base/tracked_objects.cc

   1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "base/tracked_objects.h"
   6
   7 #include <limits.h>
   8 #include <stdlib.h>
   9
  10 #include "base/atomicops.h"
  11 #include "base/base_switches.h"
  12 #include "base/command_line.h"
  13 #include "base/compiler_specific.h"
  14 #include "base/debug/leak_annotations.h"
  15 #include "base/logging.h"
  16 #include "base/process/process_handle.h"
  17 #include "base/profiler/alternate_timer.h"
  18 #include "base/strings/stringprintf.h"
  19 #include "base/third_party/valgrind/memcheck.h"
  20 #include "base/tracking_info.h"
  21
  22 using base::TimeDelta;
  23
  24 namespace base {
  25 class TimeDelta;
  26 }
  27
  28 namespace tracked_objects {
  29
  30 namespace {
  31 // Flag to compile out almost all of the task tracking code.
  32 const bool kTrackAllTaskObjects = true;
  33
  34 // TODO(jar): Evaluate the perf impact of enabling this.  If the perf impact is
  35 // negligible, enable by default.
  36 // Flag to compile out parent-child link recording.
  37 const bool kTrackParentChildLinks = false;
  38
  39 // When ThreadData is first initialized, should we start in an ACTIVE state to
  40 // record all of the startup-time tasks, or should we start up DEACTIVATED, so
  41 // that we only record after parsing the command line flag --enable-tracking.
  42 // Note that the flag may force either state, so this really controls only the
  43 // period of time up until that flag is parsed. If there is no flag seen, then
  44 // this state may prevail for much or all of the process lifetime.
  45 const ThreadData::Status kInitialStartupState =
  46     ThreadData::PROFILING_CHILDREN_ACTIVE;
  47
  48 // Control whether an alternate time source (Now() function) is supported by
  49 // the ThreadData class.  This compile time flag should be set to true if we
  50 // want other modules (such as a memory allocator, or a thread-specific CPU time
  51 // clock) to be able to provide a thread-specific Now() function.  Without this
  52 // compile-time flag, the code will only support the wall-clock time.  This flag
  53 // can be flipped to efficiently disable this path (if there is a performance
  54 // problem with its presence).
  55 static const bool kAllowAlternateTimeSourceHandling = true;
  56
  57 // Possible states of the profiler timing enabledness.
  58 enum {
  59   UNDEFINED_TIMING,
  60   ENABLED_TIMING,
  61   DISABLED_TIMING,
  62 };
  63
  64 // State of the profiler timing enabledness.
  65 base::subtle::Atomic32 g_profiler_timing_enabled = UNDEFINED_TIMING;
  66
  67 // Returns whether profiler timing is enabled. The default is true, but this may
  68 // be overridden by a command-line flag. Some platforms may programmatically set
  69 // this command-line flag to the "off" value if it's not specified.
  70 // This in turn can be overridden by explicitly calling
  71 // ThreadData::EnableProfilerTiming, say, based on a field trial.
  72 inline bool IsProfilerTimingEnabled() {
  73   // Reading |g_profiler_timing_enabled| is done without barrier because
  74   // multiple initialization is not an issue while the barrier can be relatively
  75   // costly given that this method is sometimes called in a tight loop.
  76   base::subtle::Atomic32 current_timing_enabled =
  77       base::subtle::NoBarrier_Load(&g_profiler_timing_enabled);
  78   if (current_timing_enabled == UNDEFINED_TIMING) {
  79     if (!base::CommandLine::InitializedForCurrentProcess())
  80       return true;
  81     current_timing_enabled =
  82         (base::CommandLine::ForCurrentProcess()->GetSwitchValueASCII(
  83              switches::kProfilerTiming) ==
  84          switches::kProfilerTimingDisabledValue)
  85             ? DISABLED_TIMING
  86             : ENABLED_TIMING;
  87     base::subtle::NoBarrier_Store(&g_profiler_timing_enabled,
  88                                   current_timing_enabled);
  89   }
  90   return current_timing_enabled == ENABLED_TIMING;
  91 }
  92
  93 }  // namespace
  94
  95 //------------------------------------------------------------------------------
  96 // DeathData tallies durations when a death takes place.
  97
  98 DeathData::DeathData() {
  99   Clear();
 100 }
 101
 102 DeathData::DeathData(int count) {
 103   Clear();
 104   count_ = count;
 105 }
 106
 107 // TODO(jar): I need to see if this macro to optimize branching is worth using.
 108 //
 109 // This macro has no branching, so it is surely fast, and is equivalent to:
 110 //             if (assign_it)
 111 //               target = source;
 112 // We use a macro rather than a template to force this to inline.
 113 // Related code for calculating max is discussed on the web.
 114 #define CONDITIONAL_ASSIGN(assign_it, target, source) \
 115     ((target) ^= ((target) ^ (source)) & -static_cast<int32>(assign_it))
 116
 117 void DeathData::RecordDeath(const int32 queue_duration,
 118                             const int32 run_duration,
 119                             const uint32 random_number) {
 120   // We'll just clamp at INT_MAX, but we should note this in the UI as such.
 121   if (count_ < INT_MAX)
 122     ++count_;
 123   queue_duration_sum_ += queue_duration;
 124   run_duration_sum_ += run_duration;
 125
 126   if (queue_duration_max_ < queue_duration)
 127     queue_duration_max_ = queue_duration;
 128   if (run_duration_max_ < run_duration)
 129     run_duration_max_ = run_duration;
 130
 131   // Take a uniformly distributed sample over all durations ever supplied.
 132   // The probability that we (instead) use this new sample is 1/count_.  This
 133   // results in a completely uniform selection of the sample (at least when we
 134   // don't clamp count_... but that should be inconsequentially likely).
 135   // We ignore the fact that we correlated our selection of a sample to the run
 136   // and queue times (i.e., we used them to generate random_number).
 137   CHECK_GT(count_, 0);
 138   if (0 == (random_number % count_)) {
 139     queue_duration_sample_ = queue_duration;
 140     run_duration_sample_ = run_duration;
 141   }
 142 }
 143
 144 int DeathData::count() const { return count_; }
 145
 146 int32 DeathData::run_duration_sum() const { return run_duration_sum_; }
 147
 148 int32 DeathData::run_duration_max() const { return run_duration_max_; }
 149
 150 int32 DeathData::run_duration_sample() const {
 151   return run_duration_sample_;
 152 }
 153
 154 int32 DeathData::queue_duration_sum() const {
 155   return queue_duration_sum_;
 156 }
 157
 158 int32 DeathData::queue_duration_max() const {
 159   return queue_duration_max_;
 160 }
 161
 162 int32 DeathData::queue_duration_sample() const {
 163   return queue_duration_sample_;
 164 }
 165
 166 void DeathData::Clear() {
 167   count_ = 0;
 168   run_duration_sum_ = 0;
 169   run_duration_max_ = 0;
 170   run_duration_sample_ = 0;
 171   queue_duration_sum_ = 0;
 172   queue_duration_max_ = 0;
 173   queue_duration_sample_ = 0;
 174 }
 175
 176 //------------------------------------------------------------------------------
 177 DeathDataSnapshot::DeathDataSnapshot()
 178     : count(-1),
 179       run_duration_sum(-1),
 180       run_duration_max(-1),
 181       run_duration_sample(-1),
 182       queue_duration_sum(-1),
 183       queue_duration_max(-1),
 184       queue_duration_sample(-1) {
 185 }
 186
 187 DeathDataSnapshot::DeathDataSnapshot(
 188     const tracked_objects::DeathData& death_data)
 189     : count(death_data.count()),
 190       run_duration_sum(death_data.run_duration_sum()),
 191       run_duration_max(death_data.run_duration_max()),
 192       run_duration_sample(death_data.run_duration_sample()),
 193       queue_duration_sum(death_data.queue_duration_sum()),
 194       queue_duration_max(death_data.queue_duration_max()),
 195       queue_duration_sample(death_data.queue_duration_sample()) {
 196 }
 197
 198 DeathDataSnapshot::~DeathDataSnapshot() {
 199 }
 200
 201 //------------------------------------------------------------------------------
 202 BirthOnThread::BirthOnThread(const Location& location,
 203                              const ThreadData& current)
 204     : location_(location),
 205       birth_thread_(&current) {
 206 }
 207
 208 //------------------------------------------------------------------------------
 209 BirthOnThreadSnapshot::BirthOnThreadSnapshot() {
 210 }
 211
 212 BirthOnThreadSnapshot::BirthOnThreadSnapshot(
 213     const tracked_objects::BirthOnThread& birth)
 214     : location(birth.location()),
 215       thread_name(birth.birth_thread()->thread_name()) {
 216 }
 217
 218 BirthOnThreadSnapshot::~BirthOnThreadSnapshot() {
 219 }
 220
 221 //------------------------------------------------------------------------------
 222 Births::Births(const Location& location, const ThreadData& current)
 223     : BirthOnThread(location, current),
 224       birth_count_(1) { }
 225
 226 int Births::birth_count() const { return birth_count_; }
 227
 228 void Births::RecordBirth() { ++birth_count_; }
 229
 230 //------------------------------------------------------------------------------
 231 // ThreadData maintains the central data for all births and deaths on a single
 232 // thread.
 233
 234 // TODO(jar): We should pull all these static vars together, into a struct, and
 235 // optimize layout so that we benefit from locality of reference during accesses
 236 // to them.
 237
 238 // static
 239 NowFunction* ThreadData::now_function_ = NULL;
 240
 241 // static
 242 bool ThreadData::now_function_is_time_ = false;
 243
 244 // A TLS slot which points to the ThreadData instance for the current thread. We
 245 // do a fake initialization here (zeroing out data), and then the real in-place
 246 // construction happens when we call tls_index_.Initialize().
 247 // static
 248 base::ThreadLocalStorage::StaticSlot ThreadData::tls_index_ = TLS_INITIALIZER;
 249
 250 // static
 251 int ThreadData::worker_thread_data_creation_count_ = 0;
 252
 253 // static
 254 int ThreadData::cleanup_count_ = 0;
 255
 256 // static
 257 int ThreadData::incarnation_counter_ = 0;
 258
 259 // static
 260 ThreadData* ThreadData::all_thread_data_list_head_ = NULL;
 261
 262 // static
 263 ThreadData* ThreadData::first_retired_worker_ = NULL;
 264
 265 // static
 266 base::LazyInstance<base::Lock>::Leaky
 267     ThreadData::list_lock_ = LAZY_INSTANCE_INITIALIZER;
 268
 269 // static
 270 ThreadData::Status ThreadData::status_ = ThreadData::UNINITIALIZED;
 271
 272 ThreadData::ThreadData(const std::string& suggested_name)
 273     : next_(NULL),
 274       next_retired_worker_(NULL),
 275       worker_thread_number_(0),
 276       incarnation_count_for_pool_(-1),
 277       current_stopwatch_(NULL) {
 278   DCHECK_GE(suggested_name.size(), 0u);
 279   thread_name_ = suggested_name;
 280   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 281 }
 282
 283 ThreadData::ThreadData(int thread_number)
 284     : next_(NULL),
 285       next_retired_worker_(NULL),
 286       worker_thread_number_(thread_number),
 287       incarnation_count_for_pool_(-1),
 288       current_stopwatch_(NULL) {
 289   CHECK_GT(thread_number, 0);
 290   base::StringAppendF(&thread_name_, "WorkerThread-%d", thread_number);
 291   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 292 }
 293
 294 ThreadData::~ThreadData() {}
 295
 296 void ThreadData::PushToHeadOfList() {
 297   // Toss in a hint of randomness (atop the uniniitalized value).
 298   (void)VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE(&random_number_,
 299                                                  sizeof(random_number_));
 300   MSAN_UNPOISON(&random_number_, sizeof(random_number_));
 301   random_number_ += static_cast<uint32>(this - static_cast<ThreadData*>(0));
 302   random_number_ ^= (Now() - TrackedTime()).InMilliseconds();
 303
 304   DCHECK(!next_);
 305   base::AutoLock lock(*list_lock_.Pointer());
 306   incarnation_count_for_pool_ = incarnation_counter_;
 307   next_ = all_thread_data_list_head_;
 308   all_thread_data_list_head_ = this;
 309 }
 310
 311 // static
 312 ThreadData* ThreadData::first() {
 313   base::AutoLock lock(*list_lock_.Pointer());
 314   return all_thread_data_list_head_;
 315 }
 316
 317 ThreadData* ThreadData::next() const { return next_; }
 318
 319 // static
 320 void ThreadData::InitializeThreadContext(const std::string& suggested_name) {
 321   if (!Initialize())  // Always initialize if needed.
 322     return;
 323   ThreadData* current_thread_data =
 324       reinterpret_cast<ThreadData*>(tls_index_.Get());
 325   if (current_thread_data)
 326     return;  // Browser tests instigate this.
 327   current_thread_data = new ThreadData(suggested_name);
 328   tls_index_.Set(current_thread_data);
 329 }
 330
 331 // static
 332 ThreadData* ThreadData::Get() {
 333   if (!tls_index_.initialized())
 334     return NULL;  // For unittests only.
 335   ThreadData* registered = reinterpret_cast<ThreadData*>(tls_index_.Get());
 336   if (registered)
 337     return registered;
 338
 339   // We must be a worker thread, since we didn't pre-register.
 340   ThreadData* worker_thread_data = NULL;
 341   int worker_thread_number = 0;
 342   {
 343     base::AutoLock lock(*list_lock_.Pointer());
 344     if (first_retired_worker_) {
 345       worker_thread_data = first_retired_worker_;
 346       first_retired_worker_ = first_retired_worker_->next_retired_worker_;
 347       worker_thread_data->next_retired_worker_ = NULL;
 348     } else {
 349       worker_thread_number = ++worker_thread_data_creation_count_;
 350     }
 351   }
 352
 353   // If we can't find a previously used instance, then we have to create one.
 354   if (!worker_thread_data) {
 355     DCHECK_GT(worker_thread_number, 0);
 356     worker_thread_data = new ThreadData(worker_thread_number);
 357   }
 358   DCHECK_GT(worker_thread_data->worker_thread_number_, 0);
 359
 360   tls_index_.Set(worker_thread_data);
 361   return worker_thread_data;
 362 }
 363
 364 // static
 365 void ThreadData::OnThreadTermination(void* thread_data) {
 366   DCHECK(thread_data);  // TLS should *never* call us with a NULL.
 367   // We must NOT do any allocations during this callback. There is a chance
 368   // that the allocator is no longer active on this thread.
 369   if (!kTrackAllTaskObjects)
 370     return;  // Not compiled in.
 371   reinterpret_cast<ThreadData*>(thread_data)->OnThreadTerminationCleanup();
 372 }
 373
 374 void ThreadData::OnThreadTerminationCleanup() {
 375   // The list_lock_ was created when we registered the callback, so it won't be
 376   // allocated here despite the lazy reference.
 377   base::AutoLock lock(*list_lock_.Pointer());
 378   if (incarnation_counter_ != incarnation_count_for_pool_)
 379     return;  // ThreadData was constructed in an earlier unit test.
 380   ++cleanup_count_;
 381   // Only worker threads need to be retired and reused.
 382   if (!worker_thread_number_) {
 383     return;
 384   }
 385   // We must NOT do any allocations during this callback.
 386   // Using the simple linked lists avoids all allocations.
 387   DCHECK_EQ(this->next_retired_worker_, reinterpret_cast<ThreadData*>(NULL));
 388   this->next_retired_worker_ = first_retired_worker_;
 389   first_retired_worker_ = this;
 390 }
 391
 392 // static
 393 void ThreadData::Snapshot(ProcessDataSnapshot* process_data_snapshot) {
 394   ThreadData::SnapshotCurrentPhase(
 395       &process_data_snapshot->phased_process_data_snapshots[0]);
 396 }
 397
 398 Births* ThreadData::TallyABirth(const Location& location) {
 399   BirthMap::iterator it = birth_map_.find(location);
 400   Births* child;
 401   if (it != birth_map_.end()) {
 402     child =  it->second;
 403     child->RecordBirth();
 404   } else {
 405     child = new Births(location, *this);  // Leak this.
 406     // Lock since the map may get relocated now, and other threads sometimes
 407     // snapshot it (but they lock before copying it).
 408     base::AutoLock lock(map_lock_);
 409     birth_map_[location] = child;
 410   }
 411
 412   if (kTrackParentChildLinks && status_ > PROFILING_ACTIVE &&
 413       !parent_stack_.empty()) {
 414     const Births* parent = parent_stack_.top();
 415     ParentChildPair pair(parent, child);
 416     if (parent_child_set_.find(pair) == parent_child_set_.end()) {
 417       // Lock since the map may get relocated now, and other threads sometimes
 418       // snapshot it (but they lock before copying it).
 419       base::AutoLock lock(map_lock_);
 420       parent_child_set_.insert(pair);
 421     }
 422   }
 423
 424   return child;
 425 }
 426
 427 void ThreadData::TallyADeath(const Births& birth,
 428                              int32 queue_duration,
 429                              const TaskStopwatch& stopwatch) {
 430   int32 run_duration = stopwatch.RunDurationMs();
 431
 432   // Stir in some randomness, plus add constant in case durations are zero.
 433   const uint32 kSomePrimeNumber = 2147483647;
 434   random_number_ += queue_duration + run_duration + kSomePrimeNumber;
 435   // An address is going to have some randomness to it as well ;-).
 436   random_number_ ^= static_cast<uint32>(&birth - reinterpret_cast<Births*>(0));
 437
 438   // We don't have queue durations without OS timer. OS timer is automatically
 439   // used for task-post-timing, so the use of an alternate timer implies all
 440   // queue times are invalid, unless it was explicitly said that we can trust
 441   // the alternate timer.
 442   if (kAllowAlternateTimeSourceHandling &&
 443       now_function_ &&
 444       !now_function_is_time_) {
 445     queue_duration = 0;
 446   }
 447
 448   DeathMap::iterator it = death_map_.find(&birth);
 449   DeathData* death_data;
 450   if (it != death_map_.end()) {
 451     death_data = &it->second;
 452   } else {
 453     base::AutoLock lock(map_lock_);  // Lock as the map may get relocated now.
 454     death_data = &death_map_[&birth];
 455   }  // Release lock ASAP.
 456   death_data->RecordDeath(queue_duration, run_duration, random_number_);
 457
 458   if (!kTrackParentChildLinks)
 459     return;
 460   if (!parent_stack_.empty()) {  // We might get turned off.
 461     DCHECK_EQ(parent_stack_.top(), &birth);
 462     parent_stack_.pop();
 463   }
 464 }
 465
 466 // static
 467 Births* ThreadData::TallyABirthIfActive(const Location& location) {
 468   if (!kTrackAllTaskObjects)
 469     return NULL;  // Not compiled in.
 470
 471   if (!TrackingStatus())
 472     return NULL;
 473   ThreadData* current_thread_data = Get();
 474   if (!current_thread_data)
 475     return NULL;
 476   return current_thread_data->TallyABirth(location);
 477 }
 478
 479 // static
 480 void ThreadData::TallyRunOnNamedThreadIfTracking(
 481     const base::TrackingInfo& completed_task,
 482     const TaskStopwatch& stopwatch) {
 483   if (!kTrackAllTaskObjects)
 484     return;  // Not compiled in.
 485
 486   // Even if we have been DEACTIVATED, we will process any pending births so
 487   // that our data structures (which counted the outstanding births) remain
 488   // consistent.
 489   const Births* birth = completed_task.birth_tally;
 490   if (!birth)
 491     return;
 492   ThreadData* current_thread_data = stopwatch.GetThreadData();
 493   if (!current_thread_data)
 494     return;
 495
 496   // Watch out for a race where status_ is changing, and hence one or both
 497   // of start_of_run or end_of_run is zero.  In that case, we didn't bother to
 498   // get a time value since we "weren't tracking" and we were trying to be
 499   // efficient by not calling for a genuine time value. For simplicity, we'll
 500   // use a default zero duration when we can't calculate a true value.
 501   TrackedTime start_of_run = stopwatch.StartTime();
 502   int32 queue_duration = 0;
 503   if (!start_of_run.is_null()) {
 504     queue_duration = (start_of_run - completed_task.EffectiveTimePosted())
 505         .InMilliseconds();
 506   }
 507   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 508 }
 509
 510 // static
 511 void ThreadData::TallyRunOnWorkerThreadIfTracking(
 512     const Births* birth,
 513     const TrackedTime& time_posted,
 514     const TaskStopwatch& stopwatch) {
 515   if (!kTrackAllTaskObjects)
 516     return;  // Not compiled in.
 517
 518   // Even if we have been DEACTIVATED, we will process any pending births so
 519   // that our data structures (which counted the outstanding births) remain
 520   // consistent.
 521   if (!birth)
 522     return;
 523
 524   // TODO(jar): Support the option to coalesce all worker-thread activity under
 525   // one ThreadData instance that uses locks to protect *all* access.  This will
 526   // reduce memory (making it provably bounded), but run incrementally slower
 527   // (since we'll use locks on TallyABirth and TallyADeath).  The good news is
 528   // that the locks on TallyADeath will be *after* the worker thread has run,
 529   // and hence nothing will be waiting for the completion (... besides some
 530   // other thread that might like to run).  Also, the worker threads tasks are
 531   // generally longer, and hence the cost of the lock may perchance be amortized
 532   // over the long task's lifetime.
 533   ThreadData* current_thread_data = stopwatch.GetThreadData();
 534   if (!current_thread_data)
 535     return;
 536
 537   TrackedTime start_of_run = stopwatch.StartTime();
 538   int32 queue_duration = 0;
 539   if (!start_of_run.is_null()) {
 540     queue_duration = (start_of_run - time_posted).InMilliseconds();
 541   }
 542   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 543 }
 544
 545 // static
 546 void ThreadData::TallyRunInAScopedRegionIfTracking(
 547     const Births* birth,
 548     const TaskStopwatch& stopwatch) {
 549   if (!kTrackAllTaskObjects)
 550     return;  // Not compiled in.
 551
 552   // Even if we have been DEACTIVATED, we will process any pending births so
 553   // that our data structures (which counted the outstanding births) remain
 554   // consistent.
 555   if (!birth)
 556     return;
 557
 558   ThreadData* current_thread_data = stopwatch.GetThreadData();
 559   if (!current_thread_data)
 560     return;
 561
 562   int32 queue_duration = 0;
 563   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 564 }
 565
 566 // static
 567 void ThreadData::SnapshotAllExecutedTasks(
 568     ProcessDataPhaseSnapshot* process_data_phase,
 569     BirthCountMap* birth_counts) {
 570   if (!kTrackAllTaskObjects)
 571     return;  // Not compiled in.
 572
 573   // Get an unchanging copy of a ThreadData list.
 574   ThreadData* my_list = ThreadData::first();
 575
 576   // Gather data serially.
 577   // This hackish approach *can* get some slighly corrupt tallies, as we are
 578   // grabbing values without the protection of a lock, but it has the advantage
 579   // of working even with threads that don't have message loops.  If a user
 580   // sees any strangeness, they can always just run their stats gathering a
 581   // second time.
 582   for (ThreadData* thread_data = my_list;
 583        thread_data;
 584        thread_data = thread_data->next()) {
 585     thread_data->SnapshotExecutedTasks(process_data_phase, birth_counts);
 586   }
 587 }
 588
 589 // static
 590 void ThreadData::SnapshotCurrentPhase(
 591     ProcessDataPhaseSnapshot* process_data_phase) {
 592   // Add births that have run to completion to |collected_data|.
 593   // |birth_counts| tracks the total number of births recorded at each location
 594   // for which we have not seen a death count.
 595   BirthCountMap birth_counts;
 596   ThreadData::SnapshotAllExecutedTasks(process_data_phase, &birth_counts);
 597
 598   // Add births that are still active -- i.e. objects that have tallied a birth,
 599   // but have not yet tallied a matching death, and hence must be either
 600   // running, queued up, or being held in limbo for future posting.
 601   for (const auto& birth_count : birth_counts) {
 602     if (birth_count.second > 0) {
 603       process_data_phase->tasks.push_back(TaskSnapshot(
 604           *birth_count.first, DeathData(birth_count.second), "Still_Alive"));
 605     }
 606   }
 607 }
 608
 609 void ThreadData::SnapshotExecutedTasks(
 610     ProcessDataPhaseSnapshot* process_data_phase,
 611     BirthCountMap* birth_counts) {
 612   // Get copy of data, so that the data will not change during the iterations
 613   // and processing.
 614   ThreadData::BirthMap birth_map;
 615   ThreadData::DeathMap death_map;
 616   ThreadData::ParentChildSet parent_child_set;
 617   SnapshotMaps(&birth_map, &death_map, &parent_child_set);
 618
 619   for (const auto& death : death_map) {
 620     process_data_phase->tasks.push_back(
 621         TaskSnapshot(*death.first, death.second, thread_name()));
 622     (*birth_counts)[death.first] -= death.first->birth_count();
 623   }
 624
 625   for (const auto& birth : birth_map) {
 626     (*birth_counts)[birth.second] += birth.second->birth_count();
 627   }
 628
 629   if (!kTrackParentChildLinks)
 630     return;
 631
 632   for (const auto& parent_child : parent_child_set) {
 633     process_data_phase->descendants.push_back(
 634         ParentChildPairSnapshot(parent_child));
 635   }
 636 }
 637
 638 // This may be called from another thread.
 639 void ThreadData::SnapshotMaps(BirthMap* birth_map,
 640                               DeathMap* death_map,
 641                               ParentChildSet* parent_child_set) {
 642   base::AutoLock lock(map_lock_);
 643   for (const auto& birth : birth_map_)
 644     (*birth_map)[birth.first] = birth.second;
 645   for (const auto& death : death_map_)
 646     (*death_map)[death.first] = death.second;
 647
 648   if (!kTrackParentChildLinks)
 649     return;
 650
 651   for (const auto& parent_child : parent_child_set_)
 652     parent_child_set->insert(parent_child);
 653 }
 654
 655 static void OptionallyInitializeAlternateTimer() {
 656   NowFunction* alternate_time_source = GetAlternateTimeSource();
 657   if (alternate_time_source)
 658     ThreadData::SetAlternateTimeSource(alternate_time_source);
 659 }
 660
 661 bool ThreadData::Initialize() {
 662   if (!kTrackAllTaskObjects)
 663     return false;  // Not compiled in.
 664   if (status_ >= DEACTIVATED)
 665     return true;  // Someone else did the initialization.
 666   // Due to racy lazy initialization in tests, we'll need to recheck status_
 667   // after we acquire the lock.
 668
 669   // Ensure that we don't double initialize tls.  We are called when single
 670   // threaded in the product, but some tests may be racy and lazy about our
 671   // initialization.
 672   base::AutoLock lock(*list_lock_.Pointer());
 673   if (status_ >= DEACTIVATED)
 674     return true;  // Someone raced in here and beat us.
 675
 676   // Put an alternate timer in place if the environment calls for it, such as
 677   // for tracking TCMalloc allocations.  This insertion is idempotent, so we
 678   // don't mind if there is a race, and we'd prefer not to be in a lock while
 679   // doing this work.
 680   if (kAllowAlternateTimeSourceHandling)
 681     OptionallyInitializeAlternateTimer();
 682
 683   // Perform the "real" TLS initialization now, and leave it intact through
 684   // process termination.
 685   if (!tls_index_.initialized()) {  // Testing may have initialized this.
 686     DCHECK_EQ(status_, UNINITIALIZED);
 687     tls_index_.Initialize(&ThreadData::OnThreadTermination);
 688     if (!tls_index_.initialized())
 689       return false;
 690   } else {
 691     // TLS was initialzed for us earlier.
 692     DCHECK_EQ(status_, DORMANT_DURING_TESTS);
 693   }
 694
 695   // Incarnation counter is only significant to testing, as it otherwise will
 696   // never again change in this process.
 697   ++incarnation_counter_;
 698
 699   // The lock is not critical for setting status_, but it doesn't hurt. It also
 700   // ensures that if we have a racy initialization, that we'll bail as soon as
 701   // we get the lock earlier in this method.
 702   status_ = kInitialStartupState;
 703   if (!kTrackParentChildLinks &&
 704       kInitialStartupState == PROFILING_CHILDREN_ACTIVE)
 705     status_ = PROFILING_ACTIVE;
 706   DCHECK(status_ != UNINITIALIZED);
 707   return true;
 708 }
 709
 710 // static
 711 bool ThreadData::InitializeAndSetTrackingStatus(Status status) {
 712   DCHECK_GE(status, DEACTIVATED);
 713   DCHECK_LE(status, PROFILING_CHILDREN_ACTIVE);
 714
 715   if (!Initialize())  // No-op if already initialized.
 716     return false;  // Not compiled in.
 717
 718   if (!kTrackParentChildLinks && status > DEACTIVATED)
 719     status = PROFILING_ACTIVE;
 720   status_ = status;
 721   return true;
 722 }
 723
 724 // static
 725 ThreadData::Status ThreadData::status() {
 726   return status_;
 727 }
 728
 729 // static
 730 bool ThreadData::TrackingStatus() {
 731   return status_ > DEACTIVATED;
 732 }
 733
 734 // static
 735 bool ThreadData::TrackingParentChildStatus() {
 736   return status_ >= PROFILING_CHILDREN_ACTIVE;
 737 }
 738
 739 // static
 740 void ThreadData::PrepareForStartOfRun(const Births* parent) {
 741   if (kTrackParentChildLinks && parent && status_ > PROFILING_ACTIVE) {
 742     ThreadData* current_thread_data = Get();
 743     if (current_thread_data)
 744       current_thread_data->parent_stack_.push(parent);
 745   }
 746 }
 747
 748 // static
 749 void ThreadData::SetAlternateTimeSource(NowFunction* now_function) {
 750   DCHECK(now_function);
 751   if (kAllowAlternateTimeSourceHandling)
 752     now_function_ = now_function;
 753 }
 754
 755 // static
 756 void ThreadData::EnableProfilerTiming() {
 757   base::subtle::NoBarrier_Store(&g_profiler_timing_enabled, ENABLED_TIMING);
 758 }
 759
 760 // static
 761 TrackedTime ThreadData::Now() {
 762   if (kAllowAlternateTimeSourceHandling && now_function_)
 763     return TrackedTime::FromMilliseconds((*now_function_)());
 764   if (kTrackAllTaskObjects && IsProfilerTimingEnabled() && TrackingStatus())
 765     return TrackedTime::Now();
 766   return TrackedTime();  // Super fast when disabled, or not compiled.
 767 }
 768
 769 // static
 770 void ThreadData::EnsureCleanupWasCalled(int major_threads_shutdown_count) {
 771   base::AutoLock lock(*list_lock_.Pointer());
 772   if (worker_thread_data_creation_count_ == 0)
 773     return;  // We haven't really run much, and couldn't have leaked.
 774
 775   // TODO(jar): until this is working on XP, don't run the real test.
 776 #if 0
 777   // Verify that we've at least shutdown/cleanup the major namesd threads.  The
 778   // caller should tell us how many thread shutdowns should have taken place by
 779   // now.
 780   CHECK_GT(cleanup_count_, major_threads_shutdown_count);
 781 #endif
 782 }
 783
 784 // static
 785 void ThreadData::ShutdownSingleThreadedCleanup(bool leak) {
 786   // This is only called from test code, where we need to cleanup so that
 787   // additional tests can be run.
 788   // We must be single threaded... but be careful anyway.
 789   if (!InitializeAndSetTrackingStatus(DEACTIVATED))
 790     return;
 791   ThreadData* thread_data_list;
 792   {
 793     base::AutoLock lock(*list_lock_.Pointer());
 794     thread_data_list = all_thread_data_list_head_;
 795     all_thread_data_list_head_ = NULL;
 796     ++incarnation_counter_;
 797     // To be clean, break apart the retired worker list (though we leak them).
 798     while (first_retired_worker_) {
 799       ThreadData* worker = first_retired_worker_;
 800       CHECK_GT(worker->worker_thread_number_, 0);
 801       first_retired_worker_ = worker->next_retired_worker_;
 802       worker->next_retired_worker_ = NULL;
 803     }
 804   }
 805
 806   // Put most global static back in pristine shape.
 807   worker_thread_data_creation_count_ = 0;
 808   cleanup_count_ = 0;
 809   tls_index_.Set(NULL);
 810   status_ = DORMANT_DURING_TESTS;  // Almost UNINITIALIZED.
 811
 812   // To avoid any chance of racing in unit tests, which is the only place we
 813   // call this function, we may sometimes leak all the data structures we
 814   // recovered, as they may still be in use on threads from prior tests!
 815   if (leak) {
 816     ThreadData* thread_data = thread_data_list;
 817     while (thread_data) {
 818       ANNOTATE_LEAKING_OBJECT_PTR(thread_data);
 819       thread_data = thread_data->next();
 820     }
 821     return;
 822   }
 823
 824   // When we want to cleanup (on a single thread), here is what we do.
 825
 826   // Do actual recursive delete in all ThreadData instances.
 827   while (thread_data_list) {
 828     ThreadData* next_thread_data = thread_data_list;
 829     thread_data_list = thread_data_list->next();
 830
 831     for (BirthMap::iterator it = next_thread_data->birth_map_.begin();
 832          next_thread_data->birth_map_.end() != it; ++it)
 833       delete it->second;  // Delete the Birth Records.
 834     delete next_thread_data;  // Includes all Death Records.
 835   }
 836 }
 837
 838 //------------------------------------------------------------------------------
 839 TaskStopwatch::TaskStopwatch()
 840     : wallclock_duration_ms_(0),
 841       current_thread_data_(NULL),
 842       excluded_duration_ms_(0),
 843       parent_(NULL) {
 844 #if DCHECK_IS_ON()
 845   state_ = CREATED;
 846   child_ = NULL;
 847 #endif
 848 }
 849
 850 TaskStopwatch::~TaskStopwatch() {
 851 #if DCHECK_IS_ON()
 852   DCHECK(state_ != RUNNING);
 853   DCHECK(child_ == NULL);
 854 #endif
 855 }
 856
 857 void TaskStopwatch::Start() {
 858 #if DCHECK_IS_ON()
 859   DCHECK(state_ == CREATED);
 860   state_ = RUNNING;
 861 #endif
 862
 863   start_time_ = ThreadData::Now();
 864
 865   current_thread_data_ = ThreadData::Get();
 866   if (!current_thread_data_)
 867     return;
 868
 869   parent_ = current_thread_data_->current_stopwatch_;
 870 #if DCHECK_IS_ON()
 871   if (parent_) {
 872     DCHECK(parent_->state_ == RUNNING);
 873     DCHECK(parent_->child_ == NULL);
 874     parent_->child_ = this;
 875   }
 876 #endif
 877   current_thread_data_->current_stopwatch_ = this;
 878 }
 879
 880 void TaskStopwatch::Stop() {
 881   const TrackedTime end_time = ThreadData::Now();
 882 #if DCHECK_IS_ON()
 883   DCHECK(state_ == RUNNING);
 884   state_ = STOPPED;
 885   DCHECK(child_ == NULL);
 886 #endif
 887
 888   if (!start_time_.is_null() && !end_time.is_null()) {
 889     wallclock_duration_ms_ = (end_time - start_time_).InMilliseconds();
 890   }
 891
 892   if (!current_thread_data_)
 893     return;
 894
 895   DCHECK(current_thread_data_->current_stopwatch_ == this);
 896   current_thread_data_->current_stopwatch_ = parent_;
 897   if (!parent_)
 898     return;
 899
 900 #if DCHECK_IS_ON()
 901   DCHECK(parent_->state_ == RUNNING);
 902   DCHECK(parent_->child_ == this);
 903   parent_->child_ = NULL;
 904 #endif
 905   parent_->excluded_duration_ms_ += wallclock_duration_ms_;
 906   parent_ = NULL;
 907 }
 908
 909 TrackedTime TaskStopwatch::StartTime() const {
 910 #if DCHECK_IS_ON()
 911   DCHECK(state_ != CREATED);
 912 #endif
 913
 914   return start_time_;
 915 }
 916
 917 int32 TaskStopwatch::RunDurationMs() const {
 918 #if DCHECK_IS_ON()
 919   DCHECK(state_ == STOPPED);
 920 #endif
 921
 922   return wallclock_duration_ms_ - excluded_duration_ms_;
 923 }
 924
 925 ThreadData* TaskStopwatch::GetThreadData() const {
 926 #if DCHECK_IS_ON()
 927   DCHECK(state_ != CREATED);
 928 #endif
 929
 930   return current_thread_data_;
 931 }
 932
 933 //------------------------------------------------------------------------------
 934 TaskSnapshot::TaskSnapshot() {
 935 }
 936
 937 TaskSnapshot::TaskSnapshot(const BirthOnThread& birth,
 938                            const DeathData& death_data,
 939                            const std::string& death_thread_name)
 940     : birth(birth),
 941       death_data(death_data),
 942       death_thread_name(death_thread_name) {
 943 }
 944
 945 TaskSnapshot::~TaskSnapshot() {
 946 }
 947
 948 //------------------------------------------------------------------------------
 949 // ParentChildPairSnapshot
 950
 951 ParentChildPairSnapshot::ParentChildPairSnapshot() {
 952 }
 953
 954 ParentChildPairSnapshot::ParentChildPairSnapshot(
 955     const ThreadData::ParentChildPair& parent_child)
 956     : parent(*parent_child.first),
 957       child(*parent_child.second) {
 958 }
 959
 960 ParentChildPairSnapshot::~ParentChildPairSnapshot() {
 961 }
 962
 963 //------------------------------------------------------------------------------
 964 // ProcessDataPhaseSnapshot
 965
 966 ProcessDataPhaseSnapshot::ProcessDataPhaseSnapshot() {
 967 }
 968
 969 ProcessDataPhaseSnapshot::~ProcessDataPhaseSnapshot() {
 970 }
 971
 972 //------------------------------------------------------------------------------
 973 // ProcessDataPhaseSnapshot
 974
 975 ProcessDataSnapshot::ProcessDataSnapshot()
 976 #if !defined(OS_NACL)
 977     : process_id(base::GetCurrentProcId()) {
 978 #else
 979     : process_id(base::kNullProcessId) {
 980 #endif
 981 }
 982
 983 ProcessDataSnapshot::~ProcessDataSnapshot() {
 984 }
 985
 986 }  // namespace tracked_objects