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 void Births::ForgetBirth() { --birth_count_; }
 231
 232 void Births::Clear() { birth_count_ = 0; }
 233
 234 //------------------------------------------------------------------------------
 235 // ThreadData maintains the central data for all births and deaths on a single
 236 // thread.
 237
 238 // TODO(jar): We should pull all these static vars together, into a struct, and
 239 // optimize layout so that we benefit from locality of reference during accesses
 240 // to them.
 241
 242 // static
 243 NowFunction* ThreadData::now_function_ = NULL;
 244
 245 // static
 246 bool ThreadData::now_function_is_time_ = false;
 247
 248 // A TLS slot which points to the ThreadData instance for the current thread. We
 249 // do a fake initialization here (zeroing out data), and then the real in-place
 250 // construction happens when we call tls_index_.Initialize().
 251 // static
 252 base::ThreadLocalStorage::StaticSlot ThreadData::tls_index_ = TLS_INITIALIZER;
 253
 254 // static
 255 int ThreadData::worker_thread_data_creation_count_ = 0;
 256
 257 // static
 258 int ThreadData::cleanup_count_ = 0;
 259
 260 // static
 261 int ThreadData::incarnation_counter_ = 0;
 262
 263 // static
 264 ThreadData* ThreadData::all_thread_data_list_head_ = NULL;
 265
 266 // static
 267 ThreadData* ThreadData::first_retired_worker_ = NULL;
 268
 269 // static
 270 base::LazyInstance<base::Lock>::Leaky
 271     ThreadData::list_lock_ = LAZY_INSTANCE_INITIALIZER;
 272
 273 // static
 274 ThreadData::Status ThreadData::status_ = ThreadData::UNINITIALIZED;
 275
 276 ThreadData::ThreadData(const std::string& suggested_name)
 277     : next_(NULL),
 278       next_retired_worker_(NULL),
 279       worker_thread_number_(0),
 280       incarnation_count_for_pool_(-1),
 281       current_stopwatch_(NULL) {
 282   DCHECK_GE(suggested_name.size(), 0u);
 283   thread_name_ = suggested_name;
 284   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 285 }
 286
 287 ThreadData::ThreadData(int thread_number)
 288     : next_(NULL),
 289       next_retired_worker_(NULL),
 290       worker_thread_number_(thread_number),
 291       incarnation_count_for_pool_(-1),
 292       current_stopwatch_(NULL) {
 293   CHECK_GT(thread_number, 0);
 294   base::StringAppendF(&thread_name_, "WorkerThread-%d", thread_number);
 295   PushToHeadOfList();  // Which sets real incarnation_count_for_pool_.
 296 }
 297
 298 ThreadData::~ThreadData() {}
 299
 300 void ThreadData::PushToHeadOfList() {
 301   // Toss in a hint of randomness (atop the uniniitalized value).
 302   (void)VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE(&random_number_,
 303                                                  sizeof(random_number_));
 304   MSAN_UNPOISON(&random_number_, sizeof(random_number_));
 305   random_number_ += static_cast<uint32>(this - static_cast<ThreadData*>(0));
 306   random_number_ ^= (Now() - TrackedTime()).InMilliseconds();
 307
 308   DCHECK(!next_);
 309   base::AutoLock lock(*list_lock_.Pointer());
 310   incarnation_count_for_pool_ = incarnation_counter_;
 311   next_ = all_thread_data_list_head_;
 312   all_thread_data_list_head_ = this;
 313 }
 314
 315 // static
 316 ThreadData* ThreadData::first() {
 317   base::AutoLock lock(*list_lock_.Pointer());
 318   return all_thread_data_list_head_;
 319 }
 320
 321 ThreadData* ThreadData::next() const { return next_; }
 322
 323 // static
 324 void ThreadData::InitializeThreadContext(const std::string& suggested_name) {
 325   if (!Initialize())  // Always initialize if needed.
 326     return;
 327   ThreadData* current_thread_data =
 328       reinterpret_cast<ThreadData*>(tls_index_.Get());
 329   if (current_thread_data)
 330     return;  // Browser tests instigate this.
 331   current_thread_data = new ThreadData(suggested_name);
 332   tls_index_.Set(current_thread_data);
 333 }
 334
 335 // static
 336 ThreadData* ThreadData::Get() {
 337   if (!tls_index_.initialized())
 338     return NULL;  // For unittests only.
 339   ThreadData* registered = reinterpret_cast<ThreadData*>(tls_index_.Get());
 340   if (registered)
 341     return registered;
 342
 343   // We must be a worker thread, since we didn't pre-register.
 344   ThreadData* worker_thread_data = NULL;
 345   int worker_thread_number = 0;
 346   {
 347     base::AutoLock lock(*list_lock_.Pointer());
 348     if (first_retired_worker_) {
 349       worker_thread_data = first_retired_worker_;
 350       first_retired_worker_ = first_retired_worker_->next_retired_worker_;
 351       worker_thread_data->next_retired_worker_ = NULL;
 352     } else {
 353       worker_thread_number = ++worker_thread_data_creation_count_;
 354     }
 355   }
 356
 357   // If we can't find a previously used instance, then we have to create one.
 358   if (!worker_thread_data) {
 359     DCHECK_GT(worker_thread_number, 0);
 360     worker_thread_data = new ThreadData(worker_thread_number);
 361   }
 362   DCHECK_GT(worker_thread_data->worker_thread_number_, 0);
 363
 364   tls_index_.Set(worker_thread_data);
 365   return worker_thread_data;
 366 }
 367
 368 // static
 369 void ThreadData::OnThreadTermination(void* thread_data) {
 370   DCHECK(thread_data);  // TLS should *never* call us with a NULL.
 371   // We must NOT do any allocations during this callback. There is a chance
 372   // that the allocator is no longer active on this thread.
 373   if (!kTrackAllTaskObjects)
 374     return;  // Not compiled in.
 375   reinterpret_cast<ThreadData*>(thread_data)->OnThreadTerminationCleanup();
 376 }
 377
 378 void ThreadData::OnThreadTerminationCleanup() {
 379   // The list_lock_ was created when we registered the callback, so it won't be
 380   // allocated here despite the lazy reference.
 381   base::AutoLock lock(*list_lock_.Pointer());
 382   if (incarnation_counter_ != incarnation_count_for_pool_)
 383     return;  // ThreadData was constructed in an earlier unit test.
 384   ++cleanup_count_;
 385   // Only worker threads need to be retired and reused.
 386   if (!worker_thread_number_) {
 387     return;
 388   }
 389   // We must NOT do any allocations during this callback.
 390   // Using the simple linked lists avoids all allocations.
 391   DCHECK_EQ(this->next_retired_worker_, reinterpret_cast<ThreadData*>(NULL));
 392   this->next_retired_worker_ = first_retired_worker_;
 393   first_retired_worker_ = this;
 394 }
 395
 396 // static
 397 void ThreadData::Snapshot(ProcessDataSnapshot* process_data) {
 398   // Add births that have run to completion to |collected_data|.
 399   // |birth_counts| tracks the total number of births recorded at each location
 400   // for which we have not seen a death count.
 401   BirthCountMap birth_counts;
 402   ThreadData::SnapshotAllExecutedTasks(process_data, &birth_counts);
 403
 404   // Add births that are still active -- i.e. objects that have tallied a birth,
 405   // but have not yet tallied a matching death, and hence must be either
 406   // running, queued up, or being held in limbo for future posting.
 407   for (BirthCountMap::const_iterator it = birth_counts.begin();
 408        it != birth_counts.end(); ++it) {
 409     if (it->second > 0) {
 410       process_data->tasks.push_back(
 411           TaskSnapshot(*it->first, DeathData(it->second), "Still_Alive"));
 412     }
 413   }
 414 }
 415
 416 Births* ThreadData::TallyABirth(const Location& location) {
 417   BirthMap::iterator it = birth_map_.find(location);
 418   Births* child;
 419   if (it != birth_map_.end()) {
 420     child =  it->second;
 421     child->RecordBirth();
 422   } else {
 423     child = new Births(location, *this);  // Leak this.
 424     // Lock since the map may get relocated now, and other threads sometimes
 425     // snapshot it (but they lock before copying it).
 426     base::AutoLock lock(map_lock_);
 427     birth_map_[location] = child;
 428   }
 429
 430   if (kTrackParentChildLinks && status_ > PROFILING_ACTIVE &&
 431       !parent_stack_.empty()) {
 432     const Births* parent = parent_stack_.top();
 433     ParentChildPair pair(parent, child);
 434     if (parent_child_set_.find(pair) == parent_child_set_.end()) {
 435       // Lock since the map may get relocated now, and other threads sometimes
 436       // snapshot it (but they lock before copying it).
 437       base::AutoLock lock(map_lock_);
 438       parent_child_set_.insert(pair);
 439     }
 440   }
 441
 442   return child;
 443 }
 444
 445 void ThreadData::TallyADeath(const Births& birth,
 446                              int32 queue_duration,
 447                              const TaskStopwatch& stopwatch) {
 448   int32 run_duration = stopwatch.RunDurationMs();
 449
 450   // Stir in some randomness, plus add constant in case durations are zero.
 451   const uint32 kSomePrimeNumber = 2147483647;
 452   random_number_ += queue_duration + run_duration + kSomePrimeNumber;
 453   // An address is going to have some randomness to it as well ;-).
 454   random_number_ ^= static_cast<uint32>(&birth - reinterpret_cast<Births*>(0));
 455
 456   // We don't have queue durations without OS timer. OS timer is automatically
 457   // used for task-post-timing, so the use of an alternate timer implies all
 458   // queue times are invalid, unless it was explicitly said that we can trust
 459   // the alternate timer.
 460   if (kAllowAlternateTimeSourceHandling &&
 461       now_function_ &&
 462       !now_function_is_time_) {
 463     queue_duration = 0;
 464   }
 465
 466   DeathMap::iterator it = death_map_.find(&birth);
 467   DeathData* death_data;
 468   if (it != death_map_.end()) {
 469     death_data = &it->second;
 470   } else {
 471     base::AutoLock lock(map_lock_);  // Lock as the map may get relocated now.
 472     death_data = &death_map_[&birth];
 473   }  // Release lock ASAP.
 474   death_data->RecordDeath(queue_duration, run_duration, random_number_);
 475
 476   if (!kTrackParentChildLinks)
 477     return;
 478   if (!parent_stack_.empty()) {  // We might get turned off.
 479     DCHECK_EQ(parent_stack_.top(), &birth);
 480     parent_stack_.pop();
 481   }
 482 }
 483
 484 // static
 485 Births* ThreadData::TallyABirthIfActive(const Location& location) {
 486   if (!kTrackAllTaskObjects)
 487     return NULL;  // Not compiled in.
 488
 489   if (!TrackingStatus())
 490     return NULL;
 491   ThreadData* current_thread_data = Get();
 492   if (!current_thread_data)
 493     return NULL;
 494   return current_thread_data->TallyABirth(location);
 495 }
 496
 497 // static
 498 void ThreadData::TallyRunOnNamedThreadIfTracking(
 499     const base::TrackingInfo& completed_task,
 500     const TaskStopwatch& stopwatch) {
 501   if (!kTrackAllTaskObjects)
 502     return;  // Not compiled in.
 503
 504   // Even if we have been DEACTIVATED, we will process any pending births so
 505   // that our data structures (which counted the outstanding births) remain
 506   // consistent.
 507   const Births* birth = completed_task.birth_tally;
 508   if (!birth)
 509     return;
 510   ThreadData* current_thread_data = stopwatch.GetThreadData();
 511   if (!current_thread_data)
 512     return;
 513
 514   // Watch out for a race where status_ is changing, and hence one or both
 515   // of start_of_run or end_of_run is zero.  In that case, we didn't bother to
 516   // get a time value since we "weren't tracking" and we were trying to be
 517   // efficient by not calling for a genuine time value. For simplicity, we'll
 518   // use a default zero duration when we can't calculate a true value.
 519   TrackedTime start_of_run = stopwatch.StartTime();
 520   int32 queue_duration = 0;
 521   if (!start_of_run.is_null()) {
 522     queue_duration = (start_of_run - completed_task.EffectiveTimePosted())
 523         .InMilliseconds();
 524   }
 525   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 526 }
 527
 528 // static
 529 void ThreadData::TallyRunOnWorkerThreadIfTracking(
 530     const Births* birth,
 531     const TrackedTime& time_posted,
 532     const TaskStopwatch& stopwatch) {
 533   if (!kTrackAllTaskObjects)
 534     return;  // Not compiled in.
 535
 536   // Even if we have been DEACTIVATED, we will process any pending births so
 537   // that our data structures (which counted the outstanding births) remain
 538   // consistent.
 539   if (!birth)
 540     return;
 541
 542   // TODO(jar): Support the option to coalesce all worker-thread activity under
 543   // one ThreadData instance that uses locks to protect *all* access.  This will
 544   // reduce memory (making it provably bounded), but run incrementally slower
 545   // (since we'll use locks on TallyABirth and TallyADeath).  The good news is
 546   // that the locks on TallyADeath will be *after* the worker thread has run,
 547   // and hence nothing will be waiting for the completion (... besides some
 548   // other thread that might like to run).  Also, the worker threads tasks are
 549   // generally longer, and hence the cost of the lock may perchance be amortized
 550   // over the long task's lifetime.
 551   ThreadData* current_thread_data = stopwatch.GetThreadData();
 552   if (!current_thread_data)
 553     return;
 554
 555   TrackedTime start_of_run = stopwatch.StartTime();
 556   int32 queue_duration = 0;
 557   if (!start_of_run.is_null()) {
 558     queue_duration = (start_of_run - time_posted).InMilliseconds();
 559   }
 560   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 561 }
 562
 563 // static
 564 void ThreadData::TallyRunInAScopedRegionIfTracking(
 565     const Births* birth,
 566     const TaskStopwatch& stopwatch) {
 567   if (!kTrackAllTaskObjects)
 568     return;  // Not compiled in.
 569
 570   // Even if we have been DEACTIVATED, we will process any pending births so
 571   // that our data structures (which counted the outstanding births) remain
 572   // consistent.
 573   if (!birth)
 574     return;
 575
 576   ThreadData* current_thread_data = stopwatch.GetThreadData();
 577   if (!current_thread_data)
 578     return;
 579
 580   int32 queue_duration = 0;
 581   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 582 }
 583
 584 // static
 585 void ThreadData::SnapshotAllExecutedTasks(ProcessDataSnapshot* process_data,
 586                                           BirthCountMap* birth_counts) {
 587   if (!kTrackAllTaskObjects)
 588     return;  // Not compiled in.
 589
 590   // Get an unchanging copy of a ThreadData list.
 591   ThreadData* my_list = ThreadData::first();
 592
 593   // Gather data serially.
 594   // This hackish approach *can* get some slighly corrupt tallies, as we are
 595   // grabbing values without the protection of a lock, but it has the advantage
 596   // of working even with threads that don't have message loops.  If a user
 597   // sees any strangeness, they can always just run their stats gathering a
 598   // second time.
 599   for (ThreadData* thread_data = my_list;
 600        thread_data;
 601        thread_data = thread_data->next()) {
 602     thread_data->SnapshotExecutedTasks(process_data, birth_counts);
 603   }
 604 }
 605
 606 void ThreadData::SnapshotExecutedTasks(ProcessDataSnapshot* process_data,
 607                                        BirthCountMap* birth_counts) {
 608   // Get copy of data, so that the data will not change during the iterations
 609   // and processing.
 610   ThreadData::BirthMap birth_map;
 611   ThreadData::DeathMap death_map;
 612   ThreadData::ParentChildSet parent_child_set;
 613   SnapshotMaps(&birth_map, &death_map, &parent_child_set);
 614
 615   for (ThreadData::DeathMap::const_iterator it = death_map.begin();
 616        it != death_map.end(); ++it) {
 617     process_data->tasks.push_back(
 618         TaskSnapshot(*it->first, it->second, thread_name()));
 619     (*birth_counts)[it->first] -= it->first->birth_count();
 620   }
 621
 622   for (ThreadData::BirthMap::const_iterator it = birth_map.begin();
 623        it != birth_map.end(); ++it) {
 624     (*birth_counts)[it->second] += it->second->birth_count();
 625   }
 626
 627   if (!kTrackParentChildLinks)
 628     return;
 629
 630   for (ThreadData::ParentChildSet::const_iterator it = parent_child_set.begin();
 631        it != parent_child_set.end(); ++it) {
 632     process_data->descendants.push_back(ParentChildPairSnapshot(*it));
 633   }
 634 }
 635
 636 // This may be called from another thread.
 637 void ThreadData::SnapshotMaps(BirthMap* birth_map,
 638                               DeathMap* death_map,
 639                               ParentChildSet* parent_child_set) {
 640   base::AutoLock lock(map_lock_);
 641   for (BirthMap::const_iterator it = birth_map_.begin();
 642        it != birth_map_.end(); ++it)
 643     (*birth_map)[it->first] = it->second;
 644   for (DeathMap::iterator it = death_map_.begin();
 645        it != death_map_.end(); ++it) {
 646     (*death_map)[it->first] = it->second;
 647   }
 648
 649   if (!kTrackParentChildLinks)
 650     return;
 651
 652   for (ParentChildSet::iterator it = parent_child_set_.begin();
 653        it != parent_child_set_.end(); ++it)
 654     parent_child_set->insert(*it);
 655 }
 656
 657 // static
 658 void ThreadData::ResetAllThreadData() {
 659   ThreadData* my_list = first();
 660
 661   for (ThreadData* thread_data = my_list;
 662        thread_data;
 663        thread_data = thread_data->next())
 664     thread_data->Reset();
 665 }
 666
 667 void ThreadData::Reset() {
 668   base::AutoLock lock(map_lock_);
 669   for (DeathMap::iterator it = death_map_.begin();
 670        it != death_map_.end(); ++it)
 671     it->second.Clear();
 672   for (BirthMap::iterator it = birth_map_.begin();
 673        it != birth_map_.end(); ++it)
 674     it->second->Clear();
 675 }
 676
 677 static void OptionallyInitializeAlternateTimer() {
 678   NowFunction* alternate_time_source = GetAlternateTimeSource();
 679   if (alternate_time_source)
 680     ThreadData::SetAlternateTimeSource(alternate_time_source);
 681 }
 682
 683 bool ThreadData::Initialize() {
 684   if (!kTrackAllTaskObjects)
 685     return false;  // Not compiled in.
 686   if (status_ >= DEACTIVATED)
 687     return true;  // Someone else did the initialization.
 688   // Due to racy lazy initialization in tests, we'll need to recheck status_
 689   // after we acquire the lock.
 690
 691   // Ensure that we don't double initialize tls.  We are called when single
 692   // threaded in the product, but some tests may be racy and lazy about our
 693   // initialization.
 694   base::AutoLock lock(*list_lock_.Pointer());
 695   if (status_ >= DEACTIVATED)
 696     return true;  // Someone raced in here and beat us.
 697
 698   // Put an alternate timer in place if the environment calls for it, such as
 699   // for tracking TCMalloc allocations.  This insertion is idempotent, so we
 700   // don't mind if there is a race, and we'd prefer not to be in a lock while
 701   // doing this work.
 702   if (kAllowAlternateTimeSourceHandling)
 703     OptionallyInitializeAlternateTimer();
 704
 705   // Perform the "real" TLS initialization now, and leave it intact through
 706   // process termination.
 707   if (!tls_index_.initialized()) {  // Testing may have initialized this.
 708     DCHECK_EQ(status_, UNINITIALIZED);
 709     tls_index_.Initialize(&ThreadData::OnThreadTermination);
 710     if (!tls_index_.initialized())
 711       return false;
 712   } else {
 713     // TLS was initialzed for us earlier.
 714     DCHECK_EQ(status_, DORMANT_DURING_TESTS);
 715   }
 716
 717   // Incarnation counter is only significant to testing, as it otherwise will
 718   // never again change in this process.
 719   ++incarnation_counter_;
 720
 721   // The lock is not critical for setting status_, but it doesn't hurt. It also
 722   // ensures that if we have a racy initialization, that we'll bail as soon as
 723   // we get the lock earlier in this method.
 724   status_ = kInitialStartupState;
 725   if (!kTrackParentChildLinks &&
 726       kInitialStartupState == PROFILING_CHILDREN_ACTIVE)
 727     status_ = PROFILING_ACTIVE;
 728   DCHECK(status_ != UNINITIALIZED);
 729   return true;
 730 }
 731
 732 // static
 733 bool ThreadData::InitializeAndSetTrackingStatus(Status status) {
 734   DCHECK_GE(status, DEACTIVATED);
 735   DCHECK_LE(status, PROFILING_CHILDREN_ACTIVE);
 736
 737   if (!Initialize())  // No-op if already initialized.
 738     return false;  // Not compiled in.
 739
 740   if (!kTrackParentChildLinks && status > DEACTIVATED)
 741     status = PROFILING_ACTIVE;
 742   status_ = status;
 743   return true;
 744 }
 745
 746 // static
 747 ThreadData::Status ThreadData::status() {
 748   return status_;
 749 }
 750
 751 // static
 752 bool ThreadData::TrackingStatus() {
 753   return status_ > DEACTIVATED;
 754 }
 755
 756 // static
 757 bool ThreadData::TrackingParentChildStatus() {
 758   return status_ >= PROFILING_CHILDREN_ACTIVE;
 759 }
 760
 761 // static
 762 void ThreadData::PrepareForStartOfRun(const Births* parent) {
 763   if (kTrackParentChildLinks && parent && status_ > PROFILING_ACTIVE) {
 764     ThreadData* current_thread_data = Get();
 765     if (current_thread_data)
 766       current_thread_data->parent_stack_.push(parent);
 767   }
 768 }
 769
 770 // static
 771 void ThreadData::SetAlternateTimeSource(NowFunction* now_function) {
 772   DCHECK(now_function);
 773   if (kAllowAlternateTimeSourceHandling)
 774     now_function_ = now_function;
 775 }
 776
 777 // static
 778 void ThreadData::EnableProfilerTiming() {
 779   base::subtle::NoBarrier_Store(&g_profiler_timing_enabled, ENABLED_TIMING);
 780 }
 781
 782 // static
 783 TrackedTime ThreadData::Now() {
 784   if (kAllowAlternateTimeSourceHandling && now_function_)
 785     return TrackedTime::FromMilliseconds((*now_function_)());
 786   if (kTrackAllTaskObjects && IsProfilerTimingEnabled() && TrackingStatus())
 787     return TrackedTime::Now();
 788   return TrackedTime();  // Super fast when disabled, or not compiled.
 789 }
 790
 791 // static
 792 void ThreadData::EnsureCleanupWasCalled(int major_threads_shutdown_count) {
 793   base::AutoLock lock(*list_lock_.Pointer());
 794   if (worker_thread_data_creation_count_ == 0)
 795     return;  // We haven't really run much, and couldn't have leaked.
 796
 797   // TODO(jar): until this is working on XP, don't run the real test.
 798 #if 0
 799   // Verify that we've at least shutdown/cleanup the major namesd threads.  The
 800   // caller should tell us how many thread shutdowns should have taken place by
 801   // now.
 802   CHECK_GT(cleanup_count_, major_threads_shutdown_count);
 803 #endif
 804 }
 805
 806 // static
 807 void ThreadData::ShutdownSingleThreadedCleanup(bool leak) {
 808   // This is only called from test code, where we need to cleanup so that
 809   // additional tests can be run.
 810   // We must be single threaded... but be careful anyway.
 811   if (!InitializeAndSetTrackingStatus(DEACTIVATED))
 812     return;
 813   ThreadData* thread_data_list;
 814   {
 815     base::AutoLock lock(*list_lock_.Pointer());
 816     thread_data_list = all_thread_data_list_head_;
 817     all_thread_data_list_head_ = NULL;
 818     ++incarnation_counter_;
 819     // To be clean, break apart the retired worker list (though we leak them).
 820     while (first_retired_worker_) {
 821       ThreadData* worker = first_retired_worker_;
 822       CHECK_GT(worker->worker_thread_number_, 0);
 823       first_retired_worker_ = worker->next_retired_worker_;
 824       worker->next_retired_worker_ = NULL;
 825     }
 826   }
 827
 828   // Put most global static back in pristine shape.
 829   worker_thread_data_creation_count_ = 0;
 830   cleanup_count_ = 0;
 831   tls_index_.Set(NULL);
 832   status_ = DORMANT_DURING_TESTS;  // Almost UNINITIALIZED.
 833
 834   // To avoid any chance of racing in unit tests, which is the only place we
 835   // call this function, we may sometimes leak all the data structures we
 836   // recovered, as they may still be in use on threads from prior tests!
 837   if (leak) {
 838     ThreadData* thread_data = thread_data_list;
 839     while (thread_data) {
 840       ANNOTATE_LEAKING_OBJECT_PTR(thread_data);
 841       thread_data = thread_data->next();
 842     }
 843     return;
 844   }
 845
 846   // When we want to cleanup (on a single thread), here is what we do.
 847
 848   // Do actual recursive delete in all ThreadData instances.
 849   while (thread_data_list) {
 850     ThreadData* next_thread_data = thread_data_list;
 851     thread_data_list = thread_data_list->next();
 852
 853     for (BirthMap::iterator it = next_thread_data->birth_map_.begin();
 854          next_thread_data->birth_map_.end() != it; ++it)
 855       delete it->second;  // Delete the Birth Records.
 856     delete next_thread_data;  // Includes all Death Records.
 857   }
 858 }
 859
 860 //------------------------------------------------------------------------------
 861 TaskStopwatch::TaskStopwatch()
 862     : wallclock_duration_ms_(0),
 863       current_thread_data_(NULL),
 864       excluded_duration_ms_(0),
 865       parent_(NULL) {
 866 #if DCHECK_IS_ON()
 867   state_ = CREATED;
 868   child_ = NULL;
 869 #endif
 870 }
 871
 872 TaskStopwatch::~TaskStopwatch() {
 873 #if DCHECK_IS_ON()
 874   DCHECK(state_ != RUNNING);
 875   DCHECK(child_ == NULL);
 876 #endif
 877 }
 878
 879 void TaskStopwatch::Start() {
 880 #if DCHECK_IS_ON()
 881   DCHECK(state_ == CREATED);
 882   state_ = RUNNING;
 883 #endif
 884
 885   start_time_ = ThreadData::Now();
 886
 887   current_thread_data_ = ThreadData::Get();
 888   if (!current_thread_data_)
 889     return;
 890
 891   parent_ = current_thread_data_->current_stopwatch_;
 892 #if DCHECK_IS_ON()
 893   if (parent_) {
 894     DCHECK(parent_->state_ == RUNNING);
 895     DCHECK(parent_->child_ == NULL);
 896     parent_->child_ = this;
 897   }
 898 #endif
 899   current_thread_data_->current_stopwatch_ = this;
 900 }
 901
 902 void TaskStopwatch::Stop() {
 903   const TrackedTime end_time = ThreadData::Now();
 904 #if DCHECK_IS_ON()
 905   DCHECK(state_ == RUNNING);
 906   state_ = STOPPED;
 907   DCHECK(child_ == NULL);
 908 #endif
 909
 910   if (!start_time_.is_null() && !end_time.is_null()) {
 911     wallclock_duration_ms_ = (end_time - start_time_).InMilliseconds();
 912   }
 913
 914   if (!current_thread_data_)
 915     return;
 916
 917   DCHECK(current_thread_data_->current_stopwatch_ == this);
 918   current_thread_data_->current_stopwatch_ = parent_;
 919   if (!parent_)
 920     return;
 921
 922 #if DCHECK_IS_ON()
 923   DCHECK(parent_->state_ == RUNNING);
 924   DCHECK(parent_->child_ == this);
 925   parent_->child_ = NULL;
 926 #endif
 927   parent_->excluded_duration_ms_ += wallclock_duration_ms_;
 928   parent_ = NULL;
 929 }
 930
 931 TrackedTime TaskStopwatch::StartTime() const {
 932 #if DCHECK_IS_ON()
 933   DCHECK(state_ != CREATED);
 934 #endif
 935
 936   return start_time_;
 937 }
 938
 939 int32 TaskStopwatch::RunDurationMs() const {
 940 #if DCHECK_IS_ON()
 941   DCHECK(state_ == STOPPED);
 942 #endif
 943
 944   return wallclock_duration_ms_ - excluded_duration_ms_;
 945 }
 946
 947 ThreadData* TaskStopwatch::GetThreadData() const {
 948 #if DCHECK_IS_ON()
 949   DCHECK(state_ != CREATED);
 950 #endif
 951
 952   return current_thread_data_;
 953 }
 954
 955 //------------------------------------------------------------------------------
 956 TaskSnapshot::TaskSnapshot() {
 957 }
 958
 959 TaskSnapshot::TaskSnapshot(const BirthOnThread& birth,
 960                            const DeathData& death_data,
 961                            const std::string& death_thread_name)
 962     : birth(birth),
 963       death_data(death_data),
 964       death_thread_name(death_thread_name) {
 965 }
 966
 967 TaskSnapshot::~TaskSnapshot() {
 968 }
 969
 970 //------------------------------------------------------------------------------
 971 // ParentChildPairSnapshot
 972
 973 ParentChildPairSnapshot::ParentChildPairSnapshot() {
 974 }
 975
 976 ParentChildPairSnapshot::ParentChildPairSnapshot(
 977     const ThreadData::ParentChildPair& parent_child)
 978     : parent(*parent_child.first),
 979       child(*parent_child.second) {
 980 }
 981
 982 ParentChildPairSnapshot::~ParentChildPairSnapshot() {
 983 }
 984
 985 //------------------------------------------------------------------------------
 986 // ProcessDataSnapshot
 987
 988 ProcessDataSnapshot::ProcessDataSnapshot()
 989 #if !defined(OS_NACL)
 990     : process_id(base::GetCurrentProcId()) {
 991 #else
 992     : process_id(0) {
 993 #endif
 994 }
 995
 996 ProcessDataSnapshot::~ProcessDataSnapshot() {
 997 }
 998
 999 }  // namespace tracked_objects