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 <dirent.h>
20 #include <sys/types.h>
21 #include <unistd.h>
22
23 #include <sstream>
24 #include <vector>
25
26 #include "android-base/stringprintf.h"
27 #include "backtrace/BacktraceMap.h"
28 #include "nativehelper/scoped_local_ref.h"
29 #include "nativehelper/scoped_utf_chars.h"
30
31 #include "base/aborting.h"
32 #include "base/histogram-inl.h"
33 #include "base/mutex-inl.h"
34 #include "base/systrace.h"
35 #include "base/time_utils.h"
36 #include "base/timing_logger.h"
37 #include "debugger.h"
38 #include "gc/collector/concurrent_copying.h"
39 #include "gc/gc_pause_listener.h"
40 #include "gc/heap.h"
41 #include "gc/reference_processor.h"
42 #include "gc_root.h"
43 #include "jni/jni_internal.h"
44 #include "lock_word.h"
45 #include "monitor.h"
46 #include "native_stack_dump.h"
47 #include "scoped_thread_state_change-inl.h"
48 #include "thread.h"
49 #include "trace.h"
50 #include "well_known_classes.h"
51
52 #if ART_USE_FUTEXES
53 #include "linux/futex.h"
54 #include "sys/syscall.h"
55 #ifndef SYS_futex
56 #define SYS_futex __NR_futex
57 #endif
58 #endif // ART_USE_FUTEXES
59
60 namespace art {
61
62 using android::base::StringPrintf;
63
64 static constexpr uint64_t kLongThreadSuspendThreshold = MsToNs(5);
65 // Use 0 since we want to yield to prevent blocking for an unpredictable amount of time.
66 static constexpr useconds_t kThreadSuspendInitialSleepUs = 0;
67 static constexpr useconds_t kThreadSuspendMaxYieldUs = 3000;
68 static constexpr useconds_t kThreadSuspendMaxSleepUs = 5000;
69
70 // Whether we should try to dump the native stack of unattached threads. See commit ed8b723 for
71 // some history.
72 static constexpr bool kDumpUnattachedThreadNativeStackForSigQuit = true;
73
ThreadList(uint64_t thread_suspend_timeout_ns)74 ThreadList::ThreadList(uint64_t thread_suspend_timeout_ns)
75 : suspend_all_count_(0),
76 unregistering_count_(0),
77 suspend_all_historam_("suspend all histogram", 16, 64),
78 long_suspend_(false),
79 shut_down_(false),
80 thread_suspend_timeout_ns_(thread_suspend_timeout_ns),
81 empty_checkpoint_barrier_(new Barrier(0)) {
82 CHECK(Monitor::IsValidLockWord(LockWord::FromThinLockId(kMaxThreadId, 1, 0U)));
83 }
84
~ThreadList()85 ThreadList::~ThreadList() {
86 CHECK(shut_down_);
87 }
88
ShutDown()89 void ThreadList::ShutDown() {
90 ScopedTrace trace(__PRETTY_FUNCTION__);
91 // Detach the current thread if necessary. If we failed to start, there might not be any threads.
92 // We need to detach the current thread here in case there's another thread waiting to join with
93 // us.
94 bool contains = false;
95 Thread* self = Thread::Current();
96 {
97 MutexLock mu(self, *Locks::thread_list_lock_);
98 contains = Contains(self);
99 }
100 if (contains) {
101 Runtime::Current()->DetachCurrentThread();
102 }
103 WaitForOtherNonDaemonThreadsToExit();
104 // Disable GC and wait for GC to complete in case there are still daemon threads doing
105 // allocations.
106 gc::Heap* const heap = Runtime::Current()->GetHeap();
107 heap->DisableGCForShutdown();
108 // In case a GC is in progress, wait for it to finish.
109 heap->WaitForGcToComplete(gc::kGcCauseBackground, Thread::Current());
110 // TODO: there's an unaddressed race here where a thread may attach during shutdown, see
111 // Thread::Init.
112 SuspendAllDaemonThreadsForShutdown();
113
114 shut_down_ = true;
115 }
116
Contains(Thread * thread)117 bool ThreadList::Contains(Thread* thread) {
118 return find(list_.begin(), list_.end(), thread) != list_.end();
119 }
120
Contains(pid_t tid)121 bool ThreadList::Contains(pid_t tid) {
122 for (const auto& thread : list_) {
123 if (thread->GetTid() == tid) {
124 return true;
125 }
126 }
127 return false;
128 }
129
GetLockOwner()130 pid_t ThreadList::GetLockOwner() {
131 return Locks::thread_list_lock_->GetExclusiveOwnerTid();
132 }
133
DumpNativeStacks(std::ostream & os)134 void ThreadList::DumpNativeStacks(std::ostream& os) {
135 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
136 std::unique_ptr<BacktraceMap> map(BacktraceMap::Create(getpid()));
137 for (const auto& thread : list_) {
138 os << "DUMPING THREAD " << thread->GetTid() << "\n";
139 DumpNativeStack(os, thread->GetTid(), map.get(), "\t");
140 os << "\n";
141 }
142 }
143
DumpForSigQuit(std::ostream & os)144 void ThreadList::DumpForSigQuit(std::ostream& os) {
145 {
146 ScopedObjectAccess soa(Thread::Current());
147 // Only print if we have samples.
148 if (suspend_all_historam_.SampleSize() > 0) {
149 Histogram<uint64_t>::CumulativeData data;
150 suspend_all_historam_.CreateHistogram(&data);
151 suspend_all_historam_.PrintConfidenceIntervals(os, 0.99, data); // Dump time to suspend.
152 }
153 }
154 bool dump_native_stack = Runtime::Current()->GetDumpNativeStackOnSigQuit();
155 Dump(os, dump_native_stack);
156 DumpUnattachedThreads(os, dump_native_stack && kDumpUnattachedThreadNativeStackForSigQuit);
157 }
158
DumpUnattachedThread(std::ostream & os,pid_t tid,bool dump_native_stack)159 static void DumpUnattachedThread(std::ostream& os, pid_t tid, bool dump_native_stack)
160 NO_THREAD_SAFETY_ANALYSIS {
161 // TODO: No thread safety analysis as DumpState with a null thread won't access fields, should
162 // refactor DumpState to avoid skipping analysis.
163 Thread::DumpState(os, nullptr, tid);
164 if (dump_native_stack) {
165 DumpNativeStack(os, tid, nullptr, " native: ");
166 }
167 os << std::endl;
168 }
169
DumpUnattachedThreads(std::ostream & os,bool dump_native_stack)170 void ThreadList::DumpUnattachedThreads(std::ostream& os, bool dump_native_stack) {
171 DIR* d = opendir("/proc/self/task");
172 if (!d) {
173 return;
174 }
175
176 Thread* self = Thread::Current();
177 dirent* e;
178 while ((e = readdir(d)) != nullptr) {
179 char* end;
180 pid_t tid = strtol(e->d_name, &end, 10);
181 if (!*end) {
182 bool contains;
183 {
184 MutexLock mu(self, *Locks::thread_list_lock_);
185 contains = Contains(tid);
186 }
187 if (!contains) {
188 DumpUnattachedThread(os, tid, dump_native_stack);
189 }
190 }
191 }
192 closedir(d);
193 }
194
195 // Dump checkpoint timeout in milliseconds. Larger amount on the target, since the device could be
196 // overloaded with ANR dumps.
197 static constexpr uint32_t kDumpWaitTimeout = kIsTargetBuild ? 100000 : 20000;
198
199 // A closure used by Thread::Dump.
200 class DumpCheckpoint final : public Closure {
201 public:
DumpCheckpoint(std::ostream * os,bool dump_native_stack)202 DumpCheckpoint(std::ostream* os, bool dump_native_stack)
203 : os_(os),
204 // Avoid verifying count in case a thread doesn't end up passing through the barrier.
205 // This avoids a SIGABRT that would otherwise happen in the destructor.
206 barrier_(0, /*verify_count_on_shutdown=*/false),
207 backtrace_map_(dump_native_stack ? BacktraceMap::Create(getpid()) : nullptr),
208 dump_native_stack_(dump_native_stack) {
209 if (backtrace_map_ != nullptr) {
210 backtrace_map_->SetSuffixesToIgnore(std::vector<std::string> { "oat", "odex" });
211 }
212 }
213
Run(Thread * thread)214 void Run(Thread* thread) override {
215 // Note thread and self may not be equal if thread was already suspended at the point of the
216 // request.
217 Thread* self = Thread::Current();
218 CHECK(self != nullptr);
219 std::ostringstream local_os;
220 {
221 ScopedObjectAccess soa(self);
222 thread->Dump(local_os, dump_native_stack_, backtrace_map_.get());
223 }
224 {
225 // Use the logging lock to ensure serialization when writing to the common ostream.
226 MutexLock mu(self, *Locks::logging_lock_);
227 *os_ << local_os.str() << std::endl;
228 }
229 barrier_.Pass(self);
230 }
231
WaitForThreadsToRunThroughCheckpoint(size_t threads_running_checkpoint)232 void WaitForThreadsToRunThroughCheckpoint(size_t threads_running_checkpoint) {
233 Thread* self = Thread::Current();
234 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
235 bool timed_out = barrier_.Increment(self, threads_running_checkpoint, kDumpWaitTimeout);
236 if (timed_out) {
237 // Avoid a recursive abort.
238 LOG((kIsDebugBuild && (gAborting == 0)) ? ::android::base::FATAL : ::android::base::ERROR)
239 << "Unexpected time out during dump checkpoint.";
240 }
241 }
242
243 private:
244 // The common stream that will accumulate all the dumps.
245 std::ostream* const os_;
246 // The barrier to be passed through and for the requestor to wait upon.
247 Barrier barrier_;
248 // A backtrace map, so that all threads use a shared info and don't reacquire/parse separately.
249 std::unique_ptr<BacktraceMap> backtrace_map_;
250 // Whether we should dump the native stack.
251 const bool dump_native_stack_;
252 };
253
Dump(std::ostream & os,bool dump_native_stack)254 void ThreadList::Dump(std::ostream& os, bool dump_native_stack) {
255 Thread* self = Thread::Current();
256 {
257 MutexLock mu(self, *Locks::thread_list_lock_);
258 os << "DALVIK THREADS (" << list_.size() << "):\n";
259 }
260 if (self != nullptr) {
261 DumpCheckpoint checkpoint(&os, dump_native_stack);
262 size_t threads_running_checkpoint;
263 {
264 // Use SOA to prevent deadlocks if multiple threads are calling Dump() at the same time.
265 ScopedObjectAccess soa(self);
266 threads_running_checkpoint = RunCheckpoint(&checkpoint);
267 }
268 if (threads_running_checkpoint != 0) {
269 checkpoint.WaitForThreadsToRunThroughCheckpoint(threads_running_checkpoint);
270 }
271 } else {
272 DumpUnattachedThreads(os, dump_native_stack);
273 }
274 }
275
AssertThreadsAreSuspended(Thread * self,Thread * ignore1,Thread * ignore2)276 void ThreadList::AssertThreadsAreSuspended(Thread* self, Thread* ignore1, Thread* ignore2) {
277 MutexLock mu(self, *Locks::thread_list_lock_);
278 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
279 for (const auto& thread : list_) {
280 if (thread != ignore1 && thread != ignore2) {
281 CHECK(thread->IsSuspended())
282 << "\nUnsuspended thread: <<" << *thread << "\n"
283 << "self: <<" << *Thread::Current();
284 }
285 }
286 }
287
288 #if HAVE_TIMED_RWLOCK
289 // Attempt to rectify locks so that we dump thread list with required locks before exiting.
UnsafeLogFatalForThreadSuspendAllTimeout()290 NO_RETURN static void UnsafeLogFatalForThreadSuspendAllTimeout() {
291 // Increment gAborting before doing the thread list dump since we don't want any failures from
292 // AssertThreadSuspensionIsAllowable in cases where thread suspension is not allowed.
293 // See b/69044468.
294 ++gAborting;
295 Runtime* runtime = Runtime::Current();
296 std::ostringstream ss;
297 ss << "Thread suspend timeout\n";
298 Locks::mutator_lock_->Dump(ss);
299 ss << "\n";
300 runtime->GetThreadList()->Dump(ss);
301 --gAborting;
302 LOG(FATAL) << ss.str();
303 exit(0);
304 }
305 #endif
306
307 // Unlike suspending all threads where we can wait to acquire the mutator_lock_, suspending an
308 // individual thread requires polling. delay_us is the requested sleep wait. If delay_us is 0 then
309 // we use sched_yield instead of calling usleep.
310 // Although there is the possibility, here and elsewhere, that usleep could return -1 and
311 // errno = EINTR, there should be no problem if interrupted, so we do not check.
ThreadSuspendSleep(useconds_t delay_us)312 static void ThreadSuspendSleep(useconds_t delay_us) {
313 if (delay_us == 0) {
314 sched_yield();
315 } else {
316 usleep(delay_us);
317 }
318 }
319
RunCheckpoint(Closure * checkpoint_function,Closure * callback)320 size_t ThreadList::RunCheckpoint(Closure* checkpoint_function, Closure* callback) {
321 Thread* self = Thread::Current();
322 Locks::mutator_lock_->AssertNotExclusiveHeld(self);
323 Locks::thread_list_lock_->AssertNotHeld(self);
324 Locks::thread_suspend_count_lock_->AssertNotHeld(self);
325
326 std::vector<Thread*> suspended_count_modified_threads;
327 size_t count = 0;
328 {
329 // Call a checkpoint function for each thread, threads which are suspend get their checkpoint
330 // manually called.
331 MutexLock mu(self, *Locks::thread_list_lock_);
332 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
333 count = list_.size();
334 for (const auto& thread : list_) {
335 if (thread != self) {
336 bool requested_suspend = false;
337 while (true) {
338 if (thread->RequestCheckpoint(checkpoint_function)) {
339 // This thread will run its checkpoint some time in the near future.
340 if (requested_suspend) {
341 // The suspend request is now unnecessary.
342 bool updated =
343 thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
344 DCHECK(updated);
345 requested_suspend = false;
346 }
347 break;
348 } else {
349 // The thread is probably suspended, try to make sure that it stays suspended.
350 if (thread->GetState() == kRunnable) {
351 // Spurious fail, try again.
352 continue;
353 }
354 if (!requested_suspend) {
355 bool updated =
356 thread->ModifySuspendCount(self, +1, nullptr, SuspendReason::kInternal);
357 DCHECK(updated);
358 requested_suspend = true;
359 if (thread->IsSuspended()) {
360 break;
361 }
362 // The thread raced us to become Runnable. Try to RequestCheckpoint() again.
363 } else {
364 // The thread previously raced our suspend request to become Runnable but
365 // since it is suspended again, it must honor that suspend request now.
366 DCHECK(thread->IsSuspended());
367 break;
368 }
369 }
370 }
371 if (requested_suspend) {
372 suspended_count_modified_threads.push_back(thread);
373 }
374 }
375 }
376 // Run the callback to be called inside this critical section.
377 if (callback != nullptr) {
378 callback->Run(self);
379 }
380 }
381
382 // Run the checkpoint on ourself while we wait for threads to suspend.
383 checkpoint_function->Run(self);
384
385 // Run the checkpoint on the suspended threads.
386 for (const auto& thread : suspended_count_modified_threads) {
387 // We know for sure that the thread is suspended at this point.
388 DCHECK(thread->IsSuspended());
389 checkpoint_function->Run(thread);
390 {
391 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
392 bool updated = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
393 DCHECK(updated);
394 }
395 }
396
397 {
398 // Imitate ResumeAll, threads may be waiting on Thread::resume_cond_ since we raised their
399 // suspend count. Now the suspend_count_ is lowered so we must do the broadcast.
400 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
401 Thread::resume_cond_->Broadcast(self);
402 }
403
404 return count;
405 }
406
RunEmptyCheckpoint()407 void ThreadList::RunEmptyCheckpoint() {
408 Thread* self = Thread::Current();
409 Locks::mutator_lock_->AssertNotExclusiveHeld(self);
410 Locks::thread_list_lock_->AssertNotHeld(self);
411 Locks::thread_suspend_count_lock_->AssertNotHeld(self);
412 std::vector<uint32_t> runnable_thread_ids;
413 size_t count = 0;
414 Barrier* barrier = empty_checkpoint_barrier_.get();
415 barrier->Init(self, 0);
416 {
417 MutexLock mu(self, *Locks::thread_list_lock_);
418 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
419 for (Thread* thread : list_) {
420 if (thread != self) {
421 while (true) {
422 if (thread->RequestEmptyCheckpoint()) {
423 // This thread will run an empty checkpoint (decrement the empty checkpoint barrier)
424 // some time in the near future.
425 ++count;
426 if (kIsDebugBuild) {
427 runnable_thread_ids.push_back(thread->GetThreadId());
428 }
429 break;
430 }
431 if (thread->GetState() != kRunnable) {
432 // It's seen suspended, we are done because it must not be in the middle of a mutator
433 // heap access.
434 break;
435 }
436 }
437 }
438 }
439 }
440
441 // Wake up the threads blocking for weak ref access so that they will respond to the empty
442 // checkpoint request. Otherwise we will hang as they are blocking in the kRunnable state.
443 Runtime::Current()->GetHeap()->GetReferenceProcessor()->BroadcastForSlowPath(self);
444 Runtime::Current()->BroadcastForNewSystemWeaks(/*broadcast_for_checkpoint=*/true);
445 {
446 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
447 uint64_t total_wait_time = 0;
448 bool first_iter = true;
449 while (true) {
450 // Wake up the runnable threads blocked on the mutexes that another thread, which is blocked
451 // on a weak ref access, holds (indirectly blocking for weak ref access through another thread
452 // and a mutex.) This needs to be done periodically because the thread may be preempted
453 // between the CheckEmptyCheckpointFromMutex call and the subsequent futex wait in
454 // Mutex::ExclusiveLock, etc. when the wakeup via WakeupToRespondToEmptyCheckpoint
455 // arrives. This could cause a *very rare* deadlock, if not repeated. Most of the cases are
456 // handled in the first iteration.
457 for (BaseMutex* mutex : Locks::expected_mutexes_on_weak_ref_access_) {
458 mutex->WakeupToRespondToEmptyCheckpoint();
459 }
460 static constexpr uint64_t kEmptyCheckpointPeriodicTimeoutMs = 100; // 100ms
461 static constexpr uint64_t kEmptyCheckpointTotalTimeoutMs = 600 * 1000; // 10 minutes.
462 size_t barrier_count = first_iter ? count : 0;
463 first_iter = false; // Don't add to the barrier count from the second iteration on.
464 bool timed_out = barrier->Increment(self, barrier_count, kEmptyCheckpointPeriodicTimeoutMs);
465 if (!timed_out) {
466 break; // Success
467 }
468 // This is a very rare case.
469 total_wait_time += kEmptyCheckpointPeriodicTimeoutMs;
470 if (kIsDebugBuild && total_wait_time > kEmptyCheckpointTotalTimeoutMs) {
471 std::ostringstream ss;
472 ss << "Empty checkpoint timeout\n";
473 ss << "Barrier count " << barrier->GetCount(self) << "\n";
474 ss << "Runnable thread IDs";
475 for (uint32_t tid : runnable_thread_ids) {
476 ss << " " << tid;
477 }
478 ss << "\n";
479 Locks::mutator_lock_->Dump(ss);
480 ss << "\n";
481 LOG(FATAL_WITHOUT_ABORT) << ss.str();
482 // Some threads in 'runnable_thread_ids' are probably stuck. Try to dump their stacks.
483 // Avoid using ThreadList::Dump() initially because it is likely to get stuck as well.
484 {
485 ScopedObjectAccess soa(self);
486 MutexLock mu1(self, *Locks::thread_list_lock_);
487 for (Thread* thread : GetList()) {
488 uint32_t tid = thread->GetThreadId();
489 bool is_in_runnable_thread_ids =
490 std::find(runnable_thread_ids.begin(), runnable_thread_ids.end(), tid) !=
491 runnable_thread_ids.end();
492 if (is_in_runnable_thread_ids &&
493 thread->ReadFlag(kEmptyCheckpointRequest)) {
494 // Found a runnable thread that hasn't responded to the empty checkpoint request.
495 // Assume it's stuck and safe to dump its stack.
496 thread->Dump(LOG_STREAM(FATAL_WITHOUT_ABORT),
497 /*dump_native_stack=*/ true,
498 /*backtrace_map=*/ nullptr,
499 /*force_dump_stack=*/ true);
500 }
501 }
502 }
503 LOG(FATAL_WITHOUT_ABORT)
504 << "Dumped runnable threads that haven't responded to empty checkpoint.";
505 // Now use ThreadList::Dump() to dump more threads, noting it may get stuck.
506 Dump(LOG_STREAM(FATAL_WITHOUT_ABORT));
507 LOG(FATAL) << "Dumped all threads.";
508 }
509 }
510 }
511 }
512
513 // A checkpoint/suspend-all hybrid to switch thread roots from
514 // from-space to to-space refs. Used to synchronize threads at a point
515 // to mark the initiation of marking while maintaining the to-space
516 // invariant.
FlipThreadRoots(Closure * thread_flip_visitor,Closure * flip_callback,gc::collector::GarbageCollector * collector,gc::GcPauseListener * pause_listener)517 size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
518 Closure* flip_callback,
519 gc::collector::GarbageCollector* collector,
520 gc::GcPauseListener* pause_listener) {
521 TimingLogger::ScopedTiming split("ThreadListFlip", collector->GetTimings());
522 Thread* self = Thread::Current();
523 Locks::mutator_lock_->AssertNotHeld(self);
524 Locks::thread_list_lock_->AssertNotHeld(self);
525 Locks::thread_suspend_count_lock_->AssertNotHeld(self);
526 CHECK_NE(self->GetState(), kRunnable);
527
528 collector->GetHeap()->ThreadFlipBegin(self); // Sync with JNI critical calls.
529
530 // ThreadFlipBegin happens before we suspend all the threads, so it does not count towards the
531 // pause.
532 const uint64_t suspend_start_time = NanoTime();
533 SuspendAllInternal(self, self, nullptr);
534 if (pause_listener != nullptr) {
535 pause_listener->StartPause();
536 }
537
538 // Run the flip callback for the collector.
539 Locks::mutator_lock_->ExclusiveLock(self);
540 suspend_all_historam_.AdjustAndAddValue(NanoTime() - suspend_start_time);
541 flip_callback->Run(self);
542 Locks::mutator_lock_->ExclusiveUnlock(self);
543 collector->RegisterPause(NanoTime() - suspend_start_time);
544 if (pause_listener != nullptr) {
545 pause_listener->EndPause();
546 }
547
548 // Resume runnable threads.
549 size_t runnable_thread_count = 0;
550 std::vector<Thread*> other_threads;
551 {
552 TimingLogger::ScopedTiming split2("ResumeRunnableThreads", collector->GetTimings());
553 MutexLock mu(self, *Locks::thread_list_lock_);
554 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
555 --suspend_all_count_;
556 for (const auto& thread : list_) {
557 // Set the flip function for all threads because Thread::DumpState/DumpJavaStack() (invoked by
558 // a checkpoint) may cause the flip function to be run for a runnable/suspended thread before
559 // a runnable thread runs it for itself or we run it for a suspended thread below.
560 thread->SetFlipFunction(thread_flip_visitor);
561 if (thread == self) {
562 continue;
563 }
564 // Resume early the threads that were runnable but are suspended just for this thread flip or
565 // about to transition from non-runnable (eg. kNative at the SOA entry in a JNI function) to
566 // runnable (both cases waiting inside Thread::TransitionFromSuspendedToRunnable), or waiting
567 // for the thread flip to end at the JNI critical section entry (kWaitingForGcThreadFlip),
568 ThreadState state = thread->GetState();
569 if ((state == kWaitingForGcThreadFlip || thread->IsTransitioningToRunnable()) &&
570 thread->GetSuspendCount() == 1) {
571 // The thread will resume right after the broadcast.
572 bool updated = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
573 DCHECK(updated);
574 ++runnable_thread_count;
575 } else {
576 other_threads.push_back(thread);
577 }
578 }
579 Thread::resume_cond_->Broadcast(self);
580 }
581
582 collector->GetHeap()->ThreadFlipEnd(self);
583
584 // Run the closure on the other threads and let them resume.
585 {
586 TimingLogger::ScopedTiming split3("FlipOtherThreads", collector->GetTimings());
587 ReaderMutexLock mu(self, *Locks::mutator_lock_);
588 for (const auto& thread : other_threads) {
589 Closure* flip_func = thread->GetFlipFunction();
590 if (flip_func != nullptr) {
591 flip_func->Run(thread);
592 }
593 }
594 // Run it for self.
595 Closure* flip_func = self->GetFlipFunction();
596 if (flip_func != nullptr) {
597 flip_func->Run(self);
598 }
599 }
600
601 // Resume other threads.
602 {
603 TimingLogger::ScopedTiming split4("ResumeOtherThreads", collector->GetTimings());
604 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
605 for (const auto& thread : other_threads) {
606 bool updated = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
607 DCHECK(updated);
608 }
609 Thread::resume_cond_->Broadcast(self);
610 }
611
612 return runnable_thread_count + other_threads.size() + 1; // +1 for self.
613 }
614
SuspendAll(const char * cause,bool long_suspend)615 void ThreadList::SuspendAll(const char* cause, bool long_suspend) {
616 Thread* self = Thread::Current();
617
618 if (self != nullptr) {
619 VLOG(threads) << *self << " SuspendAll for " << cause << " starting...";
620 } else {
621 VLOG(threads) << "Thread[null] SuspendAll for " << cause << " starting...";
622 }
623 {
624 ScopedTrace trace("Suspending mutator threads");
625 const uint64_t start_time = NanoTime();
626
627 SuspendAllInternal(self, self);
628 // All threads are known to have suspended (but a thread may still own the mutator lock)
629 // Make sure this thread grabs exclusive access to the mutator lock and its protected data.
630 #if HAVE_TIMED_RWLOCK
631 while (true) {
632 if (Locks::mutator_lock_->ExclusiveLockWithTimeout(self,
633 NsToMs(thread_suspend_timeout_ns_),
634 0)) {
635 break;
636 } else if (!long_suspend_) {
637 // Reading long_suspend without the mutator lock is slightly racy, in some rare cases, this
638 // could result in a thread suspend timeout.
639 // Timeout if we wait more than thread_suspend_timeout_ns_ nanoseconds.
640 UnsafeLogFatalForThreadSuspendAllTimeout();
641 }
642 }
643 #else
644 Locks::mutator_lock_->ExclusiveLock(self);
645 #endif
646
647 long_suspend_ = long_suspend;
648
649 const uint64_t end_time = NanoTime();
650 const uint64_t suspend_time = end_time - start_time;
651 suspend_all_historam_.AdjustAndAddValue(suspend_time);
652 if (suspend_time > kLongThreadSuspendThreshold) {
653 LOG(WARNING) << "Suspending all threads took: " << PrettyDuration(suspend_time);
654 }
655
656 if (kDebugLocking) {
657 // Debug check that all threads are suspended.
658 AssertThreadsAreSuspended(self, self);
659 }
660 }
661 ATraceBegin((std::string("Mutator threads suspended for ") + cause).c_str());
662
663 if (self != nullptr) {
664 VLOG(threads) << *self << " SuspendAll complete";
665 } else {
666 VLOG(threads) << "Thread[null] SuspendAll complete";
667 }
668 }
669
670 // Ensures all threads running Java suspend and that those not running Java don't start.
SuspendAllInternal(Thread * self,Thread * ignore1,Thread * ignore2,SuspendReason reason)671 void ThreadList::SuspendAllInternal(Thread* self,
672 Thread* ignore1,
673 Thread* ignore2,
674 SuspendReason reason) {
675 Locks::mutator_lock_->AssertNotExclusiveHeld(self);
676 Locks::thread_list_lock_->AssertNotHeld(self);
677 Locks::thread_suspend_count_lock_->AssertNotHeld(self);
678 if (kDebugLocking && self != nullptr) {
679 CHECK_NE(self->GetState(), kRunnable);
680 }
681
682 // First request that all threads suspend, then wait for them to suspend before
683 // returning. This suspension scheme also relies on other behaviour:
684 // 1. Threads cannot be deleted while they are suspended or have a suspend-
685 // request flag set - (see Unregister() below).
686 // 2. When threads are created, they are created in a suspended state (actually
687 // kNative) and will never begin executing Java code without first checking
688 // the suspend-request flag.
689
690 // The atomic counter for number of threads that need to pass the barrier.
691 AtomicInteger pending_threads;
692 uint32_t num_ignored = 0;
693 if (ignore1 != nullptr) {
694 ++num_ignored;
695 }
696 if (ignore2 != nullptr && ignore1 != ignore2) {
697 ++num_ignored;
698 }
699 {
700 MutexLock mu(self, *Locks::thread_list_lock_);
701 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
702 // Update global suspend all state for attaching threads.
703 ++suspend_all_count_;
704 pending_threads.store(list_.size() - num_ignored, std::memory_order_relaxed);
705 // Increment everybody's suspend count (except those that should be ignored).
706 for (const auto& thread : list_) {
707 if (thread == ignore1 || thread == ignore2) {
708 continue;
709 }
710 VLOG(threads) << "requesting thread suspend: " << *thread;
711 bool updated = thread->ModifySuspendCount(self, +1, &pending_threads, reason);
712 DCHECK(updated);
713
714 // Must install the pending_threads counter first, then check thread->IsSuspend() and clear
715 // the counter. Otherwise there's a race with Thread::TransitionFromRunnableToSuspended()
716 // that can lead a thread to miss a call to PassActiveSuspendBarriers().
717 if (thread->IsSuspended()) {
718 // Only clear the counter for the current thread.
719 thread->ClearSuspendBarrier(&pending_threads);
720 pending_threads.fetch_sub(1, std::memory_order_seq_cst);
721 }
722 }
723 }
724
725 // Wait for the barrier to be passed by all runnable threads. This wait
726 // is done with a timeout so that we can detect problems.
727 #if ART_USE_FUTEXES
728 timespec wait_timeout;
729 InitTimeSpec(false, CLOCK_MONOTONIC, NsToMs(thread_suspend_timeout_ns_), 0, &wait_timeout);
730 #endif
731 const uint64_t start_time = NanoTime();
732 while (true) {
733 int32_t cur_val = pending_threads.load(std::memory_order_relaxed);
734 if (LIKELY(cur_val > 0)) {
735 #if ART_USE_FUTEXES
736 if (futex(pending_threads.Address(), FUTEX_WAIT_PRIVATE, cur_val, &wait_timeout, nullptr, 0)
737 != 0) {
738 if ((errno == EAGAIN) || (errno == EINTR)) {
739 // EAGAIN and EINTR both indicate a spurious failure, try again from the beginning.
740 continue;
741 }
742 if (errno == ETIMEDOUT) {
743 const uint64_t wait_time = NanoTime() - start_time;
744 MutexLock mu(self, *Locks::thread_list_lock_);
745 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
746 std::ostringstream oss;
747 for (const auto& thread : list_) {
748 if (thread == ignore1 || thread == ignore2) {
749 continue;
750 }
751 if (!thread->IsSuspended()) {
752 oss << std::endl << "Thread not suspended: " << *thread;
753 }
754 }
755 LOG(kIsDebugBuild ? ::android::base::FATAL : ::android::base::ERROR)
756 << "Timed out waiting for threads to suspend, waited for "
757 << PrettyDuration(wait_time)
758 << oss.str();
759 } else {
760 PLOG(FATAL) << "futex wait failed for SuspendAllInternal()";
761 }
762 } // else re-check pending_threads in the next iteration (this may be a spurious wake-up).
763 #else
764 // Spin wait. This is likely to be slow, but on most architecture ART_USE_FUTEXES is set.
765 UNUSED(start_time);
766 #endif
767 } else {
768 CHECK_EQ(cur_val, 0);
769 break;
770 }
771 }
772 }
773
ResumeAll()774 void ThreadList::ResumeAll() {
775 Thread* self = Thread::Current();
776
777 if (self != nullptr) {
778 VLOG(threads) << *self << " ResumeAll starting";
779 } else {
780 VLOG(threads) << "Thread[null] ResumeAll starting";
781 }
782
783 ATraceEnd();
784
785 ScopedTrace trace("Resuming mutator threads");
786
787 if (kDebugLocking) {
788 // Debug check that all threads are suspended.
789 AssertThreadsAreSuspended(self, self);
790 }
791
792 long_suspend_ = false;
793
794 Locks::mutator_lock_->ExclusiveUnlock(self);
795 {
796 MutexLock mu(self, *Locks::thread_list_lock_);
797 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
798 // Update global suspend all state for attaching threads.
799 --suspend_all_count_;
800 // Decrement the suspend counts for all threads.
801 for (const auto& thread : list_) {
802 if (thread == self) {
803 continue;
804 }
805 bool updated = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
806 DCHECK(updated);
807 }
808
809 // Broadcast a notification to all suspended threads, some or all of
810 // which may choose to wake up. No need to wait for them.
811 if (self != nullptr) {
812 VLOG(threads) << *self << " ResumeAll waking others";
813 } else {
814 VLOG(threads) << "Thread[null] ResumeAll waking others";
815 }
816 Thread::resume_cond_->Broadcast(self);
817 }
818
819 if (self != nullptr) {
820 VLOG(threads) << *self << " ResumeAll complete";
821 } else {
822 VLOG(threads) << "Thread[null] ResumeAll complete";
823 }
824 }
825
Resume(Thread * thread,SuspendReason reason)826 bool ThreadList::Resume(Thread* thread, SuspendReason reason) {
827 // This assumes there was an ATraceBegin when we suspended the thread.
828 ATraceEnd();
829
830 Thread* self = Thread::Current();
831 DCHECK_NE(thread, self);
832 VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") starting..." << reason;
833
834 {
835 // To check Contains.
836 MutexLock mu(self, *Locks::thread_list_lock_);
837 // To check IsSuspended.
838 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
839 if (UNLIKELY(!thread->IsSuspended())) {
840 LOG(ERROR) << "Resume(" << reinterpret_cast<void*>(thread)
841 << ") thread not suspended";
842 return false;
843 }
844 if (!Contains(thread)) {
845 // We only expect threads within the thread-list to have been suspended otherwise we can't
846 // stop such threads from delete-ing themselves.
847 LOG(ERROR) << "Resume(" << reinterpret_cast<void*>(thread)
848 << ") thread not within thread list";
849 return false;
850 }
851 if (UNLIKELY(!thread->ModifySuspendCount(self, -1, nullptr, reason))) {
852 LOG(ERROR) << "Resume(" << reinterpret_cast<void*>(thread)
853 << ") could not modify suspend count.";
854 return false;
855 }
856 }
857
858 {
859 VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") waking others";
860 MutexLock mu(self, *Locks::thread_suspend_count_lock_);
861 Thread::resume_cond_->Broadcast(self);
862 }
863
864 VLOG(threads) << "Resume(" << reinterpret_cast<void*>(thread) << ") complete";
865 return true;
866 }
867
ThreadSuspendByPeerWarning(Thread * self,LogSeverity severity,const char * message,jobject peer)868 static void ThreadSuspendByPeerWarning(Thread* self,
869 LogSeverity severity,
870 const char* message,
871 jobject peer) {
872 JNIEnvExt* env = self->GetJniEnv();
873 ScopedLocalRef<jstring>
874 scoped_name_string(env, static_cast<jstring>(env->GetObjectField(
875 peer, WellKnownClasses::java_lang_Thread_name)));
876 ScopedUtfChars scoped_name_chars(env, scoped_name_string.get());
877 if (scoped_name_chars.c_str() == nullptr) {
878 LOG(severity) << message << ": " << peer;
879 env->ExceptionClear();
880 } else {
881 LOG(severity) << message << ": " << peer << ":" << scoped_name_chars.c_str();
882 }
883 }
884
SuspendThreadByPeer(jobject peer,bool request_suspension,SuspendReason reason,bool * timed_out)885 Thread* ThreadList::SuspendThreadByPeer(jobject peer,
886 bool request_suspension,
887 SuspendReason reason,
888 bool* timed_out) {
889 const uint64_t start_time = NanoTime();
890 useconds_t sleep_us = kThreadSuspendInitialSleepUs;
891 *timed_out = false;
892 Thread* const self = Thread::Current();
893 Thread* suspended_thread = nullptr;
894 VLOG(threads) << "SuspendThreadByPeer starting";
895 while (true) {
896 Thread* thread;
897 {
898 // Note: this will transition to runnable and potentially suspend. We ensure only one thread
899 // is requesting another suspend, to avoid deadlock, by requiring this function be called
900 // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
901 // than request thread suspension, to avoid potential cycles in threads requesting each other
902 // suspend.
903 ScopedObjectAccess soa(self);
904 MutexLock thread_list_mu(self, *Locks::thread_list_lock_);
905 thread = Thread::FromManagedThread(soa, peer);
906 if (thread == nullptr) {
907 if (suspended_thread != nullptr) {
908 MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
909 // If we incremented the suspend count but the thread reset its peer, we need to
910 // re-decrement it since it is shutting down and may deadlock the runtime in
911 // ThreadList::WaitForOtherNonDaemonThreadsToExit.
912 bool updated = suspended_thread->ModifySuspendCount(soa.Self(),
913 -1,
914 nullptr,
915 reason);
916 DCHECK(updated);
917 }
918 ThreadSuspendByPeerWarning(self,
919 ::android::base::WARNING,
920 "No such thread for suspend",
921 peer);
922 return nullptr;
923 }
924 if (!Contains(thread)) {
925 CHECK(suspended_thread == nullptr);
926 VLOG(threads) << "SuspendThreadByPeer failed for unattached thread: "
927 << reinterpret_cast<void*>(thread);
928 return nullptr;
929 }
930 VLOG(threads) << "SuspendThreadByPeer found thread: " << *thread;
931 {
932 MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
933 if (request_suspension) {
934 if (self->GetSuspendCount() > 0) {
935 // We hold the suspend count lock but another thread is trying to suspend us. Its not
936 // safe to try to suspend another thread in case we get a cycle. Start the loop again
937 // which will allow this thread to be suspended.
938 continue;
939 }
940 CHECK(suspended_thread == nullptr);
941 suspended_thread = thread;
942 bool updated = suspended_thread->ModifySuspendCount(self, +1, nullptr, reason);
943 DCHECK(updated);
944 request_suspension = false;
945 } else {
946 // If the caller isn't requesting suspension, a suspension should have already occurred.
947 CHECK_GT(thread->GetSuspendCount(), 0);
948 }
949 // IsSuspended on the current thread will fail as the current thread is changed into
950 // Runnable above. As the suspend count is now raised if this is the current thread
951 // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
952 // to just explicitly handle the current thread in the callers to this code.
953 CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
954 // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
955 // count, or else we've waited and it has self suspended) or is the current thread, we're
956 // done.
957 if (thread->IsSuspended()) {
958 VLOG(threads) << "SuspendThreadByPeer thread suspended: " << *thread;
959 if (ATraceEnabled()) {
960 std::string name;
961 thread->GetThreadName(name);
962 ATraceBegin(StringPrintf("SuspendThreadByPeer suspended %s for peer=%p", name.c_str(),
963 peer).c_str());
964 }
965 return thread;
966 }
967 const uint64_t total_delay = NanoTime() - start_time;
968 if (total_delay >= thread_suspend_timeout_ns_) {
969 ThreadSuspendByPeerWarning(self,
970 ::android::base::FATAL,
971 "Thread suspension timed out",
972 peer);
973 if (suspended_thread != nullptr) {
974 CHECK_EQ(suspended_thread, thread);
975 bool updated = suspended_thread->ModifySuspendCount(soa.Self(),
976 -1,
977 nullptr,
978 reason);
979 DCHECK(updated);
980 }
981 *timed_out = true;
982 return nullptr;
983 } else if (sleep_us == 0 &&
984 total_delay > static_cast<uint64_t>(kThreadSuspendMaxYieldUs) * 1000) {
985 // We have spun for kThreadSuspendMaxYieldUs time, switch to sleeps to prevent
986 // excessive CPU usage.
987 sleep_us = kThreadSuspendMaxYieldUs / 2;
988 }
989 }
990 // Release locks and come out of runnable state.
991 }
992 VLOG(threads) << "SuspendThreadByPeer waiting to allow thread chance to suspend";
993 ThreadSuspendSleep(sleep_us);
994 // This may stay at 0 if sleep_us == 0, but this is WAI since we want to avoid using usleep at
995 // all if possible. This shouldn't be an issue since time to suspend should always be small.
996 sleep_us = std::min(sleep_us * 2, kThreadSuspendMaxSleepUs);
997 }
998 }
999
ThreadSuspendByThreadIdWarning(LogSeverity severity,const char * message,uint32_t thread_id)1000 static void ThreadSuspendByThreadIdWarning(LogSeverity severity,
1001 const char* message,
1002 uint32_t thread_id) {
1003 LOG(severity) << StringPrintf("%s: %d", message, thread_id);
1004 }
1005
SuspendThreadByThreadId(uint32_t thread_id,SuspendReason reason,bool * timed_out)1006 Thread* ThreadList::SuspendThreadByThreadId(uint32_t thread_id,
1007 SuspendReason reason,
1008 bool* timed_out) {
1009 const uint64_t start_time = NanoTime();
1010 useconds_t sleep_us = kThreadSuspendInitialSleepUs;
1011 *timed_out = false;
1012 Thread* suspended_thread = nullptr;
1013 Thread* const self = Thread::Current();
1014 CHECK_NE(thread_id, kInvalidThreadId);
1015 VLOG(threads) << "SuspendThreadByThreadId starting";
1016 while (true) {
1017 {
1018 // Note: this will transition to runnable and potentially suspend. We ensure only one thread
1019 // is requesting another suspend, to avoid deadlock, by requiring this function be called
1020 // holding Locks::thread_list_suspend_thread_lock_. Its important this thread suspend rather
1021 // than request thread suspension, to avoid potential cycles in threads requesting each other
1022 // suspend.
1023 ScopedObjectAccess soa(self);
1024 MutexLock thread_list_mu(self, *Locks::thread_list_lock_);
1025 Thread* thread = nullptr;
1026 for (const auto& it : list_) {
1027 if (it->GetThreadId() == thread_id) {
1028 thread = it;
1029 break;
1030 }
1031 }
1032 if (thread == nullptr) {
1033 CHECK(suspended_thread == nullptr) << "Suspended thread " << suspended_thread
1034 << " no longer in thread list";
1035 // There's a race in inflating a lock and the owner giving up ownership and then dying.
1036 ThreadSuspendByThreadIdWarning(::android::base::WARNING,
1037 "No such thread id for suspend",
1038 thread_id);
1039 return nullptr;
1040 }
1041 VLOG(threads) << "SuspendThreadByThreadId found thread: " << *thread;
1042 DCHECK(Contains(thread));
1043 {
1044 MutexLock suspend_count_mu(self, *Locks::thread_suspend_count_lock_);
1045 if (suspended_thread == nullptr) {
1046 if (self->GetSuspendCount() > 0) {
1047 // We hold the suspend count lock but another thread is trying to suspend us. Its not
1048 // safe to try to suspend another thread in case we get a cycle. Start the loop again
1049 // which will allow this thread to be suspended.
1050 continue;
1051 }
1052 bool updated = thread->ModifySuspendCount(self, +1, nullptr, reason);
1053 DCHECK(updated);
1054 suspended_thread = thread;
1055 } else {
1056 CHECK_EQ(suspended_thread, thread);
1057 // If the caller isn't requesting suspension, a suspension should have already occurred.
1058 CHECK_GT(thread->GetSuspendCount(), 0);
1059 }
1060 // IsSuspended on the current thread will fail as the current thread is changed into
1061 // Runnable above. As the suspend count is now raised if this is the current thread
1062 // it will self suspend on transition to Runnable, making it hard to work with. It's simpler
1063 // to just explicitly handle the current thread in the callers to this code.
1064 CHECK_NE(thread, self) << "Attempt to suspend the current thread for the debugger";
1065 // If thread is suspended (perhaps it was already not Runnable but didn't have a suspend
1066 // count, or else we've waited and it has self suspended) or is the current thread, we're
1067 // done.
1068 if (thread->IsSuspended()) {
1069 if (ATraceEnabled()) {
1070 std::string name;
1071 thread->GetThreadName(name);
1072 ATraceBegin(StringPrintf("SuspendThreadByThreadId suspended %s id=%d",
1073 name.c_str(), thread_id).c_str());
1074 }
1075 VLOG(threads) << "SuspendThreadByThreadId thread suspended: " << *thread;
1076 return thread;
1077 }
1078 const uint64_t total_delay = NanoTime() - start_time;
1079 if (total_delay >= thread_suspend_timeout_ns_) {
1080 ThreadSuspendByThreadIdWarning(::android::base::WARNING,
1081 "Thread suspension timed out",
1082 thread_id);
1083 if (suspended_thread != nullptr) {
1084 bool updated = thread->ModifySuspendCount(soa.Self(), -1, nullptr, reason);
1085 DCHECK(updated);
1086 }
1087 *timed_out = true;
1088 return nullptr;
1089 } else if (sleep_us == 0 &&
1090 total_delay > static_cast<uint64_t>(kThreadSuspendMaxYieldUs) * 1000) {
1091 // We have spun for kThreadSuspendMaxYieldUs time, switch to sleeps to prevent
1092 // excessive CPU usage.
1093 sleep_us = kThreadSuspendMaxYieldUs / 2;
1094 }
1095 }
1096 // Release locks and come out of runnable state.
1097 }
1098 VLOG(threads) << "SuspendThreadByThreadId waiting to allow thread chance to suspend";
1099 ThreadSuspendSleep(sleep_us);
1100 sleep_us = std::min(sleep_us * 2, kThreadSuspendMaxSleepUs);
1101 }
1102 }
1103
FindThreadByThreadId(uint32_t thread_id)1104 Thread* ThreadList::FindThreadByThreadId(uint32_t thread_id) {
1105 for (const auto& thread : list_) {
1106 if (thread->GetThreadId() == thread_id) {
1107 return thread;
1108 }
1109 }
1110 return nullptr;
1111 }
1112
WaitForOtherNonDaemonThreadsToExit(bool check_no_birth)1113 void ThreadList::WaitForOtherNonDaemonThreadsToExit(bool check_no_birth) {
1114 ScopedTrace trace(__PRETTY_FUNCTION__);
1115 Thread* self = Thread::Current();
1116 Locks::mutator_lock_->AssertNotHeld(self);
1117 while (true) {
1118 Locks::runtime_shutdown_lock_->Lock(self);
1119 if (check_no_birth) {
1120 // No more threads can be born after we start to shutdown.
1121 CHECK(Runtime::Current()->IsShuttingDownLocked());
1122 CHECK_EQ(Runtime::Current()->NumberOfThreadsBeingBorn(), 0U);
1123 } else {
1124 if (Runtime::Current()->NumberOfThreadsBeingBorn() != 0U) {
1125 // Awkward. Shutdown_cond_ is private, but the only live thread may not be registered yet.
1126 // Fortunately, this is used mostly for testing, and not performance-critical.
1127 Locks::runtime_shutdown_lock_->Unlock(self);
1128 usleep(1000);
1129 continue;
1130 }
1131 }
1132 MutexLock mu(self, *Locks::thread_list_lock_);
1133 Locks::runtime_shutdown_lock_->Unlock(self);
1134 // Also wait for any threads that are unregistering to finish. This is required so that no
1135 // threads access the thread list after it is deleted. TODO: This may not work for user daemon
1136 // threads since they could unregister at the wrong time.
1137 bool done = unregistering_count_ == 0;
1138 if (done) {
1139 for (const auto& thread : list_) {
1140 if (thread != self && !thread->IsDaemon()) {
1141 done = false;
1142 break;
1143 }
1144 }
1145 }
1146 if (done) {
1147 break;
1148 }
1149 // Wait for another thread to exit before re-checking.
1150 Locks::thread_exit_cond_->Wait(self);
1151 }
1152 }
1153
SuspendAllDaemonThreadsForShutdown()1154 void ThreadList::SuspendAllDaemonThreadsForShutdown() {
1155 ScopedTrace trace(__PRETTY_FUNCTION__);
1156 Thread* self = Thread::Current();
1157 size_t daemons_left = 0;
1158 {
1159 // Tell all the daemons it's time to suspend.
1160 MutexLock mu(self, *Locks::thread_list_lock_);
1161 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1162 for (const auto& thread : list_) {
1163 // This is only run after all non-daemon threads have exited, so the remainder should all be
1164 // daemons.
1165 CHECK(thread->IsDaemon()) << *thread;
1166 if (thread != self) {
1167 bool updated = thread->ModifySuspendCount(self, +1, nullptr, SuspendReason::kInternal);
1168 DCHECK(updated);
1169 ++daemons_left;
1170 }
1171 // We are shutting down the runtime, set the JNI functions of all the JNIEnvs to be
1172 // the sleep forever one.
1173 thread->GetJniEnv()->SetFunctionsToRuntimeShutdownFunctions();
1174 }
1175 }
1176 if (daemons_left == 0) {
1177 // No threads left; safe to shut down.
1178 return;
1179 }
1180 // There is not a clean way to shut down if we have daemons left. We have no mechanism for
1181 // killing them and reclaiming thread stacks. We also have no mechanism for waiting until they
1182 // have truly finished touching the memory we are about to deallocate. We do the best we can with
1183 // timeouts.
1184 //
1185 // If we have any daemons left, wait until they are (a) suspended and (b) they are not stuck
1186 // in a place where they are about to access runtime state and are not in a runnable state.
1187 // We attempt to do the latter by just waiting long enough for things to
1188 // quiesce. Examples: Monitor code or waking up from a condition variable.
1189 //
1190 // Give the threads a chance to suspend, complaining if they're slow. (a)
1191 bool have_complained = false;
1192 static constexpr size_t kTimeoutMicroseconds = 2000 * 1000;
1193 static constexpr size_t kSleepMicroseconds = 1000;
1194 bool all_suspended = false;
1195 for (size_t i = 0; !all_suspended && i < kTimeoutMicroseconds / kSleepMicroseconds; ++i) {
1196 bool found_running = false;
1197 {
1198 MutexLock mu(self, *Locks::thread_list_lock_);
1199 for (const auto& thread : list_) {
1200 if (thread != self && thread->GetState() == kRunnable) {
1201 if (!have_complained) {
1202 LOG(WARNING) << "daemon thread not yet suspended: " << *thread;
1203 have_complained = true;
1204 }
1205 found_running = true;
1206 }
1207 }
1208 }
1209 if (found_running) {
1210 // Sleep briefly before checking again. Max total sleep time is kTimeoutMicroseconds.
1211 usleep(kSleepMicroseconds);
1212 } else {
1213 all_suspended = true;
1214 }
1215 }
1216 if (!all_suspended) {
1217 // We can get here if a daemon thread executed a fastnative native call, so that it
1218 // remained in runnable state, and then made a JNI call after we called
1219 // SetFunctionsToRuntimeShutdownFunctions(), causing it to permanently stay in a harmless
1220 // but runnable state. See b/147804269 .
1221 LOG(WARNING) << "timed out suspending all daemon threads";
1222 }
1223 // Assume all threads are either suspended or somehow wedged.
1224 // Wait again for all the now "suspended" threads to actually quiesce. (b)
1225 static constexpr size_t kDaemonSleepTime = 400'000;
1226 usleep(kDaemonSleepTime);
1227 std::list<Thread*> list_copy;
1228 {
1229 MutexLock mu(self, *Locks::thread_list_lock_);
1230 // Half-way through the wait, set the "runtime deleted" flag, causing any newly awoken
1231 // threads to immediately go back to sleep without touching memory. This prevents us from
1232 // touching deallocated memory, but it also prevents mutexes from getting released. Thus we
1233 // only do this once we're reasonably sure that no system mutexes are still held.
1234 for (const auto& thread : list_) {
1235 DCHECK(thread == self || !all_suspended || thread->GetState() != kRunnable);
1236 // In the !all_suspended case, the target is probably sleeping.
1237 thread->GetJniEnv()->SetRuntimeDeleted();
1238 // Possibly contended Mutex acquisitions are unsafe after this.
1239 // Releasing thread_list_lock_ is OK, since it can't block.
1240 }
1241 }
1242 // Finally wait for any threads woken before we set the "runtime deleted" flags to finish
1243 // touching memory.
1244 usleep(kDaemonSleepTime);
1245 #if defined(__has_feature)
1246 #if __has_feature(address_sanitizer) || __has_feature(hwaddress_sanitizer)
1247 // Sleep a bit longer with -fsanitize=address, since everything is slower.
1248 usleep(2 * kDaemonSleepTime);
1249 #endif
1250 #endif
1251 // At this point no threads should be touching our data structures anymore.
1252 }
1253
Register(Thread * self)1254 void ThreadList::Register(Thread* self) {
1255 DCHECK_EQ(self, Thread::Current());
1256 CHECK(!shut_down_);
1257
1258 if (VLOG_IS_ON(threads)) {
1259 std::ostringstream oss;
1260 self->ShortDump(oss); // We don't hold the mutator_lock_ yet and so cannot call Dump.
1261 LOG(INFO) << "ThreadList::Register() " << *self << "\n" << oss.str();
1262 }
1263
1264 // Atomically add self to the thread list and make its thread_suspend_count_ reflect ongoing
1265 // SuspendAll requests.
1266 MutexLock mu(self, *Locks::thread_list_lock_);
1267 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1268 // Modify suspend count in increments of 1 to maintain invariants in ModifySuspendCount. While
1269 // this isn't particularly efficient the suspend counts are most commonly 0 or 1.
1270 for (int delta = suspend_all_count_; delta > 0; delta--) {
1271 bool updated = self->ModifySuspendCount(self, +1, nullptr, SuspendReason::kInternal);
1272 DCHECK(updated);
1273 }
1274 CHECK(!Contains(self));
1275 list_.push_back(self);
1276 if (kUseReadBarrier) {
1277 gc::collector::ConcurrentCopying* const cc =
1278 Runtime::Current()->GetHeap()->ConcurrentCopyingCollector();
1279 // Initialize according to the state of the CC collector.
1280 self->SetIsGcMarkingAndUpdateEntrypoints(cc->IsMarking());
1281 if (cc->IsUsingReadBarrierEntrypoints()) {
1282 self->SetReadBarrierEntrypoints();
1283 }
1284 self->SetWeakRefAccessEnabled(cc->IsWeakRefAccessEnabled());
1285 }
1286 self->NotifyInTheadList();
1287 }
1288
Unregister(Thread * self)1289 void ThreadList::Unregister(Thread* self) {
1290 DCHECK_EQ(self, Thread::Current());
1291 CHECK_NE(self->GetState(), kRunnable);
1292 Locks::mutator_lock_->AssertNotHeld(self);
1293
1294 VLOG(threads) << "ThreadList::Unregister() " << *self;
1295
1296 {
1297 MutexLock mu(self, *Locks::thread_list_lock_);
1298 ++unregistering_count_;
1299 }
1300
1301 // Any time-consuming destruction, plus anything that can call back into managed code or
1302 // suspend and so on, must happen at this point, and not in ~Thread. The self->Destroy is what
1303 // causes the threads to join. It is important to do this after incrementing unregistering_count_
1304 // since we want the runtime to wait for the daemon threads to exit before deleting the thread
1305 // list.
1306 self->Destroy();
1307
1308 // If tracing, remember thread id and name before thread exits.
1309 Trace::StoreExitingThreadInfo(self);
1310
1311 uint32_t thin_lock_id = self->GetThreadId();
1312 while (true) {
1313 // Remove and delete the Thread* while holding the thread_list_lock_ and
1314 // thread_suspend_count_lock_ so that the unregistering thread cannot be suspended.
1315 // Note: deliberately not using MutexLock that could hold a stale self pointer.
1316 {
1317 MutexLock mu(self, *Locks::thread_list_lock_);
1318 if (!Contains(self)) {
1319 std::string thread_name;
1320 self->GetThreadName(thread_name);
1321 std::ostringstream os;
1322 DumpNativeStack(os, GetTid(), nullptr, " native: ", nullptr);
1323 LOG(ERROR) << "Request to unregister unattached thread " << thread_name << "\n" << os.str();
1324 break;
1325 } else {
1326 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1327 if (!self->IsSuspended()) {
1328 list_.remove(self);
1329 break;
1330 }
1331 }
1332 }
1333 // In the case where we are not suspended yet, sleep to leave other threads time to execute.
1334 // This is important if there are realtime threads. b/111277984
1335 usleep(1);
1336 // We failed to remove the thread due to a suspend request, loop and try again.
1337 }
1338 delete self;
1339
1340 // Release the thread ID after the thread is finished and deleted to avoid cases where we can
1341 // temporarily have multiple threads with the same thread id. When this occurs, it causes
1342 // problems in FindThreadByThreadId / SuspendThreadByThreadId.
1343 ReleaseThreadId(nullptr, thin_lock_id);
1344
1345 // Clear the TLS data, so that the underlying native thread is recognizably detached.
1346 // (It may wish to reattach later.)
1347 #ifdef __BIONIC__
1348 __get_tls()[TLS_SLOT_ART_THREAD_SELF] = nullptr;
1349 #else
1350 CHECK_PTHREAD_CALL(pthread_setspecific, (Thread::pthread_key_self_, nullptr), "detach self");
1351 Thread::self_tls_ = nullptr;
1352 #endif
1353
1354 // Signal that a thread just detached.
1355 MutexLock mu(nullptr, *Locks::thread_list_lock_);
1356 --unregistering_count_;
1357 Locks::thread_exit_cond_->Broadcast(nullptr);
1358 }
1359
ForEach(void (* callback)(Thread *,void *),void * context)1360 void ThreadList::ForEach(void (*callback)(Thread*, void*), void* context) {
1361 for (const auto& thread : list_) {
1362 callback(thread, context);
1363 }
1364 }
1365
VisitRootsForSuspendedThreads(RootVisitor * visitor)1366 void ThreadList::VisitRootsForSuspendedThreads(RootVisitor* visitor) {
1367 Thread* const self = Thread::Current();
1368 std::vector<Thread*> threads_to_visit;
1369
1370 // Tell threads to suspend and copy them into list.
1371 {
1372 MutexLock mu(self, *Locks::thread_list_lock_);
1373 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1374 for (Thread* thread : list_) {
1375 bool suspended = thread->ModifySuspendCount(self, +1, nullptr, SuspendReason::kInternal);
1376 DCHECK(suspended);
1377 if (thread == self || thread->IsSuspended()) {
1378 threads_to_visit.push_back(thread);
1379 } else {
1380 bool resumed = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
1381 DCHECK(resumed);
1382 }
1383 }
1384 }
1385
1386 // Visit roots without holding thread_list_lock_ and thread_suspend_count_lock_ to prevent lock
1387 // order violations.
1388 for (Thread* thread : threads_to_visit) {
1389 thread->VisitRoots(visitor, kVisitRootFlagAllRoots);
1390 }
1391
1392 // Restore suspend counts.
1393 {
1394 MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
1395 for (Thread* thread : threads_to_visit) {
1396 bool updated = thread->ModifySuspendCount(self, -1, nullptr, SuspendReason::kInternal);
1397 DCHECK(updated);
1398 }
1399 }
1400 }
1401
VisitRoots(RootVisitor * visitor,VisitRootFlags flags) const1402 void ThreadList::VisitRoots(RootVisitor* visitor, VisitRootFlags flags) const {
1403 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
1404 for (const auto& thread : list_) {
1405 thread->VisitRoots(visitor, flags);
1406 }
1407 }
1408
SweepInterpreterCaches(IsMarkedVisitor * visitor) const1409 void ThreadList::SweepInterpreterCaches(IsMarkedVisitor* visitor) const {
1410 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
1411 for (const auto& thread : list_) {
1412 thread->SweepInterpreterCache(visitor);
1413 }
1414 }
1415
VisitReflectiveTargets(ReflectiveValueVisitor * visitor) const1416 void ThreadList::VisitReflectiveTargets(ReflectiveValueVisitor *visitor) const {
1417 MutexLock mu(Thread::Current(), *Locks::thread_list_lock_);
1418 for (const auto& thread : list_) {
1419 thread->VisitReflectiveTargets(visitor);
1420 }
1421 }
1422
AllocThreadId(Thread * self)1423 uint32_t ThreadList::AllocThreadId(Thread* self) {
1424 MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
1425 for (size_t i = 0; i < allocated_ids_.size(); ++i) {
1426 if (!allocated_ids_[i]) {
1427 allocated_ids_.set(i);
1428 return i + 1; // Zero is reserved to mean "invalid".
1429 }
1430 }
1431 LOG(FATAL) << "Out of internal thread ids";
1432 UNREACHABLE();
1433 }
1434
ReleaseThreadId(Thread * self,uint32_t id)1435 void ThreadList::ReleaseThreadId(Thread* self, uint32_t id) {
1436 MutexLock mu(self, *Locks::allocated_thread_ids_lock_);
1437 --id; // Zero is reserved to mean "invalid".
1438 DCHECK(allocated_ids_[id]) << id;
1439 allocated_ids_.reset(id);
1440 }
1441
ScopedSuspendAll(const char * cause,bool long_suspend)1442 ScopedSuspendAll::ScopedSuspendAll(const char* cause, bool long_suspend) {
1443 Runtime::Current()->GetThreadList()->SuspendAll(cause, long_suspend);
1444 }
1445
~ScopedSuspendAll()1446 ScopedSuspendAll::~ScopedSuspendAll() {
1447 Runtime::Current()->GetThreadList()->ResumeAll();
1448 }
1449
1450 } // namespace art
1451