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/atomic.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.hpp" 48 #include "runtime/objectMonitor.inline.hpp" 49 #include "runtime/os.inline.hpp" 50 #include "runtime/osThread.hpp" 51 #include "runtime/safepointMechanism.inline.hpp" 52 #include "runtime/safepointVerifiers.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "runtime/stubRoutines.hpp" 55 #include "runtime/synchronizer.inline.hpp" 56 #include "runtime/threads.hpp" 57 #include "runtime/timer.hpp" 58 #include "runtime/trimNativeHeap.hpp" 59 #include "runtime/vframe.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "utilities/align.hpp" 62 #include "utilities/dtrace.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/globalCounter.inline.hpp" 65 #include "utilities/globalDefinitions.hpp" 66 #include "utilities/fastHash.hpp" 67 #include "utilities/linkedlist.hpp" 68 #include "utilities/preserveException.hpp" 69 70 class ObjectMonitorDeflationLogging; 71 72 void MonitorList::add(ObjectMonitor* m) { 73 ObjectMonitor* head; 74 do { 75 head = Atomic::load(&_head); 76 m->set_next_om(head); 77 } while (Atomic::cmpxchg(&_head, head, m) != head); 78 79 size_t count = Atomic::add(&_count, 1u, memory_order_relaxed); 80 size_t old_max; 81 do { 82 old_max = Atomic::load(&_max); 83 if (count <= old_max) { 84 break; 85 } 86 } while (Atomic::cmpxchg(&_max, old_max, count, memory_order_relaxed) != old_max); 87 } 88 89 size_t MonitorList::count() const { 90 return Atomic::load(&_count); 91 } 92 93 size_t MonitorList::max() const { 94 return Atomic::load(&_max); 95 } 96 97 class ObjectMonitorDeflationSafepointer : public StackObj { 98 JavaThread* const _current; 99 ObjectMonitorDeflationLogging* const _log; 100 101 public: 102 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 103 : _current(current), _log(log) {} 104 105 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 106 }; 107 108 // Walk the in-use list and unlink deflated ObjectMonitors. 109 // Returns the number of unlinked ObjectMonitors. 110 size_t MonitorList::unlink_deflated(size_t deflated_count, 111 GrowableArray<ObjectMonitor*>* unlinked_list, 112 ObjectMonitorDeflationSafepointer* safepointer) { 113 size_t unlinked_count = 0; 114 ObjectMonitor* prev = nullptr; 115 ObjectMonitor* m = Atomic::load_acquire(&_head); 116 117 while (m != nullptr) { 118 if (m->is_being_async_deflated()) { 119 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 120 // modify the list once per batch. The batch starts at "m". 121 size_t unlinked_batch = 0; 122 ObjectMonitor* next = m; 123 // Look for at most MonitorUnlinkBatch monitors, or the number of 124 // deflated and not unlinked monitors, whatever comes first. 125 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 126 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 127 do { 128 ObjectMonitor* next_next = next->next_om(); 129 unlinked_batch++; 130 unlinked_list->append(next); 131 next = next_next; 132 if (unlinked_batch >= unlinked_batch_limit) { 133 // Reached the max batch, so bail out of the gathering loop. 134 break; 135 } 136 if (prev == nullptr && Atomic::load(&_head) != m) { 137 // Current batch used to be at head, but it is not at head anymore. 138 // Bail out and figure out where we currently are. This avoids long 139 // walks searching for new prev during unlink under heavy list inserts. 140 break; 141 } 142 } while (next != nullptr && next->is_being_async_deflated()); 143 144 // Unlink the found batch. 145 if (prev == nullptr) { 146 // The current batch is the first batch, so there is a chance that it starts at head. 147 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 148 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next); 149 if (prev_head != m) { 150 // Something must have updated the head. Figure out the actual prev for this batch. 151 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 152 prev = n; 153 } 154 assert(prev != nullptr, "Should have found the prev for the current batch"); 155 prev->set_next_om(next); 156 } 157 } else { 158 // The current batch is preceded by another batch. This guarantees the current batch 159 // does not start at head. Unlink the entire current batch without updating the head. 160 assert(Atomic::load(&_head) != m, "Sanity"); 161 prev->set_next_om(next); 162 } 163 164 unlinked_count += unlinked_batch; 165 if (unlinked_count >= deflated_count) { 166 // Reached the max so bail out of the searching loop. 167 // There should be no more deflated monitors left. 168 break; 169 } 170 m = next; 171 } else { 172 prev = m; 173 m = m->next_om(); 174 } 175 176 // Must check for a safepoint/handshake and honor it. 177 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 178 } 179 180 #ifdef ASSERT 181 // Invariant: the code above should unlink all deflated monitors. 182 // The code that runs after this unlinking does not expect deflated monitors. 183 // Notably, attempting to deflate the already deflated monitor would break. 184 { 185 ObjectMonitor* m = Atomic::load_acquire(&_head); 186 while (m != nullptr) { 187 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 188 m = m->next_om(); 189 } 190 } 191 #endif 192 193 Atomic::sub(&_count, unlinked_count); 194 return unlinked_count; 195 } 196 197 MonitorList::Iterator MonitorList::iterator() const { 198 return Iterator(Atomic::load_acquire(&_head)); 199 } 200 201 ObjectMonitor* MonitorList::Iterator::next() { 202 ObjectMonitor* current = _current; 203 _current = current->next_om(); 204 return current; 205 } 206 207 // The "core" versions of monitor enter and exit reside in this file. 208 // The interpreter and compilers contain specialized transliterated 209 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 210 // fast_lock(...) for instance. If you make changes here, make sure to modify the 211 // interpreter, and both C1 and C2 fast-path inline locking code emission. 212 // 213 // ----------------------------------------------------------------------------- 214 215 #ifdef DTRACE_ENABLED 216 217 // Only bother with this argument setup if dtrace is available 218 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 219 220 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 221 char* bytes = nullptr; \ 222 int len = 0; \ 223 jlong jtid = SharedRuntime::get_java_tid(thread); \ 224 Symbol* klassname = obj->klass()->name(); \ 225 if (klassname != nullptr) { \ 226 bytes = (char*)klassname->bytes(); \ 227 len = klassname->utf8_length(); \ 228 } 229 230 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 231 { \ 232 if (DTraceMonitorProbes) { \ 233 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 234 HOTSPOT_MONITOR_WAIT(jtid, \ 235 (uintptr_t)(monitor), bytes, len, (millis)); \ 236 } \ 237 } 238 239 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 240 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 241 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 242 243 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 244 { \ 245 if (DTraceMonitorProbes) { \ 246 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 247 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 248 (uintptr_t)(monitor), bytes, len); \ 249 } \ 250 } 251 252 #else // ndef DTRACE_ENABLED 253 254 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 255 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 256 257 #endif // ndef DTRACE_ENABLED 258 259 // This exists only as a workaround of dtrace bug 6254741 260 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 261 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 262 return 0; 263 } 264 265 static constexpr size_t inflation_lock_count() { 266 return 256; 267 } 268 269 // Static storage for an array of PlatformMutex. 270 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 271 272 static inline PlatformMutex* inflation_lock(size_t index) { 273 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 274 } 275 276 void ObjectSynchronizer::initialize() { 277 for (size_t i = 0; i < inflation_lock_count(); i++) { 278 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 279 } 280 // Start the ceiling with the estimate for one thread. 281 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 282 283 // Start the timer for deflations, so it does not trigger immediately. 284 _last_async_deflation_time_ns = os::javaTimeNanos(); 285 286 if (LockingMode == LM_LIGHTWEIGHT) { 287 LightweightSynchronizer::initialize(); 288 } 289 } 290 291 MonitorList ObjectSynchronizer::_in_use_list; 292 // monitors_used_above_threshold() policy is as follows: 293 // 294 // The ratio of the current _in_use_list count to the ceiling is used 295 // to determine if we are above MonitorUsedDeflationThreshold and need 296 // to do an async monitor deflation cycle. The ceiling is increased by 297 // AvgMonitorsPerThreadEstimate when a thread is added to the system 298 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 299 // removed from the system. 300 // 301 // Note: If the _in_use_list max exceeds the ceiling, then 302 // monitors_used_above_threshold() will use the in_use_list max instead 303 // of the thread count derived ceiling because we have used more 304 // ObjectMonitors than the estimated average. 305 // 306 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 307 // no-progress async monitor deflation cycles in a row, then the ceiling 308 // is adjusted upwards by monitors_used_above_threshold(). 309 // 310 // Start the ceiling with the estimate for one thread in initialize() 311 // which is called after cmd line options are processed. 312 static size_t _in_use_list_ceiling = 0; 313 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 314 bool volatile ObjectSynchronizer::_is_final_audit = false; 315 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 316 static uintx _no_progress_cnt = 0; 317 static bool _no_progress_skip_increment = false; 318 319 // =====================> Quick functions 320 321 // The quick_* forms are special fast-path variants used to improve 322 // performance. In the simplest case, a "quick_*" implementation could 323 // simply return false, in which case the caller will perform the necessary 324 // state transitions and call the slow-path form. 325 // The fast-path is designed to handle frequently arising cases in an efficient 326 // manner and is just a degenerate "optimistic" variant of the slow-path. 327 // returns true -- to indicate the call was satisfied. 328 // returns false -- to indicate the call needs the services of the slow-path. 329 // A no-loitering ordinance is in effect for code in the quick_* family 330 // operators: safepoints or indefinite blocking (blocking that might span a 331 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 332 // entry. 333 // 334 // Consider: An interesting optimization is to have the JIT recognize the 335 // following common idiom: 336 // synchronized (someobj) { .... ; notify(); } 337 // That is, we find a notify() or notifyAll() call that immediately precedes 338 // the monitorexit operation. In that case the JIT could fuse the operations 339 // into a single notifyAndExit() runtime primitive. 340 341 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 342 assert(current->thread_state() == _thread_in_Java, "invariant"); 343 NoSafepointVerifier nsv; 344 if (obj == nullptr) return false; // slow-path for invalid obj 345 const markWord mark = obj->mark(); 346 347 if (LockingMode == LM_LIGHTWEIGHT) { 348 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 349 // Degenerate notify 350 // fast-locked by caller so by definition the implied waitset is empty. 351 return true; 352 } 353 } else if (LockingMode == LM_LEGACY) { 354 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 355 // Degenerate notify 356 // stack-locked by caller so by definition the implied waitset is empty. 357 return true; 358 } 359 } 360 361 if (mark.has_monitor()) { 362 ObjectMonitor* const mon = read_monitor(current, obj, mark); 363 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) { 364 // Racing with inflation/deflation go slow path 365 return false; 366 } 367 assert(mon->object() == oop(obj), "invariant"); 368 if (!mon->has_owner(current)) return false; // slow-path for IMS exception 369 370 if (mon->first_waiter() != nullptr) { 371 // We have one or more waiters. Since this is an inflated monitor 372 // that we own, we quickly notify them here and now, avoiding the slow-path. 373 if (all) { 374 mon->quick_notifyAll(current); 375 } else { 376 mon->quick_notify(current); 377 } 378 } 379 return true; 380 } 381 382 // other IMS exception states take the slow-path 383 return false; 384 } 385 386 static bool useHeavyMonitors() { 387 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390) 388 return LockingMode == LM_MONITOR; 389 #else 390 return false; 391 #endif 392 } 393 394 // The LockNode emitted directly at the synchronization site would have 395 // been too big if it were to have included support for the cases of inflated 396 // recursive enter and exit, so they go here instead. 397 // Note that we can't safely call AsyncPrintJavaStack() from within 398 // quick_enter() as our thread state remains _in_Java. 399 400 bool ObjectSynchronizer::quick_enter_legacy(oop obj, BasicLock* lock, JavaThread* current) { 401 assert(current->thread_state() == _thread_in_Java, "invariant"); 402 403 if (useHeavyMonitors()) { 404 return false; // Slow path 405 } 406 407 assert(LockingMode == LM_LEGACY, "legacy mode below"); 408 409 const markWord mark = obj->mark(); 410 411 if (mark.has_monitor()) { 412 413 ObjectMonitor* const m = read_monitor(mark); 414 // An async deflation or GC can race us before we manage to make 415 // the ObjectMonitor busy by setting the owner below. If we detect 416 // that race we just bail out to the slow-path here. 417 if (m->object_peek() == nullptr) { 418 return false; 419 } 420 421 // Lock contention and Transactional Lock Elision (TLE) diagnostics 422 // and observability 423 // Case: light contention possibly amenable to TLE 424 // Case: TLE inimical operations such as nested/recursive synchronization 425 426 if (m->has_owner(current)) { 427 m->increment_recursions(current); 428 current->inc_held_monitor_count(); 429 return true; 430 } 431 432 // This Java Monitor is inflated so obj's header will never be 433 // displaced to this thread's BasicLock. Make the displaced header 434 // non-null so this BasicLock is not seen as recursive nor as 435 // being locked. We do this unconditionally so that this thread's 436 // BasicLock cannot be mis-interpreted by any stack walkers. For 437 // performance reasons, stack walkers generally first check for 438 // stack-locking in the object's header, the second check is for 439 // recursive stack-locking in the displaced header in the BasicLock, 440 // and last are the inflated Java Monitor (ObjectMonitor) checks. 441 lock->set_displaced_header(markWord::unused_mark()); 442 443 if (!m->has_owner() && m->try_set_owner(current)) { 444 assert(m->recursions() == 0, "invariant"); 445 current->inc_held_monitor_count(); 446 return true; 447 } 448 } 449 450 // Note that we could inflate in quick_enter. 451 // This is likely a useful optimization 452 // Critically, in quick_enter() we must not: 453 // -- block indefinitely, or 454 // -- reach a safepoint 455 456 return false; // revert to slow-path 457 } 458 459 // Handle notifications when synchronizing on value based classes 460 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 461 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 462 frame last_frame = locking_thread->last_frame(); 463 bool bcp_was_adjusted = false; 464 // Don't decrement bcp if it points to the frame's first instruction. This happens when 465 // handle_sync_on_value_based_class() is called because of a synchronized method. There 466 // is no actual monitorenter instruction in the byte code in this case. 467 if (last_frame.is_interpreted_frame() && 468 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 469 // adjust bcp to point back to monitorenter so that we print the correct line numbers 470 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 471 bcp_was_adjusted = true; 472 } 473 474 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 475 ResourceMark rm; 476 stringStream ss; 477 locking_thread->print_active_stack_on(&ss); 478 char* base = (char*)strstr(ss.base(), "at"); 479 char* newline = (char*)strchr(ss.base(), '\n'); 480 if (newline != nullptr) { 481 *newline = '\0'; 482 } 483 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 484 } else { 485 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 486 ResourceMark rm; 487 Log(valuebasedclasses) vblog; 488 489 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 490 if (locking_thread->has_last_Java_frame()) { 491 LogStream info_stream(vblog.info()); 492 locking_thread->print_active_stack_on(&info_stream); 493 } else { 494 vblog.info("Cannot find the last Java frame"); 495 } 496 497 EventSyncOnValueBasedClass event; 498 if (event.should_commit()) { 499 event.set_valueBasedClass(obj->klass()); 500 event.commit(); 501 } 502 } 503 504 if (bcp_was_adjusted) { 505 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 506 } 507 } 508 509 // ----------------------------------------------------------------------------- 510 // Monitor Enter/Exit 511 512 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 513 // When called with locking_thread != Thread::current() some mechanism must synchronize 514 // the locking_thread with respect to the current thread. Currently only used when 515 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 516 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 517 518 if (LockingMode == LM_LIGHTWEIGHT) { 519 return LightweightSynchronizer::enter_for(obj, lock, locking_thread); 520 } 521 522 if (!enter_fast_impl(obj, lock, locking_thread)) { 523 // Inflated ObjectMonitor::enter_for is required 524 525 // An async deflation can race after the inflate_for() call and before 526 // enter_for() can make the ObjectMonitor busy. enter_for() returns false 527 // if we have lost the race to async deflation and we simply try again. 528 while (true) { 529 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter); 530 if (monitor->enter_for(locking_thread)) { 531 return; 532 } 533 assert(monitor->is_being_async_deflated(), "must be"); 534 } 535 } 536 } 537 538 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) { 539 if (!enter_fast_impl(obj, lock, current)) { 540 // Inflated ObjectMonitor::enter is required 541 542 // An async deflation can race after the inflate() call and before 543 // enter() can make the ObjectMonitor busy. enter() returns false if 544 // we have lost the race to async deflation and we simply try again. 545 while (true) { 546 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter); 547 if (monitor->enter(current)) { 548 return; 549 } 550 } 551 } 552 } 553 554 // The interpreter and compiler assembly code tries to lock using the fast path 555 // of this algorithm. Make sure to update that code if the following function is 556 // changed. The implementation is extremely sensitive to race condition. Be careful. 557 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 558 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer"); 559 560 if (obj->klass()->is_value_based()) { 561 handle_sync_on_value_based_class(obj, locking_thread); 562 } 563 564 locking_thread->inc_held_monitor_count(); 565 566 if (!useHeavyMonitors()) { 567 if (LockingMode == LM_LEGACY) { 568 markWord mark = obj->mark(); 569 if (mark.is_unlocked()) { 570 // Anticipate successful CAS -- the ST of the displaced mark must 571 // be visible <= the ST performed by the CAS. 572 lock->set_displaced_header(mark); 573 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 574 return true; 575 } 576 } else if (mark.has_locker() && 577 locking_thread->is_lock_owned((address) mark.locker())) { 578 assert(lock != mark.locker(), "must not re-lock the same lock"); 579 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock"); 580 lock->set_displaced_header(markWord::from_pointer(nullptr)); 581 return true; 582 } 583 584 // The object header will never be displaced to this lock, 585 // so it does not matter what the value is, except that it 586 // must be non-zero to avoid looking like a re-entrant lock, 587 // and must not look locked either. 588 lock->set_displaced_header(markWord::unused_mark()); 589 590 // Failed to fast lock. 591 return false; 592 } 593 } else if (VerifyHeavyMonitors) { 594 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 595 } 596 597 return false; 598 } 599 600 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) { 601 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer"); 602 603 if (!useHeavyMonitors()) { 604 markWord mark = object->mark(); 605 if (LockingMode == LM_LEGACY) { 606 markWord dhw = lock->displaced_header(); 607 if (dhw.value() == 0) { 608 // If the displaced header is null, then this exit matches up with 609 // a recursive enter. No real work to do here except for diagnostics. 610 #ifndef PRODUCT 611 if (mark != markWord::INFLATING()) { 612 // Only do diagnostics if we are not racing an inflation. Simply 613 // exiting a recursive enter of a Java Monitor that is being 614 // inflated is safe; see the has_monitor() comment below. 615 assert(!mark.is_unlocked(), "invariant"); 616 assert(!mark.has_locker() || 617 current->is_lock_owned((address)mark.locker()), "invariant"); 618 if (mark.has_monitor()) { 619 // The BasicLock's displaced_header is marked as a recursive 620 // enter and we have an inflated Java Monitor (ObjectMonitor). 621 // This is a special case where the Java Monitor was inflated 622 // after this thread entered the stack-lock recursively. When a 623 // Java Monitor is inflated, we cannot safely walk the Java 624 // Monitor owner's stack and update the BasicLocks because a 625 // Java Monitor can be asynchronously inflated by a thread that 626 // does not own the Java Monitor. 627 ObjectMonitor* m = read_monitor(mark); 628 assert(m->object()->mark() == mark, "invariant"); 629 assert(m->is_entered(current), "invariant"); 630 } 631 } 632 #endif 633 return; 634 } 635 636 if (mark == markWord::from_pointer(lock)) { 637 // If the object is stack-locked by the current thread, try to 638 // swing the displaced header from the BasicLock back to the mark. 639 assert(dhw.is_neutral(), "invariant"); 640 if (object->cas_set_mark(dhw, mark) == mark) { 641 return; 642 } 643 } 644 } 645 } else if (VerifyHeavyMonitors) { 646 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 647 } 648 649 // We have to take the slow-path of possible inflation and then exit. 650 // The ObjectMonitor* can't be async deflated until ownership is 651 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 652 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal); 653 assert(!monitor->has_anonymous_owner(), "must not be"); 654 monitor->exit(current); 655 } 656 657 // ----------------------------------------------------------------------------- 658 // JNI locks on java objects 659 // NOTE: must use heavy weight monitor to handle jni monitor enter 660 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 661 // Top native frames in the stack will not be seen if we attempt 662 // preemption, since we start walking from the last Java anchor. 663 NoPreemptMark npm(current); 664 665 if (obj->klass()->is_value_based()) { 666 handle_sync_on_value_based_class(obj, current); 667 } 668 669 // the current locking is from JNI instead of Java code 670 current->set_current_pending_monitor_is_from_java(false); 671 // An async deflation can race after the inflate() call and before 672 // enter() can make the ObjectMonitor busy. enter() returns false if 673 // we have lost the race to async deflation and we simply try again. 674 while (true) { 675 ObjectMonitor* monitor; 676 bool entered; 677 if (LockingMode == LM_LIGHTWEIGHT) { 678 BasicLock lock; 679 entered = LightweightSynchronizer::inflate_and_enter(obj(), &lock, inflate_cause_jni_enter, current, current) != nullptr; 680 } else { 681 monitor = inflate(current, obj(), inflate_cause_jni_enter); 682 entered = monitor->enter(current); 683 } 684 685 if (entered) { 686 current->inc_held_monitor_count(1, true); 687 break; 688 } 689 } 690 current->set_current_pending_monitor_is_from_java(true); 691 } 692 693 // NOTE: must use heavy weight monitor to handle jni monitor exit 694 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 695 JavaThread* current = THREAD; 696 697 ObjectMonitor* monitor; 698 if (LockingMode == LM_LIGHTWEIGHT) { 699 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK); 700 } else { 701 // The ObjectMonitor* can't be async deflated until ownership is 702 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 703 monitor = inflate(current, obj, inflate_cause_jni_exit); 704 } 705 // If this thread has locked the object, exit the monitor. We 706 // intentionally do not use CHECK on check_owner because we must exit the 707 // monitor even if an exception was already pending. 708 if (monitor->check_owner(THREAD)) { 709 monitor->exit(current); 710 current->dec_held_monitor_count(1, true); 711 } 712 } 713 714 // ----------------------------------------------------------------------------- 715 // Internal VM locks on java objects 716 // standard constructor, allows locking failures 717 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) { 718 _thread = thread; 719 _thread->check_for_valid_safepoint_state(); 720 _obj = obj; 721 722 if (_obj() != nullptr) { 723 ObjectSynchronizer::enter(_obj, &_lock, _thread); 724 } 725 } 726 727 ObjectLocker::~ObjectLocker() { 728 if (_obj() != nullptr) { 729 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 730 } 731 } 732 733 734 // ----------------------------------------------------------------------------- 735 // Wait/Notify/NotifyAll 736 // NOTE: must use heavy weight monitor to handle wait() 737 738 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 739 JavaThread* current = THREAD; 740 if (millis < 0) { 741 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 742 } 743 744 ObjectMonitor* monitor; 745 if (LockingMode == LM_LIGHTWEIGHT) { 746 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0); 747 } else { 748 // The ObjectMonitor* can't be async deflated because the _waiters 749 // field is incremented before ownership is dropped and decremented 750 // after ownership is regained. 751 monitor = inflate(current, obj(), inflate_cause_wait); 752 } 753 754 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 755 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 756 757 // This dummy call is in place to get around dtrace bug 6254741. Once 758 // that's fixed we can uncomment the following line, remove the call 759 // and change this function back into a "void" func. 760 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 761 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 762 return ret_code; 763 } 764 765 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 766 if (millis < 0) { 767 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 768 } 769 770 ObjectMonitor* monitor; 771 if (LockingMode == LM_LIGHTWEIGHT) { 772 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK); 773 } else { 774 monitor = inflate(THREAD, obj(), inflate_cause_wait); 775 } 776 monitor->wait(millis, false, THREAD); 777 } 778 779 780 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 781 JavaThread* current = THREAD; 782 783 markWord mark = obj->mark(); 784 if (LockingMode == LM_LIGHTWEIGHT) { 785 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 786 // Not inflated so there can't be any waiters to notify. 787 return; 788 } 789 } else if (LockingMode == LM_LEGACY) { 790 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 791 // Not inflated so there can't be any waiters to notify. 792 return; 793 } 794 } 795 796 ObjectMonitor* monitor; 797 if (LockingMode == LM_LIGHTWEIGHT) { 798 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 799 } else { 800 // The ObjectMonitor* can't be async deflated until ownership is 801 // dropped by the calling thread. 802 monitor = inflate(current, obj(), inflate_cause_notify); 803 } 804 monitor->notify(CHECK); 805 } 806 807 // NOTE: see comment of notify() 808 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 809 JavaThread* current = THREAD; 810 811 markWord mark = obj->mark(); 812 if (LockingMode == LM_LIGHTWEIGHT) { 813 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 814 // Not inflated so there can't be any waiters to notify. 815 return; 816 } 817 } else if (LockingMode == LM_LEGACY) { 818 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 819 // Not inflated so there can't be any waiters to notify. 820 return; 821 } 822 } 823 824 ObjectMonitor* monitor; 825 if (LockingMode == LM_LIGHTWEIGHT) { 826 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 827 } else { 828 // The ObjectMonitor* can't be async deflated until ownership is 829 // dropped by the calling thread. 830 monitor = inflate(current, obj(), inflate_cause_notify); 831 } 832 monitor->notifyAll(CHECK); 833 } 834 835 // ----------------------------------------------------------------------------- 836 // Hash Code handling 837 838 struct SharedGlobals { 839 char _pad_prefix[OM_CACHE_LINE_SIZE]; 840 // This is a highly shared mostly-read variable. 841 // To avoid false-sharing it needs to be the sole occupant of a cache line. 842 volatile int stw_random; 843 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 844 // Hot RW variable -- Sequester to avoid false-sharing 845 volatile int hc_sequence; 846 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 847 }; 848 849 static SharedGlobals GVars; 850 851 static markWord read_stable_mark(oop obj) { 852 markWord mark = obj->mark_acquire(); 853 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) { 854 // New lightweight locking does not use the markWord::INFLATING() protocol. 855 return mark; // normal fast-path return 856 } 857 858 int its = 0; 859 for (;;) { 860 markWord mark = obj->mark_acquire(); 861 if (!mark.is_being_inflated()) { 862 return mark; // normal fast-path return 863 } 864 865 // The object is being inflated by some other thread. 866 // The caller of read_stable_mark() must wait for inflation to complete. 867 // Avoid live-lock. 868 869 ++its; 870 if (its > 10000 || !os::is_MP()) { 871 if (its & 1) { 872 os::naked_yield(); 873 } else { 874 // Note that the following code attenuates the livelock problem but is not 875 // a complete remedy. A more complete solution would require that the inflating 876 // thread hold the associated inflation lock. The following code simply restricts 877 // the number of spinners to at most one. We'll have N-2 threads blocked 878 // on the inflationlock, 1 thread holding the inflation lock and using 879 // a yield/park strategy, and 1 thread in the midst of inflation. 880 // A more refined approach would be to change the encoding of INFLATING 881 // to allow encapsulation of a native thread pointer. Threads waiting for 882 // inflation to complete would use CAS to push themselves onto a singly linked 883 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 884 // and calling park(). When inflation was complete the thread that accomplished inflation 885 // would detach the list and set the markword to inflated with a single CAS and 886 // then for each thread on the list, set the flag and unpark() the thread. 887 888 // Index into the lock array based on the current object address. 889 static_assert(is_power_of_2(inflation_lock_count()), "must be"); 890 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1); 891 int YieldThenBlock = 0; 892 assert(ix < inflation_lock_count(), "invariant"); 893 inflation_lock(ix)->lock(); 894 while (obj->mark_acquire() == markWord::INFLATING()) { 895 // Beware: naked_yield() is advisory and has almost no effect on some platforms 896 // so we periodically call current->_ParkEvent->park(1). 897 // We use a mixed spin/yield/block mechanism. 898 if ((YieldThenBlock++) >= 16) { 899 Thread::current()->_ParkEvent->park(1); 900 } else { 901 os::naked_yield(); 902 } 903 } 904 inflation_lock(ix)->unlock(); 905 } 906 } else { 907 SpinPause(); // SMP-polite spinning 908 } 909 } 910 } 911 912 // hashCode() generation : 913 // 914 // Possibilities: 915 // * MD5Digest of {obj,stw_random} 916 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 917 // * A DES- or AES-style SBox[] mechanism 918 // * One of the Phi-based schemes, such as: 919 // 2654435761 = 2^32 * Phi (golden ratio) 920 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 921 // * A variation of Marsaglia's shift-xor RNG scheme. 922 // * (obj ^ stw_random) is appealing, but can result 923 // in undesirable regularity in the hashCode values of adjacent objects 924 // (objects allocated back-to-back, in particular). This could potentially 925 // result in hashtable collisions and reduced hashtable efficiency. 926 // There are simple ways to "diffuse" the middle address bits over the 927 // generated hashCode values: 928 929 intptr_t ObjectSynchronizer::get_next_hash(Thread* current, oop obj) { 930 intptr_t value = 0; 931 if (hashCode == 0) { 932 // This form uses global Park-Miller RNG. 933 // On MP system we'll have lots of RW access to a global, so the 934 // mechanism induces lots of coherency traffic. 935 value = os::random(); 936 } else if (hashCode == 1) { 937 // This variation has the property of being stable (idempotent) 938 // between STW operations. This can be useful in some of the 1-0 939 // synchronization schemes. 940 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 941 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 942 } else if (hashCode == 2) { 943 value = 1; // for sensitivity testing 944 } else if (hashCode == 3) { 945 value = ++GVars.hc_sequence; 946 } else if (hashCode == 4) { 947 value = cast_from_oop<intptr_t>(obj); 948 } else if (hashCode == 5) { 949 // Marsaglia's xor-shift scheme with thread-specific state 950 // This is probably the best overall implementation -- we'll 951 // likely make this the default in future releases. 952 unsigned t = current->_hashStateX; 953 t ^= (t << 11); 954 current->_hashStateX = current->_hashStateY; 955 current->_hashStateY = current->_hashStateZ; 956 current->_hashStateZ = current->_hashStateW; 957 unsigned v = current->_hashStateW; 958 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 959 current->_hashStateW = v; 960 value = v; 961 } else { 962 assert(UseCompactObjectHeaders, "Only with compact i-hash"); 963 #ifdef _LP64 964 uint64_t val = cast_from_oop<uint64_t>(obj); 965 uint32_t hash = FastHash::get_hash32((uint32_t)val, (uint32_t)(val >> 32)); 966 #else 967 uint32_t val = cast_from_oop<uint32_t>(obj); 968 uint32_t hash = FastHash::get_hash32(val, UCONST64(0xAAAAAAAA)); 969 #endif 970 value= static_cast<intptr_t>(hash); 971 } 972 973 value &= markWord::hash_mask; 974 if (hashCode != 6 && value == 0) value = 0xBAD; 975 assert(value != markWord::no_hash || hashCode == 6, "invariant"); 976 return value; 977 } 978 979 static intptr_t install_hash_code(Thread* current, oop obj) { 980 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be"); 981 982 markWord mark = obj->mark_acquire(); 983 for (;;) { 984 if (UseCompactObjectHeaders) { 985 if (mark.is_hashed()) { 986 return LightweightSynchronizer::get_hash(mark, obj); 987 } 988 intptr_t hash = ObjectSynchronizer::get_next_hash(current, obj); // get a new hash 989 markWord new_mark; 990 if (mark.is_not_hashed_expanded()) { 991 new_mark = mark.set_hashed_expanded(); 992 int offset = mark.klass()->hash_offset_in_bytes(obj, mark); 993 obj->int_field_put(offset, (jint) hash); 994 } else { 995 new_mark = mark.set_hashed_not_expanded(); 996 } 997 markWord old_mark = obj->cas_set_mark(new_mark, mark); 998 if (old_mark == mark) { 999 return hash; 1000 } 1001 mark = old_mark; 1002 } else { 1003 intptr_t hash = mark.hash(); 1004 if (hash != 0) { 1005 return hash; 1006 } 1007 1008 hash = ObjectSynchronizer::get_next_hash(current, obj); 1009 const markWord old_mark = mark; 1010 const markWord new_mark = old_mark.copy_set_hash(hash); 1011 1012 mark = obj->cas_set_mark(new_mark, old_mark); 1013 if (old_mark == mark) { 1014 return hash; 1015 } 1016 } 1017 } 1018 } 1019 1020 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 1021 if (UseObjectMonitorTable) { 1022 // Since the monitor isn't in the object header, the hash can simply be 1023 // installed in the object header. 1024 return install_hash_code(current, obj); 1025 } 1026 1027 while (true) { 1028 ObjectMonitor* monitor = nullptr; 1029 markWord temp, test; 1030 intptr_t hash; 1031 markWord mark = read_stable_mark(obj); 1032 if (VerifyHeavyMonitors) { 1033 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)"); 1034 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 1035 } 1036 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) { 1037 hash = mark.hash(); 1038 if (hash != 0) { // if it has a hash, just return it 1039 return hash; 1040 } 1041 hash = get_next_hash(current, obj); // get a new hash 1042 temp = mark.copy_set_hash(hash); // merge the hash into header 1043 // try to install the hash 1044 test = obj->cas_set_mark(temp, mark); 1045 if (test == mark) { // if the hash was installed, return it 1046 return hash; 1047 } 1048 if (LockingMode == LM_LIGHTWEIGHT) { 1049 // CAS failed, retry 1050 continue; 1051 } 1052 // Failed to install the hash. It could be that another thread 1053 // installed the hash just before our attempt or inflation has 1054 // occurred or... so we fall thru to inflate the monitor for 1055 // stability and then install the hash. 1056 } else if (mark.has_monitor()) { 1057 monitor = mark.monitor(); 1058 temp = monitor->header(); 1059 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1060 hash = temp.hash(); 1061 if (hash != 0) { 1062 // It has a hash. 1063 1064 // Separate load of dmw/header above from the loads in 1065 // is_being_async_deflated(). 1066 1067 // dmw/header and _contentions may get written by different threads. 1068 // Make sure to observe them in the same order when having several observers. 1069 OrderAccess::loadload_for_IRIW(); 1070 1071 if (monitor->is_being_async_deflated()) { 1072 // But we can't safely use the hash if we detect that async 1073 // deflation has occurred. So we attempt to restore the 1074 // header/dmw to the object's header so that we only retry 1075 // once if the deflater thread happens to be slow. 1076 monitor->install_displaced_markword_in_object(obj); 1077 continue; 1078 } 1079 return hash; 1080 } 1081 // Fall thru so we only have one place that installs the hash in 1082 // the ObjectMonitor. 1083 } else if (LockingMode == LM_LEGACY && mark.has_locker() 1084 && current->is_Java_thread() 1085 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) { 1086 // This is a stack-lock owned by the calling thread so fetch the 1087 // displaced markWord from the BasicLock on the stack. 1088 temp = mark.displaced_mark_helper(); 1089 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1090 hash = temp.hash(); 1091 if (hash != 0) { // if it has a hash, just return it 1092 return hash; 1093 } 1094 // WARNING: 1095 // The displaced header in the BasicLock on a thread's stack 1096 // is strictly immutable. It CANNOT be changed in ANY cases. 1097 // So we have to inflate the stack-lock into an ObjectMonitor 1098 // even if the current thread owns the lock. The BasicLock on 1099 // a thread's stack can be asynchronously read by other threads 1100 // during an inflate() call so any change to that stack memory 1101 // may not propagate to other threads correctly. 1102 } 1103 1104 // Inflate the monitor to set the hash. 1105 1106 // There's no need to inflate if the mark has already got a monitor. 1107 // NOTE: an async deflation can race after we get the monitor and 1108 // before we can update the ObjectMonitor's header with the hash 1109 // value below. 1110 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code); 1111 // Load ObjectMonitor's header/dmw field and see if it has a hash. 1112 mark = monitor->header(); 1113 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1114 hash = mark.hash(); 1115 if (hash == 0) { // if it does not have a hash 1116 hash = get_next_hash(current, obj); // get a new hash 1117 temp = mark.copy_set_hash(hash) ; // merge the hash into header 1118 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1119 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value()); 1120 test = markWord(v); 1121 if (test != mark) { 1122 // The attempt to update the ObjectMonitor's header/dmw field 1123 // did not work. This can happen if another thread managed to 1124 // merge in the hash just before our cmpxchg(). 1125 // If we add any new usages of the header/dmw field, this code 1126 // will need to be updated. 1127 hash = test.hash(); 1128 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1129 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1130 } 1131 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) { 1132 // If we detect that async deflation has occurred, then we 1133 // attempt to restore the header/dmw to the object's header 1134 // so that we only retry once if the deflater thread happens 1135 // to be slow. 1136 monitor->install_displaced_markword_in_object(obj); 1137 continue; 1138 } 1139 } 1140 // We finally get the hash. 1141 return hash; 1142 } 1143 } 1144 1145 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 1146 Handle h_obj) { 1147 assert(current == JavaThread::current(), "Can only be called on current thread"); 1148 oop obj = h_obj(); 1149 1150 markWord mark = read_stable_mark(obj); 1151 1152 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1153 // stack-locked case, header points into owner's stack 1154 return current->is_lock_owned((address)mark.locker()); 1155 } 1156 1157 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1158 // fast-locking case, see if lock is in current's lock stack 1159 return current->lock_stack().contains(h_obj()); 1160 } 1161 1162 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) { 1163 ObjectMonitor* monitor = read_monitor(current, obj, mark); 1164 if (monitor != nullptr) { 1165 return monitor->is_entered(current) != 0; 1166 } 1167 // Racing with inflation/deflation, retry 1168 mark = obj->mark_acquire(); 1169 1170 if (mark.is_fast_locked()) { 1171 // Some other thread fast_locked, current could not have held the lock 1172 return false; 1173 } 1174 } 1175 1176 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) { 1177 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1178 // The first stage of async deflation does not affect any field 1179 // used by this comparison so the ObjectMonitor* is usable here. 1180 ObjectMonitor* monitor = read_monitor(mark); 1181 return monitor->is_entered(current) != 0; 1182 } 1183 // Unlocked case, header in place 1184 assert(mark.is_unlocked(), "sanity check"); 1185 return false; 1186 } 1187 1188 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1189 oop obj = h_obj(); 1190 markWord mark = read_stable_mark(obj); 1191 1192 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1193 // stack-locked so header points into owner's stack. 1194 // owning_thread_from_monitor_owner() may also return null here: 1195 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker()); 1196 } 1197 1198 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1199 // fast-locked so get owner from the object. 1200 // owning_thread_from_object() may also return null here: 1201 return Threads::owning_thread_from_object(t_list, h_obj()); 1202 } 1203 1204 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) { 1205 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark); 1206 if (monitor != nullptr) { 1207 return Threads::owning_thread_from_monitor(t_list, monitor); 1208 } 1209 // Racing with inflation/deflation, retry 1210 mark = obj->mark_acquire(); 1211 1212 if (mark.is_fast_locked()) { 1213 // Some other thread fast_locked 1214 return Threads::owning_thread_from_object(t_list, h_obj()); 1215 } 1216 } 1217 1218 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) { 1219 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1220 // The first stage of async deflation does not affect any field 1221 // used by this comparison so the ObjectMonitor* is usable here. 1222 ObjectMonitor* monitor = read_monitor(mark); 1223 assert(monitor != nullptr, "monitor should be non-null"); 1224 // owning_thread_from_monitor() may also return null here: 1225 return Threads::owning_thread_from_monitor(t_list, monitor); 1226 } 1227 1228 // Unlocked case, header in place 1229 // Cannot have assertion since this object may have been 1230 // locked by another thread when reaching here. 1231 // assert(mark.is_unlocked(), "sanity check"); 1232 1233 return nullptr; 1234 } 1235 1236 // Visitors ... 1237 1238 // Iterate over all ObjectMonitors. 1239 template <typename Function> 1240 void ObjectSynchronizer::monitors_iterate(Function function) { 1241 MonitorList::Iterator iter = _in_use_list.iterator(); 1242 while (iter.has_next()) { 1243 ObjectMonitor* monitor = iter.next(); 1244 function(monitor); 1245 } 1246 } 1247 1248 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 1249 // returns true. 1250 template <typename OwnerFilter> 1251 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 1252 monitors_iterate([&](ObjectMonitor* monitor) { 1253 // This function is only called at a safepoint or when the 1254 // target thread is suspended or when the target thread is 1255 // operating on itself. The current closures in use today are 1256 // only interested in an owned ObjectMonitor and ownership 1257 // cannot be dropped under the calling contexts so the 1258 // ObjectMonitor cannot be async deflated. 1259 if (monitor->has_owner() && filter(monitor)) { 1260 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 1261 1262 closure->do_monitor(monitor); 1263 } 1264 }); 1265 } 1266 1267 // Iterate ObjectMonitors where the owner == thread; this does NOT include 1268 // ObjectMonitors where owner is set to a stack-lock address in thread. 1269 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 1270 int64_t key = ObjectMonitor::owner_id_from(thread); 1271 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 1272 return owned_monitors_iterate_filtered(closure, thread_filter); 1273 } 1274 1275 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) { 1276 int64_t key = ObjectMonitor::owner_id_from(vthread); 1277 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 1278 return owned_monitors_iterate_filtered(closure, thread_filter); 1279 } 1280 1281 // Iterate ObjectMonitors owned by any thread. 1282 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 1283 auto all_filter = [&](ObjectMonitor* monitor) { return true; }; 1284 return owned_monitors_iterate_filtered(closure, all_filter); 1285 } 1286 1287 static bool monitors_used_above_threshold(MonitorList* list) { 1288 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 1289 return false; 1290 } 1291 size_t monitors_used = list->count(); 1292 if (monitors_used == 0) { // empty list is easy 1293 return false; 1294 } 1295 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling(); 1296 // Make sure that we use a ceiling value that is not lower than 1297 // previous, not lower than the recorded max used by the system, and 1298 // not lower than the current number of monitors in use (which can 1299 // race ahead of max). The result is guaranteed > 0. 1300 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used); 1301 1302 // Check if our monitor usage is above the threshold: 1303 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 1304 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 1305 // Deflate monitors if over the threshold percentage, unless no 1306 // progress on previous deflations. 1307 bool is_above_threshold = true; 1308 1309 // Check if it's time to adjust the in_use_list_ceiling up, due 1310 // to too many async deflation attempts without any progress. 1311 if (NoAsyncDeflationProgressMax != 0 && 1312 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 1313 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 1314 size_t delta = (size_t)(ceiling * remainder) + 1; 1315 size_t new_ceiling = (ceiling > SIZE_MAX - delta) 1316 ? SIZE_MAX // Overflow, let's clamp new_ceiling. 1317 : ceiling + delta; 1318 1319 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 1320 log_info(monitorinflation)("Too many deflations without progress; " 1321 "bumping in_use_list_ceiling from %zu" 1322 " to %zu", old_ceiling, new_ceiling); 1323 _no_progress_cnt = 0; 1324 ceiling = new_ceiling; 1325 1326 // Check if our monitor usage is still above the threshold: 1327 monitor_usage = (monitors_used * 100LL) / ceiling; 1328 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold; 1329 } 1330 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu" 1331 ", monitor_usage=%zu, threshold=%d", 1332 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 1333 return is_above_threshold; 1334 } 1335 1336 return false; 1337 } 1338 1339 size_t ObjectSynchronizer::in_use_list_count() { 1340 return _in_use_list.count(); 1341 } 1342 1343 size_t ObjectSynchronizer::in_use_list_max() { 1344 return _in_use_list.max(); 1345 } 1346 1347 size_t ObjectSynchronizer::in_use_list_ceiling() { 1348 return _in_use_list_ceiling; 1349 } 1350 1351 void ObjectSynchronizer::dec_in_use_list_ceiling() { 1352 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1353 } 1354 1355 void ObjectSynchronizer::inc_in_use_list_ceiling() { 1356 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1357 } 1358 1359 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 1360 _in_use_list_ceiling = new_value; 1361 } 1362 1363 bool ObjectSynchronizer::is_async_deflation_needed() { 1364 if (is_async_deflation_requested()) { 1365 // Async deflation request. 1366 log_info(monitorinflation)("Async deflation needed: explicit request"); 1367 return true; 1368 } 1369 1370 jlong time_since_last = time_since_last_async_deflation_ms(); 1371 1372 if (AsyncDeflationInterval > 0 && 1373 time_since_last > AsyncDeflationInterval && 1374 monitors_used_above_threshold(&_in_use_list)) { 1375 // It's been longer than our specified deflate interval and there 1376 // are too many monitors in use. We don't deflate more frequently 1377 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1378 // in order to not swamp the MonitorDeflationThread. 1379 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 1380 return true; 1381 } 1382 1383 if (GuaranteedAsyncDeflationInterval > 0 && 1384 time_since_last > GuaranteedAsyncDeflationInterval) { 1385 // It's been longer than our specified guaranteed deflate interval. 1386 // We need to clean up the used monitors even if the threshold is 1387 // not reached, to keep the memory utilization at bay when many threads 1388 // touched many monitors. 1389 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) " 1390 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 1391 GuaranteedAsyncDeflationInterval, time_since_last); 1392 1393 // If this deflation has no progress, then it should not affect the no-progress 1394 // tracking, otherwise threshold heuristics would think it was triggered, experienced 1395 // no progress, and needs to backoff more aggressively. In this "no progress" case, 1396 // the generic code would bump the no-progress counter, and we compensate for that 1397 // by telling it to skip the update. 1398 // 1399 // If this deflation has progress, then it should let non-progress tracking 1400 // know about this, otherwise the threshold heuristics would kick in, potentially 1401 // experience no-progress due to aggressive cleanup by this deflation, and think 1402 // it is still in no-progress stride. In this "progress" case, the generic code would 1403 // zero the counter, and we allow it to happen. 1404 _no_progress_skip_increment = true; 1405 1406 return true; 1407 } 1408 1409 return false; 1410 } 1411 1412 void ObjectSynchronizer::request_deflate_idle_monitors() { 1413 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1414 set_is_async_deflation_requested(true); 1415 ml.notify_all(); 1416 } 1417 1418 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 1419 JavaThread* current = JavaThread::current(); 1420 bool ret_code = false; 1421 1422 jlong last_time = last_async_deflation_time_ns(); 1423 1424 request_deflate_idle_monitors(); 1425 1426 const int N_CHECKS = 5; 1427 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1428 if (last_async_deflation_time_ns() > last_time) { 1429 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1430 ret_code = true; 1431 break; 1432 } 1433 { 1434 // JavaThread has to honor the blocking protocol. 1435 ThreadBlockInVM tbivm(current); 1436 os::naked_short_sleep(999); // sleep for almost 1 second 1437 } 1438 } 1439 if (!ret_code) { 1440 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1441 } 1442 1443 return ret_code; 1444 } 1445 1446 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1447 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1448 } 1449 1450 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1451 const oop obj, 1452 ObjectSynchronizer::InflateCause cause) { 1453 assert(event != nullptr, "invariant"); 1454 const Klass* monitor_klass = obj->klass(); 1455 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) { 1456 return; 1457 } 1458 event->set_monitorClass(monitor_klass); 1459 event->set_address((uintptr_t)(void*)obj); 1460 event->set_cause((u1)cause); 1461 event->commit(); 1462 } 1463 1464 // Fast path code shared by multiple functions 1465 void ObjectSynchronizer::inflate_helper(oop obj) { 1466 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter"); 1467 markWord mark = obj->mark_acquire(); 1468 if (mark.has_monitor()) { 1469 ObjectMonitor* monitor = read_monitor(mark); 1470 markWord dmw = monitor->header(); 1471 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1472 return; 1473 } 1474 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal); 1475 } 1476 1477 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) { 1478 assert(current == Thread::current(), "must be"); 1479 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter"); 1480 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause); 1481 } 1482 1483 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) { 1484 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be"); 1485 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for"); 1486 return inflate_impl(thread, obj, cause); 1487 } 1488 1489 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) { 1490 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or 1491 // is suspended throughout the call by some other mechanism. 1492 // The thread might be nullptr when called from a non JavaThread. (As may still be 1493 // the case from FastHashCode). However it is only important for correctness that the 1494 // thread is set when called from ObjectSynchronizer::enter from the owning thread, 1495 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit. 1496 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl"); 1497 EventJavaMonitorInflate event; 1498 1499 for (;;) { 1500 const markWord mark = object->mark_acquire(); 1501 1502 // The mark can be in one of the following states: 1503 // * inflated - If the ObjectMonitor owner is anonymous and the 1504 // locking_thread owns the object lock, then we 1505 // make the locking_thread the ObjectMonitor owner. 1506 // * stack-locked - Coerce it to inflated from stack-locked. 1507 // * INFLATING - Busy wait for conversion from stack-locked to 1508 // inflated. 1509 // * unlocked - Aggressively inflate the object. 1510 1511 // CASE: inflated 1512 if (mark.has_monitor()) { 1513 ObjectMonitor* inf = mark.monitor(); 1514 markWord dmw = inf->header(); 1515 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1516 if (inf->has_anonymous_owner() && locking_thread != nullptr) { 1517 assert(LockingMode == LM_LEGACY, "invariant"); 1518 if (locking_thread->is_lock_owned((address)inf->stack_locker())) { 1519 inf->set_stack_locker(nullptr); 1520 inf->set_owner_from_anonymous(locking_thread); 1521 } 1522 } 1523 return inf; 1524 } 1525 1526 // CASE: inflation in progress - inflating over a stack-lock. 1527 // Some other thread is converting from stack-locked to inflated. 1528 // Only that thread can complete inflation -- other threads must wait. 1529 // The INFLATING value is transient. 1530 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1531 // We could always eliminate polling by parking the thread on some auxiliary list. 1532 if (mark == markWord::INFLATING()) { 1533 read_stable_mark(object); 1534 continue; 1535 } 1536 1537 // CASE: stack-locked 1538 // Could be stack-locked either by current or by some other thread. 1539 // 1540 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting 1541 // to install INFLATING into the mark word. We originally installed INFLATING, 1542 // allocated the ObjectMonitor, and then finally STed the address of the 1543 // ObjectMonitor into the mark. This was correct, but artificially lengthened 1544 // the interval in which INFLATING appeared in the mark, thus increasing 1545 // the odds of inflation contention. If we lose the race to set INFLATING, 1546 // then we just delete the ObjectMonitor and loop around again. 1547 // 1548 LogStreamHandle(Trace, monitorinflation) lsh; 1549 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1550 ObjectMonitor* m = new ObjectMonitor(object); 1551 // Optimistically prepare the ObjectMonitor - anticipate successful CAS 1552 // We do this before the CAS in order to minimize the length of time 1553 // in which INFLATING appears in the mark. 1554 1555 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1556 if (cmp != mark) { 1557 delete m; 1558 continue; // Interference -- just retry 1559 } 1560 1561 // We've successfully installed INFLATING (0) into the mark-word. 1562 // This is the only case where 0 will appear in a mark-word. 1563 // Only the singular thread that successfully swings the mark-word 1564 // to 0 can perform (or more precisely, complete) inflation. 1565 // 1566 // Why do we CAS a 0 into the mark-word instead of just CASing the 1567 // mark-word from the stack-locked value directly to the new inflated state? 1568 // Consider what happens when a thread unlocks a stack-locked object. 1569 // It attempts to use CAS to swing the displaced header value from the 1570 // on-stack BasicLock back into the object header. Recall also that the 1571 // header value (hash code, etc) can reside in (a) the object header, or 1572 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1573 // header in an ObjectMonitor. The inflate() routine must copy the header 1574 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1575 // the while preserving the hashCode stability invariants. If the owner 1576 // decides to release the lock while the value is 0, the unlock will fail 1577 // and control will eventually pass from slow_exit() to inflate. The owner 1578 // will then spin, waiting for the 0 value to disappear. Put another way, 1579 // the 0 causes the owner to stall if the owner happens to try to 1580 // drop the lock (restoring the header from the BasicLock to the object) 1581 // while inflation is in-progress. This protocol avoids races that might 1582 // would otherwise permit hashCode values to change or "flicker" for an object. 1583 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1584 // 0 serves as a "BUSY" inflate-in-progress indicator. 1585 1586 1587 // fetch the displaced mark from the owner's stack. 1588 // The owner can't die or unwind past the lock while our INFLATING 1589 // object is in the mark. Furthermore the owner can't complete 1590 // an unlock on the object, either. 1591 markWord dmw = mark.displaced_mark_helper(); 1592 // Catch if the object's header is not neutral (not locked and 1593 // not marked is what we care about here). 1594 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1595 1596 // Setup monitor fields to proper values -- prepare the monitor 1597 m->set_header(dmw); 1598 1599 // Note that a thread can inflate an object 1600 // that it has stack-locked -- as might happen in wait() -- directly 1601 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom. 1602 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) { 1603 m->set_owner(locking_thread); 1604 } else { 1605 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack, 1606 // and set the stack locker field in the monitor. 1607 m->set_stack_locker(mark.locker()); 1608 m->set_anonymous_owner(); 1609 } 1610 // TODO-FIXME: assert BasicLock->dhw != 0. 1611 1612 // Must preserve store ordering. The monitor state must 1613 // be stable at the time of publishing the monitor address. 1614 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1615 // Release semantics so that above set_object() is seen first. 1616 object->release_set_mark(markWord::encode(m)); 1617 1618 // Once ObjectMonitor is configured and the object is associated 1619 // with the ObjectMonitor, it is safe to allow async deflation: 1620 _in_use_list.add(m); 1621 1622 if (log_is_enabled(Trace, monitorinflation)) { 1623 ResourceMark rm; 1624 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1625 INTPTR_FORMAT ", type='%s'", p2i(object), 1626 object->mark().value(), object->klass()->external_name()); 1627 } 1628 if (event.should_commit()) { 1629 post_monitor_inflate_event(&event, object, cause); 1630 } 1631 return m; 1632 } 1633 1634 // CASE: unlocked 1635 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1636 // If we know we're inflating for entry it's better to inflate by swinging a 1637 // pre-locked ObjectMonitor pointer into the object header. A successful 1638 // CAS inflates the object *and* confers ownership to the inflating thread. 1639 // In the current implementation we use a 2-step mechanism where we CAS() 1640 // to inflate and then CAS() again to try to swing _owner from null to current. 1641 // An inflateTry() method that we could call from enter() would be useful. 1642 1643 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1644 ObjectMonitor* m = new ObjectMonitor(object); 1645 // prepare m for installation - set monitor to initial state 1646 m->set_header(mark); 1647 1648 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1649 delete m; 1650 m = nullptr; 1651 continue; 1652 // interference - the markword changed - just retry. 1653 // The state-transitions are one-way, so there's no chance of 1654 // live-lock -- "Inflated" is an absorbing state. 1655 } 1656 1657 // Once the ObjectMonitor is configured and object is associated 1658 // with the ObjectMonitor, it is safe to allow async deflation: 1659 _in_use_list.add(m); 1660 1661 if (log_is_enabled(Trace, monitorinflation)) { 1662 ResourceMark rm; 1663 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark=" 1664 INTPTR_FORMAT ", type='%s'", p2i(object), 1665 object->mark().value(), object->klass()->external_name()); 1666 } 1667 if (event.should_commit()) { 1668 post_monitor_inflate_event(&event, object, cause); 1669 } 1670 return m; 1671 } 1672 } 1673 1674 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1675 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1676 // 1677 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1678 MonitorList::Iterator iter = _in_use_list.iterator(); 1679 size_t deflated_count = 0; 1680 Thread* current = Thread::current(); 1681 1682 while (iter.has_next()) { 1683 if (deflated_count >= (size_t)MonitorDeflationMax) { 1684 break; 1685 } 1686 ObjectMonitor* mid = iter.next(); 1687 if (mid->deflate_monitor(current)) { 1688 deflated_count++; 1689 } 1690 1691 // Must check for a safepoint/handshake and honor it. 1692 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1693 } 1694 1695 return deflated_count; 1696 } 1697 1698 class HandshakeForDeflation : public HandshakeClosure { 1699 public: 1700 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {} 1701 1702 void do_thread(Thread* thread) { 1703 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread=" 1704 INTPTR_FORMAT, p2i(thread)); 1705 if (thread->is_Java_thread()) { 1706 // Clear OM cache 1707 JavaThread* jt = JavaThread::cast(thread); 1708 jt->om_clear_monitor_cache(); 1709 } 1710 } 1711 }; 1712 1713 class VM_RendezvousGCThreads : public VM_Operation { 1714 public: 1715 bool evaluate_at_safepoint() const override { return false; } 1716 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1717 void doit() override { 1718 Universe::heap()->safepoint_synchronize_begin(); 1719 Universe::heap()->safepoint_synchronize_end(); 1720 }; 1721 }; 1722 1723 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1724 ObjectMonitorDeflationSafepointer* safepointer) { 1725 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1726 size_t deleted_count = 0; 1727 for (ObjectMonitor* monitor: *delete_list) { 1728 delete monitor; 1729 deleted_count++; 1730 // A JavaThread must check for a safepoint/handshake and honor it. 1731 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1732 } 1733 return deleted_count; 1734 } 1735 1736 class ObjectMonitorDeflationLogging: public StackObj { 1737 LogStreamHandle(Debug, monitorinflation) _debug; 1738 LogStreamHandle(Info, monitorinflation) _info; 1739 LogStream* _stream; 1740 elapsedTimer _timer; 1741 1742 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1743 size_t count() const { return ObjectSynchronizer::in_use_list_count(); } 1744 size_t max() const { return ObjectSynchronizer::in_use_list_max(); } 1745 1746 public: 1747 ObjectMonitorDeflationLogging() 1748 : _debug(), _info(), _stream(nullptr) { 1749 if (_debug.is_enabled()) { 1750 _stream = &_debug; 1751 } else if (_info.is_enabled()) { 1752 _stream = &_info; 1753 } 1754 } 1755 1756 void begin() { 1757 if (_stream != nullptr) { 1758 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1759 ceiling(), count(), max()); 1760 _timer.start(); 1761 } 1762 } 1763 1764 void before_handshake(size_t unlinked_count) { 1765 if (_stream != nullptr) { 1766 _timer.stop(); 1767 _stream->print_cr("before handshaking: unlinked_count=%zu" 1768 ", in_use_list stats: ceiling=%zu, count=" 1769 "%zu, max=%zu", 1770 unlinked_count, ceiling(), count(), max()); 1771 } 1772 } 1773 1774 void after_handshake() { 1775 if (_stream != nullptr) { 1776 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1777 "%zu, count=%zu, max=%zu", 1778 ceiling(), count(), max()); 1779 _timer.start(); 1780 } 1781 } 1782 1783 void end(size_t deflated_count, size_t unlinked_count) { 1784 if (_stream != nullptr) { 1785 _timer.stop(); 1786 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1787 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs", 1788 deflated_count, unlinked_count, _timer.seconds()); 1789 } 1790 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1791 ceiling(), count(), max()); 1792 } 1793 } 1794 1795 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1796 if (_stream != nullptr) { 1797 _timer.stop(); 1798 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling=" 1799 "%zu, count=%zu, max=%zu", 1800 op_name, cnt_name, cnt, ceiling(), count(), max()); 1801 } 1802 } 1803 1804 void after_block_for_safepoint(const char* op_name) { 1805 if (_stream != nullptr) { 1806 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu" 1807 ", count=%zu, max=%zu", op_name, 1808 ceiling(), count(), max()); 1809 _timer.start(); 1810 } 1811 } 1812 }; 1813 1814 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1815 if (!SafepointMechanism::should_process(_current)) { 1816 return; 1817 } 1818 1819 // A safepoint/handshake has started. 1820 _log->before_block_for_safepoint(op_name, count_name, counter); 1821 1822 { 1823 // Honor block request. 1824 ThreadBlockInVM tbivm(_current); 1825 } 1826 1827 _log->after_block_for_safepoint(op_name); 1828 } 1829 1830 // This function is called by the MonitorDeflationThread to deflate 1831 // ObjectMonitors. 1832 size_t ObjectSynchronizer::deflate_idle_monitors() { 1833 JavaThread* current = JavaThread::current(); 1834 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1835 1836 // The async deflation request has been processed. 1837 _last_async_deflation_time_ns = os::javaTimeNanos(); 1838 set_is_async_deflation_requested(false); 1839 1840 ObjectMonitorDeflationLogging log; 1841 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1842 1843 log.begin(); 1844 1845 // Deflate some idle ObjectMonitors. 1846 size_t deflated_count = deflate_monitor_list(&safepointer); 1847 1848 // Unlink the deflated ObjectMonitors from the in-use list. 1849 size_t unlinked_count = 0; 1850 size_t deleted_count = 0; 1851 if (deflated_count > 0) { 1852 ResourceMark rm(current); 1853 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1854 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1855 1856 #ifdef ASSERT 1857 if (UseObjectMonitorTable) { 1858 for (ObjectMonitor* monitor : delete_list) { 1859 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed"); 1860 } 1861 } 1862 #endif 1863 1864 log.before_handshake(unlinked_count); 1865 1866 // A JavaThread needs to handshake in order to safely free the 1867 // ObjectMonitors that were deflated in this cycle. 1868 HandshakeForDeflation hfd_hc; 1869 Handshake::execute(&hfd_hc); 1870 // Also, we sync and desync GC threads around the handshake, so that they can 1871 // safely read the mark-word and look-through to the object-monitor, without 1872 // being afraid that the object-monitor is going away. 1873 VM_RendezvousGCThreads sync_gc; 1874 VMThread::execute(&sync_gc); 1875 1876 log.after_handshake(); 1877 1878 // After the handshake, safely free the ObjectMonitors that were 1879 // deflated and unlinked in this cycle. 1880 1881 // Delete the unlinked ObjectMonitors. 1882 deleted_count = delete_monitors(&delete_list, &safepointer); 1883 assert(unlinked_count == deleted_count, "must be"); 1884 } 1885 1886 log.end(deflated_count, unlinked_count); 1887 1888 GVars.stw_random = os::random(); 1889 1890 if (deflated_count != 0) { 1891 _no_progress_cnt = 0; 1892 } else if (_no_progress_skip_increment) { 1893 _no_progress_skip_increment = false; 1894 } else { 1895 _no_progress_cnt++; 1896 } 1897 1898 return deflated_count; 1899 } 1900 1901 // Monitor cleanup on JavaThread::exit 1902 1903 // Iterate through monitor cache and attempt to release thread's monitors 1904 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1905 private: 1906 JavaThread* _thread; 1907 1908 public: 1909 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1910 void do_monitor(ObjectMonitor* mid) { 1911 intx rec = mid->complete_exit(_thread); 1912 _thread->dec_held_monitor_count(rec + 1); 1913 } 1914 }; 1915 1916 // Release all inflated monitors owned by current thread. Lightweight monitors are 1917 // ignored. This is meant to be called during JNI thread detach which assumes 1918 // all remaining monitors are heavyweight. All exceptions are swallowed. 1919 // Scanning the extant monitor list can be time consuming. 1920 // A simple optimization is to add a per-thread flag that indicates a thread 1921 // called jni_monitorenter() during its lifetime. 1922 // 1923 // Instead of NoSafepointVerifier it might be cheaper to 1924 // use an idiom of the form: 1925 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1926 // <code that must not run at safepoint> 1927 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1928 // Since the tests are extremely cheap we could leave them enabled 1929 // for normal product builds. 1930 1931 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1932 assert(current == JavaThread::current(), "must be current Java thread"); 1933 NoSafepointVerifier nsv; 1934 ReleaseJavaMonitorsClosure rjmc(current); 1935 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1936 assert(!current->has_pending_exception(), "Should not be possible"); 1937 current->clear_pending_exception(); 1938 assert(current->held_monitor_count() == 0, "Should not be possible"); 1939 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1940 current->clear_jni_monitor_count(); 1941 } 1942 1943 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1944 switch (cause) { 1945 case inflate_cause_vm_internal: return "VM Internal"; 1946 case inflate_cause_monitor_enter: return "Monitor Enter"; 1947 case inflate_cause_wait: return "Monitor Wait"; 1948 case inflate_cause_notify: return "Monitor Notify"; 1949 case inflate_cause_hash_code: return "Monitor Hash Code"; 1950 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1951 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1952 default: 1953 ShouldNotReachHere(); 1954 } 1955 return "Unknown"; 1956 } 1957 1958 //------------------------------------------------------------------------------ 1959 // Debugging code 1960 1961 u_char* ObjectSynchronizer::get_gvars_addr() { 1962 return (u_char*)&GVars; 1963 } 1964 1965 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1966 return (u_char*)&GVars.hc_sequence; 1967 } 1968 1969 size_t ObjectSynchronizer::get_gvars_size() { 1970 return sizeof(SharedGlobals); 1971 } 1972 1973 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1974 return (u_char*)&GVars.stw_random; 1975 } 1976 1977 // Do the final audit and print of ObjectMonitor stats; must be done 1978 // by the VMThread at VM exit time. 1979 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1980 assert(Thread::current()->is_VM_thread(), "sanity check"); 1981 1982 if (is_final_audit()) { // Only do the audit once. 1983 return; 1984 } 1985 set_is_final_audit(); 1986 log_info(monitorinflation)("Starting the final audit."); 1987 1988 if (log_is_enabled(Info, monitorinflation)) { 1989 LogStreamHandle(Info, monitorinflation) ls; 1990 audit_and_print_stats(&ls, true /* on_exit */); 1991 } 1992 } 1993 1994 // This function can be called by the MonitorDeflationThread or it can be called when 1995 // we are trying to exit the VM. The list walker functions can run in parallel with 1996 // the other list operations. 1997 // Calls to this function can be added in various places as a debugging 1998 // aid. 1999 // 2000 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) { 2001 int error_cnt = 0; 2002 2003 ls->print_cr("Checking in_use_list:"); 2004 chk_in_use_list(ls, &error_cnt); 2005 2006 if (error_cnt == 0) { 2007 ls->print_cr("No errors found in in_use_list checks."); 2008 } else { 2009 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 2010 } 2011 2012 // When exiting, only log the interesting entries at the Info level. 2013 // When called at intervals by the MonitorDeflationThread, log output 2014 // at the Trace level since there can be a lot of it. 2015 if (!on_exit && log_is_enabled(Trace, monitorinflation)) { 2016 LogStreamHandle(Trace, monitorinflation) ls_tr; 2017 log_in_use_monitor_details(&ls_tr, true /* log_all */); 2018 } else if (on_exit) { 2019 log_in_use_monitor_details(ls, false /* log_all */); 2020 } 2021 2022 ls->flush(); 2023 2024 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2025 } 2026 2027 // Check the in_use_list; log the results of the checks. 2028 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 2029 size_t l_in_use_count = _in_use_list.count(); 2030 size_t l_in_use_max = _in_use_list.max(); 2031 out->print_cr("count=%zu, max=%zu", l_in_use_count, 2032 l_in_use_max); 2033 2034 size_t ck_in_use_count = 0; 2035 MonitorList::Iterator iter = _in_use_list.iterator(); 2036 while (iter.has_next()) { 2037 ObjectMonitor* mid = iter.next(); 2038 chk_in_use_entry(mid, out, error_cnt_p); 2039 ck_in_use_count++; 2040 } 2041 2042 if (l_in_use_count == ck_in_use_count) { 2043 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu", 2044 l_in_use_count, ck_in_use_count); 2045 } else { 2046 out->print_cr("WARNING: in_use_count=%zu is not equal to " 2047 "ck_in_use_count=%zu", l_in_use_count, 2048 ck_in_use_count); 2049 } 2050 2051 size_t ck_in_use_max = _in_use_list.max(); 2052 if (l_in_use_max == ck_in_use_max) { 2053 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu", 2054 l_in_use_max, ck_in_use_max); 2055 } else { 2056 out->print_cr("WARNING: in_use_max=%zu is not equal to " 2057 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max); 2058 } 2059 } 2060 2061 // Check an in-use monitor entry; log any errors. 2062 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 2063 int* error_cnt_p) { 2064 if (n->owner_is_DEFLATER_MARKER()) { 2065 // This could happen when monitor deflation blocks for a safepoint. 2066 return; 2067 } 2068 2069 2070 if (n->metadata() == 0) { 2071 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 2072 "have non-null _metadata (header/hash) field.", p2i(n)); 2073 *error_cnt_p = *error_cnt_p + 1; 2074 } 2075 2076 const oop obj = n->object_peek(); 2077 if (obj == nullptr) { 2078 return; 2079 } 2080 2081 const markWord mark = obj->mark(); 2082 if (!mark.has_monitor()) { 2083 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2084 "object does not think it has a monitor: obj=" 2085 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2086 p2i(obj), mark.value()); 2087 *error_cnt_p = *error_cnt_p + 1; 2088 return; 2089 } 2090 2091 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark); 2092 if (n != obj_mon) { 2093 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2094 "object does not refer to the same monitor: obj=" 2095 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2096 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2097 *error_cnt_p = *error_cnt_p + 1; 2098 } 2099 } 2100 2101 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 2102 // flags indicate why the entry is in-use, 'object' and 'object type' 2103 // indicate the associated object and its type. 2104 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 2105 if (_in_use_list.count() > 0) { 2106 stringStream ss; 2107 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)"); 2108 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2109 out->print_cr("%18s %s %18s %18s", 2110 "monitor", "BHL", "object", "object type"); 2111 out->print_cr("================== === ================== =================="); 2112 2113 auto is_interesting = [&](ObjectMonitor* monitor) { 2114 return log_all || monitor->has_owner() || monitor->is_busy(); 2115 }; 2116 2117 monitors_iterate([&](ObjectMonitor* monitor) { 2118 if (is_interesting(monitor)) { 2119 const oop obj = monitor->object_peek(); 2120 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash(); 2121 ResourceMark rm; 2122 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 2123 monitor->is_busy(), hash != 0, monitor->has_owner(), 2124 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 2125 if (monitor->is_busy()) { 2126 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 2127 ss.reset(); 2128 } 2129 out->cr(); 2130 } 2131 }); 2132 } 2133 2134 out->flush(); 2135 }