1 /* 2 * Copyright (c) 1998, 2023, 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/suspendibleThreadSet.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/handles.inline.hpp" 40 #include "runtime/handshake.hpp" 41 #include "runtime/interfaceSupport.inline.hpp" 42 #include "runtime/javaThread.hpp" 43 #include "runtime/lockStack.inline.hpp" 44 #include "runtime/mutexLocker.hpp" 45 #include "runtime/objectMonitor.hpp" 46 #include "runtime/objectMonitor.inline.hpp" 47 #include "runtime/os.inline.hpp" 48 #include "runtime/osThread.hpp" 49 #include "runtime/perfData.hpp" 50 #include "runtime/safepointMechanism.inline.hpp" 51 #include "runtime/safepointVerifiers.hpp" 52 #include "runtime/sharedRuntime.hpp" 53 #include "runtime/stubRoutines.hpp" 54 #include "runtime/synchronizer.hpp" 55 #include "runtime/threads.hpp" 56 #include "runtime/timer.hpp" 57 #include "runtime/trimNativeHeap.hpp" 58 #include "runtime/vframe.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "utilities/align.hpp" 61 #include "utilities/dtrace.hpp" 62 #include "utilities/events.hpp" 63 #include "utilities/linkedlist.hpp" 64 #include "utilities/preserveException.hpp" 65 66 class ObjectMonitorsHashtable::PtrList : 67 public LinkedListImpl<ObjectMonitor*, 68 AnyObj::C_HEAP, mtThread, 69 AllocFailStrategy::RETURN_NULL> {}; 70 71 class CleanupObjectMonitorsHashtable: StackObj { 72 public: 73 bool do_entry(void*& key, ObjectMonitorsHashtable::PtrList*& list) { 74 list->clear(); // clear the LinkListNodes 75 delete list; // then delete the LinkedList 76 return true; 77 } 78 }; 79 80 ObjectMonitorsHashtable::~ObjectMonitorsHashtable() { 81 CleanupObjectMonitorsHashtable cleanup; 82 _ptrs->unlink(&cleanup); // cleanup the LinkedLists 83 delete _ptrs; // then delete the hash table 84 } 85 86 void ObjectMonitorsHashtable::add_entry(void* key, ObjectMonitor* om) { 87 ObjectMonitorsHashtable::PtrList* list = get_entry(key); 88 if (list == nullptr) { 89 // Create new list and add it to the hash table: 90 list = new (mtThread) ObjectMonitorsHashtable::PtrList; 91 add_entry(key, list); 92 } 93 list->add(om); // Add the ObjectMonitor to the list. 94 _om_count++; 95 } 96 97 bool ObjectMonitorsHashtable::has_entry(void* key, ObjectMonitor* om) { 98 ObjectMonitorsHashtable::PtrList* list = get_entry(key); 99 if (list == nullptr || list->find(om) == nullptr) { 100 return false; 101 } 102 return true; 103 } 104 105 void MonitorList::add(ObjectMonitor* m) { 106 ObjectMonitor* head; 107 do { 108 head = Atomic::load(&_head); 109 m->set_next_om(head); 110 } while (Atomic::cmpxchg(&_head, head, m) != head); 111 112 size_t count = Atomic::add(&_count, 1u); 113 if (count > max()) { 114 Atomic::inc(&_max); 115 } 116 } 117 118 size_t MonitorList::count() const { 119 return Atomic::load(&_count); 120 } 121 122 size_t MonitorList::max() const { 123 return Atomic::load(&_max); 124 } 125 126 // Walk the in-use list and unlink (at most MonitorDeflationMax) deflated 127 // ObjectMonitors. Returns the number of unlinked ObjectMonitors. 128 size_t MonitorList::unlink_deflated(Thread* current, LogStream* ls, 129 elapsedTimer* timer_p, 130 GrowableArray<ObjectMonitor*>* unlinked_list) { 131 size_t unlinked_count = 0; 132 ObjectMonitor* prev = nullptr; 133 ObjectMonitor* head = Atomic::load_acquire(&_head); 134 ObjectMonitor* m = head; 135 // The in-use list head can be null during the final audit. 136 while (m != nullptr) { 137 if (m->is_being_async_deflated()) { 138 // Find next live ObjectMonitor. 139 ObjectMonitor* next = m; 140 do { 141 ObjectMonitor* next_next = next->next_om(); 142 unlinked_count++; 143 unlinked_list->append(next); 144 next = next_next; 145 if (unlinked_count >= (size_t)MonitorDeflationMax) { 146 // Reached the max so bail out on the gathering loop. 147 break; 148 } 149 } while (next != nullptr && next->is_being_async_deflated()); 150 if (prev == nullptr) { 151 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, head, next); 152 if (prev_head != head) { 153 // Find new prev ObjectMonitor that just got inserted. 154 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 155 prev = n; 156 } 157 prev->set_next_om(next); 158 } 159 } else { 160 prev->set_next_om(next); 161 } 162 if (unlinked_count >= (size_t)MonitorDeflationMax) { 163 // Reached the max so bail out on the searching loop. 164 break; 165 } 166 m = next; 167 } else { 168 prev = m; 169 m = m->next_om(); 170 } 171 172 if (current->is_Java_thread()) { 173 // A JavaThread must check for a safepoint/handshake and honor it. 174 ObjectSynchronizer::chk_for_block_req(JavaThread::cast(current), "unlinking", 175 "unlinked_count", unlinked_count, 176 ls, timer_p); 177 } 178 } 179 Atomic::sub(&_count, unlinked_count); 180 return unlinked_count; 181 } 182 183 MonitorList::Iterator MonitorList::iterator() const { 184 return Iterator(Atomic::load_acquire(&_head)); 185 } 186 187 ObjectMonitor* MonitorList::Iterator::next() { 188 ObjectMonitor* current = _current; 189 _current = current->next_om(); 190 return current; 191 } 192 193 // The "core" versions of monitor enter and exit reside in this file. 194 // The interpreter and compilers contain specialized transliterated 195 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 196 // fast_lock(...) for instance. If you make changes here, make sure to modify the 197 // interpreter, and both C1 and C2 fast-path inline locking code emission. 198 // 199 // ----------------------------------------------------------------------------- 200 201 #ifdef DTRACE_ENABLED 202 203 // Only bother with this argument setup if dtrace is available 204 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 205 206 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 207 char* bytes = nullptr; \ 208 int len = 0; \ 209 jlong jtid = SharedRuntime::get_java_tid(thread); \ 210 Symbol* klassname = obj->klass()->name(); \ 211 if (klassname != nullptr) { \ 212 bytes = (char*)klassname->bytes(); \ 213 len = klassname->utf8_length(); \ 214 } 215 216 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 217 { \ 218 if (DTraceMonitorProbes) { \ 219 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 220 HOTSPOT_MONITOR_WAIT(jtid, \ 221 (uintptr_t)(monitor), bytes, len, (millis)); \ 222 } \ 223 } 224 225 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 226 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 227 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 228 229 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 230 { \ 231 if (DTraceMonitorProbes) { \ 232 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 233 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 234 (uintptr_t)(monitor), bytes, len); \ 235 } \ 236 } 237 238 #else // ndef DTRACE_ENABLED 239 240 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 241 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 242 243 #endif // ndef DTRACE_ENABLED 244 245 // This exists only as a workaround of dtrace bug 6254741 246 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 247 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 248 return 0; 249 } 250 251 static constexpr size_t inflation_lock_count() { 252 return 256; 253 } 254 255 // Static storage for an array of PlatformMutex. 256 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 257 258 static inline PlatformMutex* inflation_lock(size_t index) { 259 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 260 } 261 262 void ObjectSynchronizer::initialize() { 263 for (size_t i = 0; i < inflation_lock_count(); i++) { 264 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 265 } 266 // Start the ceiling with the estimate for one thread. 267 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 268 269 // Start the timer for deflations, so it does not trigger immediately. 270 _last_async_deflation_time_ns = os::javaTimeNanos(); 271 } 272 273 MonitorList ObjectSynchronizer::_in_use_list; 274 // monitors_used_above_threshold() policy is as follows: 275 // 276 // The ratio of the current _in_use_list count to the ceiling is used 277 // to determine if we are above MonitorUsedDeflationThreshold and need 278 // to do an async monitor deflation cycle. The ceiling is increased by 279 // AvgMonitorsPerThreadEstimate when a thread is added to the system 280 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 281 // removed from the system. 282 // 283 // Note: If the _in_use_list max exceeds the ceiling, then 284 // monitors_used_above_threshold() will use the in_use_list max instead 285 // of the thread count derived ceiling because we have used more 286 // ObjectMonitors than the estimated average. 287 // 288 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 289 // no-progress async monitor deflation cycles in a row, then the ceiling 290 // is adjusted upwards by monitors_used_above_threshold(). 291 // 292 // Start the ceiling with the estimate for one thread in initialize() 293 // which is called after cmd line options are processed. 294 static size_t _in_use_list_ceiling = 0; 295 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 296 bool volatile ObjectSynchronizer::_is_final_audit = false; 297 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 298 static uintx _no_progress_cnt = 0; 299 static bool _no_progress_skip_increment = false; 300 301 // =====================> Quick functions 302 303 // The quick_* forms are special fast-path variants used to improve 304 // performance. In the simplest case, a "quick_*" implementation could 305 // simply return false, in which case the caller will perform the necessary 306 // state transitions and call the slow-path form. 307 // The fast-path is designed to handle frequently arising cases in an efficient 308 // manner and is just a degenerate "optimistic" variant of the slow-path. 309 // returns true -- to indicate the call was satisfied. 310 // returns false -- to indicate the call needs the services of the slow-path. 311 // A no-loitering ordinance is in effect for code in the quick_* family 312 // operators: safepoints or indefinite blocking (blocking that might span a 313 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 314 // entry. 315 // 316 // Consider: An interesting optimization is to have the JIT recognize the 317 // following common idiom: 318 // synchronized (someobj) { .... ; notify(); } 319 // That is, we find a notify() or notifyAll() call that immediately precedes 320 // the monitorexit operation. In that case the JIT could fuse the operations 321 // into a single notifyAndExit() runtime primitive. 322 323 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 324 assert(current->thread_state() == _thread_in_Java, "invariant"); 325 NoSafepointVerifier nsv; 326 if (obj == nullptr) return false; // slow-path for invalid obj 327 const markWord mark = obj->mark(); 328 329 if (LockingMode == LM_LIGHTWEIGHT) { 330 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 331 // Degenerate notify 332 // fast-locked by caller so by definition the implied waitset is empty. 333 return true; 334 } 335 } else if (LockingMode == LM_LEGACY) { 336 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 337 // Degenerate notify 338 // stack-locked by caller so by definition the implied waitset is empty. 339 return true; 340 } 341 } 342 343 if (mark.has_monitor()) { 344 ObjectMonitor* const mon = mark.monitor(); 345 assert(mon->object() == oop(obj), "invariant"); 346 if (mon->owner() != current) return false; // slow-path for IMS exception 347 348 if (mon->first_waiter() != nullptr) { 349 // We have one or more waiters. Since this is an inflated monitor 350 // that we own, we can transfer one or more threads from the waitset 351 // to the entrylist here and now, avoiding the slow-path. 352 if (all) { 353 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current); 354 } else { 355 DTRACE_MONITOR_PROBE(notify, mon, obj, current); 356 } 357 int free_count = 0; 358 do { 359 mon->INotify(current); 360 ++free_count; 361 } while (mon->first_waiter() != nullptr && all); 362 OM_PERFDATA_OP(Notifications, inc(free_count)); 363 } 364 return true; 365 } 366 367 // other IMS exception states take the slow-path 368 return false; 369 } 370 371 372 // The LockNode emitted directly at the synchronization site would have 373 // been too big if it were to have included support for the cases of inflated 374 // recursive enter and exit, so they go here instead. 375 // Note that we can't safely call AsyncPrintJavaStack() from within 376 // quick_enter() as our thread state remains _in_Java. 377 378 bool ObjectSynchronizer::quick_enter(oop obj, JavaThread* current, 379 BasicLock * lock) { 380 assert(current->thread_state() == _thread_in_Java, "invariant"); 381 NoSafepointVerifier nsv; 382 if (obj == nullptr) return false; // Need to throw NPE 383 384 if (obj->klass()->is_value_based()) { 385 return false; 386 } 387 388 const markWord mark = obj->mark(); 389 390 if (mark.has_monitor()) { 391 ObjectMonitor* const m = mark.monitor(); 392 // An async deflation or GC can race us before we manage to make 393 // the ObjectMonitor busy by setting the owner below. If we detect 394 // that race we just bail out to the slow-path here. 395 if (m->object_peek() == nullptr) { 396 return false; 397 } 398 JavaThread* const owner = static_cast<JavaThread*>(m->owner_raw()); 399 400 // Lock contention and Transactional Lock Elision (TLE) diagnostics 401 // and observability 402 // Case: light contention possibly amenable to TLE 403 // Case: TLE inimical operations such as nested/recursive synchronization 404 405 if (owner == current) { 406 m->_recursions++; 407 current->inc_held_monitor_count(); 408 return true; 409 } 410 411 if (LockingMode != LM_LIGHTWEIGHT) { 412 // This Java Monitor is inflated so obj's header will never be 413 // displaced to this thread's BasicLock. Make the displaced header 414 // non-null so this BasicLock is not seen as recursive nor as 415 // being locked. We do this unconditionally so that this thread's 416 // BasicLock cannot be mis-interpreted by any stack walkers. For 417 // performance reasons, stack walkers generally first check for 418 // stack-locking in the object's header, the second check is for 419 // recursive stack-locking in the displaced header in the BasicLock, 420 // and last are the inflated Java Monitor (ObjectMonitor) checks. 421 lock->set_displaced_header(markWord::unused_mark()); 422 } 423 424 if (owner == nullptr && m->try_set_owner_from(nullptr, current) == nullptr) { 425 assert(m->_recursions == 0, "invariant"); 426 current->inc_held_monitor_count(); 427 return true; 428 } 429 } 430 431 // Note that we could inflate in quick_enter. 432 // This is likely a useful optimization 433 // Critically, in quick_enter() we must not: 434 // -- block indefinitely, or 435 // -- reach a safepoint 436 437 return false; // revert to slow-path 438 } 439 440 // Handle notifications when synchronizing on value based classes 441 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* current) { 442 frame last_frame = current->last_frame(); 443 bool bcp_was_adjusted = false; 444 // Don't decrement bcp if it points to the frame's first instruction. This happens when 445 // handle_sync_on_value_based_class() is called because of a synchronized method. There 446 // is no actual monitorenter instruction in the byte code in this case. 447 if (last_frame.is_interpreted_frame() && 448 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 449 // adjust bcp to point back to monitorenter so that we print the correct line numbers 450 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 451 bcp_was_adjusted = true; 452 } 453 454 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 455 ResourceMark rm(current); 456 stringStream ss; 457 current->print_active_stack_on(&ss); 458 char* base = (char*)strstr(ss.base(), "at"); 459 char* newline = (char*)strchr(ss.base(), '\n'); 460 if (newline != nullptr) { 461 *newline = '\0'; 462 } 463 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 464 } else { 465 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 466 ResourceMark rm(current); 467 Log(valuebasedclasses) vblog; 468 469 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 470 if (current->has_last_Java_frame()) { 471 LogStream info_stream(vblog.info()); 472 current->print_active_stack_on(&info_stream); 473 } else { 474 vblog.info("Cannot find the last Java frame"); 475 } 476 477 EventSyncOnValueBasedClass event; 478 if (event.should_commit()) { 479 event.set_valueBasedClass(obj->klass()); 480 event.commit(); 481 } 482 } 483 484 if (bcp_was_adjusted) { 485 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 486 } 487 } 488 489 static bool useHeavyMonitors() { 490 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390) 491 return LockingMode == LM_MONITOR; 492 #else 493 return false; 494 #endif 495 } 496 497 // ----------------------------------------------------------------------------- 498 // Monitor Enter/Exit 499 // The interpreter and compiler assembly code tries to lock using the fast path 500 // of this algorithm. Make sure to update that code if the following function is 501 // changed. The implementation is extremely sensitive to race condition. Be careful. 502 503 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, JavaThread* current) { 504 if (obj->klass()->is_value_based()) { 505 handle_sync_on_value_based_class(obj, current); 506 } 507 508 current->inc_held_monitor_count(); 509 510 if (!useHeavyMonitors()) { 511 if (LockingMode == LM_LIGHTWEIGHT) { 512 // Fast-locking does not use the 'lock' argument. 513 LockStack& lock_stack = current->lock_stack(); 514 if (lock_stack.can_push()) { 515 markWord mark = obj()->mark_acquire(); 516 if (mark.is_neutral()) { 517 assert(!lock_stack.contains(obj()), "thread must not already hold the lock"); 518 // Try to swing into 'fast-locked' state. 519 markWord locked_mark = mark.set_fast_locked(); 520 markWord old_mark = obj()->cas_set_mark(locked_mark, mark); 521 if (old_mark == mark) { 522 // Successfully fast-locked, push object to lock-stack and return. 523 lock_stack.push(obj()); 524 return; 525 } 526 } 527 } 528 // All other paths fall-through to inflate-enter. 529 } else if (LockingMode == LM_LEGACY) { 530 markWord mark = obj->mark(); 531 if (mark.is_neutral()) { 532 // Anticipate successful CAS -- the ST of the displaced mark must 533 // be visible <= the ST performed by the CAS. 534 lock->set_displaced_header(mark); 535 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 536 return; 537 } 538 // Fall through to inflate() ... 539 } else if (mark.has_locker() && 540 current->is_lock_owned((address) mark.locker())) { 541 assert(lock != mark.locker(), "must not re-lock the same lock"); 542 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock"); 543 lock->set_displaced_header(markWord::from_pointer(nullptr)); 544 return; 545 } 546 547 // The object header will never be displaced to this lock, 548 // so it does not matter what the value is, except that it 549 // must be non-zero to avoid looking like a re-entrant lock, 550 // and must not look locked either. 551 lock->set_displaced_header(markWord::unused_mark()); 552 } 553 } else if (VerifyHeavyMonitors) { 554 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 555 } 556 557 // An async deflation can race after the inflate() call and before 558 // enter() can make the ObjectMonitor busy. enter() returns false if 559 // we have lost the race to async deflation and we simply try again. 560 while (true) { 561 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter); 562 if (monitor->enter(current)) { 563 return; 564 } 565 } 566 } 567 568 void ObjectSynchronizer::exit(oop object, BasicLock* lock, JavaThread* current) { 569 current->dec_held_monitor_count(); 570 571 if (!useHeavyMonitors()) { 572 markWord mark = object->mark(); 573 if (LockingMode == LM_LIGHTWEIGHT) { 574 // Fast-locking does not use the 'lock' argument. 575 if (mark.is_fast_locked()) { 576 markWord unlocked_mark = mark.set_unlocked(); 577 markWord old_mark = object->cas_set_mark(unlocked_mark, mark); 578 if (old_mark != mark) { 579 // Another thread won the CAS, it must have inflated the monitor. 580 // It can only have installed an anonymously locked monitor at this point. 581 // Fetch that monitor, set owner correctly to this thread, and 582 // exit it (allowing waiting threads to enter). 583 assert(old_mark.has_monitor(), "must have monitor"); 584 ObjectMonitor* monitor = old_mark.monitor(); 585 assert(monitor->is_owner_anonymous(), "must be anonymous owner"); 586 monitor->set_owner_from_anonymous(current); 587 monitor->exit(current); 588 } 589 LockStack& lock_stack = current->lock_stack(); 590 lock_stack.remove(object); 591 return; 592 } 593 } else if (LockingMode == LM_LEGACY) { 594 markWord dhw = lock->displaced_header(); 595 if (dhw.value() == 0) { 596 // If the displaced header is null, then this exit matches up with 597 // a recursive enter. No real work to do here except for diagnostics. 598 #ifndef PRODUCT 599 if (mark != markWord::INFLATING()) { 600 // Only do diagnostics if we are not racing an inflation. Simply 601 // exiting a recursive enter of a Java Monitor that is being 602 // inflated is safe; see the has_monitor() comment below. 603 assert(!mark.is_neutral(), "invariant"); 604 assert(!mark.has_locker() || 605 current->is_lock_owned((address)mark.locker()), "invariant"); 606 if (mark.has_monitor()) { 607 // The BasicLock's displaced_header is marked as a recursive 608 // enter and we have an inflated Java Monitor (ObjectMonitor). 609 // This is a special case where the Java Monitor was inflated 610 // after this thread entered the stack-lock recursively. When a 611 // Java Monitor is inflated, we cannot safely walk the Java 612 // Monitor owner's stack and update the BasicLocks because a 613 // Java Monitor can be asynchronously inflated by a thread that 614 // does not own the Java Monitor. 615 ObjectMonitor* m = mark.monitor(); 616 assert(m->object()->mark() == mark, "invariant"); 617 assert(m->is_entered(current), "invariant"); 618 } 619 } 620 #endif 621 return; 622 } 623 624 if (mark == markWord::from_pointer(lock)) { 625 // If the object is stack-locked by the current thread, try to 626 // swing the displaced header from the BasicLock back to the mark. 627 assert(dhw.is_neutral(), "invariant"); 628 if (object->cas_set_mark(dhw, mark) == mark) { 629 return; 630 } 631 } 632 } 633 } else if (VerifyHeavyMonitors) { 634 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 635 } 636 637 // We have to take the slow-path of possible inflation and then exit. 638 // The ObjectMonitor* can't be async deflated until ownership is 639 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 640 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal); 641 if (LockingMode == LM_LIGHTWEIGHT && monitor->is_owner_anonymous()) { 642 // It must be owned by us. Pop lock object from lock stack. 643 LockStack& lock_stack = current->lock_stack(); 644 oop popped = lock_stack.pop(); 645 assert(popped == object, "must be owned by this thread"); 646 monitor->set_owner_from_anonymous(current); 647 } 648 monitor->exit(current); 649 } 650 651 // ----------------------------------------------------------------------------- 652 // JNI locks on java objects 653 // NOTE: must use heavy weight monitor to handle jni monitor enter 654 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 655 if (obj->klass()->is_value_based()) { 656 handle_sync_on_value_based_class(obj, current); 657 } 658 659 // the current locking is from JNI instead of Java code 660 current->set_current_pending_monitor_is_from_java(false); 661 // An async deflation can race after the inflate() call and before 662 // enter() can make the ObjectMonitor busy. enter() returns false if 663 // we have lost the race to async deflation and we simply try again. 664 while (true) { 665 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_jni_enter); 666 if (monitor->enter(current)) { 667 current->inc_held_monitor_count(1, true); 668 break; 669 } 670 } 671 current->set_current_pending_monitor_is_from_java(true); 672 } 673 674 // NOTE: must use heavy weight monitor to handle jni monitor exit 675 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 676 JavaThread* current = THREAD; 677 678 // The ObjectMonitor* can't be async deflated until ownership is 679 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 680 ObjectMonitor* monitor = inflate(current, obj, inflate_cause_jni_exit); 681 // If this thread has locked the object, exit the monitor. We 682 // intentionally do not use CHECK on check_owner because we must exit the 683 // monitor even if an exception was already pending. 684 if (monitor->check_owner(THREAD)) { 685 monitor->exit(current); 686 current->dec_held_monitor_count(1, true); 687 } 688 } 689 690 // ----------------------------------------------------------------------------- 691 // Internal VM locks on java objects 692 // standard constructor, allows locking failures 693 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) { 694 _thread = thread; 695 _thread->check_for_valid_safepoint_state(); 696 _obj = obj; 697 698 if (_obj() != nullptr) { 699 ObjectSynchronizer::enter(_obj, &_lock, _thread); 700 } 701 } 702 703 ObjectLocker::~ObjectLocker() { 704 if (_obj() != nullptr) { 705 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 706 } 707 } 708 709 710 // ----------------------------------------------------------------------------- 711 // Wait/Notify/NotifyAll 712 // NOTE: must use heavy weight monitor to handle wait() 713 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 714 JavaThread* current = THREAD; 715 if (millis < 0) { 716 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 717 } 718 // The ObjectMonitor* can't be async deflated because the _waiters 719 // field is incremented before ownership is dropped and decremented 720 // after ownership is regained. 721 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait); 722 723 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 724 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 725 726 // This dummy call is in place to get around dtrace bug 6254741. Once 727 // that's fixed we can uncomment the following line, remove the call 728 // and change this function back into a "void" func. 729 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 730 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 731 return ret_code; 732 } 733 734 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 735 JavaThread* current = THREAD; 736 737 markWord mark = obj->mark(); 738 if (LockingMode == LM_LIGHTWEIGHT) { 739 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 740 // Not inflated so there can't be any waiters to notify. 741 return; 742 } 743 } else if (LockingMode == LM_LEGACY) { 744 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 745 // Not inflated so there can't be any waiters to notify. 746 return; 747 } 748 } 749 // The ObjectMonitor* can't be async deflated until ownership is 750 // dropped by the calling thread. 751 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 752 monitor->notify(CHECK); 753 } 754 755 // NOTE: see comment of notify() 756 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 757 JavaThread* current = THREAD; 758 759 markWord mark = obj->mark(); 760 if (LockingMode == LM_LIGHTWEIGHT) { 761 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 762 // Not inflated so there can't be any waiters to notify. 763 return; 764 } 765 } else if (LockingMode == LM_LEGACY) { 766 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 767 // Not inflated so there can't be any waiters to notify. 768 return; 769 } 770 } 771 // The ObjectMonitor* can't be async deflated until ownership is 772 // dropped by the calling thread. 773 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 774 monitor->notifyAll(CHECK); 775 } 776 777 // ----------------------------------------------------------------------------- 778 // Hash Code handling 779 780 struct SharedGlobals { 781 char _pad_prefix[OM_CACHE_LINE_SIZE]; 782 // This is a highly shared mostly-read variable. 783 // To avoid false-sharing it needs to be the sole occupant of a cache line. 784 volatile int stw_random; 785 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 786 // Hot RW variable -- Sequester to avoid false-sharing 787 volatile int hc_sequence; 788 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 789 }; 790 791 static SharedGlobals GVars; 792 793 static markWord read_stable_mark(oop obj) { 794 markWord mark = obj->mark_acquire(); 795 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) { 796 // New lightweight locking does not use the markWord::INFLATING() protocol. 797 return mark; // normal fast-path return 798 } 799 800 int its = 0; 801 for (;;) { 802 markWord mark = obj->mark_acquire(); 803 if (!mark.is_being_inflated()) { 804 return mark; // normal fast-path return 805 } 806 807 // The object is being inflated by some other thread. 808 // The caller of read_stable_mark() must wait for inflation to complete. 809 // Avoid live-lock. 810 811 ++its; 812 if (its > 10000 || !os::is_MP()) { 813 if (its & 1) { 814 os::naked_yield(); 815 } else { 816 // Note that the following code attenuates the livelock problem but is not 817 // a complete remedy. A more complete solution would require that the inflating 818 // thread hold the associated inflation lock. The following code simply restricts 819 // the number of spinners to at most one. We'll have N-2 threads blocked 820 // on the inflationlock, 1 thread holding the inflation lock and using 821 // a yield/park strategy, and 1 thread in the midst of inflation. 822 // A more refined approach would be to change the encoding of INFLATING 823 // to allow encapsulation of a native thread pointer. Threads waiting for 824 // inflation to complete would use CAS to push themselves onto a singly linked 825 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 826 // and calling park(). When inflation was complete the thread that accomplished inflation 827 // would detach the list and set the markword to inflated with a single CAS and 828 // then for each thread on the list, set the flag and unpark() the thread. 829 830 // Index into the lock array based on the current object address. 831 static_assert(is_power_of_2(inflation_lock_count()), "must be"); 832 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1); 833 int YieldThenBlock = 0; 834 assert(ix < inflation_lock_count(), "invariant"); 835 inflation_lock(ix)->lock(); 836 while (obj->mark_acquire() == markWord::INFLATING()) { 837 // Beware: naked_yield() is advisory and has almost no effect on some platforms 838 // so we periodically call current->_ParkEvent->park(1). 839 // We use a mixed spin/yield/block mechanism. 840 if ((YieldThenBlock++) >= 16) { 841 Thread::current()->_ParkEvent->park(1); 842 } else { 843 os::naked_yield(); 844 } 845 } 846 inflation_lock(ix)->unlock(); 847 } 848 } else { 849 SpinPause(); // SMP-polite spinning 850 } 851 } 852 } 853 854 // hashCode() generation : 855 // 856 // Possibilities: 857 // * MD5Digest of {obj,stw_random} 858 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 859 // * A DES- or AES-style SBox[] mechanism 860 // * One of the Phi-based schemes, such as: 861 // 2654435761 = 2^32 * Phi (golden ratio) 862 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 863 // * A variation of Marsaglia's shift-xor RNG scheme. 864 // * (obj ^ stw_random) is appealing, but can result 865 // in undesirable regularity in the hashCode values of adjacent objects 866 // (objects allocated back-to-back, in particular). This could potentially 867 // result in hashtable collisions and reduced hashtable efficiency. 868 // There are simple ways to "diffuse" the middle address bits over the 869 // generated hashCode values: 870 871 static inline intptr_t get_next_hash(Thread* current, oop obj) { 872 intptr_t value = 0; 873 if (hashCode == 0) { 874 // This form uses global Park-Miller RNG. 875 // On MP system we'll have lots of RW access to a global, so the 876 // mechanism induces lots of coherency traffic. 877 value = os::random(); 878 } else if (hashCode == 1) { 879 // This variation has the property of being stable (idempotent) 880 // between STW operations. This can be useful in some of the 1-0 881 // synchronization schemes. 882 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 883 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 884 } else if (hashCode == 2) { 885 value = 1; // for sensitivity testing 886 } else if (hashCode == 3) { 887 value = ++GVars.hc_sequence; 888 } else if (hashCode == 4) { 889 value = cast_from_oop<intptr_t>(obj); 890 } else { 891 // Marsaglia's xor-shift scheme with thread-specific state 892 // This is probably the best overall implementation -- we'll 893 // likely make this the default in future releases. 894 unsigned t = current->_hashStateX; 895 t ^= (t << 11); 896 current->_hashStateX = current->_hashStateY; 897 current->_hashStateY = current->_hashStateZ; 898 current->_hashStateZ = current->_hashStateW; 899 unsigned v = current->_hashStateW; 900 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 901 current->_hashStateW = v; 902 value = v; 903 } 904 905 value &= markWord::hash_mask; 906 if (value == 0) value = 0xBAD; 907 assert(value != markWord::no_hash, "invariant"); 908 return value; 909 } 910 911 // Can be called from non JavaThreads (e.g., VMThread) for FastHashCode 912 // calculations as part of JVM/TI tagging. 913 static bool is_lock_owned(Thread* thread, oop obj) { 914 assert(LockingMode == LM_LIGHTWEIGHT, "only call this with new lightweight locking enabled"); 915 return thread->is_Java_thread() ? JavaThread::cast(thread)->lock_stack().contains(obj) : false; 916 } 917 918 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 919 920 while (true) { 921 ObjectMonitor* monitor = nullptr; 922 markWord temp, test; 923 intptr_t hash; 924 markWord mark = read_stable_mark(obj); 925 if (VerifyHeavyMonitors) { 926 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)"); 927 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 928 } 929 if (mark.is_neutral()) { // if this is a normal header 930 hash = mark.hash(); 931 if (hash != 0) { // if it has a hash, just return it 932 return hash; 933 } 934 hash = get_next_hash(current, obj); // get a new hash 935 temp = mark.copy_set_hash(hash); // merge the hash into header 936 // try to install the hash 937 test = obj->cas_set_mark(temp, mark); 938 if (test == mark) { // if the hash was installed, return it 939 return hash; 940 } 941 // Failed to install the hash. It could be that another thread 942 // installed the hash just before our attempt or inflation has 943 // occurred or... so we fall thru to inflate the monitor for 944 // stability and then install the hash. 945 } else if (mark.has_monitor()) { 946 monitor = mark.monitor(); 947 temp = monitor->header(); 948 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 949 hash = temp.hash(); 950 if (hash != 0) { 951 // It has a hash. 952 953 // Separate load of dmw/header above from the loads in 954 // is_being_async_deflated(). 955 956 // dmw/header and _contentions may get written by different threads. 957 // Make sure to observe them in the same order when having several observers. 958 OrderAccess::loadload_for_IRIW(); 959 960 if (monitor->is_being_async_deflated()) { 961 // But we can't safely use the hash if we detect that async 962 // deflation has occurred. So we attempt to restore the 963 // header/dmw to the object's header so that we only retry 964 // once if the deflater thread happens to be slow. 965 monitor->install_displaced_markword_in_object(obj); 966 continue; 967 } 968 return hash; 969 } 970 // Fall thru so we only have one place that installs the hash in 971 // the ObjectMonitor. 972 } else if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked() && is_lock_owned(current, obj)) { 973 // This is a fast-lock owned by the calling thread so use the 974 // markWord from the object. 975 hash = mark.hash(); 976 if (hash != 0) { // if it has a hash, just return it 977 return hash; 978 } 979 } else if (LockingMode == LM_LEGACY && mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 980 // This is a stack-lock owned by the calling thread so fetch the 981 // displaced markWord from the BasicLock on the stack. 982 temp = mark.displaced_mark_helper(); 983 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 984 hash = temp.hash(); 985 if (hash != 0) { // if it has a hash, just return it 986 return hash; 987 } 988 // WARNING: 989 // The displaced header in the BasicLock on a thread's stack 990 // is strictly immutable. It CANNOT be changed in ANY cases. 991 // So we have to inflate the stack-lock into an ObjectMonitor 992 // even if the current thread owns the lock. The BasicLock on 993 // a thread's stack can be asynchronously read by other threads 994 // during an inflate() call so any change to that stack memory 995 // may not propagate to other threads correctly. 996 } 997 998 // Inflate the monitor to set the hash. 999 1000 // An async deflation can race after the inflate() call and before we 1001 // can update the ObjectMonitor's header with the hash value below. 1002 monitor = inflate(current, obj, inflate_cause_hash_code); 1003 // Load ObjectMonitor's header/dmw field and see if it has a hash. 1004 mark = monitor->header(); 1005 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1006 hash = mark.hash(); 1007 if (hash == 0) { // if it does not have a hash 1008 hash = get_next_hash(current, obj); // get a new hash 1009 temp = mark.copy_set_hash(hash) ; // merge the hash into header 1010 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1011 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value()); 1012 test = markWord(v); 1013 if (test != mark) { 1014 // The attempt to update the ObjectMonitor's header/dmw field 1015 // did not work. This can happen if another thread managed to 1016 // merge in the hash just before our cmpxchg(). 1017 // If we add any new usages of the header/dmw field, this code 1018 // will need to be updated. 1019 hash = test.hash(); 1020 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1021 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1022 } 1023 if (monitor->is_being_async_deflated()) { 1024 // If we detect that async deflation has occurred, then we 1025 // attempt to restore the header/dmw to the object's header 1026 // so that we only retry once if the deflater thread happens 1027 // to be slow. 1028 monitor->install_displaced_markword_in_object(obj); 1029 continue; 1030 } 1031 } 1032 // We finally get the hash. 1033 return hash; 1034 } 1035 } 1036 1037 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 1038 Handle h_obj) { 1039 assert(current == JavaThread::current(), "Can only be called on current thread"); 1040 oop obj = h_obj(); 1041 1042 markWord mark = read_stable_mark(obj); 1043 1044 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1045 // stack-locked case, header points into owner's stack 1046 return current->is_lock_owned((address)mark.locker()); 1047 } 1048 1049 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1050 // fast-locking case, see if lock is in current's lock stack 1051 return current->lock_stack().contains(h_obj()); 1052 } 1053 1054 if (mark.has_monitor()) { 1055 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1056 // The first stage of async deflation does not affect any field 1057 // used by this comparison so the ObjectMonitor* is usable here. 1058 ObjectMonitor* monitor = mark.monitor(); 1059 return monitor->is_entered(current) != 0; 1060 } 1061 // Unlocked case, header in place 1062 assert(mark.is_neutral(), "sanity check"); 1063 return false; 1064 } 1065 1066 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1067 oop obj = h_obj(); 1068 markWord mark = read_stable_mark(obj); 1069 1070 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1071 // stack-locked so header points into owner's stack. 1072 // owning_thread_from_monitor_owner() may also return null here: 1073 return Threads::owning_thread_from_monitor_owner(t_list, (address) mark.locker()); 1074 } 1075 1076 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1077 // fast-locked so get owner from the object. 1078 // owning_thread_from_object() may also return null here: 1079 return Threads::owning_thread_from_object(t_list, h_obj()); 1080 } 1081 1082 if (mark.has_monitor()) { 1083 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1084 // The first stage of async deflation does not affect any field 1085 // used by this comparison so the ObjectMonitor* is usable here. 1086 ObjectMonitor* monitor = mark.monitor(); 1087 assert(monitor != nullptr, "monitor should be non-null"); 1088 // owning_thread_from_monitor() may also return null here: 1089 return Threads::owning_thread_from_monitor(t_list, monitor); 1090 } 1091 1092 // Unlocked case, header in place 1093 // Cannot have assertion since this object may have been 1094 // locked by another thread when reaching here. 1095 // assert(mark.is_neutral(), "sanity check"); 1096 1097 return nullptr; 1098 } 1099 1100 // Visitors ... 1101 1102 // Iterate ObjectMonitors where the owner == thread; this does NOT include 1103 // ObjectMonitors where owner is set to a stack-lock address in thread. 1104 // 1105 // This version of monitors_iterate() works with the in-use monitor list. 1106 // 1107 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 1108 MonitorList::Iterator iter = _in_use_list.iterator(); 1109 while (iter.has_next()) { 1110 ObjectMonitor* mid = iter.next(); 1111 if (mid->owner() != thread) { 1112 // Not owned by the target thread and intentionally skips when owner 1113 // is set to a stack-lock address in the target thread. 1114 continue; 1115 } 1116 if (!mid->is_being_async_deflated() && mid->object_peek() != nullptr) { 1117 // Only process with closure if the object is set. 1118 1119 // monitors_iterate() is only called at a safepoint or when the 1120 // target thread is suspended or when the target thread is 1121 // operating on itself. The current closures in use today are 1122 // only interested in an owned ObjectMonitor and ownership 1123 // cannot be dropped under the calling contexts so the 1124 // ObjectMonitor cannot be async deflated. 1125 closure->do_monitor(mid); 1126 } 1127 } 1128 } 1129 1130 // This version of monitors_iterate() works with the specified linked list. 1131 // 1132 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure, 1133 ObjectMonitorsHashtable::PtrList* list, 1134 JavaThread* thread) { 1135 typedef LinkedListIterator<ObjectMonitor*> ObjectMonitorIterator; 1136 ObjectMonitorIterator iter(list->head()); 1137 while (!iter.is_empty()) { 1138 ObjectMonitor* mid = *iter.next(); 1139 // Owner set to a stack-lock address in thread should never be seen here: 1140 assert(mid->owner() == thread, "must be"); 1141 if (!mid->is_being_async_deflated() && mid->object_peek() != nullptr) { 1142 // Only process with closure if the object is set. 1143 1144 // monitors_iterate() is only called at a safepoint or when the 1145 // target thread is suspended or when the target thread is 1146 // operating on itself. The current closures in use today are 1147 // only interested in an owned ObjectMonitor and ownership 1148 // cannot be dropped under the calling contexts so the 1149 // ObjectMonitor cannot be async deflated. 1150 closure->do_monitor(mid); 1151 } 1152 } 1153 } 1154 1155 static bool monitors_used_above_threshold(MonitorList* list) { 1156 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 1157 return false; 1158 } 1159 // Start with ceiling based on a per-thread estimate: 1160 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling(); 1161 size_t old_ceiling = ceiling; 1162 if (ceiling < list->max()) { 1163 // The max used by the system has exceeded the ceiling so use that: 1164 ceiling = list->max(); 1165 } 1166 size_t monitors_used = list->count(); 1167 if (monitors_used == 0) { // empty list is easy 1168 return false; 1169 } 1170 if (NoAsyncDeflationProgressMax != 0 && 1171 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 1172 float remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 1173 size_t new_ceiling = ceiling + (ceiling * remainder) + 1; 1174 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 1175 log_info(monitorinflation)("Too many deflations without progress; " 1176 "bumping in_use_list_ceiling from " SIZE_FORMAT 1177 " to " SIZE_FORMAT, old_ceiling, new_ceiling); 1178 _no_progress_cnt = 0; 1179 ceiling = new_ceiling; 1180 } 1181 1182 // Check if our monitor usage is above the threshold: 1183 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 1184 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 1185 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT 1186 ", monitor_usage=" SIZE_FORMAT ", threshold=" INTX_FORMAT, 1187 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 1188 return true; 1189 } 1190 1191 return false; 1192 } 1193 1194 size_t ObjectSynchronizer::in_use_list_ceiling() { 1195 return _in_use_list_ceiling; 1196 } 1197 1198 void ObjectSynchronizer::dec_in_use_list_ceiling() { 1199 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1200 } 1201 1202 void ObjectSynchronizer::inc_in_use_list_ceiling() { 1203 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1204 } 1205 1206 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 1207 _in_use_list_ceiling = new_value; 1208 } 1209 1210 bool ObjectSynchronizer::is_async_deflation_needed() { 1211 if (is_async_deflation_requested()) { 1212 // Async deflation request. 1213 log_info(monitorinflation)("Async deflation needed: explicit request"); 1214 return true; 1215 } 1216 1217 jlong time_since_last = time_since_last_async_deflation_ms(); 1218 1219 if (AsyncDeflationInterval > 0 && 1220 time_since_last > AsyncDeflationInterval && 1221 monitors_used_above_threshold(&_in_use_list)) { 1222 // It's been longer than our specified deflate interval and there 1223 // are too many monitors in use. We don't deflate more frequently 1224 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1225 // in order to not swamp the MonitorDeflationThread. 1226 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 1227 return true; 1228 } 1229 1230 if (GuaranteedAsyncDeflationInterval > 0 && 1231 time_since_last > GuaranteedAsyncDeflationInterval) { 1232 // It's been longer than our specified guaranteed deflate interval. 1233 // We need to clean up the used monitors even if the threshold is 1234 // not reached, to keep the memory utilization at bay when many threads 1235 // touched many monitors. 1236 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) " 1237 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 1238 GuaranteedAsyncDeflationInterval, time_since_last); 1239 1240 // If this deflation has no progress, then it should not affect the no-progress 1241 // tracking, otherwise threshold heuristics would think it was triggered, experienced 1242 // no progress, and needs to backoff more aggressively. In this "no progress" case, 1243 // the generic code would bump the no-progress counter, and we compensate for that 1244 // by telling it to skip the update. 1245 // 1246 // If this deflation has progress, then it should let non-progress tracking 1247 // know about this, otherwise the threshold heuristics would kick in, potentially 1248 // experience no-progress due to aggressive cleanup by this deflation, and think 1249 // it is still in no-progress stride. In this "progress" case, the generic code would 1250 // zero the counter, and we allow it to happen. 1251 _no_progress_skip_increment = true; 1252 1253 return true; 1254 } 1255 1256 return false; 1257 } 1258 1259 bool ObjectSynchronizer::request_deflate_idle_monitors() { 1260 JavaThread* current = JavaThread::current(); 1261 bool ret_code = false; 1262 1263 jlong last_time = last_async_deflation_time_ns(); 1264 set_is_async_deflation_requested(true); 1265 { 1266 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1267 ml.notify_all(); 1268 } 1269 const int N_CHECKS = 5; 1270 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1271 if (last_async_deflation_time_ns() > last_time) { 1272 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1273 ret_code = true; 1274 break; 1275 } 1276 { 1277 // JavaThread has to honor the blocking protocol. 1278 ThreadBlockInVM tbivm(current); 1279 os::naked_short_sleep(999); // sleep for almost 1 second 1280 } 1281 } 1282 if (!ret_code) { 1283 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1284 } 1285 1286 return ret_code; 1287 } 1288 1289 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1290 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1291 } 1292 1293 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1294 const oop obj, 1295 ObjectSynchronizer::InflateCause cause) { 1296 assert(event != nullptr, "invariant"); 1297 event->set_monitorClass(obj->klass()); 1298 event->set_address((uintptr_t)(void*)obj); 1299 event->set_cause((u1)cause); 1300 event->commit(); 1301 } 1302 1303 // Fast path code shared by multiple functions 1304 void ObjectSynchronizer::inflate_helper(oop obj) { 1305 markWord mark = obj->mark_acquire(); 1306 if (mark.has_monitor()) { 1307 ObjectMonitor* monitor = mark.monitor(); 1308 markWord dmw = monitor->header(); 1309 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1310 return; 1311 } 1312 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal); 1313 } 1314 1315 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop object, 1316 const InflateCause cause) { 1317 EventJavaMonitorInflate event; 1318 1319 for (;;) { 1320 const markWord mark = object->mark_acquire(); 1321 1322 // The mark can be in one of the following states: 1323 // * inflated - Just return if using stack-locking. 1324 // If using fast-locking and the ObjectMonitor owner 1325 // is anonymous and the current thread owns the 1326 // object lock, then we make the current thread the 1327 // ObjectMonitor owner and remove the lock from the 1328 // current thread's lock stack. 1329 // * fast-locked - Coerce it to inflated from fast-locked. 1330 // * stack-locked - Coerce it to inflated from stack-locked. 1331 // * INFLATING - Busy wait for conversion from stack-locked to 1332 // inflated. 1333 // * neutral - Aggressively inflate the object. 1334 1335 // CASE: inflated 1336 if (mark.has_monitor()) { 1337 ObjectMonitor* inf = mark.monitor(); 1338 markWord dmw = inf->header(); 1339 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1340 if (LockingMode == LM_LIGHTWEIGHT && inf->is_owner_anonymous() && is_lock_owned(current, object)) { 1341 inf->set_owner_from_anonymous(current); 1342 JavaThread::cast(current)->lock_stack().remove(object); 1343 } 1344 return inf; 1345 } 1346 1347 if (LockingMode != LM_LIGHTWEIGHT) { 1348 // New lightweight locking does not use INFLATING. 1349 // CASE: inflation in progress - inflating over a stack-lock. 1350 // Some other thread is converting from stack-locked to inflated. 1351 // Only that thread can complete inflation -- other threads must wait. 1352 // The INFLATING value is transient. 1353 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1354 // We could always eliminate polling by parking the thread on some auxiliary list. 1355 if (mark == markWord::INFLATING()) { 1356 read_stable_mark(object); 1357 continue; 1358 } 1359 } 1360 1361 // CASE: fast-locked 1362 // Could be fast-locked either by current or by some other thread. 1363 // 1364 // Note that we allocate the ObjectMonitor speculatively, _before_ 1365 // attempting to set the object's mark to the new ObjectMonitor. If 1366 // this thread owns the monitor, then we set the ObjectMonitor's 1367 // owner to this thread. Otherwise, we set the ObjectMonitor's owner 1368 // to anonymous. If we lose the race to set the object's mark to the 1369 // new ObjectMonitor, then we just delete it and loop around again. 1370 // 1371 LogStreamHandle(Trace, monitorinflation) lsh; 1372 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1373 ObjectMonitor* monitor = new ObjectMonitor(object); 1374 monitor->set_header(mark.set_unlocked()); 1375 bool own = is_lock_owned(current, object); 1376 if (own) { 1377 // Owned by us. 1378 monitor->set_owner_from(nullptr, current); 1379 } else { 1380 // Owned by somebody else. 1381 monitor->set_owner_anonymous(); 1382 } 1383 markWord monitor_mark = markWord::encode(monitor); 1384 markWord old_mark = object->cas_set_mark(monitor_mark, mark); 1385 if (old_mark == mark) { 1386 // Success! Return inflated monitor. 1387 if (own) { 1388 JavaThread::cast(current)->lock_stack().remove(object); 1389 } 1390 // Once the ObjectMonitor is configured and object is associated 1391 // with the ObjectMonitor, it is safe to allow async deflation: 1392 _in_use_list.add(monitor); 1393 1394 // Hopefully the performance counters are allocated on distinct 1395 // cache lines to avoid false sharing on MP systems ... 1396 OM_PERFDATA_OP(Inflations, inc()); 1397 if (log_is_enabled(Trace, monitorinflation)) { 1398 ResourceMark rm(current); 1399 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1400 INTPTR_FORMAT ", type='%s'", p2i(object), 1401 object->mark().value(), object->klass()->external_name()); 1402 } 1403 if (event.should_commit()) { 1404 post_monitor_inflate_event(&event, object, cause); 1405 } 1406 return monitor; 1407 } else { 1408 delete monitor; 1409 continue; // Interference -- just retry 1410 } 1411 } 1412 1413 // CASE: stack-locked 1414 // Could be stack-locked either by current or by some other thread. 1415 // 1416 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting 1417 // to install INFLATING into the mark word. We originally installed INFLATING, 1418 // allocated the ObjectMonitor, and then finally STed the address of the 1419 // ObjectMonitor into the mark. This was correct, but artificially lengthened 1420 // the interval in which INFLATING appeared in the mark, thus increasing 1421 // the odds of inflation contention. If we lose the race to set INFLATING, 1422 // then we just delete the ObjectMonitor and loop around again. 1423 // 1424 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1425 assert(LockingMode != LM_LIGHTWEIGHT, "cannot happen with new lightweight locking"); 1426 ObjectMonitor* m = new ObjectMonitor(object); 1427 // Optimistically prepare the ObjectMonitor - anticipate successful CAS 1428 // We do this before the CAS in order to minimize the length of time 1429 // in which INFLATING appears in the mark. 1430 1431 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1432 if (cmp != mark) { 1433 delete m; 1434 continue; // Interference -- just retry 1435 } 1436 1437 // We've successfully installed INFLATING (0) into the mark-word. 1438 // This is the only case where 0 will appear in a mark-word. 1439 // Only the singular thread that successfully swings the mark-word 1440 // to 0 can perform (or more precisely, complete) inflation. 1441 // 1442 // Why do we CAS a 0 into the mark-word instead of just CASing the 1443 // mark-word from the stack-locked value directly to the new inflated state? 1444 // Consider what happens when a thread unlocks a stack-locked object. 1445 // It attempts to use CAS to swing the displaced header value from the 1446 // on-stack BasicLock back into the object header. Recall also that the 1447 // header value (hash code, etc) can reside in (a) the object header, or 1448 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1449 // header in an ObjectMonitor. The inflate() routine must copy the header 1450 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1451 // the while preserving the hashCode stability invariants. If the owner 1452 // decides to release the lock while the value is 0, the unlock will fail 1453 // and control will eventually pass from slow_exit() to inflate. The owner 1454 // will then spin, waiting for the 0 value to disappear. Put another way, 1455 // the 0 causes the owner to stall if the owner happens to try to 1456 // drop the lock (restoring the header from the BasicLock to the object) 1457 // while inflation is in-progress. This protocol avoids races that might 1458 // would otherwise permit hashCode values to change or "flicker" for an object. 1459 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1460 // 0 serves as a "BUSY" inflate-in-progress indicator. 1461 1462 1463 // fetch the displaced mark from the owner's stack. 1464 // The owner can't die or unwind past the lock while our INFLATING 1465 // object is in the mark. Furthermore the owner can't complete 1466 // an unlock on the object, either. 1467 markWord dmw = mark.displaced_mark_helper(); 1468 // Catch if the object's header is not neutral (not locked and 1469 // not marked is what we care about here). 1470 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1471 1472 // Setup monitor fields to proper values -- prepare the monitor 1473 m->set_header(dmw); 1474 1475 // Optimization: if the mark.locker stack address is associated 1476 // with this thread we could simply set m->_owner = current. 1477 // Note that a thread can inflate an object 1478 // that it has stack-locked -- as might happen in wait() -- directly 1479 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom. 1480 m->set_owner_from(nullptr, mark.locker()); 1481 // TODO-FIXME: assert BasicLock->dhw != 0. 1482 1483 // Must preserve store ordering. The monitor state must 1484 // be stable at the time of publishing the monitor address. 1485 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1486 // Release semantics so that above set_object() is seen first. 1487 object->release_set_mark(markWord::encode(m)); 1488 1489 // Once ObjectMonitor is configured and the object is associated 1490 // with the ObjectMonitor, it is safe to allow async deflation: 1491 _in_use_list.add(m); 1492 1493 // Hopefully the performance counters are allocated on distinct cache lines 1494 // to avoid false sharing on MP systems ... 1495 OM_PERFDATA_OP(Inflations, inc()); 1496 if (log_is_enabled(Trace, monitorinflation)) { 1497 ResourceMark rm(current); 1498 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1499 INTPTR_FORMAT ", type='%s'", p2i(object), 1500 object->mark().value(), object->klass()->external_name()); 1501 } 1502 if (event.should_commit()) { 1503 post_monitor_inflate_event(&event, object, cause); 1504 } 1505 return m; 1506 } 1507 1508 // CASE: neutral 1509 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1510 // If we know we're inflating for entry it's better to inflate by swinging a 1511 // pre-locked ObjectMonitor pointer into the object header. A successful 1512 // CAS inflates the object *and* confers ownership to the inflating thread. 1513 // In the current implementation we use a 2-step mechanism where we CAS() 1514 // to inflate and then CAS() again to try to swing _owner from null to current. 1515 // An inflateTry() method that we could call from enter() would be useful. 1516 1517 // Catch if the object's header is not neutral (not locked and 1518 // not marked is what we care about here). 1519 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1520 ObjectMonitor* m = new ObjectMonitor(object); 1521 // prepare m for installation - set monitor to initial state 1522 m->set_header(mark); 1523 1524 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1525 delete m; 1526 m = nullptr; 1527 continue; 1528 // interference - the markword changed - just retry. 1529 // The state-transitions are one-way, so there's no chance of 1530 // live-lock -- "Inflated" is an absorbing state. 1531 } 1532 1533 // Once the ObjectMonitor is configured and object is associated 1534 // with the ObjectMonitor, it is safe to allow async deflation: 1535 _in_use_list.add(m); 1536 1537 // Hopefully the performance counters are allocated on distinct 1538 // cache lines to avoid false sharing on MP systems ... 1539 OM_PERFDATA_OP(Inflations, inc()); 1540 if (log_is_enabled(Trace, monitorinflation)) { 1541 ResourceMark rm(current); 1542 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 1543 INTPTR_FORMAT ", type='%s'", p2i(object), 1544 object->mark().value(), object->klass()->external_name()); 1545 } 1546 if (event.should_commit()) { 1547 post_monitor_inflate_event(&event, object, cause); 1548 } 1549 return m; 1550 } 1551 } 1552 1553 void ObjectSynchronizer::chk_for_block_req(JavaThread* current, const char* op_name, 1554 const char* cnt_name, size_t cnt, 1555 LogStream* ls, elapsedTimer* timer_p) { 1556 if (!SafepointMechanism::should_process(current)) { 1557 return; 1558 } 1559 1560 // A safepoint/handshake has started. 1561 if (ls != nullptr) { 1562 timer_p->stop(); 1563 ls->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling=" 1564 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1565 op_name, cnt_name, cnt, in_use_list_ceiling(), 1566 _in_use_list.count(), _in_use_list.max()); 1567 } 1568 1569 { 1570 // Honor block request. 1571 ThreadBlockInVM tbivm(current); 1572 } 1573 1574 if (ls != nullptr) { 1575 ls->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT 1576 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name, 1577 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max()); 1578 timer_p->start(); 1579 } 1580 } 1581 1582 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1583 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1584 // 1585 // If table != nullptr, we gather owned ObjectMonitors indexed by the 1586 // owner in the table. Please note that ObjectMonitors where the owner 1587 // is set to a stack-lock address are NOT associated with the JavaThread 1588 // that holds that stack-lock. All of the current consumers of 1589 // ObjectMonitorsHashtable info only care about JNI locked monitors and 1590 // those do not have the owner set to a stack-lock address. 1591 // 1592 size_t ObjectSynchronizer::deflate_monitor_list(Thread* current, LogStream* ls, 1593 elapsedTimer* timer_p, 1594 ObjectMonitorsHashtable* table) { 1595 MonitorList::Iterator iter = _in_use_list.iterator(); 1596 size_t deflated_count = 0; 1597 1598 while (iter.has_next()) { 1599 if (deflated_count >= (size_t)MonitorDeflationMax) { 1600 break; 1601 } 1602 ObjectMonitor* mid = iter.next(); 1603 if (mid->deflate_monitor()) { 1604 deflated_count++; 1605 } else if (table != nullptr) { 1606 // The caller is interested in the owned ObjectMonitors. This does 1607 // not include when owner is set to a stack-lock address in thread. 1608 // This also does not capture unowned ObjectMonitors that cannot be 1609 // deflated because of a waiter. 1610 void* key = mid->owner(); 1611 // Since deflate_idle_monitors() and deflate_monitor_list() can be 1612 // called more than once, we have to make sure the entry has not 1613 // already been added. 1614 if (key != nullptr && !table->has_entry(key, mid)) { 1615 table->add_entry(key, mid); 1616 } 1617 } 1618 1619 if (current->is_Java_thread()) { 1620 // A JavaThread must check for a safepoint/handshake and honor it. 1621 chk_for_block_req(JavaThread::cast(current), "deflation", "deflated_count", 1622 deflated_count, ls, timer_p); 1623 } 1624 } 1625 1626 return deflated_count; 1627 } 1628 1629 class HandshakeForDeflation : public HandshakeClosure { 1630 public: 1631 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {} 1632 1633 void do_thread(Thread* thread) { 1634 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread=" 1635 INTPTR_FORMAT, p2i(thread)); 1636 } 1637 }; 1638 1639 class VM_RendezvousGCThreads : public VM_Operation { 1640 public: 1641 bool evaluate_at_safepoint() const override { return false; } 1642 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1643 void doit() override { 1644 SuspendibleThreadSet::synchronize(); 1645 SuspendibleThreadSet::desynchronize(); 1646 }; 1647 }; 1648 1649 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list) { 1650 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1651 size_t count = 0; 1652 for (ObjectMonitor* monitor: *delete_list) { 1653 delete monitor; 1654 count++; 1655 } 1656 return count; 1657 } 1658 1659 // This function is called by the MonitorDeflationThread to deflate 1660 // ObjectMonitors. It is also called via do_final_audit_and_print_stats() 1661 // and VM_ThreadDump::doit() by the VMThread. 1662 size_t ObjectSynchronizer::deflate_idle_monitors(ObjectMonitorsHashtable* table) { 1663 Thread* current = Thread::current(); 1664 if (current->is_Java_thread()) { 1665 // The async deflation request has been processed. 1666 _last_async_deflation_time_ns = os::javaTimeNanos(); 1667 set_is_async_deflation_requested(false); 1668 } 1669 1670 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1671 LogStreamHandle(Info, monitorinflation) lsh_info; 1672 LogStream* ls = nullptr; 1673 if (log_is_enabled(Debug, monitorinflation)) { 1674 ls = &lsh_debug; 1675 } else if (log_is_enabled(Info, monitorinflation)) { 1676 ls = &lsh_info; 1677 } 1678 1679 elapsedTimer timer; 1680 if (ls != nullptr) { 1681 ls->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1682 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max()); 1683 timer.start(); 1684 } 1685 1686 // Deflate some idle ObjectMonitors. 1687 size_t deflated_count = deflate_monitor_list(current, ls, &timer, table); 1688 size_t unlinked_count = 0; 1689 size_t deleted_count = 0; 1690 if (deflated_count > 0 || is_final_audit()) { 1691 // There are ObjectMonitors that have been deflated or this is the 1692 // final audit and all the remaining ObjectMonitors have been 1693 // deflated, BUT the MonitorDeflationThread blocked for the final 1694 // safepoint during unlinking. 1695 1696 // Unlink deflated ObjectMonitors from the in-use list. 1697 ResourceMark rm; 1698 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1699 unlinked_count = _in_use_list.unlink_deflated(current, ls, &timer, &delete_list); 1700 if (current->is_monitor_deflation_thread()) { 1701 if (ls != nullptr) { 1702 timer.stop(); 1703 ls->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT 1704 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count=" 1705 SIZE_FORMAT ", max=" SIZE_FORMAT, 1706 unlinked_count, in_use_list_ceiling(), 1707 _in_use_list.count(), _in_use_list.max()); 1708 } 1709 1710 // A JavaThread needs to handshake in order to safely free the 1711 // ObjectMonitors that were deflated in this cycle. 1712 HandshakeForDeflation hfd_hc; 1713 Handshake::execute(&hfd_hc); 1714 // Also, we sync and desync GC threads around the handshake, so that they can 1715 // safely read the mark-word and look-through to the object-monitor, without 1716 // being afraid that the object-monitor is going away. 1717 VM_RendezvousGCThreads sync_gc; 1718 VMThread::execute(&sync_gc); 1719 1720 if (ls != nullptr) { 1721 ls->print_cr("after handshaking: in_use_list stats: ceiling=" 1722 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1723 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max()); 1724 timer.start(); 1725 } 1726 } else { 1727 // This is not a monitor deflation thread. 1728 // No handshake or rendezvous is needed when we are already at safepoint. 1729 assert_at_safepoint(); 1730 } 1731 1732 // After the handshake, safely free the ObjectMonitors that were 1733 // deflated and unlinked in this cycle. 1734 if (current->is_Java_thread()) { 1735 if (ls != NULL) { 1736 timer.stop(); 1737 ls->print_cr("before setting blocked: unlinked_count=" SIZE_FORMAT 1738 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count=" 1739 SIZE_FORMAT ", max=" SIZE_FORMAT, 1740 unlinked_count, in_use_list_ceiling(), 1741 _in_use_list.count(), _in_use_list.max()); 1742 } 1743 // Mark the calling JavaThread blocked (safepoint safe) while we free 1744 // the ObjectMonitors so we don't delay safepoints whilst doing that. 1745 ThreadBlockInVM tbivm(JavaThread::cast(current)); 1746 if (ls != NULL) { 1747 ls->print_cr("after setting blocked: in_use_list stats: ceiling=" 1748 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1749 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max()); 1750 timer.start(); 1751 } 1752 deleted_count = delete_monitors(&delete_list); 1753 // ThreadBlockInVM is destroyed here 1754 } else { 1755 // A non-JavaThread can just free the ObjectMonitors: 1756 deleted_count = delete_monitors(&delete_list); 1757 } 1758 assert(unlinked_count == deleted_count, "must be"); 1759 } 1760 1761 if (ls != nullptr) { 1762 timer.stop(); 1763 if (deflated_count != 0 || unlinked_count != 0 || log_is_enabled(Debug, monitorinflation)) { 1764 ls->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs", 1765 deflated_count, unlinked_count, timer.seconds()); 1766 } 1767 ls->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1768 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max()); 1769 if (table != nullptr) { 1770 ls->print_cr("ObjectMonitorsHashtable: key_count=" SIZE_FORMAT ", om_count=" SIZE_FORMAT, 1771 table->key_count(), table->om_count()); 1772 } 1773 } 1774 1775 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count())); 1776 OM_PERFDATA_OP(Deflations, inc(deflated_count)); 1777 1778 GVars.stw_random = os::random(); 1779 1780 if (deflated_count != 0) { 1781 _no_progress_cnt = 0; 1782 } else if (_no_progress_skip_increment) { 1783 _no_progress_skip_increment = false; 1784 } else { 1785 _no_progress_cnt++; 1786 } 1787 1788 return deflated_count; 1789 } 1790 1791 // Monitor cleanup on JavaThread::exit 1792 1793 // Iterate through monitor cache and attempt to release thread's monitors 1794 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1795 private: 1796 JavaThread* _thread; 1797 1798 public: 1799 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1800 void do_monitor(ObjectMonitor* mid) { 1801 intx rec = mid->complete_exit(_thread); 1802 _thread->dec_held_monitor_count(rec + 1); 1803 } 1804 }; 1805 1806 // Release all inflated monitors owned by current thread. Lightweight monitors are 1807 // ignored. This is meant to be called during JNI thread detach which assumes 1808 // all remaining monitors are heavyweight. All exceptions are swallowed. 1809 // Scanning the extant monitor list can be time consuming. 1810 // A simple optimization is to add a per-thread flag that indicates a thread 1811 // called jni_monitorenter() during its lifetime. 1812 // 1813 // Instead of NoSafepointVerifier it might be cheaper to 1814 // use an idiom of the form: 1815 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1816 // <code that must not run at safepoint> 1817 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1818 // Since the tests are extremely cheap we could leave them enabled 1819 // for normal product builds. 1820 1821 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1822 assert(current == JavaThread::current(), "must be current Java thread"); 1823 NoSafepointVerifier nsv; 1824 ReleaseJavaMonitorsClosure rjmc(current); 1825 ObjectSynchronizer::monitors_iterate(&rjmc, current); 1826 assert(!current->has_pending_exception(), "Should not be possible"); 1827 current->clear_pending_exception(); 1828 assert(current->held_monitor_count() == 0, "Should not be possible"); 1829 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1830 current->clear_jni_monitor_count(); 1831 } 1832 1833 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1834 switch (cause) { 1835 case inflate_cause_vm_internal: return "VM Internal"; 1836 case inflate_cause_monitor_enter: return "Monitor Enter"; 1837 case inflate_cause_wait: return "Monitor Wait"; 1838 case inflate_cause_notify: return "Monitor Notify"; 1839 case inflate_cause_hash_code: return "Monitor Hash Code"; 1840 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1841 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1842 default: 1843 ShouldNotReachHere(); 1844 } 1845 return "Unknown"; 1846 } 1847 1848 //------------------------------------------------------------------------------ 1849 // Debugging code 1850 1851 u_char* ObjectSynchronizer::get_gvars_addr() { 1852 return (u_char*)&GVars; 1853 } 1854 1855 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1856 return (u_char*)&GVars.hc_sequence; 1857 } 1858 1859 size_t ObjectSynchronizer::get_gvars_size() { 1860 return sizeof(SharedGlobals); 1861 } 1862 1863 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1864 return (u_char*)&GVars.stw_random; 1865 } 1866 1867 // Do the final audit and print of ObjectMonitor stats; must be done 1868 // by the VMThread at VM exit time. 1869 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1870 assert(Thread::current()->is_VM_thread(), "sanity check"); 1871 1872 if (is_final_audit()) { // Only do the audit once. 1873 return; 1874 } 1875 set_is_final_audit(); 1876 log_info(monitorinflation)("Starting the final audit."); 1877 1878 if (log_is_enabled(Info, monitorinflation)) { 1879 // Do deflations in order to reduce the in-use monitor population 1880 // that is reported by ObjectSynchronizer::log_in_use_monitor_details() 1881 // which is called by ObjectSynchronizer::audit_and_print_stats(). 1882 while (deflate_idle_monitors(/* ObjectMonitorsHashtable is not needed here */ nullptr) > 0) { 1883 ; // empty 1884 } 1885 // The other audit_and_print_stats() call is done at the Debug 1886 // level at a safepoint in SafepointSynchronize::do_cleanup_tasks. 1887 audit_and_print_stats(true /* on_exit */); 1888 } 1889 } 1890 1891 // This function can be called at a safepoint or it can be called when 1892 // we are trying to exit the VM. When we are trying to exit the VM, the 1893 // list walker functions can run in parallel with the other list 1894 // operations so spin-locking is used for safety. 1895 // 1896 // Calls to this function can be added in various places as a debugging 1897 // aid; pass 'true' for the 'on_exit' parameter to have in-use monitor 1898 // details logged at the Info level and 'false' for the 'on_exit' 1899 // parameter to have in-use monitor details logged at the Trace level. 1900 // 1901 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) { 1902 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant"); 1903 1904 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1905 LogStreamHandle(Info, monitorinflation) lsh_info; 1906 LogStreamHandle(Trace, monitorinflation) lsh_trace; 1907 LogStream* ls = nullptr; 1908 if (log_is_enabled(Trace, monitorinflation)) { 1909 ls = &lsh_trace; 1910 } else if (log_is_enabled(Debug, monitorinflation)) { 1911 ls = &lsh_debug; 1912 } else if (log_is_enabled(Info, monitorinflation)) { 1913 ls = &lsh_info; 1914 } 1915 assert(ls != nullptr, "sanity check"); 1916 1917 int error_cnt = 0; 1918 1919 ls->print_cr("Checking in_use_list:"); 1920 chk_in_use_list(ls, &error_cnt); 1921 1922 if (error_cnt == 0) { 1923 ls->print_cr("No errors found in in_use_list checks."); 1924 } else { 1925 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 1926 } 1927 1928 if ((on_exit && log_is_enabled(Info, monitorinflation)) || 1929 (!on_exit && log_is_enabled(Trace, monitorinflation))) { 1930 // When exiting this log output is at the Info level. When called 1931 // at a safepoint, this log output is at the Trace level since 1932 // there can be a lot of it. 1933 log_in_use_monitor_details(ls); 1934 } 1935 1936 ls->flush(); 1937 1938 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 1939 } 1940 1941 // Check the in_use_list; log the results of the checks. 1942 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 1943 size_t l_in_use_count = _in_use_list.count(); 1944 size_t l_in_use_max = _in_use_list.max(); 1945 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count, 1946 l_in_use_max); 1947 1948 size_t ck_in_use_count = 0; 1949 MonitorList::Iterator iter = _in_use_list.iterator(); 1950 while (iter.has_next()) { 1951 ObjectMonitor* mid = iter.next(); 1952 chk_in_use_entry(mid, out, error_cnt_p); 1953 ck_in_use_count++; 1954 } 1955 1956 if (l_in_use_count == ck_in_use_count) { 1957 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count=" 1958 SIZE_FORMAT, l_in_use_count, ck_in_use_count); 1959 } else { 1960 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to " 1961 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count, 1962 ck_in_use_count); 1963 } 1964 1965 size_t ck_in_use_max = _in_use_list.max(); 1966 if (l_in_use_max == ck_in_use_max) { 1967 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max=" 1968 SIZE_FORMAT, l_in_use_max, ck_in_use_max); 1969 } else { 1970 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to " 1971 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max); 1972 } 1973 } 1974 1975 // Check an in-use monitor entry; log any errors. 1976 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 1977 int* error_cnt_p) { 1978 if (n->owner_is_DEFLATER_MARKER()) { 1979 // This should not happen, but if it does, it is not fatal. 1980 out->print_cr("WARNING: monitor=" INTPTR_FORMAT ": in-use monitor is " 1981 "deflated.", p2i(n)); 1982 return; 1983 } 1984 if (n->header().value() == 0) { 1985 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 1986 "have non-null _header field.", p2i(n)); 1987 *error_cnt_p = *error_cnt_p + 1; 1988 } 1989 const oop obj = n->object_peek(); 1990 if (obj != nullptr) { 1991 const markWord mark = obj->mark(); 1992 if (!mark.has_monitor()) { 1993 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 1994 "object does not think it has a monitor: obj=" 1995 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 1996 p2i(obj), mark.value()); 1997 *error_cnt_p = *error_cnt_p + 1; 1998 } 1999 ObjectMonitor* const obj_mon = mark.monitor(); 2000 if (n != obj_mon) { 2001 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2002 "object does not refer to the same monitor: obj=" 2003 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2004 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2005 *error_cnt_p = *error_cnt_p + 1; 2006 } 2007 } 2008 } 2009 2010 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 2011 // flags indicate why the entry is in-use, 'object' and 'object type' 2012 // indicate the associated object and its type. 2013 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out) { 2014 stringStream ss; 2015 if (_in_use_list.count() > 0) { 2016 out->print_cr("In-use monitor info:"); 2017 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2018 out->print_cr("%18s %s %18s %18s", 2019 "monitor", "BHL", "object", "object type"); 2020 out->print_cr("================== === ================== =================="); 2021 MonitorList::Iterator iter = _in_use_list.iterator(); 2022 while (iter.has_next()) { 2023 ObjectMonitor* mid = iter.next(); 2024 const oop obj = mid->object_peek(); 2025 const markWord mark = mid->header(); 2026 ResourceMark rm; 2027 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(mid), 2028 mid->is_busy(), mark.hash() != 0, mid->owner() != nullptr, 2029 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 2030 if (mid->is_busy()) { 2031 out->print(" (%s)", mid->is_busy_to_string(&ss)); 2032 ss.reset(); 2033 } 2034 out->cr(); 2035 } 2036 } 2037 2038 out->flush(); 2039 }