>();
510 } while (k->IsArrayClass());
511
512 DCHECK(kBootClassLoader || !k->IsPrimitive());
513 if (kBootClassLoader && UNLIKELY(k->IsPrimitive())) {
514 size_t index = enum_cast<size_t>(k->GetPrimitiveType());
515 DCHECK_LT(index, max_primitive_array_dimensions_.size());
516 if (dim > max_primitive_array_dimensions_[index]) {
517 // Enforce an upper limit of 255 for primitive array dimensions.
518 max_primitive_array_dimensions_[index] =
519 std::min<size_t>(dim, std::numeric_limits<uint8_t>::max());
520 }
521 return true;
522 }
523
524 // Attribute the array class to the defining dex file of the element class.
525 DCHECK_EQ(klass->GetCopiedMethodsStartOffset(), 0u);
526 DCHECK(klass->GetMethodsPtr() == nullptr);
527 } else {
528 // Non-array class. There is no need to collect primitive types.
529 DCHECK(kBootClassLoader || !k->IsPrimitive());
530 if (kBootClassLoader && UNLIKELY(klass->IsPrimitive())) {
531 DCHECK(profile_boot_class_path_);
532 DCHECK_EQ(klass->NumMethods(), 0u); // No methods to collect.
533 return true;
534 }
535 }
536
537 if (!k->IsResolved() || k->IsProxyClass()) {
538 return true;
539 }
540
541 const DexFile& dex_file = k->GetDexFile();
542 dex::TypeIndex type_index = k->GetDexTypeIndex();
543 uint32_t copied_methods_start = klass->GetCopiedMethodsStartOffset();
544 LengthPrefixedArray<ArtMethod>* methods = klass->GetMethodsPtr();
545
546 DexFileRecords* dex_file_records;
547 auto it = dex_file_records_map_.find(&dex_file);
548 if (it != dex_file_records_map_.end()) {
549 dex_file_records = it->second;
550 } else {
551 dex_file_records = new (&allocator_) DexFileRecords(&allocator_);
552 dex_file_records_map_.insert(std::make_pair(&dex_file, dex_file_records));
553 }
554 dex_file_records->class_records.push_back(
555 ClassRecord{type_index, dim, copied_methods_start, methods});
556 return true;
557 });
558 }
559
CollectClasses(Thread * self)560 void ProfileSaver::GetClassesAndMethodsHelper::CollectClasses(Thread* self) {
561 ScopedTrace trace(__PRETTY_FUNCTION__);
562
563 // Collect class loaders into a `VariableSizedHandleScope` to prevent contention
564 // problems on the class_linker_classes_lock. Hold those class loaders in
565 // a member variable to keep them alive and prevent unloading their classes,
566 // so that methods referenced in collected `DexFileRecords` remain valid.
567 class_loaders_.emplace(self);
568 {
569 GetClassLoadersVisitor class_loader_visitor(&class_loaders_.value());
570 ClassLinker* const class_linker = Runtime::Current()->GetClassLinker();
571 ReaderMutexLock mu(self, *Locks::classlinker_classes_lock_);
572 class_linker->VisitClassLoaders(&class_loader_visitor);
573 }
574
575 // Collect classes and their method array pointers.
576 if (profile_boot_class_path_) {
577 // Collect classes from the boot class loader since visit classloaders doesn't visit it.
578 CollectInternal</*kBootClassLoader=*/ true>(/*class_loader=*/ nullptr);
579 }
580 {
581 CollectInternalVisitor visitor(this);
582 class_loaders_->VisitRoots(visitor);
583 }
584
585 // Attribute copied methods to defining dex files while holding the mutator lock.
586 for (const auto& entry : dex_file_records_map_) {
587 const DexFile* dex_file = entry.first;
588 DexFileRecords* dex_file_records = entry.second;
589
590 for (const ClassRecord& class_record : dex_file_records->class_records) {
591 LengthPrefixedArray<ArtMethod>* methods = class_record.methods;
592 if (methods == nullptr) {
593 continue;
594 }
595 const size_t methods_size = methods->size();
596 for (size_t index = class_record.copied_methods_start; index != methods_size; ++index) {
597 // Note: Using `ArtMethod` array with implicit `kRuntimePointerSize`.
598 ArtMethod& method = methods->At(index);
599 DCHECK(method.IsCopied());
600 DCHECK(!method.IsNative());
601 if (method.IsInvokable()) {
602 const DexFile* method_dex_file = method.GetDexFile();
603 DexFileRecords* method_dex_file_records = dex_file_records;
604 if (method_dex_file != dex_file) {
605 auto it = dex_file_records_map_.find(method_dex_file);
606 if (it == dex_file_records_map_.end()) {
607 // We have not seen any class in the dex file that defines the interface with this
608 // copied method. This can happen if the interface is in the boot class path and
609 // we are not profiling boot class path; or when we first visit classes for the
610 // interface's defining class loader before it has any resolved classes and then
611 // the interface is resolved and an implementing class is defined in a child class
612 // loader before we visit that child class loader's classes.
613 continue;
614 }
615 method_dex_file_records = it->second;
616 }
617 method_dex_file_records->copied_methods.push_back(&method);
618 }
619 }
620 }
621 }
622 }
623
UpdateProfile(const std::set<std::string> & locations,ProfileCompilationInfo * profile_info)624 void ProfileSaver::GetClassesAndMethodsHelper::UpdateProfile(const std::set<std::string>& locations,
625 ProfileCompilationInfo* profile_info) {
626 // Move members to local variables to allow the compiler to optimize this properly.
627 const bool startup = startup_;
628 const uint32_t hot_method_sample_threshold = hot_method_sample_threshold_;
629 const uint32_t base_flags =
630 (startup ? Hotness::kFlagStartup : Hotness::kFlagPostStartup) | extra_flags_;
631
632 // Collect the number of hot and sampled methods.
633 size_t number_of_hot_methods = 0u;
634 size_t number_of_sampled_methods = 0u;
635
636 uint16_t initial_value = Runtime::Current()->GetJITOptions()->GetWarmupThreshold();
637 auto get_method_flags = [&](ArtMethod& method) {
638 // Mark methods as hot if they have more than hot_method_sample_threshold
639 // samples. This means they will get compiled by the compiler driver.
640 if (method.PreviouslyWarm() ||
641 method.CounterHasReached(hot_method_sample_threshold, initial_value)) {
642 ++number_of_hot_methods;
643 return enum_cast<ProfileCompilationInfo::MethodHotness::Flag>(base_flags | Hotness::kFlagHot);
644 } else if (method.CounterHasChanged(initial_value)) {
645 ++number_of_sampled_methods;
646 return enum_cast<ProfileCompilationInfo::MethodHotness::Flag>(base_flags);
647 } else {
648 return enum_cast<ProfileCompilationInfo::MethodHotness::Flag>(0u);
649 }
650 };
651
652 // Use a single string for array descriptors to avoid too many reallocations.
653 std::string array_class_descriptor;
654
655 // Process classes and methods.
656 for (const auto& entry : dex_file_records_map_) {
657 const DexFile* dex_file = entry.first;
658 const DexFileRecords* dex_file_records = entry.second;
659
660 // Check if this is a profiled dex file.
661 const std::string base_location = DexFileLoader::GetBaseLocation(dex_file->GetLocation());
662 if (locations.find(base_location) == locations.end()) {
663 continue;
664 }
665
666 // Get the profile index.
667 ProfileCompilationInfo::ProfileIndexType profile_index =
668 profile_info->FindOrAddDexFile(*dex_file, annotation_);
669 if (profile_index == ProfileCompilationInfo::MaxProfileIndex()) {
670 // Error adding dex file to the `profile_info`.
671 continue;
672 }
673
674 for (const ClassRecord& class_record : dex_file_records->class_records) {
675 if (class_record.array_dimension != 0u) {
676 DCHECK(ShouldCollectClasses(startup));
677 DCHECK(class_record.methods == nullptr); // No methods to process.
678 array_class_descriptor.assign(class_record.array_dimension, '[');
679 array_class_descriptor += dex_file->StringByTypeIdx(class_record.type_index);
680 dex::TypeIndex type_index =
681 profile_info->FindOrCreateTypeIndex(*dex_file, array_class_descriptor.c_str());
682 if (type_index.IsValid()) {
683 profile_info->AddClass(profile_index, type_index);
684 }
685 } else {
686 // Non-array class.
687 if (ShouldCollectClasses(startup)) {
688 profile_info->AddClass(profile_index, class_record.type_index);
689 }
690 const size_t num_declared_methods = class_record.copied_methods_start;
691 LengthPrefixedArray<ArtMethod>* methods = class_record.methods;
692 for (size_t index = 0; index != num_declared_methods; ++index) {
693 // Note: Using `ArtMethod` array with implicit `kRuntimePointerSize`.
694 ArtMethod& method = methods->At(index);
695 DCHECK(!method.IsCopied());
696 // We do not record native methods. Once we AOT-compile the app,
697 // all native methods shall have their JNI stubs compiled.
698 if (method.IsInvokable() && !method.IsNative()) {
699 ProfileCompilationInfo::MethodHotness::Flag flags = get_method_flags(method);
700 if (flags != 0u) {
701 profile_info->AddMethod(profile_index, method.GetDexMethodIndex(), flags);
702 }
703 }
704 }
705 }
706 }
707
708 for (ArtMethod* method : dex_file_records->copied_methods) {
709 DCHECK(method->IsCopied());
710 DCHECK(method->IsInvokable());
711 DCHECK(!method->IsNative());
712 ProfileCompilationInfo::MethodHotness::Flag flags = get_method_flags(*method);
713 if (flags != 0u) {
714 profile_info->AddMethod(profile_index, method->GetDexMethodIndex(), flags);
715 }
716 }
717 }
718
719 if (profile_boot_class_path_) {
720 // Attribute primitive arrays to the first dex file in the boot class path (should
721 // be core-oj). We collect primitive array types to know the needed dimensions.
722 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
723 DCHECK(!class_linker->GetBootClassPath().empty());
724 const DexFile* dex_file = class_linker->GetBootClassPath().front();
725 ProfileCompilationInfo::ProfileIndexType profile_index =
726 profile_info->FindOrAddDexFile(*dex_file, annotation_);
727 if (profile_index != ProfileCompilationInfo::MaxProfileIndex()) {
728 for (size_t i = 0; i != max_primitive_array_dimensions_.size(); ++i) {
729 size_t max_dim = max_primitive_array_dimensions_[i];
730 // Insert descriptors for all dimensions up to `max_dim`.
731 for (size_t dim = 1; dim <= max_dim; ++dim) {
732 array_class_descriptor.assign(dim, '[');
733 array_class_descriptor += Primitive::Descriptor(enum_cast<Primitive::Type>(i));
734 dex::TypeIndex type_index =
735 profile_info->FindOrCreateTypeIndex(*dex_file, array_class_descriptor.c_str());
736 if (type_index.IsValid()) {
737 profile_info->AddClass(profile_index, type_index);
738 }
739 }
740 }
741 } else {
742 // Error adding dex file to the `profile_info`.
743 }
744 } else {
745 DCHECK(std::all_of(max_primitive_array_dimensions_.begin(),
746 max_primitive_array_dimensions_.end(),
747 [](uint8_t dim) { return dim == 0u; }));
748 }
749
750 // Store the number of hot and sampled methods.
751 number_of_hot_methods_ = number_of_hot_methods;
752 number_of_sampled_methods_ = number_of_sampled_methods;
753 }
754
FetchAndCacheResolvedClassesAndMethods(bool startup)755 void ProfileSaver::FetchAndCacheResolvedClassesAndMethods(bool startup) {
756 ScopedTrace trace(__PRETTY_FUNCTION__);
757 const uint64_t start_time = NanoTime();
758
759 // Resolve any new registered locations.
760 ResolveTrackedLocations();
761
762 Thread* const self = Thread::Current();
763 pthread_t profiler_pthread;
764 {
765 MutexLock mu(self, *Locks::profiler_lock_);
766 profiler_pthread = profiler_pthread_;
767 }
768
769 uint32_t hot_method_sample_threshold = 0u;
770 size_t number_of_hot_methods = 0u;
771 size_t number_of_sampled_methods = 0u;
772 {
773 // Restore profile saver thread priority while holding the mutator lock. This helps
774 // prevent priority inversions blocking the GC for long periods of time.
775 // Only restore default priority if we are the profile saver thread. Other threads
776 // that call this are threads calling Stop and the signal catcher (for SIGUSR1).
777 std::optional<ScopedDefaultPriority> sdp = std::nullopt;
778 if (pthread_self() == profiler_pthread) {
779 sdp.emplace(profiler_pthread);
780 }
781
782 ScopedObjectAccess soa(self);
783 GetClassesAndMethodsHelper helper(startup, options_, GetProfileSampleAnnotation());
784 hot_method_sample_threshold = helper.GetHotMethodSampleThreshold();
785 helper.CollectClasses(self);
786
787 // Release the mutator lock. We shall need to re-acquire the lock for a moment to
788 // destroy the `VariableSizedHandleScope` inside the `helper` which shall be
789 // conveniently handled by destroying `sts`, then `helper` and then `soa`.
790 ScopedThreadSuspension sts(self, ThreadState::kNative);
791 // Get back to the previous thread priority. We shall not increase the priority
792 // for the short time we need to re-acquire mutator lock for `helper` destructor.
793 sdp.reset();
794
795 MutexLock mu(self, *Locks::profiler_lock_);
796 for (const auto& it : tracked_dex_base_locations_) {
797 const std::string& filename = it.first;
798 auto info_it = profile_cache_.find(filename);
799 if (info_it == profile_cache_.end()) {
800 info_it = profile_cache_.Put(
801 filename,
802 new ProfileCompilationInfo(
803 Runtime::Current()->GetArenaPool(), options_.GetProfileBootClassPath()));
804 }
805 ProfileCompilationInfo* cached_info = info_it->second;
806
807 const std::set<std::string>& locations = it.second;
808 VLOG(profiler) << "Locations for " << it.first << " " << android::base::Join(locations, ':');
809 helper.UpdateProfile(locations, cached_info);
810
811 // Update statistics. Note that a method shall be counted for each
812 // tracked location that covers the dex file where it is defined.
813 number_of_hot_methods += helper.GetNumberOfHotMethods();
814 number_of_sampled_methods += helper.GetNumberOfSampledMethods();
815 }
816 }
817 VLOG(profiler) << "Profile saver recorded " << number_of_hot_methods
818 << " hot methods and " << number_of_sampled_methods
819 << " sampled methods with threshold " << hot_method_sample_threshold
820 << " in " << PrettyDuration(NanoTime() - start_time);
821 }
822
ProcessProfilingInfo(bool force_save,bool skip_class_and_method_fetching,uint16_t * number_of_new_methods)823 bool ProfileSaver::ProcessProfilingInfo(
824 bool force_save,
825 bool skip_class_and_method_fetching,
826 /*out*/uint16_t* number_of_new_methods) {
827 ScopedTrace trace(__PRETTY_FUNCTION__);
828
829 // Resolve any new registered locations.
830 ResolveTrackedLocations();
831
832 SafeMap<std::string, std::set<std::string>> tracked_locations;
833 {
834 // Make a copy so that we don't hold the lock while doing I/O.
835 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
836 tracked_locations = tracked_dex_base_locations_;
837 }
838
839 bool profile_file_saved = false;
840 if (number_of_new_methods != nullptr) {
841 *number_of_new_methods = 0;
842 }
843
844 if (!skip_class_and_method_fetching) {
845 // We only need to do this once, not once per dex location.
846 // TODO: Figure out a way to only do it when stuff has changed? It takes 30-50ms.
847 FetchAndCacheResolvedClassesAndMethods(/*startup=*/ false);
848 }
849
850 for (const auto& it : tracked_locations) {
851 if (!force_save && ShuttingDown(Thread::Current())) {
852 // The ProfileSaver is in shutdown mode, meaning a stop request was made and
853 // we need to exit cleanly (by waiting for the saver thread to finish). Unless
854 // we have a request for a forced save, do not do any processing so that we
855 // speed up the exit.
856 return true;
857 }
858 const std::string& filename = it.first;
859 const std::set<std::string>& locations = it.second;
860 VLOG(profiler) << "Tracked filename " << filename << " locations "
861 << android::base::Join(locations, ":");
862
863 std::vector<ProfileMethodInfo> profile_methods;
864 {
865 ScopedObjectAccess soa(Thread::Current());
866 jit_code_cache_->GetProfiledMethods(locations, profile_methods);
867 total_number_of_code_cache_queries_++;
868 }
869 {
870 ProfileCompilationInfo info(Runtime::Current()->GetArenaPool(),
871 /*for_boot_image=*/options_.GetProfileBootClassPath());
872 // Load the existing profile before saving.
873 // If the file is updated between `Load` and `Save`, the update will be lost. This is
874 // acceptable. The main reason is that the lost entries will eventually come back if the user
875 // keeps using the same methods, or they won't be needed if the user doesn't use the same
876 // methods again.
877 if (!info.Load(filename, /*clear_if_invalid=*/true)) {
878 LOG(WARNING) << "Could not forcefully load profile " << filename;
879 continue;
880 }
881
882 uint64_t last_save_number_of_methods = info.GetNumberOfMethods();
883 uint64_t last_save_number_of_classes = info.GetNumberOfResolvedClasses();
884 VLOG(profiler) << "last_save_number_of_methods=" << last_save_number_of_methods
885 << " last_save_number_of_classes=" << last_save_number_of_classes
886 << " number of profiled methods=" << profile_methods.size();
887
888 // Try to add the method data. Note this may fail is the profile loaded from disk contains
889 // outdated data (e.g. the previous profiled dex files might have been updated).
890 // If this happens we clear the profile data and for the save to ensure the file is cleared.
891 if (!info.AddMethods(
892 profile_methods,
893 AnnotateSampleFlags(Hotness::kFlagHot | Hotness::kFlagPostStartup),
894 GetProfileSampleAnnotation())) {
895 LOG(WARNING) << "Could not add methods to the existing profiler. "
896 << "Clearing the profile data.";
897 info.ClearData();
898 force_save = true;
899 }
900
901 {
902 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
903 auto profile_cache_it = profile_cache_.find(filename);
904 if (profile_cache_it != profile_cache_.end()) {
905 if (!info.MergeWith(*(profile_cache_it->second))) {
906 LOG(WARNING) << "Could not merge the profile. Clearing the profile data.";
907 info.ClearData();
908 force_save = true;
909 }
910 } else if (VLOG_IS_ON(profiler)) {
911 LOG(INFO) << "Failed to find cached profile for " << filename;
912 for (auto&& pair : profile_cache_) {
913 LOG(INFO) << "Cached profile " << pair.first;
914 }
915 }
916
917 int64_t delta_number_of_methods =
918 info.GetNumberOfMethods() - last_save_number_of_methods;
919 int64_t delta_number_of_classes =
920 info.GetNumberOfResolvedClasses() - last_save_number_of_classes;
921
922 if (!force_save &&
923 delta_number_of_methods < options_.GetMinMethodsToSave() &&
924 delta_number_of_classes < options_.GetMinClassesToSave()) {
925 VLOG(profiler) << "Not enough information to save to: " << filename
926 << " Number of methods: " << delta_number_of_methods
927 << " Number of classes: " << delta_number_of_classes;
928 total_number_of_skipped_writes_++;
929 continue;
930 }
931
932 if (number_of_new_methods != nullptr) {
933 *number_of_new_methods =
934 std::max(static_cast<uint16_t>(delta_number_of_methods),
935 *number_of_new_methods);
936 }
937 uint64_t bytes_written;
938 // Force the save. In case the profile data is corrupted or the profile
939 // has the wrong version this will "fix" the file to the correct format.
940 if (info.Save(filename, &bytes_written)) {
941 // We managed to save the profile. Clear the cache stored during startup.
942 if (profile_cache_it != profile_cache_.end()) {
943 ProfileCompilationInfo *cached_info = profile_cache_it->second;
944 profile_cache_.erase(profile_cache_it);
945 delete cached_info;
946 }
947 if (bytes_written > 0) {
948 total_number_of_writes_++;
949 total_bytes_written_ += bytes_written;
950 profile_file_saved = true;
951 } else {
952 // At this point we could still have avoided the write.
953 // We load and merge the data from the file lazily at its first ever
954 // save attempt. So, whatever we are trying to save could already be
955 // in the file.
956 total_number_of_skipped_writes_++;
957 }
958 } else {
959 LOG(WARNING) << "Could not save profiling info to " << filename;
960 total_number_of_failed_writes_++;
961 }
962 }
963 }
964 }
965
966 // Trim the maps to madvise the pages used for profile info.
967 // It is unlikely we will need them again in the near feature.
968 Runtime::Current()->GetArenaPool()->TrimMaps();
969
970 return profile_file_saved;
971 }
972
RunProfileSaverThread(void * arg)973 void* ProfileSaver::RunProfileSaverThread(void* arg) {
974 Runtime* runtime = Runtime::Current();
975
976 bool attached = runtime->AttachCurrentThread("Profile Saver",
977 /*as_daemon=*/true,
978 runtime->GetSystemThreadGroup(),
979 /*create_peer=*/true);
980 if (!attached) {
981 CHECK(runtime->IsShuttingDown(Thread::Current()));
982 return nullptr;
983 }
984
985 {
986 Locks::profiler_lock_->ExclusiveLock(Thread::Current());
987 CHECK_EQ(reinterpret_cast<ProfileSaver*>(arg), instance_);
988 instance_->Run();
989 }
990
991 runtime->DetachCurrentThread();
992 VLOG(profiler) << "Profile saver shutdown";
993 return nullptr;
994 }
995
ShouldProfileLocation(const std::string & location,bool profile_aot_code)996 static bool ShouldProfileLocation(const std::string& location, bool profile_aot_code) {
997 if (profile_aot_code) {
998 // If we have to profile all the code, irrespective of its compilation state, return true
999 // right away.
1000 return true;
1001 }
1002
1003 OatFileManager& oat_manager = Runtime::Current()->GetOatFileManager();
1004 const OatFile* oat_file = oat_manager.FindOpenedOatFileFromDexLocation(location);
1005 if (oat_file == nullptr) {
1006 // This can happen if we fallback to run code directly from the APK.
1007 // Profile it with the hope that the background dexopt will get us back into
1008 // a good state.
1009 VLOG(profiler) << "Asked to profile a location without an oat file:" << location;
1010 return true;
1011 }
1012 CompilerFilter::Filter filter = oat_file->GetCompilerFilter();
1013 if ((filter == CompilerFilter::kSpeed) || (filter == CompilerFilter::kEverything)) {
1014 VLOG(profiler)
1015 << "Skip profiling oat file because it's already speed|everything compiled: "
1016 << location << " oat location: " << oat_file->GetLocation();
1017 return false;
1018 }
1019 return true;
1020 }
1021
Start(const ProfileSaverOptions & options,const std::string & output_filename,jit::JitCodeCache * jit_code_cache,const std::vector<std::string> & code_paths,const std::string & ref_profile_filename)1022 void ProfileSaver::Start(const ProfileSaverOptions& options,
1023 const std::string& output_filename,
1024 jit::JitCodeCache* jit_code_cache,
1025 const std::vector<std::string>& code_paths,
1026 const std::string& ref_profile_filename) {
1027 Runtime* const runtime = Runtime::Current();
1028 DCHECK(options.IsEnabled());
1029 DCHECK(runtime->GetJit() != nullptr);
1030 DCHECK(!output_filename.empty());
1031 DCHECK(jit_code_cache != nullptr);
1032
1033 std::vector<std::string> code_paths_to_profile;
1034 for (const std::string& location : code_paths) {
1035 if (ShouldProfileLocation(location, options.GetProfileAOTCode())) {
1036 VLOG(profiler) << "Code path to profile " << location;
1037 code_paths_to_profile.push_back(location);
1038 }
1039 }
1040
1041 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1042 // Support getting profile samples for the boot class path. This will be used to generate the boot
1043 // image profile. The intention is to use this code to generate to boot image but not use it in
1044 // production. b/37966211
1045 if (options.GetProfileBootClassPath()) {
1046 std::set<std::string> code_paths_keys;
1047 for (const std::string& location : code_paths) {
1048 // Use the profile base key for checking file uniqueness (as it is constructed solely based
1049 // on the location and ignores other metadata like origin package).
1050 code_paths_keys.insert(ProfileCompilationInfo::GetProfileDexFileBaseKey(location));
1051 }
1052 for (const DexFile* dex_file : runtime->GetClassLinker()->GetBootClassPath()) {
1053 // Don't check ShouldProfileLocation since the boot class path may be speed compiled.
1054 const std::string& location = dex_file->GetLocation();
1055 const std::string key = ProfileCompilationInfo::GetProfileDexFileBaseKey(location);
1056 VLOG(profiler) << "Registering boot dex file " << location;
1057 if (code_paths_keys.find(key) != code_paths_keys.end()) {
1058 LOG(WARNING) << "Boot class path location key conflicts with code path " << location;
1059 } else if (instance_ == nullptr) {
1060 // Only add the boot class path once since Start may be called multiple times for secondary
1061 // dexes.
1062 // We still do the collision check above. This handles any secondary dexes that conflict
1063 // with the boot class path dex files.
1064 code_paths_to_profile.push_back(location);
1065 }
1066 }
1067 }
1068 if (code_paths_to_profile.empty()) {
1069 VLOG(profiler) << "No code paths should be profiled.";
1070 return;
1071 }
1072
1073 if (instance_ != nullptr) {
1074 // If we already have an instance, make sure it uses the same jit_code_cache.
1075 // This may be called multiple times via Runtime::registerAppInfo (e.g. for
1076 // apps which share the same runtime).
1077 DCHECK_EQ(instance_->jit_code_cache_, jit_code_cache);
1078 // Add the code_paths to the tracked locations.
1079 instance_->AddTrackedLocations(output_filename, code_paths_to_profile, ref_profile_filename);
1080 return;
1081 }
1082
1083 VLOG(profiler) << "Starting profile saver using output file: " << output_filename
1084 << ". Tracking: " << android::base::Join(code_paths_to_profile, ':')
1085 << ". With reference profile: " << ref_profile_filename;
1086
1087 instance_ = new ProfileSaver(options, jit_code_cache);
1088 instance_->AddTrackedLocations(output_filename, code_paths_to_profile, ref_profile_filename);
1089
1090 // Create a new thread which does the saving.
1091 CHECK_PTHREAD_CALL(
1092 pthread_create,
1093 (&profiler_pthread_, nullptr, &RunProfileSaverThread, reinterpret_cast<void*>(instance_)),
1094 "Profile saver thread");
1095
1096 SetProfileSaverThreadPriority(profiler_pthread_, kProfileSaverPthreadPriority);
1097 }
1098
Stop(bool dump_info)1099 void ProfileSaver::Stop(bool dump_info) {
1100 ProfileSaver* profile_saver = nullptr;
1101 pthread_t profiler_pthread = 0U;
1102
1103 {
1104 MutexLock profiler_mutex(Thread::Current(), *Locks::profiler_lock_);
1105 VLOG(profiler) << "Stopping profile saver thread";
1106 profile_saver = instance_;
1107 profiler_pthread = profiler_pthread_;
1108 if (instance_ == nullptr) {
1109 DCHECK(false) << "Tried to stop a profile saver which was not started";
1110 return;
1111 }
1112 if (instance_->shutting_down_) {
1113 DCHECK(false) << "Tried to stop the profile saver twice";
1114 return;
1115 }
1116 instance_->shutting_down_ = true;
1117 }
1118
1119 {
1120 // Wake up the saver thread if it is sleeping to allow for a clean exit.
1121 MutexLock wait_mutex(Thread::Current(), profile_saver->wait_lock_);
1122 profile_saver->period_condition_.Signal(Thread::Current());
1123 }
1124
1125 // Force save everything before destroying the thread since we want profiler_pthread_ to remain
1126 // valid.
1127 profile_saver->ProcessProfilingInfo(
1128 /*force_ save=*/ true,
1129 /*skip_class_and_method_fetching=*/ false,
1130 /*number_of_new_methods=*/ nullptr);
1131
1132 // Wait for the saver thread to stop.
1133 CHECK_PTHREAD_CALL(pthread_join, (profiler_pthread, nullptr), "profile saver thread shutdown");
1134
1135 {
1136 MutexLock profiler_mutex(Thread::Current(), *Locks::profiler_lock_);
1137 if (dump_info) {
1138 instance_->DumpInfo(LOG_STREAM(INFO));
1139 }
1140 instance_ = nullptr;
1141 profiler_pthread_ = 0U;
1142 }
1143 delete profile_saver;
1144 }
1145
ShuttingDown(Thread * self)1146 bool ProfileSaver::ShuttingDown(Thread* self) {
1147 MutexLock mu(self, *Locks::profiler_lock_);
1148 return shutting_down_;
1149 }
1150
IsStarted()1151 bool ProfileSaver::IsStarted() {
1152 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1153 return instance_ != nullptr;
1154 }
1155
AddTrackedLocationsToMap(const std::string & output_filename,const std::vector<std::string> & code_paths,SafeMap<std::string,std::set<std::string>> * map)1156 static void AddTrackedLocationsToMap(const std::string& output_filename,
1157 const std::vector<std::string>& code_paths,
1158 SafeMap<std::string, std::set<std::string>>* map) {
1159 std::vector<std::string> code_paths_and_filenames;
1160 // The dex locations are sometimes set to the filename instead of the full path.
1161 // So make sure we have both "locations" when tracking what needs to be profiled.
1162 // - apps + system server have filenames
1163 // - boot classpath elements have full paths
1164
1165 // TODO(calin, ngeoffray, vmarko) This is an workaround for using filanames as
1166 // dex locations - needed to prebuilt with a partial boot image
1167 // (commit: c4a924d8c74241057d957d360bf31cd5cd0e4f9c).
1168 // We should find a better way which allows us to do the tracking based on full paths.
1169 for (const std::string& path : code_paths) {
1170 size_t last_sep_index = path.find_last_of('/');
1171 if (last_sep_index == path.size() - 1) {
1172 // Should not happen, but anyone can register code paths so better be prepared and ignore
1173 // such locations.
1174 continue;
1175 }
1176 std::string filename = last_sep_index == std::string::npos
1177 ? path
1178 : path.substr(last_sep_index + 1);
1179
1180 code_paths_and_filenames.push_back(path);
1181 code_paths_and_filenames.push_back(filename);
1182 }
1183
1184 auto it = map->find(output_filename);
1185 if (it == map->end()) {
1186 map->Put(
1187 output_filename,
1188 std::set<std::string>(code_paths_and_filenames.begin(), code_paths_and_filenames.end()));
1189 } else {
1190 it->second.insert(code_paths_and_filenames.begin(), code_paths_and_filenames.end());
1191 }
1192 }
1193
AddTrackedLocations(const std::string & output_filename,const std::vector<std::string> & code_paths,const std::string & ref_profile_filename)1194 void ProfileSaver::AddTrackedLocations(const std::string& output_filename,
1195 const std::vector<std::string>& code_paths,
1196 const std::string& ref_profile_filename) {
1197 // Register the output profile and its reference profile.
1198 auto it = tracked_profiles_.find(output_filename);
1199 if (it == tracked_profiles_.end()) {
1200 tracked_profiles_.Put(output_filename, ref_profile_filename);
1201 }
1202
1203 // Add the code paths to the list of tracked location.
1204 AddTrackedLocationsToMap(output_filename, code_paths, &tracked_dex_base_locations_);
1205 // The code paths may contain symlinks which could fool the profiler.
1206 // If the dex file is compiled with an absolute location but loaded with symlink
1207 // the profiler could skip the dex due to location mismatch.
1208 // To avoid this, we add the code paths to the temporary cache of 'to_be_resolved'
1209 // locations. When the profiler thread executes we will resolve the paths to their
1210 // real paths.
1211 // Note that we delay taking the realpath to avoid spending more time than needed
1212 // when registering location (as it is done during app launch).
1213 AddTrackedLocationsToMap(output_filename,
1214 code_paths,
1215 &tracked_dex_base_locations_to_be_resolved_);
1216 }
1217
DumpInstanceInfo(std::ostream & os)1218 void ProfileSaver::DumpInstanceInfo(std::ostream& os) {
1219 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1220 if (instance_ != nullptr) {
1221 instance_->DumpInfo(os);
1222 }
1223 }
1224
DumpInfo(std::ostream & os)1225 void ProfileSaver::DumpInfo(std::ostream& os) {
1226 os << "ProfileSaver total_bytes_written=" << total_bytes_written_ << '\n'
1227 << "ProfileSaver total_number_of_writes=" << total_number_of_writes_ << '\n'
1228 << "ProfileSaver total_number_of_code_cache_queries="
1229 << total_number_of_code_cache_queries_ << '\n'
1230 << "ProfileSaver total_number_of_skipped_writes=" << total_number_of_skipped_writes_ << '\n'
1231 << "ProfileSaver total_number_of_failed_writes=" << total_number_of_failed_writes_ << '\n'
1232 << "ProfileSaver total_ms_of_sleep=" << total_ms_of_sleep_ << '\n'
1233 << "ProfileSaver total_ms_of_work=" << NsToMs(total_ns_of_work_) << '\n'
1234 << "ProfileSaver total_number_of_hot_spikes=" << total_number_of_hot_spikes_ << '\n'
1235 << "ProfileSaver total_number_of_wake_ups=" << total_number_of_wake_ups_ << '\n';
1236 }
1237
1238
ForceProcessProfiles()1239 void ProfileSaver::ForceProcessProfiles() {
1240 ProfileSaver* saver = nullptr;
1241 {
1242 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1243 saver = instance_;
1244 }
1245 // TODO(calin): this is not actually thread safe as the instance_ may have been deleted,
1246 // but we only use this in testing when we now this won't happen.
1247 // Refactor the way we handle the instance so that we don't end up in this situation.
1248 if (saver != nullptr) {
1249 saver->ProcessProfilingInfo(
1250 /*force_save=*/ true,
1251 /*skip_class_and_method_fetching=*/ false,
1252 /*number_of_new_methods=*/ nullptr);
1253 }
1254 }
1255
ResolveTrackedLocations()1256 void ProfileSaver::ResolveTrackedLocations() {
1257 SafeMap<std::string, std::set<std::string>> locations_to_be_resolved;
1258 {
1259 // Make a copy so that we don't hold the lock while doing I/O.
1260 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1261 locations_to_be_resolved = tracked_dex_base_locations_to_be_resolved_;
1262 tracked_dex_base_locations_to_be_resolved_.clear();
1263 }
1264
1265 // Resolve the locations.
1266 SafeMap<std::string, std::vector<std::string>> resolved_locations_map;
1267 for (const auto& it : locations_to_be_resolved) {
1268 const std::string& filename = it.first;
1269 const std::set<std::string>& locations = it.second;
1270 auto resolved_locations_it = resolved_locations_map.Put(
1271 filename,
1272 std::vector<std::string>(locations.size()));
1273
1274 for (const auto& location : locations) {
1275 UniqueCPtr<const char[]> location_real(realpath(location.c_str(), nullptr));
1276 // Note that it's ok if we cannot get the real path.
1277 if (location_real != nullptr) {
1278 resolved_locations_it->second.emplace_back(location_real.get());
1279 }
1280 }
1281 }
1282
1283 // Add the resolved locations to the tracked collection.
1284 MutexLock mu(Thread::Current(), *Locks::profiler_lock_);
1285 for (const auto& it : resolved_locations_map) {
1286 AddTrackedLocationsToMap(it.first, it.second, &tracked_dex_base_locations_);
1287 }
1288 }
1289
GetProfileSampleAnnotation()1290 ProfileCompilationInfo::ProfileSampleAnnotation ProfileSaver::GetProfileSampleAnnotation() {
1291 // Ideally, this would be cached in the ProfileSaver class, when we start the thread.
1292 // However the profile is initialized before the process package name is set and fixing this
1293 // would require unnecessary complex synchronizations.
1294 std::string package_name = Runtime::Current()->GetProcessPackageName();
1295 if (package_name.empty()) {
1296 package_name = "unknown";
1297 }
1298 // We only use annotation for the boot image profiles. Regular apps do not use the extra
1299 // metadata and as such there is no need to pay the cost (storage and computational)
1300 // that comes with the annotations.
1301 return options_.GetProfileBootClassPath()
1302 ? ProfileCompilationInfo::ProfileSampleAnnotation(package_name)
1303 : ProfileCompilationInfo::ProfileSampleAnnotation::kNone;
1304 }
1305
GetExtraMethodHotnessFlags(const ProfileSaverOptions & options)1306 uint32_t ProfileSaver::GetExtraMethodHotnessFlags(const ProfileSaverOptions& options) {
1307 // We only add the extra flags for the boot image profile because individual apps do not use
1308 // this information.
1309 if (options.GetProfileBootClassPath()) {
1310 return Is64BitInstructionSet(Runtime::Current()->GetInstructionSet())
1311 ? Hotness::kFlag64bit
1312 : Hotness::kFlag32bit;
1313 } else {
1314 return 0u;
1315 }
1316 }
1317
AnnotateSampleFlags(uint32_t flags)1318 Hotness::Flag ProfileSaver::AnnotateSampleFlags(uint32_t flags) {
1319 uint32_t extra_flags = GetExtraMethodHotnessFlags(options_);
1320 return static_cast<Hotness::Flag>(flags | extra_flags);
1321 }
1322
1323 } // namespace art
1324