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 <math.h>
   8 #include <stdlib.h>
   9
  10 #include "base/format_macros.h"
  11 #include "base/memory/scoped_ptr.h"
  12 #include "base/process_util.h"
  13 #include "base/profiler/alternate_timer.h"
  14 #include "base/stringprintf.h"
  15 #include "base/third_party/valgrind/memcheck.h"
  16 #include "base/threading/thread_restrictions.h"
  17 #include "base/port.h"
  18
  19 using base::TimeDelta;
  20
  21 namespace tracked_objects {
  22
  23 namespace {
  24 // Flag to compile out almost all of the task tracking code.
  25 const bool kTrackAllTaskObjects = true;
  26
  27 // TODO(jar): Evaluate the perf impact of enabling this.  If the perf impact is
  28 // negligible, enable by default.
  29 // Flag to compile out parent-child link recording.
  30 const bool kTrackParentChildLinks = false;
  31
  32 // When ThreadData is first initialized, should we start in an ACTIVE state to
  33 // record all of the startup-time tasks, or should we start up DEACTIVATED, so
  34 // that we only record after parsing the command line flag --enable-tracking.
  35 // Note that the flag may force either state, so this really controls only the
  36 // period of time up until that flag is parsed. If there is no flag seen, then
  37 // this state may prevail for much or all of the process lifetime.
  38 const ThreadData::Status kInitialStartupState =
  39     ThreadData::PROFILING_CHILDREN_ACTIVE;
  40
  41 // Control whether an alternate time source (Now() function) is supported by
  42 // the ThreadData class.  This compile time flag should be set to true if we
  43 // want other modules (such as a memory allocator, or a thread-specific CPU time
  44 // clock) to be able to provide a thread-specific Now() function.  Without this
  45 // compile-time flag, the code will only support the wall-clock time.  This flag
  46 // can be flipped to efficiently disable this path (if there is a performance
  47 // problem with its presence).
  48 static const bool kAllowAlternateTimeSourceHandling = true;
  49
  50 }  // namespace
  51
  52 //------------------------------------------------------------------------------
  53 // DeathData tallies durations when a death takes place.
  54
  55 DeathData::DeathData() {
  56   Clear();
  57 }
  58
  59 DeathData::DeathData(int count) {
  60   Clear();
  61   count_ = count;
  62 }
  63
  64 // TODO(jar): I need to see if this macro to optimize branching is worth using.
  65 //
  66 // This macro has no branching, so it is surely fast, and is equivalent to:
  67 //             if (assign_it)
  68 //               target = source;
  69 // We use a macro rather than a template to force this to inline.
  70 // Related code for calculating max is discussed on the web.
  71 #define CONDITIONAL_ASSIGN(assign_it, target, source) \
  72     ((target) ^= ((target) ^ (source)) & -static_cast<int32>(assign_it))
  73
  74 void DeathData::RecordDeath(const int32 queue_duration,
  75                             const int32 run_duration,
  76                             int32 random_number) {
  77   ++count_;
  78   queue_duration_sum_ += queue_duration;
  79   run_duration_sum_ += run_duration;
  80
  81   if (queue_duration_max_ < queue_duration)
  82     queue_duration_max_ = queue_duration;
  83   if (run_duration_max_ < run_duration)
  84     run_duration_max_ = run_duration;
  85
  86   // Take a uniformly distributed sample over all durations ever supplied.
  87   // The probability that we (instead) use this new sample is 1/count_.  This
  88   // results in a completely uniform selection of the sample (at least when we
  89   // don't clamp count_... but that should be inconsequentially likely).
  90   // We ignore the fact that we correlated our selection of a sample to the run
  91   // and queue times (i.e., we used them to generate random_number).
  92   if (count_ <= 0) {  // Handle wrapping of count_, such as in bug 138961.
  93     CHECK_GE(count_ - 1, 0);  // Detect memory corruption.
  94     // We'll just clamp at INT_MAX, but we should note this in the UI as such.
  95     count_ = INT_MAX;
  96   }
  97   if (0 == (random_number % count_)) {
  98     queue_duration_sample_ = queue_duration;
  99     run_duration_sample_ = run_duration;
 100   }
 101 }
 102
 103 int DeathData::count() const { return count_; }
 104
 105 int32 DeathData::run_duration_sum() const { return run_duration_sum_; }
 106
 107 int32 DeathData::run_duration_max() const { return run_duration_max_; }
 108
 109 int32 DeathData::run_duration_sample() const {
 110   return run_duration_sample_;
 111 }
 112
 113 int32 DeathData::queue_duration_sum() const {
 114   return queue_duration_sum_;
 115 }
 116
 117 int32 DeathData::queue_duration_max() const {
 118   return queue_duration_max_;
 119 }
 120
 121 int32 DeathData::queue_duration_sample() const {
 122   return queue_duration_sample_;
 123 }
 124
 125 void DeathData::ResetMax() {
 126   run_duration_max_ = 0;
 127   queue_duration_max_ = 0;
 128 }
 129
 130 void DeathData::Clear() {
 131   count_ = 0;
 132   run_duration_sum_ = 0;
 133   run_duration_max_ = 0;
 134   run_duration_sample_ = 0;
 135   queue_duration_sum_ = 0;
 136   queue_duration_max_ = 0;
 137   queue_duration_sample_ = 0;
 138 }
 139
 140 //------------------------------------------------------------------------------
 141 DeathDataSnapshot::DeathDataSnapshot()
 142     : count(-1),
 143       run_duration_sum(-1),
 144       run_duration_max(-1),
 145       run_duration_sample(-1),
 146       queue_duration_sum(-1),
 147       queue_duration_max(-1),
 148       queue_duration_sample(-1) {
 149 }
 150
 151 DeathDataSnapshot::DeathDataSnapshot(
 152     const tracked_objects::DeathData& death_data)
 153     : count(death_data.count()),
 154       run_duration_sum(death_data.run_duration_sum()),
 155       run_duration_max(death_data.run_duration_max()),
 156       run_duration_sample(death_data.run_duration_sample()),
 157       queue_duration_sum(death_data.queue_duration_sum()),
 158       queue_duration_max(death_data.queue_duration_max()),
 159       queue_duration_sample(death_data.queue_duration_sample()) {
 160 }
 161
 162 DeathDataSnapshot::~DeathDataSnapshot() {
 163 }
 164
 165 //------------------------------------------------------------------------------
 166 BirthOnThread::BirthOnThread(const Location& location,
 167                              const ThreadData& current)
 168     : location_(location),
 169       birth_thread_(&current) {
 170 }
 171
 172 //------------------------------------------------------------------------------
 173 BirthOnThreadSnapshot::BirthOnThreadSnapshot() {
 174 }
 175
 176 BirthOnThreadSnapshot::BirthOnThreadSnapshot(
 177     const tracked_objects::BirthOnThread& birth)
 178     : location(birth.location()),
 179       thread_name(birth.birth_thread()->thread_name()) {
 180 }
 181
 182 BirthOnThreadSnapshot::~BirthOnThreadSnapshot() {
 183 }
 184
 185 //------------------------------------------------------------------------------
 186 Births::Births(const Location& location, const ThreadData& current)
 187     : BirthOnThread(location, current),
 188       birth_count_(1) { }
 189
 190 int Births::birth_count() const { return birth_count_; }
 191
 192 void Births::RecordBirth() { ++birth_count_; }
 193
 194 void Births::ForgetBirth() { --birth_count_; }
 195
 196 void Births::Clear() { birth_count_ = 0; }
 197
 198 //------------------------------------------------------------------------------
 199 // ThreadData maintains the central data for all births and deaths on a single
 200 // thread.
 201
 202 // TODO(jar): We should pull all these static vars together, into a struct, and
 203 // optimize layout so that we benefit from locality of reference during accesses
 204 // to them.
 205
 206 // static
 207 NowFunction* ThreadData::now_function_ = NULL;
 208
 209 // A TLS slot which points to the ThreadData instance for the current thread. We
 210 // do a fake initialization here (zeroing out data), and then the real in-place
 211 // construction happens when we call tls_index_.Initialize().
 212 // static
 213 base::ThreadLocalStorage::StaticSlot ThreadData::tls_index_ = TLS_INITIALIZER;
 214
 215 // static
 216 int ThreadData::worker_thread_data_creation_count_ = 0;
 217
 218 // static
 219 int ThreadData::cleanup_count_ = 0;
 220
 221 // static
 222 int ThreadData::incarnation_counter_ = 0;
 223
 224 // static
 225 ThreadData* ThreadData::all_thread_data_list_head_ = NULL;
 226
 227 // static
 228 ThreadData* ThreadData::first_retired_worker_ = NULL;
 229
 230 // static
 231 base::LazyInstance<base::Lock>::Leaky
 232     ThreadData::list_lock_ = LAZY_INSTANCE_INITIALIZER;
 233
 234 // static
 235 ThreadData::Status ThreadData::status_ = ThreadData::UNINITIALIZED;
 236
 237 ThreadData::ThreadData(const std::string& suggested_name)
 238     : next_(NULL),
 239       next_retired_worker_(NULL),
 240       worker_thread_number_(0),
 241       incarnation_count_for_pool_(-1) {
 242   DCHECK_GE(suggested_name.size(), 0u);
 243   thread_name_ = suggested_name;
 244   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 245 }
 246
 247 ThreadData::ThreadData(int thread_number)
 248     : next_(NULL),
 249       next_retired_worker_(NULL),
 250       worker_thread_number_(thread_number),
 251       incarnation_count_for_pool_(-1)  {
 252   CHECK_GT(thread_number, 0);
 253   base::StringAppendF(&thread_name_, "WorkerThread-%d", thread_number);
 254   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 255 }
 256
 257 ThreadData::~ThreadData() {}
 258
 259 void ThreadData::PushToHeadOfList() {
 260   // Toss in a hint of randomness (atop the uniniitalized value).
 261   (void)VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE(&random_number_,
 262                                                  sizeof(random_number_));
 263   random_number_ += static_cast<int32>(this - static_cast<ThreadData*>(0));
 264   random_number_ ^= (Now() - TrackedTime()).InMilliseconds();
 265
 266   DCHECK(!next_);
 267   base::AutoLock lock(*list_lock_.Pointer());
 268   incarnation_count_for_pool_ = incarnation_counter_;
 269   next_ = all_thread_data_list_head_;
 270   all_thread_data_list_head_ = this;
 271 }
 272
 273 // static
 274 ThreadData* ThreadData::first() {
 275   base::AutoLock lock(*list_lock_.Pointer());
 276   return all_thread_data_list_head_;
 277 }
 278
 279 ThreadData* ThreadData::next() const { return next_; }
 280
 281 // static
 282 void ThreadData::InitializeThreadContext(const std::string& suggested_name) {
 283   if (!Initialize())  // Always initialize if needed.
 284     return;
 285   ThreadData* current_thread_data =
 286       reinterpret_cast<ThreadData*>(tls_index_.Get());
 287   if (current_thread_data)
 288     return;  // Browser tests instigate this.
 289   current_thread_data = new ThreadData(suggested_name);
 290   tls_index_.Set(current_thread_data);
 291 }
 292
 293 // static
 294 ThreadData* ThreadData::Get() {
 295   if (!tls_index_.initialized())
 296     return NULL;  // For unittests only.
 297   ThreadData* registered = reinterpret_cast<ThreadData*>(tls_index_.Get());
 298   if (registered)
 299     return registered;
 300
 301   // We must be a worker thread, since we didn't pre-register.
 302   ThreadData* worker_thread_data = NULL;
 303   int worker_thread_number = 0;
 304   {
 305     base::AutoLock lock(*list_lock_.Pointer());
 306     if (first_retired_worker_) {
 307       worker_thread_data = first_retired_worker_;
 308       first_retired_worker_ = first_retired_worker_->next_retired_worker_;
 309       worker_thread_data->next_retired_worker_ = NULL;
 310     } else {
 311       worker_thread_number = ++worker_thread_data_creation_count_;
 312     }
 313   }
 314
 315   // If we can't find a previously used instance, then we have to create one.
 316   if (!worker_thread_data) {
 317     DCHECK_GT(worker_thread_number, 0);
 318     worker_thread_data = new ThreadData(worker_thread_number);
 319   }
 320   DCHECK_GT(worker_thread_data->worker_thread_number_, 0);
 321
 322   tls_index_.Set(worker_thread_data);
 323   return worker_thread_data;
 324 }
 325
 326 // static
 327 void ThreadData::OnThreadTermination(void* thread_data) {
 328   DCHECK(thread_data);  // TLS should *never* call us with a NULL.
 329   // We must NOT do any allocations during this callback. There is a chance
 330   // that the allocator is no longer active on this thread.
 331   if (!kTrackAllTaskObjects)
 332     return;  // Not compiled in.
 333   reinterpret_cast<ThreadData*>(thread_data)->OnThreadTerminationCleanup();
 334 }
 335
 336 void ThreadData::OnThreadTerminationCleanup() {
 337   // The list_lock_ was created when we registered the callback, so it won't be
 338   // allocated here despite the lazy reference.
 339   base::AutoLock lock(*list_lock_.Pointer());
 340   if (incarnation_counter_ != incarnation_count_for_pool_)
 341     return;  // ThreadData was constructed in an earlier unit test.
 342   ++cleanup_count_;
 343   // Only worker threads need to be retired and reused.
 344   if (!worker_thread_number_) {
 345     return;
 346   }
 347   // We must NOT do any allocations during this callback.
 348   // Using the simple linked lists avoids all allocations.
 349   DCHECK_EQ(this->next_retired_worker_, reinterpret_cast<ThreadData*>(NULL));
 350   this->next_retired_worker_ = first_retired_worker_;
 351   first_retired_worker_ = this;
 352 }
 353
 354 // static
 355 void ThreadData::Snapshot(bool reset_max, ProcessDataSnapshot* process_data) {
 356   // Add births that have run to completion to |collected_data|.
 357   // |birth_counts| tracks the total number of births recorded at each location
 358   // for which we have not seen a death count.
 359   BirthCountMap birth_counts;
 360   ThreadData::SnapshotAllExecutedTasks(reset_max, process_data, &birth_counts);
 361
 362   // Add births that are still active -- i.e. objects that have tallied a birth,
 363   // but have not yet tallied a matching death, and hence must be either
 364   // running, queued up, or being held in limbo for future posting.
 365   for (BirthCountMap::const_iterator it = birth_counts.begin();
 366        it != birth_counts.end(); ++it) {
 367     if (it->second > 0) {
 368       process_data->tasks.push_back(
 369           TaskSnapshot(*it->first, DeathData(it->second), "Still_Alive"));
 370     }
 371   }
 372 }
 373
 374 Births* ThreadData::TallyABirth(const Location& location) {
 375   BirthMap::iterator it = birth_map_.find(location);
 376   Births* child;
 377   if (it != birth_map_.end()) {
 378     child =  it->second;
 379     child->RecordBirth();
 380   } else {
 381     child = new Births(location, *this);  // Leak this.
 382     // Lock since the map may get relocated now, and other threads sometimes
 383     // snapshot it (but they lock before copying it).
 384     base::AutoLock lock(map_lock_);
 385     birth_map_[location] = child;
 386   }
 387
 388   if (kTrackParentChildLinks && status_ > PROFILING_ACTIVE &&
 389       !parent_stack_.empty()) {
 390     const Births* parent = parent_stack_.top();
 391     ParentChildPair pair(parent, child);
 392     if (parent_child_set_.find(pair) == parent_child_set_.end()) {
 393       // Lock since the map may get relocated now, and other threads sometimes
 394       // snapshot it (but they lock before copying it).
 395       base::AutoLock lock(map_lock_);
 396       parent_child_set_.insert(pair);
 397     }
 398   }
 399
 400   return child;
 401 }
 402
 403 void ThreadData::TallyADeath(const Births& birth,
 404                              int32 queue_duration,
 405                              int32 run_duration) {
 406   // Stir in some randomness, plus add constant in case durations are zero.
 407   const int32 kSomePrimeNumber = 2147483647;
 408   random_number_ += queue_duration + run_duration + kSomePrimeNumber;
 409   // An address is going to have some randomness to it as well ;-).
 410   random_number_ ^= static_cast<int32>(&birth - reinterpret_cast<Births*>(0));
 411
 412   // We don't have queue durations without OS timer. OS timer is automatically
 413   // used for task-post-timing, so the use of an alternate timer implies all
 414   // queue times are invalid.
 415   if (kAllowAlternateTimeSourceHandling && now_function_)
 416     queue_duration = 0;
 417
 418   DeathMap::iterator it = death_map_.find(&birth);
 419   DeathData* death_data;
 420   if (it != death_map_.end()) {
 421     death_data = &it->second;
 422   } else {
 423     base::AutoLock lock(map_lock_);  // Lock as the map may get relocated now.
 424     death_data = &death_map_[&birth];
 425   }  // Release lock ASAP.
 426   death_data->RecordDeath(queue_duration, run_duration, random_number_);
 427
 428   if (!kTrackParentChildLinks)
 429     return;
 430   if (!parent_stack_.empty()) {  // We might get turned off.
 431     DCHECK_EQ(parent_stack_.top(), &birth);
 432     parent_stack_.pop();
 433   }
 434 }
 435
 436 // static
 437 Births* ThreadData::TallyABirthIfActive(const Location& location) {
 438   if (!kTrackAllTaskObjects)
 439     return NULL;  // Not compiled in.
 440
 441   if (!TrackingStatus())
 442     return NULL;
 443   ThreadData* current_thread_data = Get();
 444   if (!current_thread_data)
 445     return NULL;
 446   return current_thread_data->TallyABirth(location);
 447 }
 448
 449 // static
 450 void ThreadData::TallyRunOnNamedThreadIfTracking(
 451     const base::TrackingInfo& completed_task,
 452     const TrackedTime& start_of_run,
 453     const TrackedTime& end_of_run) {
 454   if (!kTrackAllTaskObjects)
 455     return;  // Not compiled in.
 456
 457   // Even if we have been DEACTIVATED, we will process any pending births so
 458   // that our data structures (which counted the outstanding births) remain
 459   // consistent.
 460   const Births* birth = completed_task.birth_tally;
 461   if (!birth)
 462     return;
 463   ThreadData* current_thread_data = Get();
 464   if (!current_thread_data)
 465     return;
 466
 467   // To avoid conflating our stats with the delay duration in a PostDelayedTask,
 468   // we identify such tasks, and replace their post_time with the time they
 469   // were scheduled (requested?) to emerge from the delayed task queue. This
 470   // means that queueing delay for such tasks will show how long they went
 471   // unserviced, after they *could* be serviced.  This is the same stat as we
 472   // have for non-delayed tasks, and we consistently call it queueing delay.
 473   TrackedTime effective_post_time = completed_task.delayed_run_time.is_null()
 474       ? tracked_objects::TrackedTime(completed_task.time_posted)
 475       : tracked_objects::TrackedTime(completed_task.delayed_run_time);
 476
 477   // Watch out for a race where status_ is changing, and hence one or both
 478   // of start_of_run or end_of_run is zero.  In that case, we didn't bother to
 479   // get a time value since we "weren't tracking" and we were trying to be
 480   // efficient by not calling for a genuine time value. For simplicity, we'll
 481   // use a default zero duration when we can't calculate a true value.
 482   int32 queue_duration = 0;
 483   int32 run_duration = 0;
 484   if (!start_of_run.is_null()) {
 485     queue_duration = (start_of_run - effective_post_time).InMilliseconds();
 486     if (!end_of_run.is_null())
 487       run_duration = (end_of_run - start_of_run).InMilliseconds();
 488   }
 489   current_thread_data->TallyADeath(*birth, queue_duration, run_duration);
 490 }
 491
 492 // static
 493 void ThreadData::TallyRunOnWorkerThreadIfTracking(
 494     const Births* birth,
 495     const TrackedTime& time_posted,
 496     const TrackedTime& start_of_run,
 497     const TrackedTime& end_of_run) {
 498   if (!kTrackAllTaskObjects)
 499     return;  // Not compiled in.
 500
 501   // Even if we have been DEACTIVATED, we will process any pending births so
 502   // that our data structures (which counted the outstanding births) remain
 503   // consistent.
 504   if (!birth)
 505     return;
 506
 507   // TODO(jar): Support the option to coalesce all worker-thread activity under
 508   // one ThreadData instance that uses locks to protect *all* access.  This will
 509   // reduce memory (making it provably bounded), but run incrementally slower
 510   // (since we'll use locks on TallyBirth and TallyDeath).  The good news is
 511   // that the locks on TallyDeath will be *after* the worker thread has run, and
 512   // hence nothing will be waiting for the completion (... besides some other
 513   // thread that might like to run).  Also, the worker threads tasks are
 514   // generally longer, and hence the cost of the lock may perchance be amortized
 515   // over the long task's lifetime.
 516   ThreadData* current_thread_data = Get();
 517   if (!current_thread_data)
 518     return;
 519
 520   int32 queue_duration = 0;
 521   int32 run_duration = 0;
 522   if (!start_of_run.is_null()) {
 523     queue_duration = (start_of_run - time_posted).InMilliseconds();
 524     if (!end_of_run.is_null())
 525       run_duration = (end_of_run - start_of_run).InMilliseconds();
 526   }
 527   current_thread_data->TallyADeath(*birth, queue_duration, run_duration);
 528 }
 529
 530 // static
 531 void ThreadData::TallyRunInAScopedRegionIfTracking(
 532     const Births* birth,
 533     const TrackedTime& start_of_run,
 534     const TrackedTime& end_of_run) {
 535   if (!kTrackAllTaskObjects)
 536     return;  // Not compiled in.
 537
 538   // Even if we have been DEACTIVATED, we will process any pending births so
 539   // that our data structures (which counted the outstanding births) remain
 540   // consistent.
 541   if (!birth)
 542     return;
 543
 544   ThreadData* current_thread_data = Get();
 545   if (!current_thread_data)
 546     return;
 547
 548   int32 queue_duration = 0;
 549   int32 run_duration = 0;
 550   if (!start_of_run.is_null() && !end_of_run.is_null())
 551     run_duration = (end_of_run - start_of_run).InMilliseconds();
 552   current_thread_data->TallyADeath(*birth, queue_duration, run_duration);
 553 }
 554
 555 const std::string ThreadData::thread_name() const { return thread_name_; }
 556
 557 // static
 558 void ThreadData::SnapshotAllExecutedTasks(bool reset_max,
 559                                           ProcessDataSnapshot* process_data,
 560                                           BirthCountMap* birth_counts) {
 561   if (!kTrackAllTaskObjects)
 562     return;  // Not compiled in.
 563
 564   // Get an unchanging copy of a ThreadData list.
 565   ThreadData* my_list = ThreadData::first();
 566
 567   // Gather data serially.
 568   // This hackish approach *can* get some slighly corrupt tallies, as we are
 569   // grabbing values without the protection of a lock, but it has the advantage
 570   // of working even with threads that don't have message loops.  If a user
 571   // sees any strangeness, they can always just run their stats gathering a
 572   // second time.
 573   for (ThreadData* thread_data = my_list;
 574        thread_data;
 575        thread_data = thread_data->next()) {
 576     thread_data->SnapshotExecutedTasks(reset_max, process_data, birth_counts);
 577   }
 578 }
 579
 580 void ThreadData::SnapshotExecutedTasks(bool reset_max,
 581                                        ProcessDataSnapshot* process_data,
 582                                        BirthCountMap* birth_counts) {
 583   // Get copy of data, so that the data will not change during the iterations
 584   // and processing.
 585   ThreadData::BirthMap birth_map;
 586   ThreadData::DeathMap death_map;
 587   ThreadData::ParentChildSet parent_child_set;
 588   SnapshotMaps(reset_max, &birth_map, &death_map, &parent_child_set);
 589
 590   for (ThreadData::DeathMap::const_iterator it = death_map.begin();
 591        it != death_map.end(); ++it) {
 592     process_data->tasks.push_back(
 593         TaskSnapshot(*it->first, it->second, thread_name()));
 594     (*birth_counts)[it->first] -= it->first->birth_count();
 595   }
 596
 597   for (ThreadData::BirthMap::const_iterator it = birth_map.begin();
 598        it != birth_map.end(); ++it) {
 599     (*birth_counts)[it->second] += it->second->birth_count();
 600   }
 601
 602   if (!kTrackParentChildLinks)
 603     return;
 604
 605   for (ThreadData::ParentChildSet::const_iterator it = parent_child_set.begin();
 606        it != parent_child_set.end(); ++it) {
 607     process_data->descendants.push_back(ParentChildPairSnapshot(*it));
 608   }
 609 }
 610
 611 // This may be called from another thread.
 612 void ThreadData::SnapshotMaps(bool reset_max,
 613                               BirthMap* birth_map,
 614                               DeathMap* death_map,
 615                               ParentChildSet* parent_child_set) {
 616   base::AutoLock lock(map_lock_);
 617   for (BirthMap::const_iterator it = birth_map_.begin();
 618        it != birth_map_.end(); ++it)
 619     (*birth_map)[it->first] = it->second;
 620   for (DeathMap::iterator it = death_map_.begin();
 621        it != death_map_.end(); ++it) {
 622     (*death_map)[it->first] = it->second;
 623     if (reset_max)
 624       it->second.ResetMax();
 625   }
 626
 627   if (!kTrackParentChildLinks)
 628     return;
 629
 630   for (ParentChildSet::iterator it = parent_child_set_.begin();
 631        it != parent_child_set_.end(); ++it)
 632     parent_child_set->insert(*it);
 633 }
 634
 635 // static
 636 void ThreadData::ResetAllThreadData() {
 637   ThreadData* my_list = first();
 638
 639   for (ThreadData* thread_data = my_list;
 640        thread_data;
 641        thread_data = thread_data->next())
 642     thread_data->Reset();
 643 }
 644
 645 void ThreadData::Reset() {
 646   base::AutoLock lock(map_lock_);
 647   for (DeathMap::iterator it = death_map_.begin();
 648        it != death_map_.end(); ++it)
 649     it->second.Clear();
 650   for (BirthMap::iterator it = birth_map_.begin();
 651        it != birth_map_.end(); ++it)
 652     it->second->Clear();
 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 TrackedTime ThreadData::NowForStartOfRun(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   return Now();
 747 }
 748
 749 // static
 750 TrackedTime ThreadData::NowForEndOfRun() {
 751   return Now();
 752 }
 753
 754 // static
 755 void ThreadData::SetAlternateTimeSource(NowFunction* now_function) {
 756   DCHECK(now_function);
 757   if (kAllowAlternateTimeSourceHandling)
 758     now_function_ = now_function;
 759 }
 760
 761 // static
 762 TrackedTime ThreadData::Now() {
 763   if (kAllowAlternateTimeSourceHandling && now_function_)
 764     return TrackedTime::FromMilliseconds((*now_function_)());
 765   if (kTrackAllTaskObjects && TrackingStatus())
 766     return TrackedTime::Now();
 767   return TrackedTime();  // Super fast when disabled, or not compiled.
 768 }
 769
 770 // static
 771 void ThreadData::EnsureCleanupWasCalled(int major_threads_shutdown_count) {
 772   base::AutoLock lock(*list_lock_.Pointer());
 773   if (worker_thread_data_creation_count_ == 0)
 774     return;  // We haven't really run much, and couldn't have leaked.
 775   // Verify that we've at least shutdown/cleanup the major namesd threads.  The
 776   // caller should tell us how many thread shutdowns should have taken place by
 777   // now.
 778   return;  // TODO(jar): until this is working on XP, don't run the real test.
 779   CHECK_GT(cleanup_count_, major_threads_shutdown_count);
 780 }
 781
 782 // static
 783 void ThreadData::ShutdownSingleThreadedCleanup(bool leak) {
 784   // This is only called from test code, where we need to cleanup so that
 785   // additional tests can be run.
 786   // We must be single threaded... but be careful anyway.
 787   if (!InitializeAndSetTrackingStatus(DEACTIVATED))
 788     return;
 789   ThreadData* thread_data_list;
 790   {
 791     base::AutoLock lock(*list_lock_.Pointer());
 792     thread_data_list = all_thread_data_list_head_;
 793     all_thread_data_list_head_ = NULL;
 794     ++incarnation_counter_;
 795     // To be clean, break apart the retired worker list (though we leak them).
 796     while (first_retired_worker_) {
 797       ThreadData* worker = first_retired_worker_;
 798       CHECK_GT(worker->worker_thread_number_, 0);
 799       first_retired_worker_ = worker->next_retired_worker_;
 800       worker->next_retired_worker_ = NULL;
 801     }
 802   }
 803
 804   // Put most global static back in pristine shape.
 805   worker_thread_data_creation_count_ = 0;
 806   cleanup_count_ = 0;
 807   tls_index_.Set(NULL);
 808   status_ = DORMANT_DURING_TESTS;  // Almost UNINITIALIZED.
 809
 810   // To avoid any chance of racing in unit tests, which is the only place we
 811   // call this function, we may sometimes leak all the data structures we
 812   // recovered, as they may still be in use on threads from prior tests!
 813   if (leak)
 814     return;
 815
 816   // When we want to cleanup (on a single thread), here is what we do.
 817
 818   // Do actual recursive delete in all ThreadData instances.
 819   while (thread_data_list) {
 820     ThreadData* next_thread_data = thread_data_list;
 821     thread_data_list = thread_data_list->next();
 822
 823     for (BirthMap::iterator it = next_thread_data->birth_map_.begin();
 824          next_thread_data->birth_map_.end() != it; ++it)
 825       delete it->second;  // Delete the Birth Records.
 826     delete next_thread_data;  // Includes all Death Records.
 827   }
 828 }
 829
 830 //------------------------------------------------------------------------------
 831 TaskSnapshot::TaskSnapshot() {
 832 }
 833
 834 TaskSnapshot::TaskSnapshot(const BirthOnThread& birth,
 835                            const DeathData& death_data,
 836                            const std::string& death_thread_name)
 837     : birth(birth),
 838       death_data(death_data),
 839       death_thread_name(death_thread_name) {
 840 }
 841
 842 TaskSnapshot::~TaskSnapshot() {
 843 }
 844
 845 //------------------------------------------------------------------------------
 846 // ParentChildPairSnapshot
 847
 848 ParentChildPairSnapshot::ParentChildPairSnapshot() {
 849 }
 850
 851 ParentChildPairSnapshot::ParentChildPairSnapshot(
 852     const ThreadData::ParentChildPair& parent_child)
 853     : parent(*parent_child.first),
 854       child(*parent_child.second) {
 855 }
 856
 857 ParentChildPairSnapshot::~ParentChildPairSnapshot() {
 858 }
 859
 860 //------------------------------------------------------------------------------
 861 // ProcessDataSnapshot
 862
 863 ProcessDataSnapshot::ProcessDataSnapshot()
 864 #if !defined(OS_NACL)
 865     : process_id(base::GetCurrentProcId()) {
 866 #else
 867     : process_id(0) {
 868 #endif
 869 }
 870
 871 ProcessDataSnapshot::~ProcessDataSnapshot() {
 872 }
 873
 874 }  // namespace tracked_objects