• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <backtrace/BacktraceMap.h>
20 #include <dirent.h>
21 #include <ScopedLocalRef.h>
22 #include <ScopedUtfChars.h>
23 #include <sys/types.h>
24 #include <unistd.h>
25 
26 #include <sstream>
27 
28 #include "base/histogram-inl.h"
29 #include "base/mutex-inl.h"
30 #include "base/systrace.h"
31 #include "base/time_utils.h"
32 #include "base/timing_logger.h"
33 #include "debugger.h"
34 #include "gc/collector/concurrent_copying.h"
35 #include "jni_internal.h"
36 #include "lock_word.h"
37 #include "monitor.h"
38 #include "scoped_thread_state_change.h"
39 #include "thread.h"
40 #include "trace.h"
41 #include "well_known_classes.h"
42 
43 #if ART_USE_FUTEXES
44 #include "linux/futex.h"
45 #include "sys/syscall.h"
46 #ifndef SYS_futex
47 #define SYS_futex __NR_futex
48 #endif
49 #endif  // ART_USE_FUTEXES
50 
51 namespace art {
52 
53 static constexpr uint64_t kLongThreadSuspendThreshold = MsToNs(5);
54 static constexpr uint64_t kThreadSuspendTimeoutMs = 30 * 1000;  // 30s.
55 // Use 0 since we want to yield to prevent blocking for an unpredictable amount of time.
56 static constexpr useconds_t kThreadSuspendInitialSleepUs = 0;
57 static constexpr useconds_t kThreadSuspendMaxYieldUs = 3000;
58 static constexpr useconds_t kThreadSuspendMaxSleepUs = 5000;
59 
60 // Whether we should try to dump the native stack of unattached threads. See commit ed8b723 for
61 // some history.
62 // Turned off again. b/29248079
63 static constexpr bool kDumpUnattachedThreadNativeStack = false;
64 
ThreadList()65 ThreadList::ThreadList()
66     : suspend_all_count_(0),
67       debug_suspend_all_count_(0),
68       unregistering_count_(0),
69       suspend_all_historam_("suspend all histogram", 16, 64),
70       long_suspend_(false) {
71   CHECK(Monitor::IsValidLockWord(LockWord::FromThinLockId(kMaxThreadId, 1, 0U)));
72 }
73 
~ThreadList()74 ThreadList::~ThreadList() {
75   ScopedTrace trace(__PRETTY_FUNCTION__);
76   // Detach the current thread if necessary. If we failed to start, there might not be any threads.
77   // We need to detach the current thread here in case there's another thread waiting to join with
78   // us.
79   bool contains = false;
80   Thread* self = Thread::Current();
81   {
82     MutexLock mu(self, *Locks::thread_list_lock_);
83     contains = Contains(self);
84   }
85   if (contains) {
86     Runtime::Current()->DetachCurrentThread();
87   }
88   WaitForOtherNonDaemonThreadsToExit();
89   // Disable GC and wait for GC to complete in case there are still daemon threads doing
90   // allocations.
91   gc::Heap* const heap = Runtime::Current()->GetHeap();
92   heap->DisableGCForShutdown();
93   // In case a GC is in progress, wait for it to finish.
94   heap->WaitForGcToComplete(gc::kGcCauseBackground, Thread::Current());
95   // TODO: there's an unaddressed race here where a thread may attach during shutdown, see
96   //       Thread::Init.
97   SuspendAllDaemonThreadsForShutdown();
98 }
99 
Contains(Thread * thread)100 bool ThreadList::Contains(Thread* thread) {
101   return find(list_.begin(), list_.end(), thread) != list_.end();
102 }
103 
Contains(pid_t tid)104 bool ThreadList::Contains(pid_t tid) {
105   for (const auto& thread : list_) {
106     if (thread->GetTid() == tid) {
107       return true;
108     }
109   }
110   return false;
111 }
112 
GetLockOwner()113 pid_t ThreadList::GetLockOwner() {
114   return Locks::thread_list_lock_->GetExclusiveOwnerTid();
115 }
116 
DumpNativeStacks(std::ostream & os)117 void ThreadList::DumpNativeStacks(std::ostream& os) {
118   MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
119   std::unique_ptr<BacktraceMap> map(BacktraceMap::Create(getpid()));
120   for (const auto& thread : list_) {
121     os << "DUMPING THREAD " << thread->GetTid() << "\n";
122     DumpNativeStack(os, thread->GetTid(), map.get(), "\t");
123     os << "\n";
124   }
125 }
126 
DumpForSigQuit(std::ostream & os)127 void ThreadList::DumpForSigQuit(std::ostream& os) {
128   {
129     ScopedObjectAccess soa(Thread::Current());
130     // Only print if we have samples.
131     if (suspend_all_historam_.SampleSize() > 0) {
132       Histogram<uint64_t>::CumulativeData data;
133       suspend_all_historam_.CreateHistogram(&data);
134       suspend_all_historam_.PrintConfidenceIntervals(os, 0.99, data);  // Dump time to suspend.
135     }
136   }
137   bool dump_native_stack = Runtime::Current()->GetDumpNativeStackOnSigQuit();
138   Dump(os, dump_native_stack);
139   DumpUnattachedThreads(os, dump_native_stack);
140 }
141 
DumpUnattachedThread(std::ostream & os,pid_t tid,bool dump_native_stack)142 static void DumpUnattachedThread(std::ostream& os, pid_t tid, bool dump_native_stack)
143     NO_THREAD_SAFETY_ANALYSIS {
144   // TODO: No thread safety analysis as DumpState with a null thread won't access fields, should
145   // refactor DumpState to avoid skipping analysis.
146   Thread::DumpState(os, nullptr, tid);
147   DumpKernelStack(os, tid, "  kernel: ", false);
148   if (dump_native_stack && kDumpUnattachedThreadNativeStack) {
149     DumpNativeStack(os, tid, nullptr, "  native: ");
150   }
151   os << "\n";
152 }
153 
DumpUnattachedThreads(std::ostream & os,bool dump_native_stack)154 void ThreadList::DumpUnattachedThreads(std::ostream& os, bool dump_native_stack) {
155   DIR* d = opendir("/proc/self/task");
156   if (!d) {
157     return;
158   }
159 
160   Thread* self = Thread::Current();
161   dirent* e;
162   while ((e = readdir(d)) != nullptr) {
163     char* end;
164     pid_t tid = strtol(e->d_name, &end, 10);
165     if (!*end) {
166       bool contains;
167       {
168         MutexLock mu(self, *Locks::thread_list_lock_);
169         contains = Contains(tid);
170       }
171       if (!contains) {
172         DumpUnattachedThread(os, tid, dump_native_stack);
173       }
174     }
175   }
176   closedir(d);
177 }
178 
179 // Dump checkpoint timeout in milliseconds. Larger amount on the target, since the device could be
180 // overloaded with ANR dumps.
181 static constexpr uint32_t kDumpWaitTimeout = kIsTargetBuild ? 100000 : 20000;
182 
183 // A closure used by Thread::Dump.
184 class DumpCheckpoint FINAL : public Closure {
185  public:
DumpCheckpoint(std::ostream * os,bool dump_native_stack)186   DumpCheckpoint(std::ostream* os, bool dump_native_stack)
187       : os_(os),
188         barrier_(0),
189         backtrace_map_(dump_native_stack ? BacktraceMap::Create(getpid()) : nullptr),
190         dump_native_stack_(dump_native_stack) {}
191 
Run(Thread * thread)192   void Run(Thread* thread) OVERRIDE {
193     // Note thread and self may not be equal if thread was already suspended at the point of the
194     // request.
195     Thread* self = Thread::Current();
196     std::ostringstream local_os;
197     {
198       ScopedObjectAccess soa(self);
199       thread->Dump(local_os, dump_native_stack_, backtrace_map_.get());
200     }
201     local_os << "\n";
202     {
203       // Use the logging lock to ensure serialization when writing to the common ostream.
204       MutexLock mu(self, *Locks::logging_lock_);
205       *os_ << local_os.str();
206     }
207     barrier_.Pass(self);
208   }
209 
WaitForThreadsToRunThroughCheckpoint(size_t threads_running_checkpoint)210   void WaitForThreadsToRunThroughCheckpoint(size_t threads_running_checkpoint) {
211     Thread* self = Thread::Current();
212     ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
213     bool timed_out = barrier_.Increment(self, threads_running_checkpoint, kDumpWaitTimeout);
214     if (timed_out) {
215       // Avoid a recursive abort.
216       LOG((kIsDebugBuild && (gAborting == 0)) ? FATAL : ERROR)
217           << "Unexpected time out during dump checkpoint.";
218     }
219   }
220 
221  private:
222   // The common stream that will accumulate all the dumps.
223   std::ostream* const os_;
224   // The barrier to be passed through and for the requestor to wait upon.
225   Barrier barrier_;
226   // A backtrace map, so that all threads use a shared info and don't reacquire/parse separately.
227   std::unique_ptr<BacktraceMap> backtrace_map_;
228   // Whether we should dump the native stack.
229   const bool dump_native_stack_;
230 };
231 
Dump(std::ostream & os,bool dump_native_stack)232 void ThreadList::Dump(std::ostream& os, bool dump_native_stack) {
233   {
234     MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
235     os << "DALVIK THREADS (" << list_.size() << "):\n";
236   }
237   DumpCheckpoint checkpoint(&os, dump_native_stack);
238   size_t threads_running_checkpoint;
239   {
240     // Use SOA to prevent deadlocks if multiple threads are calling Dump() at the same time.
241     ScopedObjectAccess soa(Thread::Current());
242     threads_running_checkpoint = RunCheckpoint(&checkpoint);
243   }
244   if (threads_running_checkpoint != 0) {
245     checkpoint.WaitForThreadsToRunThroughCheckpoint(threads_running_checkpoint);
246   }
247 }
248 
AssertThreadsAreSuspended(Thread * self,Thread * ignore1,Thread * ignore2)249 void ThreadList::AssertThreadsAreSuspended(Thread* self, Thread* ignore1, Thread* ignore2) {
250   MutexLock mu(self, *Locks::thread_list_lock_);
251   MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
252   for (const auto& thread : list_) {
253     if (thread != ignore1 && thread != ignore2) {
254       CHECK(thread->IsSuspended())
255             << "\nUnsuspended thread: <<" << *thread << "\n"
256             << "self: <<" << *Thread::Current();
257     }
258   }
259 }
260 
261 #if HAVE_TIMED_RWLOCK
262 // Attempt to rectify locks so that we dump thread list with required locks before exiting.
UnsafeLogFatalForThreadSuspendAllTimeout()263 NO_RETURN static void UnsafeLogFatalForThreadSuspendAllTimeout() {
264   Runtime* runtime = Runtime::Current();
265   std::ostringstream ss;
266   ss << "Thread suspend timeout\n";
267   Locks::mutator_lock_->Dump(ss);
268   ss << "\n";
269   runtime->GetThreadList()->Dump(ss);
270   LOG(FATAL) << ss.str();
271   exit(0);
272 }
273 #endif
274 
275 // Unlike suspending all threads where we can wait to acquire the mutator_lock_, suspending an
276 // individual thread requires polling. delay_us is the requested sleep wait. If delay_us is 0 then
277 // we use sched_yield instead of calling usleep.
ThreadSuspendSleep(useconds_t delay_us)278 static void ThreadSuspendSleep(useconds_t delay_us) {
279   if (delay_us == 0) {
280     sched_yield();
281   } else {
282     usleep(delay_us);
283   }
284 }
285 
RunCheckpoint(Closure * checkpoint_function)286 size_t ThreadList::RunCheckpoint(Closure* checkpoint_function) {
287   Thread* self = Thread::Current();
288   Locks::mutator_lock_->AssertNotExclusiveHeld(self);
289   Locks::thread_list_lock_->AssertNotHeld(self);
290   Locks::thread_suspend_count_lock_->AssertNotHeld(self);
291 
292   std::vector<Thread*> suspended_count_modified_threads;
293   size_t count = 0;
294   {
295     // Call a checkpoint function for each thread, threads which are suspend get their checkpoint
296     // manually called.
297     MutexLock mu(self, *Locks::thread_list_lock_);
298     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
299     count = list_.size();
300     for (const auto& thread : list_) {
301       if (thread != self) {
302         while (true) {
303           if (thread->RequestCheckpoint(checkpoint_function)) {
304             // This thread will run its checkpoint some time in the near future.
305             break;
306           } else {
307             // We are probably suspended, try to make sure that we stay suspended.
308             // The thread switched back to runnable.
309             if (thread->GetState() == kRunnable) {
310               // Spurious fail, try again.
311               continue;
312             }
313             thread->ModifySuspendCount(self, +1, nullptr, false);
314             suspended_count_modified_threads.push_back(thread);
315             break;
316           }
317         }
318       }
319     }
320   }
321 
322   // Run the checkpoint on ourself while we wait for threads to suspend.
323   checkpoint_function->Run(self);
324 
325   // Run the checkpoint on the suspended threads.
326   for (const auto& thread : suspended_count_modified_threads) {
327     if (!thread->IsSuspended()) {
328       if (ATRACE_ENABLED()) {
329         std::ostringstream oss;
330         thread->ShortDump(oss);
331         ATRACE_BEGIN((std::string("Waiting for suspension of thread ") + oss.str()).c_str());
332       }
333       // Busy wait until the thread is suspended.
334       const uint64_t start_time = NanoTime();
335       do {
336         ThreadSuspendSleep(kThreadSuspendInitialSleepUs);
337       } while (!thread->IsSuspended());
338       const uint64_t total_delay = NanoTime() - start_time;
339       // Shouldn't need to wait for longer than 1000 microseconds.
340       constexpr uint64_t kLongWaitThreshold = MsToNs(1);
341       ATRACE_END();
342       if (UNLIKELY(total_delay > kLongWaitThreshold)) {
343         LOG(WARNING) << "Long wait of " << PrettyDuration(total_delay) << " for "
344             << *thread << " suspension!";
345       }
346     }
347     // We know for sure that the thread is suspended at this point.
348     checkpoint_function->Run(thread);
349     {
350       MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
351       thread->ModifySuspendCount(self, -1, nullptr, false);
352     }
353   }
354 
355   {
356     // Imitate ResumeAll, threads may be waiting on Thread::resume_cond_ since we raised their
357     // suspend count. Now the suspend_count_ is lowered so we must do the broadcast.
358     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
359     Thread::resume_cond_->Broadcast(self);
360   }
361 
362   return count;
363 }
364 
365 // Request that a checkpoint function be run on all active (non-suspended)
366 // threads.  Returns the number of successful requests.
RunCheckpointOnRunnableThreads(Closure * checkpoint_function)367 size_t ThreadList::RunCheckpointOnRunnableThreads(Closure* checkpoint_function) {
368   Thread* self = Thread::Current();
369   Locks::mutator_lock_->AssertNotExclusiveHeld(self);
370   Locks::thread_list_lock_->AssertNotHeld(self);
371   Locks::thread_suspend_count_lock_->AssertNotHeld(self);
372   CHECK_NE(self->GetState(), kRunnable);
373 
374   size_t count = 0;
375   {
376     // Call a checkpoint function for each non-suspended thread.
377     MutexLock mu(self, *Locks::thread_list_lock_);
378     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
379     for (const auto& thread : list_) {
380       if (thread != self) {
381         if (thread->RequestCheckpoint(checkpoint_function)) {
382           // This thread will run its checkpoint some time in the near future.
383           count++;
384         }
385       }
386     }
387   }
388 
389   // Return the number of threads that will run the checkpoint function.
390   return count;
391 }
392 
393 // A checkpoint/suspend-all hybrid to switch thread roots from
394 // from-space to to-space refs. Used to synchronize threads at a point
395 // to mark the initiation of marking while maintaining the to-space
396 // invariant.
FlipThreadRoots(Closure * thread_flip_visitor,Closure * flip_callback,gc::collector::GarbageCollector * collector)397 size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
398                                    Closure* flip_callback,
399                                    gc::collector::GarbageCollector* collector) {
400   TimingLogger::ScopedTiming split("ThreadListFlip", collector->GetTimings());
401   const uint64_t start_time = NanoTime();
402   Thread* self = Thread::Current();
403   Locks::mutator_lock_->AssertNotHeld(self);
404   Locks::thread_list_lock_->AssertNotHeld(self);
405   Locks::thread_suspend_count_lock_->AssertNotHeld(self);
406   CHECK_NE(self->GetState(), kRunnable);
407 
408   SuspendAllInternal(self, self, nullptr);
409 
410   // Run the flip callback for the collector.
411   Locks::mutator_lock_->ExclusiveLock(self);
412   flip_callback->Run(self);
413   Locks::mutator_lock_->ExclusiveUnlock(self);
414   collector->RegisterPause(NanoTime() - start_time);
415 
416   // Resume runnable threads.
417   std::vector<Thread*> runnable_threads;
418   std::vector<Thread*> other_threads;
419   {
420     MutexLock mu(self, *Locks::thread_list_lock_);
421     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
422     --suspend_all_count_;
423     for (const auto& thread : list_) {
424       if (thread == self) {
425         continue;
426       }
427       // Set the flip function for both runnable and suspended threads
428       // because Thread::DumpState/DumpJavaStack() (invoked by a
429       // checkpoint) may cause the flip function to be run for a
430       // runnable/suspended thread before a runnable threads runs it
431       // for itself or we run it for a suspended thread below.
432       thread->SetFlipFunction(thread_flip_visitor);
433       if (thread->IsSuspendedAtSuspendCheck()) {
434         // The thread will resume right after the broadcast.
435         thread->ModifySuspendCount(self, -1, nullptr, false);
436         runnable_threads.push_back(thread);
437       } else {
438         other_threads.push_back(thread);
439       }
440     }
441     Thread::resume_cond_->Broadcast(self);
442   }
443 
444   // Run the closure on the other threads and let them resume.
445   {
446     ReaderMutexLock mu(self, *Locks::mutator_lock_);
447     for (const auto& thread : other_threads) {
448       Closure* flip_func = thread->GetFlipFunction();
449       if (flip_func != nullptr) {
450         flip_func->Run(thread);
451       }
452     }
453     // Run it for self.
454     thread_flip_visitor->Run(self);
455   }
456 
457   // Resume other threads.
458   {
459     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
460     for (const auto& thread : other_threads) {
461       thread->ModifySuspendCount(self, -1, nullptr, false);
462     }
463     Thread::resume_cond_->Broadcast(self);
464   }
465 
466   return runnable_threads.size() + other_threads.size() + 1;  // +1 for self.
467 }
468 
SuspendAll(const char * cause,bool long_suspend)469 void ThreadList::SuspendAll(const char* cause, bool long_suspend) {
470   Thread* self = Thread::Current();
471 
472   if (self != nullptr) {
473     VLOG(threads) << *self << " SuspendAll for " << cause << " starting...";
474   } else {
475     VLOG(threads) << "Thread[null] SuspendAll for " << cause << " starting...";
476   }
477   {
478     ScopedTrace trace("Suspending mutator threads");
479     const uint64_t start_time = NanoTime();
480 
481     SuspendAllInternal(self, self);
482     // All threads are known to have suspended (but a thread may still own the mutator lock)
483     // Make sure this thread grabs exclusive access to the mutator lock and its protected data.
484 #if HAVE_TIMED_RWLOCK
485     while (true) {
486       if (Locks::mutator_lock_->ExclusiveLockWithTimeout(self, kThreadSuspendTimeoutMs, 0)) {
487         break;
488       } else if (!long_suspend_) {
489         // Reading long_suspend without the mutator lock is slightly racy, in some rare cases, this
490         // could result in a thread suspend timeout.
491         // Timeout if we wait more than kThreadSuspendTimeoutMs seconds.
492         UnsafeLogFatalForThreadSuspendAllTimeout();
493       }
494     }
495 #else
496     Locks::mutator_lock_->ExclusiveLock(self);
497 #endif
498 
499     long_suspend_ = long_suspend;
500 
501     const uint64_t end_time = NanoTime();
502     const uint64_t suspend_time = end_time - start_time;
503     suspend_all_historam_.AdjustAndAddValue(suspend_time);
504     if (suspend_time > kLongThreadSuspendThreshold) {
505       LOG(WARNING) << "Suspending all threads took: " << PrettyDuration(suspend_time);
506     }
507 
508     if (kDebugLocking) {
509       // Debug check that all threads are suspended.
510       AssertThreadsAreSuspended(self, self);
511     }
512   }
513   ATRACE_BEGIN((std::string("Mutator threads suspended for ") + cause).c_str());
514 
515   if (self != nullptr) {
516     VLOG(threads) << *self << " SuspendAll complete";
517   } else {
518     VLOG(threads) << "Thread[null] SuspendAll complete";
519   }
520 }
521 
522 // Ensures all threads running Java suspend and that those not running Java don't start.
523 // Debugger thread might be set to kRunnable for a short period of time after the
524 // SuspendAllInternal. This is safe because it will be set back to suspended state before
525 // the SuspendAll returns.
SuspendAllInternal(Thread * self,Thread * ignore1,Thread * ignore2,bool debug_suspend)526 void ThreadList::SuspendAllInternal(Thread* self,
527                                     Thread* ignore1,
528                                     Thread* ignore2,
529                                     bool debug_suspend) {
530   Locks::mutator_lock_->AssertNotExclusiveHeld(self);
531   Locks::thread_list_lock_->AssertNotHeld(self);
532   Locks::thread_suspend_count_lock_->AssertNotHeld(self);
533   if (kDebugLocking && self != nullptr) {
534     CHECK_NE(self->GetState(), kRunnable);
535   }
536 
537   // First request that all threads suspend, then wait for them to suspend before
538   // returning. This suspension scheme also relies on other behaviour:
539   // 1. Threads cannot be deleted while they are suspended or have a suspend-
540   //    request flag set - (see Unregister() below).
541   // 2. When threads are created, they are created in a suspended state (actually
542   //    kNative) and will never begin executing Java code without first checking
543   //    the suspend-request flag.
544 
545   // The atomic counter for number of threads that need to pass the barrier.
546   AtomicInteger pending_threads;
547   uint32_t num_ignored = 0;
548   if (ignore1 != nullptr) {
549     ++num_ignored;
550   }
551   if (ignore2 != nullptr && ignore1 != ignore2) {
552     ++num_ignored;
553   }
554   {
555     MutexLock mu(self, *Locks::thread_list_lock_);
556     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
557     // Update global suspend all state for attaching threads.
558     ++suspend_all_count_;
559     if (debug_suspend)
560       ++debug_suspend_all_count_;
561     pending_threads.StoreRelaxed(list_.size() - num_ignored);
562     // Increment everybody's suspend count (except those that should be ignored).
563     for (const auto& thread : list_) {
564       if (thread == ignore1 || thread == ignore2) {
565         continue;
566       }
567       VLOG(threads) << "requesting thread suspend: " << *thread;
568       while (true) {
569         if (LIKELY(thread->ModifySuspendCount(self, +1, &pending_threads, debug_suspend))) {
570           break;
571         } else {
572           // Failure means the list of active_suspend_barriers is full, we should release the
573           // thread_suspend_count_lock_ (to avoid deadlock) and wait till the target thread has
574           // executed Thread::PassActiveSuspendBarriers(). Note that we could not simply wait for
575           // the thread to change to a suspended state, because it might need to run checkpoint
576           // function before the state change, which also needs thread_suspend_count_lock_.
577 
578           // This is very unlikely to happen since more than kMaxSuspendBarriers threads need to
579           // execute SuspendAllInternal() simultaneously, and target thread stays in kRunnable
580           // in the mean time.
581           Locks::thread_suspend_count_lock_->ExclusiveUnlock(self);
582           NanoSleep(100000);
583           Locks::thread_suspend_count_lock_->ExclusiveLock(self);
584         }
585       }
586 
587       // Must install the pending_threads counter first, then check thread->IsSuspend() and clear
588       // the counter. Otherwise there's a race with Thread::TransitionFromRunnableToSuspended()
589       // that can lead a thread to miss a call to PassActiveSuspendBarriers().
590       if (thread->IsSuspended()) {
591         // Only clear the counter for the current thread.
592         thread->ClearSuspendBarrier(&pending_threads);
593         pending_threads.FetchAndSubSequentiallyConsistent(1);
594       }
595     }
596   }
597 
598   // Wait for the barrier to be passed by all runnable threads. This wait
599   // is done with a timeout so that we can detect problems.
600 #if ART_USE_FUTEXES
601   timespec wait_timeout;
602   InitTimeSpec(true, CLOCK_MONOTONIC, 10000, 0, &wait_timeout);
603 #endif
604   while (true) {
605     int32_t cur_val = pending_threads.LoadRelaxed();
606     if (LIKELY(cur_val > 0)) {
607 #if ART_USE_FUTEXES
608       if (futex(pending_threads.Address(), FUTEX_WAIT, cur_val, &wait_timeout, nullptr, 0) != 0) {
609         // EAGAIN and EINTR both indicate a spurious failure, try again from the beginning.
610         if ((errno != EAGAIN) && (errno != EINTR)) {
611           if (errno == ETIMEDOUT) {
612             LOG(kIsDebugBuild ? FATAL : ERROR) << "Unexpected time out during suspend all.";
613           } else {
614             PLOG(FATAL) << "futex wait failed for SuspendAllInternal()";
615           }
616         }
617       } else {
618         cur_val = pending_threads.LoadRelaxed();
619         CHECK_EQ(cur_val, 0);
620         break;
621       }
622 #else
623       // Spin wait. This is likely to be slow, but on most architecture ART_USE_FUTEXES is set.
624 #endif
625     } else {
626       CHECK_EQ(cur_val, 0);
627       break;
628     }
629   }
630 }
631 
ResumeAll()632 void ThreadList::ResumeAll() {
633   Thread* self = Thread::Current();
634 
635   if (self != nullptr) {
636     VLOG(threads) << *self << " ResumeAll starting";
637   } else {
638     VLOG(threads) << "Thread[null] ResumeAll starting";
639   }
640 
641   ATRACE_END();
642 
643   ScopedTrace trace("Resuming mutator threads");
644 
645   if (kDebugLocking) {
646     // Debug check that all threads are suspended.
647     AssertThreadsAreSuspended(self, self);
648   }
649 
650   long_suspend_ = false;
651 
652   Locks::mutator_lock_->ExclusiveUnlock(self);
653   {
654     MutexLock mu(self, *Locks::thread_list_lock_);
655     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
656     // Update global suspend all state for attaching threads.
657     --suspend_all_count_;
658     // Decrement the suspend counts for all threads.
659     for (const auto& thread : list_) {
660       if (thread == self) {
661         continue;
662       }
663       thread->ModifySuspendCount(self, -1, nullptr, false);
664     }
665 
666     // Broadcast a notification to all suspended threads, some or all of
667     // which may choose to wake up.  No need to wait for them.
668     if (self != nullptr) {
669       VLOG(threads) << *self << " ResumeAll waking others";
670     } else {
671       VLOG(threads) << "Thread[null] ResumeAll waking others";
672     }
673     Thread::resume_cond_->Broadcast(self);
674   }
675 
676   if (self != nullptr) {
677     VLOG(threads) << *self << " ResumeAll complete";
678   } else {
679     VLOG(threads) << "Thread[null] ResumeAll complete";
680   }
681 }
682 
Resume(Thread * thread,bool for_debugger)683 void ThreadList::Resume(Thread* thread, bool for_debugger) {
684   // This assumes there was an ATRACE_BEGIN when we suspended the thread.
685   ATRACE_END();
686 
687   Thread* self = Thread::Current();
688   DCHECK_NE(thread, self);
689   VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") starting..."
690       << (for_debugger ? " (debugger)" : "");
691 
692   {
693     // To check Contains.
694     MutexLock mu(self, *Locks::thread_list_lock_);
695     // To check IsSuspended.
696     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
697     DCHECK(thread->IsSuspended());
698     if (!Contains(thread)) {
699       // We only expect threads within the thread-list to have been suspended otherwise we can't
700       // stop such threads from delete-ing themselves.
701       LOG(ERROR) << "Resume(" << reinterpret_cast<void*>(thread)
702           << ") thread not within thread list";
703       return;
704     }
705     thread->ModifySuspendCount(self, -1, nullptr, for_debugger);
706   }
707 
708   {
709     VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") waking others";
710     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
711     Thread::resume_cond_->Broadcast(self);
712   }
713 
714   VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") complete";
715 }
716 
ThreadSuspendByPeerWarning(Thread * self,LogSeverity severity,const char * message,jobject peer)717 static void ThreadSuspendByPeerWarning(Thread* self,
718                                        LogSeverity severity,
719                                        const char* message,
720                                        jobject peer) {
721   JNIEnvExt* env = self->GetJniEnv();
722   ScopedLocalRef<jstring>
723       scoped_name_string(env, static_cast<jstring>(env->GetObjectField(
724           peer, WellKnownClasses::java_lang_Thread_name)));
725   ScopedUtfChars scoped_name_chars(env, scoped_name_string.get());
726   if (scoped_name_chars.c_str() == nullptr) {
727       LOG(severity) << message << ": " << peer;
728       env->ExceptionClear();
729   } else {
730       LOG(severity) << message << ": " << peer << ":" << scoped_name_chars.c_str();
731   }
732 }
733 
SuspendThreadByPeer(jobject peer,bool request_suspension,bool debug_suspension,bool * timed_out)734 Thread* ThreadList::SuspendThreadByPeer(jobject peer,
735                                         bool request_suspension,
736                                         bool debug_suspension,
737                                         bool* timed_out) {
738   const uint64_t start_time = NanoTime();
739   useconds_t sleep_us = kThreadSuspendInitialSleepUs;
740   *timed_out = false;
741   Thread* const self = Thread::Current();
742   Thread* suspended_thread = nullptr;
743   VLOG(threads) << "SuspendThreadByPeer starting";
744   while (true) {
745     Thread* thread;
746     {
747       // Note: this will transition to runnable and potentially suspend. We ensure only one thread
748       // is requesting another suspend, to avoid deadlock, by requiring this function be called
749       // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
750       // than request thread suspension, to avoid potential cycles in threads requesting each other
751       // suspend.
752       ScopedObjectAccess soa(self);
753       MutexLock thread_list_mu(self, *Locks::thread_list_lock_);
754       thread = Thread::FromManagedThread(soa, peer);
755       if (thread == nullptr) {
756         if (suspended_thread != nullptr) {
757           MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
758           // If we incremented the suspend count but the thread reset its peer, we need to
759           // re-decrement it since it is shutting down and may deadlock the runtime in
760           // ThreadList::WaitForOtherNonDaemonThreadsToExit.
761           suspended_thread->ModifySuspendCount(soa.Self(), -1, nullptr, debug_suspension);
762         }
763         ThreadSuspendByPeerWarning(self, WARNING, "No such thread for suspend", peer);
764         return nullptr;
765       }
766       if (!Contains(thread)) {
767         CHECK(suspended_thread == nullptr);
768         VLOG(threads) << "SuspendThreadByPeer failed for unattached thread: "
769             << reinterpret_cast<void*>(thread);
770         return nullptr;
771       }
772       VLOG(threads) << "SuspendThreadByPeer found thread: " << *thread;
773       {
774         MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
775         if (request_suspension) {
776           if (self->GetSuspendCount() > 0) {
777             // We hold the suspend count lock but another thread is trying to suspend us. Its not
778             // safe to try to suspend another thread in case we get a cycle. Start the loop again
779             // which will allow this thread to be suspended.
780             continue;
781           }
782           CHECK(suspended_thread == nullptr);
783           suspended_thread = thread;
784           suspended_thread->ModifySuspendCount(self, +1, nullptr, debug_suspension);
785           request_suspension = false;
786         } else {
787           // If the caller isn't requesting suspension, a suspension should have already occurred.
788           CHECK_GT(thread->GetSuspendCount(), 0);
789         }
790         // IsSuspended on the current thread will fail as the current thread is changed into
791         // Runnable above. As the suspend count is now raised if this is the current thread
792         // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
793         // to just explicitly handle the current thread in the callers to this code.
794         CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
795         // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
796         // count, or else we've waited and it has self suspended) or is the current thread, we're
797         // done.
798         if (thread->IsSuspended()) {
799           VLOG(threads) << "SuspendThreadByPeer thread suspended: " << *thread;
800           if (ATRACE_ENABLED()) {
801             std::string name;
802             thread->GetThreadName(name);
803             ATRACE_BEGIN(StringPrintf("SuspendThreadByPeer suspended %s for peer=%p", name.c_str(),
804                                       peer).c_str());
805           }
806           return thread;
807         }
808         const uint64_t total_delay = NanoTime() - start_time;
809         if (total_delay >= MsToNs(kThreadSuspendTimeoutMs)) {
810           ThreadSuspendByPeerWarning(self, FATAL, "Thread suspension timed out", peer);
811           if (suspended_thread != nullptr) {
812             CHECK_EQ(suspended_thread, thread);
813             suspended_thread->ModifySuspendCount(soa.Self(), -1, nullptr, debug_suspension);
814           }
815           *timed_out = true;
816           return nullptr;
817         } else if (sleep_us == 0 &&
818             total_delay > static_cast<uint64_t>(kThreadSuspendMaxYieldUs) * 1000) {
819           // We have spun for kThreadSuspendMaxYieldUs time, switch to sleeps to prevent
820           // excessive CPU usage.
821           sleep_us = kThreadSuspendMaxYieldUs / 2;
822         }
823       }
824       // Release locks and come out of runnable state.
825     }
826     VLOG(threads) << "SuspendThreadByPeer waiting to allow thread chance to suspend";
827     ThreadSuspendSleep(sleep_us);
828     // This may stay at 0 if sleep_us == 0, but this is WAI since we want to avoid using usleep at
829     // all if possible. This shouldn't be an issue since time to suspend should always be small.
830     sleep_us = std::min(sleep_us * 2, kThreadSuspendMaxSleepUs);
831   }
832 }
833 
ThreadSuspendByThreadIdWarning(LogSeverity severity,const char * message,uint32_t thread_id)834 static void ThreadSuspendByThreadIdWarning(LogSeverity severity,
835                                            const char* message,
836                                            uint32_t thread_id) {
837   LOG(severity) << StringPrintf("%s: %d", message, thread_id);
838 }
839 
SuspendThreadByThreadId(uint32_t thread_id,bool debug_suspension,bool * timed_out)840 Thread* ThreadList::SuspendThreadByThreadId(uint32_t thread_id,
841                                             bool debug_suspension,
842                                             bool* timed_out) {
843   const uint64_t start_time = NanoTime();
844   useconds_t sleep_us = kThreadSuspendInitialSleepUs;
845   *timed_out = false;
846   Thread* suspended_thread = nullptr;
847   Thread* const self = Thread::Current();
848   CHECK_NE(thread_id, kInvalidThreadId);
849   VLOG(threads) << "SuspendThreadByThreadId starting";
850   while (true) {
851     {
852       // Note: this will transition to runnable and potentially suspend. We ensure only one thread
853       // is requesting another suspend, to avoid deadlock, by requiring this function be called
854       // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
855       // than request thread suspension, to avoid potential cycles in threads requesting each other
856       // suspend.
857       ScopedObjectAccess soa(self);
858       MutexLock thread_list_mu(self, *Locks::thread_list_lock_);
859       Thread* thread = nullptr;
860       for (const auto& it : list_) {
861         if (it->GetThreadId() == thread_id) {
862           thread = it;
863           break;
864         }
865       }
866       if (thread == nullptr) {
867         CHECK(suspended_thread == nullptr) << "Suspended thread " << suspended_thread
868             << " no longer in thread list";
869         // There's a race in inflating a lock and the owner giving up ownership and then dying.
870         ThreadSuspendByThreadIdWarning(WARNING, "No such thread id for suspend", thread_id);
871         return nullptr;
872       }
873       VLOG(threads) << "SuspendThreadByThreadId found thread: " << *thread;
874       DCHECK(Contains(thread));
875       {
876         MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
877         if (suspended_thread == nullptr) {
878           if (self->GetSuspendCount() > 0) {
879             // We hold the suspend count lock but another thread is trying to suspend us. Its not
880             // safe to try to suspend another thread in case we get a cycle. Start the loop again
881             // which will allow this thread to be suspended.
882             continue;
883           }
884           thread->ModifySuspendCount(self, +1, nullptr, debug_suspension);
885           suspended_thread = thread;
886         } else {
887           CHECK_EQ(suspended_thread, thread);
888           // If the caller isn't requesting suspension, a suspension should have already occurred.
889           CHECK_GT(thread->GetSuspendCount(), 0);
890         }
891         // IsSuspended on the current thread will fail as the current thread is changed into
892         // Runnable above. As the suspend count is now raised if this is the current thread
893         // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
894         // to just explicitly handle the current thread in the callers to this code.
895         CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
896         // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
897         // count, or else we've waited and it has self suspended) or is the current thread, we're
898         // done.
899         if (thread->IsSuspended()) {
900           if (ATRACE_ENABLED()) {
901             std::string name;
902             thread->GetThreadName(name);
903             ATRACE_BEGIN(StringPrintf("SuspendThreadByThreadId suspended %s id=%d",
904                                       name.c_str(), thread_id).c_str());
905           }
906           VLOG(threads) << "SuspendThreadByThreadId thread suspended: " << *thread;
907           return thread;
908         }
909         const uint64_t total_delay = NanoTime() - start_time;
910         if (total_delay >= MsToNs(kThreadSuspendTimeoutMs)) {
911           ThreadSuspendByThreadIdWarning(WARNING, "Thread suspension timed out", thread_id);
912           if (suspended_thread != nullptr) {
913             thread->ModifySuspendCount(soa.Self(), -1, nullptr, debug_suspension);
914           }
915           *timed_out = true;
916           return nullptr;
917         } else if (sleep_us == 0 &&
918             total_delay > static_cast<uint64_t>(kThreadSuspendMaxYieldUs) * 1000) {
919           // We have spun for kThreadSuspendMaxYieldUs time, switch to sleeps to prevent
920           // excessive CPU usage.
921           sleep_us = kThreadSuspendMaxYieldUs / 2;
922         }
923       }
924       // Release locks and come out of runnable state.
925     }
926     VLOG(threads) << "SuspendThreadByThreadId waiting to allow thread chance to suspend";
927     ThreadSuspendSleep(sleep_us);
928     sleep_us = std::min(sleep_us * 2, kThreadSuspendMaxSleepUs);
929   }
930 }
931 
FindThreadByThreadId(uint32_t thread_id)932 Thread* ThreadList::FindThreadByThreadId(uint32_t thread_id) {
933   for (const auto& thread : list_) {
934     if (thread->GetThreadId() == thread_id) {
935       return thread;
936     }
937   }
938   return nullptr;
939 }
940 
SuspendAllForDebugger()941 void ThreadList::SuspendAllForDebugger() {
942   Thread* self = Thread::Current();
943   Thread* debug_thread = Dbg::GetDebugThread();
944 
945   VLOG(threads) << *self << " SuspendAllForDebugger starting...";
946 
947   SuspendAllInternal(self, self, debug_thread, true);
948   // Block on the mutator lock until all Runnable threads release their share of access then
949   // immediately unlock again.
950 #if HAVE_TIMED_RWLOCK
951   // Timeout if we wait more than 30 seconds.
952   if (!Locks::mutator_lock_->ExclusiveLockWithTimeout(self, 30 * 1000, 0)) {
953     UnsafeLogFatalForThreadSuspendAllTimeout();
954   } else {
955     Locks::mutator_lock_->ExclusiveUnlock(self);
956   }
957 #else
958   Locks::mutator_lock_->ExclusiveLock(self);
959   Locks::mutator_lock_->ExclusiveUnlock(self);
960 #endif
961   // Disabled for the following race condition:
962   // Thread 1 calls SuspendAllForDebugger, gets preempted after pulsing the mutator lock.
963   // Thread 2 calls SuspendAll and SetStateUnsafe (perhaps from Dbg::Disconnected).
964   // Thread 1 fails assertion that all threads are suspended due to thread 2 being in a runnable
965   // state (from SetStateUnsafe).
966   // AssertThreadsAreSuspended(self, self, debug_thread);
967 
968   VLOG(threads) << *self << " SuspendAllForDebugger complete";
969 }
970 
SuspendSelfForDebugger()971 void ThreadList::SuspendSelfForDebugger() {
972   Thread* const self = Thread::Current();
973   self->SetReadyForDebugInvoke(true);
974 
975   // The debugger thread must not suspend itself due to debugger activity!
976   Thread* debug_thread = Dbg::GetDebugThread();
977   CHECK(self != debug_thread);
978   CHECK_NE(self->GetState(), kRunnable);
979   Locks::mutator_lock_->AssertNotHeld(self);
980 
981   // The debugger may have detached while we were executing an invoke request. In that case, we
982   // must not suspend ourself.
983   DebugInvokeReq* pReq = self->GetInvokeReq();
984   const bool skip_thread_suspension = (pReq != nullptr && !Dbg::IsDebuggerActive());
985   if (!skip_thread_suspension) {
986     // Collisions with other suspends aren't really interesting. We want
987     // to ensure that we're the only one fiddling with the suspend count
988     // though.
989     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
990     self->ModifySuspendCount(self, +1, nullptr, true);
991     CHECK_GT(self->GetSuspendCount(), 0);
992 
993     VLOG(threads) << *self << " self-suspending (debugger)";
994   } else {
995     // We must no longer be subject to debugger suspension.
996     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
997     CHECK_EQ(self->GetDebugSuspendCount(), 0) << "Debugger detached without resuming us";
998 
999     VLOG(threads) << *self << " not self-suspending because debugger detached during invoke";
1000   }
1001 
1002   // If the debugger requested an invoke, we need to send the reply and clear the request.
1003   if (pReq != nullptr) {
1004     Dbg::FinishInvokeMethod(pReq);
1005     self->ClearDebugInvokeReq();
1006     pReq = nullptr;  // object has been deleted, clear it for safety.
1007   }
1008 
1009   // Tell JDWP that we've completed suspension. The JDWP thread can't
1010   // tell us to resume before we're fully asleep because we hold the
1011   // suspend count lock.
1012   Dbg::ClearWaitForEventThread();
1013 
1014   {
1015     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
1016     while (self->GetSuspendCount() != 0) {
1017       Thread::resume_cond_->Wait(self);
1018       if (self->GetSuspendCount() != 0) {
1019         // The condition was signaled but we're still suspended. This
1020         // can happen when we suspend then resume all threads to
1021         // update instrumentation or compute monitor info. This can
1022         // also happen if the debugger lets go while a SIGQUIT thread
1023         // dump event is pending (assuming SignalCatcher was resumed for
1024         // just long enough to try to grab the thread-suspend lock).
1025         VLOG(jdwp) << *self << " still suspended after undo "
1026                    << "(suspend count=" << self->GetSuspendCount() << ", "
1027                    << "debug suspend count=" << self->GetDebugSuspendCount() << ")";
1028       }
1029     }
1030     CHECK_EQ(self->GetSuspendCount(), 0);
1031   }
1032 
1033   self->SetReadyForDebugInvoke(false);
1034   VLOG(threads) << *self << " self-reviving (debugger)";
1035 }
1036 
ResumeAllForDebugger()1037 void ThreadList::ResumeAllForDebugger() {
1038   Thread* self = Thread::Current();
1039   Thread* debug_thread = Dbg::GetDebugThread();
1040 
1041   VLOG(threads) << *self << " ResumeAllForDebugger starting...";
1042 
1043   // Threads can't resume if we exclusively hold the mutator lock.
1044   Locks::mutator_lock_->AssertNotExclusiveHeld(self);
1045 
1046   {
1047     MutexLock thread_list_mu(self, *Locks::thread_list_lock_);
1048     {
1049       MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
1050       // Update global suspend all state for attaching threads.
1051       DCHECK_GE(suspend_all_count_, debug_suspend_all_count_);
1052       if (debug_suspend_all_count_ > 0) {
1053         --suspend_all_count_;
1054         --debug_suspend_all_count_;
1055       } else {
1056         // We've been asked to resume all threads without being asked to
1057         // suspend them all before. That may happen if a debugger tries
1058         // to resume some suspended threads (with suspend count == 1)
1059         // at once with a VirtualMachine.Resume command. Let's print a
1060         // warning.
1061         LOG(WARNING) << "Debugger attempted to resume all threads without "
1062                      << "having suspended them all before.";
1063       }
1064       // Decrement everybody's suspend count (except our own).
1065       for (const auto& thread : list_) {
1066         if (thread == self || thread == debug_thread) {
1067           continue;
1068         }
1069         if (thread->GetDebugSuspendCount() == 0) {
1070           // This thread may have been individually resumed with ThreadReference.Resume.
1071           continue;
1072         }
1073         VLOG(threads) << "requesting thread resume: " << *thread;
1074         thread->ModifySuspendCount(self, -1, nullptr, true);
1075       }
1076     }
1077   }
1078 
1079   {
1080     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
1081     Thread::resume_cond_->Broadcast(self);
1082   }
1083 
1084   VLOG(threads) << *self << " ResumeAllForDebugger complete";
1085 }
1086 
UndoDebuggerSuspensions()1087 void ThreadList::UndoDebuggerSuspensions() {
1088   Thread* self = Thread::Current();
1089 
1090   VLOG(threads) << *self << " UndoDebuggerSuspensions starting";
1091 
1092   {
1093     MutexLock mu(self, *Locks::thread_list_lock_);
1094     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1095     // Update global suspend all state for attaching threads.
1096     suspend_all_count_ -= debug_suspend_all_count_;
1097     debug_suspend_all_count_ = 0;
1098     // Update running threads.
1099     for (const auto& thread : list_) {
1100       if (thread == self || thread->GetDebugSuspendCount() == 0) {
1101         continue;
1102       }
1103       thread->ModifySuspendCount(self, -thread->GetDebugSuspendCount(), nullptr, true);
1104     }
1105   }
1106 
1107   {
1108     MutexLock mu(self, *Locks::thread_suspend_count_lock_);
1109     Thread::resume_cond_->Broadcast(self);
1110   }
1111 
1112   VLOG(threads) << "UndoDebuggerSuspensions(" << *self << ") complete";
1113 }
1114 
WaitForOtherNonDaemonThreadsToExit()1115 void ThreadList::WaitForOtherNonDaemonThreadsToExit() {
1116   ScopedTrace trace(__PRETTY_FUNCTION__);
1117   Thread* self = Thread::Current();
1118   Locks::mutator_lock_->AssertNotHeld(self);
1119   while (true) {
1120     {
1121       // No more threads can be born after we start to shutdown.
1122       MutexLock mu(self, *Locks::runtime_shutdown_lock_);
1123       CHECK(Runtime::Current()->IsShuttingDownLocked());
1124       CHECK_EQ(Runtime::Current()->NumberOfThreadsBeingBorn(), 0U);
1125     }
1126     MutexLock mu(self, *Locks::thread_list_lock_);
1127     // Also wait for any threads that are unregistering to finish. This is required so that no
1128     // threads access the thread list after it is deleted. TODO: This may not work for user daemon
1129     // threads since they could unregister at the wrong time.
1130     bool done = unregistering_count_ == 0;
1131     if (done) {
1132       for (const auto& thread : list_) {
1133         if (thread != self && !thread->IsDaemon()) {
1134           done = false;
1135           break;
1136         }
1137       }
1138     }
1139     if (done) {
1140       break;
1141     }
1142     // Wait for another thread to exit before re-checking.
1143     Locks::thread_exit_cond_->Wait(self);
1144   }
1145 }
1146 
SuspendAllDaemonThreadsForShutdown()1147 void ThreadList::SuspendAllDaemonThreadsForShutdown() {
1148   ScopedTrace trace(__PRETTY_FUNCTION__);
1149   Thread* self = Thread::Current();
1150   MutexLock mu(self, *Locks::thread_list_lock_);
1151   size_t daemons_left = 0;
1152   {  // Tell all the daemons it's time to suspend.
1153     MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1154     for (const auto& thread : list_) {
1155       // This is only run after all non-daemon threads have exited, so the remainder should all be
1156       // daemons.
1157       CHECK(thread->IsDaemon()) << *thread;
1158       if (thread != self) {
1159         thread->ModifySuspendCount(self, +1, nullptr, false);
1160         ++daemons_left;
1161       }
1162       // We are shutting down the runtime, set the JNI functions of all the JNIEnvs to be
1163       // the sleep forever one.
1164       thread->GetJniEnv()->SetFunctionsToRuntimeShutdownFunctions();
1165     }
1166   }
1167   // If we have any daemons left, wait 200ms to ensure they are not stuck in a place where they
1168   // are about to access runtime state and are not in a runnable state. Examples: Monitor code
1169   // or waking up from a condition variable. TODO: Try and see if there is a better way to wait
1170   // for daemon threads to be in a blocked state.
1171   if (daemons_left > 0) {
1172     static constexpr size_t kDaemonSleepTime = 200 * 1000;
1173     usleep(kDaemonSleepTime);
1174   }
1175   // Give the threads a chance to suspend, complaining if they're slow.
1176   bool have_complained = false;
1177   static constexpr size_t kTimeoutMicroseconds = 2000 * 1000;
1178   static constexpr size_t kSleepMicroseconds = 1000;
1179   for (size_t i = 0; i < kTimeoutMicroseconds / kSleepMicroseconds; ++i) {
1180     bool all_suspended = true;
1181     for (const auto& thread : list_) {
1182       if (thread != self && thread->GetState() == kRunnable) {
1183         if (!have_complained) {
1184           LOG(WARNING) << "daemon thread not yet suspended: " << *thread;
1185           have_complained = true;
1186         }
1187         all_suspended = false;
1188       }
1189     }
1190     if (all_suspended) {
1191       return;
1192     }
1193     usleep(kSleepMicroseconds);
1194   }
1195   LOG(WARNING) << "timed out suspending all daemon threads";
1196 }
1197 
Register(Thread * self)1198 void ThreadList::Register(Thread* self) {
1199   DCHECK_EQ(self, Thread::Current());
1200 
1201   if (VLOG_IS_ON(threads)) {
1202     std::ostringstream oss;
1203     self->ShortDump(oss);  // We don't hold the mutator_lock_ yet and so cannot call Dump.
1204     LOG(INFO) << "ThreadList::Register() " << *self  << "\n" << oss.str();
1205   }
1206 
1207   // Atomically add self to the thread list and make its thread_suspend_count_ reflect ongoing
1208   // SuspendAll requests.
1209   MutexLock mu(self, *Locks::thread_list_lock_);
1210   MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1211   CHECK_GE(suspend_all_count_, debug_suspend_all_count_);
1212   // Modify suspend count in increments of 1 to maintain invariants in ModifySuspendCount. While
1213   // this isn't particularly efficient the suspend counts are most commonly 0 or 1.
1214   for (int delta = debug_suspend_all_count_; delta > 0; delta--) {
1215     self->ModifySuspendCount(self, +1, nullptr, true);
1216   }
1217   for (int delta = suspend_all_count_ - debug_suspend_all_count_; delta > 0; delta--) {
1218     self->ModifySuspendCount(self, +1, nullptr, false);
1219   }
1220   CHECK(!Contains(self));
1221   list_.push_back(self);
1222   if (kUseReadBarrier) {
1223     // Initialize according to the state of the CC collector.
1224     bool is_gc_marking =
1225         Runtime::Current()->GetHeap()->ConcurrentCopyingCollector()->IsMarking();
1226     self->SetIsGcMarking(is_gc_marking);
1227     bool weak_ref_access_enabled =
1228         Runtime::Current()->GetHeap()->ConcurrentCopyingCollector()->IsWeakRefAccessEnabled();
1229     self->SetWeakRefAccessEnabled(weak_ref_access_enabled);
1230   }
1231 }
1232 
Unregister(Thread * self)1233 void ThreadList::Unregister(Thread* self) {
1234   DCHECK_EQ(self, Thread::Current());
1235   CHECK_NE(self->GetState(), kRunnable);
1236   Locks::mutator_lock_->AssertNotHeld(self);
1237 
1238   VLOG(threads) << "ThreadList::Unregister() " << *self;
1239 
1240   {
1241     MutexLock mu(self, *Locks::thread_list_lock_);
1242     ++unregistering_count_;
1243   }
1244 
1245   // Any time-consuming destruction, plus anything that can call back into managed code or
1246   // suspend and so on, must happen at this point, and not in ~Thread. The self->Destroy is what
1247   // causes the threads to join. It is important to do this after incrementing unregistering_count_
1248   // since we want the runtime to wait for the daemon threads to exit before deleting the thread
1249   // list.
1250   self->Destroy();
1251 
1252   // If tracing, remember thread id and name before thread exits.
1253   Trace::StoreExitingThreadInfo(self);
1254 
1255   uint32_t thin_lock_id = self->GetThreadId();
1256   while (true) {
1257     // Remove and delete the Thread* while holding the thread_list_lock_ and
1258     // thread_suspend_count_lock_ so that the unregistering thread cannot be suspended.
1259     // Note: deliberately not using MutexLock that could hold a stale self pointer.
1260     MutexLock mu(self, *Locks::thread_list_lock_);
1261     if (!Contains(self)) {
1262       std::string thread_name;
1263       self->GetThreadName(thread_name);
1264       std::ostringstream os;
1265       DumpNativeStack(os, GetTid(), nullptr, "  native: ", nullptr);
1266       LOG(ERROR) << "Request to unregister unattached thread " << thread_name << "\n" << os.str();
1267       break;
1268     } else {
1269       MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1270       if (!self->IsSuspended()) {
1271         list_.remove(self);
1272         break;
1273       }
1274     }
1275     // We failed to remove the thread due to a suspend request, loop and try again.
1276   }
1277   delete self;
1278 
1279   // Release the thread ID after the thread is finished and deleted to avoid cases where we can
1280   // temporarily have multiple threads with the same thread id. When this occurs, it causes
1281   // problems in FindThreadByThreadId / SuspendThreadByThreadId.
1282   ReleaseThreadId(nullptr, thin_lock_id);
1283 
1284   // Clear the TLS data, so that the underlying native thread is recognizably detached.
1285   // (It may wish to reattach later.)
1286 #ifdef __ANDROID__
1287   __get_tls()[TLS_SLOT_ART_THREAD_SELF] = nullptr;
1288 #else
1289   CHECK_PTHREAD_CALL(pthread_setspecific, (Thread::pthread_key_self_, nullptr), "detach self");
1290 #endif
1291 
1292   // Signal that a thread just detached.
1293   MutexLock mu(nullptr, *Locks::thread_list_lock_);
1294   --unregistering_count_;
1295   Locks::thread_exit_cond_->Broadcast(nullptr);
1296 }
1297 
ForEach(void (* callback)(Thread *,void *),void * context)1298 void ThreadList::ForEach(void (*callback)(Thread*, void*), void* context) {
1299   for (const auto& thread : list_) {
1300     callback(thread, context);
1301   }
1302 }
1303 
VisitRoots(RootVisitor * visitor) const1304 void ThreadList::VisitRoots(RootVisitor* visitor) const {
1305   MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
1306   for (const auto& thread : list_) {
1307     thread->VisitRoots(visitor);
1308   }
1309 }
1310 
AllocThreadId(Thread * self)1311 uint32_t ThreadList::AllocThreadId(Thread* self) {
1312   MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
1313   for (size_t i = 0; i < allocated_ids_.size(); ++i) {
1314     if (!allocated_ids_[i]) {
1315       allocated_ids_.set(i);
1316       return i + 1;  // Zero is reserved to mean "invalid".
1317     }
1318   }
1319   LOG(FATAL) << "Out of internal thread ids";
1320   return 0;
1321 }
1322 
ReleaseThreadId(Thread * self,uint32_t id)1323 void ThreadList::ReleaseThreadId(Thread* self, uint32_t id) {
1324   MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
1325   --id;  // Zero is reserved to mean "invalid".
1326   DCHECK(allocated_ids_[id]) << id;
1327   allocated_ids_.reset(id);
1328 }
1329 
ScopedSuspendAll(const char * cause,bool long_suspend)1330 ScopedSuspendAll::ScopedSuspendAll(const char* cause, bool long_suspend) {
1331   Runtime::Current()->GetThreadList()->SuspendAll(cause, long_suspend);
1332 }
1333 
~ScopedSuspendAll()1334 ScopedSuspendAll::~ScopedSuspendAll() {
1335   Runtime::Current()->GetThreadList()->ResumeAll();
1336 }
1337 
1338 }  // namespace art
1339