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/perfData.hpp" 52 #include "runtime/safepointMechanism.inline.hpp" 53 #include "runtime/safepointVerifiers.hpp" 54 #include "runtime/sharedRuntime.hpp" 55 #include "runtime/stubRoutines.hpp" 56 #include "runtime/synchronizer.inline.hpp" 57 #include "runtime/threads.hpp" 58 #include "runtime/timer.hpp" 59 #include "runtime/trimNativeHeap.hpp" 60 #include "runtime/vframe.hpp" 61 #include "runtime/vmThread.hpp" 62 #include "utilities/align.hpp" 63 #include "utilities/dtrace.hpp" 64 #include "utilities/events.hpp" 65 #include "utilities/globalCounter.inline.hpp" 66 #include "utilities/globalDefinitions.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); 80 if (count > max()) { 81 Atomic::inc(&_max); 82 } 83 } 84 85 size_t MonitorList::count() const { 86 return Atomic::load(&_count); 87 } 88 89 size_t MonitorList::max() const { 90 return Atomic::load(&_max); 91 } 92 93 class ObjectMonitorDeflationSafepointer : public StackObj { 94 JavaThread* const _current; 95 ObjectMonitorDeflationLogging* const _log; 96 97 public: 98 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 99 : _current(current), _log(log) {} 100 101 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 102 }; 103 104 // Walk the in-use list and unlink deflated ObjectMonitors. 105 // Returns the number of unlinked ObjectMonitors. 106 size_t MonitorList::unlink_deflated(size_t deflated_count, 107 GrowableArray<ObjectMonitor*>* unlinked_list, 108 ObjectMonitorDeflationSafepointer* safepointer) { 109 size_t unlinked_count = 0; 110 ObjectMonitor* prev = nullptr; 111 ObjectMonitor* m = Atomic::load_acquire(&_head); 112 113 while (m != nullptr) { 114 if (m->is_being_async_deflated()) { 115 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 116 // modify the list once per batch. The batch starts at "m". 117 size_t unlinked_batch = 0; 118 ObjectMonitor* next = m; 119 // Look for at most MonitorUnlinkBatch monitors, or the number of 120 // deflated and not unlinked monitors, whatever comes first. 121 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 122 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 123 do { 124 ObjectMonitor* next_next = next->next_om(); 125 unlinked_batch++; 126 unlinked_list->append(next); 127 next = next_next; 128 if (unlinked_batch >= unlinked_batch_limit) { 129 // Reached the max batch, so bail out of the gathering loop. 130 break; 131 } 132 if (prev == nullptr && Atomic::load(&_head) != m) { 133 // Current batch used to be at head, but it is not at head anymore. 134 // Bail out and figure out where we currently are. This avoids long 135 // walks searching for new prev during unlink under heavy list inserts. 136 break; 137 } 138 } while (next != nullptr && next->is_being_async_deflated()); 139 140 // Unlink the found batch. 141 if (prev == nullptr) { 142 // The current batch is the first batch, so there is a chance that it starts at head. 143 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 144 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next); 145 if (prev_head != m) { 146 // Something must have updated the head. Figure out the actual prev for this batch. 147 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 148 prev = n; 149 } 150 assert(prev != nullptr, "Should have found the prev for the current batch"); 151 prev->set_next_om(next); 152 } 153 } else { 154 // The current batch is preceded by another batch. This guarantees the current batch 155 // does not start at head. Unlink the entire current batch without updating the head. 156 assert(Atomic::load(&_head) != m, "Sanity"); 157 prev->set_next_om(next); 158 } 159 160 unlinked_count += unlinked_batch; 161 if (unlinked_count >= deflated_count) { 162 // Reached the max so bail out of the searching loop. 163 // There should be no more deflated monitors left. 164 break; 165 } 166 m = next; 167 } else { 168 prev = m; 169 m = m->next_om(); 170 } 171 172 // Must check for a safepoint/handshake and honor it. 173 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 174 } 175 176 #ifdef ASSERT 177 // Invariant: the code above should unlink all deflated monitors. 178 // The code that runs after this unlinking does not expect deflated monitors. 179 // Notably, attempting to deflate the already deflated monitor would break. 180 { 181 ObjectMonitor* m = Atomic::load_acquire(&_head); 182 while (m != nullptr) { 183 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 184 m = m->next_om(); 185 } 186 } 187 #endif 188 189 Atomic::sub(&_count, unlinked_count); 190 return unlinked_count; 191 } 192 193 MonitorList::Iterator MonitorList::iterator() const { 194 return Iterator(Atomic::load_acquire(&_head)); 195 } 196 197 ObjectMonitor* MonitorList::Iterator::next() { 198 ObjectMonitor* current = _current; 199 _current = current->next_om(); 200 return current; 201 } 202 203 // The "core" versions of monitor enter and exit reside in this file. 204 // The interpreter and compilers contain specialized transliterated 205 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 206 // fast_lock(...) for instance. If you make changes here, make sure to modify the 207 // interpreter, and both C1 and C2 fast-path inline locking code emission. 208 // 209 // ----------------------------------------------------------------------------- 210 211 #ifdef DTRACE_ENABLED 212 213 // Only bother with this argument setup if dtrace is available 214 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 215 216 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 217 char* bytes = nullptr; \ 218 int len = 0; \ 219 jlong jtid = SharedRuntime::get_java_tid(thread); \ 220 Symbol* klassname = obj->klass()->name(); \ 221 if (klassname != nullptr) { \ 222 bytes = (char*)klassname->bytes(); \ 223 len = klassname->utf8_length(); \ 224 } 225 226 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 227 { \ 228 if (DTraceMonitorProbes) { \ 229 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 230 HOTSPOT_MONITOR_WAIT(jtid, \ 231 (uintptr_t)(monitor), bytes, len, (millis)); \ 232 } \ 233 } 234 235 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 236 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 237 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 238 239 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 240 { \ 241 if (DTraceMonitorProbes) { \ 242 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 243 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 244 (uintptr_t)(monitor), bytes, len); \ 245 } \ 246 } 247 248 #else // ndef DTRACE_ENABLED 249 250 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 251 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 252 253 #endif // ndef DTRACE_ENABLED 254 255 // This exists only as a workaround of dtrace bug 6254741 256 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 257 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 258 return 0; 259 } 260 261 static constexpr size_t inflation_lock_count() { 262 return 256; 263 } 264 265 // Static storage for an array of PlatformMutex. 266 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 267 268 static inline PlatformMutex* inflation_lock(size_t index) { 269 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 270 } 271 272 void ObjectSynchronizer::initialize() { 273 for (size_t i = 0; i < inflation_lock_count(); i++) { 274 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 275 } 276 // Start the ceiling with the estimate for one thread. 277 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 278 279 // Start the timer for deflations, so it does not trigger immediately. 280 _last_async_deflation_time_ns = os::javaTimeNanos(); 281 282 if (LockingMode == LM_LIGHTWEIGHT) { 283 LightweightSynchronizer::initialize(); 284 } 285 } 286 287 MonitorList ObjectSynchronizer::_in_use_list; 288 // monitors_used_above_threshold() policy is as follows: 289 // 290 // The ratio of the current _in_use_list count to the ceiling is used 291 // to determine if we are above MonitorUsedDeflationThreshold and need 292 // to do an async monitor deflation cycle. The ceiling is increased by 293 // AvgMonitorsPerThreadEstimate when a thread is added to the system 294 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 295 // removed from the system. 296 // 297 // Note: If the _in_use_list max exceeds the ceiling, then 298 // monitors_used_above_threshold() will use the in_use_list max instead 299 // of the thread count derived ceiling because we have used more 300 // ObjectMonitors than the estimated average. 301 // 302 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 303 // no-progress async monitor deflation cycles in a row, then the ceiling 304 // is adjusted upwards by monitors_used_above_threshold(). 305 // 306 // Start the ceiling with the estimate for one thread in initialize() 307 // which is called after cmd line options are processed. 308 static size_t _in_use_list_ceiling = 0; 309 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 310 bool volatile ObjectSynchronizer::_is_final_audit = false; 311 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 312 static uintx _no_progress_cnt = 0; 313 static bool _no_progress_skip_increment = false; 314 315 // These checks are required for wait, notify and exit to avoid inflating the monitor to 316 // find out this inline type object cannot be locked. 317 #define CHECK_THROW_NOSYNC_IMSE(obj) \ 318 if (EnableValhalla && (obj)->mark().is_inline_type()) { \ 319 JavaThread* THREAD = current; \ 320 ResourceMark rm(THREAD); \ 321 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), obj->klass()->external_name()); \ 322 } 323 324 #define CHECK_THROW_NOSYNC_IMSE_0(obj) \ 325 if (EnableValhalla && (obj)->mark().is_inline_type()) { \ 326 JavaThread* THREAD = current; \ 327 ResourceMark rm(THREAD); \ 328 THROW_MSG_0(vmSymbols::java_lang_IllegalMonitorStateException(), obj->klass()->external_name()); \ 329 } 330 331 // =====================> Quick functions 332 333 // The quick_* forms are special fast-path variants used to improve 334 // performance. In the simplest case, a "quick_*" implementation could 335 // simply return false, in which case the caller will perform the necessary 336 // state transitions and call the slow-path form. 337 // The fast-path is designed to handle frequently arising cases in an efficient 338 // manner and is just a degenerate "optimistic" variant of the slow-path. 339 // returns true -- to indicate the call was satisfied. 340 // returns false -- to indicate the call needs the services of the slow-path. 341 // A no-loitering ordinance is in effect for code in the quick_* family 342 // operators: safepoints or indefinite blocking (blocking that might span a 343 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 344 // entry. 345 // 346 // Consider: An interesting optimization is to have the JIT recognize the 347 // following common idiom: 348 // synchronized (someobj) { .... ; notify(); } 349 // That is, we find a notify() or notifyAll() call that immediately precedes 350 // the monitorexit operation. In that case the JIT could fuse the operations 351 // into a single notifyAndExit() runtime primitive. 352 353 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 354 assert(current->thread_state() == _thread_in_Java, "invariant"); 355 NoSafepointVerifier nsv; 356 if (obj == nullptr) return false; // slow-path for invalid obj 357 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "monitor op on inline type"); 358 const markWord mark = obj->mark(); 359 360 if (LockingMode == LM_LIGHTWEIGHT) { 361 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 362 // Degenerate notify 363 // fast-locked by caller so by definition the implied waitset is empty. 364 return true; 365 } 366 } else if (LockingMode == LM_LEGACY) { 367 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 368 // Degenerate notify 369 // stack-locked by caller so by definition the implied waitset is empty. 370 return true; 371 } 372 } 373 374 if (mark.has_monitor()) { 375 ObjectMonitor* const mon = read_monitor(current, obj, mark); 376 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) { 377 // Racing with inflation/deflation go slow path 378 return false; 379 } 380 assert(mon->object() == oop(obj), "invariant"); 381 if (!mon->has_owner(current)) return false; // slow-path for IMS exception 382 383 if (mon->first_waiter() != nullptr) { 384 // We have one or more waiters. Since this is an inflated monitor 385 // that we own, we can transfer one or more threads from the waitset 386 // to the entrylist here and now, avoiding the slow-path. 387 if (all) { 388 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current); 389 } else { 390 DTRACE_MONITOR_PROBE(notify, mon, obj, current); 391 } 392 int free_count = 0; 393 do { 394 mon->INotify(current); 395 ++free_count; 396 } while (mon->first_waiter() != nullptr && all); 397 OM_PERFDATA_OP(Notifications, inc(free_count)); 398 } 399 return true; 400 } 401 402 // other IMS exception states take the slow-path 403 return false; 404 } 405 406 static bool useHeavyMonitors() { 407 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390) 408 return LockingMode == LM_MONITOR; 409 #else 410 return false; 411 #endif 412 } 413 414 // The LockNode emitted directly at the synchronization site would have 415 // been too big if it were to have included support for the cases of inflated 416 // recursive enter and exit, so they go here instead. 417 // Note that we can't safely call AsyncPrintJavaStack() from within 418 // quick_enter() as our thread state remains _in_Java. 419 420 bool ObjectSynchronizer::quick_enter_legacy(oop obj, BasicLock* lock, JavaThread* current) { 421 assert(current->thread_state() == _thread_in_Java, "invariant"); 422 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "monitor op on inline type"); 423 424 if (useHeavyMonitors()) { 425 return false; // Slow path 426 } 427 428 if (LockingMode == LM_LIGHTWEIGHT) { 429 return LightweightSynchronizer::quick_enter(obj, lock, current); 430 } 431 432 assert(LockingMode == LM_LEGACY, "legacy mode below"); 433 434 const markWord mark = obj->mark(); 435 436 if (mark.has_monitor()) { 437 438 ObjectMonitor* const m = read_monitor(mark); 439 // An async deflation or GC can race us before we manage to make 440 // the ObjectMonitor busy by setting the owner below. If we detect 441 // that race we just bail out to the slow-path here. 442 if (m->object_peek() == nullptr) { 443 return false; 444 } 445 446 // Lock contention and Transactional Lock Elision (TLE) diagnostics 447 // and observability 448 // Case: light contention possibly amenable to TLE 449 // Case: TLE inimical operations such as nested/recursive synchronization 450 451 if (m->has_owner(current)) { 452 m->_recursions++; 453 current->inc_held_monitor_count(); 454 return true; 455 } 456 457 // This Java Monitor is inflated so obj's header will never be 458 // displaced to this thread's BasicLock. Make the displaced header 459 // non-null so this BasicLock is not seen as recursive nor as 460 // being locked. We do this unconditionally so that this thread's 461 // BasicLock cannot be mis-interpreted by any stack walkers. For 462 // performance reasons, stack walkers generally first check for 463 // stack-locking in the object's header, the second check is for 464 // recursive stack-locking in the displaced header in the BasicLock, 465 // and last are the inflated Java Monitor (ObjectMonitor) checks. 466 lock->set_displaced_header(markWord::unused_mark()); 467 468 if (!m->has_owner() && m->try_set_owner(current)) { 469 assert(m->_recursions == 0, "invariant"); 470 current->inc_held_monitor_count(); 471 return true; 472 } 473 } 474 475 // Note that we could inflate in quick_enter. 476 // This is likely a useful optimization 477 // Critically, in quick_enter() we must not: 478 // -- block indefinitely, or 479 // -- reach a safepoint 480 481 return false; // revert to slow-path 482 } 483 484 // Handle notifications when synchronizing on value based classes 485 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 486 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 487 frame last_frame = locking_thread->last_frame(); 488 bool bcp_was_adjusted = false; 489 // Don't decrement bcp if it points to the frame's first instruction. This happens when 490 // handle_sync_on_value_based_class() is called because of a synchronized method. There 491 // is no actual monitorenter instruction in the byte code in this case. 492 if (last_frame.is_interpreted_frame() && 493 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 494 // adjust bcp to point back to monitorenter so that we print the correct line numbers 495 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 496 bcp_was_adjusted = true; 497 } 498 499 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 500 ResourceMark rm; 501 stringStream ss; 502 locking_thread->print_active_stack_on(&ss); 503 char* base = (char*)strstr(ss.base(), "at"); 504 char* newline = (char*)strchr(ss.base(), '\n'); 505 if (newline != nullptr) { 506 *newline = '\0'; 507 } 508 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 509 } else { 510 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 511 ResourceMark rm; 512 Log(valuebasedclasses) vblog; 513 514 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 515 if (locking_thread->has_last_Java_frame()) { 516 LogStream info_stream(vblog.info()); 517 locking_thread->print_active_stack_on(&info_stream); 518 } else { 519 vblog.info("Cannot find the last Java frame"); 520 } 521 522 EventSyncOnValueBasedClass event; 523 if (event.should_commit()) { 524 event.set_valueBasedClass(obj->klass()); 525 event.commit(); 526 } 527 } 528 529 if (bcp_was_adjusted) { 530 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 531 } 532 } 533 534 // ----------------------------------------------------------------------------- 535 // Monitor Enter/Exit 536 537 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 538 // When called with locking_thread != Thread::current() some mechanism must synchronize 539 // the locking_thread with respect to the current thread. Currently only used when 540 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 541 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 542 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "JITed code should never have locked an instance of a value class"); 543 544 if (LockingMode == LM_LIGHTWEIGHT) { 545 return LightweightSynchronizer::enter_for(obj, lock, locking_thread); 546 } 547 548 if (!enter_fast_impl(obj, lock, locking_thread)) { 549 // Inflated ObjectMonitor::enter_for is required 550 551 // An async deflation can race after the inflate_for() call and before 552 // enter_for() can make the ObjectMonitor busy. enter_for() returns false 553 // if we have lost the race to async deflation and we simply try again. 554 while (true) { 555 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter); 556 if (monitor->enter_for(locking_thread)) { 557 return; 558 } 559 assert(monitor->is_being_async_deflated(), "must be"); 560 } 561 } 562 } 563 564 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) { 565 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "This method should never be called on an instance of an inline class"); 566 if (!enter_fast_impl(obj, lock, current)) { 567 // Inflated ObjectMonitor::enter is required 568 569 // An async deflation can race after the inflate() call and before 570 // enter() can make the ObjectMonitor busy. enter() returns false if 571 // we have lost the race to async deflation and we simply try again. 572 while (true) { 573 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter); 574 if (monitor->enter(current)) { 575 return; 576 } 577 } 578 } 579 } 580 581 // The interpreter and compiler assembly code tries to lock using the fast path 582 // of this algorithm. Make sure to update that code if the following function is 583 // changed. The implementation is extremely sensitive to race condition. Be careful. 584 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 585 guarantee(!EnableValhalla || !obj->klass()->is_inline_klass(), "Attempt to inflate inline type"); 586 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer"); 587 588 if (obj->klass()->is_value_based()) { 589 handle_sync_on_value_based_class(obj, locking_thread); 590 } 591 592 locking_thread->inc_held_monitor_count(); 593 594 if (!useHeavyMonitors()) { 595 if (LockingMode == LM_LEGACY) { 596 markWord mark = obj->mark(); 597 if (mark.is_unlocked()) { 598 // Anticipate successful CAS -- the ST of the displaced mark must 599 // be visible <= the ST performed by the CAS. 600 lock->set_displaced_header(mark); 601 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 602 return true; 603 } 604 } else if (mark.has_locker() && 605 locking_thread->is_lock_owned((address) mark.locker())) { 606 assert(lock != mark.locker(), "must not re-lock the same lock"); 607 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock"); 608 lock->set_displaced_header(markWord::from_pointer(nullptr)); 609 return true; 610 } 611 612 // The object header will never be displaced to this lock, 613 // so it does not matter what the value is, except that it 614 // must be non-zero to avoid looking like a re-entrant lock, 615 // and must not look locked either. 616 lock->set_displaced_header(markWord::unused_mark()); 617 618 // Failed to fast lock. 619 return false; 620 } 621 } else if (VerifyHeavyMonitors) { 622 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 623 } 624 625 return false; 626 } 627 628 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) { 629 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer"); 630 631 if (!useHeavyMonitors()) { 632 markWord mark = object->mark(); 633 if (EnableValhalla && mark.is_inline_type()) { 634 return; 635 } 636 if (LockingMode == LM_LEGACY) { 637 markWord dhw = lock->displaced_header(); 638 if (dhw.value() == 0) { 639 // If the displaced header is null, then this exit matches up with 640 // a recursive enter. No real work to do here except for diagnostics. 641 #ifndef PRODUCT 642 if (mark != markWord::INFLATING()) { 643 // Only do diagnostics if we are not racing an inflation. Simply 644 // exiting a recursive enter of a Java Monitor that is being 645 // inflated is safe; see the has_monitor() comment below. 646 assert(!mark.is_unlocked(), "invariant"); 647 assert(!mark.has_locker() || 648 current->is_lock_owned((address)mark.locker()), "invariant"); 649 if (mark.has_monitor()) { 650 // The BasicLock's displaced_header is marked as a recursive 651 // enter and we have an inflated Java Monitor (ObjectMonitor). 652 // This is a special case where the Java Monitor was inflated 653 // after this thread entered the stack-lock recursively. When a 654 // Java Monitor is inflated, we cannot safely walk the Java 655 // Monitor owner's stack and update the BasicLocks because a 656 // Java Monitor can be asynchronously inflated by a thread that 657 // does not own the Java Monitor. 658 ObjectMonitor* m = read_monitor(mark); 659 assert(m->object()->mark() == mark, "invariant"); 660 assert(m->is_entered(current), "invariant"); 661 } 662 } 663 #endif 664 return; 665 } 666 667 if (mark == markWord::from_pointer(lock)) { 668 // If the object is stack-locked by the current thread, try to 669 // swing the displaced header from the BasicLock back to the mark. 670 assert(dhw.is_neutral(), "invariant"); 671 if (object->cas_set_mark(dhw, mark) == mark) { 672 return; 673 } 674 } 675 } 676 } else if (VerifyHeavyMonitors) { 677 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 678 } 679 680 // We have to take the slow-path of possible inflation and then exit. 681 // The ObjectMonitor* can't be async deflated until ownership is 682 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 683 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal); 684 assert(!monitor->has_anonymous_owner(), "must not be"); 685 monitor->exit(current); 686 } 687 688 // ----------------------------------------------------------------------------- 689 // JNI locks on java objects 690 // NOTE: must use heavy weight monitor to handle jni monitor enter 691 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 692 JavaThread* THREAD = current; 693 // Top native frames in the stack will not be seen if we attempt 694 // preemption, since we start walking from the last Java anchor. 695 NoPreemptMark npm(current); 696 697 if (obj->klass()->is_value_based()) { 698 handle_sync_on_value_based_class(obj, current); 699 } 700 701 if (EnableValhalla && obj->klass()->is_inline_klass()) { 702 ResourceMark rm(THREAD); 703 const char* desc = "Cannot synchronize on an instance of value class "; 704 const char* className = obj->klass()->external_name(); 705 size_t msglen = strlen(desc) + strlen(className) + 1; 706 char* message = NEW_RESOURCE_ARRAY(char, msglen); 707 assert(message != nullptr, "NEW_RESOURCE_ARRAY should have called vm_exit_out_of_memory and not return nullptr"); 708 THROW_MSG(vmSymbols::java_lang_IdentityException(), className); 709 } 710 711 // the current locking is from JNI instead of Java code 712 current->set_current_pending_monitor_is_from_java(false); 713 // An async deflation can race after the inflate() call and before 714 // enter() can make the ObjectMonitor busy. enter() returns false if 715 // we have lost the race to async deflation and we simply try again. 716 while (true) { 717 ObjectMonitor* monitor; 718 bool entered; 719 if (LockingMode == LM_LIGHTWEIGHT) { 720 entered = LightweightSynchronizer::inflate_and_enter(obj(), inflate_cause_jni_enter, current, current) != nullptr; 721 } else { 722 monitor = inflate(current, obj(), inflate_cause_jni_enter); 723 entered = monitor->enter(current); 724 } 725 726 if (entered) { 727 current->inc_held_monitor_count(1, true); 728 break; 729 } 730 } 731 current->set_current_pending_monitor_is_from_java(true); 732 } 733 734 // NOTE: must use heavy weight monitor to handle jni monitor exit 735 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 736 JavaThread* current = THREAD; 737 CHECK_THROW_NOSYNC_IMSE(obj); 738 739 ObjectMonitor* monitor; 740 if (LockingMode == LM_LIGHTWEIGHT) { 741 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK); 742 } else { 743 // The ObjectMonitor* can't be async deflated until ownership is 744 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 745 monitor = inflate(current, obj, inflate_cause_jni_exit); 746 } 747 // If this thread has locked the object, exit the monitor. We 748 // intentionally do not use CHECK on check_owner because we must exit the 749 // monitor even if an exception was already pending. 750 if (monitor->check_owner(THREAD)) { 751 monitor->exit(current); 752 current->dec_held_monitor_count(1, true); 753 } 754 } 755 756 // ----------------------------------------------------------------------------- 757 // Internal VM locks on java objects 758 // standard constructor, allows locking failures 759 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) { 760 _thread = thread; 761 _thread->check_for_valid_safepoint_state(); 762 _obj = obj; 763 764 if (_obj() != nullptr) { 765 ObjectSynchronizer::enter(_obj, &_lock, _thread); 766 } 767 } 768 769 ObjectLocker::~ObjectLocker() { 770 if (_obj() != nullptr) { 771 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 772 } 773 } 774 775 776 // ----------------------------------------------------------------------------- 777 // Wait/Notify/NotifyAll 778 // NOTE: must use heavy weight monitor to handle wait() 779 780 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 781 JavaThread* current = THREAD; 782 CHECK_THROW_NOSYNC_IMSE_0(obj); 783 if (millis < 0) { 784 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 785 } 786 787 ObjectMonitor* monitor; 788 if (LockingMode == LM_LIGHTWEIGHT) { 789 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0); 790 } else { 791 // The ObjectMonitor* can't be async deflated because the _waiters 792 // field is incremented before ownership is dropped and decremented 793 // after ownership is regained. 794 monitor = inflate(current, obj(), inflate_cause_wait); 795 } 796 797 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 798 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 799 800 // This dummy call is in place to get around dtrace bug 6254741. Once 801 // that's fixed we can uncomment the following line, remove the call 802 // and change this function back into a "void" func. 803 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 804 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 805 return ret_code; 806 } 807 808 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 809 if (millis < 0) { 810 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 811 } 812 813 ObjectMonitor* monitor; 814 if (LockingMode == LM_LIGHTWEIGHT) { 815 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK); 816 } else { 817 monitor = inflate(THREAD, obj(), inflate_cause_wait); 818 } 819 monitor->wait(millis, false, THREAD); 820 } 821 822 823 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 824 JavaThread* current = THREAD; 825 CHECK_THROW_NOSYNC_IMSE(obj); 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->notify(CHECK); 849 } 850 851 // NOTE: see comment of notify() 852 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 853 JavaThread* current = THREAD; 854 CHECK_THROW_NOSYNC_IMSE(obj); 855 856 markWord mark = obj->mark(); 857 if (LockingMode == LM_LIGHTWEIGHT) { 858 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 859 // Not inflated so there can't be any waiters to notify. 860 return; 861 } 862 } else if (LockingMode == LM_LEGACY) { 863 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 864 // Not inflated so there can't be any waiters to notify. 865 return; 866 } 867 } 868 869 ObjectMonitor* monitor; 870 if (LockingMode == LM_LIGHTWEIGHT) { 871 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 872 } else { 873 // The ObjectMonitor* can't be async deflated until ownership is 874 // dropped by the calling thread. 875 monitor = inflate(current, obj(), inflate_cause_notify); 876 } 877 monitor->notifyAll(CHECK); 878 } 879 880 // ----------------------------------------------------------------------------- 881 // Hash Code handling 882 883 struct SharedGlobals { 884 char _pad_prefix[OM_CACHE_LINE_SIZE]; 885 // This is a highly shared mostly-read variable. 886 // To avoid false-sharing it needs to be the sole occupant of a cache line. 887 volatile int stw_random; 888 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 889 // Hot RW variable -- Sequester to avoid false-sharing 890 volatile int hc_sequence; 891 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 892 }; 893 894 static SharedGlobals GVars; 895 896 static markWord read_stable_mark(oop obj) { 897 markWord mark = obj->mark_acquire(); 898 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) { 899 // New lightweight locking does not use the markWord::INFLATING() protocol. 900 return mark; // normal fast-path return 901 } 902 903 int its = 0; 904 for (;;) { 905 markWord mark = obj->mark_acquire(); 906 if (!mark.is_being_inflated()) { 907 return mark; // normal fast-path return 908 } 909 910 // The object is being inflated by some other thread. 911 // The caller of read_stable_mark() must wait for inflation to complete. 912 // Avoid live-lock. 913 914 ++its; 915 if (its > 10000 || !os::is_MP()) { 916 if (its & 1) { 917 os::naked_yield(); 918 } else { 919 // Note that the following code attenuates the livelock problem but is not 920 // a complete remedy. A more complete solution would require that the inflating 921 // thread hold the associated inflation lock. The following code simply restricts 922 // the number of spinners to at most one. We'll have N-2 threads blocked 923 // on the inflationlock, 1 thread holding the inflation lock and using 924 // a yield/park strategy, and 1 thread in the midst of inflation. 925 // A more refined approach would be to change the encoding of INFLATING 926 // to allow encapsulation of a native thread pointer. Threads waiting for 927 // inflation to complete would use CAS to push themselves onto a singly linked 928 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 929 // and calling park(). When inflation was complete the thread that accomplished inflation 930 // would detach the list and set the markword to inflated with a single CAS and 931 // then for each thread on the list, set the flag and unpark() the thread. 932 933 // Index into the lock array based on the current object address. 934 static_assert(is_power_of_2(inflation_lock_count()), "must be"); 935 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1); 936 int YieldThenBlock = 0; 937 assert(ix < inflation_lock_count(), "invariant"); 938 inflation_lock(ix)->lock(); 939 while (obj->mark_acquire() == markWord::INFLATING()) { 940 // Beware: naked_yield() is advisory and has almost no effect on some platforms 941 // so we periodically call current->_ParkEvent->park(1). 942 // We use a mixed spin/yield/block mechanism. 943 if ((YieldThenBlock++) >= 16) { 944 Thread::current()->_ParkEvent->park(1); 945 } else { 946 os::naked_yield(); 947 } 948 } 949 inflation_lock(ix)->unlock(); 950 } 951 } else { 952 SpinPause(); // SMP-polite spinning 953 } 954 } 955 } 956 957 // hashCode() generation : 958 // 959 // Possibilities: 960 // * MD5Digest of {obj,stw_random} 961 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 962 // * A DES- or AES-style SBox[] mechanism 963 // * One of the Phi-based schemes, such as: 964 // 2654435761 = 2^32 * Phi (golden ratio) 965 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 966 // * A variation of Marsaglia's shift-xor RNG scheme. 967 // * (obj ^ stw_random) is appealing, but can result 968 // in undesirable regularity in the hashCode values of adjacent objects 969 // (objects allocated back-to-back, in particular). This could potentially 970 // result in hashtable collisions and reduced hashtable efficiency. 971 // There are simple ways to "diffuse" the middle address bits over the 972 // generated hashCode values: 973 974 static intptr_t get_next_hash(Thread* current, oop obj) { 975 intptr_t value = 0; 976 if (hashCode == 0) { 977 // This form uses global Park-Miller RNG. 978 // On MP system we'll have lots of RW access to a global, so the 979 // mechanism induces lots of coherency traffic. 980 value = os::random(); 981 } else if (hashCode == 1) { 982 // This variation has the property of being stable (idempotent) 983 // between STW operations. This can be useful in some of the 1-0 984 // synchronization schemes. 985 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 986 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 987 } else if (hashCode == 2) { 988 value = 1; // for sensitivity testing 989 } else if (hashCode == 3) { 990 value = ++GVars.hc_sequence; 991 } else if (hashCode == 4) { 992 value = cast_from_oop<intptr_t>(obj); 993 } else { 994 // Marsaglia's xor-shift scheme with thread-specific state 995 // This is probably the best overall implementation -- we'll 996 // likely make this the default in future releases. 997 unsigned t = current->_hashStateX; 998 t ^= (t << 11); 999 current->_hashStateX = current->_hashStateY; 1000 current->_hashStateY = current->_hashStateZ; 1001 current->_hashStateZ = current->_hashStateW; 1002 unsigned v = current->_hashStateW; 1003 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 1004 current->_hashStateW = v; 1005 value = v; 1006 } 1007 1008 value &= markWord::hash_mask; 1009 if (value == 0) value = 0xBAD; 1010 assert(value != markWord::no_hash, "invariant"); 1011 return value; 1012 } 1013 1014 static intptr_t install_hash_code(Thread* current, oop obj) { 1015 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be"); 1016 1017 markWord mark = obj->mark_acquire(); 1018 for (;;) { 1019 intptr_t hash = mark.hash(); 1020 if (hash != 0) { 1021 return hash; 1022 } 1023 1024 hash = get_next_hash(current, obj); 1025 const markWord old_mark = mark; 1026 const markWord new_mark = old_mark.copy_set_hash(hash); 1027 1028 mark = obj->cas_set_mark(new_mark, old_mark); 1029 if (old_mark == mark) { 1030 return hash; 1031 } 1032 } 1033 } 1034 1035 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 1036 if (EnableValhalla && obj->klass()->is_inline_klass()) { 1037 // VM should be calling bootstrap method 1038 ShouldNotReachHere(); 1039 } 1040 if (UseObjectMonitorTable) { 1041 // Since the monitor isn't in the object header, the hash can simply be 1042 // installed in the object header. 1043 return install_hash_code(current, obj); 1044 } 1045 1046 while (true) { 1047 ObjectMonitor* monitor = nullptr; 1048 markWord temp, test; 1049 intptr_t hash; 1050 markWord mark = read_stable_mark(obj); 1051 if (VerifyHeavyMonitors) { 1052 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)"); 1053 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 1054 } 1055 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) { 1056 hash = mark.hash(); 1057 if (hash != 0) { // if it has a hash, just return it 1058 return hash; 1059 } 1060 hash = get_next_hash(current, obj); // get a new hash 1061 temp = mark.copy_set_hash(hash); // merge the hash into header 1062 // try to install the hash 1063 test = obj->cas_set_mark(temp, mark); 1064 if (test == mark) { // if the hash was installed, return it 1065 return hash; 1066 } 1067 if (LockingMode == LM_LIGHTWEIGHT) { 1068 // CAS failed, retry 1069 continue; 1070 } 1071 // Failed to install the hash. It could be that another thread 1072 // installed the hash just before our attempt or inflation has 1073 // occurred or... so we fall thru to inflate the monitor for 1074 // stability and then install the hash. 1075 } else if (mark.has_monitor()) { 1076 monitor = mark.monitor(); 1077 temp = monitor->header(); 1078 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1079 hash = temp.hash(); 1080 if (hash != 0) { 1081 // It has a hash. 1082 1083 // Separate load of dmw/header above from the loads in 1084 // is_being_async_deflated(). 1085 1086 // dmw/header and _contentions may get written by different threads. 1087 // Make sure to observe them in the same order when having several observers. 1088 OrderAccess::loadload_for_IRIW(); 1089 1090 if (monitor->is_being_async_deflated()) { 1091 // But we can't safely use the hash if we detect that async 1092 // deflation has occurred. So we attempt to restore the 1093 // header/dmw to the object's header so that we only retry 1094 // once if the deflater thread happens to be slow. 1095 monitor->install_displaced_markword_in_object(obj); 1096 continue; 1097 } 1098 return hash; 1099 } 1100 // Fall thru so we only have one place that installs the hash in 1101 // the ObjectMonitor. 1102 } else if (LockingMode == LM_LEGACY && mark.has_locker() 1103 && current->is_Java_thread() 1104 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) { 1105 // This is a stack-lock owned by the calling thread so fetch the 1106 // displaced markWord from the BasicLock on the stack. 1107 temp = mark.displaced_mark_helper(); 1108 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1109 hash = temp.hash(); 1110 if (hash != 0) { // if it has a hash, just return it 1111 return hash; 1112 } 1113 // WARNING: 1114 // The displaced header in the BasicLock on a thread's stack 1115 // is strictly immutable. It CANNOT be changed in ANY cases. 1116 // So we have to inflate the stack-lock into an ObjectMonitor 1117 // even if the current thread owns the lock. The BasicLock on 1118 // a thread's stack can be asynchronously read by other threads 1119 // during an inflate() call so any change to that stack memory 1120 // may not propagate to other threads correctly. 1121 } 1122 1123 // Inflate the monitor to set the hash. 1124 1125 // There's no need to inflate if the mark has already got a monitor. 1126 // NOTE: an async deflation can race after we get the monitor and 1127 // before we can update the ObjectMonitor's header with the hash 1128 // value below. 1129 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code); 1130 // Load ObjectMonitor's header/dmw field and see if it has a hash. 1131 mark = monitor->header(); 1132 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1133 hash = mark.hash(); 1134 if (hash == 0) { // if it does not have a hash 1135 hash = get_next_hash(current, obj); // get a new hash 1136 temp = mark.copy_set_hash(hash) ; // merge the hash into header 1137 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1138 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value()); 1139 test = markWord(v); 1140 if (test != mark) { 1141 // The attempt to update the ObjectMonitor's header/dmw field 1142 // did not work. This can happen if another thread managed to 1143 // merge in the hash just before our cmpxchg(). 1144 // If we add any new usages of the header/dmw field, this code 1145 // will need to be updated. 1146 hash = test.hash(); 1147 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1148 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1149 } 1150 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) { 1151 // If we detect that async deflation has occurred, then we 1152 // attempt to restore the header/dmw to the object's header 1153 // so that we only retry once if the deflater thread happens 1154 // to be slow. 1155 monitor->install_displaced_markword_in_object(obj); 1156 continue; 1157 } 1158 } 1159 // We finally get the hash. 1160 return hash; 1161 } 1162 } 1163 1164 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 1165 Handle h_obj) { 1166 if (EnableValhalla && h_obj->mark().is_inline_type()) { 1167 return false; 1168 } 1169 assert(current == JavaThread::current(), "Can only be called on current thread"); 1170 oop obj = h_obj(); 1171 1172 markWord mark = read_stable_mark(obj); 1173 1174 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1175 // stack-locked case, header points into owner's stack 1176 return current->is_lock_owned((address)mark.locker()); 1177 } 1178 1179 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1180 // fast-locking case, see if lock is in current's lock stack 1181 return current->lock_stack().contains(h_obj()); 1182 } 1183 1184 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) { 1185 ObjectMonitor* monitor = read_monitor(current, obj, mark); 1186 if (monitor != nullptr) { 1187 return monitor->is_entered(current) != 0; 1188 } 1189 // Racing with inflation/deflation, retry 1190 mark = obj->mark_acquire(); 1191 1192 if (mark.is_fast_locked()) { 1193 // Some other thread fast_locked, current could not have held the lock 1194 return false; 1195 } 1196 } 1197 1198 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) { 1199 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1200 // The first stage of async deflation does not affect any field 1201 // used by this comparison so the ObjectMonitor* is usable here. 1202 ObjectMonitor* monitor = read_monitor(mark); 1203 return monitor->is_entered(current) != 0; 1204 } 1205 // Unlocked case, header in place 1206 assert(mark.is_unlocked(), "sanity check"); 1207 return false; 1208 } 1209 1210 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1211 oop obj = h_obj(); 1212 markWord mark = read_stable_mark(obj); 1213 1214 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1215 // stack-locked so header points into owner's stack. 1216 // owning_thread_from_monitor_owner() may also return null here: 1217 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker()); 1218 } 1219 1220 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1221 // fast-locked so get owner from the object. 1222 // owning_thread_from_object() may also return null here: 1223 return Threads::owning_thread_from_object(t_list, h_obj()); 1224 } 1225 1226 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) { 1227 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark); 1228 if (monitor != nullptr) { 1229 return Threads::owning_thread_from_monitor(t_list, monitor); 1230 } 1231 // Racing with inflation/deflation, retry 1232 mark = obj->mark_acquire(); 1233 1234 if (mark.is_fast_locked()) { 1235 // Some other thread fast_locked 1236 return Threads::owning_thread_from_object(t_list, h_obj()); 1237 } 1238 } 1239 1240 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) { 1241 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1242 // The first stage of async deflation does not affect any field 1243 // used by this comparison so the ObjectMonitor* is usable here. 1244 ObjectMonitor* monitor = read_monitor(mark); 1245 assert(monitor != nullptr, "monitor should be non-null"); 1246 // owning_thread_from_monitor() may also return null here: 1247 return Threads::owning_thread_from_monitor(t_list, monitor); 1248 } 1249 1250 // Unlocked case, header in place 1251 // Cannot have assertion since this object may have been 1252 // locked by another thread when reaching here. 1253 // assert(mark.is_unlocked(), "sanity check"); 1254 1255 return nullptr; 1256 } 1257 1258 // Visitors ... 1259 1260 // Iterate over all ObjectMonitors. 1261 template <typename Function> 1262 void ObjectSynchronizer::monitors_iterate(Function function) { 1263 MonitorList::Iterator iter = _in_use_list.iterator(); 1264 while (iter.has_next()) { 1265 ObjectMonitor* monitor = iter.next(); 1266 function(monitor); 1267 } 1268 } 1269 1270 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 1271 // returns true. 1272 template <typename OwnerFilter> 1273 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 1274 monitors_iterate([&](ObjectMonitor* monitor) { 1275 // This function is only called at a safepoint or when the 1276 // target thread is suspended or when the target thread is 1277 // operating on itself. The current closures in use today are 1278 // only interested in an owned ObjectMonitor and ownership 1279 // cannot be dropped under the calling contexts so the 1280 // ObjectMonitor cannot be async deflated. 1281 if (monitor->has_owner() && filter(monitor)) { 1282 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 1283 1284 closure->do_monitor(monitor); 1285 } 1286 }); 1287 } 1288 1289 // Iterate ObjectMonitors where the owner == thread; this does NOT include 1290 // ObjectMonitors where owner is set to a stack-lock address in thread. 1291 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 1292 int64_t key = ObjectMonitor::owner_id_from(thread); 1293 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 1294 return owned_monitors_iterate_filtered(closure, thread_filter); 1295 } 1296 1297 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) { 1298 int64_t key = ObjectMonitor::owner_id_from(vthread); 1299 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 1300 return owned_monitors_iterate_filtered(closure, thread_filter); 1301 } 1302 1303 // Iterate ObjectMonitors owned by any thread. 1304 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 1305 auto all_filter = [&](ObjectMonitor* monitor) { return true; }; 1306 return owned_monitors_iterate_filtered(closure, all_filter); 1307 } 1308 1309 static bool monitors_used_above_threshold(MonitorList* list) { 1310 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 1311 return false; 1312 } 1313 size_t monitors_used = list->count(); 1314 if (monitors_used == 0) { // empty list is easy 1315 return false; 1316 } 1317 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling(); 1318 // Make sure that we use a ceiling value that is not lower than 1319 // previous, not lower than the recorded max used by the system, and 1320 // not lower than the current number of monitors in use (which can 1321 // race ahead of max). The result is guaranteed > 0. 1322 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used); 1323 1324 // Check if our monitor usage is above the threshold: 1325 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 1326 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 1327 // Deflate monitors if over the threshold percentage, unless no 1328 // progress on previous deflations. 1329 bool is_above_threshold = true; 1330 1331 // Check if it's time to adjust the in_use_list_ceiling up, due 1332 // to too many async deflation attempts without any progress. 1333 if (NoAsyncDeflationProgressMax != 0 && 1334 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 1335 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 1336 size_t delta = (size_t)(ceiling * remainder) + 1; 1337 size_t new_ceiling = (ceiling > SIZE_MAX - delta) 1338 ? SIZE_MAX // Overflow, let's clamp new_ceiling. 1339 : ceiling + delta; 1340 1341 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 1342 log_info(monitorinflation)("Too many deflations without progress; " 1343 "bumping in_use_list_ceiling from %zu" 1344 " to %zu", old_ceiling, new_ceiling); 1345 _no_progress_cnt = 0; 1346 ceiling = new_ceiling; 1347 1348 // Check if our monitor usage is still above the threshold: 1349 monitor_usage = (monitors_used * 100LL) / ceiling; 1350 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold; 1351 } 1352 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu" 1353 ", monitor_usage=%zu, threshold=%d", 1354 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 1355 return is_above_threshold; 1356 } 1357 1358 return false; 1359 } 1360 1361 size_t ObjectSynchronizer::in_use_list_ceiling() { 1362 return _in_use_list_ceiling; 1363 } 1364 1365 void ObjectSynchronizer::dec_in_use_list_ceiling() { 1366 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1367 } 1368 1369 void ObjectSynchronizer::inc_in_use_list_ceiling() { 1370 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1371 } 1372 1373 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 1374 _in_use_list_ceiling = new_value; 1375 } 1376 1377 bool ObjectSynchronizer::is_async_deflation_needed() { 1378 if (is_async_deflation_requested()) { 1379 // Async deflation request. 1380 log_info(monitorinflation)("Async deflation needed: explicit request"); 1381 return true; 1382 } 1383 1384 jlong time_since_last = time_since_last_async_deflation_ms(); 1385 1386 if (AsyncDeflationInterval > 0 && 1387 time_since_last > AsyncDeflationInterval && 1388 monitors_used_above_threshold(&_in_use_list)) { 1389 // It's been longer than our specified deflate interval and there 1390 // are too many monitors in use. We don't deflate more frequently 1391 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1392 // in order to not swamp the MonitorDeflationThread. 1393 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 1394 return true; 1395 } 1396 1397 if (GuaranteedAsyncDeflationInterval > 0 && 1398 time_since_last > GuaranteedAsyncDeflationInterval) { 1399 // It's been longer than our specified guaranteed deflate interval. 1400 // We need to clean up the used monitors even if the threshold is 1401 // not reached, to keep the memory utilization at bay when many threads 1402 // touched many monitors. 1403 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) " 1404 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 1405 GuaranteedAsyncDeflationInterval, time_since_last); 1406 1407 // If this deflation has no progress, then it should not affect the no-progress 1408 // tracking, otherwise threshold heuristics would think it was triggered, experienced 1409 // no progress, and needs to backoff more aggressively. In this "no progress" case, 1410 // the generic code would bump the no-progress counter, and we compensate for that 1411 // by telling it to skip the update. 1412 // 1413 // If this deflation has progress, then it should let non-progress tracking 1414 // know about this, otherwise the threshold heuristics would kick in, potentially 1415 // experience no-progress due to aggressive cleanup by this deflation, and think 1416 // it is still in no-progress stride. In this "progress" case, the generic code would 1417 // zero the counter, and we allow it to happen. 1418 _no_progress_skip_increment = true; 1419 1420 return true; 1421 } 1422 1423 return false; 1424 } 1425 1426 void ObjectSynchronizer::request_deflate_idle_monitors() { 1427 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1428 set_is_async_deflation_requested(true); 1429 ml.notify_all(); 1430 } 1431 1432 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 1433 JavaThread* current = JavaThread::current(); 1434 bool ret_code = false; 1435 1436 jlong last_time = last_async_deflation_time_ns(); 1437 1438 request_deflate_idle_monitors(); 1439 1440 const int N_CHECKS = 5; 1441 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1442 if (last_async_deflation_time_ns() > last_time) { 1443 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1444 ret_code = true; 1445 break; 1446 } 1447 { 1448 // JavaThread has to honor the blocking protocol. 1449 ThreadBlockInVM tbivm(current); 1450 os::naked_short_sleep(999); // sleep for almost 1 second 1451 } 1452 } 1453 if (!ret_code) { 1454 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1455 } 1456 1457 return ret_code; 1458 } 1459 1460 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1461 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1462 } 1463 1464 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1465 const oop obj, 1466 ObjectSynchronizer::InflateCause cause) { 1467 assert(event != nullptr, "invariant"); 1468 event->set_monitorClass(obj->klass()); 1469 event->set_address((uintptr_t)(void*)obj); 1470 event->set_cause((u1)cause); 1471 event->commit(); 1472 } 1473 1474 // Fast path code shared by multiple functions 1475 void ObjectSynchronizer::inflate_helper(oop obj) { 1476 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter"); 1477 markWord mark = obj->mark_acquire(); 1478 if (mark.has_monitor()) { 1479 ObjectMonitor* monitor = read_monitor(mark); 1480 markWord dmw = monitor->header(); 1481 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1482 return; 1483 } 1484 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal); 1485 } 1486 1487 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) { 1488 assert(current == Thread::current(), "must be"); 1489 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter"); 1490 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause); 1491 } 1492 1493 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) { 1494 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be"); 1495 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for"); 1496 return inflate_impl(thread, obj, cause); 1497 } 1498 1499 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) { 1500 if (EnableValhalla) { 1501 guarantee(!object->klass()->is_inline_klass(), "Attempt to inflate inline type"); 1502 } 1503 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or 1504 // is suspended throughout the call by some other mechanism. 1505 // The thread might be nullptr when called from a non JavaThread. (As may still be 1506 // the case from FastHashCode). However it is only important for correctness that the 1507 // thread is set when called from ObjectSynchronizer::enter from the owning thread, 1508 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit. 1509 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl"); 1510 EventJavaMonitorInflate event; 1511 1512 for (;;) { 1513 const markWord mark = object->mark_acquire(); 1514 1515 // The mark can be in one of the following states: 1516 // * inflated - If the ObjectMonitor owner is anonymous and the 1517 // locking_thread owns the object lock, then we 1518 // make the locking_thread the ObjectMonitor owner. 1519 // * stack-locked - Coerce it to inflated from stack-locked. 1520 // * INFLATING - Busy wait for conversion from stack-locked to 1521 // inflated. 1522 // * unlocked - Aggressively inflate the object. 1523 1524 // CASE: inflated 1525 if (mark.has_monitor()) { 1526 ObjectMonitor* inf = mark.monitor(); 1527 markWord dmw = inf->header(); 1528 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1529 if (inf->has_anonymous_owner() && locking_thread != nullptr) { 1530 assert(LockingMode == LM_LEGACY, "invariant"); 1531 if (locking_thread->is_lock_owned((address)inf->stack_locker())) { 1532 inf->set_stack_locker(nullptr); 1533 inf->set_owner_from_anonymous(locking_thread); 1534 } 1535 } 1536 return inf; 1537 } 1538 1539 // CASE: inflation in progress - inflating over a stack-lock. 1540 // Some other thread is converting from stack-locked to inflated. 1541 // Only that thread can complete inflation -- other threads must wait. 1542 // The INFLATING value is transient. 1543 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1544 // We could always eliminate polling by parking the thread on some auxiliary list. 1545 if (mark == markWord::INFLATING()) { 1546 read_stable_mark(object); 1547 continue; 1548 } 1549 1550 // CASE: stack-locked 1551 // Could be stack-locked either by current or by some other thread. 1552 // 1553 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting 1554 // to install INFLATING into the mark word. We originally installed INFLATING, 1555 // allocated the ObjectMonitor, and then finally STed the address of the 1556 // ObjectMonitor into the mark. This was correct, but artificially lengthened 1557 // the interval in which INFLATING appeared in the mark, thus increasing 1558 // the odds of inflation contention. If we lose the race to set INFLATING, 1559 // then we just delete the ObjectMonitor and loop around again. 1560 // 1561 LogStreamHandle(Trace, monitorinflation) lsh; 1562 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1563 ObjectMonitor* m = new ObjectMonitor(object); 1564 // Optimistically prepare the ObjectMonitor - anticipate successful CAS 1565 // We do this before the CAS in order to minimize the length of time 1566 // in which INFLATING appears in the mark. 1567 1568 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1569 if (cmp != mark) { 1570 delete m; 1571 continue; // Interference -- just retry 1572 } 1573 1574 // We've successfully installed INFLATING (0) into the mark-word. 1575 // This is the only case where 0 will appear in a mark-word. 1576 // Only the singular thread that successfully swings the mark-word 1577 // to 0 can perform (or more precisely, complete) inflation. 1578 // 1579 // Why do we CAS a 0 into the mark-word instead of just CASing the 1580 // mark-word from the stack-locked value directly to the new inflated state? 1581 // Consider what happens when a thread unlocks a stack-locked object. 1582 // It attempts to use CAS to swing the displaced header value from the 1583 // on-stack BasicLock back into the object header. Recall also that the 1584 // header value (hash code, etc) can reside in (a) the object header, or 1585 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1586 // header in an ObjectMonitor. The inflate() routine must copy the header 1587 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1588 // the while preserving the hashCode stability invariants. If the owner 1589 // decides to release the lock while the value is 0, the unlock will fail 1590 // and control will eventually pass from slow_exit() to inflate. The owner 1591 // will then spin, waiting for the 0 value to disappear. Put another way, 1592 // the 0 causes the owner to stall if the owner happens to try to 1593 // drop the lock (restoring the header from the BasicLock to the object) 1594 // while inflation is in-progress. This protocol avoids races that might 1595 // would otherwise permit hashCode values to change or "flicker" for an object. 1596 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1597 // 0 serves as a "BUSY" inflate-in-progress indicator. 1598 1599 1600 // fetch the displaced mark from the owner's stack. 1601 // The owner can't die or unwind past the lock while our INFLATING 1602 // object is in the mark. Furthermore the owner can't complete 1603 // an unlock on the object, either. 1604 markWord dmw = mark.displaced_mark_helper(); 1605 // Catch if the object's header is not neutral (not locked and 1606 // not marked is what we care about here). 1607 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1608 1609 // Setup monitor fields to proper values -- prepare the monitor 1610 m->set_header(dmw); 1611 1612 // Note that a thread can inflate an object 1613 // that it has stack-locked -- as might happen in wait() -- directly 1614 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom. 1615 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) { 1616 m->set_owner(locking_thread); 1617 } else { 1618 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack, 1619 // and set the stack locker field in the monitor. 1620 m->set_stack_locker(mark.locker()); 1621 m->set_anonymous_owner(); 1622 } 1623 // TODO-FIXME: assert BasicLock->dhw != 0. 1624 1625 // Must preserve store ordering. The monitor state must 1626 // be stable at the time of publishing the monitor address. 1627 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1628 // Release semantics so that above set_object() is seen first. 1629 object->release_set_mark(markWord::encode(m)); 1630 1631 // Once ObjectMonitor is configured and the object is associated 1632 // with the ObjectMonitor, it is safe to allow async deflation: 1633 _in_use_list.add(m); 1634 1635 // Hopefully the performance counters are allocated on distinct cache lines 1636 // to avoid false sharing on MP systems ... 1637 OM_PERFDATA_OP(Inflations, inc()); 1638 if (log_is_enabled(Trace, monitorinflation)) { 1639 ResourceMark rm; 1640 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1641 INTPTR_FORMAT ", type='%s'", p2i(object), 1642 object->mark().value(), object->klass()->external_name()); 1643 } 1644 if (event.should_commit()) { 1645 post_monitor_inflate_event(&event, object, cause); 1646 } 1647 return m; 1648 } 1649 1650 // CASE: unlocked 1651 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1652 // If we know we're inflating for entry it's better to inflate by swinging a 1653 // pre-locked ObjectMonitor pointer into the object header. A successful 1654 // CAS inflates the object *and* confers ownership to the inflating thread. 1655 // In the current implementation we use a 2-step mechanism where we CAS() 1656 // to inflate and then CAS() again to try to swing _owner from null to current. 1657 // An inflateTry() method that we could call from enter() would be useful. 1658 1659 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1660 ObjectMonitor* m = new ObjectMonitor(object); 1661 // prepare m for installation - set monitor to initial state 1662 m->set_header(mark); 1663 1664 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1665 delete m; 1666 m = nullptr; 1667 continue; 1668 // interference - the markword changed - just retry. 1669 // The state-transitions are one-way, so there's no chance of 1670 // live-lock -- "Inflated" is an absorbing state. 1671 } 1672 1673 // Once the ObjectMonitor is configured and object is associated 1674 // with the ObjectMonitor, it is safe to allow async deflation: 1675 _in_use_list.add(m); 1676 1677 // Hopefully the performance counters are allocated on distinct 1678 // cache lines to avoid false sharing on MP systems ... 1679 OM_PERFDATA_OP(Inflations, inc()); 1680 if (log_is_enabled(Trace, monitorinflation)) { 1681 ResourceMark rm; 1682 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark=" 1683 INTPTR_FORMAT ", type='%s'", p2i(object), 1684 object->mark().value(), object->klass()->external_name()); 1685 } 1686 if (event.should_commit()) { 1687 post_monitor_inflate_event(&event, object, cause); 1688 } 1689 return m; 1690 } 1691 } 1692 1693 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1694 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1695 // 1696 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1697 MonitorList::Iterator iter = _in_use_list.iterator(); 1698 size_t deflated_count = 0; 1699 Thread* current = Thread::current(); 1700 1701 while (iter.has_next()) { 1702 if (deflated_count >= (size_t)MonitorDeflationMax) { 1703 break; 1704 } 1705 ObjectMonitor* mid = iter.next(); 1706 if (mid->deflate_monitor(current)) { 1707 deflated_count++; 1708 } 1709 1710 // Must check for a safepoint/handshake and honor it. 1711 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1712 } 1713 1714 return deflated_count; 1715 } 1716 1717 class HandshakeForDeflation : public HandshakeClosure { 1718 public: 1719 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {} 1720 1721 void do_thread(Thread* thread) { 1722 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread=" 1723 INTPTR_FORMAT, p2i(thread)); 1724 if (thread->is_Java_thread()) { 1725 // Clear OM cache 1726 JavaThread* jt = JavaThread::cast(thread); 1727 jt->om_clear_monitor_cache(); 1728 } 1729 } 1730 }; 1731 1732 class VM_RendezvousGCThreads : public VM_Operation { 1733 public: 1734 bool evaluate_at_safepoint() const override { return false; } 1735 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1736 void doit() override { 1737 Universe::heap()->safepoint_synchronize_begin(); 1738 Universe::heap()->safepoint_synchronize_end(); 1739 }; 1740 }; 1741 1742 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1743 ObjectMonitorDeflationSafepointer* safepointer) { 1744 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1745 size_t deleted_count = 0; 1746 for (ObjectMonitor* monitor: *delete_list) { 1747 delete monitor; 1748 deleted_count++; 1749 // A JavaThread must check for a safepoint/handshake and honor it. 1750 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1751 } 1752 return deleted_count; 1753 } 1754 1755 class ObjectMonitorDeflationLogging: public StackObj { 1756 LogStreamHandle(Debug, monitorinflation) _debug; 1757 LogStreamHandle(Info, monitorinflation) _info; 1758 LogStream* _stream; 1759 elapsedTimer _timer; 1760 1761 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1762 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); } 1763 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); } 1764 1765 public: 1766 ObjectMonitorDeflationLogging() 1767 : _debug(), _info(), _stream(nullptr) { 1768 if (_debug.is_enabled()) { 1769 _stream = &_debug; 1770 } else if (_info.is_enabled()) { 1771 _stream = &_info; 1772 } 1773 } 1774 1775 void begin() { 1776 if (_stream != nullptr) { 1777 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1778 ceiling(), count(), max()); 1779 _timer.start(); 1780 } 1781 } 1782 1783 void before_handshake(size_t unlinked_count) { 1784 if (_stream != nullptr) { 1785 _timer.stop(); 1786 _stream->print_cr("before handshaking: unlinked_count=%zu" 1787 ", in_use_list stats: ceiling=%zu, count=" 1788 "%zu, max=%zu", 1789 unlinked_count, ceiling(), count(), max()); 1790 } 1791 } 1792 1793 void after_handshake() { 1794 if (_stream != nullptr) { 1795 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1796 "%zu, count=%zu, max=%zu", 1797 ceiling(), count(), max()); 1798 _timer.start(); 1799 } 1800 } 1801 1802 void end(size_t deflated_count, size_t unlinked_count) { 1803 if (_stream != nullptr) { 1804 _timer.stop(); 1805 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1806 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs", 1807 deflated_count, unlinked_count, _timer.seconds()); 1808 } 1809 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1810 ceiling(), count(), max()); 1811 } 1812 } 1813 1814 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1815 if (_stream != nullptr) { 1816 _timer.stop(); 1817 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling=" 1818 "%zu, count=%zu, max=%zu", 1819 op_name, cnt_name, cnt, ceiling(), count(), max()); 1820 } 1821 } 1822 1823 void after_block_for_safepoint(const char* op_name) { 1824 if (_stream != nullptr) { 1825 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu" 1826 ", count=%zu, max=%zu", op_name, 1827 ceiling(), count(), max()); 1828 _timer.start(); 1829 } 1830 } 1831 }; 1832 1833 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1834 if (!SafepointMechanism::should_process(_current)) { 1835 return; 1836 } 1837 1838 // A safepoint/handshake has started. 1839 _log->before_block_for_safepoint(op_name, count_name, counter); 1840 1841 { 1842 // Honor block request. 1843 ThreadBlockInVM tbivm(_current); 1844 } 1845 1846 _log->after_block_for_safepoint(op_name); 1847 } 1848 1849 // This function is called by the MonitorDeflationThread to deflate 1850 // ObjectMonitors. 1851 size_t ObjectSynchronizer::deflate_idle_monitors() { 1852 JavaThread* current = JavaThread::current(); 1853 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1854 1855 // The async deflation request has been processed. 1856 _last_async_deflation_time_ns = os::javaTimeNanos(); 1857 set_is_async_deflation_requested(false); 1858 1859 ObjectMonitorDeflationLogging log; 1860 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1861 1862 log.begin(); 1863 1864 // Deflate some idle ObjectMonitors. 1865 size_t deflated_count = deflate_monitor_list(&safepointer); 1866 1867 // Unlink the deflated ObjectMonitors from the in-use list. 1868 size_t unlinked_count = 0; 1869 size_t deleted_count = 0; 1870 if (deflated_count > 0) { 1871 ResourceMark rm(current); 1872 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1873 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1874 1875 #ifdef ASSERT 1876 if (UseObjectMonitorTable) { 1877 for (ObjectMonitor* monitor : delete_list) { 1878 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed"); 1879 } 1880 } 1881 #endif 1882 1883 log.before_handshake(unlinked_count); 1884 1885 // A JavaThread needs to handshake in order to safely free the 1886 // ObjectMonitors that were deflated in this cycle. 1887 HandshakeForDeflation hfd_hc; 1888 Handshake::execute(&hfd_hc); 1889 // Also, we sync and desync GC threads around the handshake, so that they can 1890 // safely read the mark-word and look-through to the object-monitor, without 1891 // being afraid that the object-monitor is going away. 1892 VM_RendezvousGCThreads sync_gc; 1893 VMThread::execute(&sync_gc); 1894 1895 log.after_handshake(); 1896 1897 // After the handshake, safely free the ObjectMonitors that were 1898 // deflated and unlinked in this cycle. 1899 1900 // Delete the unlinked ObjectMonitors. 1901 deleted_count = delete_monitors(&delete_list, &safepointer); 1902 assert(unlinked_count == deleted_count, "must be"); 1903 } 1904 1905 log.end(deflated_count, unlinked_count); 1906 1907 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count())); 1908 OM_PERFDATA_OP(Deflations, inc(deflated_count)); 1909 1910 GVars.stw_random = os::random(); 1911 1912 if (deflated_count != 0) { 1913 _no_progress_cnt = 0; 1914 } else if (_no_progress_skip_increment) { 1915 _no_progress_skip_increment = false; 1916 } else { 1917 _no_progress_cnt++; 1918 } 1919 1920 return deflated_count; 1921 } 1922 1923 // Monitor cleanup on JavaThread::exit 1924 1925 // Iterate through monitor cache and attempt to release thread's monitors 1926 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1927 private: 1928 JavaThread* _thread; 1929 1930 public: 1931 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1932 void do_monitor(ObjectMonitor* mid) { 1933 intx rec = mid->complete_exit(_thread); 1934 _thread->dec_held_monitor_count(rec + 1); 1935 } 1936 }; 1937 1938 // Release all inflated monitors owned by current thread. Lightweight monitors are 1939 // ignored. This is meant to be called during JNI thread detach which assumes 1940 // all remaining monitors are heavyweight. All exceptions are swallowed. 1941 // Scanning the extant monitor list can be time consuming. 1942 // A simple optimization is to add a per-thread flag that indicates a thread 1943 // called jni_monitorenter() during its lifetime. 1944 // 1945 // Instead of NoSafepointVerifier it might be cheaper to 1946 // use an idiom of the form: 1947 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1948 // <code that must not run at safepoint> 1949 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1950 // Since the tests are extremely cheap we could leave them enabled 1951 // for normal product builds. 1952 1953 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1954 assert(current == JavaThread::current(), "must be current Java thread"); 1955 NoSafepointVerifier nsv; 1956 ReleaseJavaMonitorsClosure rjmc(current); 1957 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1958 assert(!current->has_pending_exception(), "Should not be possible"); 1959 current->clear_pending_exception(); 1960 assert(current->held_monitor_count() == 0, "Should not be possible"); 1961 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1962 current->clear_jni_monitor_count(); 1963 } 1964 1965 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1966 switch (cause) { 1967 case inflate_cause_vm_internal: return "VM Internal"; 1968 case inflate_cause_monitor_enter: return "Monitor Enter"; 1969 case inflate_cause_wait: return "Monitor Wait"; 1970 case inflate_cause_notify: return "Monitor Notify"; 1971 case inflate_cause_hash_code: return "Monitor Hash Code"; 1972 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1973 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1974 default: 1975 ShouldNotReachHere(); 1976 } 1977 return "Unknown"; 1978 } 1979 1980 //------------------------------------------------------------------------------ 1981 // Debugging code 1982 1983 u_char* ObjectSynchronizer::get_gvars_addr() { 1984 return (u_char*)&GVars; 1985 } 1986 1987 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1988 return (u_char*)&GVars.hc_sequence; 1989 } 1990 1991 size_t ObjectSynchronizer::get_gvars_size() { 1992 return sizeof(SharedGlobals); 1993 } 1994 1995 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1996 return (u_char*)&GVars.stw_random; 1997 } 1998 1999 // Do the final audit and print of ObjectMonitor stats; must be done 2000 // by the VMThread at VM exit time. 2001 void ObjectSynchronizer::do_final_audit_and_print_stats() { 2002 assert(Thread::current()->is_VM_thread(), "sanity check"); 2003 2004 if (is_final_audit()) { // Only do the audit once. 2005 return; 2006 } 2007 set_is_final_audit(); 2008 log_info(monitorinflation)("Starting the final audit."); 2009 2010 if (log_is_enabled(Info, monitorinflation)) { 2011 LogStreamHandle(Info, monitorinflation) ls; 2012 audit_and_print_stats(&ls, true /* on_exit */); 2013 } 2014 } 2015 2016 // This function can be called by the MonitorDeflationThread or it can be called when 2017 // we are trying to exit the VM. The list walker functions can run in parallel with 2018 // the other list operations. 2019 // Calls to this function can be added in various places as a debugging 2020 // aid. 2021 // 2022 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) { 2023 int error_cnt = 0; 2024 2025 ls->print_cr("Checking in_use_list:"); 2026 chk_in_use_list(ls, &error_cnt); 2027 2028 if (error_cnt == 0) { 2029 ls->print_cr("No errors found in in_use_list checks."); 2030 } else { 2031 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 2032 } 2033 2034 // When exiting, only log the interesting entries at the Info level. 2035 // When called at intervals by the MonitorDeflationThread, log output 2036 // at the Trace level since there can be a lot of it. 2037 if (!on_exit && log_is_enabled(Trace, monitorinflation)) { 2038 LogStreamHandle(Trace, monitorinflation) ls_tr; 2039 log_in_use_monitor_details(&ls_tr, true /* log_all */); 2040 } else if (on_exit) { 2041 log_in_use_monitor_details(ls, false /* log_all */); 2042 } 2043 2044 ls->flush(); 2045 2046 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2047 } 2048 2049 // Check the in_use_list; log the results of the checks. 2050 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 2051 size_t l_in_use_count = _in_use_list.count(); 2052 size_t l_in_use_max = _in_use_list.max(); 2053 out->print_cr("count=%zu, max=%zu", l_in_use_count, 2054 l_in_use_max); 2055 2056 size_t ck_in_use_count = 0; 2057 MonitorList::Iterator iter = _in_use_list.iterator(); 2058 while (iter.has_next()) { 2059 ObjectMonitor* mid = iter.next(); 2060 chk_in_use_entry(mid, out, error_cnt_p); 2061 ck_in_use_count++; 2062 } 2063 2064 if (l_in_use_count == ck_in_use_count) { 2065 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu", 2066 l_in_use_count, ck_in_use_count); 2067 } else { 2068 out->print_cr("WARNING: in_use_count=%zu is not equal to " 2069 "ck_in_use_count=%zu", l_in_use_count, 2070 ck_in_use_count); 2071 } 2072 2073 size_t ck_in_use_max = _in_use_list.max(); 2074 if (l_in_use_max == ck_in_use_max) { 2075 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu", 2076 l_in_use_max, ck_in_use_max); 2077 } else { 2078 out->print_cr("WARNING: in_use_max=%zu is not equal to " 2079 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max); 2080 } 2081 } 2082 2083 // Check an in-use monitor entry; log any errors. 2084 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 2085 int* error_cnt_p) { 2086 if (n->owner_is_DEFLATER_MARKER()) { 2087 // This could happen when monitor deflation blocks for a safepoint. 2088 return; 2089 } 2090 2091 2092 if (n->metadata() == 0) { 2093 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 2094 "have non-null _metadata (header/hash) field.", p2i(n)); 2095 *error_cnt_p = *error_cnt_p + 1; 2096 } 2097 2098 const oop obj = n->object_peek(); 2099 if (obj == nullptr) { 2100 return; 2101 } 2102 2103 const markWord mark = obj->mark(); 2104 if (!mark.has_monitor()) { 2105 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2106 "object does not think it has a monitor: obj=" 2107 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2108 p2i(obj), mark.value()); 2109 *error_cnt_p = *error_cnt_p + 1; 2110 return; 2111 } 2112 2113 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark); 2114 if (n != obj_mon) { 2115 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2116 "object does not refer to the same monitor: obj=" 2117 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2118 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2119 *error_cnt_p = *error_cnt_p + 1; 2120 } 2121 } 2122 2123 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 2124 // flags indicate why the entry is in-use, 'object' and 'object type' 2125 // indicate the associated object and its type. 2126 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 2127 if (_in_use_list.count() > 0) { 2128 stringStream ss; 2129 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)"); 2130 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2131 out->print_cr("%18s %s %18s %18s", 2132 "monitor", "BHL", "object", "object type"); 2133 out->print_cr("================== === ================== =================="); 2134 2135 auto is_interesting = [&](ObjectMonitor* monitor) { 2136 return log_all || monitor->has_owner() || monitor->is_busy(); 2137 }; 2138 2139 monitors_iterate([&](ObjectMonitor* monitor) { 2140 if (is_interesting(monitor)) { 2141 const oop obj = monitor->object_peek(); 2142 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash(); 2143 ResourceMark rm; 2144 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 2145 monitor->is_busy(), hash != 0, monitor->has_owner(), 2146 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 2147 if (monitor->is_busy()) { 2148 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 2149 ss.reset(); 2150 } 2151 out->cr(); 2152 } 2153 }); 2154 } 2155 2156 out->flush(); 2157 }