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 #ifndef ART_RUNTIME_THREAD_LIST_H_ 18 #define ART_RUNTIME_THREAD_LIST_H_ 19 20 #include <bitset> 21 #include <list> 22 #include <vector> 23 24 #include "barrier.h" 25 #include "base/histogram.h" 26 #include "base/mutex.h" 27 #include "base/macros.h" 28 #include "base/value_object.h" 29 #include "jni.h" 30 #include "reflective_handle_scope.h" 31 #include "suspend_reason.h" 32 #include "thread_state.h" 33 34 namespace art HIDDEN { 35 namespace gc { 36 namespace collector { 37 class GarbageCollector; 38 } // namespace collector 39 class GcPauseListener; 40 } // namespace gc 41 class Closure; 42 class IsMarkedVisitor; 43 class RootVisitor; 44 class Thread; 45 class TimingLogger; 46 enum VisitRootFlags : uint8_t; 47 48 class ThreadList { 49 public: 50 static constexpr uint32_t kMaxThreadId = 0xFFFF; 51 static constexpr uint32_t kInvalidThreadId = 0; 52 static constexpr uint32_t kMainThreadId = 1; 53 static constexpr uint64_t kDefaultThreadSuspendTimeout = 54 kIsDebugBuild ? 2'000'000'000ull : 4'000'000'000ull; 55 // We fail more aggressively in debug builds to catch potential issues early. 56 // The number of times we may retry when we find ourselves in a suspend-unfriendly state. 57 static constexpr int kMaxSuspendRetries = kIsDebugBuild ? 500 : 5000; 58 static constexpr useconds_t kThreadSuspendSleepUs = 100; 59 60 explicit ThreadList(uint64_t thread_suspend_timeout_ns); 61 ~ThreadList(); 62 63 void ShutDown(); 64 65 void DumpForSigQuit(std::ostream& os) 66 REQUIRES(!Locks::thread_list_lock_, !Locks::mutator_lock_); 67 // For thread suspend timeout dumps. 68 EXPORT void Dump(std::ostream& os, bool dump_native_stack = true) 69 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 70 pid_t GetLockOwner(); // For SignalCatcher. 71 72 // Thread suspension support. 73 EXPORT void ResumeAll() 74 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_) 75 UNLOCK_FUNCTION(Locks::mutator_lock_); 76 EXPORT bool Resume(Thread* thread, SuspendReason reason = SuspendReason::kInternal) 77 REQUIRES(!Locks::thread_suspend_count_lock_) WARN_UNUSED; 78 79 // Suspends all other threads and gets exclusive access to the mutator lock. 80 // If long_suspend is true, then other threads who try to suspend will never timeout. 81 // long_suspend is currenly used for hprof since large heaps take a long time. 82 EXPORT void SuspendAll(const char* cause, bool long_suspend = false) 83 EXCLUSIVE_LOCK_FUNCTION(Locks::mutator_lock_) 84 REQUIRES(!Locks::thread_list_lock_, 85 !Locks::thread_suspend_count_lock_, 86 !Locks::mutator_lock_); 87 88 // Suspend a thread using a peer, typically used by the debugger. Returns the thread on success, 89 // else null. The peer is used to identify the thread to avoid races with the thread terminating. 90 EXPORT Thread* SuspendThreadByPeer(jobject peer, SuspendReason reason) 91 REQUIRES(!Locks::mutator_lock_, 92 !Locks::thread_list_lock_, 93 !Locks::thread_suspend_count_lock_); 94 95 // Suspend a thread using its thread id, typically used by lock/monitor inflation. Returns the 96 // thread on success else null. The thread id is used to identify the thread to avoid races with 97 // the thread terminating. Note that as thread ids are recycled this may not suspend the expected 98 // thread, that may be terminating. 'attempt_of_4' is zero if this is the only 99 // attempt, or 1..4 to try 4 times with fractional timeouts. 100 // TODO: Reconsider the use of thread_id, now that we have ThreadExitFlag. 101 Thread* SuspendThreadByThreadId(uint32_t thread_id, SuspendReason reason, int attempt_of_4 = 0) 102 REQUIRES(!Locks::mutator_lock_, 103 !Locks::thread_list_lock_, 104 !Locks::thread_suspend_count_lock_); 105 106 // Find an existing thread (or self) by its thread id (not tid). 107 EXPORT Thread* FindThreadByThreadId(uint32_t thread_id) REQUIRES(Locks::thread_list_lock_); 108 109 // Find an existing thread (or self) by its tid (not thread id). 110 Thread* FindThreadByTid(int tid) REQUIRES(Locks::thread_list_lock_); 111 112 // Does the thread list still contain the given thread, or one at the same address? 113 // Used by Monitor to provide (mostly accurate) debugging information. 114 bool Contains(Thread* thread) REQUIRES(Locks::thread_list_lock_); 115 116 // Run a checkpoint on all threads. Return the total number of threads for which the checkpoint 117 // function has been or will be called. 118 // Running threads are not suspended but run the checkpoint inside of the suspend check. The 119 // return value includes already suspended threads for b/24191051. Runs or requests the 120 // callback, if non-null, inside the thread_list_lock critical section after determining the 121 // runnable/suspended states of the threads. Does not wait for completion of the checkpoint 122 // function in running threads. If the caller holds the mutator lock, then all instances of the 123 // checkpoint function are run with the mutator lock. If the caller does not hold the mutator 124 // lock (see mutator_gc_coord.md) then, since the checkpoint code may not acquire or release the 125 // mutator lock, the checkpoint will have no way to access Java data. 126 // TODO: Is it possible to just require the mutator lock here? 127 EXPORT size_t RunCheckpoint(Closure* checkpoint_function, 128 Closure* callback = nullptr, 129 bool allow_lock_checking = true) 130 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 131 132 // Convenience version of the above to disable lock checking inside Run function. Hopefully this 133 // and the third parameter above will eventually disappear. 134 size_t RunCheckpointUnchecked(Closure* checkpoint_function, Closure* callback = nullptr) 135 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_) { 136 return RunCheckpoint(checkpoint_function, callback, false); 137 } 138 139 // Run an empty checkpoint on threads. Wait until threads pass the next suspend point or are 140 // suspended. This is used to ensure that the threads finish or aren't in the middle of an 141 // in-flight mutator heap access (eg. a read barrier.) Runnable threads will respond by 142 // decrementing the empty checkpoint barrier count. This works even when the weak ref access is 143 // disabled. Only one concurrent use is currently supported. 144 void RunEmptyCheckpoint() 145 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 146 147 // Used to flip thread roots from from-space refs to to-space refs. Used only by the concurrent 148 // moving collectors during a GC, and hence cannot be called from multiple threads concurrently. 149 // 150 // Briefly suspends all threads to atomically install a checkpoint-like thread_flip_visitor 151 // function to be run on each thread. Run flip_callback while threads are suspended. 152 // Thread_flip_visitors are run by each thread before it becomes runnable, or by us. We do not 153 // return until all thread_flip_visitors have been run. 154 void FlipThreadRoots(Closure* thread_flip_visitor, 155 Closure* flip_callback, 156 gc::collector::GarbageCollector* collector, 157 gc::GcPauseListener* pause_listener) 158 REQUIRES(!Locks::mutator_lock_, 159 !Locks::thread_list_lock_, 160 !Locks::thread_suspend_count_lock_); 161 162 // Iterates over all the threads. 163 EXPORT void ForEach(void (*callback)(Thread*, void*), void* context) 164 REQUIRES(Locks::thread_list_lock_); 165 166 template<typename CallBack> ForEach(CallBack cb)167 void ForEach(CallBack cb) REQUIRES(Locks::thread_list_lock_) { 168 ForEach([](Thread* t, void* ctx) REQUIRES(Locks::thread_list_lock_) { 169 (*reinterpret_cast<CallBack*>(ctx))(t); 170 }, &cb); 171 } 172 173 // Add/remove current thread from list. 174 void Register(Thread* self) 175 REQUIRES(Locks::runtime_shutdown_lock_) 176 REQUIRES(!Locks::mutator_lock_, 177 !Locks::thread_list_lock_, 178 !Locks::thread_suspend_count_lock_); 179 void Unregister(Thread* self, bool should_run_callbacks) 180 REQUIRES(!Locks::mutator_lock_, 181 !Locks::thread_list_lock_, 182 !Locks::thread_suspend_count_lock_); 183 184 // Wait until there are no Unregister() requests in flight. Only makes sense when we know that 185 // no new calls can be made. e.g. because we're the last thread. 186 void WaitForUnregisterToComplete(Thread* self) REQUIRES(Locks::thread_list_lock_); 187 188 void VisitRoots(RootVisitor* visitor, VisitRootFlags flags) const 189 REQUIRES_SHARED(Locks::mutator_lock_); 190 191 void VisitRootsForSuspendedThreads(RootVisitor* visitor) 192 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_) 193 REQUIRES_SHARED(Locks::mutator_lock_); 194 195 void VisitReflectiveTargets(ReflectiveValueVisitor* visitor) const REQUIRES(Locks::mutator_lock_); 196 197 EXPORT void SweepInterpreterCaches(IsMarkedVisitor* visitor) const 198 REQUIRES(Locks::mutator_lock_, !Locks::thread_list_lock_); 199 200 // Return a copy of the thread list. GetList()201 std::list<Thread*> GetList() REQUIRES(Locks::thread_list_lock_) { 202 return list_; 203 } 204 Size()205 size_t Size() REQUIRES(Locks::thread_list_lock_) { return list_.size(); } 206 CheckOnly1Thread(Thread * self)207 void CheckOnly1Thread(Thread* self) REQUIRES(!Locks::thread_list_lock_) { 208 MutexLock mu(self, *Locks::thread_list_lock_); 209 CHECK_EQ(Size(), 1u); 210 } 211 212 void DumpNativeStacks(std::ostream& os) 213 REQUIRES(!Locks::thread_list_lock_); 214 EmptyCheckpointBarrier()215 Barrier* EmptyCheckpointBarrier() { 216 return empty_checkpoint_barrier_.get(); 217 } 218 219 void WaitForOtherNonDaemonThreadsToExit(bool check_no_birth = true) 220 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_, 221 !Locks::mutator_lock_); 222 223 // Wait for suspend barrier to reach zero. Return a string possibly containing diagnostic 224 // information on timeout, nothing on success. The argument t specifies a thread to monitor for 225 // the diagnostic information. If 0 is passed, we return an empty string on timeout. Normally 226 // the caller does not hold the mutator lock. See the comment at the call in 227 // RequestSynchronousCheckpoint for the only exception. 228 std::optional<std::string> WaitForSuspendBarrier(AtomicInteger* barrier, 229 pid_t t = 0, 230 int attempt_of_4 = 0) 231 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 232 233 private: 234 uint32_t AllocThreadId(Thread* self); 235 void ReleaseThreadId(Thread* self, uint32_t id) REQUIRES(!Locks::allocated_thread_ids_lock_); 236 237 void DumpUnattachedThreads(std::ostream& os, bool dump_native_stack) 238 REQUIRES(!Locks::thread_list_lock_); 239 240 void SuspendAllDaemonThreadsForShutdown() 241 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 242 243 void ResumeAllInternal(Thread* self) 244 REQUIRES(Locks::thread_list_lock_, Locks::thread_suspend_count_lock_) 245 UNLOCK_FUNCTION(Locks::mutator_lock_); 246 247 // Helper to actually suspend a single thread. This is called with thread_list_lock_ held and 248 // the caller guarantees that *thread is valid until that is released. We "release the mutator 249 // lock", by switching to self_state. 'attempt_of_4' is 0 if we only attempt once, and 1..4 if 250 // we are going to try 4 times with a quarter of the full timeout. 'func_name' is used only to 251 // identify ourselves for logging. 252 bool SuspendThread(Thread* self, 253 Thread* thread, 254 SuspendReason reason, 255 ThreadState self_state, 256 const char* func_name, 257 int attempt_of_4) RELEASE(Locks::thread_list_lock_) 258 RELEASE_SHARED(Locks::mutator_lock_); 259 260 void SuspendAllInternal(Thread* self, SuspendReason reason = SuspendReason::kInternal) 261 REQUIRES(!Locks::thread_list_lock_, 262 !Locks::thread_suspend_count_lock_, 263 !Locks::mutator_lock_); 264 265 void AssertOtherThreadsAreSuspended(Thread* self) 266 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_); 267 268 std::bitset<kMaxThreadId> allocated_ids_ GUARDED_BY(Locks::allocated_thread_ids_lock_); 269 270 // The actual list of all threads. 271 std::list<Thread*> list_ GUARDED_BY(Locks::thread_list_lock_); 272 273 // Ongoing suspend all requests, used to ensure threads added to list_ respect SuspendAll, and 274 // to ensure that only one SuspendAll ot FlipThreadRoots call is active at a time. The value is 275 // always either 0 or 1. Thread_suspend_count_lock must be held continuously while these two 276 // functions modify suspend counts of all other threads and modify suspend_all_count_ . 277 int suspend_all_count_ GUARDED_BY(Locks::thread_suspend_count_lock_); 278 279 // Number of threads unregistering, ~ThreadList blocks until this hits 0. 280 int unregistering_count_ GUARDED_BY(Locks::thread_list_lock_); 281 282 // Thread suspend time histogram. Only modified when all the threads are suspended, so guarding 283 // by mutator lock ensures no thread can read when another thread is modifying it. 284 Histogram<uint64_t> suspend_all_histogram_ GUARDED_BY(Locks::mutator_lock_); 285 286 // Whether or not the current thread suspension is long. 287 bool long_suspend_; 288 289 // Whether the shutdown function has been called. This is checked in the destructor. It is an 290 // error to destroy a ThreadList instance without first calling ShutDown(). 291 bool shut_down_; 292 293 // Thread suspension timeout in nanoseconds. 294 const uint64_t thread_suspend_timeout_ns_; 295 296 std::unique_ptr<Barrier> empty_checkpoint_barrier_; 297 298 friend class Thread; 299 300 friend class Mutex; 301 friend class BaseMutex; 302 303 DISALLOW_COPY_AND_ASSIGN(ThreadList); 304 }; 305 306 // Helper for suspending all threads and getting exclusive access to the mutator lock. 307 class ScopedSuspendAll : public ValueObject { 308 public: 309 EXPORT explicit ScopedSuspendAll(const char* cause, bool long_suspend = false) 310 EXCLUSIVE_LOCK_FUNCTION(Locks::mutator_lock_) 311 REQUIRES(!Locks::thread_list_lock_, 312 !Locks::thread_suspend_count_lock_, 313 !Locks::mutator_lock_); 314 // No REQUIRES(mutator_lock_) since the unlock function already asserts this. 315 EXPORT ~ScopedSuspendAll() 316 REQUIRES(!Locks::thread_list_lock_, !Locks::thread_suspend_count_lock_) 317 UNLOCK_FUNCTION(Locks::mutator_lock_); 318 }; 319 320 } // namespace art 321 322 #endif // ART_RUNTIME_THREAD_LIST_H_ 323