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