1 /* 2 * Copyright (c) 1998, 2025, 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 "classfile/vmSymbols.hpp" 26 #include "gc/shared/collectedHeap.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "logging/log.hpp" 29 #include "logging/logStream.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/padded.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "memory/universe.hpp" 34 #include "oops/markWord.hpp" 35 #include "oops/oop.inline.hpp" 36 #include "runtime/atomicAccess.hpp" 37 #include "runtime/basicLock.inline.hpp" 38 #include "runtime/frame.inline.hpp" 39 #include "runtime/globals.hpp" 40 #include "runtime/handles.inline.hpp" 41 #include "runtime/handshake.hpp" 42 #include "runtime/interfaceSupport.inline.hpp" 43 #include "runtime/javaThread.hpp" 44 #include "runtime/lightweightSynchronizer.hpp" 45 #include "runtime/lockStack.inline.hpp" 46 #include "runtime/mutexLocker.hpp" 47 #include "runtime/objectMonitor.inline.hpp" 48 #include "runtime/os.inline.hpp" 49 #include "runtime/osThread.hpp" 50 #include "runtime/safepointMechanism.inline.hpp" 51 #include "runtime/safepointVerifiers.hpp" 52 #include "runtime/sharedRuntime.hpp" 53 #include "runtime/stubRoutines.hpp" 54 #include "runtime/synchronizer.inline.hpp" 55 #include "runtime/threads.hpp" 56 #include "runtime/timer.hpp" 57 #include "runtime/trimNativeHeap.hpp" 58 #include "runtime/vframe.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "utilities/align.hpp" 61 #include "utilities/dtrace.hpp" 62 #include "utilities/events.hpp" 63 #include "utilities/globalCounter.inline.hpp" 64 #include "utilities/globalDefinitions.hpp" 65 #include "utilities/linkedlist.hpp" 66 #include "utilities/preserveException.hpp" 67 68 class ObjectMonitorDeflationLogging; 69 70 void MonitorList::add(ObjectMonitor* m) { 71 ObjectMonitor* head; 72 do { 73 head = AtomicAccess::load(&_head); 74 m->set_next_om(head); 75 } while (AtomicAccess::cmpxchg(&_head, head, m) != head); 76 77 size_t count = AtomicAccess::add(&_count, 1u, memory_order_relaxed); 78 size_t old_max; 79 do { 80 old_max = AtomicAccess::load(&_max); 81 if (count <= old_max) { 82 break; 83 } 84 } while (AtomicAccess::cmpxchg(&_max, old_max, count, memory_order_relaxed) != old_max); 85 } 86 87 size_t MonitorList::count() const { 88 return AtomicAccess::load(&_count); 89 } 90 91 size_t MonitorList::max() const { 92 return AtomicAccess::load(&_max); 93 } 94 95 class ObjectMonitorDeflationSafepointer : public StackObj { 96 JavaThread* const _current; 97 ObjectMonitorDeflationLogging* const _log; 98 99 public: 100 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 101 : _current(current), _log(log) {} 102 103 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 104 }; 105 106 // Walk the in-use list and unlink deflated ObjectMonitors. 107 // Returns the number of unlinked ObjectMonitors. 108 size_t MonitorList::unlink_deflated(size_t deflated_count, 109 GrowableArray<ObjectMonitor*>* unlinked_list, 110 ObjectMonitorDeflationSafepointer* safepointer) { 111 size_t unlinked_count = 0; 112 ObjectMonitor* prev = nullptr; 113 ObjectMonitor* m = AtomicAccess::load_acquire(&_head); 114 115 while (m != nullptr) { 116 if (m->is_being_async_deflated()) { 117 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 118 // modify the list once per batch. The batch starts at "m". 119 size_t unlinked_batch = 0; 120 ObjectMonitor* next = m; 121 // Look for at most MonitorUnlinkBatch monitors, or the number of 122 // deflated and not unlinked monitors, whatever comes first. 123 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 124 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 125 do { 126 ObjectMonitor* next_next = next->next_om(); 127 unlinked_batch++; 128 unlinked_list->append(next); 129 next = next_next; 130 if (unlinked_batch >= unlinked_batch_limit) { 131 // Reached the max batch, so bail out of the gathering loop. 132 break; 133 } 134 if (prev == nullptr && AtomicAccess::load(&_head) != m) { 135 // Current batch used to be at head, but it is not at head anymore. 136 // Bail out and figure out where we currently are. This avoids long 137 // walks searching for new prev during unlink under heavy list inserts. 138 break; 139 } 140 } while (next != nullptr && next->is_being_async_deflated()); 141 142 // Unlink the found batch. 143 if (prev == nullptr) { 144 // The current batch is the first batch, so there is a chance that it starts at head. 145 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 146 ObjectMonitor* prev_head = AtomicAccess::cmpxchg(&_head, m, next); 147 if (prev_head != m) { 148 // Something must have updated the head. Figure out the actual prev for this batch. 149 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 150 prev = n; 151 } 152 assert(prev != nullptr, "Should have found the prev for the current batch"); 153 prev->set_next_om(next); 154 } 155 } else { 156 // The current batch is preceded by another batch. This guarantees the current batch 157 // does not start at head. Unlink the entire current batch without updating the head. 158 assert(AtomicAccess::load(&_head) != m, "Sanity"); 159 prev->set_next_om(next); 160 } 161 162 unlinked_count += unlinked_batch; 163 if (unlinked_count >= deflated_count) { 164 // Reached the max so bail out of the searching loop. 165 // There should be no more deflated monitors left. 166 break; 167 } 168 m = next; 169 } else { 170 prev = m; 171 m = m->next_om(); 172 } 173 174 // Must check for a safepoint/handshake and honor it. 175 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 176 } 177 178 #ifdef ASSERT 179 // Invariant: the code above should unlink all deflated monitors. 180 // The code that runs after this unlinking does not expect deflated monitors. 181 // Notably, attempting to deflate the already deflated monitor would break. 182 { 183 ObjectMonitor* m = AtomicAccess::load_acquire(&_head); 184 while (m != nullptr) { 185 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 186 m = m->next_om(); 187 } 188 } 189 #endif 190 191 AtomicAccess::sub(&_count, unlinked_count); 192 return unlinked_count; 193 } 194 195 MonitorList::Iterator MonitorList::iterator() const { 196 return Iterator(AtomicAccess::load_acquire(&_head)); 197 } 198 199 ObjectMonitor* MonitorList::Iterator::next() { 200 ObjectMonitor* current = _current; 201 _current = current->next_om(); 202 return current; 203 } 204 205 // The "core" versions of monitor enter and exit reside in this file. 206 // The interpreter and compilers contain specialized transliterated 207 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 208 // fast_lock(...) for instance. If you make changes here, make sure to modify the 209 // interpreter, and both C1 and C2 fast-path inline locking code emission. 210 // 211 // ----------------------------------------------------------------------------- 212 213 #ifdef DTRACE_ENABLED 214 215 // Only bother with this argument setup if dtrace is available 216 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 217 218 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 219 char* bytes = nullptr; \ 220 int len = 0; \ 221 jlong jtid = SharedRuntime::get_java_tid(thread); \ 222 Symbol* klassname = obj->klass()->name(); \ 223 if (klassname != nullptr) { \ 224 bytes = (char*)klassname->bytes(); \ 225 len = klassname->utf8_length(); \ 226 } 227 228 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 229 { \ 230 if (DTraceMonitorProbes) { \ 231 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 232 HOTSPOT_MONITOR_WAIT(jtid, \ 233 (uintptr_t)(monitor), bytes, len, (millis)); \ 234 } \ 235 } 236 237 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 238 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 239 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 240 241 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 242 { \ 243 if (DTraceMonitorProbes) { \ 244 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 245 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 246 (uintptr_t)(monitor), bytes, len); \ 247 } \ 248 } 249 250 #else // ndef DTRACE_ENABLED 251 252 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 253 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 254 255 #endif // ndef DTRACE_ENABLED 256 257 // This exists only as a workaround of dtrace bug 6254741 258 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 259 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 260 return 0; 261 } 262 263 static constexpr size_t inflation_lock_count() { 264 return 256; 265 } 266 267 // Static storage for an array of PlatformMutex. 268 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 269 270 static inline PlatformMutex* inflation_lock(size_t index) { 271 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 272 } 273 274 void ObjectSynchronizer::initialize() { 275 for (size_t i = 0; i < inflation_lock_count(); i++) { 276 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 277 } 278 // Start the ceiling with the estimate for one thread. 279 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 280 281 // Start the timer for deflations, so it does not trigger immediately. 282 _last_async_deflation_time_ns = os::javaTimeNanos(); 283 284 LightweightSynchronizer::initialize(); 285 } 286 287 MonitorList ObjectSynchronizer::_in_use_list; 288 // monitors_used_above_threshold() policy is as follows: 289 // 290 // The ratio of the current _in_use_list count to the ceiling is used 291 // to determine if we are above MonitorUsedDeflationThreshold and need 292 // to do an async monitor deflation cycle. The ceiling is increased by 293 // AvgMonitorsPerThreadEstimate when a thread is added to the system 294 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 295 // removed from the system. 296 // 297 // Note: If the _in_use_list max exceeds the ceiling, then 298 // monitors_used_above_threshold() will use the in_use_list max instead 299 // of the thread count derived ceiling because we have used more 300 // ObjectMonitors than the estimated average. 301 // 302 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 303 // no-progress async monitor deflation cycles in a row, then the ceiling 304 // is adjusted upwards by monitors_used_above_threshold(). 305 // 306 // Start the ceiling with the estimate for one thread in initialize() 307 // which is called after cmd line options are processed. 308 static size_t _in_use_list_ceiling = 0; 309 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 310 bool volatile ObjectSynchronizer::_is_final_audit = false; 311 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 312 static uintx _no_progress_cnt = 0; 313 static bool _no_progress_skip_increment = false; 314 315 // =====================> Quick functions 316 317 // The quick_* forms are special fast-path variants used to improve 318 // performance. In the simplest case, a "quick_*" implementation could 319 // simply return false, in which case the caller will perform the necessary 320 // state transitions and call the slow-path form. 321 // The fast-path is designed to handle frequently arising cases in an efficient 322 // manner and is just a degenerate "optimistic" variant of the slow-path. 323 // returns true -- to indicate the call was satisfied. 324 // returns false -- to indicate the call needs the services of the slow-path. 325 // A no-loitering ordinance is in effect for code in the quick_* family 326 // operators: safepoints or indefinite blocking (blocking that might span a 327 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 328 // entry. 329 // 330 // Consider: An interesting optimization is to have the JIT recognize the 331 // following common idiom: 332 // synchronized (someobj) { .... ; notify(); } 333 // That is, we find a notify() or notifyAll() call that immediately precedes 334 // the monitorexit operation. In that case the JIT could fuse the operations 335 // into a single notifyAndExit() runtime primitive. 336 337 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 338 assert(current->thread_state() == _thread_in_Java, "invariant"); 339 NoSafepointVerifier nsv; 340 if (obj == nullptr) return false; // slow-path for invalid obj 341 const markWord mark = obj->mark(); 342 343 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 344 // Degenerate notify 345 // fast-locked by caller so by definition the implied waitset is empty. 346 return true; 347 } 348 349 if (mark.has_monitor()) { 350 ObjectMonitor* const mon = read_monitor(current, obj, mark); 351 if (mon == nullptr) { 352 // Racing with inflation/deflation go slow path 353 return false; 354 } 355 assert(mon->object() == oop(obj), "invariant"); 356 if (!mon->has_owner(current)) return false; // slow-path for IMS exception 357 358 if (mon->first_waiter() != nullptr) { 359 // We have one or more waiters. Since this is an inflated monitor 360 // that we own, we quickly notify them here and now, avoiding the slow-path. 361 if (all) { 362 mon->quick_notifyAll(current); 363 } else { 364 mon->quick_notify(current); 365 } 366 } 367 return true; 368 } 369 370 // other IMS exception states take the slow-path 371 return false; 372 } 373 374 // Handle notifications when synchronizing on value based classes 375 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 376 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 377 frame last_frame = locking_thread->last_frame(); 378 bool bcp_was_adjusted = false; 379 // Don't decrement bcp if it points to the frame's first instruction. This happens when 380 // handle_sync_on_value_based_class() is called because of a synchronized method. There 381 // is no actual monitorenter instruction in the byte code in this case. 382 if (last_frame.is_interpreted_frame() && 383 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 384 // adjust bcp to point back to monitorenter so that we print the correct line numbers 385 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 386 bcp_was_adjusted = true; 387 } 388 389 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 390 ResourceMark rm; 391 stringStream ss; 392 locking_thread->print_active_stack_on(&ss); 393 char* base = (char*)strstr(ss.base(), "at"); 394 char* newline = (char*)strchr(ss.base(), '\n'); 395 if (newline != nullptr) { 396 *newline = '\0'; 397 } 398 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 399 } else { 400 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 401 ResourceMark rm; 402 Log(valuebasedclasses) vblog; 403 404 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 405 if (locking_thread->has_last_Java_frame()) { 406 LogStream info_stream(vblog.info()); 407 locking_thread->print_active_stack_on(&info_stream); 408 } else { 409 vblog.info("Cannot find the last Java frame"); 410 } 411 412 EventSyncOnValueBasedClass event; 413 if (event.should_commit()) { 414 event.set_valueBasedClass(obj->klass()); 415 event.commit(); 416 } 417 } 418 419 if (bcp_was_adjusted) { 420 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 421 } 422 } 423 424 // ----------------------------------------------------------------------------- 425 // Monitor Enter/Exit 426 427 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 428 // When called with locking_thread != Thread::current() some mechanism must synchronize 429 // the locking_thread with respect to the current thread. Currently only used when 430 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 431 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 432 return LightweightSynchronizer::enter_for(obj, lock, locking_thread); 433 } 434 435 // ----------------------------------------------------------------------------- 436 // JNI locks on java objects 437 // NOTE: must use heavy weight monitor to handle jni monitor enter 438 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 439 // Top native frames in the stack will not be seen if we attempt 440 // preemption, since we start walking from the last Java anchor. 441 NoPreemptMark npm(current); 442 443 if (obj->klass()->is_value_based()) { 444 handle_sync_on_value_based_class(obj, current); 445 } 446 447 // the current locking is from JNI instead of Java code 448 current->set_current_pending_monitor_is_from_java(false); 449 // An async deflation can race after the inflate() call and before 450 // enter() can make the ObjectMonitor busy. enter() returns false if 451 // we have lost the race to async deflation and we simply try again. 452 while (true) { 453 BasicLock lock; 454 if (LightweightSynchronizer::inflate_and_enter(obj(), &lock, inflate_cause_jni_enter, current, current) != nullptr) { 455 break; 456 } 457 } 458 current->set_current_pending_monitor_is_from_java(true); 459 } 460 461 // NOTE: must use heavy weight monitor to handle jni monitor exit 462 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 463 JavaThread* current = THREAD; 464 465 ObjectMonitor* monitor; 466 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK); 467 // If this thread has locked the object, exit the monitor. We 468 // intentionally do not use CHECK on check_owner because we must exit the 469 // monitor even if an exception was already pending. 470 if (monitor->check_owner(THREAD)) { 471 monitor->exit(current); 472 } 473 } 474 475 // ----------------------------------------------------------------------------- 476 // Internal VM locks on java objects 477 // standard constructor, allows locking failures 478 ObjectLocker::ObjectLocker(Handle obj, TRAPS) : _thread(THREAD), _obj(obj), 479 _npm(_thread, _thread->at_preemptable_init() /* ignore_mark */), _skip_exit(false) { 480 assert(!_thread->preempting(), ""); 481 482 _thread->check_for_valid_safepoint_state(); 483 484 if (_obj() != nullptr) { 485 ObjectSynchronizer::enter(_obj, &_lock, _thread); 486 487 if (_thread->preempting()) { 488 // If preemption was cancelled we acquired the monitor after freezing 489 // the frames. Redoing the vm call laterĀ in thaw will require us to 490 // release it since the call should look like the original one. We 491 // do it in ~ObjectLocker to reduce the window of time we hold the 492 // monitor since we can't do anything useful with it now, and would 493 // otherwise just force other vthreads to preempt in case they try 494 // to acquire this monitor. 495 _skip_exit = !_thread->preemption_cancelled(); 496 ObjectSynchronizer::read_monitor(_thread, _obj())->set_object_strong(); 497 _thread->set_pending_preempted_exception(); 498 499 } 500 } 501 } 502 503 ObjectLocker::~ObjectLocker() { 504 if (_obj() != nullptr && !_skip_exit) { 505 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 506 } 507 } 508 509 void ObjectLocker::wait_uninterruptibly(TRAPS) { 510 ObjectSynchronizer::waitUninterruptibly(_obj, 0, _thread); 511 if (_thread->preempting()) { 512 _skip_exit = true; 513 ObjectSynchronizer::read_monitor(_thread, _obj())->set_object_strong(); 514 _thread->set_pending_preempted_exception(); 515 } 516 } 517 518 // ----------------------------------------------------------------------------- 519 // Wait/Notify/NotifyAll 520 // NOTE: must use heavy weight monitor to handle wait() 521 522 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 523 JavaThread* current = THREAD; 524 if (millis < 0) { 525 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 526 } 527 528 ObjectMonitor* monitor; 529 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0); 530 531 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 532 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 533 534 // This dummy call is in place to get around dtrace bug 6254741. Once 535 // that's fixed we can uncomment the following line, remove the call 536 // and change this function back into a "void" func. 537 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 538 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 539 return ret_code; 540 } 541 542 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 543 assert(millis >= 0, "timeout value is negative"); 544 545 ObjectMonitor* monitor; 546 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK); 547 monitor->wait(millis, false, THREAD); 548 } 549 550 551 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 552 JavaThread* current = THREAD; 553 554 markWord mark = obj->mark(); 555 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 556 // Not inflated so there can't be any waiters to notify. 557 return; 558 } 559 ObjectMonitor* monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 560 monitor->notify(CHECK); 561 } 562 563 // NOTE: see comment of notify() 564 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 565 JavaThread* current = THREAD; 566 567 markWord mark = obj->mark(); 568 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 569 // Not inflated so there can't be any waiters to notify. 570 return; 571 } 572 573 ObjectMonitor* monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 574 monitor->notifyAll(CHECK); 575 } 576 577 // ----------------------------------------------------------------------------- 578 // Hash Code handling 579 580 struct SharedGlobals { 581 char _pad_prefix[OM_CACHE_LINE_SIZE]; 582 // This is a highly shared mostly-read variable. 583 // To avoid false-sharing it needs to be the sole occupant of a cache line. 584 volatile int stw_random; 585 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 586 // Hot RW variable -- Sequester to avoid false-sharing 587 volatile int hc_sequence; 588 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 589 }; 590 591 static SharedGlobals GVars; 592 593 // hashCode() generation : 594 // 595 // Possibilities: 596 // * MD5Digest of {obj,stw_random} 597 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 598 // * A DES- or AES-style SBox[] mechanism 599 // * One of the Phi-based schemes, such as: 600 // 2654435761 = 2^32 * Phi (golden ratio) 601 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 602 // * A variation of Marsaglia's shift-xor RNG scheme. 603 // * (obj ^ stw_random) is appealing, but can result 604 // in undesirable regularity in the hashCode values of adjacent objects 605 // (objects allocated back-to-back, in particular). This could potentially 606 // result in hashtable collisions and reduced hashtable efficiency. 607 // There are simple ways to "diffuse" the middle address bits over the 608 // generated hashCode values: 609 610 static intptr_t get_next_hash(Thread* current, oop obj) { 611 intptr_t value = 0; 612 if (hashCode == 0) { 613 // This form uses global Park-Miller RNG. 614 // On MP system we'll have lots of RW access to a global, so the 615 // mechanism induces lots of coherency traffic. 616 value = os::random(); 617 } else if (hashCode == 1) { 618 // This variation has the property of being stable (idempotent) 619 // between STW operations. This can be useful in some of the 1-0 620 // synchronization schemes. 621 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 622 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 623 } else if (hashCode == 2) { 624 value = 1; // for sensitivity testing 625 } else if (hashCode == 3) { 626 value = ++GVars.hc_sequence; 627 } else if (hashCode == 4) { 628 value = cast_from_oop<intptr_t>(obj); 629 } else { 630 // Marsaglia's xor-shift scheme with thread-specific state 631 // This is probably the best overall implementation -- we'll 632 // likely make this the default in future releases. 633 unsigned t = current->_hashStateX; 634 t ^= (t << 11); 635 current->_hashStateX = current->_hashStateY; 636 current->_hashStateY = current->_hashStateZ; 637 current->_hashStateZ = current->_hashStateW; 638 unsigned v = current->_hashStateW; 639 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 640 current->_hashStateW = v; 641 value = v; 642 } 643 644 value &= markWord::hash_mask; 645 if (value == 0) value = 0xBAD; 646 assert(value != markWord::no_hash, "invariant"); 647 return value; 648 } 649 650 static intptr_t install_hash_code(Thread* current, oop obj) { 651 assert(UseObjectMonitorTable, "must be"); 652 653 markWord mark = obj->mark_acquire(); 654 for (;;) { 655 intptr_t hash = mark.hash(); 656 if (hash != 0) { 657 return hash; 658 } 659 660 hash = get_next_hash(current, obj); 661 const markWord old_mark = mark; 662 const markWord new_mark = old_mark.copy_set_hash(hash); 663 664 mark = obj->cas_set_mark(new_mark, old_mark); 665 if (old_mark == mark) { 666 return hash; 667 } 668 } 669 } 670 671 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 672 if (UseObjectMonitorTable) { 673 // Since the monitor isn't in the object header, the hash can simply be 674 // installed in the object header. 675 return install_hash_code(current, obj); 676 } 677 678 while (true) { 679 ObjectMonitor* monitor = nullptr; 680 markWord temp, test; 681 intptr_t hash; 682 markWord mark = obj->mark_acquire(); 683 if (mark.is_unlocked() || mark.is_fast_locked()) { 684 hash = mark.hash(); 685 if (hash != 0) { // if it has a hash, just return it 686 return hash; 687 } 688 hash = get_next_hash(current, obj); // get a new hash 689 temp = mark.copy_set_hash(hash); // merge the hash into header 690 // try to install the hash 691 test = obj->cas_set_mark(temp, mark); 692 if (test == mark) { // if the hash was installed, return it 693 return hash; 694 } 695 // CAS failed, retry 696 continue; 697 698 // Failed to install the hash. It could be that another thread 699 // installed the hash just before our attempt or inflation has 700 // occurred or... so we fall thru to inflate the monitor for 701 // stability and then install the hash. 702 } else if (mark.has_monitor()) { 703 monitor = mark.monitor(); 704 temp = monitor->header(); 705 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 706 hash = temp.hash(); 707 if (hash != 0) { 708 // It has a hash. 709 710 // Separate load of dmw/header above from the loads in 711 // is_being_async_deflated(). 712 713 // dmw/header and _contentions may get written by different threads. 714 // Make sure to observe them in the same order when having several observers. 715 OrderAccess::loadload_for_IRIW(); 716 717 if (monitor->is_being_async_deflated()) { 718 // But we can't safely use the hash if we detect that async 719 // deflation has occurred. So we attempt to restore the 720 // header/dmw to the object's header so that we only retry 721 // once if the deflater thread happens to be slow. 722 monitor->install_displaced_markword_in_object(obj); 723 continue; 724 } 725 return hash; 726 } 727 // Fall thru so we only have one place that installs the hash in 728 // the ObjectMonitor. 729 } 730 731 // NOTE: an async deflation can race after we get the monitor and 732 // before we can update the ObjectMonitor's header with the hash 733 // value below. 734 assert(mark.has_monitor(), "must be"); 735 monitor = mark.monitor(); 736 737 // Load ObjectMonitor's header/dmw field and see if it has a hash. 738 mark = monitor->header(); 739 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 740 hash = mark.hash(); 741 if (hash == 0) { // if it does not have a hash 742 hash = get_next_hash(current, obj); // get a new hash 743 temp = mark.copy_set_hash(hash) ; // merge the hash into header 744 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 745 uintptr_t v = AtomicAccess::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value()); 746 test = markWord(v); 747 if (test != mark) { 748 // The attempt to update the ObjectMonitor's header/dmw field 749 // did not work. This can happen if another thread managed to 750 // merge in the hash just before our cmpxchg(). 751 // If we add any new usages of the header/dmw field, this code 752 // will need to be updated. 753 hash = test.hash(); 754 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 755 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 756 } 757 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) { 758 // If we detect that async deflation has occurred, then we 759 // attempt to restore the header/dmw to the object's header 760 // so that we only retry once if the deflater thread happens 761 // to be slow. 762 monitor->install_displaced_markword_in_object(obj); 763 continue; 764 } 765 } 766 // We finally get the hash. 767 return hash; 768 } 769 } 770 771 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 772 Handle h_obj) { 773 assert(current == JavaThread::current(), "Can only be called on current thread"); 774 oop obj = h_obj(); 775 776 markWord mark = obj->mark_acquire(); 777 778 if (mark.is_fast_locked()) { 779 // fast-locking case, see if lock is in current's lock stack 780 return current->lock_stack().contains(h_obj()); 781 } 782 783 while (mark.has_monitor()) { 784 ObjectMonitor* monitor = read_monitor(current, obj, mark); 785 if (monitor != nullptr) { 786 return monitor->is_entered(current) != 0; 787 } 788 // Racing with inflation/deflation, retry 789 mark = obj->mark_acquire(); 790 791 if (mark.is_fast_locked()) { 792 // Some other thread fast_locked, current could not have held the lock 793 return false; 794 } 795 } 796 797 // Unlocked case, header in place 798 assert(mark.is_unlocked(), "sanity check"); 799 return false; 800 } 801 802 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 803 oop obj = h_obj(); 804 markWord mark = obj->mark_acquire(); 805 806 if (mark.is_fast_locked()) { 807 // fast-locked so get owner from the object. 808 // owning_thread_from_object() may also return null here: 809 return Threads::owning_thread_from_object(t_list, h_obj()); 810 } 811 812 while (mark.has_monitor()) { 813 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark); 814 if (monitor != nullptr) { 815 return Threads::owning_thread_from_monitor(t_list, monitor); 816 } 817 // Racing with inflation/deflation, retry 818 mark = obj->mark_acquire(); 819 820 if (mark.is_fast_locked()) { 821 // Some other thread fast_locked 822 return Threads::owning_thread_from_object(t_list, h_obj()); 823 } 824 } 825 826 // Unlocked case, header in place 827 // Cannot have assertion since this object may have been 828 // locked by another thread when reaching here. 829 // assert(mark.is_unlocked(), "sanity check"); 830 831 return nullptr; 832 } 833 834 // Visitors ... 835 836 // Iterate over all ObjectMonitors. 837 template <typename Function> 838 void ObjectSynchronizer::monitors_iterate(Function function) { 839 MonitorList::Iterator iter = _in_use_list.iterator(); 840 while (iter.has_next()) { 841 ObjectMonitor* monitor = iter.next(); 842 function(monitor); 843 } 844 } 845 846 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 847 // returns true. 848 template <typename OwnerFilter> 849 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 850 monitors_iterate([&](ObjectMonitor* monitor) { 851 // This function is only called at a safepoint or when the 852 // target thread is suspended or when the target thread is 853 // operating on itself. The current closures in use today are 854 // only interested in an owned ObjectMonitor and ownership 855 // cannot be dropped under the calling contexts so the 856 // ObjectMonitor cannot be async deflated. 857 if (monitor->has_owner() && filter(monitor)) { 858 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 859 860 closure->do_monitor(monitor); 861 } 862 }); 863 } 864 865 // Iterate ObjectMonitors where the owner == thread; this does NOT include 866 // ObjectMonitors where owner is set to a stack-lock address in thread. 867 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 868 int64_t key = ObjectMonitor::owner_id_from(thread); 869 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 870 return owned_monitors_iterate_filtered(closure, thread_filter); 871 } 872 873 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) { 874 int64_t key = ObjectMonitor::owner_id_from(vthread); 875 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 876 return owned_monitors_iterate_filtered(closure, thread_filter); 877 } 878 879 // Iterate ObjectMonitors owned by any thread. 880 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 881 auto all_filter = [&](ObjectMonitor* monitor) { return true; }; 882 return owned_monitors_iterate_filtered(closure, all_filter); 883 } 884 885 static bool monitors_used_above_threshold(MonitorList* list) { 886 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 887 return false; 888 } 889 size_t monitors_used = list->count(); 890 if (monitors_used == 0) { // empty list is easy 891 return false; 892 } 893 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling(); 894 // Make sure that we use a ceiling value that is not lower than 895 // previous, not lower than the recorded max used by the system, and 896 // not lower than the current number of monitors in use (which can 897 // race ahead of max). The result is guaranteed > 0. 898 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used); 899 900 // Check if our monitor usage is above the threshold: 901 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 902 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 903 // Deflate monitors if over the threshold percentage, unless no 904 // progress on previous deflations. 905 bool is_above_threshold = true; 906 907 // Check if it's time to adjust the in_use_list_ceiling up, due 908 // to too many async deflation attempts without any progress. 909 if (NoAsyncDeflationProgressMax != 0 && 910 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 911 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 912 size_t delta = (size_t)(ceiling * remainder) + 1; 913 size_t new_ceiling = (ceiling > SIZE_MAX - delta) 914 ? SIZE_MAX // Overflow, let's clamp new_ceiling. 915 : ceiling + delta; 916 917 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 918 log_info(monitorinflation)("Too many deflations without progress; " 919 "bumping in_use_list_ceiling from %zu" 920 " to %zu", old_ceiling, new_ceiling); 921 _no_progress_cnt = 0; 922 ceiling = new_ceiling; 923 924 // Check if our monitor usage is still above the threshold: 925 monitor_usage = (monitors_used * 100LL) / ceiling; 926 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold; 927 } 928 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu" 929 ", monitor_usage=%zu, threshold=%d", 930 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 931 return is_above_threshold; 932 } 933 934 return false; 935 } 936 937 size_t ObjectSynchronizer::in_use_list_count() { 938 return _in_use_list.count(); 939 } 940 941 size_t ObjectSynchronizer::in_use_list_max() { 942 return _in_use_list.max(); 943 } 944 945 size_t ObjectSynchronizer::in_use_list_ceiling() { 946 return _in_use_list_ceiling; 947 } 948 949 void ObjectSynchronizer::dec_in_use_list_ceiling() { 950 AtomicAccess::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 951 } 952 953 void ObjectSynchronizer::inc_in_use_list_ceiling() { 954 AtomicAccess::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 955 } 956 957 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 958 _in_use_list_ceiling = new_value; 959 } 960 961 bool ObjectSynchronizer::is_async_deflation_needed() { 962 if (is_async_deflation_requested()) { 963 // Async deflation request. 964 log_info(monitorinflation)("Async deflation needed: explicit request"); 965 return true; 966 } 967 968 jlong time_since_last = time_since_last_async_deflation_ms(); 969 970 if (AsyncDeflationInterval > 0 && 971 time_since_last > AsyncDeflationInterval && 972 monitors_used_above_threshold(&_in_use_list)) { 973 // It's been longer than our specified deflate interval and there 974 // are too many monitors in use. We don't deflate more frequently 975 // than AsyncDeflationInterval (unless is_async_deflation_requested) 976 // in order to not swamp the MonitorDeflationThread. 977 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 978 return true; 979 } 980 981 if (GuaranteedAsyncDeflationInterval > 0 && 982 time_since_last > GuaranteedAsyncDeflationInterval) { 983 // It's been longer than our specified guaranteed deflate interval. 984 // We need to clean up the used monitors even if the threshold is 985 // not reached, to keep the memory utilization at bay when many threads 986 // touched many monitors. 987 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) " 988 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 989 GuaranteedAsyncDeflationInterval, time_since_last); 990 991 // If this deflation has no progress, then it should not affect the no-progress 992 // tracking, otherwise threshold heuristics would think it was triggered, experienced 993 // no progress, and needs to backoff more aggressively. In this "no progress" case, 994 // the generic code would bump the no-progress counter, and we compensate for that 995 // by telling it to skip the update. 996 // 997 // If this deflation has progress, then it should let non-progress tracking 998 // know about this, otherwise the threshold heuristics would kick in, potentially 999 // experience no-progress due to aggressive cleanup by this deflation, and think 1000 // it is still in no-progress stride. In this "progress" case, the generic code would 1001 // zero the counter, and we allow it to happen. 1002 _no_progress_skip_increment = true; 1003 1004 return true; 1005 } 1006 1007 return false; 1008 } 1009 1010 void ObjectSynchronizer::request_deflate_idle_monitors() { 1011 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1012 set_is_async_deflation_requested(true); 1013 ml.notify_all(); 1014 } 1015 1016 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 1017 JavaThread* current = JavaThread::current(); 1018 bool ret_code = false; 1019 1020 jlong last_time = last_async_deflation_time_ns(); 1021 1022 request_deflate_idle_monitors(); 1023 1024 const int N_CHECKS = 5; 1025 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1026 if (last_async_deflation_time_ns() > last_time) { 1027 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1028 ret_code = true; 1029 break; 1030 } 1031 { 1032 // JavaThread has to honor the blocking protocol. 1033 ThreadBlockInVM tbivm(current); 1034 os::naked_short_sleep(999); // sleep for almost 1 second 1035 } 1036 } 1037 if (!ret_code) { 1038 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1039 } 1040 1041 return ret_code; 1042 } 1043 1044 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1045 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1046 } 1047 1048 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1049 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1050 // 1051 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1052 MonitorList::Iterator iter = _in_use_list.iterator(); 1053 size_t deflated_count = 0; 1054 Thread* current = Thread::current(); 1055 1056 while (iter.has_next()) { 1057 if (deflated_count >= (size_t)MonitorDeflationMax) { 1058 break; 1059 } 1060 ObjectMonitor* mid = iter.next(); 1061 if (mid->deflate_monitor(current)) { 1062 deflated_count++; 1063 } 1064 1065 // Must check for a safepoint/handshake and honor it. 1066 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1067 } 1068 1069 return deflated_count; 1070 } 1071 1072 class DeflationHandshakeClosure : public HandshakeClosure { 1073 public: 1074 DeflationHandshakeClosure() : HandshakeClosure("DeflationHandshakeClosure") {} 1075 1076 void do_thread(Thread* thread) { 1077 log_trace(monitorinflation)("DeflationHandshakeClosure::do_thread: thread=" 1078 INTPTR_FORMAT, p2i(thread)); 1079 if (thread->is_Java_thread()) { 1080 // Clear OM cache 1081 JavaThread* jt = JavaThread::cast(thread); 1082 jt->om_clear_monitor_cache(); 1083 } 1084 } 1085 }; 1086 1087 class VM_RendezvousGCThreads : public VM_Operation { 1088 public: 1089 bool evaluate_at_safepoint() const override { return false; } 1090 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1091 void doit() override { 1092 Universe::heap()->safepoint_synchronize_begin(); 1093 Universe::heap()->safepoint_synchronize_end(); 1094 }; 1095 }; 1096 1097 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1098 ObjectMonitorDeflationSafepointer* safepointer) { 1099 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1100 size_t deleted_count = 0; 1101 for (ObjectMonitor* monitor: *delete_list) { 1102 delete monitor; 1103 deleted_count++; 1104 // A JavaThread must check for a safepoint/handshake and honor it. 1105 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1106 } 1107 return deleted_count; 1108 } 1109 1110 class ObjectMonitorDeflationLogging: public StackObj { 1111 LogStreamHandle(Debug, monitorinflation) _debug; 1112 LogStreamHandle(Info, monitorinflation) _info; 1113 LogStream* _stream; 1114 elapsedTimer _timer; 1115 1116 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1117 size_t count() const { return ObjectSynchronizer::in_use_list_count(); } 1118 size_t max() const { return ObjectSynchronizer::in_use_list_max(); } 1119 1120 public: 1121 ObjectMonitorDeflationLogging() 1122 : _debug(), _info(), _stream(nullptr) { 1123 if (_debug.is_enabled()) { 1124 _stream = &_debug; 1125 } else if (_info.is_enabled()) { 1126 _stream = &_info; 1127 } 1128 } 1129 1130 void begin() { 1131 if (_stream != nullptr) { 1132 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1133 ceiling(), count(), max()); 1134 _timer.start(); 1135 } 1136 } 1137 1138 void before_handshake(size_t unlinked_count) { 1139 if (_stream != nullptr) { 1140 _timer.stop(); 1141 _stream->print_cr("before handshaking: unlinked_count=%zu" 1142 ", in_use_list stats: ceiling=%zu, count=" 1143 "%zu, max=%zu", 1144 unlinked_count, ceiling(), count(), max()); 1145 } 1146 } 1147 1148 void after_handshake() { 1149 if (_stream != nullptr) { 1150 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1151 "%zu, count=%zu, max=%zu", 1152 ceiling(), count(), max()); 1153 _timer.start(); 1154 } 1155 } 1156 1157 void end(size_t deflated_count, size_t unlinked_count) { 1158 if (_stream != nullptr) { 1159 _timer.stop(); 1160 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1161 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs", 1162 deflated_count, unlinked_count, _timer.seconds()); 1163 } 1164 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1165 ceiling(), count(), max()); 1166 } 1167 } 1168 1169 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1170 if (_stream != nullptr) { 1171 _timer.stop(); 1172 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling=" 1173 "%zu, count=%zu, max=%zu", 1174 op_name, cnt_name, cnt, ceiling(), count(), max()); 1175 } 1176 } 1177 1178 void after_block_for_safepoint(const char* op_name) { 1179 if (_stream != nullptr) { 1180 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu" 1181 ", count=%zu, max=%zu", op_name, 1182 ceiling(), count(), max()); 1183 _timer.start(); 1184 } 1185 } 1186 }; 1187 1188 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1189 if (!SafepointMechanism::should_process(_current)) { 1190 return; 1191 } 1192 1193 // A safepoint/handshake has started. 1194 _log->before_block_for_safepoint(op_name, count_name, counter); 1195 1196 { 1197 // Honor block request. 1198 ThreadBlockInVM tbivm(_current); 1199 } 1200 1201 _log->after_block_for_safepoint(op_name); 1202 } 1203 1204 // This function is called by the MonitorDeflationThread to deflate 1205 // ObjectMonitors. 1206 size_t ObjectSynchronizer::deflate_idle_monitors() { 1207 JavaThread* current = JavaThread::current(); 1208 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1209 1210 // The async deflation request has been processed. 1211 _last_async_deflation_time_ns = os::javaTimeNanos(); 1212 set_is_async_deflation_requested(false); 1213 1214 ObjectMonitorDeflationLogging log; 1215 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1216 1217 log.begin(); 1218 1219 // Deflate some idle ObjectMonitors. 1220 size_t deflated_count = deflate_monitor_list(&safepointer); 1221 1222 // Unlink the deflated ObjectMonitors from the in-use list. 1223 size_t unlinked_count = 0; 1224 size_t deleted_count = 0; 1225 if (deflated_count > 0) { 1226 ResourceMark rm(current); 1227 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1228 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1229 1230 #ifdef ASSERT 1231 if (UseObjectMonitorTable) { 1232 for (ObjectMonitor* monitor : delete_list) { 1233 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed"); 1234 } 1235 } 1236 #endif 1237 1238 log.before_handshake(unlinked_count); 1239 1240 // A JavaThread needs to handshake in order to safely free the 1241 // ObjectMonitors that were deflated in this cycle. 1242 DeflationHandshakeClosure dhc; 1243 Handshake::execute(&dhc); 1244 // Also, we sync and desync GC threads around the handshake, so that they can 1245 // safely read the mark-word and look-through to the object-monitor, without 1246 // being afraid that the object-monitor is going away. 1247 VM_RendezvousGCThreads sync_gc; 1248 VMThread::execute(&sync_gc); 1249 1250 log.after_handshake(); 1251 1252 // After the handshake, safely free the ObjectMonitors that were 1253 // deflated and unlinked in this cycle. 1254 1255 // Delete the unlinked ObjectMonitors. 1256 deleted_count = delete_monitors(&delete_list, &safepointer); 1257 assert(unlinked_count == deleted_count, "must be"); 1258 } 1259 1260 log.end(deflated_count, unlinked_count); 1261 1262 GVars.stw_random = os::random(); 1263 1264 if (deflated_count != 0) { 1265 _no_progress_cnt = 0; 1266 } else if (_no_progress_skip_increment) { 1267 _no_progress_skip_increment = false; 1268 } else { 1269 _no_progress_cnt++; 1270 } 1271 1272 return deflated_count; 1273 } 1274 1275 // Monitor cleanup on JavaThread::exit 1276 1277 // Iterate through monitor cache and attempt to release thread's monitors 1278 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1279 private: 1280 JavaThread* _thread; 1281 1282 public: 1283 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1284 void do_monitor(ObjectMonitor* mid) { 1285 mid->complete_exit(_thread); 1286 } 1287 }; 1288 1289 // Release all inflated monitors owned by current thread. Lightweight monitors are 1290 // ignored. This is meant to be called during JNI thread detach which assumes 1291 // all remaining monitors are heavyweight. All exceptions are swallowed. 1292 // Scanning the extant monitor list can be time consuming. 1293 // A simple optimization is to add a per-thread flag that indicates a thread 1294 // called jni_monitorenter() during its lifetime. 1295 // 1296 // Instead of NoSafepointVerifier it might be cheaper to 1297 // use an idiom of the form: 1298 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1299 // <code that must not run at safepoint> 1300 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1301 // Since the tests are extremely cheap we could leave them enabled 1302 // for normal product builds. 1303 1304 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1305 assert(current == JavaThread::current(), "must be current Java thread"); 1306 NoSafepointVerifier nsv; 1307 ReleaseJavaMonitorsClosure rjmc(current); 1308 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1309 assert(!current->has_pending_exception(), "Should not be possible"); 1310 current->clear_pending_exception(); 1311 } 1312 1313 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1314 switch (cause) { 1315 case inflate_cause_vm_internal: return "VM Internal"; 1316 case inflate_cause_monitor_enter: return "Monitor Enter"; 1317 case inflate_cause_wait: return "Monitor Wait"; 1318 case inflate_cause_notify: return "Monitor Notify"; 1319 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1320 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1321 default: 1322 ShouldNotReachHere(); 1323 } 1324 return "Unknown"; 1325 } 1326 1327 //------------------------------------------------------------------------------ 1328 // Debugging code 1329 1330 u_char* ObjectSynchronizer::get_gvars_addr() { 1331 return (u_char*)&GVars; 1332 } 1333 1334 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1335 return (u_char*)&GVars.hc_sequence; 1336 } 1337 1338 size_t ObjectSynchronizer::get_gvars_size() { 1339 return sizeof(SharedGlobals); 1340 } 1341 1342 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1343 return (u_char*)&GVars.stw_random; 1344 } 1345 1346 // Do the final audit and print of ObjectMonitor stats; must be done 1347 // by the VMThread at VM exit time. 1348 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1349 assert(Thread::current()->is_VM_thread(), "sanity check"); 1350 1351 if (is_final_audit()) { // Only do the audit once. 1352 return; 1353 } 1354 set_is_final_audit(); 1355 log_info(monitorinflation)("Starting the final audit."); 1356 1357 if (log_is_enabled(Info, monitorinflation)) { 1358 LogStreamHandle(Info, monitorinflation) ls; 1359 audit_and_print_stats(&ls, true /* on_exit */); 1360 } 1361 } 1362 1363 // This function can be called by the MonitorDeflationThread or it can be called when 1364 // we are trying to exit the VM. The list walker functions can run in parallel with 1365 // the other list operations. 1366 // Calls to this function can be added in various places as a debugging 1367 // aid. 1368 // 1369 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) { 1370 int error_cnt = 0; 1371 1372 ls->print_cr("Checking in_use_list:"); 1373 chk_in_use_list(ls, &error_cnt); 1374 1375 if (error_cnt == 0) { 1376 ls->print_cr("No errors found in in_use_list checks."); 1377 } else { 1378 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 1379 } 1380 1381 // When exiting, only log the interesting entries at the Info level. 1382 // When called at intervals by the MonitorDeflationThread, log output 1383 // at the Trace level since there can be a lot of it. 1384 if (!on_exit && log_is_enabled(Trace, monitorinflation)) { 1385 LogStreamHandle(Trace, monitorinflation) ls_tr; 1386 log_in_use_monitor_details(&ls_tr, true /* log_all */); 1387 } else if (on_exit) { 1388 log_in_use_monitor_details(ls, false /* log_all */); 1389 } 1390 1391 ls->flush(); 1392 1393 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 1394 } 1395 1396 // Check the in_use_list; log the results of the checks. 1397 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 1398 size_t l_in_use_count = _in_use_list.count(); 1399 size_t l_in_use_max = _in_use_list.max(); 1400 out->print_cr("count=%zu, max=%zu", l_in_use_count, 1401 l_in_use_max); 1402 1403 size_t ck_in_use_count = 0; 1404 MonitorList::Iterator iter = _in_use_list.iterator(); 1405 while (iter.has_next()) { 1406 ObjectMonitor* mid = iter.next(); 1407 chk_in_use_entry(mid, out, error_cnt_p); 1408 ck_in_use_count++; 1409 } 1410 1411 if (l_in_use_count == ck_in_use_count) { 1412 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu", 1413 l_in_use_count, ck_in_use_count); 1414 } else { 1415 out->print_cr("WARNING: in_use_count=%zu is not equal to " 1416 "ck_in_use_count=%zu", l_in_use_count, 1417 ck_in_use_count); 1418 } 1419 1420 size_t ck_in_use_max = _in_use_list.max(); 1421 if (l_in_use_max == ck_in_use_max) { 1422 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu", 1423 l_in_use_max, ck_in_use_max); 1424 } else { 1425 out->print_cr("WARNING: in_use_max=%zu is not equal to " 1426 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max); 1427 } 1428 } 1429 1430 // Check an in-use monitor entry; log any errors. 1431 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 1432 int* error_cnt_p) { 1433 if (n->owner_is_DEFLATER_MARKER()) { 1434 // This could happen when monitor deflation blocks for a safepoint. 1435 return; 1436 } 1437 1438 1439 if (n->metadata() == 0) { 1440 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 1441 "have non-null _metadata (header/hash) field.", p2i(n)); 1442 *error_cnt_p = *error_cnt_p + 1; 1443 } 1444 1445 const oop obj = n->object_peek(); 1446 if (obj == nullptr) { 1447 return; 1448 } 1449 1450 const markWord mark = obj->mark(); 1451 if (!mark.has_monitor()) { 1452 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 1453 "object does not think it has a monitor: obj=" 1454 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 1455 p2i(obj), mark.value()); 1456 *error_cnt_p = *error_cnt_p + 1; 1457 return; 1458 } 1459 1460 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark); 1461 if (n != obj_mon) { 1462 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 1463 "object does not refer to the same monitor: obj=" 1464 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 1465 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 1466 *error_cnt_p = *error_cnt_p + 1; 1467 } 1468 } 1469 1470 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 1471 // flags indicate why the entry is in-use, 'object' and 'object type' 1472 // indicate the associated object and its type. 1473 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 1474 if (_in_use_list.count() > 0) { 1475 stringStream ss; 1476 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)"); 1477 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 1478 out->print_cr("%18s %s %18s %18s", 1479 "monitor", "BHL", "object", "object type"); 1480 out->print_cr("================== === ================== =================="); 1481 1482 auto is_interesting = [&](ObjectMonitor* monitor) { 1483 return log_all || monitor->has_owner() || monitor->is_busy(); 1484 }; 1485 1486 monitors_iterate([&](ObjectMonitor* monitor) { 1487 if (is_interesting(monitor)) { 1488 const oop obj = monitor->object_peek(); 1489 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash(); 1490 ResourceMark rm; 1491 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 1492 monitor->is_busy(), hash != 0, monitor->has_owner(), 1493 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 1494 if (monitor->is_busy()) { 1495 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 1496 ss.reset(); 1497 } 1498 out->cr(); 1499 } 1500 }); 1501 } 1502 1503 out->flush(); 1504 }