1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderDataGraph.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "code/codeCache.hpp" 30 #include "code/icBuffer.hpp" 31 #include "code/nmethod.hpp" 32 #include "code/pcDesc.hpp" 33 #include "code/scopeDesc.hpp" 34 #include "compiler/compilationPolicy.hpp" 35 #include "gc/shared/collectedHeap.hpp" 36 #include "gc/shared/gcLocker.hpp" 37 #include "gc/shared/oopStorage.hpp" 38 #include "gc/shared/strongRootsScope.hpp" 39 #include "gc/shared/workerThread.hpp" 40 #include "gc/shared/workerUtils.hpp" 41 #include "interpreter/interpreter.hpp" 42 #include "jfr/jfrEvents.hpp" 43 #include "logging/log.hpp" 44 #include "logging/logStream.hpp" 45 #include "memory/resourceArea.hpp" 46 #include "memory/universe.hpp" 47 #include "oops/oop.inline.hpp" 48 #include "oops/symbol.hpp" 49 #include "oops/inlineKlass.hpp" 50 #include "runtime/atomic.hpp" 51 #include "runtime/deoptimization.hpp" 52 #include "runtime/frame.inline.hpp" 53 #include "runtime/globals.hpp" 54 #include "runtime/handles.inline.hpp" 55 #include "runtime/interfaceSupport.inline.hpp" 56 #include "runtime/javaThread.inline.hpp" 57 #include "runtime/mutexLocker.hpp" 58 #include "runtime/orderAccess.hpp" 59 #include "runtime/osThread.hpp" 60 #include "runtime/safepoint.hpp" 61 #include "runtime/safepointMechanism.inline.hpp" 62 #include "runtime/signature.hpp" 63 #include "runtime/stackWatermarkSet.inline.hpp" 64 #include "runtime/stubCodeGenerator.hpp" 65 #include "runtime/stubRoutines.hpp" 66 #include "runtime/synchronizer.hpp" 67 #include "runtime/threads.hpp" 68 #include "runtime/threadSMR.hpp" 69 #include "runtime/threadWXSetters.inline.hpp" 70 #include "runtime/timerTrace.hpp" 71 #include "services/runtimeService.hpp" 72 #include "utilities/events.hpp" 73 #include "utilities/macros.hpp" 74 #include "utilities/systemMemoryBarrier.hpp" 75 76 static void post_safepoint_begin_event(EventSafepointBegin& event, 77 uint64_t safepoint_id, 78 int thread_count, 79 int critical_thread_count) { 80 if (event.should_commit()) { 81 event.set_safepointId(safepoint_id); 82 event.set_totalThreadCount(thread_count); 83 event.set_jniCriticalThreadCount(critical_thread_count); 84 event.commit(); 85 } 86 } 87 88 static void post_safepoint_cleanup_event(EventSafepointCleanup& event, uint64_t safepoint_id) { 89 if (event.should_commit()) { 90 event.set_safepointId(safepoint_id); 91 event.commit(); 92 } 93 } 94 95 static void post_safepoint_synchronize_event(EventSafepointStateSynchronization& event, 96 uint64_t safepoint_id, 97 int initial_number_of_threads, 98 int threads_waiting_to_block, 99 uint64_t iterations) { 100 if (event.should_commit()) { 101 event.set_safepointId(safepoint_id); 102 event.set_initialThreadCount(initial_number_of_threads); 103 event.set_runningThreadCount(threads_waiting_to_block); 104 event.set_iterations(iterations); 105 event.commit(); 106 } 107 } 108 109 static void post_safepoint_cleanup_task_event(EventSafepointCleanupTask& event, 110 uint64_t safepoint_id, 111 const char* name) { 112 if (event.should_commit()) { 113 event.set_safepointId(safepoint_id); 114 event.set_name(name); 115 event.commit(); 116 } 117 } 118 119 static void post_safepoint_end_event(EventSafepointEnd& event, uint64_t safepoint_id) { 120 if (event.should_commit()) { 121 event.set_safepointId(safepoint_id); 122 event.commit(); 123 } 124 } 125 126 // SafepointCheck 127 SafepointStateTracker::SafepointStateTracker(uint64_t safepoint_id, bool at_safepoint) 128 : _safepoint_id(safepoint_id), _at_safepoint(at_safepoint) {} 129 130 bool SafepointStateTracker::safepoint_state_changed() { 131 return _safepoint_id != SafepointSynchronize::safepoint_id() || 132 _at_safepoint != SafepointSynchronize::is_at_safepoint(); 133 } 134 135 // -------------------------------------------------------------------------------------------------- 136 // Implementation of Safepoint begin/end 137 138 SafepointSynchronize::SynchronizeState volatile SafepointSynchronize::_state = SafepointSynchronize::_not_synchronized; 139 int SafepointSynchronize::_waiting_to_block = 0; 140 volatile uint64_t SafepointSynchronize::_safepoint_counter = 0; 141 uint64_t SafepointSynchronize::_safepoint_id = 0; 142 const uint64_t SafepointSynchronize::InactiveSafepointCounter = 0; 143 int SafepointSynchronize::_current_jni_active_count = 0; 144 145 WaitBarrier* SafepointSynchronize::_wait_barrier; 146 147 static bool timeout_error_printed = false; 148 149 // Statistic related 150 static jlong _safepoint_begin_time = 0; 151 static volatile int _nof_threads_hit_polling_page = 0; 152 153 void SafepointSynchronize::init(Thread* vmthread) { 154 // WaitBarrier should never be destroyed since we will have 155 // threads waiting on it while exiting. 156 _wait_barrier = new WaitBarrier(vmthread); 157 SafepointTracing::init(); 158 } 159 160 void SafepointSynchronize::increment_jni_active_count() { 161 assert(Thread::current()->is_VM_thread(), "Only VM thread may increment"); 162 ++_current_jni_active_count; 163 } 164 165 void SafepointSynchronize::decrement_waiting_to_block() { 166 assert(_waiting_to_block > 0, "sanity check"); 167 assert(Thread::current()->is_VM_thread(), "Only VM thread may decrement"); 168 --_waiting_to_block; 169 } 170 171 bool SafepointSynchronize::thread_not_running(ThreadSafepointState *cur_state) { 172 if (!cur_state->is_running()) { 173 // Robustness: asserted in the caller, but handle/tolerate it for release bits. 174 LogTarget(Error, safepoint) lt; 175 if (lt.is_enabled()) { 176 ResourceMark rm; 177 LogStream ls(lt); 178 ls.print("Illegal initial state detected: "); 179 cur_state->print_on(&ls); 180 } 181 return true; 182 } 183 cur_state->examine_state_of_thread(SafepointSynchronize::safepoint_counter()); 184 if (!cur_state->is_running()) { 185 return true; 186 } 187 LogTarget(Trace, safepoint) lt; 188 if (lt.is_enabled()) { 189 ResourceMark rm; 190 LogStream ls(lt); 191 cur_state->print_on(&ls); 192 } 193 return false; 194 } 195 196 #ifdef ASSERT 197 static void assert_list_is_valid(const ThreadSafepointState* tss_head, int still_running) { 198 int a = 0; 199 const ThreadSafepointState *tmp_tss = tss_head; 200 while (tmp_tss != nullptr) { 201 ++a; 202 assert(tmp_tss->is_running(), "Illegal initial state"); 203 tmp_tss = tmp_tss->get_next(); 204 } 205 assert(a == still_running, "Must be the same"); 206 } 207 #endif // ASSERT 208 209 static void back_off(int64_t start_time) { 210 // We start with fine-grained nanosleeping until a millisecond has 211 // passed, at which point we resort to plain naked_short_sleep. 212 if (os::javaTimeNanos() - start_time < NANOSECS_PER_MILLISEC) { 213 os::naked_short_nanosleep(10 * (NANOUNITS / MICROUNITS)); 214 } else { 215 os::naked_short_sleep(1); 216 } 217 } 218 219 int SafepointSynchronize::synchronize_threads(jlong safepoint_limit_time, int nof_threads, int* initial_running) 220 { 221 JavaThreadIteratorWithHandle jtiwh; 222 223 #ifdef ASSERT 224 for (; JavaThread *cur = jtiwh.next(); ) { 225 assert(cur->safepoint_state()->is_running(), "Illegal initial state"); 226 } 227 jtiwh.rewind(); 228 #endif // ASSERT 229 230 // Iterate through all threads until it has been determined how to stop them all at a safepoint. 231 int still_running = nof_threads; 232 ThreadSafepointState *tss_head = nullptr; 233 ThreadSafepointState **p_prev = &tss_head; 234 for (; JavaThread *cur = jtiwh.next(); ) { 235 ThreadSafepointState *cur_tss = cur->safepoint_state(); 236 assert(cur_tss->get_next() == nullptr, "Must be null"); 237 if (thread_not_running(cur_tss)) { 238 --still_running; 239 } else { 240 *p_prev = cur_tss; 241 p_prev = cur_tss->next_ptr(); 242 } 243 } 244 *p_prev = nullptr; 245 246 DEBUG_ONLY(assert_list_is_valid(tss_head, still_running);) 247 248 *initial_running = still_running; 249 250 // If there is no thread still running, we are already done. 251 if (still_running <= 0) { 252 assert(tss_head == nullptr, "Must be empty"); 253 return 1; 254 } 255 256 int iterations = 1; // The first iteration is above. 257 int64_t start_time = os::javaTimeNanos(); 258 259 do { 260 // Check if this has taken too long: 261 if (SafepointTimeout && safepoint_limit_time < os::javaTimeNanos()) { 262 print_safepoint_timeout(); 263 } 264 265 p_prev = &tss_head; 266 ThreadSafepointState *cur_tss = tss_head; 267 while (cur_tss != nullptr) { 268 assert(cur_tss->is_running(), "Illegal initial state"); 269 if (thread_not_running(cur_tss)) { 270 --still_running; 271 *p_prev = nullptr; 272 ThreadSafepointState *tmp = cur_tss; 273 cur_tss = cur_tss->get_next(); 274 tmp->set_next(nullptr); 275 } else { 276 *p_prev = cur_tss; 277 p_prev = cur_tss->next_ptr(); 278 cur_tss = cur_tss->get_next(); 279 } 280 } 281 282 DEBUG_ONLY(assert_list_is_valid(tss_head, still_running);) 283 284 if (still_running > 0) { 285 back_off(start_time); 286 } 287 288 iterations++; 289 } while (still_running > 0); 290 291 assert(tss_head == nullptr, "Must be empty"); 292 293 return iterations; 294 } 295 296 void SafepointSynchronize::arm_safepoint() { 297 // Begin the process of bringing the system to a safepoint. 298 // Java threads can be in several different states and are 299 // stopped by different mechanisms: 300 // 301 // 1. Running interpreted 302 // When executing branching/returning byte codes interpreter 303 // checks if the poll is armed, if so blocks in SS::block(). 304 // 2. Running in native code 305 // When returning from the native code, a Java thread must check 306 // the safepoint _state to see if we must block. If the 307 // VM thread sees a Java thread in native, it does 308 // not wait for this thread to block. The order of the memory 309 // writes and reads of both the safepoint state and the Java 310 // threads state is critical. In order to guarantee that the 311 // memory writes are serialized with respect to each other, 312 // the VM thread issues a memory barrier instruction. 313 // 3. Running compiled Code 314 // Compiled code reads the local polling page that 315 // is set to fault if we are trying to get to a safepoint. 316 // 4. Blocked 317 // A thread which is blocked will not be allowed to return from the 318 // block condition until the safepoint operation is complete. 319 // 5. In VM or Transitioning between states 320 // If a Java thread is currently running in the VM or transitioning 321 // between states, the safepointing code will poll the thread state 322 // until the thread blocks itself when it attempts transitions to a 323 // new state or locking a safepoint checked monitor. 324 325 // We must never miss a thread with correct safepoint id, so we must make sure we arm 326 // the wait barrier for the next safepoint id/counter. 327 // Arming must be done after resetting _current_jni_active_count, _waiting_to_block. 328 _wait_barrier->arm(static_cast<int>(_safepoint_counter + 1)); 329 330 assert((_safepoint_counter & 0x1) == 0, "must be even"); 331 // The store to _safepoint_counter must happen after any stores in arming. 332 Atomic::release_store(&_safepoint_counter, _safepoint_counter + 1); 333 334 // We are synchronizing 335 OrderAccess::storestore(); // Ordered with _safepoint_counter 336 _state = _synchronizing; 337 338 // Arming the per thread poll while having _state != _not_synchronized means safepointing 339 log_trace(safepoint)("Setting thread local yield flag for threads"); 340 OrderAccess::storestore(); // storestore, global state -> local state 341 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *cur = jtiwh.next(); ) { 342 // Make sure the threads start polling, it is time to yield. 343 SafepointMechanism::arm_local_poll(cur); 344 } 345 if (UseSystemMemoryBarrier) { 346 SystemMemoryBarrier::emit(); // storestore|storeload, global state -> local state 347 } else { 348 OrderAccess::fence(); // storestore|storeload, global state -> local state 349 } 350 } 351 352 // Roll all threads forward to a safepoint and suspend them all 353 void SafepointSynchronize::begin() { 354 assert(Thread::current()->is_VM_thread(), "Only VM thread may execute a safepoint"); 355 356 EventSafepointBegin begin_event; 357 SafepointTracing::begin(VMThread::vm_op_type()); 358 359 Universe::heap()->safepoint_synchronize_begin(); 360 361 // By getting the Threads_lock, we assure that no threads are about to start or 362 // exit. It is released again in SafepointSynchronize::end(). 363 Threads_lock->lock(); 364 365 assert( _state == _not_synchronized, "trying to safepoint synchronize with wrong state"); 366 367 int nof_threads = Threads::number_of_threads(); 368 369 _nof_threads_hit_polling_page = 0; 370 371 log_debug(safepoint)("Safepoint synchronization initiated using %s wait barrier. (%d threads)", _wait_barrier->description(), nof_threads); 372 373 // Reset the count of active JNI critical threads 374 _current_jni_active_count = 0; 375 376 // Set number of threads to wait for 377 _waiting_to_block = nof_threads; 378 379 jlong safepoint_limit_time = 0; 380 if (SafepointTimeout) { 381 // Set the limit time, so that it can be compared to see if this has taken 382 // too long to complete. 383 safepoint_limit_time = SafepointTracing::start_of_safepoint() + (jlong)(SafepointTimeoutDelay * NANOSECS_PER_MILLISEC); 384 timeout_error_printed = false; 385 } 386 387 EventSafepointStateSynchronization sync_event; 388 int initial_running = 0; 389 390 // Arms the safepoint, _current_jni_active_count and _waiting_to_block must be set before. 391 arm_safepoint(); 392 393 // Will spin until all threads are safe. 394 int iterations = synchronize_threads(safepoint_limit_time, nof_threads, &initial_running); 395 assert(_waiting_to_block == 0, "No thread should be running"); 396 397 #ifndef PRODUCT 398 // Mark all threads 399 if (VerifyCrossModifyFence) { 400 JavaThreadIteratorWithHandle jtiwh; 401 for (; JavaThread *cur = jtiwh.next(); ) { 402 cur->set_requires_cross_modify_fence(true); 403 } 404 } 405 406 if (safepoint_limit_time != 0) { 407 jlong current_time = os::javaTimeNanos(); 408 if (safepoint_limit_time < current_time) { 409 log_warning(safepoint)("# SafepointSynchronize: Finished after " 410 INT64_FORMAT_W(6) " ms", 411 (int64_t)(current_time - SafepointTracing::start_of_safepoint()) / (NANOUNITS / MILLIUNITS)); 412 } 413 } 414 #endif 415 416 assert(Threads_lock->owned_by_self(), "must hold Threads_lock"); 417 418 // Record state 419 _state = _synchronized; 420 421 OrderAccess::fence(); 422 423 // Set the new id 424 ++_safepoint_id; 425 426 #ifdef ASSERT 427 // Make sure all the threads were visited. 428 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *cur = jtiwh.next(); ) { 429 assert(cur->was_visited_for_critical_count(_safepoint_counter), "missed a thread"); 430 } 431 #endif // ASSERT 432 433 // Update the count of active JNI critical regions 434 GCLocker::set_jni_lock_count(_current_jni_active_count); 435 436 post_safepoint_synchronize_event(sync_event, 437 _safepoint_id, 438 initial_running, 439 _waiting_to_block, iterations); 440 441 SafepointTracing::synchronized(nof_threads, initial_running, _nof_threads_hit_polling_page); 442 443 // We do the safepoint cleanup first since a GC related safepoint 444 // needs cleanup to be completed before running the GC op. 445 EventSafepointCleanup cleanup_event; 446 do_cleanup_tasks(); 447 post_safepoint_cleanup_event(cleanup_event, _safepoint_id); 448 449 post_safepoint_begin_event(begin_event, _safepoint_id, nof_threads, _current_jni_active_count); 450 SafepointTracing::cleanup(); 451 } 452 453 void SafepointSynchronize::disarm_safepoint() { 454 uint64_t active_safepoint_counter = _safepoint_counter; 455 { 456 JavaThreadIteratorWithHandle jtiwh; 457 #ifdef ASSERT 458 // A pending_exception cannot be installed during a safepoint. The threads 459 // may install an async exception after they come back from a safepoint into 460 // pending_exception after they unblock. But that should happen later. 461 for (; JavaThread *cur = jtiwh.next(); ) { 462 assert (!(cur->has_pending_exception() && 463 cur->safepoint_state()->is_at_poll_safepoint()), 464 "safepoint installed a pending exception"); 465 } 466 #endif // ASSERT 467 468 OrderAccess::fence(); // keep read and write of _state from floating up 469 assert(_state == _synchronized, "must be synchronized before ending safepoint synchronization"); 470 471 // Change state first to _not_synchronized. 472 // No threads should see _synchronized when running. 473 _state = _not_synchronized; 474 475 // Set the next dormant (even) safepoint id. 476 assert((_safepoint_counter & 0x1) == 1, "must be odd"); 477 Atomic::release_store(&_safepoint_counter, _safepoint_counter + 1); 478 479 OrderAccess::fence(); // Keep the local state from floating up. 480 481 jtiwh.rewind(); 482 for (; JavaThread *current = jtiwh.next(); ) { 483 // Clear the visited flag to ensure that the critical counts are collected properly. 484 DEBUG_ONLY(current->reset_visited_for_critical_count(active_safepoint_counter);) 485 ThreadSafepointState* cur_state = current->safepoint_state(); 486 assert(!cur_state->is_running(), "Thread not suspended at safepoint"); 487 cur_state->restart(); // TSS _running 488 assert(cur_state->is_running(), "safepoint state has not been reset"); 489 } 490 } // ~JavaThreadIteratorWithHandle 491 492 // Release threads lock, so threads can be created/destroyed again. 493 Threads_lock->unlock(); 494 495 // Wake threads after local state is correctly set. 496 _wait_barrier->disarm(); 497 } 498 499 // Wake up all threads, so they are ready to resume execution after the safepoint 500 // operation has been carried out 501 void SafepointSynchronize::end() { 502 assert(Threads_lock->owned_by_self(), "must hold Threads_lock"); 503 EventSafepointEnd event; 504 assert(Thread::current()->is_VM_thread(), "Only VM thread can execute a safepoint"); 505 506 disarm_safepoint(); 507 508 Universe::heap()->safepoint_synchronize_end(); 509 510 SafepointTracing::end(); 511 512 post_safepoint_end_event(event, safepoint_id()); 513 } 514 515 bool SafepointSynchronize::is_cleanup_needed() { 516 // Need a safepoint if some inline cache buffers is non-empty 517 if (!InlineCacheBuffer::is_empty()) return true; 518 if (StringTable::needs_rehashing()) return true; 519 if (SymbolTable::needs_rehashing()) return true; 520 return false; 521 } 522 523 class ParallelCleanupTask : public WorkerTask { 524 private: 525 SubTasksDone _subtasks; 526 bool _do_lazy_roots; 527 528 class Tracer { 529 private: 530 const char* _name; 531 EventSafepointCleanupTask _event; 532 TraceTime _timer; 533 534 public: 535 Tracer(const char* name) : 536 _name(name), 537 _event(), 538 _timer(name, TRACETIME_LOG(Info, safepoint, cleanup)) {} 539 ~Tracer() { 540 post_safepoint_cleanup_task_event(_event, SafepointSynchronize::safepoint_id(), _name); 541 } 542 }; 543 544 public: 545 ParallelCleanupTask() : 546 WorkerTask("Parallel Safepoint Cleanup"), 547 _subtasks(SafepointSynchronize::SAFEPOINT_CLEANUP_NUM_TASKS), 548 _do_lazy_roots(!VMThread::vm_operation()->skip_thread_oop_barriers() && 549 Universe::heap()->uses_stack_watermark_barrier()) {} 550 551 uint expected_num_workers() const { 552 uint workers = 0; 553 554 if (SymbolTable::rehash_table_expects_safepoint_rehashing()) { 555 workers++; 556 } 557 558 if (StringTable::rehash_table_expects_safepoint_rehashing()) { 559 workers++; 560 } 561 562 if (InlineCacheBuffer::needs_update_inline_caches()) { 563 workers++; 564 } 565 566 if (_do_lazy_roots) { 567 workers++; 568 } 569 570 return MAX2<uint>(1, workers); 571 } 572 573 void work(uint worker_id) { 574 // These tasks are ordered by relative length of time to execute so that potentially longer tasks start first. 575 if (_subtasks.try_claim_task(SafepointSynchronize::SAFEPOINT_CLEANUP_SYMBOL_TABLE_REHASH)) { 576 if (SymbolTable::needs_rehashing()) { 577 Tracer t("rehashing symbol table"); 578 SymbolTable::rehash_table(); 579 } 580 } 581 582 if (_subtasks.try_claim_task(SafepointSynchronize::SAFEPOINT_CLEANUP_STRING_TABLE_REHASH)) { 583 if (StringTable::needs_rehashing()) { 584 Tracer t("rehashing string table"); 585 StringTable::rehash_table(); 586 } 587 } 588 589 if (_subtasks.try_claim_task(SafepointSynchronize::SAFEPOINT_CLEANUP_LAZY_ROOT_PROCESSING)) { 590 if (_do_lazy_roots) { 591 Tracer t("lazy partial thread root processing"); 592 class LazyRootClosure : public ThreadClosure { 593 public: 594 void do_thread(Thread* thread) { 595 StackWatermarkSet::start_processing(JavaThread::cast(thread), StackWatermarkKind::gc); 596 } 597 }; 598 LazyRootClosure cl; 599 Threads::java_threads_do(&cl); 600 } 601 } 602 603 if (_subtasks.try_claim_task(SafepointSynchronize::SAFEPOINT_CLEANUP_UPDATE_INLINE_CACHES)) { 604 Tracer t("updating inline caches"); 605 InlineCacheBuffer::update_inline_caches(); 606 } 607 608 if (_subtasks.try_claim_task(SafepointSynchronize::SAFEPOINT_CLEANUP_REQUEST_OOPSTORAGE_CLEANUP)) { 609 // Don't bother reporting event or time for this very short operation. 610 // To have any utility we'd also want to report whether needed. 611 OopStorage::trigger_cleanup_if_needed(); 612 } 613 614 _subtasks.all_tasks_claimed(); 615 } 616 }; 617 618 // Various cleaning tasks that should be done periodically at safepoints. 619 void SafepointSynchronize::do_cleanup_tasks() { 620 621 TraceTime timer("safepoint cleanup tasks", TRACETIME_LOG(Info, safepoint, cleanup)); 622 623 CollectedHeap* heap = Universe::heap(); 624 assert(heap != nullptr, "heap not initialized yet?"); 625 ParallelCleanupTask cleanup; 626 WorkerThreads* cleanup_workers = heap->safepoint_workers(); 627 const uint expected_num_workers = cleanup.expected_num_workers(); 628 if (cleanup_workers != nullptr && expected_num_workers > 1) { 629 // Parallel cleanup using GC provided thread pool. 630 const uint num_workers = MIN2(expected_num_workers, cleanup_workers->active_workers()); 631 cleanup_workers->run_task(&cleanup, num_workers); 632 } else { 633 // Serial cleanup using VMThread. 634 cleanup.work(0); 635 } 636 637 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 638 639 if (log_is_enabled(Debug, monitorinflation)) { 640 // The VMThread calls do_final_audit_and_print_stats() which calls 641 // audit_and_print_stats() at the Info level at VM exit time. 642 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */); 643 } 644 } 645 646 // Methods for determining if a JavaThread is safepoint safe. 647 648 // False means unsafe with undetermined state. 649 // True means a determined state, but it may be an unsafe state. 650 // If called from a non-safepoint context safepoint_count MUST be InactiveSafepointCounter. 651 bool SafepointSynchronize::try_stable_load_state(JavaThreadState *state, JavaThread *thread, uint64_t safepoint_count) { 652 assert((safepoint_count != InactiveSafepointCounter && 653 Thread::current() == (Thread*)VMThread::vm_thread() && 654 SafepointSynchronize::_state != _not_synchronized) 655 || safepoint_count == InactiveSafepointCounter, "Invalid check"); 656 657 // To handle the thread_blocked state on the backedge of the WaitBarrier from 658 // previous safepoint and reading the reset value (0/InactiveSafepointCounter) we 659 // re-read state after we read thread safepoint id. The JavaThread changes its 660 // thread state from thread_blocked before resetting safepoint id to 0. 661 // This guarantees the second read will be from an updated thread state. It can 662 // either be different state making this an unsafe state or it can see blocked 663 // again. When we see blocked twice with a 0 safepoint id, either: 664 // - It is normally blocked, e.g. on Mutex, TBIVM. 665 // - It was in SS:block(), looped around to SS:block() and is blocked on the WaitBarrier. 666 // - It was in SS:block() but now on a Mutex. 667 // All of these cases are safe. 668 669 *state = thread->thread_state(); 670 OrderAccess::loadload(); 671 uint64_t sid = thread->safepoint_state()->get_safepoint_id(); // Load acquire 672 if (sid != InactiveSafepointCounter && sid != safepoint_count) { 673 // In an old safepoint, state not relevant. 674 return false; 675 } 676 return *state == thread->thread_state(); 677 } 678 679 static bool safepoint_safe_with(JavaThread *thread, JavaThreadState state) { 680 switch(state) { 681 case _thread_in_native: 682 // native threads are safe if they have no java stack or have walkable stack 683 return !thread->has_last_Java_frame() || thread->frame_anchor()->walkable(); 684 685 case _thread_blocked: 686 // On wait_barrier or blocked. 687 // Blocked threads should already have walkable stack. 688 assert(!thread->has_last_Java_frame() || thread->frame_anchor()->walkable(), "blocked and not walkable"); 689 return true; 690 691 default: 692 return false; 693 } 694 } 695 696 bool SafepointSynchronize::handshake_safe(JavaThread *thread) { 697 if (thread->is_terminated()) { 698 return true; 699 } 700 JavaThreadState stable_state; 701 if (try_stable_load_state(&stable_state, thread, InactiveSafepointCounter)) { 702 return safepoint_safe_with(thread, stable_state); 703 } 704 return false; 705 } 706 707 708 // ------------------------------------------------------------------------------------------------------- 709 // Implementation of Safepoint blocking point 710 711 void SafepointSynchronize::block(JavaThread *thread) { 712 assert(thread != nullptr, "thread must be set"); 713 714 // Threads shouldn't block if they are in the middle of printing, but... 715 ttyLocker::break_tty_lock_for_safepoint(os::current_thread_id()); 716 717 // Only bail from the block() call if the thread is gone from the 718 // thread list; starting to exit should still block. 719 if (thread->is_terminated()) { 720 // block current thread if we come here from native code when VM is gone 721 thread->block_if_vm_exited(); 722 723 // otherwise do nothing 724 return; 725 } 726 727 JavaThreadState state = thread->thread_state(); 728 thread->frame_anchor()->make_walkable(); 729 730 uint64_t safepoint_id = SafepointSynchronize::safepoint_counter(); 731 732 // We have no idea where the VMThread is, it might even be at next safepoint. 733 // So we can miss this poll, but stop at next. 734 735 // Load dependent store, it must not pass loading of safepoint_id. 736 thread->safepoint_state()->set_safepoint_id(safepoint_id); // Release store 737 738 // This part we can skip if we notice we miss or are in a future safepoint. 739 OrderAccess::storestore(); 740 // Load in wait barrier should not float up 741 thread->set_thread_state_fence(_thread_blocked); 742 743 _wait_barrier->wait(static_cast<int>(safepoint_id)); 744 assert(_state != _synchronized, "Can't be"); 745 746 // If barrier is disarmed stop store from floating above loads in barrier. 747 OrderAccess::loadstore(); 748 thread->set_thread_state(state); 749 750 // Then we reset the safepoint id to inactive. 751 thread->safepoint_state()->reset_safepoint_id(); // Release store 752 753 OrderAccess::fence(); 754 755 guarantee(thread->safepoint_state()->get_safepoint_id() == InactiveSafepointCounter, 756 "The safepoint id should be set only in block path"); 757 758 // cross_modify_fence is done by SafepointMechanism::process_if_requested 759 // which is the only caller here. 760 } 761 762 // ------------------------------------------------------------------------------------------------------ 763 // Exception handlers 764 765 766 void SafepointSynchronize::handle_polling_page_exception(JavaThread *thread) { 767 assert(thread->thread_state() == _thread_in_Java, "should come from Java code"); 768 thread->set_thread_state(_thread_in_vm); 769 770 // Enable WXWrite: the function is called implicitly from java code. 771 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, thread)); 772 773 if (log_is_enabled(Info, safepoint, stats)) { 774 Atomic::inc(&_nof_threads_hit_polling_page); 775 } 776 777 ThreadSafepointState* state = thread->safepoint_state(); 778 779 state->handle_polling_page_exception(); 780 781 thread->set_thread_state(_thread_in_Java); 782 } 783 784 785 void SafepointSynchronize::print_safepoint_timeout() { 786 if (!timeout_error_printed) { 787 timeout_error_printed = true; 788 // Print out the thread info which didn't reach the safepoint for debugging 789 // purposes (useful when there are lots of threads in the debugger). 790 LogTarget(Warning, safepoint) lt; 791 if (lt.is_enabled()) { 792 ResourceMark rm; 793 LogStream ls(lt); 794 795 ls.cr(); 796 ls.print_cr("# SafepointSynchronize::begin: Timeout detected:"); 797 ls.print_cr("# SafepointSynchronize::begin: Timed out while spinning to reach a safepoint."); 798 ls.print_cr("# SafepointSynchronize::begin: Threads which did not reach the safepoint:"); 799 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *cur_thread = jtiwh.next(); ) { 800 if (cur_thread->safepoint_state()->is_running()) { 801 ls.print("# "); 802 cur_thread->print_on(&ls); 803 ls.cr(); 804 } 805 } 806 ls.print_cr("# SafepointSynchronize::begin: (End of list)"); 807 } 808 } 809 810 // To debug the long safepoint, specify both AbortVMOnSafepointTimeout & 811 // ShowMessageBoxOnError. 812 if (AbortVMOnSafepointTimeout) { 813 // Send the blocking thread a signal to terminate and write an error file. 814 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *cur_thread = jtiwh.next(); ) { 815 if (cur_thread->safepoint_state()->is_running()) { 816 if (!os::signal_thread(cur_thread, SIGILL, "blocking a safepoint")) { 817 break; // Could not send signal. Report fatal error. 818 } 819 // Give cur_thread a chance to report the error and terminate the VM. 820 os::naked_sleep(3000); 821 } 822 } 823 fatal("Safepoint sync time longer than %.6f ms detected when executing %s.", 824 SafepointTimeoutDelay, VMThread::vm_operation()->name()); 825 } 826 } 827 828 // ------------------------------------------------------------------------------------------------------- 829 // Implementation of ThreadSafepointState 830 831 ThreadSafepointState::ThreadSafepointState(JavaThread *thread) 832 : _at_poll_safepoint(false), _thread(thread), _safepoint_safe(false), 833 _safepoint_id(SafepointSynchronize::InactiveSafepointCounter), _next(nullptr) { 834 } 835 836 void ThreadSafepointState::create(JavaThread *thread) { 837 ThreadSafepointState *state = new ThreadSafepointState(thread); 838 thread->set_safepoint_state(state); 839 } 840 841 void ThreadSafepointState::destroy(JavaThread *thread) { 842 if (thread->safepoint_state()) { 843 delete(thread->safepoint_state()); 844 thread->set_safepoint_state(nullptr); 845 } 846 } 847 848 uint64_t ThreadSafepointState::get_safepoint_id() const { 849 return Atomic::load_acquire(&_safepoint_id); 850 } 851 852 void ThreadSafepointState::reset_safepoint_id() { 853 Atomic::release_store(&_safepoint_id, SafepointSynchronize::InactiveSafepointCounter); 854 } 855 856 void ThreadSafepointState::set_safepoint_id(uint64_t safepoint_id) { 857 Atomic::release_store(&_safepoint_id, safepoint_id); 858 } 859 860 void ThreadSafepointState::examine_state_of_thread(uint64_t safepoint_count) { 861 assert(is_running(), "better be running or just have hit safepoint poll"); 862 863 JavaThreadState stable_state; 864 if (!SafepointSynchronize::try_stable_load_state(&stable_state, _thread, safepoint_count)) { 865 // We could not get stable state of the JavaThread. 866 // Consider it running and just return. 867 return; 868 } 869 870 if (safepoint_safe_with(_thread, stable_state)) { 871 account_safe_thread(); 872 return; 873 } 874 875 // All other thread states will continue to run until they 876 // transition and self-block in state _blocked 877 // Safepoint polling in compiled code causes the Java threads to do the same. 878 // Note: new threads may require a malloc so they must be allowed to finish 879 880 assert(is_running(), "examine_state_of_thread on non-running thread"); 881 return; 882 } 883 884 void ThreadSafepointState::account_safe_thread() { 885 SafepointSynchronize::decrement_waiting_to_block(); 886 if (_thread->in_critical()) { 887 // Notice that this thread is in a critical section 888 SafepointSynchronize::increment_jni_active_count(); 889 } 890 DEBUG_ONLY(_thread->set_visited_for_critical_count(SafepointSynchronize::safepoint_counter());) 891 assert(!_safepoint_safe, "Must be unsafe before safe"); 892 _safepoint_safe = true; 893 } 894 895 void ThreadSafepointState::restart() { 896 assert(_safepoint_safe, "Must be safe before unsafe"); 897 _safepoint_safe = false; 898 } 899 900 void ThreadSafepointState::print_on(outputStream *st) const { 901 const char *s = _safepoint_safe ? "_at_safepoint" : "_running"; 902 903 st->print_cr("Thread: " INTPTR_FORMAT 904 " [0x%2x] State: %s _at_poll_safepoint %d", 905 p2i(_thread), _thread->osthread()->thread_id(), s, _at_poll_safepoint); 906 907 _thread->print_thread_state_on(st); 908 } 909 910 // --------------------------------------------------------------------------------------------------------------------- 911 912 // Process pending operation. 913 void ThreadSafepointState::handle_polling_page_exception() { 914 JavaThread* self = thread(); 915 assert(self == JavaThread::current(), "must be self"); 916 917 // Step 1: Find the nmethod from the return address 918 address real_return_addr = self->saved_exception_pc(); 919 920 CodeBlob *cb = CodeCache::find_blob(real_return_addr); 921 assert(cb != nullptr && cb->is_compiled(), "return address should be in nmethod"); 922 CompiledMethod* nm = (CompiledMethod*)cb; 923 924 // Find frame of caller 925 frame stub_fr = self->last_frame(); 926 CodeBlob* stub_cb = stub_fr.cb(); 927 assert(stub_cb->is_safepoint_stub(), "must be a safepoint stub"); 928 RegisterMap map(self, 929 RegisterMap::UpdateMap::include, 930 RegisterMap::ProcessFrames::skip, 931 RegisterMap::WalkContinuation::skip); 932 frame caller_fr = stub_fr.sender(&map); 933 934 // Should only be poll_return or poll 935 assert( nm->is_at_poll_or_poll_return(real_return_addr), "should not be at call" ); 936 937 // This is a poll immediately before a return. The exception handling code 938 // has already had the effect of causing the return to occur, so the execution 939 // will continue immediately after the call. In addition, the oopmap at the 940 // return point does not mark the return value as an oop (if it is), so 941 // it needs a handle here to be updated. 942 if( nm->is_at_poll_return(real_return_addr) ) { 943 ResourceMark rm; 944 // See if return type is an oop. 945 Method* method = nm->method(); 946 bool return_oop = method->is_returning_oop(); 947 HandleMark hm(self); 948 949 GrowableArray<Handle> return_values; 950 InlineKlass* vk = nullptr; 951 if (return_oop && InlineTypeReturnedAsFields && 952 (method->result_type() == T_PRIMITIVE_OBJECT || method->result_type() == T_OBJECT)) { 953 // Check if an inline type is returned as fields 954 vk = InlineKlass::returned_inline_klass(map); 955 if (vk != nullptr) { 956 // We're at a safepoint at the return of a method that returns 957 // multiple values. We must make sure we preserve the oop values 958 // across the safepoint. 959 assert(vk == method->returns_inline_type(thread()), "bad inline klass"); 960 vk->save_oop_fields(map, return_values); 961 return_oop = false; 962 } 963 } 964 965 if (return_oop) { 966 // The oop result has been saved on the stack together with all 967 // the other registers. In order to preserve it over GCs we need 968 // to keep it in a handle. 969 oop result = caller_fr.saved_oop_result(&map); 970 assert(oopDesc::is_oop_or_null(result), "must be oop"); 971 return_values.push(Handle(self, result)); 972 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 973 } 974 975 // We get here if compiled return polls found a reason to call into the VM. 976 // One condition for that is that the top frame is not yet safe to use. 977 // The following stack watermark barrier poll will catch such situations. 978 StackWatermarkSet::after_unwind(self); 979 980 // Process pending operation 981 SafepointMechanism::process_if_requested_with_exit_check(self, true /* check asyncs */); 982 983 // restore oop result, if any 984 if (return_oop) { 985 assert(return_values.length() == 1, "only one return value"); 986 caller_fr.set_saved_oop_result(&map, return_values.pop()()); 987 } else if (vk != nullptr) { 988 vk->restore_oop_results(map, return_values); 989 } 990 } 991 992 // This is a safepoint poll. Verify the return address and block. 993 else { 994 995 // verify the blob built the "return address" correctly 996 assert(real_return_addr == caller_fr.pc(), "must match"); 997 998 set_at_poll_safepoint(true); 999 // Process pending operation 1000 // We never deliver an async exception at a polling point as the 1001 // compiler may not have an exception handler for it (polling at 1002 // a return point is ok though). We will check for a pending async 1003 // exception below and deoptimize if needed. We also cannot deoptimize 1004 // and still install the exception here because live registers needed 1005 // during deoptimization are clobbered by the exception path. The 1006 // exception will just be delivered once we get into the interpreter. 1007 SafepointMechanism::process_if_requested_with_exit_check(self, false /* check asyncs */); 1008 set_at_poll_safepoint(false); 1009 1010 if (self->has_async_exception_condition()) { 1011 Deoptimization::deoptimize_frame(self, caller_fr.id()); 1012 log_info(exceptions)("deferred async exception at compiled safepoint"); 1013 } 1014 1015 // If an exception has been installed we must verify that the top frame wasn't deoptimized. 1016 if (self->has_pending_exception() ) { 1017 RegisterMap map(self, 1018 RegisterMap::UpdateMap::include, 1019 RegisterMap::ProcessFrames::skip, 1020 RegisterMap::WalkContinuation::skip); 1021 frame caller_fr = stub_fr.sender(&map); 1022 if (caller_fr.is_deoptimized_frame()) { 1023 // The exception path will destroy registers that are still 1024 // live and will be needed during deoptimization, so if we 1025 // have an exception now things are messed up. We only check 1026 // at this scope because for a poll return it is ok to deoptimize 1027 // while having a pending exception since the call we are returning 1028 // from already collides with exception handling registers and 1029 // so there is no issue (the exception handling path kills call 1030 // result registers but this is ok since the exception kills 1031 // the result anyway). 1032 fatal("Exception installed and deoptimization is pending"); 1033 } 1034 } 1035 } 1036 } 1037 1038 1039 // ------------------------------------------------------------------------------------------------------- 1040 // Implementation of SafepointTracing 1041 1042 jlong SafepointTracing::_last_safepoint_begin_time_ns = 0; 1043 jlong SafepointTracing::_last_safepoint_sync_time_ns = 0; 1044 jlong SafepointTracing::_last_safepoint_cleanup_time_ns = 0; 1045 jlong SafepointTracing::_last_safepoint_end_time_ns = 0; 1046 jlong SafepointTracing::_last_app_time_ns = 0; 1047 int SafepointTracing::_nof_threads = 0; 1048 int SafepointTracing::_nof_running = 0; 1049 int SafepointTracing::_page_trap = 0; 1050 VM_Operation::VMOp_Type SafepointTracing::_current_type; 1051 jlong SafepointTracing::_max_sync_time = 0; 1052 jlong SafepointTracing::_max_cleanup_time = 0; 1053 jlong SafepointTracing::_max_vmop_time = 0; 1054 uint64_t SafepointTracing::_op_count[VM_Operation::VMOp_Terminating] = {0}; 1055 1056 void SafepointTracing::init() { 1057 // Application start 1058 _last_safepoint_end_time_ns = os::javaTimeNanos(); 1059 } 1060 1061 // Helper method to print the header. 1062 static void print_header(outputStream* st) { 1063 // The number of spaces is significant here, and should match the format 1064 // specifiers in print_statistics(). 1065 1066 st->print("VM Operation " 1067 "[ threads: total initial_running ]" 1068 "[ time: sync cleanup vmop total ]"); 1069 1070 st->print_cr(" page_trap_count"); 1071 } 1072 1073 // This prints a nice table. To get the statistics to not shift due to the logging uptime 1074 // decorator, use the option as: -Xlog:safepoint+stats:[outputfile]:none 1075 void SafepointTracing::statistics_log() { 1076 LogTarget(Info, safepoint, stats) lt; 1077 assert (lt.is_enabled(), "should only be called when printing statistics is enabled"); 1078 LogStream ls(lt); 1079 1080 static int _cur_stat_index = 0; 1081 1082 // Print header every 30 entries 1083 if ((_cur_stat_index % 30) == 0) { 1084 print_header(&ls); 1085 _cur_stat_index = 1; // wrap 1086 } else { 1087 _cur_stat_index++; 1088 } 1089 1090 ls.print("%-28s [ " 1091 INT32_FORMAT_W(8) " " INT32_FORMAT_W(8) " " 1092 "]", 1093 VM_Operation::name(_current_type), 1094 _nof_threads, 1095 _nof_running); 1096 ls.print("[ " 1097 INT64_FORMAT_W(10) " " INT64_FORMAT_W(10) " " 1098 INT64_FORMAT_W(10) " " INT64_FORMAT_W(10) " ]", 1099 (int64_t)(_last_safepoint_sync_time_ns - _last_safepoint_begin_time_ns), 1100 (int64_t)(_last_safepoint_cleanup_time_ns - _last_safepoint_sync_time_ns), 1101 (int64_t)(_last_safepoint_end_time_ns - _last_safepoint_cleanup_time_ns), 1102 (int64_t)(_last_safepoint_end_time_ns - _last_safepoint_begin_time_ns)); 1103 1104 ls.print_cr(INT32_FORMAT_W(16), _page_trap); 1105 } 1106 1107 // This method will be called when VM exits. This tries to summarize the sampling. 1108 // Current thread may already be deleted, so don't use ResourceMark. 1109 void SafepointTracing::statistics_exit_log() { 1110 if (!log_is_enabled(Info, safepoint, stats)) { 1111 return; 1112 } 1113 for (int index = 0; index < VM_Operation::VMOp_Terminating; index++) { 1114 if (_op_count[index] != 0) { 1115 log_info(safepoint, stats)("%-28s" UINT64_FORMAT_W(10), VM_Operation::name(index), 1116 _op_count[index]); 1117 } 1118 } 1119 1120 log_info(safepoint, stats)("Maximum sync time " INT64_FORMAT" ns", 1121 (int64_t)(_max_sync_time)); 1122 log_info(safepoint, stats)("Maximum cleanup time " INT64_FORMAT" ns", 1123 (int64_t)(_max_cleanup_time)); 1124 log_info(safepoint, stats)("Maximum vm operation time (except for Exit VM operation) " 1125 INT64_FORMAT " ns", 1126 (int64_t)(_max_vmop_time)); 1127 } 1128 1129 void SafepointTracing::begin(VM_Operation::VMOp_Type type) { 1130 _op_count[type]++; 1131 _current_type = type; 1132 1133 // update the time stamp to begin recording safepoint time 1134 _last_safepoint_begin_time_ns = os::javaTimeNanos(); 1135 _last_safepoint_sync_time_ns = 0; 1136 _last_safepoint_cleanup_time_ns = 0; 1137 1138 _last_app_time_ns = _last_safepoint_begin_time_ns - _last_safepoint_end_time_ns; 1139 _last_safepoint_end_time_ns = 0; 1140 1141 RuntimeService::record_safepoint_begin(_last_app_time_ns); 1142 } 1143 1144 void SafepointTracing::synchronized(int nof_threads, int nof_running, int traps) { 1145 _last_safepoint_sync_time_ns = os::javaTimeNanos(); 1146 _nof_threads = nof_threads; 1147 _nof_running = nof_running; 1148 _page_trap = traps; 1149 RuntimeService::record_safepoint_synchronized(_last_safepoint_sync_time_ns - _last_safepoint_begin_time_ns); 1150 } 1151 1152 void SafepointTracing::cleanup() { 1153 _last_safepoint_cleanup_time_ns = os::javaTimeNanos(); 1154 } 1155 1156 void SafepointTracing::end() { 1157 _last_safepoint_end_time_ns = os::javaTimeNanos(); 1158 1159 if (_max_sync_time < (_last_safepoint_sync_time_ns - _last_safepoint_begin_time_ns)) { 1160 _max_sync_time = _last_safepoint_sync_time_ns - _last_safepoint_begin_time_ns; 1161 } 1162 if (_max_cleanup_time < (_last_safepoint_cleanup_time_ns - _last_safepoint_sync_time_ns)) { 1163 _max_cleanup_time = _last_safepoint_cleanup_time_ns - _last_safepoint_sync_time_ns; 1164 } 1165 if (_max_vmop_time < (_last_safepoint_end_time_ns - _last_safepoint_sync_time_ns)) { 1166 _max_vmop_time = _last_safepoint_end_time_ns - _last_safepoint_sync_time_ns; 1167 } 1168 if (log_is_enabled(Info, safepoint, stats)) { 1169 statistics_log(); 1170 } 1171 1172 log_info(safepoint)( 1173 "Safepoint \"%s\", " 1174 "Time since last: " JLONG_FORMAT " ns, " 1175 "Reaching safepoint: " JLONG_FORMAT " ns, " 1176 "Cleanup: " JLONG_FORMAT " ns, " 1177 "At safepoint: " JLONG_FORMAT " ns, " 1178 "Total: " JLONG_FORMAT " ns", 1179 VM_Operation::name(_current_type), 1180 _last_app_time_ns, 1181 _last_safepoint_sync_time_ns - _last_safepoint_begin_time_ns, 1182 _last_safepoint_cleanup_time_ns - _last_safepoint_sync_time_ns, 1183 _last_safepoint_end_time_ns - _last_safepoint_cleanup_time_ns, 1184 _last_safepoint_end_time_ns - _last_safepoint_begin_time_ns 1185 ); 1186 1187 RuntimeService::record_safepoint_end(_last_safepoint_end_time_ns - _last_safepoint_sync_time_ns); 1188 }