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