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) {
 394   // Add births that have run to completion to |collected_data|.
 395   // |birth_counts| tracks the total number of births recorded at each location
 396   // for which we have not seen a death count.
 397   BirthCountMap birth_counts;
 398   ThreadData::SnapshotAllExecutedTasks(process_data, &birth_counts);
 399
 400   // Add births that are still active -- i.e. objects that have tallied a birth,
 401   // but have not yet tallied a matching death, and hence must be either
 402   // running, queued up, or being held in limbo for future posting.
 403   for (BirthCountMap::const_iterator it = birth_counts.begin();
 404        it != birth_counts.end(); ++it) {
 405     if (it->second > 0) {
 406       process_data->tasks.push_back(
 407           TaskSnapshot(*it->first, DeathData(it->second), "Still_Alive"));
 408     }
 409   }
 410 }
 411
 412 Births* ThreadData::TallyABirth(const Location& location) {
 413   BirthMap::iterator it = birth_map_.find(location);
 414   Births* child;
 415   if (it != birth_map_.end()) {
 416     child =  it->second;
 417     child->RecordBirth();
 418   } else {
 419     child = new Births(location, *this);  // Leak this.
 420     // Lock since the map may get relocated now, and other threads sometimes
 421     // snapshot it (but they lock before copying it).
 422     base::AutoLock lock(map_lock_);
 423     birth_map_[location] = child;
 424   }
 425
 426   if (kTrackParentChildLinks && status_ > PROFILING_ACTIVE &&
 427       !parent_stack_.empty()) {
 428     const Births* parent = parent_stack_.top();
 429     ParentChildPair pair(parent, child);
 430     if (parent_child_set_.find(pair) == parent_child_set_.end()) {
 431       // Lock since the map may get relocated now, and other threads sometimes
 432       // snapshot it (but they lock before copying it).
 433       base::AutoLock lock(map_lock_);
 434       parent_child_set_.insert(pair);
 435     }
 436   }
 437
 438   return child;
 439 }
 440
 441 void ThreadData::TallyADeath(const Births& birth,
 442                              int32 queue_duration,
 443                              const TaskStopwatch& stopwatch) {
 444   int32 run_duration = stopwatch.RunDurationMs();
 445
 446   // Stir in some randomness, plus add constant in case durations are zero.
 447   const uint32 kSomePrimeNumber = 2147483647;
 448   random_number_ += queue_duration + run_duration + kSomePrimeNumber;
 449   // An address is going to have some randomness to it as well ;-).
 450   random_number_ ^= static_cast<uint32>(&birth - reinterpret_cast<Births*>(0));
 451
 452   // We don't have queue durations without OS timer. OS timer is automatically
 453   // used for task-post-timing, so the use of an alternate timer implies all
 454   // queue times are invalid, unless it was explicitly said that we can trust
 455   // the alternate timer.
 456   if (kAllowAlternateTimeSourceHandling &&
 457       now_function_ &&
 458       !now_function_is_time_) {
 459     queue_duration = 0;
 460   }
 461
 462   DeathMap::iterator it = death_map_.find(&birth);
 463   DeathData* death_data;
 464   if (it != death_map_.end()) {
 465     death_data = &it->second;
 466   } else {
 467     base::AutoLock lock(map_lock_);  // Lock as the map may get relocated now.
 468     death_data = &death_map_[&birth];
 469   }  // Release lock ASAP.
 470   death_data->RecordDeath(queue_duration, run_duration, random_number_);
 471
 472   if (!kTrackParentChildLinks)
 473     return;
 474   if (!parent_stack_.empty()) {  // We might get turned off.
 475     DCHECK_EQ(parent_stack_.top(), &birth);
 476     parent_stack_.pop();
 477   }
 478 }
 479
 480 // static
 481 Births* ThreadData::TallyABirthIfActive(const Location& location) {
 482   if (!kTrackAllTaskObjects)
 483     return NULL;  // Not compiled in.
 484
 485   if (!TrackingStatus())
 486     return NULL;
 487   ThreadData* current_thread_data = Get();
 488   if (!current_thread_data)
 489     return NULL;
 490   return current_thread_data->TallyABirth(location);
 491 }
 492
 493 // static
 494 void ThreadData::TallyRunOnNamedThreadIfTracking(
 495     const base::TrackingInfo& completed_task,
 496     const TaskStopwatch& stopwatch) {
 497   if (!kTrackAllTaskObjects)
 498     return;  // Not compiled in.
 499
 500   // Even if we have been DEACTIVATED, we will process any pending births so
 501   // that our data structures (which counted the outstanding births) remain
 502   // consistent.
 503   const Births* birth = completed_task.birth_tally;
 504   if (!birth)
 505     return;
 506   ThreadData* current_thread_data = stopwatch.GetThreadData();
 507   if (!current_thread_data)
 508     return;
 509
 510   // Watch out for a race where status_ is changing, and hence one or both
 511   // of start_of_run or end_of_run is zero.  In that case, we didn't bother to
 512   // get a time value since we "weren't tracking" and we were trying to be
 513   // efficient by not calling for a genuine time value. For simplicity, we'll
 514   // use a default zero duration when we can't calculate a true value.
 515   TrackedTime start_of_run = stopwatch.StartTime();
 516   int32 queue_duration = 0;
 517   if (!start_of_run.is_null()) {
 518     queue_duration = (start_of_run - completed_task.EffectiveTimePosted())
 519         .InMilliseconds();
 520   }
 521   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 522 }
 523
 524 // static
 525 void ThreadData::TallyRunOnWorkerThreadIfTracking(
 526     const Births* birth,
 527     const TrackedTime& time_posted,
 528     const TaskStopwatch& stopwatch) {
 529   if (!kTrackAllTaskObjects)
 530     return;  // Not compiled in.
 531
 532   // Even if we have been DEACTIVATED, we will process any pending births so
 533   // that our data structures (which counted the outstanding births) remain
 534   // consistent.
 535   if (!birth)
 536     return;
 537
 538   // TODO(jar): Support the option to coalesce all worker-thread activity under
 539   // one ThreadData instance that uses locks to protect *all* access.  This will
 540   // reduce memory (making it provably bounded), but run incrementally slower
 541   // (since we'll use locks on TallyABirth and TallyADeath).  The good news is
 542   // that the locks on TallyADeath will be *after* the worker thread has run,
 543   // and hence nothing will be waiting for the completion (... besides some
 544   // other thread that might like to run).  Also, the worker threads tasks are
 545   // generally longer, and hence the cost of the lock may perchance be amortized
 546   // over the long task's lifetime.
 547   ThreadData* current_thread_data = stopwatch.GetThreadData();
 548   if (!current_thread_data)
 549     return;
 550
 551   TrackedTime start_of_run = stopwatch.StartTime();
 552   int32 queue_duration = 0;
 553   if (!start_of_run.is_null()) {
 554     queue_duration = (start_of_run - time_posted).InMilliseconds();
 555   }
 556   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 557 }
 558
 559 // static
 560 void ThreadData::TallyRunInAScopedRegionIfTracking(
 561     const Births* birth,
 562     const TaskStopwatch& stopwatch) {
 563   if (!kTrackAllTaskObjects)
 564     return;  // Not compiled in.
 565
 566   // Even if we have been DEACTIVATED, we will process any pending births so
 567   // that our data structures (which counted the outstanding births) remain
 568   // consistent.
 569   if (!birth)
 570     return;
 571
 572   ThreadData* current_thread_data = stopwatch.GetThreadData();
 573   if (!current_thread_data)
 574     return;
 575
 576   int32 queue_duration = 0;
 577   current_thread_data->TallyADeath(*birth, queue_duration, stopwatch);
 578 }
 579
 580 // static
 581 void ThreadData::SnapshotAllExecutedTasks(ProcessDataSnapshot* process_data,
 582                                           BirthCountMap* birth_counts) {
 583   if (!kTrackAllTaskObjects)
 584     return;  // Not compiled in.
 585
 586   // Get an unchanging copy of a ThreadData list.
 587   ThreadData* my_list = ThreadData::first();
 588
 589   // Gather data serially.
 590   // This hackish approach *can* get some slighly corrupt tallies, as we are
 591   // grabbing values without the protection of a lock, but it has the advantage
 592   // of working even with threads that don't have message loops.  If a user
 593   // sees any strangeness, they can always just run their stats gathering a
 594   // second time.
 595   for (ThreadData* thread_data = my_list;
 596        thread_data;
 597        thread_data = thread_data->next()) {
 598     thread_data->SnapshotExecutedTasks(process_data, birth_counts);
 599   }
 600 }
 601
 602 void ThreadData::SnapshotExecutedTasks(ProcessDataSnapshot* process_data,
 603                                        BirthCountMap* birth_counts) {
 604   // Get copy of data, so that the data will not change during the iterations
 605   // and processing.
 606   ThreadData::BirthMap birth_map;
 607   ThreadData::DeathMap death_map;
 608   ThreadData::ParentChildSet parent_child_set;
 609   SnapshotMaps(&birth_map, &death_map, &parent_child_set);
 610
 611   for (ThreadData::DeathMap::const_iterator it = death_map.begin();
 612        it != death_map.end(); ++it) {
 613     process_data->tasks.push_back(
 614         TaskSnapshot(*it->first, it->second, thread_name()));
 615     (*birth_counts)[it->first] -= it->first->birth_count();
 616   }
 617
 618   for (ThreadData::BirthMap::const_iterator it = birth_map.begin();
 619        it != birth_map.end(); ++it) {
 620     (*birth_counts)[it->second] += it->second->birth_count();
 621   }
 622
 623   if (!kTrackParentChildLinks)
 624     return;
 625
 626   for (ThreadData::ParentChildSet::const_iterator it = parent_child_set.begin();
 627        it != parent_child_set.end(); ++it) {
 628     process_data->descendants.push_back(ParentChildPairSnapshot(*it));
 629   }
 630 }
 631
 632 // This may be called from another thread.
 633 void ThreadData::SnapshotMaps(BirthMap* birth_map,
 634                               DeathMap* death_map,
 635                               ParentChildSet* parent_child_set) {
 636   base::AutoLock lock(map_lock_);
 637   for (BirthMap::const_iterator it = birth_map_.begin();
 638        it != birth_map_.end(); ++it)
 639     (*birth_map)[it->first] = it->second;
 640   for (DeathMap::iterator it = death_map_.begin();
 641        it != death_map_.end(); ++it) {
 642     (*death_map)[it->first] = it->second;
 643   }
 644
 645   if (!kTrackParentChildLinks)
 646     return;
 647
 648   for (ParentChildSet::iterator it = parent_child_set_.begin();
 649        it != parent_child_set_.end(); ++it)
 650     parent_child_set->insert(*it);
 651 }
 652
 653 static void OptionallyInitializeAlternateTimer() {
 654   NowFunction* alternate_time_source = GetAlternateTimeSource();
 655   if (alternate_time_source)
 656     ThreadData::SetAlternateTimeSource(alternate_time_source);
 657 }
 658
 659 bool ThreadData::Initialize() {
 660   if (!kTrackAllTaskObjects)
 661     return false;  // Not compiled in.
 662   if (status_ >= DEACTIVATED)
 663     return true;  // Someone else did the initialization.
 664   // Due to racy lazy initialization in tests, we'll need to recheck status_
 665   // after we acquire the lock.
 666
 667   // Ensure that we don't double initialize tls.  We are called when single
 668   // threaded in the product, but some tests may be racy and lazy about our
 669   // initialization.
 670   base::AutoLock lock(*list_lock_.Pointer());
 671   if (status_ >= DEACTIVATED)
 672     return true;  // Someone raced in here and beat us.
 673
 674   // Put an alternate timer in place if the environment calls for it, such as
 675   // for tracking TCMalloc allocations.  This insertion is idempotent, so we
 676   // don't mind if there is a race, and we'd prefer not to be in a lock while
 677   // doing this work.
 678   if (kAllowAlternateTimeSourceHandling)
 679     OptionallyInitializeAlternateTimer();
 680
 681   // Perform the "real" TLS initialization now, and leave it intact through
 682   // process termination.
 683   if (!tls_index_.initialized()) {  // Testing may have initialized this.
 684     DCHECK_EQ(status_, UNINITIALIZED);
 685     tls_index_.Initialize(&ThreadData::OnThreadTermination);
 686     if (!tls_index_.initialized())
 687       return false;
 688   } else {
 689     // TLS was initialzed for us earlier.
 690     DCHECK_EQ(status_, DORMANT_DURING_TESTS);
 691   }
 692
 693   // Incarnation counter is only significant to testing, as it otherwise will
 694   // never again change in this process.
 695   ++incarnation_counter_;
 696
 697   // The lock is not critical for setting status_, but it doesn't hurt. It also
 698   // ensures that if we have a racy initialization, that we'll bail as soon as
 699   // we get the lock earlier in this method.
 700   status_ = kInitialStartupState;
 701   if (!kTrackParentChildLinks &&
 702       kInitialStartupState == PROFILING_CHILDREN_ACTIVE)
 703     status_ = PROFILING_ACTIVE;
 704   DCHECK(status_ != UNINITIALIZED);
 705   return true;
 706 }
 707
 708 // static
 709 bool ThreadData::InitializeAndSetTrackingStatus(Status status) {
 710   DCHECK_GE(status, DEACTIVATED);
 711   DCHECK_LE(status, PROFILING_CHILDREN_ACTIVE);
 712
 713   if (!Initialize())  // No-op if already initialized.
 714     return false;  // Not compiled in.
 715
 716   if (!kTrackParentChildLinks && status > DEACTIVATED)
 717     status = PROFILING_ACTIVE;
 718   status_ = status;
 719   return true;
 720 }
 721
 722 // static
 723 ThreadData::Status ThreadData::status() {
 724   return status_;
 725 }
 726
 727 // static
 728 bool ThreadData::TrackingStatus() {
 729   return status_ > DEACTIVATED;
 730 }
 731
 732 // static
 733 bool ThreadData::TrackingParentChildStatus() {
 734   return status_ >= PROFILING_CHILDREN_ACTIVE;
 735 }
 736
 737 // static
 738 void ThreadData::PrepareForStartOfRun(const Births* parent) {
 739   if (kTrackParentChildLinks && parent && status_ > PROFILING_ACTIVE) {
 740     ThreadData* current_thread_data = Get();
 741     if (current_thread_data)
 742       current_thread_data->parent_stack_.push(parent);
 743   }
 744 }
 745
 746 // static
 747 void ThreadData::SetAlternateTimeSource(NowFunction* now_function) {
 748   DCHECK(now_function);
 749   if (kAllowAlternateTimeSourceHandling)
 750     now_function_ = now_function;
 751 }
 752
 753 // static
 754 void ThreadData::EnableProfilerTiming() {
 755   base::subtle::NoBarrier_Store(&g_profiler_timing_enabled, ENABLED_TIMING);
 756 }
 757
 758 // static
 759 TrackedTime ThreadData::Now() {
 760   if (kAllowAlternateTimeSourceHandling && now_function_)
 761     return TrackedTime::FromMilliseconds((*now_function_)());
 762   if (kTrackAllTaskObjects && IsProfilerTimingEnabled() && TrackingStatus())
 763     return TrackedTime::Now();
 764   return TrackedTime();  // Super fast when disabled, or not compiled.
 765 }
 766
 767 // static
 768 void ThreadData::EnsureCleanupWasCalled(int major_threads_shutdown_count) {
 769   base::AutoLock lock(*list_lock_.Pointer());
 770   if (worker_thread_data_creation_count_ == 0)
 771     return;  // We haven't really run much, and couldn't have leaked.
 772
 773   // TODO(jar): until this is working on XP, don't run the real test.
 774 #if 0
 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   CHECK_GT(cleanup_count_, major_threads_shutdown_count);
 779 #endif
 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     ThreadData* thread_data = thread_data_list;
 815     while (thread_data) {
 816       ANNOTATE_LEAKING_OBJECT_PTR(thread_data);
 817       thread_data = thread_data->next();
 818     }
 819     return;
 820   }
 821
 822   // When we want to cleanup (on a single thread), here is what we do.
 823
 824   // Do actual recursive delete in all ThreadData instances.
 825   while (thread_data_list) {
 826     ThreadData* next_thread_data = thread_data_list;
 827     thread_data_list = thread_data_list->next();
 828
 829     for (BirthMap::iterator it = next_thread_data->birth_map_.begin();
 830          next_thread_data->birth_map_.end() != it; ++it)
 831       delete it->second;  // Delete the Birth Records.
 832     delete next_thread_data;  // Includes all Death Records.
 833   }
 834 }
 835
 836 //------------------------------------------------------------------------------
 837 TaskStopwatch::TaskStopwatch()
 838     : wallclock_duration_ms_(0),
 839       current_thread_data_(NULL),
 840       excluded_duration_ms_(0),
 841       parent_(NULL) {
 842 #if DCHECK_IS_ON()
 843   state_ = CREATED;
 844   child_ = NULL;
 845 #endif
 846 }
 847
 848 TaskStopwatch::~TaskStopwatch() {
 849 #if DCHECK_IS_ON()
 850   DCHECK(state_ != RUNNING);
 851   DCHECK(child_ == NULL);
 852 #endif
 853 }
 854
 855 void TaskStopwatch::Start() {
 856 #if DCHECK_IS_ON()
 857   DCHECK(state_ == CREATED);
 858   state_ = RUNNING;
 859 #endif
 860
 861   start_time_ = ThreadData::Now();
 862
 863   current_thread_data_ = ThreadData::Get();
 864   if (!current_thread_data_)
 865     return;
 866
 867   parent_ = current_thread_data_->current_stopwatch_;
 868 #if DCHECK_IS_ON()
 869   if (parent_) {
 870     DCHECK(parent_->state_ == RUNNING);
 871     DCHECK(parent_->child_ == NULL);
 872     parent_->child_ = this;
 873   }
 874 #endif
 875   current_thread_data_->current_stopwatch_ = this;
 876 }
 877
 878 void TaskStopwatch::Stop() {
 879   const TrackedTime end_time = ThreadData::Now();
 880 #if DCHECK_IS_ON()
 881   DCHECK(state_ == RUNNING);
 882   state_ = STOPPED;
 883   DCHECK(child_ == NULL);
 884 #endif
 885
 886   if (!start_time_.is_null() && !end_time.is_null()) {
 887     wallclock_duration_ms_ = (end_time - start_time_).InMilliseconds();
 888   }
 889
 890   if (!current_thread_data_)
 891     return;
 892
 893   DCHECK(current_thread_data_->current_stopwatch_ == this);
 894   current_thread_data_->current_stopwatch_ = parent_;
 895   if (!parent_)
 896     return;
 897
 898 #if DCHECK_IS_ON()
 899   DCHECK(parent_->state_ == RUNNING);
 900   DCHECK(parent_->child_ == this);
 901   parent_->child_ = NULL;
 902 #endif
 903   parent_->excluded_duration_ms_ += wallclock_duration_ms_;
 904   parent_ = NULL;
 905 }
 906
 907 TrackedTime TaskStopwatch::StartTime() const {
 908 #if DCHECK_IS_ON()
 909   DCHECK(state_ != CREATED);
 910 #endif
 911
 912   return start_time_;
 913 }
 914
 915 int32 TaskStopwatch::RunDurationMs() const {
 916 #if DCHECK_IS_ON()
 917   DCHECK(state_ == STOPPED);
 918 #endif
 919
 920   return wallclock_duration_ms_ - excluded_duration_ms_;
 921 }
 922
 923 ThreadData* TaskStopwatch::GetThreadData() const {
 924 #if DCHECK_IS_ON()
 925   DCHECK(state_ != CREATED);
 926 #endif
 927
 928   return current_thread_data_;
 929 }
 930
 931 //------------------------------------------------------------------------------
 932 TaskSnapshot::TaskSnapshot() {
 933 }
 934
 935 TaskSnapshot::TaskSnapshot(const BirthOnThread& birth,
 936                            const DeathData& death_data,
 937                            const std::string& death_thread_name)
 938     : birth(birth),
 939       death_data(death_data),
 940       death_thread_name(death_thread_name) {
 941 }
 942
 943 TaskSnapshot::~TaskSnapshot() {
 944 }
 945
 946 //------------------------------------------------------------------------------
 947 // ParentChildPairSnapshot
 948
 949 ParentChildPairSnapshot::ParentChildPairSnapshot() {
 950 }
 951
 952 ParentChildPairSnapshot::ParentChildPairSnapshot(
 953     const ThreadData::ParentChildPair& parent_child)
 954     : parent(*parent_child.first),
 955       child(*parent_child.second) {
 956 }
 957
 958 ParentChildPairSnapshot::~ParentChildPairSnapshot() {
 959 }
 960
 961 //------------------------------------------------------------------------------
 962 // ProcessDataSnapshot
 963
 964 ProcessDataSnapshot::ProcessDataSnapshot()
 965 #if !defined(OS_NACL)
 966     : process_id(base::GetCurrentProcId()) {
 967 #else
 968     : process_id(0) {
 969 #endif
 970 }
 971
 972 ProcessDataSnapshot::~ProcessDataSnapshot() {
 973 }
 974
 975 }  // namespace tracked_objects