1 /* 2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "gc/shared/collectedHeap.hpp" 28 #include "jfr/jfrEvents.hpp" 29 #include "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/padded.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "memory/universe.hpp" 35 #include "oops/markWord.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "runtime/atomic.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/lockStack.inline.hpp" 45 #include "runtime/mutexLocker.hpp" 46 #include "runtime/objectMonitor.hpp" 47 #include "runtime/objectMonitor.inline.hpp" 48 #include "runtime/os.inline.hpp" 49 #include "runtime/osThread.hpp" 50 #include "runtime/perfData.hpp" 51 #include "runtime/safepointMechanism.inline.hpp" 52 #include "runtime/safepointVerifiers.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "runtime/stubRoutines.hpp" 55 #include "runtime/synchronizer.hpp" 56 #include "runtime/threads.hpp" 57 #include "runtime/timer.hpp" 58 #include "runtime/trimNativeHeap.hpp" 59 #include "runtime/vframe.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "utilities/align.hpp" 62 #include "utilities/dtrace.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/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); 78 if (count > max()) { 79 Atomic::inc(&_max); 80 } 81 } 82 83 size_t MonitorList::count() const { 84 return Atomic::load(&_count); 85 } 86 87 size_t MonitorList::max() const { 88 return Atomic::load(&_max); 89 } 90 91 class ObjectMonitorDeflationSafepointer : public StackObj { 92 JavaThread* const _current; 93 ObjectMonitorDeflationLogging* const _log; 94 95 public: 96 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 97 : _current(current), _log(log) {} 98 99 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 100 }; 101 102 // Walk the in-use list and unlink deflated ObjectMonitors. 103 // Returns the number of unlinked ObjectMonitors. 104 size_t MonitorList::unlink_deflated(size_t deflated_count, 105 GrowableArray<ObjectMonitor*>* unlinked_list, 106 ObjectMonitorDeflationSafepointer* safepointer) { 107 size_t unlinked_count = 0; 108 ObjectMonitor* prev = nullptr; 109 ObjectMonitor* m = Atomic::load_acquire(&_head); 110 111 while (m != nullptr) { 112 if (m->is_being_async_deflated()) { 113 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 114 // modify the list once per batch. The batch starts at "m". 115 size_t unlinked_batch = 0; 116 ObjectMonitor* next = m; 117 // Look for at most MonitorUnlinkBatch monitors, or the number of 118 // deflated and not unlinked monitors, whatever comes first. 119 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 120 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 121 do { 122 ObjectMonitor* next_next = next->next_om(); 123 unlinked_batch++; 124 unlinked_list->append(next); 125 next = next_next; 126 if (unlinked_batch >= unlinked_batch_limit) { 127 // Reached the max batch, so bail out of the gathering loop. 128 break; 129 } 130 if (prev == nullptr && Atomic::load(&_head) != m) { 131 // Current batch used to be at head, but it is not at head anymore. 132 // Bail out and figure out where we currently are. This avoids long 133 // walks searching for new prev during unlink under heavy list inserts. 134 break; 135 } 136 } while (next != nullptr && next->is_being_async_deflated()); 137 138 // Unlink the found batch. 139 if (prev == nullptr) { 140 // The current batch is the first batch, so there is a chance that it starts at head. 141 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 142 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next); 143 if (prev_head != m) { 144 // Something must have updated the head. Figure out the actual prev for this batch. 145 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 146 prev = n; 147 } 148 assert(prev != nullptr, "Should have found the prev for the current batch"); 149 prev->set_next_om(next); 150 } 151 } else { 152 // The current batch is preceded by another batch. This guarantees the current batch 153 // does not start at head. Unlink the entire current batch without updating the head. 154 assert(Atomic::load(&_head) != m, "Sanity"); 155 prev->set_next_om(next); 156 } 157 158 unlinked_count += unlinked_batch; 159 if (unlinked_count >= deflated_count) { 160 // Reached the max so bail out of the searching loop. 161 // There should be no more deflated monitors left. 162 break; 163 } 164 m = next; 165 } else { 166 prev = m; 167 m = m->next_om(); 168 } 169 170 // Must check for a safepoint/handshake and honor it. 171 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 172 } 173 174 #ifdef ASSERT 175 // Invariant: the code above should unlink all deflated monitors. 176 // The code that runs after this unlinking does not expect deflated monitors. 177 // Notably, attempting to deflate the already deflated monitor would break. 178 { 179 ObjectMonitor* m = Atomic::load_acquire(&_head); 180 while (m != nullptr) { 181 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 182 m = m->next_om(); 183 } 184 } 185 #endif 186 187 Atomic::sub(&_count, unlinked_count); 188 return unlinked_count; 189 } 190 191 MonitorList::Iterator MonitorList::iterator() const { 192 return Iterator(Atomic::load_acquire(&_head)); 193 } 194 195 ObjectMonitor* MonitorList::Iterator::next() { 196 ObjectMonitor* current = _current; 197 _current = current->next_om(); 198 return current; 199 } 200 201 // The "core" versions of monitor enter and exit reside in this file. 202 // The interpreter and compilers contain specialized transliterated 203 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 204 // fast_lock(...) for instance. If you make changes here, make sure to modify the 205 // interpreter, and both C1 and C2 fast-path inline locking code emission. 206 // 207 // ----------------------------------------------------------------------------- 208 209 #ifdef DTRACE_ENABLED 210 211 // Only bother with this argument setup if dtrace is available 212 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 213 214 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 215 char* bytes = nullptr; \ 216 int len = 0; \ 217 jlong jtid = SharedRuntime::get_java_tid(thread); \ 218 Symbol* klassname = obj->klass()->name(); \ 219 if (klassname != nullptr) { \ 220 bytes = (char*)klassname->bytes(); \ 221 len = klassname->utf8_length(); \ 222 } 223 224 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 225 { \ 226 if (DTraceMonitorProbes) { \ 227 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 228 HOTSPOT_MONITOR_WAIT(jtid, \ 229 (uintptr_t)(monitor), bytes, len, (millis)); \ 230 } \ 231 } 232 233 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 234 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 235 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 236 237 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 238 { \ 239 if (DTraceMonitorProbes) { \ 240 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 241 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 242 (uintptr_t)(monitor), bytes, len); \ 243 } \ 244 } 245 246 #else // ndef DTRACE_ENABLED 247 248 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 249 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 250 251 #endif // ndef DTRACE_ENABLED 252 253 // This exists only as a workaround of dtrace bug 6254741 254 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 255 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 256 return 0; 257 } 258 259 static constexpr size_t inflation_lock_count() { 260 return 256; 261 } 262 263 // Static storage for an array of PlatformMutex. 264 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 265 266 static inline PlatformMutex* inflation_lock(size_t index) { 267 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 268 } 269 270 void ObjectSynchronizer::initialize() { 271 for (size_t i = 0; i < inflation_lock_count(); i++) { 272 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 273 } 274 // Start the ceiling with the estimate for one thread. 275 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 276 277 // Start the timer for deflations, so it does not trigger immediately. 278 _last_async_deflation_time_ns = os::javaTimeNanos(); 279 } 280 281 MonitorList ObjectSynchronizer::_in_use_list; 282 // monitors_used_above_threshold() policy is as follows: 283 // 284 // The ratio of the current _in_use_list count to the ceiling is used 285 // to determine if we are above MonitorUsedDeflationThreshold and need 286 // to do an async monitor deflation cycle. The ceiling is increased by 287 // AvgMonitorsPerThreadEstimate when a thread is added to the system 288 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 289 // removed from the system. 290 // 291 // Note: If the _in_use_list max exceeds the ceiling, then 292 // monitors_used_above_threshold() will use the in_use_list max instead 293 // of the thread count derived ceiling because we have used more 294 // ObjectMonitors than the estimated average. 295 // 296 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 297 // no-progress async monitor deflation cycles in a row, then the ceiling 298 // is adjusted upwards by monitors_used_above_threshold(). 299 // 300 // Start the ceiling with the estimate for one thread in initialize() 301 // which is called after cmd line options are processed. 302 static size_t _in_use_list_ceiling = 0; 303 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 304 bool volatile ObjectSynchronizer::_is_final_audit = false; 305 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 306 static uintx _no_progress_cnt = 0; 307 static bool _no_progress_skip_increment = false; 308 309 // =====================> Quick functions 310 311 // The quick_* forms are special fast-path variants used to improve 312 // performance. In the simplest case, a "quick_*" implementation could 313 // simply return false, in which case the caller will perform the necessary 314 // state transitions and call the slow-path form. 315 // The fast-path is designed to handle frequently arising cases in an efficient 316 // manner and is just a degenerate "optimistic" variant of the slow-path. 317 // returns true -- to indicate the call was satisfied. 318 // returns false -- to indicate the call needs the services of the slow-path. 319 // A no-loitering ordinance is in effect for code in the quick_* family 320 // operators: safepoints or indefinite blocking (blocking that might span a 321 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 322 // entry. 323 // 324 // Consider: An interesting optimization is to have the JIT recognize the 325 // following common idiom: 326 // synchronized (someobj) { .... ; notify(); } 327 // That is, we find a notify() or notifyAll() call that immediately precedes 328 // the monitorexit operation. In that case the JIT could fuse the operations 329 // into a single notifyAndExit() runtime primitive. 330 331 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 332 assert(current->thread_state() == _thread_in_Java, "invariant"); 333 NoSafepointVerifier nsv; 334 if (obj == nullptr) return false; // slow-path for invalid obj 335 const markWord mark = obj->mark(); 336 337 if (LockingMode == LM_LIGHTWEIGHT) { 338 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 339 // Degenerate notify 340 // fast-locked by caller so by definition the implied waitset is empty. 341 return true; 342 } 343 } else if (LockingMode == LM_LEGACY) { 344 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 345 // Degenerate notify 346 // stack-locked by caller so by definition the implied waitset is empty. 347 return true; 348 } 349 } 350 351 if (mark.has_monitor()) { 352 ObjectMonitor* const mon = mark.monitor(); 353 assert(mon->object() == oop(obj), "invariant"); 354 if (mon->owner() != current) return false; // slow-path for IMS exception 355 356 if (mon->first_waiter() != nullptr) { 357 // We have one or more waiters. Since this is an inflated monitor 358 // that we own, we can transfer one or more threads from the waitset 359 // to the entrylist here and now, avoiding the slow-path. 360 if (all) { 361 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current); 362 } else { 363 DTRACE_MONITOR_PROBE(notify, mon, obj, current); 364 } 365 int free_count = 0; 366 do { 367 mon->INotify(current); 368 ++free_count; 369 } while (mon->first_waiter() != nullptr && all); 370 OM_PERFDATA_OP(Notifications, inc(free_count)); 371 } 372 return true; 373 } 374 375 // other IMS exception states take the slow-path 376 return false; 377 } 378 379 380 // The LockNode emitted directly at the synchronization site would have 381 // been too big if it were to have included support for the cases of inflated 382 // recursive enter and exit, so they go here instead. 383 // Note that we can't safely call AsyncPrintJavaStack() from within 384 // quick_enter() as our thread state remains _in_Java. 385 386 bool ObjectSynchronizer::quick_enter(oop obj, JavaThread* current, 387 BasicLock * lock) { 388 assert(current->thread_state() == _thread_in_Java, "invariant"); 389 NoSafepointVerifier nsv; 390 if (obj == nullptr) return false; // Need to throw NPE 391 392 if (obj->klass()->is_value_based()) { 393 return false; 394 } 395 396 if (LockingMode == LM_LIGHTWEIGHT) { 397 LockStack& lock_stack = current->lock_stack(); 398 if (lock_stack.is_full()) { 399 // Always go into runtime if the lock stack is full. 400 return false; 401 } 402 if (lock_stack.try_recursive_enter(obj)) { 403 // Recursive lock successful. 404 current->inc_held_monitor_count(); 405 return true; 406 } 407 } 408 409 const markWord mark = obj->mark(); 410 411 if (mark.has_monitor()) { 412 ObjectMonitor* const m = mark.monitor(); 413 // An async deflation or GC can race us before we manage to make 414 // the ObjectMonitor busy by setting the owner below. If we detect 415 // that race we just bail out to the slow-path here. 416 if (m->object_peek() == nullptr) { 417 return false; 418 } 419 JavaThread* const owner = static_cast<JavaThread*>(m->owner_raw()); 420 421 // Lock contention and Transactional Lock Elision (TLE) diagnostics 422 // and observability 423 // Case: light contention possibly amenable to TLE 424 // Case: TLE inimical operations such as nested/recursive synchronization 425 426 if (owner == current) { 427 m->_recursions++; 428 current->inc_held_monitor_count(); 429 return true; 430 } 431 432 if (LockingMode != LM_LIGHTWEIGHT) { 433 // This Java Monitor is inflated so obj's header will never be 434 // displaced to this thread's BasicLock. Make the displaced header 435 // non-null so this BasicLock is not seen as recursive nor as 436 // being locked. We do this unconditionally so that this thread's 437 // BasicLock cannot be mis-interpreted by any stack walkers. For 438 // performance reasons, stack walkers generally first check for 439 // stack-locking in the object's header, the second check is for 440 // recursive stack-locking in the displaced header in the BasicLock, 441 // and last are the inflated Java Monitor (ObjectMonitor) checks. 442 lock->set_displaced_header(markWord::unused_mark()); 443 } 444 445 if (owner == nullptr && m->try_set_owner_from(nullptr, current) == nullptr) { 446 assert(m->_recursions == 0, "invariant"); 447 current->inc_held_monitor_count(); 448 return true; 449 } 450 } 451 452 // Note that we could inflate in quick_enter. 453 // This is likely a useful optimization 454 // Critically, in quick_enter() we must not: 455 // -- block indefinitely, or 456 // -- reach a safepoint 457 458 return false; // revert to slow-path 459 } 460 461 // Handle notifications when synchronizing on value based classes 462 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 463 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 464 frame last_frame = locking_thread->last_frame(); 465 bool bcp_was_adjusted = false; 466 // Don't decrement bcp if it points to the frame's first instruction. This happens when 467 // handle_sync_on_value_based_class() is called because of a synchronized method. There 468 // is no actual monitorenter instruction in the byte code in this case. 469 if (last_frame.is_interpreted_frame() && 470 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 471 // adjust bcp to point back to monitorenter so that we print the correct line numbers 472 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 473 bcp_was_adjusted = true; 474 } 475 476 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 477 ResourceMark rm; 478 stringStream ss; 479 locking_thread->print_active_stack_on(&ss); 480 char* base = (char*)strstr(ss.base(), "at"); 481 char* newline = (char*)strchr(ss.base(), '\n'); 482 if (newline != nullptr) { 483 *newline = '\0'; 484 } 485 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 486 } else { 487 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 488 ResourceMark rm; 489 Log(valuebasedclasses) vblog; 490 491 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 492 if (locking_thread->has_last_Java_frame()) { 493 LogStream info_stream(vblog.info()); 494 locking_thread->print_active_stack_on(&info_stream); 495 } else { 496 vblog.info("Cannot find the last Java frame"); 497 } 498 499 EventSyncOnValueBasedClass event; 500 if (event.should_commit()) { 501 event.set_valueBasedClass(obj->klass()); 502 event.commit(); 503 } 504 } 505 506 if (bcp_was_adjusted) { 507 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 508 } 509 } 510 511 static bool useHeavyMonitors() { 512 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390) 513 return LockingMode == LM_MONITOR; 514 #else 515 return false; 516 #endif 517 } 518 519 // ----------------------------------------------------------------------------- 520 // Monitor Enter/Exit 521 522 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 523 // When called with locking_thread != Thread::current() some mechanism must synchronize 524 // the locking_thread with respect to the current thread. Currently only used when 525 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 526 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 527 if (!enter_fast_impl(obj, lock, locking_thread)) { 528 // Inflated ObjectMonitor::enter_for is required 529 530 // An async deflation can race after the inflate_for() call and before 531 // enter_for() can make the ObjectMonitor busy. enter_for() returns false 532 // if we have lost the race to async deflation and we simply try again. 533 while (true) { 534 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter); 535 if (monitor->enter_for(locking_thread)) { 536 return; 537 } 538 assert(monitor->is_being_async_deflated(), "must be"); 539 } 540 } 541 } 542 543 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, JavaThread* current) { 544 assert(current == Thread::current(), "must be"); 545 if (!enter_fast_impl(obj, lock, current)) { 546 // Inflated ObjectMonitor::enter is required 547 548 // An async deflation can race after the inflate() call and before 549 // enter() can make the ObjectMonitor busy. enter() returns false if 550 // we have lost the race to async deflation and we simply try again. 551 while (true) { 552 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter); 553 if (monitor->enter(current)) { 554 return; 555 } 556 } 557 } 558 } 559 560 // The interpreter and compiler assembly code tries to lock using the fast path 561 // of this algorithm. Make sure to update that code if the following function is 562 // changed. The implementation is extremely sensitive to race condition. Be careful. 563 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 564 565 if (obj->klass()->is_value_based()) { 566 handle_sync_on_value_based_class(obj, locking_thread); 567 } 568 569 locking_thread->inc_held_monitor_count(); 570 571 if (!useHeavyMonitors()) { 572 if (LockingMode == LM_LIGHTWEIGHT) { 573 // Fast-locking does not use the 'lock' argument. 574 LockStack& lock_stack = locking_thread->lock_stack(); 575 if (lock_stack.is_full()) { 576 // We unconditionally make room on the lock stack by inflating 577 // the least recently locked object on the lock stack. 578 579 // About the choice to inflate least recently locked object. 580 // First we must chose to inflate a lock, either some lock on 581 // the lock-stack or the lock that is currently being entered 582 // (which may or may not be on the lock-stack). 583 // Second the best lock to inflate is a lock which is entered 584 // in a control flow where there are only a very few locks being 585 // used, as the costly part of inflated locking is inflation, 586 // not locking. But this property is entirely program dependent. 587 // Third inflating the lock currently being entered on when it 588 // is not present on the lock-stack will result in a still full 589 // lock-stack. This creates a scenario where every deeper nested 590 // monitorenter must call into the runtime. 591 // The rational here is as follows: 592 // Because we cannot (currently) figure out the second, and want 593 // to avoid the third, we inflate a lock on the lock-stack. 594 // The least recently locked lock is chosen as it is the lock 595 // with the longest critical section. 596 597 log_info(monitorinflation)("LockStack capacity exceeded, inflating."); 598 ObjectMonitor* monitor = inflate_for(locking_thread, lock_stack.bottom(), inflate_cause_vm_internal); 599 assert(monitor->owner() == Thread::current(), "must be owner=" PTR_FORMAT " current=" PTR_FORMAT " mark=" PTR_FORMAT, 600 p2i(monitor->owner()), p2i(Thread::current()), monitor->object()->mark_acquire().value()); 601 assert(!lock_stack.is_full(), "must have made room here"); 602 } 603 604 markWord mark = obj()->mark_acquire(); 605 while (mark.is_neutral()) { 606 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications. 607 // Try to swing into 'fast-locked' state. 608 assert(!lock_stack.contains(obj()), "thread must not already hold the lock"); 609 const markWord locked_mark = mark.set_fast_locked(); 610 const markWord old_mark = obj()->cas_set_mark(locked_mark, mark); 611 if (old_mark == mark) { 612 // Successfully fast-locked, push object to lock-stack and return. 613 lock_stack.push(obj()); 614 return true; 615 } 616 mark = old_mark; 617 } 618 619 if (mark.is_fast_locked() && lock_stack.try_recursive_enter(obj())) { 620 // Recursive lock successful. 621 return true; 622 } 623 624 // Failed to fast lock. 625 return false; 626 } else if (LockingMode == LM_LEGACY) { 627 markWord mark = obj->mark(); 628 if (mark.is_neutral()) { 629 // Anticipate successful CAS -- the ST of the displaced mark must 630 // be visible <= the ST performed by the CAS. 631 lock->set_displaced_header(mark); 632 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 633 return true; 634 } 635 } else if (mark.has_locker() && 636 locking_thread->is_lock_owned((address) mark.locker())) { 637 assert(lock != mark.locker(), "must not re-lock the same lock"); 638 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock"); 639 lock->set_displaced_header(markWord::from_pointer(nullptr)); 640 return true; 641 } 642 643 // The object header will never be displaced to this lock, 644 // so it does not matter what the value is, except that it 645 // must be non-zero to avoid looking like a re-entrant lock, 646 // and must not look locked either. 647 lock->set_displaced_header(markWord::unused_mark()); 648 649 // Failed to fast lock. 650 return false; 651 } 652 } else if (VerifyHeavyMonitors) { 653 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 654 } 655 656 return false; 657 } 658 659 void ObjectSynchronizer::exit(oop object, BasicLock* lock, JavaThread* current) { 660 current->dec_held_monitor_count(); 661 662 if (!useHeavyMonitors()) { 663 markWord mark = object->mark(); 664 if (LockingMode == LM_LIGHTWEIGHT) { 665 // Fast-locking does not use the 'lock' argument. 666 LockStack& lock_stack = current->lock_stack(); 667 if (mark.is_fast_locked() && lock_stack.try_recursive_exit(object)) { 668 // Recursively unlocked. 669 return; 670 } 671 672 if (mark.is_fast_locked() && lock_stack.is_recursive(object)) { 673 // This lock is recursive but is not at the top of the lock stack so we're 674 // doing an unbalanced exit. We have to fall thru to inflation below and 675 // let ObjectMonitor::exit() do the unlock. 676 } else { 677 while (mark.is_fast_locked()) { 678 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications. 679 const markWord unlocked_mark = mark.set_unlocked(); 680 const markWord old_mark = object->cas_set_mark(unlocked_mark, mark); 681 if (old_mark == mark) { 682 size_t recursions = lock_stack.remove(object) - 1; 683 assert(recursions == 0, "must not be recursive here"); 684 return; 685 } 686 mark = old_mark; 687 } 688 } 689 } else if (LockingMode == LM_LEGACY) { 690 markWord dhw = lock->displaced_header(); 691 if (dhw.value() == 0) { 692 // If the displaced header is null, then this exit matches up with 693 // a recursive enter. No real work to do here except for diagnostics. 694 #ifndef PRODUCT 695 if (mark != markWord::INFLATING()) { 696 // Only do diagnostics if we are not racing an inflation. Simply 697 // exiting a recursive enter of a Java Monitor that is being 698 // inflated is safe; see the has_monitor() comment below. 699 assert(!mark.is_neutral(), "invariant"); 700 assert(!mark.has_locker() || 701 current->is_lock_owned((address)mark.locker()), "invariant"); 702 if (mark.has_monitor()) { 703 // The BasicLock's displaced_header is marked as a recursive 704 // enter and we have an inflated Java Monitor (ObjectMonitor). 705 // This is a special case where the Java Monitor was inflated 706 // after this thread entered the stack-lock recursively. When a 707 // Java Monitor is inflated, we cannot safely walk the Java 708 // Monitor owner's stack and update the BasicLocks because a 709 // Java Monitor can be asynchronously inflated by a thread that 710 // does not own the Java Monitor. 711 ObjectMonitor* m = mark.monitor(); 712 assert(m->object()->mark() == mark, "invariant"); 713 assert(m->is_entered(current), "invariant"); 714 } 715 } 716 #endif 717 return; 718 } 719 720 if (mark == markWord::from_pointer(lock)) { 721 // If the object is stack-locked by the current thread, try to 722 // swing the displaced header from the BasicLock back to the mark. 723 assert(dhw.is_neutral(), "invariant"); 724 if (object->cas_set_mark(dhw, mark) == mark) { 725 return; 726 } 727 } 728 } 729 } else if (VerifyHeavyMonitors) { 730 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 731 } 732 733 // We have to take the slow-path of possible inflation and then exit. 734 // The ObjectMonitor* can't be async deflated until ownership is 735 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 736 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal); 737 assert(!monitor->is_owner_anonymous(), "must not be"); 738 monitor->exit(current); 739 } 740 741 // ----------------------------------------------------------------------------- 742 // JNI locks on java objects 743 // NOTE: must use heavy weight monitor to handle jni monitor enter 744 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 745 if (obj->klass()->is_value_based()) { 746 handle_sync_on_value_based_class(obj, current); 747 } 748 749 // the current locking is from JNI instead of Java code 750 current->set_current_pending_monitor_is_from_java(false); 751 // An async deflation can race after the inflate() call and before 752 // enter() can make the ObjectMonitor busy. enter() returns false if 753 // we have lost the race to async deflation and we simply try again. 754 while (true) { 755 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_jni_enter); 756 if (monitor->enter(current)) { 757 current->inc_held_monitor_count(1, true); 758 break; 759 } 760 } 761 current->set_current_pending_monitor_is_from_java(true); 762 } 763 764 // NOTE: must use heavy weight monitor to handle jni monitor exit 765 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 766 JavaThread* current = THREAD; 767 768 // The ObjectMonitor* can't be async deflated until ownership is 769 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 770 ObjectMonitor* monitor = inflate(current, obj, inflate_cause_jni_exit); 771 // If this thread has locked the object, exit the monitor. We 772 // intentionally do not use CHECK on check_owner because we must exit the 773 // monitor even if an exception was already pending. 774 if (monitor->check_owner(THREAD)) { 775 monitor->exit(current); 776 current->dec_held_monitor_count(1, true); 777 } 778 } 779 780 // ----------------------------------------------------------------------------- 781 // Internal VM locks on java objects 782 // standard constructor, allows locking failures 783 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) { 784 _thread = thread; 785 _thread->check_for_valid_safepoint_state(); 786 _obj = obj; 787 788 if (_obj() != nullptr) { 789 ObjectSynchronizer::enter(_obj, &_lock, _thread); 790 } 791 } 792 793 ObjectLocker::~ObjectLocker() { 794 if (_obj() != nullptr) { 795 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 796 } 797 } 798 799 800 // ----------------------------------------------------------------------------- 801 // Wait/Notify/NotifyAll 802 // NOTE: must use heavy weight monitor to handle wait() 803 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 804 JavaThread* current = THREAD; 805 if (millis < 0) { 806 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 807 } 808 // The ObjectMonitor* can't be async deflated because the _waiters 809 // field is incremented before ownership is dropped and decremented 810 // after ownership is regained. 811 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait); 812 813 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 814 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 815 816 // This dummy call is in place to get around dtrace bug 6254741. Once 817 // that's fixed we can uncomment the following line, remove the call 818 // and change this function back into a "void" func. 819 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 820 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 821 return ret_code; 822 } 823 824 void ObjectSynchronizer::notify(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 // The ObjectMonitor* can't be async deflated until ownership is 840 // dropped by the calling thread. 841 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 842 monitor->notify(CHECK); 843 } 844 845 // NOTE: see comment of notify() 846 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 847 JavaThread* current = THREAD; 848 849 markWord mark = obj->mark(); 850 if (LockingMode == LM_LIGHTWEIGHT) { 851 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 852 // Not inflated so there can't be any waiters to notify. 853 return; 854 } 855 } else if (LockingMode == LM_LEGACY) { 856 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 857 // Not inflated so there can't be any waiters to notify. 858 return; 859 } 860 } 861 // The ObjectMonitor* can't be async deflated until ownership is 862 // dropped by the calling thread. 863 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 864 monitor->notifyAll(CHECK); 865 } 866 867 // ----------------------------------------------------------------------------- 868 // Hash Code handling 869 870 struct SharedGlobals { 871 char _pad_prefix[OM_CACHE_LINE_SIZE]; 872 // This is a highly shared mostly-read variable. 873 // To avoid false-sharing it needs to be the sole occupant of a cache line. 874 volatile int stw_random; 875 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 876 // Hot RW variable -- Sequester to avoid false-sharing 877 volatile int hc_sequence; 878 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 879 }; 880 881 static SharedGlobals GVars; 882 883 static markWord read_stable_mark(oop obj) { 884 markWord mark = obj->mark_acquire(); 885 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) { 886 // New lightweight locking does not use the markWord::INFLATING() protocol. 887 return mark; // normal fast-path return 888 } 889 890 int its = 0; 891 for (;;) { 892 markWord mark = obj->mark_acquire(); 893 if (!mark.is_being_inflated()) { 894 return mark; // normal fast-path return 895 } 896 897 // The object is being inflated by some other thread. 898 // The caller of read_stable_mark() must wait for inflation to complete. 899 // Avoid live-lock. 900 901 ++its; 902 if (its > 10000 || !os::is_MP()) { 903 if (its & 1) { 904 os::naked_yield(); 905 } else { 906 // Note that the following code attenuates the livelock problem but is not 907 // a complete remedy. A more complete solution would require that the inflating 908 // thread hold the associated inflation lock. The following code simply restricts 909 // the number of spinners to at most one. We'll have N-2 threads blocked 910 // on the inflationlock, 1 thread holding the inflation lock and using 911 // a yield/park strategy, and 1 thread in the midst of inflation. 912 // A more refined approach would be to change the encoding of INFLATING 913 // to allow encapsulation of a native thread pointer. Threads waiting for 914 // inflation to complete would use CAS to push themselves onto a singly linked 915 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 916 // and calling park(). When inflation was complete the thread that accomplished inflation 917 // would detach the list and set the markword to inflated with a single CAS and 918 // then for each thread on the list, set the flag and unpark() the thread. 919 920 // Index into the lock array based on the current object address. 921 static_assert(is_power_of_2(inflation_lock_count()), "must be"); 922 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1); 923 int YieldThenBlock = 0; 924 assert(ix < inflation_lock_count(), "invariant"); 925 inflation_lock(ix)->lock(); 926 while (obj->mark_acquire() == markWord::INFLATING()) { 927 // Beware: naked_yield() is advisory and has almost no effect on some platforms 928 // so we periodically call current->_ParkEvent->park(1). 929 // We use a mixed spin/yield/block mechanism. 930 if ((YieldThenBlock++) >= 16) { 931 Thread::current()->_ParkEvent->park(1); 932 } else { 933 os::naked_yield(); 934 } 935 } 936 inflation_lock(ix)->unlock(); 937 } 938 } else { 939 SpinPause(); // SMP-polite spinning 940 } 941 } 942 } 943 944 // hashCode() generation : 945 // 946 // Possibilities: 947 // * MD5Digest of {obj,stw_random} 948 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 949 // * A DES- or AES-style SBox[] mechanism 950 // * One of the Phi-based schemes, such as: 951 // 2654435761 = 2^32 * Phi (golden ratio) 952 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 953 // * A variation of Marsaglia's shift-xor RNG scheme. 954 // * (obj ^ stw_random) is appealing, but can result 955 // in undesirable regularity in the hashCode values of adjacent objects 956 // (objects allocated back-to-back, in particular). This could potentially 957 // result in hashtable collisions and reduced hashtable efficiency. 958 // There are simple ways to "diffuse" the middle address bits over the 959 // generated hashCode values: 960 961 static inline intptr_t get_next_hash(Thread* current, oop obj) { 962 intptr_t value = 0; 963 if (hashCode == 0) { 964 // This form uses global Park-Miller RNG. 965 // On MP system we'll have lots of RW access to a global, so the 966 // mechanism induces lots of coherency traffic. 967 value = os::random(); 968 } else if (hashCode == 1) { 969 // This variation has the property of being stable (idempotent) 970 // between STW operations. This can be useful in some of the 1-0 971 // synchronization schemes. 972 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 973 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 974 } else if (hashCode == 2) { 975 value = 1; // for sensitivity testing 976 } else if (hashCode == 3) { 977 value = ++GVars.hc_sequence; 978 } else if (hashCode == 4) { 979 value = cast_from_oop<intptr_t>(obj); 980 } else { 981 // Marsaglia's xor-shift scheme with thread-specific state 982 // This is probably the best overall implementation -- we'll 983 // likely make this the default in future releases. 984 unsigned t = current->_hashStateX; 985 t ^= (t << 11); 986 current->_hashStateX = current->_hashStateY; 987 current->_hashStateY = current->_hashStateZ; 988 current->_hashStateZ = current->_hashStateW; 989 unsigned v = current->_hashStateW; 990 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 991 current->_hashStateW = v; 992 value = v; 993 } 994 995 value &= markWord::hash_mask; 996 if (value == 0) value = 0xBAD; 997 assert(value != markWord::no_hash, "invariant"); 998 return value; 999 } 1000 1001 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 1002 1003 while (true) { 1004 ObjectMonitor* monitor = nullptr; 1005 markWord temp, test; 1006 intptr_t hash; 1007 markWord mark = read_stable_mark(obj); 1008 if (VerifyHeavyMonitors) { 1009 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)"); 1010 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 1011 } 1012 if (mark.is_neutral() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) { 1013 hash = mark.hash(); 1014 if (hash != 0) { // if it has a hash, just return it 1015 return hash; 1016 } 1017 hash = get_next_hash(current, obj); // get a new hash 1018 temp = mark.copy_set_hash(hash); // merge the hash into header 1019 // try to install the hash 1020 test = obj->cas_set_mark(temp, mark); 1021 if (test == mark) { // if the hash was installed, return it 1022 return hash; 1023 } 1024 if (LockingMode == LM_LIGHTWEIGHT) { 1025 // CAS failed, retry 1026 continue; 1027 } 1028 // Failed to install the hash. It could be that another thread 1029 // installed the hash just before our attempt or inflation has 1030 // occurred or... so we fall thru to inflate the monitor for 1031 // stability and then install the hash. 1032 } else if (mark.has_monitor()) { 1033 monitor = mark.monitor(); 1034 temp = monitor->header(); 1035 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1036 hash = temp.hash(); 1037 if (hash != 0) { 1038 // It has a hash. 1039 1040 // Separate load of dmw/header above from the loads in 1041 // is_being_async_deflated(). 1042 1043 // dmw/header and _contentions may get written by different threads. 1044 // Make sure to observe them in the same order when having several observers. 1045 OrderAccess::loadload_for_IRIW(); 1046 1047 if (monitor->is_being_async_deflated()) { 1048 // But we can't safely use the hash if we detect that async 1049 // deflation has occurred. So we attempt to restore the 1050 // header/dmw to the object's header so that we only retry 1051 // once if the deflater thread happens to be slow. 1052 monitor->install_displaced_markword_in_object(obj); 1053 continue; 1054 } 1055 return hash; 1056 } 1057 // Fall thru so we only have one place that installs the hash in 1058 // the ObjectMonitor. 1059 } else if (LockingMode == LM_LEGACY && mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 1060 // This is a stack-lock owned by the calling thread so fetch the 1061 // displaced markWord from the BasicLock on the stack. 1062 temp = mark.displaced_mark_helper(); 1063 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1064 hash = temp.hash(); 1065 if (hash != 0) { // if it has a hash, just return it 1066 return hash; 1067 } 1068 // WARNING: 1069 // The displaced header in the BasicLock on a thread's stack 1070 // is strictly immutable. It CANNOT be changed in ANY cases. 1071 // So we have to inflate the stack-lock into an ObjectMonitor 1072 // even if the current thread owns the lock. The BasicLock on 1073 // a thread's stack can be asynchronously read by other threads 1074 // during an inflate() call so any change to that stack memory 1075 // may not propagate to other threads correctly. 1076 } 1077 1078 // Inflate the monitor to set the hash. 1079 1080 // An async deflation can race after the inflate() call and before we 1081 // can update the ObjectMonitor's header with the hash value below. 1082 monitor = inflate(current, obj, inflate_cause_hash_code); 1083 // Load ObjectMonitor's header/dmw field and see if it has a hash. 1084 mark = monitor->header(); 1085 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1086 hash = mark.hash(); 1087 if (hash == 0) { // if it does not have a hash 1088 hash = get_next_hash(current, obj); // get a new hash 1089 temp = mark.copy_set_hash(hash) ; // merge the hash into header 1090 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1091 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value()); 1092 test = markWord(v); 1093 if (test != mark) { 1094 // The attempt to update the ObjectMonitor's header/dmw field 1095 // did not work. This can happen if another thread managed to 1096 // merge in the hash just before our cmpxchg(). 1097 // If we add any new usages of the header/dmw field, this code 1098 // will need to be updated. 1099 hash = test.hash(); 1100 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1101 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1102 } 1103 if (monitor->is_being_async_deflated()) { 1104 // If we detect that async deflation has occurred, then we 1105 // attempt to restore the header/dmw to the object's header 1106 // so that we only retry once if the deflater thread happens 1107 // to be slow. 1108 monitor->install_displaced_markword_in_object(obj); 1109 continue; 1110 } 1111 } 1112 // We finally get the hash. 1113 return hash; 1114 } 1115 } 1116 1117 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 1118 Handle h_obj) { 1119 assert(current == JavaThread::current(), "Can only be called on current thread"); 1120 oop obj = h_obj(); 1121 1122 markWord mark = read_stable_mark(obj); 1123 1124 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1125 // stack-locked case, header points into owner's stack 1126 return current->is_lock_owned((address)mark.locker()); 1127 } 1128 1129 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1130 // fast-locking case, see if lock is in current's lock stack 1131 return current->lock_stack().contains(h_obj()); 1132 } 1133 1134 if (mark.has_monitor()) { 1135 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1136 // The first stage of async deflation does not affect any field 1137 // used by this comparison so the ObjectMonitor* is usable here. 1138 ObjectMonitor* monitor = mark.monitor(); 1139 return monitor->is_entered(current) != 0; 1140 } 1141 // Unlocked case, header in place 1142 assert(mark.is_neutral(), "sanity check"); 1143 return false; 1144 } 1145 1146 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1147 oop obj = h_obj(); 1148 markWord mark = read_stable_mark(obj); 1149 1150 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1151 // stack-locked so header points into owner's stack. 1152 // owning_thread_from_monitor_owner() may also return null here: 1153 return Threads::owning_thread_from_monitor_owner(t_list, (address) mark.locker()); 1154 } 1155 1156 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1157 // fast-locked so get owner from the object. 1158 // owning_thread_from_object() may also return null here: 1159 return Threads::owning_thread_from_object(t_list, h_obj()); 1160 } 1161 1162 if (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 = mark.monitor(); 1167 assert(monitor != nullptr, "monitor should be non-null"); 1168 // owning_thread_from_monitor() may also return null here: 1169 return Threads::owning_thread_from_monitor(t_list, monitor); 1170 } 1171 1172 // Unlocked case, header in place 1173 // Cannot have assertion since this object may have been 1174 // locked by another thread when reaching here. 1175 // assert(mark.is_neutral(), "sanity check"); 1176 1177 return nullptr; 1178 } 1179 1180 // Visitors ... 1181 1182 // Iterate over all ObjectMonitors. 1183 template <typename Function> 1184 void ObjectSynchronizer::monitors_iterate(Function function) { 1185 MonitorList::Iterator iter = _in_use_list.iterator(); 1186 while (iter.has_next()) { 1187 ObjectMonitor* monitor = iter.next(); 1188 function(monitor); 1189 } 1190 } 1191 1192 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 1193 // returns true. 1194 template <typename OwnerFilter> 1195 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 1196 monitors_iterate([&](ObjectMonitor* monitor) { 1197 // This function is only called at a safepoint or when the 1198 // target thread is suspended or when the target thread is 1199 // operating on itself. The current closures in use today are 1200 // only interested in an owned ObjectMonitor and ownership 1201 // cannot be dropped under the calling contexts so the 1202 // ObjectMonitor cannot be async deflated. 1203 if (monitor->has_owner() && filter(monitor->owner_raw())) { 1204 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 1205 1206 closure->do_monitor(monitor); 1207 } 1208 }); 1209 } 1210 1211 // Iterate ObjectMonitors where the owner == thread; this does NOT include 1212 // ObjectMonitors where owner is set to a stack-lock address in thread. 1213 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 1214 auto thread_filter = [&](void* owner) { return owner == thread; }; 1215 return owned_monitors_iterate_filtered(closure, thread_filter); 1216 } 1217 1218 // Iterate ObjectMonitors owned by any thread. 1219 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 1220 auto all_filter = [&](void* owner) { return true; }; 1221 return owned_monitors_iterate_filtered(closure, all_filter); 1222 } 1223 1224 static bool monitors_used_above_threshold(MonitorList* list) { 1225 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 1226 return false; 1227 } 1228 // Start with ceiling based on a per-thread estimate: 1229 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling(); 1230 size_t old_ceiling = ceiling; 1231 if (ceiling < list->max()) { 1232 // The max used by the system has exceeded the ceiling so use that: 1233 ceiling = list->max(); 1234 } 1235 size_t monitors_used = list->count(); 1236 if (monitors_used == 0) { // empty list is easy 1237 return false; 1238 } 1239 if (NoAsyncDeflationProgressMax != 0 && 1240 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 1241 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 1242 size_t new_ceiling = ceiling + (size_t)((double)ceiling * remainder) + 1; 1243 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 1244 log_info(monitorinflation)("Too many deflations without progress; " 1245 "bumping in_use_list_ceiling from " SIZE_FORMAT 1246 " to " SIZE_FORMAT, old_ceiling, new_ceiling); 1247 _no_progress_cnt = 0; 1248 ceiling = new_ceiling; 1249 } 1250 1251 // Check if our monitor usage is above the threshold: 1252 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 1253 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 1254 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT 1255 ", monitor_usage=" SIZE_FORMAT ", threshold=%d", 1256 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 1257 return true; 1258 } 1259 1260 return false; 1261 } 1262 1263 size_t ObjectSynchronizer::in_use_list_ceiling() { 1264 return _in_use_list_ceiling; 1265 } 1266 1267 void ObjectSynchronizer::dec_in_use_list_ceiling() { 1268 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1269 } 1270 1271 void ObjectSynchronizer::inc_in_use_list_ceiling() { 1272 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1273 } 1274 1275 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 1276 _in_use_list_ceiling = new_value; 1277 } 1278 1279 bool ObjectSynchronizer::is_async_deflation_needed() { 1280 if (is_async_deflation_requested()) { 1281 // Async deflation request. 1282 log_info(monitorinflation)("Async deflation needed: explicit request"); 1283 return true; 1284 } 1285 1286 jlong time_since_last = time_since_last_async_deflation_ms(); 1287 1288 if (AsyncDeflationInterval > 0 && 1289 time_since_last > AsyncDeflationInterval && 1290 monitors_used_above_threshold(&_in_use_list)) { 1291 // It's been longer than our specified deflate interval and there 1292 // are too many monitors in use. We don't deflate more frequently 1293 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1294 // in order to not swamp the MonitorDeflationThread. 1295 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 1296 return true; 1297 } 1298 1299 if (GuaranteedAsyncDeflationInterval > 0 && 1300 time_since_last > GuaranteedAsyncDeflationInterval) { 1301 // It's been longer than our specified guaranteed deflate interval. 1302 // We need to clean up the used monitors even if the threshold is 1303 // not reached, to keep the memory utilization at bay when many threads 1304 // touched many monitors. 1305 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) " 1306 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 1307 GuaranteedAsyncDeflationInterval, time_since_last); 1308 1309 // If this deflation has no progress, then it should not affect the no-progress 1310 // tracking, otherwise threshold heuristics would think it was triggered, experienced 1311 // no progress, and needs to backoff more aggressively. In this "no progress" case, 1312 // the generic code would bump the no-progress counter, and we compensate for that 1313 // by telling it to skip the update. 1314 // 1315 // If this deflation has progress, then it should let non-progress tracking 1316 // know about this, otherwise the threshold heuristics would kick in, potentially 1317 // experience no-progress due to aggressive cleanup by this deflation, and think 1318 // it is still in no-progress stride. In this "progress" case, the generic code would 1319 // zero the counter, and we allow it to happen. 1320 _no_progress_skip_increment = true; 1321 1322 return true; 1323 } 1324 1325 return false; 1326 } 1327 1328 void ObjectSynchronizer::request_deflate_idle_monitors() { 1329 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1330 set_is_async_deflation_requested(true); 1331 ml.notify_all(); 1332 } 1333 1334 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 1335 JavaThread* current = JavaThread::current(); 1336 bool ret_code = false; 1337 1338 jlong last_time = last_async_deflation_time_ns(); 1339 1340 request_deflate_idle_monitors(); 1341 1342 const int N_CHECKS = 5; 1343 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1344 if (last_async_deflation_time_ns() > last_time) { 1345 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1346 ret_code = true; 1347 break; 1348 } 1349 { 1350 // JavaThread has to honor the blocking protocol. 1351 ThreadBlockInVM tbivm(current); 1352 os::naked_short_sleep(999); // sleep for almost 1 second 1353 } 1354 } 1355 if (!ret_code) { 1356 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1357 } 1358 1359 return ret_code; 1360 } 1361 1362 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1363 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1364 } 1365 1366 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1367 const oop obj, 1368 ObjectSynchronizer::InflateCause cause) { 1369 assert(event != nullptr, "invariant"); 1370 event->set_monitorClass(obj->klass()); 1371 event->set_address((uintptr_t)(void*)obj); 1372 event->set_cause((u1)cause); 1373 event->commit(); 1374 } 1375 1376 // Fast path code shared by multiple functions 1377 void ObjectSynchronizer::inflate_helper(oop obj) { 1378 markWord mark = obj->mark_acquire(); 1379 if (mark.has_monitor()) { 1380 ObjectMonitor* monitor = mark.monitor(); 1381 markWord dmw = monitor->header(); 1382 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1383 return; 1384 } 1385 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal); 1386 } 1387 1388 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) { 1389 assert(current == Thread::current(), "must be"); 1390 if (LockingMode == LM_LIGHTWEIGHT && current->is_Java_thread()) { 1391 return inflate_impl(JavaThread::cast(current), obj, cause); 1392 } 1393 return inflate_impl(nullptr, obj, cause); 1394 } 1395 1396 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) { 1397 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be"); 1398 return inflate_impl(thread, obj, cause); 1399 } 1400 1401 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* inflating_thread, oop object, const InflateCause cause) { 1402 // The JavaThread* inflating_thread parameter is only used by LM_LIGHTWEIGHT and requires 1403 // that the inflating_thread == Thread::current() or is suspended throughout the call by 1404 // some other mechanism. 1405 // Even with LM_LIGHTWEIGHT the thread might be nullptr when called from a non 1406 // JavaThread. (As may still be the case from FastHashCode). However it is only 1407 // important for the correctness of the LM_LIGHTWEIGHT algorithm that the thread 1408 // is set when called from ObjectSynchronizer::enter from the owning thread, 1409 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit. 1410 EventJavaMonitorInflate event; 1411 1412 for (;;) { 1413 const markWord mark = object->mark_acquire(); 1414 1415 // The mark can be in one of the following states: 1416 // * inflated - Just return if using stack-locking. 1417 // If using fast-locking and the ObjectMonitor owner 1418 // is anonymous and the inflating_thread owns the 1419 // object lock, then we make the inflating_thread 1420 // the ObjectMonitor owner and remove the lock from 1421 // the inflating_thread's lock stack. 1422 // * fast-locked - Coerce it to inflated from fast-locked. 1423 // * stack-locked - Coerce it to inflated from stack-locked. 1424 // * INFLATING - Busy wait for conversion from stack-locked to 1425 // inflated. 1426 // * neutral - Aggressively inflate the object. 1427 1428 // CASE: inflated 1429 if (mark.has_monitor()) { 1430 ObjectMonitor* inf = mark.monitor(); 1431 markWord dmw = inf->header(); 1432 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1433 if (LockingMode == LM_LIGHTWEIGHT && inf->is_owner_anonymous() && 1434 inflating_thread != nullptr && inflating_thread->lock_stack().contains(object)) { 1435 inf->set_owner_from_anonymous(inflating_thread); 1436 size_t removed = inflating_thread->lock_stack().remove(object); 1437 inf->set_recursions(removed - 1); 1438 } 1439 return inf; 1440 } 1441 1442 if (LockingMode != LM_LIGHTWEIGHT) { 1443 // New lightweight locking does not use INFLATING. 1444 // CASE: inflation in progress - inflating over a stack-lock. 1445 // Some other thread is converting from stack-locked to inflated. 1446 // Only that thread can complete inflation -- other threads must wait. 1447 // The INFLATING value is transient. 1448 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1449 // We could always eliminate polling by parking the thread on some auxiliary list. 1450 if (mark == markWord::INFLATING()) { 1451 read_stable_mark(object); 1452 continue; 1453 } 1454 } 1455 1456 // CASE: fast-locked 1457 // Could be fast-locked either by the inflating_thread or by some other thread. 1458 // 1459 // Note that we allocate the ObjectMonitor speculatively, _before_ 1460 // attempting to set the object's mark to the new ObjectMonitor. If 1461 // the inflating_thread owns the monitor, then we set the ObjectMonitor's 1462 // owner to the inflating_thread. Otherwise, we set the ObjectMonitor's owner 1463 // to anonymous. If we lose the race to set the object's mark to the 1464 // new ObjectMonitor, then we just delete it and loop around again. 1465 // 1466 LogStreamHandle(Trace, monitorinflation) lsh; 1467 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1468 ObjectMonitor* monitor = new ObjectMonitor(object); 1469 monitor->set_header(mark.set_unlocked()); 1470 bool own = inflating_thread != nullptr && inflating_thread->lock_stack().contains(object); 1471 if (own) { 1472 // Owned by inflating_thread. 1473 monitor->set_owner_from(nullptr, inflating_thread); 1474 } else { 1475 // Owned by somebody else. 1476 monitor->set_owner_anonymous(); 1477 } 1478 markWord monitor_mark = markWord::encode(monitor); 1479 markWord old_mark = object->cas_set_mark(monitor_mark, mark); 1480 if (old_mark == mark) { 1481 // Success! Return inflated monitor. 1482 if (own) { 1483 size_t removed = inflating_thread->lock_stack().remove(object); 1484 monitor->set_recursions(removed - 1); 1485 } 1486 // Once the ObjectMonitor is configured and object is associated 1487 // with the ObjectMonitor, it is safe to allow async deflation: 1488 _in_use_list.add(monitor); 1489 1490 // Hopefully the performance counters are allocated on distinct 1491 // cache lines to avoid false sharing on MP systems ... 1492 OM_PERFDATA_OP(Inflations, inc()); 1493 if (log_is_enabled(Trace, monitorinflation)) { 1494 ResourceMark rm; 1495 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1496 INTPTR_FORMAT ", type='%s'", p2i(object), 1497 object->mark().value(), object->klass()->external_name()); 1498 } 1499 if (event.should_commit()) { 1500 post_monitor_inflate_event(&event, object, cause); 1501 } 1502 return monitor; 1503 } else { 1504 delete monitor; 1505 continue; // Interference -- just retry 1506 } 1507 } 1508 1509 // CASE: stack-locked 1510 // Could be stack-locked either by current or by some other thread. 1511 // 1512 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting 1513 // to install INFLATING into the mark word. We originally installed INFLATING, 1514 // allocated the ObjectMonitor, and then finally STed the address of the 1515 // ObjectMonitor into the mark. This was correct, but artificially lengthened 1516 // the interval in which INFLATING appeared in the mark, thus increasing 1517 // the odds of inflation contention. If we lose the race to set INFLATING, 1518 // then we just delete the ObjectMonitor and loop around again. 1519 // 1520 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1521 assert(LockingMode != LM_LIGHTWEIGHT, "cannot happen with new lightweight locking"); 1522 ObjectMonitor* m = new ObjectMonitor(object); 1523 // Optimistically prepare the ObjectMonitor - anticipate successful CAS 1524 // We do this before the CAS in order to minimize the length of time 1525 // in which INFLATING appears in the mark. 1526 1527 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1528 if (cmp != mark) { 1529 delete m; 1530 continue; // Interference -- just retry 1531 } 1532 1533 // We've successfully installed INFLATING (0) into the mark-word. 1534 // This is the only case where 0 will appear in a mark-word. 1535 // Only the singular thread that successfully swings the mark-word 1536 // to 0 can perform (or more precisely, complete) inflation. 1537 // 1538 // Why do we CAS a 0 into the mark-word instead of just CASing the 1539 // mark-word from the stack-locked value directly to the new inflated state? 1540 // Consider what happens when a thread unlocks a stack-locked object. 1541 // It attempts to use CAS to swing the displaced header value from the 1542 // on-stack BasicLock back into the object header. Recall also that the 1543 // header value (hash code, etc) can reside in (a) the object header, or 1544 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1545 // header in an ObjectMonitor. The inflate() routine must copy the header 1546 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1547 // the while preserving the hashCode stability invariants. If the owner 1548 // decides to release the lock while the value is 0, the unlock will fail 1549 // and control will eventually pass from slow_exit() to inflate. The owner 1550 // will then spin, waiting for the 0 value to disappear. Put another way, 1551 // the 0 causes the owner to stall if the owner happens to try to 1552 // drop the lock (restoring the header from the BasicLock to the object) 1553 // while inflation is in-progress. This protocol avoids races that might 1554 // would otherwise permit hashCode values to change or "flicker" for an object. 1555 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1556 // 0 serves as a "BUSY" inflate-in-progress indicator. 1557 1558 1559 // fetch the displaced mark from the owner's stack. 1560 // The owner can't die or unwind past the lock while our INFLATING 1561 // object is in the mark. Furthermore the owner can't complete 1562 // an unlock on the object, either. 1563 markWord dmw = mark.displaced_mark_helper(); 1564 // Catch if the object's header is not neutral (not locked and 1565 // not marked is what we care about here). 1566 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1567 1568 // Setup monitor fields to proper values -- prepare the monitor 1569 m->set_header(dmw); 1570 1571 // Optimization: if the mark.locker stack address is associated 1572 // with this thread we could simply set m->_owner = current. 1573 // Note that a thread can inflate an object 1574 // that it has stack-locked -- as might happen in wait() -- directly 1575 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom. 1576 m->set_owner_from(nullptr, mark.locker()); 1577 // TODO-FIXME: assert BasicLock->dhw != 0. 1578 1579 // Must preserve store ordering. The monitor state must 1580 // be stable at the time of publishing the monitor address. 1581 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1582 // Release semantics so that above set_object() is seen first. 1583 object->release_set_mark(markWord::encode(m)); 1584 1585 // Once ObjectMonitor is configured and the object is associated 1586 // with the ObjectMonitor, it is safe to allow async deflation: 1587 _in_use_list.add(m); 1588 1589 // Hopefully the performance counters are allocated on distinct cache lines 1590 // to avoid false sharing on MP systems ... 1591 OM_PERFDATA_OP(Inflations, inc()); 1592 if (log_is_enabled(Trace, monitorinflation)) { 1593 ResourceMark rm; 1594 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1595 INTPTR_FORMAT ", type='%s'", p2i(object), 1596 object->mark().value(), object->klass()->external_name()); 1597 } 1598 if (event.should_commit()) { 1599 post_monitor_inflate_event(&event, object, cause); 1600 } 1601 return m; 1602 } 1603 1604 // CASE: neutral 1605 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1606 // If we know we're inflating for entry it's better to inflate by swinging a 1607 // pre-locked ObjectMonitor pointer into the object header. A successful 1608 // CAS inflates the object *and* confers ownership to the inflating thread. 1609 // In the current implementation we use a 2-step mechanism where we CAS() 1610 // to inflate and then CAS() again to try to swing _owner from null to current. 1611 // An inflateTry() method that we could call from enter() would be useful. 1612 1613 // Catch if the object's header is not neutral (not locked and 1614 // not marked is what we care about here). 1615 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1616 ObjectMonitor* m = new ObjectMonitor(object); 1617 // prepare m for installation - set monitor to initial state 1618 m->set_header(mark); 1619 1620 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1621 delete m; 1622 m = nullptr; 1623 continue; 1624 // interference - the markword changed - just retry. 1625 // The state-transitions are one-way, so there's no chance of 1626 // live-lock -- "Inflated" is an absorbing state. 1627 } 1628 1629 // Once the ObjectMonitor is configured and object is associated 1630 // with the ObjectMonitor, it is safe to allow async deflation: 1631 _in_use_list.add(m); 1632 1633 // Hopefully the performance counters are allocated on distinct 1634 // cache lines to avoid false sharing on MP systems ... 1635 OM_PERFDATA_OP(Inflations, inc()); 1636 if (log_is_enabled(Trace, monitorinflation)) { 1637 ResourceMark rm; 1638 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 1639 INTPTR_FORMAT ", type='%s'", p2i(object), 1640 object->mark().value(), object->klass()->external_name()); 1641 } 1642 if (event.should_commit()) { 1643 post_monitor_inflate_event(&event, object, cause); 1644 } 1645 return m; 1646 } 1647 } 1648 1649 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1650 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1651 // 1652 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1653 MonitorList::Iterator iter = _in_use_list.iterator(); 1654 size_t deflated_count = 0; 1655 1656 while (iter.has_next()) { 1657 if (deflated_count >= (size_t)MonitorDeflationMax) { 1658 break; 1659 } 1660 ObjectMonitor* mid = iter.next(); 1661 if (mid->deflate_monitor()) { 1662 deflated_count++; 1663 } 1664 1665 // Must check for a safepoint/handshake and honor it. 1666 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1667 } 1668 1669 return deflated_count; 1670 } 1671 1672 class HandshakeForDeflation : public HandshakeClosure { 1673 public: 1674 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {} 1675 1676 void do_thread(Thread* thread) { 1677 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread=" 1678 INTPTR_FORMAT, p2i(thread)); 1679 } 1680 }; 1681 1682 class VM_RendezvousGCThreads : public VM_Operation { 1683 public: 1684 bool evaluate_at_safepoint() const override { return false; } 1685 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1686 void doit() override { 1687 Universe::heap()->safepoint_synchronize_begin(); 1688 Universe::heap()->safepoint_synchronize_end(); 1689 }; 1690 }; 1691 1692 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1693 ObjectMonitorDeflationSafepointer* safepointer) { 1694 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1695 size_t deleted_count = 0; 1696 for (ObjectMonitor* monitor: *delete_list) { 1697 delete monitor; 1698 deleted_count++; 1699 // A JavaThread must check for a safepoint/handshake and honor it. 1700 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1701 } 1702 return deleted_count; 1703 } 1704 1705 class ObjectMonitorDeflationLogging: public StackObj { 1706 LogStreamHandle(Debug, monitorinflation) _debug; 1707 LogStreamHandle(Info, monitorinflation) _info; 1708 LogStream* _stream; 1709 elapsedTimer _timer; 1710 1711 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1712 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); } 1713 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); } 1714 1715 public: 1716 ObjectMonitorDeflationLogging() 1717 : _debug(), _info(), _stream(nullptr) { 1718 if (_debug.is_enabled()) { 1719 _stream = &_debug; 1720 } else if (_info.is_enabled()) { 1721 _stream = &_info; 1722 } 1723 } 1724 1725 void begin() { 1726 if (_stream != nullptr) { 1727 _stream->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1728 ceiling(), count(), max()); 1729 _timer.start(); 1730 } 1731 } 1732 1733 void before_handshake(size_t unlinked_count) { 1734 if (_stream != nullptr) { 1735 _timer.stop(); 1736 _stream->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT 1737 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count=" 1738 SIZE_FORMAT ", max=" SIZE_FORMAT, 1739 unlinked_count, ceiling(), count(), max()); 1740 } 1741 } 1742 1743 void after_handshake() { 1744 if (_stream != nullptr) { 1745 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1746 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1747 ceiling(), count(), max()); 1748 _timer.start(); 1749 } 1750 } 1751 1752 void end(size_t deflated_count, size_t unlinked_count) { 1753 if (_stream != nullptr) { 1754 _timer.stop(); 1755 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1756 _stream->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs", 1757 deflated_count, unlinked_count, _timer.seconds()); 1758 } 1759 _stream->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1760 ceiling(), count(), max()); 1761 } 1762 } 1763 1764 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1765 if (_stream != nullptr) { 1766 _timer.stop(); 1767 _stream->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling=" 1768 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1769 op_name, cnt_name, cnt, ceiling(), count(), max()); 1770 } 1771 } 1772 1773 void after_block_for_safepoint(const char* op_name) { 1774 if (_stream != nullptr) { 1775 _stream->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT 1776 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name, 1777 ceiling(), count(), max()); 1778 _timer.start(); 1779 } 1780 } 1781 }; 1782 1783 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1784 if (!SafepointMechanism::should_process(_current)) { 1785 return; 1786 } 1787 1788 // A safepoint/handshake has started. 1789 _log->before_block_for_safepoint(op_name, count_name, counter); 1790 1791 { 1792 // Honor block request. 1793 ThreadBlockInVM tbivm(_current); 1794 } 1795 1796 _log->after_block_for_safepoint(op_name); 1797 } 1798 1799 // This function is called by the MonitorDeflationThread to deflate 1800 // ObjectMonitors. 1801 size_t ObjectSynchronizer::deflate_idle_monitors() { 1802 JavaThread* current = JavaThread::current(); 1803 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1804 1805 // The async deflation request has been processed. 1806 _last_async_deflation_time_ns = os::javaTimeNanos(); 1807 set_is_async_deflation_requested(false); 1808 1809 ObjectMonitorDeflationLogging log; 1810 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1811 1812 log.begin(); 1813 1814 // Deflate some idle ObjectMonitors. 1815 size_t deflated_count = deflate_monitor_list(&safepointer); 1816 1817 // Unlink the deflated ObjectMonitors from the in-use list. 1818 size_t unlinked_count = 0; 1819 size_t deleted_count = 0; 1820 if (deflated_count > 0) { 1821 ResourceMark rm(current); 1822 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1823 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1824 1825 log.before_handshake(unlinked_count); 1826 1827 // A JavaThread needs to handshake in order to safely free the 1828 // ObjectMonitors that were deflated in this cycle. 1829 HandshakeForDeflation hfd_hc; 1830 Handshake::execute(&hfd_hc); 1831 // Also, we sync and desync GC threads around the handshake, so that they can 1832 // safely read the mark-word and look-through to the object-monitor, without 1833 // being afraid that the object-monitor is going away. 1834 VM_RendezvousGCThreads sync_gc; 1835 VMThread::execute(&sync_gc); 1836 1837 log.after_handshake(); 1838 1839 // After the handshake, safely free the ObjectMonitors that were 1840 // deflated and unlinked in this cycle. 1841 1842 // Delete the unlinked ObjectMonitors. 1843 deleted_count = delete_monitors(&delete_list, &safepointer); 1844 assert(unlinked_count == deleted_count, "must be"); 1845 } 1846 1847 log.end(deflated_count, unlinked_count); 1848 1849 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count())); 1850 OM_PERFDATA_OP(Deflations, inc(deflated_count)); 1851 1852 GVars.stw_random = os::random(); 1853 1854 if (deflated_count != 0) { 1855 _no_progress_cnt = 0; 1856 } else if (_no_progress_skip_increment) { 1857 _no_progress_skip_increment = false; 1858 } else { 1859 _no_progress_cnt++; 1860 } 1861 1862 return deflated_count; 1863 } 1864 1865 // Monitor cleanup on JavaThread::exit 1866 1867 // Iterate through monitor cache and attempt to release thread's monitors 1868 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1869 private: 1870 JavaThread* _thread; 1871 1872 public: 1873 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1874 void do_monitor(ObjectMonitor* mid) { 1875 intx rec = mid->complete_exit(_thread); 1876 _thread->dec_held_monitor_count(rec + 1); 1877 } 1878 }; 1879 1880 // Release all inflated monitors owned by current thread. Lightweight monitors are 1881 // ignored. This is meant to be called during JNI thread detach which assumes 1882 // all remaining monitors are heavyweight. All exceptions are swallowed. 1883 // Scanning the extant monitor list can be time consuming. 1884 // A simple optimization is to add a per-thread flag that indicates a thread 1885 // called jni_monitorenter() during its lifetime. 1886 // 1887 // Instead of NoSafepointVerifier it might be cheaper to 1888 // use an idiom of the form: 1889 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1890 // <code that must not run at safepoint> 1891 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1892 // Since the tests are extremely cheap we could leave them enabled 1893 // for normal product builds. 1894 1895 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1896 assert(current == JavaThread::current(), "must be current Java thread"); 1897 NoSafepointVerifier nsv; 1898 ReleaseJavaMonitorsClosure rjmc(current); 1899 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1900 assert(!current->has_pending_exception(), "Should not be possible"); 1901 current->clear_pending_exception(); 1902 assert(current->held_monitor_count() == 0, "Should not be possible"); 1903 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1904 current->clear_jni_monitor_count(); 1905 } 1906 1907 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1908 switch (cause) { 1909 case inflate_cause_vm_internal: return "VM Internal"; 1910 case inflate_cause_monitor_enter: return "Monitor Enter"; 1911 case inflate_cause_wait: return "Monitor Wait"; 1912 case inflate_cause_notify: return "Monitor Notify"; 1913 case inflate_cause_hash_code: return "Monitor Hash Code"; 1914 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1915 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1916 default: 1917 ShouldNotReachHere(); 1918 } 1919 return "Unknown"; 1920 } 1921 1922 //------------------------------------------------------------------------------ 1923 // Debugging code 1924 1925 u_char* ObjectSynchronizer::get_gvars_addr() { 1926 return (u_char*)&GVars; 1927 } 1928 1929 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1930 return (u_char*)&GVars.hc_sequence; 1931 } 1932 1933 size_t ObjectSynchronizer::get_gvars_size() { 1934 return sizeof(SharedGlobals); 1935 } 1936 1937 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1938 return (u_char*)&GVars.stw_random; 1939 } 1940 1941 // Do the final audit and print of ObjectMonitor stats; must be done 1942 // by the VMThread at VM exit time. 1943 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1944 assert(Thread::current()->is_VM_thread(), "sanity check"); 1945 1946 if (is_final_audit()) { // Only do the audit once. 1947 return; 1948 } 1949 set_is_final_audit(); 1950 log_info(monitorinflation)("Starting the final audit."); 1951 1952 if (log_is_enabled(Info, monitorinflation)) { 1953 LogStreamHandle(Info, monitorinflation) ls; 1954 audit_and_print_stats(&ls, true /* on_exit */); 1955 } 1956 } 1957 1958 // This function can be called by the MonitorDeflationThread or it can be called when 1959 // we are trying to exit the VM. The list walker functions can run in parallel with 1960 // the other list operations. 1961 // Calls to this function can be added in various places as a debugging 1962 // aid. 1963 // 1964 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) { 1965 int error_cnt = 0; 1966 1967 ls->print_cr("Checking in_use_list:"); 1968 chk_in_use_list(ls, &error_cnt); 1969 1970 if (error_cnt == 0) { 1971 ls->print_cr("No errors found in in_use_list checks."); 1972 } else { 1973 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 1974 } 1975 1976 // When exiting, only log the interesting entries at the Info level. 1977 // When called at intervals by the MonitorDeflationThread, log output 1978 // at the Trace level since there can be a lot of it. 1979 if (!on_exit && log_is_enabled(Trace, monitorinflation)) { 1980 LogStreamHandle(Trace, monitorinflation) ls_tr; 1981 log_in_use_monitor_details(&ls_tr, true /* log_all */); 1982 } else if (on_exit) { 1983 log_in_use_monitor_details(ls, false /* log_all */); 1984 } 1985 1986 ls->flush(); 1987 1988 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 1989 } 1990 1991 // Check the in_use_list; log the results of the checks. 1992 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 1993 size_t l_in_use_count = _in_use_list.count(); 1994 size_t l_in_use_max = _in_use_list.max(); 1995 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count, 1996 l_in_use_max); 1997 1998 size_t ck_in_use_count = 0; 1999 MonitorList::Iterator iter = _in_use_list.iterator(); 2000 while (iter.has_next()) { 2001 ObjectMonitor* mid = iter.next(); 2002 chk_in_use_entry(mid, out, error_cnt_p); 2003 ck_in_use_count++; 2004 } 2005 2006 if (l_in_use_count == ck_in_use_count) { 2007 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count=" 2008 SIZE_FORMAT, l_in_use_count, ck_in_use_count); 2009 } else { 2010 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to " 2011 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count, 2012 ck_in_use_count); 2013 } 2014 2015 size_t ck_in_use_max = _in_use_list.max(); 2016 if (l_in_use_max == ck_in_use_max) { 2017 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max=" 2018 SIZE_FORMAT, l_in_use_max, ck_in_use_max); 2019 } else { 2020 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to " 2021 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max); 2022 } 2023 } 2024 2025 // Check an in-use monitor entry; log any errors. 2026 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 2027 int* error_cnt_p) { 2028 if (n->owner_is_DEFLATER_MARKER()) { 2029 // This could happen when monitor deflation blocks for a safepoint. 2030 return; 2031 } 2032 2033 if (n->header().value() == 0) { 2034 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 2035 "have non-null _header field.", p2i(n)); 2036 *error_cnt_p = *error_cnt_p + 1; 2037 } 2038 const oop obj = n->object_peek(); 2039 if (obj != nullptr) { 2040 const markWord mark = obj->mark(); 2041 if (!mark.has_monitor()) { 2042 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2043 "object does not think it has a monitor: obj=" 2044 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2045 p2i(obj), mark.value()); 2046 *error_cnt_p = *error_cnt_p + 1; 2047 } 2048 ObjectMonitor* const obj_mon = mark.monitor(); 2049 if (n != obj_mon) { 2050 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2051 "object does not refer to the same monitor: obj=" 2052 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2053 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2054 *error_cnt_p = *error_cnt_p + 1; 2055 } 2056 } 2057 } 2058 2059 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 2060 // flags indicate why the entry is in-use, 'object' and 'object type' 2061 // indicate the associated object and its type. 2062 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 2063 if (_in_use_list.count() > 0) { 2064 stringStream ss; 2065 out->print_cr("In-use monitor info:"); 2066 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2067 out->print_cr("%18s %s %18s %18s", 2068 "monitor", "BHL", "object", "object type"); 2069 out->print_cr("================== === ================== =================="); 2070 2071 auto is_interesting = [&](ObjectMonitor* monitor) { 2072 return log_all || monitor->has_owner() || monitor->is_busy(); 2073 }; 2074 2075 monitors_iterate([&](ObjectMonitor* monitor) { 2076 if (is_interesting(monitor)) { 2077 const oop obj = monitor->object_peek(); 2078 const markWord mark = monitor->header(); 2079 ResourceMark rm; 2080 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 2081 monitor->is_busy(), mark.hash() != 0, monitor->owner() != nullptr, 2082 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 2083 if (monitor->is_busy()) { 2084 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 2085 ss.reset(); 2086 } 2087 out->cr(); 2088 } 2089 }); 2090 } 2091 2092 out->flush(); 2093 }