thread_list.cc revision dcfd6cad9670ca3ae460a4af0bbe7d88cae21a3c
1/*
2 * Copyright (C) 2011 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "thread_list.h"
18
19#define ATRACE_TAG ATRACE_TAG_DALVIK
20
21#include <cutils/trace.h>
22#include <dirent.h>
23#include <ScopedLocalRef.h>
24#include <ScopedUtfChars.h>
25#include <sys/types.h>
26#include <unistd.h>
27
28#include "base/mutex.h"
29#include "base/mutex-inl.h"
30#include "base/timing_logger.h"
31#include "debugger.h"
32#include "jni_internal.h"
33#include "lock_word.h"
34#include "monitor.h"
35#include "scoped_thread_state_change.h"
36#include "thread.h"
37#include "utils.h"
38#include "well_known_classes.h"
39
40namespace art {
41
42static constexpr uint64_t kLongThreadSuspendThreshold = MsToNs(5);
43
44ThreadList::ThreadList()
45    : suspend_all_count_(0), debug_suspend_all_count_(0),
46      thread_exit_cond_("thread exit condition variable", *Locks::thread_list_lock_) {
47  CHECK(Monitor::IsValidLockWord(LockWord::FromThinLockId(kMaxThreadId, 1)));
48}
49
50ThreadList::~ThreadList() {
51  // Detach the current thread if necessary. If we failed to start, there might not be any threads.
52  // We need to detach the current thread here in case there's another thread waiting to join with
53  // us.
54  if (Contains(Thread::Current())) {
55    Runtime::Current()->DetachCurrentThread();
56  }
57
58  WaitForOtherNonDaemonThreadsToExit();
59  // TODO: there's an unaddressed race here where a thread may attach during shutdown, see
60  //       Thread::Init.
61  SuspendAllDaemonThreads();
62}
63
64bool ThreadList::Contains(Thread* thread) {
65  return find(list_.begin(), list_.end(), thread) != list_.end();
66}
67
68bool ThreadList::Contains(pid_t tid) {
69  for (const auto& thread : list_) {
70    if (thread->GetTid() == tid) {
71      return true;
72    }
73  }
74  return false;
75}
76
77pid_t ThreadList::GetLockOwner() {
78  return Locks::thread_list_lock_->GetExclusiveOwnerTid();
79}
80
81void ThreadList::DumpNativeStacks(std::ostream& os) {
82  MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
83  for (const auto& thread : list_) {
84    os << "DUMPING THREAD " << thread->GetTid() << "\n";
85    DumpNativeStack(os, thread->GetTid(), "\t");
86    os << "\n";
87  }
88}
89
90void ThreadList::DumpForSigQuit(std::ostream& os) {
91  {
92    MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
93    DumpLocked(os);
94  }
95  DumpUnattachedThreads(os);
96}
97
98static void DumpUnattachedThread(std::ostream& os, pid_t tid) NO_THREAD_SAFETY_ANALYSIS {
99  // TODO: No thread safety analysis as DumpState with a NULL thread won't access fields, should
100  // refactor DumpState to avoid skipping analysis.
101  Thread::DumpState(os, NULL, tid);
102  DumpKernelStack(os, tid, "  kernel: ", false);
103  // TODO: Reenable this when the native code in system_server can handle it.
104  // Currently "adb shell kill -3 `pid system_server`" will cause it to exit.
105  if (false) {
106    DumpNativeStack(os, tid, "  native: ");
107  }
108  os << "\n";
109}
110
111void ThreadList::DumpUnattachedThreads(std::ostream& os) {
112  DIR* d = opendir("/proc/self/task");
113  if (!d) {
114    return;
115  }
116
117  Thread* self = Thread::Current();
118  dirent* e;
119  while ((e = readdir(d)) != NULL) {
120    char* end;
121    pid_t tid = strtol(e->d_name, &end, 10);
122    if (!*end) {
123      bool contains;
124      {
125        MutexLock mu(self, *Locks::thread_list_lock_);
126        contains = Contains(tid);
127      }
128      if (!contains) {
129        DumpUnattachedThread(os, tid);
130      }
131    }
132  }
133  closedir(d);
134}
135
136void ThreadList::DumpLocked(std::ostream& os) {
137  os << "DALVIK THREADS (" << list_.size() << "):\n";
138  for (const auto& thread : list_) {
139    thread->Dump(os);
140    os << "\n";
141  }
142}
143
144void ThreadList::AssertThreadsAreSuspended(Thread* self, Thread* ignore1, Thread* ignore2) {
145  MutexLock mu(self, *Locks::thread_list_lock_);
146  MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
147  for (const auto& thread : list_) {
148    if (thread != ignore1 && thread != ignore2) {
149      CHECK(thread->IsSuspended())
150            << "\nUnsuspended thread: <<" << *thread << "\n"
151            << "self: <<" << *Thread::Current();
152    }
153  }
154}
155
156#if HAVE_TIMED_RWLOCK
157// Attempt to rectify locks so that we dump thread list with required locks before exiting.
158static void UnsafeLogFatalForThreadSuspendAllTimeout() NO_THREAD_SAFETY_ANALYSIS __attribute__((noreturn));
159static void UnsafeLogFatalForThreadSuspendAllTimeout() {
160  Runtime* runtime = Runtime::Current();
161  std::ostringstream ss;
162  ss << "Thread suspend timeout\n";
163  runtime->GetThreadList()->DumpLocked(ss);
164  LOG(FATAL) << ss.str();
165  exit(0);
166}
167#endif
168
169// Unlike suspending all threads where we can wait to acquire the mutator_lock_, suspending an
170// individual thread requires polling. delay_us is the requested sleep and total_delay_us
171// accumulates the total time spent sleeping for timeouts. The first sleep is just a yield,
172// subsequently sleeps increase delay_us from 1ms to 500ms by doubling.
173static void ThreadSuspendSleep(Thread* self, useconds_t* delay_us, useconds_t* total_delay_us) {
174  useconds_t new_delay_us = (*delay_us) * 2;
175  CHECK_GE(new_delay_us, *delay_us);
176  if (new_delay_us < 500000) {  // Don't allow sleeping to be more than 0.5s.
177    *delay_us = new_delay_us;
178  }
179  if (*delay_us == 0) {
180    sched_yield();
181    // Default to 1 milliseconds (note that this gets multiplied by 2 before the first sleep).
182    *delay_us = 500;
183  } else {
184    usleep(*delay_us);
185    *total_delay_us += *delay_us;
186  }
187}
188
189size_t ThreadList::RunCheckpoint(Closure* checkpoint_function) {
190  Thread* self = Thread::Current();
191  Locks::mutator_lock_->AssertNotExclusiveHeld(self);
192  Locks::thread_list_lock_->AssertNotHeld(self);
193  Locks::thread_suspend_count_lock_->AssertNotHeld(self);
194  if (kDebugLocking) {
195    CHECK_NE(self->GetState(), kRunnable);
196  }
197
198  std::vector<Thread*> suspended_count_modified_threads;
199  size_t count = 0;
200  {
201    // Call a checkpoint function for each thread, threads which are suspend get their checkpoint
202    // manually called.
203    MutexLock mu(self, *Locks::thread_list_lock_);
204    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
205    for (const auto& thread : list_) {
206      if (thread != self) {
207        while (true) {
208          if (thread->RequestCheckpoint(checkpoint_function)) {
209            // This thread will run its checkpoint some time in the near future.
210            count++;
211            break;
212          } else {
213            // We are probably suspended, try to make sure that we stay suspended.
214            // The thread switched back to runnable.
215            if (thread->GetState() == kRunnable) {
216              // Spurious fail, try again.
217              continue;
218            }
219            thread->ModifySuspendCount(self, +1, false);
220            suspended_count_modified_threads.push_back(thread);
221            break;
222          }
223        }
224      }
225    }
226  }
227
228  // Run the checkpoint on ourself while we wait for threads to suspend.
229  checkpoint_function->Run(self);
230
231  // Run the checkpoint on the suspended threads.
232  for (const auto& thread : suspended_count_modified_threads) {
233    if (!thread->IsSuspended()) {
234      // Wait until the thread is suspended.
235      useconds_t total_delay_us = 0;
236      do {
237        useconds_t delay_us = 100;
238        ThreadSuspendSleep(self, &delay_us, &total_delay_us);
239      } while (!thread->IsSuspended());
240      // Shouldn't need to wait for longer than 1000 microseconds.
241      constexpr useconds_t kLongWaitThresholdUS = 1000;
242      if (UNLIKELY(total_delay_us > kLongWaitThresholdUS)) {
243        LOG(WARNING) << "Waited " << total_delay_us << " us for thread suspend!";
244      }
245    }
246    // We know for sure that the thread is suspended at this point.
247    checkpoint_function->Run(thread);
248    {
249      MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
250      thread->ModifySuspendCount(self, -1, false);
251    }
252  }
253
254  {
255    // Imitate ResumeAll, threads may be waiting on Thread::resume_cond_ since we raised their
256    // suspend count. Now the suspend_count_ is lowered so we must do the broadcast.
257    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
258    Thread::resume_cond_->Broadcast(self);
259  }
260
261  // Add one for self.
262  return count + suspended_count_modified_threads.size() + 1;
263}
264
265// Request that a checkpoint function be run on all active (non-suspended)
266// threads.  Returns the number of successful requests.
267size_t ThreadList::RunCheckpointOnRunnableThreads(Closure* checkpoint_function) {
268  Thread* self = Thread::Current();
269  if (kIsDebugBuild) {
270    Locks::mutator_lock_->AssertNotExclusiveHeld(self);
271    Locks::thread_list_lock_->AssertNotHeld(self);
272    Locks::thread_suspend_count_lock_->AssertNotHeld(self);
273    CHECK_NE(self->GetState(), kRunnable);
274  }
275
276  size_t count = 0;
277  {
278    // Call a checkpoint function for each non-suspended thread.
279    MutexLock mu(self, *Locks::thread_list_lock_);
280    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
281    for (const auto& thread : list_) {
282      if (thread != self) {
283        if (thread->RequestCheckpoint(checkpoint_function)) {
284          // This thread will run its checkpoint some time in the near future.
285          count++;
286        }
287      }
288    }
289  }
290
291  // Return the number of threads that will run the checkpoint function.
292  return count;
293}
294
295void ThreadList::SuspendAll() {
296  Thread* self = Thread::Current();
297
298  if (self != nullptr) {
299    VLOG(threads) << *self << " SuspendAll starting...";
300  } else {
301    VLOG(threads) << "Thread[null] SuspendAll starting...";
302  }
303  ATRACE_BEGIN("Suspending mutator threads");
304  uint64_t start_time = NanoTime();
305
306  Locks::mutator_lock_->AssertNotHeld(self);
307  Locks::thread_list_lock_->AssertNotHeld(self);
308  Locks::thread_suspend_count_lock_->AssertNotHeld(self);
309  if (kDebugLocking && self != nullptr) {
310    CHECK_NE(self->GetState(), kRunnable);
311  }
312  {
313    MutexLock mu(self, *Locks::thread_list_lock_);
314    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
315    // Update global suspend all state for attaching threads.
316    ++suspend_all_count_;
317    // Increment everybody's suspend count (except our own).
318    for (const auto& thread : list_) {
319      if (thread == self) {
320        continue;
321      }
322      VLOG(threads) << "requesting thread suspend: " << *thread;
323      thread->ModifySuspendCount(self, +1, false);
324    }
325  }
326
327  // Block on the mutator lock until all Runnable threads release their share of access.
328#if HAVE_TIMED_RWLOCK
329  // Timeout if we wait more than 30 seconds.
330  if (!Locks::mutator_lock_->ExclusiveLockWithTimeout(self, 30 * 1000, 0)) {
331    UnsafeLogFatalForThreadSuspendAllTimeout();
332  }
333#else
334  Locks::mutator_lock_->ExclusiveLock(self);
335#endif
336
337  uint64_t end_time = NanoTime();
338  if (end_time - start_time > kLongThreadSuspendThreshold) {
339    LOG(WARNING) << "Suspending all threads took: " << PrettyDuration(end_time - start_time);
340  }
341
342  if (kDebugLocking) {
343    // Debug check that all threads are suspended.
344    AssertThreadsAreSuspended(self, self);
345  }
346
347  ATRACE_END();
348  ATRACE_BEGIN("Mutator threads suspended");
349
350  if (self != nullptr) {
351    VLOG(threads) << *self << " SuspendAll complete";
352  } else {
353    VLOG(threads) << "Thread[null] SuspendAll complete";
354  }
355}
356
357void ThreadList::ResumeAll() {
358  Thread* self = Thread::Current();
359
360  if (self != nullptr) {
361    VLOG(threads) << *self << " ResumeAll starting";
362  } else {
363    VLOG(threads) << "Thread[null] ResumeAll starting";
364  }
365
366  ATRACE_END();
367  ATRACE_BEGIN("Resuming mutator threads");
368
369  if (kDebugLocking) {
370    // Debug check that all threads are suspended.
371    AssertThreadsAreSuspended(self, self);
372  }
373
374  Locks::mutator_lock_->ExclusiveUnlock(self);
375  {
376    MutexLock mu(self, *Locks::thread_list_lock_);
377    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
378    // Update global suspend all state for attaching threads.
379    --suspend_all_count_;
380    // Decrement the suspend counts for all threads.
381    for (const auto& thread : list_) {
382      if (thread == self) {
383        continue;
384      }
385      thread->ModifySuspendCount(self, -1, false);
386    }
387
388    // Broadcast a notification to all suspended threads, some or all of
389    // which may choose to wake up.  No need to wait for them.
390    if (self != nullptr) {
391      VLOG(threads) << *self << " ResumeAll waking others";
392    } else {
393      VLOG(threads) << "Thread[null] ResumeAll waking others";
394    }
395    Thread::resume_cond_->Broadcast(self);
396  }
397  ATRACE_END();
398
399  if (self != nullptr) {
400    VLOG(threads) << *self << " ResumeAll complete";
401  } else {
402    VLOG(threads) << "Thread[null] ResumeAll complete";
403  }
404}
405
406void ThreadList::Resume(Thread* thread, bool for_debugger) {
407  Thread* self = Thread::Current();
408  DCHECK_NE(thread, self);
409  VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") starting..."
410      << (for_debugger ? " (debugger)" : "");
411
412  {
413    // To check Contains.
414    MutexLock mu(self, *Locks::thread_list_lock_);
415    // To check IsSuspended.
416    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
417    DCHECK(thread->IsSuspended());
418    if (!Contains(thread)) {
419      // We only expect threads within the thread-list to have been suspended otherwise we can't
420      // stop such threads from delete-ing themselves.
421      LOG(ERROR) << "Resume(" << reinterpret_cast<void*>(thread)
422          << ") thread not within thread list";
423      return;
424    }
425    thread->ModifySuspendCount(self, -1, for_debugger);
426  }
427
428  {
429    VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") waking others";
430    MutexLock mu(self, *Locks::thread_suspend_count_lock_);
431    Thread::resume_cond_->Broadcast(self);
432  }
433
434  VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") complete";
435}
436
437static void ThreadSuspendByPeerWarning(Thread* self, int level, const char* message, jobject peer) {
438  JNIEnvExt* env = self->GetJniEnv();
439  ScopedLocalRef<jstring>
440      scoped_name_string(env, (jstring)env->GetObjectField(peer,
441                                                          WellKnownClasses::java_lang_Thread_name));
442  ScopedUtfChars scoped_name_chars(env, scoped_name_string.get());
443  if (scoped_name_chars.c_str() == NULL) {
444      LOG(level) << message << ": " << peer;
445      env->ExceptionClear();
446  } else {
447      LOG(level) << message << ": " << peer << ":" << scoped_name_chars.c_str();
448  }
449}
450
451Thread* ThreadList::SuspendThreadByPeer(jobject peer, bool request_suspension,
452                                        bool debug_suspension, bool* timed_out) {
453  static const useconds_t kTimeoutUs = 30 * 1000000;  // 30s.
454  useconds_t total_delay_us = 0;
455  useconds_t delay_us = 0;
456  bool did_suspend_request = false;
457  *timed_out = false;
458  Thread* self = Thread::Current();
459  VLOG(threads) << "SuspendThreadByPeer starting";
460  while (true) {
461    Thread* thread;
462    {
463      // Note: this will transition to runnable and potentially suspend. We ensure only one thread
464      // is requesting another suspend, to avoid deadlock, by requiring this function be called
465      // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
466      // than request thread suspension, to avoid potential cycles in threads requesting each other
467      // suspend.
468      ScopedObjectAccess soa(self);
469      MutexLock mu(self, *Locks::thread_list_lock_);
470      thread = Thread::FromManagedThread(soa, peer);
471      if (thread == nullptr) {
472        ThreadSuspendByPeerWarning(self, WARNING, "No such thread for suspend", peer);
473        return nullptr;
474      }
475      if (!Contains(thread)) {
476        VLOG(threads) << "SuspendThreadByPeer failed for unattached thread: "
477            << reinterpret_cast<void*>(thread);
478        return nullptr;
479      }
480      VLOG(threads) << "SuspendThreadByPeer found thread: " << *thread;
481      {
482        MutexLock mu(self, *Locks::thread_suspend_count_lock_);
483        if (request_suspension) {
484          thread->ModifySuspendCount(self, +1, debug_suspension);
485          request_suspension = false;
486          did_suspend_request = true;
487        } else {
488          // If the caller isn't requesting suspension, a suspension should have already occurred.
489          CHECK_GT(thread->GetSuspendCount(), 0);
490        }
491        // IsSuspended on the current thread will fail as the current thread is changed into
492        // Runnable above. As the suspend count is now raised if this is the current thread
493        // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
494        // to just explicitly handle the current thread in the callers to this code.
495        CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
496        // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
497        // count, or else we've waited and it has self suspended) or is the current thread, we're
498        // done.
499        if (thread->IsSuspended()) {
500          VLOG(threads) << "SuspendThreadByPeer thread suspended: " << *thread;
501          return thread;
502        }
503        if (total_delay_us >= kTimeoutUs) {
504          ThreadSuspendByPeerWarning(self, FATAL, "Thread suspension timed out", peer);
505          if (did_suspend_request) {
506            thread->ModifySuspendCount(soa.Self(), -1, debug_suspension);
507          }
508          *timed_out = true;
509          return nullptr;
510        }
511      }
512      // Release locks and come out of runnable state.
513    }
514    VLOG(threads) << "SuspendThreadByPeer sleeping to allow thread chance to suspend";
515    ThreadSuspendSleep(self, &delay_us, &total_delay_us);
516  }
517}
518
519static void ThreadSuspendByThreadIdWarning(int level, const char* message, uint32_t thread_id) {
520  LOG(level) << StringPrintf("%s: %d", message, thread_id);
521}
522
523Thread* ThreadList::SuspendThreadByThreadId(uint32_t thread_id, bool debug_suspension,
524                                            bool* timed_out) {
525  static const useconds_t kTimeoutUs = 30 * 1000000;  // 30s.
526  useconds_t total_delay_us = 0;
527  useconds_t delay_us = 0;
528  *timed_out = false;
529  Thread* suspended_thread = nullptr;
530  Thread* self = Thread::Current();
531  CHECK_NE(thread_id, kInvalidThreadId);
532  VLOG(threads) << "SuspendThreadByThreadId starting";
533  while (true) {
534    {
535      // Note: this will transition to runnable and potentially suspend. We ensure only one thread
536      // is requesting another suspend, to avoid deadlock, by requiring this function be called
537      // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
538      // than request thread suspension, to avoid potential cycles in threads requesting each other
539      // suspend.
540      ScopedObjectAccess soa(self);
541      MutexLock mu(self, *Locks::thread_list_lock_);
542      Thread* thread = nullptr;
543      for (const auto& it : list_) {
544        if (it->GetThreadId() == thread_id) {
545          thread = it;
546          break;
547        }
548      }
549      if (thread == nullptr) {
550        CHECK(suspended_thread == nullptr) << "Suspended thread " << suspended_thread
551            << " no longer in thread list";
552        // There's a race in inflating a lock and the owner giving up ownership and then dying.
553        ThreadSuspendByThreadIdWarning(WARNING, "No such thread id for suspend", thread_id);
554        return nullptr;
555      }
556      VLOG(threads) << "SuspendThreadByThreadId found thread: " << *thread;
557      DCHECK(Contains(thread));
558      {
559        MutexLock mu(self, *Locks::thread_suspend_count_lock_);
560        if (suspended_thread == nullptr) {
561          thread->ModifySuspendCount(self, +1, debug_suspension);
562          suspended_thread = thread;
563        } else {
564          CHECK_EQ(suspended_thread, thread);
565          // If the caller isn't requesting suspension, a suspension should have already occurred.
566          CHECK_GT(thread->GetSuspendCount(), 0);
567        }
568        // IsSuspended on the current thread will fail as the current thread is changed into
569        // Runnable above. As the suspend count is now raised if this is the current thread
570        // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
571        // to just explicitly handle the current thread in the callers to this code.
572        CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
573        // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
574        // count, or else we've waited and it has self suspended) or is the current thread, we're
575        // done.
576        if (thread->IsSuspended()) {
577          VLOG(threads) << "SuspendThreadByThreadId thread suspended: " << *thread;
578          return thread;
579        }
580        if (total_delay_us >= kTimeoutUs) {
581          ThreadSuspendByThreadIdWarning(WARNING, "Thread suspension timed out", thread_id);
582          if (suspended_thread != nullptr) {
583            thread->ModifySuspendCount(soa.Self(), -1, debug_suspension);
584          }
585          *timed_out = true;
586          return nullptr;
587        }
588      }
589      // Release locks and come out of runnable state.
590    }
591    VLOG(threads) << "SuspendThreadByThreadId sleeping to allow thread chance to suspend";
592    ThreadSuspendSleep(self, &delay_us, &total_delay_us);
593  }
594}
595
596Thread* ThreadList::FindThreadByThreadId(uint32_t thin_lock_id) {
597  Thread* self = Thread::Current();
598  MutexLock mu(self, *Locks::thread_list_lock_);
599  for (const auto& thread : list_) {
600    if (thread->GetThreadId() == thin_lock_id) {
601      CHECK(thread == self || thread->IsSuspended());
602      return thread;
603    }
604  }
605  return NULL;
606}
607
608void ThreadList::SuspendAllForDebugger() {
609  Thread* self = Thread::Current();
610  Thread* debug_thread = Dbg::GetDebugThread();
611
612  VLOG(threads) << *self << " SuspendAllForDebugger starting...";
613
614  {
615    MutexLock mu(self, *Locks::thread_list_lock_);
616    {
617      MutexLock mu(self, *Locks::thread_suspend_count_lock_);
618      // Update global suspend all state for attaching threads.
619      ++suspend_all_count_;
620      ++debug_suspend_all_count_;
621      // Increment everybody's suspend count (except our own).
622      for (const auto& thread : list_) {
623        if (thread == self || thread == debug_thread) {
624          continue;
625        }
626        VLOG(threads) << "requesting thread suspend: " << *thread;
627        thread->ModifySuspendCount(self, +1, true);
628      }
629    }
630  }
631
632  // Block on the mutator lock until all Runnable threads release their share of access then
633  // immediately unlock again.
634#if HAVE_TIMED_RWLOCK
635  // Timeout if we wait more than 30 seconds.
636  if (!Locks::mutator_lock_->ExclusiveLockWithTimeout(self, 30 * 1000, 0)) {
637    UnsafeLogFatalForThreadSuspendAllTimeout();
638  } else {
639    Locks::mutator_lock_->ExclusiveUnlock(self);
640  }
641#else
642  Locks::mutator_lock_->ExclusiveLock(self);
643  Locks::mutator_lock_->ExclusiveUnlock(self);
644#endif
645  AssertThreadsAreSuspended(self, self, debug_thread);
646
647  VLOG(threads) << *self << " SuspendAllForDebugger complete";
648}
649
650void ThreadList::SuspendSelfForDebugger() {
651  Thread* self = Thread::Current();
652
653  // The debugger thread must not suspend itself due to debugger activity!
654  Thread* debug_thread = Dbg::GetDebugThread();
655  CHECK(debug_thread != NULL);
656  CHECK(self != debug_thread);
657  CHECK_NE(self->GetState(), kRunnable);
658  Locks::mutator_lock_->AssertNotHeld(self);
659
660  {
661    // Collisions with other suspends aren't really interesting. We want
662    // to ensure that we're the only one fiddling with the suspend count
663    // though.
664    MutexLock mu(self, *Locks::thread_suspend_count_lock_);
665    self->ModifySuspendCount(self, +1, true);
666    CHECK_GT(self->GetSuspendCount(), 0);
667  }
668
669  VLOG(threads) << *self << " self-suspending (debugger)";
670
671  // Tell JDWP we've completed invocation and are ready to suspend.
672  DebugInvokeReq* pReq = self->GetInvokeReq();
673  DCHECK(pReq != NULL);
674  if (pReq->invoke_needed) {
675    // Clear this before signaling.
676    pReq->Clear();
677
678    VLOG(jdwp) << "invoke complete, signaling";
679    MutexLock mu(self, pReq->lock);
680    pReq->cond.Signal(self);
681  }
682
683  // Tell JDWP that we've completed suspension. The JDWP thread can't
684  // tell us to resume before we're fully asleep because we hold the
685  // suspend count lock.
686  Dbg::ClearWaitForEventThread();
687
688  {
689    MutexLock mu(self, *Locks::thread_suspend_count_lock_);
690    while (self->GetSuspendCount() != 0) {
691      Thread::resume_cond_->Wait(self);
692      if (self->GetSuspendCount() != 0) {
693        // The condition was signaled but we're still suspended. This
694        // can happen when we suspend then resume all threads to
695        // update instrumentation or compute monitor info. This can
696        // also happen if the debugger lets go while a SIGQUIT thread
697        // dump event is pending (assuming SignalCatcher was resumed for
698        // just long enough to try to grab the thread-suspend lock).
699        VLOG(jdwp) << *self << " still suspended after undo "
700                   << "(suspend count=" << self->GetSuspendCount() << ", "
701                   << "debug suspend count=" << self->GetDebugSuspendCount() << ")";
702      }
703    }
704    CHECK_EQ(self->GetSuspendCount(), 0);
705  }
706
707  VLOG(threads) << *self << " self-reviving (debugger)";
708}
709
710void ThreadList::UndoDebuggerSuspensions() {
711  Thread* self = Thread::Current();
712
713  VLOG(threads) << *self << " UndoDebuggerSuspensions starting";
714
715  {
716    MutexLock mu(self, *Locks::thread_list_lock_);
717    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
718    // Update global suspend all state for attaching threads.
719    suspend_all_count_ -= debug_suspend_all_count_;
720    debug_suspend_all_count_ = 0;
721    // Update running threads.
722    for (const auto& thread : list_) {
723      if (thread == self || thread->GetDebugSuspendCount() == 0) {
724        continue;
725      }
726      thread->ModifySuspendCount(self, -thread->GetDebugSuspendCount(), true);
727    }
728  }
729
730  {
731    MutexLock mu(self, *Locks::thread_suspend_count_lock_);
732    Thread::resume_cond_->Broadcast(self);
733  }
734
735  VLOG(threads) << "UndoDebuggerSuspensions(" << *self << ") complete";
736}
737
738void ThreadList::WaitForOtherNonDaemonThreadsToExit() {
739  Thread* self = Thread::Current();
740  Locks::mutator_lock_->AssertNotHeld(self);
741  bool all_threads_are_daemons;
742  do {
743    {
744      // No more threads can be born after we start to shutdown.
745      MutexLock mu(self, *Locks::runtime_shutdown_lock_);
746      CHECK(Runtime::Current()->IsShuttingDownLocked());
747      CHECK_EQ(Runtime::Current()->NumberOfThreadsBeingBorn(), 0U);
748    }
749    all_threads_are_daemons = true;
750    MutexLock mu(self, *Locks::thread_list_lock_);
751    for (const auto& thread : list_) {
752      if (thread != self && !thread->IsDaemon()) {
753        all_threads_are_daemons = false;
754        break;
755      }
756    }
757    if (!all_threads_are_daemons) {
758      // Wait for another thread to exit before re-checking.
759      thread_exit_cond_.Wait(self);
760    }
761  } while (!all_threads_are_daemons);
762}
763
764void ThreadList::SuspendAllDaemonThreads() {
765  Thread* self = Thread::Current();
766  MutexLock mu(self, *Locks::thread_list_lock_);
767  {  // Tell all the daemons it's time to suspend.
768    MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
769    for (const auto& thread : list_) {
770      // This is only run after all non-daemon threads have exited, so the remainder should all be
771      // daemons.
772      CHECK(thread->IsDaemon()) << *thread;
773      if (thread != self) {
774        thread->ModifySuspendCount(self, +1, false);
775      }
776    }
777  }
778  // Give the threads a chance to suspend, complaining if they're slow.
779  bool have_complained = false;
780  for (int i = 0; i < 10; ++i) {
781    usleep(200 * 1000);
782    bool all_suspended = true;
783    for (const auto& thread : list_) {
784      if (thread != self && thread->GetState() == kRunnable) {
785        if (!have_complained) {
786          LOG(WARNING) << "daemon thread not yet suspended: " << *thread;
787          have_complained = true;
788        }
789        all_suspended = false;
790      }
791    }
792    if (all_suspended) {
793      return;
794    }
795  }
796  LOG(ERROR) << "suspend all daemons failed";
797}
798void ThreadList::Register(Thread* self) {
799  DCHECK_EQ(self, Thread::Current());
800
801  if (VLOG_IS_ON(threads)) {
802    std::ostringstream oss;
803    self->ShortDump(oss);  // We don't hold the mutator_lock_ yet and so cannot call Dump.
804    LOG(INFO) << "ThreadList::Register() " << *self  << "\n" << oss.str();
805  }
806
807  // Atomically add self to the thread list and make its thread_suspend_count_ reflect ongoing
808  // SuspendAll requests.
809  MutexLock mu(self, *Locks::thread_list_lock_);
810  MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
811  CHECK_GE(suspend_all_count_, debug_suspend_all_count_);
812  // Modify suspend count in increments of 1 to maintain invariants in ModifySuspendCount. While
813  // this isn't particularly efficient the suspend counts are most commonly 0 or 1.
814  for (int delta = debug_suspend_all_count_; delta > 0; delta--) {
815    self->ModifySuspendCount(self, +1, true);
816  }
817  for (int delta = suspend_all_count_ - debug_suspend_all_count_; delta > 0; delta--) {
818    self->ModifySuspendCount(self, +1, false);
819  }
820  CHECK(!Contains(self));
821  list_.push_back(self);
822}
823
824void ThreadList::Unregister(Thread* self) {
825  DCHECK_EQ(self, Thread::Current());
826
827  VLOG(threads) << "ThreadList::Unregister() " << *self;
828
829  // Any time-consuming destruction, plus anything that can call back into managed code or
830  // suspend and so on, must happen at this point, and not in ~Thread.
831  self->Destroy();
832
833  uint32_t thin_lock_id = self->GetThreadId();
834  while (self != nullptr) {
835    // Remove and delete the Thread* while holding the thread_list_lock_ and
836    // thread_suspend_count_lock_ so that the unregistering thread cannot be suspended.
837    // Note: deliberately not using MutexLock that could hold a stale self pointer.
838    Locks::thread_list_lock_->ExclusiveLock(self);
839    if (!Contains(self)) {
840      std::ostringstream os;
841      DumpNativeStack(os, GetTid(), "  native: ", nullptr);
842      LOG(ERROR) << "Request to unregister unattached thread\n" << os.str();
843      self = nullptr;
844    } else {
845      // Note: we don't take the thread_suspend_count_lock_ here as to be suspending a thread other
846      // than yourself you need to hold the thread_list_lock_ (see Thread::ModifySuspendCount).
847      if (!self->IsSuspended()) {
848        list_.remove(self);
849        delete self;
850        self = nullptr;
851      }
852    }
853    Locks::thread_list_lock_->ExclusiveUnlock(self);
854  }
855  // Release the thread ID after the thread is finished and deleted to avoid cases where we can
856  // temporarily have multiple threads with the same thread id. When this occurs, it causes
857  // problems in FindThreadByThreadId / SuspendThreadByThreadId.
858  ReleaseThreadId(nullptr, thin_lock_id);
859
860  // Clear the TLS data, so that the underlying native thread is recognizably detached.
861  // (It may wish to reattach later.)
862  CHECK_PTHREAD_CALL(pthread_setspecific, (Thread::pthread_key_self_, NULL), "detach self");
863
864  // Signal that a thread just detached.
865  MutexLock mu(NULL, *Locks::thread_list_lock_);
866  thread_exit_cond_.Signal(NULL);
867}
868
869void ThreadList::ForEach(void (*callback)(Thread*, void*), void* context) {
870  for (const auto& thread : list_) {
871    callback(thread, context);
872  }
873}
874
875void ThreadList::VisitRoots(RootCallback* callback, void* arg) const {
876  MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
877  for (const auto& thread : list_) {
878    thread->VisitRoots(callback, arg);
879  }
880}
881
882class VerifyRootWrapperArg {
883 public:
884  VerifyRootWrapperArg(VerifyRootCallback* callback, void* arg) : callback_(callback), arg_(arg) {
885  }
886  VerifyRootCallback* const callback_;
887  void* const arg_;
888};
889
890static void VerifyRootWrapperCallback(mirror::Object** root, void* arg, uint32_t /*thread_id*/,
891                                      RootType root_type) {
892  VerifyRootWrapperArg* wrapperArg = reinterpret_cast<VerifyRootWrapperArg*>(arg);
893  wrapperArg->callback_(*root, wrapperArg->arg_, 0, NULL, root_type);
894}
895
896void ThreadList::VerifyRoots(VerifyRootCallback* callback, void* arg) const {
897  VerifyRootWrapperArg wrapper(callback, arg);
898  MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
899  for (const auto& thread : list_) {
900    thread->VisitRoots(VerifyRootWrapperCallback, &wrapper);
901  }
902}
903
904uint32_t ThreadList::AllocThreadId(Thread* self) {
905  MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
906  for (size_t i = 0; i < allocated_ids_.size(); ++i) {
907    if (!allocated_ids_[i]) {
908      allocated_ids_.set(i);
909      return i + 1;  // Zero is reserved to mean "invalid".
910    }
911  }
912  LOG(FATAL) << "Out of internal thread ids";
913  return 0;
914}
915
916void ThreadList::ReleaseThreadId(Thread* self, uint32_t id) {
917  MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
918  --id;  // Zero is reserved to mean "invalid".
919  DCHECK(allocated_ids_[id]) << id;
920  allocated_ids_.reset(id);
921}
922
923}  // namespace art
924