1 /* 2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "gc/shared/collectedHeap.hpp" 28 #include "jfr/jfrEvents.hpp" 29 #include "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/padded.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "memory/universe.hpp" 35 #include "oops/markWord.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "runtime/atomic.hpp" 38 #include "runtime/frame.inline.hpp" 39 #include "runtime/globals.hpp" 40 #include "runtime/handles.inline.hpp" 41 #include "runtime/handshake.hpp" 42 #include "runtime/interfaceSupport.inline.hpp" 43 #include "runtime/javaThread.hpp" 44 #include "runtime/lockStack.inline.hpp" 45 #include "runtime/mutexLocker.hpp" 46 #include "runtime/objectMonitor.hpp" 47 #include "runtime/objectMonitor.inline.hpp" 48 #include "runtime/os.inline.hpp" 49 #include "runtime/osThread.hpp" 50 #include "runtime/perfData.hpp" 51 #include "runtime/safepointMechanism.inline.hpp" 52 #include "runtime/safepointVerifiers.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "runtime/stubRoutines.hpp" 55 #include "runtime/synchronizer.hpp" 56 #include "runtime/threads.hpp" 57 #include "runtime/timer.hpp" 58 #include "runtime/trimNativeHeap.hpp" 59 #include "runtime/vframe.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "utilities/align.hpp" 62 #include "utilities/dtrace.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/globalDefinitions.hpp" 65 #include "utilities/linkedlist.hpp" 66 #include "utilities/preserveException.hpp" 67 68 class ObjectMonitorDeflationLogging; 69 70 void MonitorList::add(ObjectMonitor* m) { 71 ObjectMonitor* head; 72 do { 73 head = Atomic::load(&_head); 74 m->set_next_om(head); 75 } while (Atomic::cmpxchg(&_head, head, m) != head); 76 77 size_t count = Atomic::add(&_count, 1u); 78 if (count > max()) { 79 Atomic::inc(&_max); 80 } 81 } 82 83 size_t MonitorList::count() const { 84 return Atomic::load(&_count); 85 } 86 87 size_t MonitorList::max() const { 88 return Atomic::load(&_max); 89 } 90 91 class ObjectMonitorDeflationSafepointer : public StackObj { 92 JavaThread* const _current; 93 ObjectMonitorDeflationLogging* const _log; 94 95 public: 96 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 97 : _current(current), _log(log) {} 98 99 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 100 }; 101 102 // Walk the in-use list and unlink deflated ObjectMonitors. 103 // Returns the number of unlinked ObjectMonitors. 104 size_t MonitorList::unlink_deflated(size_t deflated_count, 105 GrowableArray<ObjectMonitor*>* unlinked_list, 106 ObjectMonitorDeflationSafepointer* safepointer) { 107 size_t unlinked_count = 0; 108 ObjectMonitor* prev = nullptr; 109 ObjectMonitor* m = Atomic::load_acquire(&_head); 110 111 while (m != nullptr) { 112 if (m->is_being_async_deflated()) { 113 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 114 // modify the list once per batch. The batch starts at "m". 115 size_t unlinked_batch = 0; 116 ObjectMonitor* next = m; 117 // Look for at most MonitorUnlinkBatch monitors, or the number of 118 // deflated and not unlinked monitors, whatever comes first. 119 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 120 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 121 do { 122 ObjectMonitor* next_next = next->next_om(); 123 unlinked_batch++; 124 unlinked_list->append(next); 125 next = next_next; 126 if (unlinked_batch >= unlinked_batch_limit) { 127 // Reached the max batch, so bail out of the gathering loop. 128 break; 129 } 130 if (prev == nullptr && Atomic::load(&_head) != m) { 131 // Current batch used to be at head, but it is not at head anymore. 132 // Bail out and figure out where we currently are. This avoids long 133 // walks searching for new prev during unlink under heavy list inserts. 134 break; 135 } 136 } while (next != nullptr && next->is_being_async_deflated()); 137 138 // Unlink the found batch. 139 if (prev == nullptr) { 140 // The current batch is the first batch, so there is a chance that it starts at head. 141 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 142 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next); 143 if (prev_head != m) { 144 // Something must have updated the head. Figure out the actual prev for this batch. 145 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 146 prev = n; 147 } 148 assert(prev != nullptr, "Should have found the prev for the current batch"); 149 prev->set_next_om(next); 150 } 151 } else { 152 // The current batch is preceded by another batch. This guarantees the current batch 153 // does not start at head. Unlink the entire current batch without updating the head. 154 assert(Atomic::load(&_head) != m, "Sanity"); 155 prev->set_next_om(next); 156 } 157 158 unlinked_count += unlinked_batch; 159 if (unlinked_count >= deflated_count) { 160 // Reached the max so bail out of the searching loop. 161 // There should be no more deflated monitors left. 162 break; 163 } 164 m = next; 165 } else { 166 prev = m; 167 m = m->next_om(); 168 } 169 170 // Must check for a safepoint/handshake and honor it. 171 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 172 } 173 174 #ifdef ASSERT 175 // Invariant: the code above should unlink all deflated monitors. 176 // The code that runs after this unlinking does not expect deflated monitors. 177 // Notably, attempting to deflate the already deflated monitor would break. 178 { 179 ObjectMonitor* m = Atomic::load_acquire(&_head); 180 while (m != nullptr) { 181 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 182 m = m->next_om(); 183 } 184 } 185 #endif 186 187 Atomic::sub(&_count, unlinked_count); 188 return unlinked_count; 189 } 190 191 MonitorList::Iterator MonitorList::iterator() const { 192 return Iterator(Atomic::load_acquire(&_head)); 193 } 194 195 ObjectMonitor* MonitorList::Iterator::next() { 196 ObjectMonitor* current = _current; 197 _current = current->next_om(); 198 return current; 199 } 200 201 // The "core" versions of monitor enter and exit reside in this file. 202 // The interpreter and compilers contain specialized transliterated 203 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 204 // fast_lock(...) for instance. If you make changes here, make sure to modify the 205 // interpreter, and both C1 and C2 fast-path inline locking code emission. 206 // 207 // ----------------------------------------------------------------------------- 208 209 #ifdef DTRACE_ENABLED 210 211 // Only bother with this argument setup if dtrace is available 212 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 213 214 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 215 char* bytes = nullptr; \ 216 int len = 0; \ 217 jlong jtid = SharedRuntime::get_java_tid(thread); \ 218 Symbol* klassname = obj->klass()->name(); \ 219 if (klassname != nullptr) { \ 220 bytes = (char*)klassname->bytes(); \ 221 len = klassname->utf8_length(); \ 222 } 223 224 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 225 { \ 226 if (DTraceMonitorProbes) { \ 227 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 228 HOTSPOT_MONITOR_WAIT(jtid, \ 229 (uintptr_t)(monitor), bytes, len, (millis)); \ 230 } \ 231 } 232 233 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 234 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 235 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 236 237 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 238 { \ 239 if (DTraceMonitorProbes) { \ 240 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 241 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 242 (uintptr_t)(monitor), bytes, len); \ 243 } \ 244 } 245 246 #else // ndef DTRACE_ENABLED 247 248 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 249 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 250 251 #endif // ndef DTRACE_ENABLED 252 253 // This exists only as a workaround of dtrace bug 6254741 254 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 255 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 256 return 0; 257 } 258 259 static constexpr size_t inflation_lock_count() { 260 return 256; 261 } 262 263 // Static storage for an array of PlatformMutex. 264 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 265 266 static inline PlatformMutex* inflation_lock(size_t index) { 267 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 268 } 269 270 void ObjectSynchronizer::initialize() { 271 for (size_t i = 0; i < inflation_lock_count(); i++) { 272 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 273 } 274 // Start the ceiling with the estimate for one thread. 275 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 276 277 // Start the timer for deflations, so it does not trigger immediately. 278 _last_async_deflation_time_ns = os::javaTimeNanos(); 279 } 280 281 MonitorList ObjectSynchronizer::_in_use_list; 282 // monitors_used_above_threshold() policy is as follows: 283 // 284 // The ratio of the current _in_use_list count to the ceiling is used 285 // to determine if we are above MonitorUsedDeflationThreshold and need 286 // to do an async monitor deflation cycle. The ceiling is increased by 287 // AvgMonitorsPerThreadEstimate when a thread is added to the system 288 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 289 // removed from the system. 290 // 291 // Note: If the _in_use_list max exceeds the ceiling, then 292 // monitors_used_above_threshold() will use the in_use_list max instead 293 // of the thread count derived ceiling because we have used more 294 // ObjectMonitors than the estimated average. 295 // 296 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 297 // no-progress async monitor deflation cycles in a row, then the ceiling 298 // is adjusted upwards by monitors_used_above_threshold(). 299 // 300 // Start the ceiling with the estimate for one thread in initialize() 301 // which is called after cmd line options are processed. 302 static size_t _in_use_list_ceiling = 0; 303 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 304 bool volatile ObjectSynchronizer::_is_final_audit = false; 305 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 306 static uintx _no_progress_cnt = 0; 307 static bool _no_progress_skip_increment = false; 308 309 // These checks are required for wait, notify and exit to avoid inflating the monitor to 310 // find out this inline type object cannot be locked. 311 #define CHECK_THROW_NOSYNC_IMSE(obj) \ 312 if (EnableValhalla && (obj)->mark().is_inline_type()) { \ 313 JavaThread* THREAD = current; \ 314 ResourceMark rm(THREAD); \ 315 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), obj->klass()->external_name()); \ 316 } 317 318 #define CHECK_THROW_NOSYNC_IMSE_0(obj) \ 319 if (EnableValhalla && (obj)->mark().is_inline_type()) { \ 320 JavaThread* THREAD = current; \ 321 ResourceMark rm(THREAD); \ 322 THROW_MSG_0(vmSymbols::java_lang_IllegalMonitorStateException(), obj->klass()->external_name()); \ 323 } 324 325 // =====================> Quick functions 326 327 // The quick_* forms are special fast-path variants used to improve 328 // performance. In the simplest case, a "quick_*" implementation could 329 // simply return false, in which case the caller will perform the necessary 330 // state transitions and call the slow-path form. 331 // The fast-path is designed to handle frequently arising cases in an efficient 332 // manner and is just a degenerate "optimistic" variant of the slow-path. 333 // returns true -- to indicate the call was satisfied. 334 // returns false -- to indicate the call needs the services of the slow-path. 335 // A no-loitering ordinance is in effect for code in the quick_* family 336 // operators: safepoints or indefinite blocking (blocking that might span a 337 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 338 // entry. 339 // 340 // Consider: An interesting optimization is to have the JIT recognize the 341 // following common idiom: 342 // synchronized (someobj) { .... ; notify(); } 343 // That is, we find a notify() or notifyAll() call that immediately precedes 344 // the monitorexit operation. In that case the JIT could fuse the operations 345 // into a single notifyAndExit() runtime primitive. 346 347 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 348 assert(current->thread_state() == _thread_in_Java, "invariant"); 349 NoSafepointVerifier nsv; 350 if (obj == nullptr) return false; // slow-path for invalid obj 351 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "monitor op on inline type"); 352 const markWord mark = obj->mark(); 353 354 if (LockingMode == LM_LIGHTWEIGHT) { 355 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 356 // Degenerate notify 357 // fast-locked by caller so by definition the implied waitset is empty. 358 return true; 359 } 360 } else if (LockingMode == LM_LEGACY) { 361 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 362 // Degenerate notify 363 // stack-locked by caller so by definition the implied waitset is empty. 364 return true; 365 } 366 } 367 368 if (mark.has_monitor()) { 369 ObjectMonitor* const mon = mark.monitor(); 370 assert(mon->object() == oop(obj), "invariant"); 371 if (mon->owner() != current) return false; // slow-path for IMS exception 372 373 if (mon->first_waiter() != nullptr) { 374 // We have one or more waiters. Since this is an inflated monitor 375 // that we own, we can transfer one or more threads from the waitset 376 // to the entrylist here and now, avoiding the slow-path. 377 if (all) { 378 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current); 379 } else { 380 DTRACE_MONITOR_PROBE(notify, mon, obj, current); 381 } 382 int free_count = 0; 383 do { 384 mon->INotify(current); 385 ++free_count; 386 } while (mon->first_waiter() != nullptr && all); 387 OM_PERFDATA_OP(Notifications, inc(free_count)); 388 } 389 return true; 390 } 391 392 // other IMS exception states take the slow-path 393 return false; 394 } 395 396 397 // The LockNode emitted directly at the synchronization site would have 398 // been too big if it were to have included support for the cases of inflated 399 // recursive enter and exit, so they go here instead. 400 // Note that we can't safely call AsyncPrintJavaStack() from within 401 // quick_enter() as our thread state remains _in_Java. 402 403 bool ObjectSynchronizer::quick_enter(oop obj, JavaThread* current, 404 BasicLock * lock) { 405 assert(current->thread_state() == _thread_in_Java, "invariant"); 406 NoSafepointVerifier nsv; 407 if (obj == nullptr) return false; // Need to throw NPE 408 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "monitor op on inline type"); 409 410 if (obj->klass()->is_value_based()) { 411 return false; 412 } 413 414 if (LockingMode == LM_LIGHTWEIGHT) { 415 LockStack& lock_stack = current->lock_stack(); 416 if (lock_stack.is_full()) { 417 // Always go into runtime if the lock stack is full. 418 return false; 419 } 420 if (lock_stack.try_recursive_enter(obj)) { 421 // Recursive lock successful. 422 current->inc_held_monitor_count(); 423 return true; 424 } 425 } 426 427 const markWord mark = obj->mark(); 428 429 if (mark.has_monitor()) { 430 ObjectMonitor* const m = mark.monitor(); 431 // An async deflation or GC can race us before we manage to make 432 // the ObjectMonitor busy by setting the owner below. If we detect 433 // that race we just bail out to the slow-path here. 434 if (m->object_peek() == nullptr) { 435 return false; 436 } 437 JavaThread* const owner = static_cast<JavaThread*>(m->owner_raw()); 438 439 // Lock contention and Transactional Lock Elision (TLE) diagnostics 440 // and observability 441 // Case: light contention possibly amenable to TLE 442 // Case: TLE inimical operations such as nested/recursive synchronization 443 444 if (owner == current) { 445 m->_recursions++; 446 current->inc_held_monitor_count(); 447 return true; 448 } 449 450 if (LockingMode != LM_LIGHTWEIGHT) { 451 // This Java Monitor is inflated so obj's header will never be 452 // displaced to this thread's BasicLock. Make the displaced header 453 // non-null so this BasicLock is not seen as recursive nor as 454 // being locked. We do this unconditionally so that this thread's 455 // BasicLock cannot be mis-interpreted by any stack walkers. For 456 // performance reasons, stack walkers generally first check for 457 // stack-locking in the object's header, the second check is for 458 // recursive stack-locking in the displaced header in the BasicLock, 459 // and last are the inflated Java Monitor (ObjectMonitor) checks. 460 lock->set_displaced_header(markWord::unused_mark()); 461 } 462 463 if (owner == nullptr && m->try_set_owner_from(nullptr, current) == nullptr) { 464 assert(m->_recursions == 0, "invariant"); 465 current->inc_held_monitor_count(); 466 return true; 467 } 468 } 469 470 // Note that we could inflate in quick_enter. 471 // This is likely a useful optimization 472 // Critically, in quick_enter() we must not: 473 // -- block indefinitely, or 474 // -- reach a safepoint 475 476 return false; // revert to slow-path 477 } 478 479 // Handle notifications when synchronizing on value based classes 480 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 481 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 482 frame last_frame = locking_thread->last_frame(); 483 bool bcp_was_adjusted = false; 484 // Don't decrement bcp if it points to the frame's first instruction. This happens when 485 // handle_sync_on_value_based_class() is called because of a synchronized method. There 486 // is no actual monitorenter instruction in the byte code in this case. 487 if (last_frame.is_interpreted_frame() && 488 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 489 // adjust bcp to point back to monitorenter so that we print the correct line numbers 490 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 491 bcp_was_adjusted = true; 492 } 493 494 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 495 ResourceMark rm; 496 stringStream ss; 497 locking_thread->print_active_stack_on(&ss); 498 char* base = (char*)strstr(ss.base(), "at"); 499 char* newline = (char*)strchr(ss.base(), '\n'); 500 if (newline != nullptr) { 501 *newline = '\0'; 502 } 503 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 504 } else { 505 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 506 ResourceMark rm; 507 Log(valuebasedclasses) vblog; 508 509 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 510 if (locking_thread->has_last_Java_frame()) { 511 LogStream info_stream(vblog.info()); 512 locking_thread->print_active_stack_on(&info_stream); 513 } else { 514 vblog.info("Cannot find the last Java frame"); 515 } 516 517 EventSyncOnValueBasedClass event; 518 if (event.should_commit()) { 519 event.set_valueBasedClass(obj->klass()); 520 event.commit(); 521 } 522 } 523 524 if (bcp_was_adjusted) { 525 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 526 } 527 } 528 529 static bool useHeavyMonitors() { 530 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390) 531 return LockingMode == LM_MONITOR; 532 #else 533 return false; 534 #endif 535 } 536 537 // ----------------------------------------------------------------------------- 538 // Monitor Enter/Exit 539 540 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 541 // When called with locking_thread != Thread::current() some mechanism must synchronize 542 // the locking_thread with respect to the current thread. Currently only used when 543 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 544 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 545 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "JITed code should never have locked an instance of a value class"); 546 if (!enter_fast_impl(obj, lock, locking_thread)) { 547 // Inflated ObjectMonitor::enter_for is required 548 549 // An async deflation can race after the inflate_for() call and before 550 // enter_for() can make the ObjectMonitor busy. enter_for() returns false 551 // if we have lost the race to async deflation and we simply try again. 552 while (true) { 553 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter); 554 if (monitor->enter_for(locking_thread)) { 555 return; 556 } 557 assert(monitor->is_being_async_deflated(), "must be"); 558 } 559 } 560 } 561 562 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, JavaThread* current) { 563 assert(current == Thread::current(), "must be"); 564 assert(!EnableValhalla || !obj->klass()->is_inline_klass(), "This method should never be called on an instance of an inline class"); 565 if (!enter_fast_impl(obj, lock, current)) { 566 // Inflated ObjectMonitor::enter is required 567 568 // An async deflation can race after the inflate() call and before 569 // enter() can make the ObjectMonitor busy. enter() returns false if 570 // we have lost the race to async deflation and we simply try again. 571 while (true) { 572 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter); 573 if (monitor->enter(current)) { 574 return; 575 } 576 } 577 } 578 } 579 580 // The interpreter and compiler assembly code tries to lock using the fast path 581 // of this algorithm. Make sure to update that code if the following function is 582 // changed. The implementation is extremely sensitive to race condition. Be careful. 583 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 584 guarantee(!EnableValhalla || !obj->klass()->is_inline_klass(), "Attempt to inflate inline type"); 585 if (obj->klass()->is_value_based()) { 586 handle_sync_on_value_based_class(obj, locking_thread); 587 } 588 589 locking_thread->inc_held_monitor_count(); 590 591 if (!useHeavyMonitors()) { 592 if (LockingMode == LM_LIGHTWEIGHT) { 593 // Fast-locking does not use the 'lock' argument. 594 LockStack& lock_stack = locking_thread->lock_stack(); 595 if (lock_stack.is_full()) { 596 // We unconditionally make room on the lock stack by inflating 597 // the least recently locked object on the lock stack. 598 599 // About the choice to inflate least recently locked object. 600 // First we must chose to inflate a lock, either some lock on 601 // the lock-stack or the lock that is currently being entered 602 // (which may or may not be on the lock-stack). 603 // Second the best lock to inflate is a lock which is entered 604 // in a control flow where there are only a very few locks being 605 // used, as the costly part of inflated locking is inflation, 606 // not locking. But this property is entirely program dependent. 607 // Third inflating the lock currently being entered on when it 608 // is not present on the lock-stack will result in a still full 609 // lock-stack. This creates a scenario where every deeper nested 610 // monitorenter must call into the runtime. 611 // The rational here is as follows: 612 // Because we cannot (currently) figure out the second, and want 613 // to avoid the third, we inflate a lock on the lock-stack. 614 // The least recently locked lock is chosen as it is the lock 615 // with the longest critical section. 616 617 log_info(monitorinflation)("LockStack capacity exceeded, inflating."); 618 ObjectMonitor* monitor = inflate_for(locking_thread, lock_stack.bottom(), inflate_cause_vm_internal); 619 assert(monitor->owner() == Thread::current(), "must be owner=" PTR_FORMAT " current=" PTR_FORMAT " mark=" PTR_FORMAT, 620 p2i(monitor->owner()), p2i(Thread::current()), monitor->object()->mark_acquire().value()); 621 assert(!lock_stack.is_full(), "must have made room here"); 622 } 623 624 markWord mark = obj()->mark_acquire(); 625 while (mark.is_neutral()) { 626 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications. 627 // Try to swing into 'fast-locked' state. 628 assert(!lock_stack.contains(obj()), "thread must not already hold the lock"); 629 const markWord locked_mark = mark.set_fast_locked(); 630 const markWord old_mark = obj()->cas_set_mark(locked_mark, mark); 631 if (old_mark == mark) { 632 // Successfully fast-locked, push object to lock-stack and return. 633 lock_stack.push(obj()); 634 return true; 635 } 636 mark = old_mark; 637 } 638 639 if (mark.is_fast_locked() && lock_stack.try_recursive_enter(obj())) { 640 // Recursive lock successful. 641 return true; 642 } 643 644 // Failed to fast lock. 645 return false; 646 } else if (LockingMode == LM_LEGACY) { 647 markWord mark = obj->mark(); 648 if (mark.is_neutral()) { 649 // Anticipate successful CAS -- the ST of the displaced mark must 650 // be visible <= the ST performed by the CAS. 651 lock->set_displaced_header(mark); 652 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 653 return true; 654 } 655 } else if (mark.has_locker() && 656 locking_thread->is_lock_owned((address) mark.locker())) { 657 assert(lock != mark.locker(), "must not re-lock the same lock"); 658 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock"); 659 lock->set_displaced_header(markWord::from_pointer(nullptr)); 660 return true; 661 } 662 663 // The object header will never be displaced to this lock, 664 // so it does not matter what the value is, except that it 665 // must be non-zero to avoid looking like a re-entrant lock, 666 // and must not look locked either. 667 lock->set_displaced_header(markWord::unused_mark()); 668 669 // Failed to fast lock. 670 return false; 671 } 672 } else if (VerifyHeavyMonitors) { 673 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 674 } 675 676 return false; 677 } 678 679 void ObjectSynchronizer::exit(oop object, BasicLock* lock, JavaThread* current) { 680 current->dec_held_monitor_count(); 681 682 if (!useHeavyMonitors()) { 683 markWord mark = object->mark(); 684 if (EnableValhalla && mark.is_inline_type()) { 685 return; 686 } 687 if (LockingMode == LM_LIGHTWEIGHT) { 688 // Fast-locking does not use the 'lock' argument. 689 LockStack& lock_stack = current->lock_stack(); 690 if (mark.is_fast_locked() && lock_stack.try_recursive_exit(object)) { 691 // Recursively unlocked. 692 return; 693 } 694 695 if (mark.is_fast_locked() && lock_stack.is_recursive(object)) { 696 // This lock is recursive but is not at the top of the lock stack so we're 697 // doing an unbalanced exit. We have to fall thru to inflation below and 698 // let ObjectMonitor::exit() do the unlock. 699 } else { 700 while (mark.is_fast_locked()) { 701 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications. 702 const markWord unlocked_mark = mark.set_unlocked(); 703 const markWord old_mark = object->cas_set_mark(unlocked_mark, mark); 704 if (old_mark == mark) { 705 size_t recursions = lock_stack.remove(object) - 1; 706 assert(recursions == 0, "must not be recursive here"); 707 return; 708 } 709 mark = old_mark; 710 } 711 } 712 } else if (LockingMode == LM_LEGACY) { 713 markWord dhw = lock->displaced_header(); 714 if (dhw.value() == 0) { 715 // If the displaced header is null, then this exit matches up with 716 // a recursive enter. No real work to do here except for diagnostics. 717 #ifndef PRODUCT 718 if (mark != markWord::INFLATING()) { 719 // Only do diagnostics if we are not racing an inflation. Simply 720 // exiting a recursive enter of a Java Monitor that is being 721 // inflated is safe; see the has_monitor() comment below. 722 assert(!mark.is_neutral(), "invariant"); 723 assert(!mark.has_locker() || 724 current->is_lock_owned((address)mark.locker()), "invariant"); 725 if (mark.has_monitor()) { 726 // The BasicLock's displaced_header is marked as a recursive 727 // enter and we have an inflated Java Monitor (ObjectMonitor). 728 // This is a special case where the Java Monitor was inflated 729 // after this thread entered the stack-lock recursively. When a 730 // Java Monitor is inflated, we cannot safely walk the Java 731 // Monitor owner's stack and update the BasicLocks because a 732 // Java Monitor can be asynchronously inflated by a thread that 733 // does not own the Java Monitor. 734 ObjectMonitor* m = mark.monitor(); 735 assert(m->object()->mark() == mark, "invariant"); 736 assert(m->is_entered(current), "invariant"); 737 } 738 } 739 #endif 740 return; 741 } 742 743 if (mark == markWord::from_pointer(lock)) { 744 // If the object is stack-locked by the current thread, try to 745 // swing the displaced header from the BasicLock back to the mark. 746 assert(dhw.is_neutral(), "invariant"); 747 if (object->cas_set_mark(dhw, mark) == mark) { 748 return; 749 } 750 } 751 } 752 } else if (VerifyHeavyMonitors) { 753 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 754 } 755 756 // We have to take the slow-path of possible inflation and then exit. 757 // The ObjectMonitor* can't be async deflated until ownership is 758 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 759 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal); 760 assert(!monitor->is_owner_anonymous(), "must not be"); 761 monitor->exit(current); 762 } 763 764 // ----------------------------------------------------------------------------- 765 // JNI locks on java objects 766 // NOTE: must use heavy weight monitor to handle jni monitor enter 767 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 768 JavaThread* THREAD = current; 769 if (obj->klass()->is_value_based()) { 770 handle_sync_on_value_based_class(obj, current); 771 } 772 773 if (EnableValhalla && obj->klass()->is_inline_klass()) { 774 ResourceMark rm(THREAD); 775 const char* desc = "Cannot synchronize on an instance of value class "; 776 const char* className = obj->klass()->external_name(); 777 size_t msglen = strlen(desc) + strlen(className) + 1; 778 char* message = NEW_RESOURCE_ARRAY(char, msglen); 779 assert(message != nullptr, "NEW_RESOURCE_ARRAY should have called vm_exit_out_of_memory and not return nullptr"); 780 THROW_MSG(vmSymbols::java_lang_IdentityException(), className); 781 } 782 783 // the current locking is from JNI instead of Java code 784 current->set_current_pending_monitor_is_from_java(false); 785 // An async deflation can race after the inflate() call and before 786 // enter() can make the ObjectMonitor busy. enter() returns false if 787 // we have lost the race to async deflation and we simply try again. 788 while (true) { 789 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_jni_enter); 790 if (monitor->enter(current)) { 791 current->inc_held_monitor_count(1, true); 792 break; 793 } 794 } 795 current->set_current_pending_monitor_is_from_java(true); 796 } 797 798 // NOTE: must use heavy weight monitor to handle jni monitor exit 799 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 800 JavaThread* current = THREAD; 801 CHECK_THROW_NOSYNC_IMSE(obj); 802 803 // The ObjectMonitor* can't be async deflated until ownership is 804 // dropped inside exit() and the ObjectMonitor* must be !is_busy(). 805 ObjectMonitor* monitor = inflate(current, obj, inflate_cause_jni_exit); 806 // If this thread has locked the object, exit the monitor. We 807 // intentionally do not use CHECK on check_owner because we must exit the 808 // monitor even if an exception was already pending. 809 if (monitor->check_owner(THREAD)) { 810 monitor->exit(current); 811 current->dec_held_monitor_count(1, true); 812 } 813 } 814 815 // ----------------------------------------------------------------------------- 816 // Internal VM locks on java objects 817 // standard constructor, allows locking failures 818 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) { 819 _thread = thread; 820 _thread->check_for_valid_safepoint_state(); 821 _obj = obj; 822 823 if (_obj() != nullptr) { 824 ObjectSynchronizer::enter(_obj, &_lock, _thread); 825 } 826 } 827 828 ObjectLocker::~ObjectLocker() { 829 if (_obj() != nullptr) { 830 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 831 } 832 } 833 834 835 // ----------------------------------------------------------------------------- 836 // Wait/Notify/NotifyAll 837 // NOTE: must use heavy weight monitor to handle wait() 838 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 839 JavaThread* current = THREAD; 840 CHECK_THROW_NOSYNC_IMSE_0(obj); 841 if (millis < 0) { 842 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 843 } 844 // The ObjectMonitor* can't be async deflated because the _waiters 845 // field is incremented before ownership is dropped and decremented 846 // after ownership is regained. 847 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait); 848 849 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 850 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 851 852 // This dummy call is in place to get around dtrace bug 6254741. Once 853 // that's fixed we can uncomment the following line, remove the call 854 // and change this function back into a "void" func. 855 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 856 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 857 return ret_code; 858 } 859 860 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 861 JavaThread* current = THREAD; 862 CHECK_THROW_NOSYNC_IMSE(obj); 863 864 markWord mark = obj->mark(); 865 if (LockingMode == LM_LIGHTWEIGHT) { 866 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 867 // Not inflated so there can't be any waiters to notify. 868 return; 869 } 870 } else if (LockingMode == LM_LEGACY) { 871 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 872 // Not inflated so there can't be any waiters to notify. 873 return; 874 } 875 } 876 // The ObjectMonitor* can't be async deflated until ownership is 877 // dropped by the calling thread. 878 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 879 monitor->notify(CHECK); 880 } 881 882 // NOTE: see comment of notify() 883 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 884 JavaThread* current = THREAD; 885 CHECK_THROW_NOSYNC_IMSE(obj); 886 887 markWord mark = obj->mark(); 888 if (LockingMode == LM_LIGHTWEIGHT) { 889 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 890 // Not inflated so there can't be any waiters to notify. 891 return; 892 } 893 } else if (LockingMode == LM_LEGACY) { 894 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 895 // Not inflated so there can't be any waiters to notify. 896 return; 897 } 898 } 899 // The ObjectMonitor* can't be async deflated until ownership is 900 // dropped by the calling thread. 901 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify); 902 monitor->notifyAll(CHECK); 903 } 904 905 // ----------------------------------------------------------------------------- 906 // Hash Code handling 907 908 struct SharedGlobals { 909 char _pad_prefix[OM_CACHE_LINE_SIZE]; 910 // This is a highly shared mostly-read variable. 911 // To avoid false-sharing it needs to be the sole occupant of a cache line. 912 volatile int stw_random; 913 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 914 // Hot RW variable -- Sequester to avoid false-sharing 915 volatile int hc_sequence; 916 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 917 }; 918 919 static SharedGlobals GVars; 920 921 static markWord read_stable_mark(oop obj) { 922 markWord mark = obj->mark_acquire(); 923 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) { 924 // New lightweight locking does not use the markWord::INFLATING() protocol. 925 return mark; // normal fast-path return 926 } 927 928 int its = 0; 929 for (;;) { 930 markWord mark = obj->mark_acquire(); 931 if (!mark.is_being_inflated()) { 932 return mark; // normal fast-path return 933 } 934 935 // The object is being inflated by some other thread. 936 // The caller of read_stable_mark() must wait for inflation to complete. 937 // Avoid live-lock. 938 939 ++its; 940 if (its > 10000 || !os::is_MP()) { 941 if (its & 1) { 942 os::naked_yield(); 943 } else { 944 // Note that the following code attenuates the livelock problem but is not 945 // a complete remedy. A more complete solution would require that the inflating 946 // thread hold the associated inflation lock. The following code simply restricts 947 // the number of spinners to at most one. We'll have N-2 threads blocked 948 // on the inflationlock, 1 thread holding the inflation lock and using 949 // a yield/park strategy, and 1 thread in the midst of inflation. 950 // A more refined approach would be to change the encoding of INFLATING 951 // to allow encapsulation of a native thread pointer. Threads waiting for 952 // inflation to complete would use CAS to push themselves onto a singly linked 953 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 954 // and calling park(). When inflation was complete the thread that accomplished inflation 955 // would detach the list and set the markword to inflated with a single CAS and 956 // then for each thread on the list, set the flag and unpark() the thread. 957 958 // Index into the lock array based on the current object address. 959 static_assert(is_power_of_2(inflation_lock_count()), "must be"); 960 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1); 961 int YieldThenBlock = 0; 962 assert(ix < inflation_lock_count(), "invariant"); 963 inflation_lock(ix)->lock(); 964 while (obj->mark_acquire() == markWord::INFLATING()) { 965 // Beware: naked_yield() is advisory and has almost no effect on some platforms 966 // so we periodically call current->_ParkEvent->park(1). 967 // We use a mixed spin/yield/block mechanism. 968 if ((YieldThenBlock++) >= 16) { 969 Thread::current()->_ParkEvent->park(1); 970 } else { 971 os::naked_yield(); 972 } 973 } 974 inflation_lock(ix)->unlock(); 975 } 976 } else { 977 SpinPause(); // SMP-polite spinning 978 } 979 } 980 } 981 982 // hashCode() generation : 983 // 984 // Possibilities: 985 // * MD5Digest of {obj,stw_random} 986 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 987 // * A DES- or AES-style SBox[] mechanism 988 // * One of the Phi-based schemes, such as: 989 // 2654435761 = 2^32 * Phi (golden ratio) 990 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 991 // * A variation of Marsaglia's shift-xor RNG scheme. 992 // * (obj ^ stw_random) is appealing, but can result 993 // in undesirable regularity in the hashCode values of adjacent objects 994 // (objects allocated back-to-back, in particular). This could potentially 995 // result in hashtable collisions and reduced hashtable efficiency. 996 // There are simple ways to "diffuse" the middle address bits over the 997 // generated hashCode values: 998 999 static inline intptr_t get_next_hash(Thread* current, oop obj) { 1000 intptr_t value = 0; 1001 if (hashCode == 0) { 1002 // This form uses global Park-Miller RNG. 1003 // On MP system we'll have lots of RW access to a global, so the 1004 // mechanism induces lots of coherency traffic. 1005 value = os::random(); 1006 } else if (hashCode == 1) { 1007 // This variation has the property of being stable (idempotent) 1008 // between STW operations. This can be useful in some of the 1-0 1009 // synchronization schemes. 1010 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 1011 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 1012 } else if (hashCode == 2) { 1013 value = 1; // for sensitivity testing 1014 } else if (hashCode == 3) { 1015 value = ++GVars.hc_sequence; 1016 } else if (hashCode == 4) { 1017 value = cast_from_oop<intptr_t>(obj); 1018 } else { 1019 // Marsaglia's xor-shift scheme with thread-specific state 1020 // This is probably the best overall implementation -- we'll 1021 // likely make this the default in future releases. 1022 unsigned t = current->_hashStateX; 1023 t ^= (t << 11); 1024 current->_hashStateX = current->_hashStateY; 1025 current->_hashStateY = current->_hashStateZ; 1026 current->_hashStateZ = current->_hashStateW; 1027 unsigned v = current->_hashStateW; 1028 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 1029 current->_hashStateW = v; 1030 value = v; 1031 } 1032 1033 value &= markWord::hash_mask; 1034 if (value == 0) value = 0xBAD; 1035 assert(value != markWord::no_hash, "invariant"); 1036 return value; 1037 } 1038 1039 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 1040 if (EnableValhalla && obj->klass()->is_inline_klass()) { 1041 // VM should be calling bootstrap method 1042 ShouldNotReachHere(); 1043 } 1044 1045 while (true) { 1046 ObjectMonitor* monitor = nullptr; 1047 markWord temp, test; 1048 intptr_t hash; 1049 markWord mark = read_stable_mark(obj); 1050 if (VerifyHeavyMonitors) { 1051 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)"); 1052 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked"); 1053 } 1054 if (mark.is_neutral() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) { 1055 hash = mark.hash(); 1056 if (hash != 0) { // if it has a hash, just return it 1057 return hash; 1058 } 1059 hash = get_next_hash(current, obj); // get a new hash 1060 temp = mark.copy_set_hash(hash); // merge the hash into header 1061 // try to install the hash 1062 test = obj->cas_set_mark(temp, mark); 1063 if (test == mark) { // if the hash was installed, return it 1064 return hash; 1065 } 1066 if (LockingMode == LM_LIGHTWEIGHT) { 1067 // CAS failed, retry 1068 continue; 1069 } 1070 // Failed to install the hash. It could be that another thread 1071 // installed the hash just before our attempt or inflation has 1072 // occurred or... so we fall thru to inflate the monitor for 1073 // stability and then install the hash. 1074 } else if (mark.has_monitor()) { 1075 monitor = mark.monitor(); 1076 temp = monitor->header(); 1077 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1078 hash = temp.hash(); 1079 if (hash != 0) { 1080 // It has a hash. 1081 1082 // Separate load of dmw/header above from the loads in 1083 // is_being_async_deflated(). 1084 1085 // dmw/header and _contentions may get written by different threads. 1086 // Make sure to observe them in the same order when having several observers. 1087 OrderAccess::loadload_for_IRIW(); 1088 1089 if (monitor->is_being_async_deflated()) { 1090 // But we can't safely use the hash if we detect that async 1091 // deflation has occurred. So we attempt to restore the 1092 // header/dmw to the object's header so that we only retry 1093 // once if the deflater thread happens to be slow. 1094 monitor->install_displaced_markword_in_object(obj); 1095 continue; 1096 } 1097 return hash; 1098 } 1099 // Fall thru so we only have one place that installs the hash in 1100 // the ObjectMonitor. 1101 } else if (LockingMode == LM_LEGACY && mark.has_locker() && current->is_lock_owned((address)mark.locker())) { 1102 // This is a stack-lock owned by the calling thread so fetch the 1103 // displaced markWord from the BasicLock on the stack. 1104 temp = mark.displaced_mark_helper(); 1105 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1106 hash = temp.hash(); 1107 if (hash != 0) { // if it has a hash, just return it 1108 return hash; 1109 } 1110 // WARNING: 1111 // The displaced header in the BasicLock on a thread's stack 1112 // is strictly immutable. It CANNOT be changed in ANY cases. 1113 // So we have to inflate the stack-lock into an ObjectMonitor 1114 // even if the current thread owns the lock. The BasicLock on 1115 // a thread's stack can be asynchronously read by other threads 1116 // during an inflate() call so any change to that stack memory 1117 // may not propagate to other threads correctly. 1118 } 1119 1120 // Inflate the monitor to set the hash. 1121 1122 // An async deflation can race after the inflate() call and before we 1123 // can update the ObjectMonitor's header with the hash value below. 1124 monitor = inflate(current, obj, inflate_cause_hash_code); 1125 // Load ObjectMonitor's header/dmw field and see if it has a hash. 1126 mark = monitor->header(); 1127 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1128 hash = mark.hash(); 1129 if (hash == 0) { // if it does not have a hash 1130 hash = get_next_hash(current, obj); // get a new hash 1131 temp = mark.copy_set_hash(hash) ; // merge the hash into header 1132 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1133 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value()); 1134 test = markWord(v); 1135 if (test != mark) { 1136 // The attempt to update the ObjectMonitor's header/dmw field 1137 // did not work. This can happen if another thread managed to 1138 // merge in the hash just before our cmpxchg(). 1139 // If we add any new usages of the header/dmw field, this code 1140 // will need to be updated. 1141 hash = test.hash(); 1142 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1143 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1144 } 1145 if (monitor->is_being_async_deflated()) { 1146 // If we detect that async deflation has occurred, then we 1147 // attempt to restore the header/dmw to the object's header 1148 // so that we only retry once if the deflater thread happens 1149 // to be slow. 1150 monitor->install_displaced_markword_in_object(obj); 1151 continue; 1152 } 1153 } 1154 // We finally get the hash. 1155 return hash; 1156 } 1157 } 1158 1159 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 1160 Handle h_obj) { 1161 if (EnableValhalla && h_obj->mark().is_inline_type()) { 1162 return false; 1163 } 1164 assert(current == JavaThread::current(), "Can only be called on current thread"); 1165 oop obj = h_obj(); 1166 1167 markWord mark = read_stable_mark(obj); 1168 1169 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1170 // stack-locked case, header points into owner's stack 1171 return current->is_lock_owned((address)mark.locker()); 1172 } 1173 1174 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1175 // fast-locking case, see if lock is in current's lock stack 1176 return current->lock_stack().contains(h_obj()); 1177 } 1178 1179 if (mark.has_monitor()) { 1180 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1181 // The first stage of async deflation does not affect any field 1182 // used by this comparison so the ObjectMonitor* is usable here. 1183 ObjectMonitor* monitor = mark.monitor(); 1184 return monitor->is_entered(current) != 0; 1185 } 1186 // Unlocked case, header in place 1187 assert(mark.is_neutral(), "sanity check"); 1188 return false; 1189 } 1190 1191 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1192 oop obj = h_obj(); 1193 markWord mark = read_stable_mark(obj); 1194 1195 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1196 // stack-locked so header points into owner's stack. 1197 // owning_thread_from_monitor_owner() may also return null here: 1198 return Threads::owning_thread_from_monitor_owner(t_list, (address) mark.locker()); 1199 } 1200 1201 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1202 // fast-locked so get owner from the object. 1203 // owning_thread_from_object() may also return null here: 1204 return Threads::owning_thread_from_object(t_list, h_obj()); 1205 } 1206 1207 if (mark.has_monitor()) { 1208 // Inflated monitor so header points to ObjectMonitor (tagged pointer). 1209 // The first stage of async deflation does not affect any field 1210 // used by this comparison so the ObjectMonitor* is usable here. 1211 ObjectMonitor* monitor = mark.monitor(); 1212 assert(monitor != nullptr, "monitor should be non-null"); 1213 // owning_thread_from_monitor() may also return null here: 1214 return Threads::owning_thread_from_monitor(t_list, monitor); 1215 } 1216 1217 // Unlocked case, header in place 1218 // Cannot have assertion since this object may have been 1219 // locked by another thread when reaching here. 1220 // assert(mark.is_neutral(), "sanity check"); 1221 1222 return nullptr; 1223 } 1224 1225 // Visitors ... 1226 1227 // Iterate over all ObjectMonitors. 1228 template <typename Function> 1229 void ObjectSynchronizer::monitors_iterate(Function function) { 1230 MonitorList::Iterator iter = _in_use_list.iterator(); 1231 while (iter.has_next()) { 1232 ObjectMonitor* monitor = iter.next(); 1233 function(monitor); 1234 } 1235 } 1236 1237 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 1238 // returns true. 1239 template <typename OwnerFilter> 1240 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 1241 monitors_iterate([&](ObjectMonitor* monitor) { 1242 // This function is only called at a safepoint or when the 1243 // target thread is suspended or when the target thread is 1244 // operating on itself. The current closures in use today are 1245 // only interested in an owned ObjectMonitor and ownership 1246 // cannot be dropped under the calling contexts so the 1247 // ObjectMonitor cannot be async deflated. 1248 if (monitor->has_owner() && filter(monitor->owner_raw())) { 1249 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 1250 1251 closure->do_monitor(monitor); 1252 } 1253 }); 1254 } 1255 1256 // Iterate ObjectMonitors where the owner == thread; this does NOT include 1257 // ObjectMonitors where owner is set to a stack-lock address in thread. 1258 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 1259 auto thread_filter = [&](void* owner) { return owner == thread; }; 1260 return owned_monitors_iterate_filtered(closure, thread_filter); 1261 } 1262 1263 // Iterate ObjectMonitors owned by any thread. 1264 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 1265 auto all_filter = [&](void* owner) { return true; }; 1266 return owned_monitors_iterate_filtered(closure, all_filter); 1267 } 1268 1269 static bool monitors_used_above_threshold(MonitorList* list) { 1270 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 1271 return false; 1272 } 1273 // Start with ceiling based on a per-thread estimate: 1274 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling(); 1275 size_t old_ceiling = ceiling; 1276 if (ceiling < list->max()) { 1277 // The max used by the system has exceeded the ceiling so use that: 1278 ceiling = list->max(); 1279 } 1280 size_t monitors_used = list->count(); 1281 if (monitors_used == 0) { // empty list is easy 1282 return false; 1283 } 1284 if (NoAsyncDeflationProgressMax != 0 && 1285 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 1286 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 1287 size_t new_ceiling = ceiling + (size_t)((double)ceiling * remainder) + 1; 1288 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 1289 log_info(monitorinflation)("Too many deflations without progress; " 1290 "bumping in_use_list_ceiling from " SIZE_FORMAT 1291 " to " SIZE_FORMAT, old_ceiling, new_ceiling); 1292 _no_progress_cnt = 0; 1293 ceiling = new_ceiling; 1294 } 1295 1296 // Check if our monitor usage is above the threshold: 1297 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 1298 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 1299 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT 1300 ", monitor_usage=" SIZE_FORMAT ", threshold=%d", 1301 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 1302 return true; 1303 } 1304 1305 return false; 1306 } 1307 1308 size_t ObjectSynchronizer::in_use_list_ceiling() { 1309 return _in_use_list_ceiling; 1310 } 1311 1312 void ObjectSynchronizer::dec_in_use_list_ceiling() { 1313 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1314 } 1315 1316 void ObjectSynchronizer::inc_in_use_list_ceiling() { 1317 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 1318 } 1319 1320 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 1321 _in_use_list_ceiling = new_value; 1322 } 1323 1324 bool ObjectSynchronizer::is_async_deflation_needed() { 1325 if (is_async_deflation_requested()) { 1326 // Async deflation request. 1327 log_info(monitorinflation)("Async deflation needed: explicit request"); 1328 return true; 1329 } 1330 1331 jlong time_since_last = time_since_last_async_deflation_ms(); 1332 1333 if (AsyncDeflationInterval > 0 && 1334 time_since_last > AsyncDeflationInterval && 1335 monitors_used_above_threshold(&_in_use_list)) { 1336 // It's been longer than our specified deflate interval and there 1337 // are too many monitors in use. We don't deflate more frequently 1338 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1339 // in order to not swamp the MonitorDeflationThread. 1340 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 1341 return true; 1342 } 1343 1344 if (GuaranteedAsyncDeflationInterval > 0 && 1345 time_since_last > GuaranteedAsyncDeflationInterval) { 1346 // It's been longer than our specified guaranteed deflate interval. 1347 // We need to clean up the used monitors even if the threshold is 1348 // not reached, to keep the memory utilization at bay when many threads 1349 // touched many monitors. 1350 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) " 1351 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 1352 GuaranteedAsyncDeflationInterval, time_since_last); 1353 1354 // If this deflation has no progress, then it should not affect the no-progress 1355 // tracking, otherwise threshold heuristics would think it was triggered, experienced 1356 // no progress, and needs to backoff more aggressively. In this "no progress" case, 1357 // the generic code would bump the no-progress counter, and we compensate for that 1358 // by telling it to skip the update. 1359 // 1360 // If this deflation has progress, then it should let non-progress tracking 1361 // know about this, otherwise the threshold heuristics would kick in, potentially 1362 // experience no-progress due to aggressive cleanup by this deflation, and think 1363 // it is still in no-progress stride. In this "progress" case, the generic code would 1364 // zero the counter, and we allow it to happen. 1365 _no_progress_skip_increment = true; 1366 1367 return true; 1368 } 1369 1370 return false; 1371 } 1372 1373 void ObjectSynchronizer::request_deflate_idle_monitors() { 1374 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 1375 set_is_async_deflation_requested(true); 1376 ml.notify_all(); 1377 } 1378 1379 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 1380 JavaThread* current = JavaThread::current(); 1381 bool ret_code = false; 1382 1383 jlong last_time = last_async_deflation_time_ns(); 1384 1385 request_deflate_idle_monitors(); 1386 1387 const int N_CHECKS = 5; 1388 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1389 if (last_async_deflation_time_ns() > last_time) { 1390 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1391 ret_code = true; 1392 break; 1393 } 1394 { 1395 // JavaThread has to honor the blocking protocol. 1396 ThreadBlockInVM tbivm(current); 1397 os::naked_short_sleep(999); // sleep for almost 1 second 1398 } 1399 } 1400 if (!ret_code) { 1401 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1402 } 1403 1404 return ret_code; 1405 } 1406 1407 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1408 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1409 } 1410 1411 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1412 const oop obj, 1413 ObjectSynchronizer::InflateCause cause) { 1414 assert(event != nullptr, "invariant"); 1415 event->set_monitorClass(obj->klass()); 1416 event->set_address((uintptr_t)(void*)obj); 1417 event->set_cause((u1)cause); 1418 event->commit(); 1419 } 1420 1421 // Fast path code shared by multiple functions 1422 void ObjectSynchronizer::inflate_helper(oop obj) { 1423 markWord mark = obj->mark_acquire(); 1424 if (mark.has_monitor()) { 1425 ObjectMonitor* monitor = mark.monitor(); 1426 markWord dmw = monitor->header(); 1427 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1428 return; 1429 } 1430 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal); 1431 } 1432 1433 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) { 1434 assert(current == Thread::current(), "must be"); 1435 if (LockingMode == LM_LIGHTWEIGHT && current->is_Java_thread()) { 1436 return inflate_impl(JavaThread::cast(current), obj, cause); 1437 } 1438 return inflate_impl(nullptr, obj, cause); 1439 } 1440 1441 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) { 1442 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be"); 1443 return inflate_impl(thread, obj, cause); 1444 } 1445 1446 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* inflating_thread, oop object, const InflateCause cause) { 1447 if (EnableValhalla) { 1448 guarantee(!object->klass()->is_inline_klass(), "Attempt to inflate inline type"); 1449 } 1450 // The JavaThread* inflating_thread parameter is only used by LM_LIGHTWEIGHT and requires 1451 // that the inflating_thread == Thread::current() or is suspended throughout the call by 1452 // some other mechanism. 1453 // Even with LM_LIGHTWEIGHT the thread might be nullptr when called from a non 1454 // JavaThread. (As may still be the case from FastHashCode). However it is only 1455 // important for the correctness of the LM_LIGHTWEIGHT algorithm that the thread 1456 // is set when called from ObjectSynchronizer::enter from the owning thread, 1457 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit. 1458 EventJavaMonitorInflate event; 1459 1460 for (;;) { 1461 const markWord mark = object->mark_acquire(); 1462 1463 // The mark can be in one of the following states: 1464 // * inflated - Just return if using stack-locking. 1465 // If using fast-locking and the ObjectMonitor owner 1466 // is anonymous and the inflating_thread owns the 1467 // object lock, then we make the inflating_thread 1468 // the ObjectMonitor owner and remove the lock from 1469 // the inflating_thread's lock stack. 1470 // * fast-locked - Coerce it to inflated from fast-locked. 1471 // * stack-locked - Coerce it to inflated from stack-locked. 1472 // * INFLATING - Busy wait for conversion from stack-locked to 1473 // inflated. 1474 // * neutral - Aggressively inflate the object. 1475 1476 // CASE: inflated 1477 if (mark.has_monitor()) { 1478 ObjectMonitor* inf = mark.monitor(); 1479 markWord dmw = inf->header(); 1480 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1481 if (LockingMode == LM_LIGHTWEIGHT && inf->is_owner_anonymous() && 1482 inflating_thread != nullptr && inflating_thread->lock_stack().contains(object)) { 1483 inf->set_owner_from_anonymous(inflating_thread); 1484 size_t removed = inflating_thread->lock_stack().remove(object); 1485 inf->set_recursions(removed - 1); 1486 } 1487 return inf; 1488 } 1489 1490 if (LockingMode != LM_LIGHTWEIGHT) { 1491 // New lightweight locking does not use INFLATING. 1492 // CASE: inflation in progress - inflating over a stack-lock. 1493 // Some other thread is converting from stack-locked to inflated. 1494 // Only that thread can complete inflation -- other threads must wait. 1495 // The INFLATING value is transient. 1496 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1497 // We could always eliminate polling by parking the thread on some auxiliary list. 1498 if (mark == markWord::INFLATING()) { 1499 read_stable_mark(object); 1500 continue; 1501 } 1502 } 1503 1504 // CASE: fast-locked 1505 // Could be fast-locked either by the inflating_thread or by some other thread. 1506 // 1507 // Note that we allocate the ObjectMonitor speculatively, _before_ 1508 // attempting to set the object's mark to the new ObjectMonitor. If 1509 // the inflating_thread owns the monitor, then we set the ObjectMonitor's 1510 // owner to the inflating_thread. Otherwise, we set the ObjectMonitor's owner 1511 // to anonymous. If we lose the race to set the object's mark to the 1512 // new ObjectMonitor, then we just delete it and loop around again. 1513 // 1514 LogStreamHandle(Trace, monitorinflation) lsh; 1515 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) { 1516 ObjectMonitor* monitor = new ObjectMonitor(object); 1517 monitor->set_header(mark.set_unlocked()); 1518 bool own = inflating_thread != nullptr && inflating_thread->lock_stack().contains(object); 1519 if (own) { 1520 // Owned by inflating_thread. 1521 monitor->set_owner_from(nullptr, inflating_thread); 1522 } else { 1523 // Owned by somebody else. 1524 monitor->set_owner_anonymous(); 1525 } 1526 markWord monitor_mark = markWord::encode(monitor); 1527 markWord old_mark = object->cas_set_mark(monitor_mark, mark); 1528 if (old_mark == mark) { 1529 // Success! Return inflated monitor. 1530 if (own) { 1531 size_t removed = inflating_thread->lock_stack().remove(object); 1532 monitor->set_recursions(removed - 1); 1533 } 1534 // Once the ObjectMonitor is configured and object is associated 1535 // with the ObjectMonitor, it is safe to allow async deflation: 1536 _in_use_list.add(monitor); 1537 1538 // Hopefully the performance counters are allocated on distinct 1539 // cache lines to avoid false sharing on MP systems ... 1540 OM_PERFDATA_OP(Inflations, inc()); 1541 if (log_is_enabled(Trace, monitorinflation)) { 1542 ResourceMark rm; 1543 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1544 INTPTR_FORMAT ", type='%s'", p2i(object), 1545 object->mark().value(), object->klass()->external_name()); 1546 } 1547 if (event.should_commit()) { 1548 post_monitor_inflate_event(&event, object, cause); 1549 } 1550 return monitor; 1551 } else { 1552 delete monitor; 1553 continue; // Interference -- just retry 1554 } 1555 } 1556 1557 // CASE: stack-locked 1558 // Could be stack-locked either by current or by some other thread. 1559 // 1560 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting 1561 // to install INFLATING into the mark word. We originally installed INFLATING, 1562 // allocated the ObjectMonitor, and then finally STed the address of the 1563 // ObjectMonitor into the mark. This was correct, but artificially lengthened 1564 // the interval in which INFLATING appeared in the mark, thus increasing 1565 // the odds of inflation contention. If we lose the race to set INFLATING, 1566 // then we just delete the ObjectMonitor and loop around again. 1567 // 1568 if (LockingMode == LM_LEGACY && mark.has_locker()) { 1569 assert(LockingMode != LM_LIGHTWEIGHT, "cannot happen with new lightweight locking"); 1570 ObjectMonitor* m = new ObjectMonitor(object); 1571 // Optimistically prepare the ObjectMonitor - anticipate successful CAS 1572 // We do this before the CAS in order to minimize the length of time 1573 // in which INFLATING appears in the mark. 1574 1575 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1576 if (cmp != mark) { 1577 delete m; 1578 continue; // Interference -- just retry 1579 } 1580 1581 // We've successfully installed INFLATING (0) into the mark-word. 1582 // This is the only case where 0 will appear in a mark-word. 1583 // Only the singular thread that successfully swings the mark-word 1584 // to 0 can perform (or more precisely, complete) inflation. 1585 // 1586 // Why do we CAS a 0 into the mark-word instead of just CASing the 1587 // mark-word from the stack-locked value directly to the new inflated state? 1588 // Consider what happens when a thread unlocks a stack-locked object. 1589 // It attempts to use CAS to swing the displaced header value from the 1590 // on-stack BasicLock back into the object header. Recall also that the 1591 // header value (hash code, etc) can reside in (a) the object header, or 1592 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1593 // header in an ObjectMonitor. The inflate() routine must copy the header 1594 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1595 // the while preserving the hashCode stability invariants. If the owner 1596 // decides to release the lock while the value is 0, the unlock will fail 1597 // and control will eventually pass from slow_exit() to inflate. The owner 1598 // will then spin, waiting for the 0 value to disappear. Put another way, 1599 // the 0 causes the owner to stall if the owner happens to try to 1600 // drop the lock (restoring the header from the BasicLock to the object) 1601 // while inflation is in-progress. This protocol avoids races that might 1602 // would otherwise permit hashCode values to change or "flicker" for an object. 1603 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1604 // 0 serves as a "BUSY" inflate-in-progress indicator. 1605 1606 1607 // fetch the displaced mark from the owner's stack. 1608 // The owner can't die or unwind past the lock while our INFLATING 1609 // object is in the mark. Furthermore the owner can't complete 1610 // an unlock on the object, either. 1611 markWord dmw = mark.displaced_mark_helper(); 1612 // Catch if the object's header is not neutral (not locked and 1613 // not marked is what we care about here). 1614 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1615 1616 // Setup monitor fields to proper values -- prepare the monitor 1617 m->set_header(dmw); 1618 1619 // Optimization: if the mark.locker stack address is associated 1620 // with this thread we could simply set m->_owner = current. 1621 // Note that a thread can inflate an object 1622 // that it has stack-locked -- as might happen in wait() -- directly 1623 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom. 1624 m->set_owner_from(nullptr, mark.locker()); 1625 // TODO-FIXME: assert BasicLock->dhw != 0. 1626 1627 // Must preserve store ordering. The monitor state must 1628 // be stable at the time of publishing the monitor address. 1629 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1630 // Release semantics so that above set_object() is seen first. 1631 object->release_set_mark(markWord::encode(m)); 1632 1633 // Once ObjectMonitor is configured and the object is associated 1634 // with the ObjectMonitor, it is safe to allow async deflation: 1635 _in_use_list.add(m); 1636 1637 // Hopefully the performance counters are allocated on distinct cache lines 1638 // to avoid false sharing on MP systems ... 1639 OM_PERFDATA_OP(Inflations, inc()); 1640 if (log_is_enabled(Trace, monitorinflation)) { 1641 ResourceMark rm; 1642 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1643 INTPTR_FORMAT ", type='%s'", p2i(object), 1644 object->mark().value(), object->klass()->external_name()); 1645 } 1646 if (event.should_commit()) { 1647 post_monitor_inflate_event(&event, object, cause); 1648 } 1649 return m; 1650 } 1651 1652 // CASE: neutral 1653 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1654 // If we know we're inflating for entry it's better to inflate by swinging a 1655 // pre-locked ObjectMonitor pointer into the object header. A successful 1656 // CAS inflates the object *and* confers ownership to the inflating thread. 1657 // In the current implementation we use a 2-step mechanism where we CAS() 1658 // to inflate and then CAS() again to try to swing _owner from null to current. 1659 // An inflateTry() method that we could call from enter() would be useful. 1660 1661 // Catch if the object's header is not neutral (not locked and 1662 // not marked is what we care about here). 1663 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1664 ObjectMonitor* m = new ObjectMonitor(object); 1665 // prepare m for installation - set monitor to initial state 1666 m->set_header(mark); 1667 1668 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1669 delete m; 1670 m = nullptr; 1671 continue; 1672 // interference - the markword changed - just retry. 1673 // The state-transitions are one-way, so there's no chance of 1674 // live-lock -- "Inflated" is an absorbing state. 1675 } 1676 1677 // Once the ObjectMonitor is configured and object is associated 1678 // with the ObjectMonitor, it is safe to allow async deflation: 1679 _in_use_list.add(m); 1680 1681 // Hopefully the performance counters are allocated on distinct 1682 // cache lines to avoid false sharing on MP systems ... 1683 OM_PERFDATA_OP(Inflations, inc()); 1684 if (log_is_enabled(Trace, monitorinflation)) { 1685 ResourceMark rm; 1686 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 1687 INTPTR_FORMAT ", type='%s'", p2i(object), 1688 object->mark().value(), object->klass()->external_name()); 1689 } 1690 if (event.should_commit()) { 1691 post_monitor_inflate_event(&event, object, cause); 1692 } 1693 return m; 1694 } 1695 } 1696 1697 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1698 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1699 // 1700 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1701 MonitorList::Iterator iter = _in_use_list.iterator(); 1702 size_t deflated_count = 0; 1703 1704 while (iter.has_next()) { 1705 if (deflated_count >= (size_t)MonitorDeflationMax) { 1706 break; 1707 } 1708 ObjectMonitor* mid = iter.next(); 1709 if (mid->deflate_monitor()) { 1710 deflated_count++; 1711 } 1712 1713 // Must check for a safepoint/handshake and honor it. 1714 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1715 } 1716 1717 return deflated_count; 1718 } 1719 1720 class HandshakeForDeflation : public HandshakeClosure { 1721 public: 1722 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {} 1723 1724 void do_thread(Thread* thread) { 1725 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread=" 1726 INTPTR_FORMAT, p2i(thread)); 1727 } 1728 }; 1729 1730 class VM_RendezvousGCThreads : public VM_Operation { 1731 public: 1732 bool evaluate_at_safepoint() const override { return false; } 1733 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1734 void doit() override { 1735 Universe::heap()->safepoint_synchronize_begin(); 1736 Universe::heap()->safepoint_synchronize_end(); 1737 }; 1738 }; 1739 1740 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1741 ObjectMonitorDeflationSafepointer* safepointer) { 1742 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1743 size_t deleted_count = 0; 1744 for (ObjectMonitor* monitor: *delete_list) { 1745 delete monitor; 1746 deleted_count++; 1747 // A JavaThread must check for a safepoint/handshake and honor it. 1748 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1749 } 1750 return deleted_count; 1751 } 1752 1753 class ObjectMonitorDeflationLogging: public StackObj { 1754 LogStreamHandle(Debug, monitorinflation) _debug; 1755 LogStreamHandle(Info, monitorinflation) _info; 1756 LogStream* _stream; 1757 elapsedTimer _timer; 1758 1759 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1760 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); } 1761 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); } 1762 1763 public: 1764 ObjectMonitorDeflationLogging() 1765 : _debug(), _info(), _stream(nullptr) { 1766 if (_debug.is_enabled()) { 1767 _stream = &_debug; 1768 } else if (_info.is_enabled()) { 1769 _stream = &_info; 1770 } 1771 } 1772 1773 void begin() { 1774 if (_stream != nullptr) { 1775 _stream->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1776 ceiling(), count(), max()); 1777 _timer.start(); 1778 } 1779 } 1780 1781 void before_handshake(size_t unlinked_count) { 1782 if (_stream != nullptr) { 1783 _timer.stop(); 1784 _stream->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT 1785 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count=" 1786 SIZE_FORMAT ", max=" SIZE_FORMAT, 1787 unlinked_count, ceiling(), count(), max()); 1788 } 1789 } 1790 1791 void after_handshake() { 1792 if (_stream != nullptr) { 1793 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1794 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1795 ceiling(), count(), max()); 1796 _timer.start(); 1797 } 1798 } 1799 1800 void end(size_t deflated_count, size_t unlinked_count) { 1801 if (_stream != nullptr) { 1802 _timer.stop(); 1803 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1804 _stream->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs", 1805 deflated_count, unlinked_count, _timer.seconds()); 1806 } 1807 _stream->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1808 ceiling(), count(), max()); 1809 } 1810 } 1811 1812 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1813 if (_stream != nullptr) { 1814 _timer.stop(); 1815 _stream->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling=" 1816 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, 1817 op_name, cnt_name, cnt, ceiling(), count(), max()); 1818 } 1819 } 1820 1821 void after_block_for_safepoint(const char* op_name) { 1822 if (_stream != nullptr) { 1823 _stream->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT 1824 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name, 1825 ceiling(), count(), max()); 1826 _timer.start(); 1827 } 1828 } 1829 }; 1830 1831 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1832 if (!SafepointMechanism::should_process(_current)) { 1833 return; 1834 } 1835 1836 // A safepoint/handshake has started. 1837 _log->before_block_for_safepoint(op_name, count_name, counter); 1838 1839 { 1840 // Honor block request. 1841 ThreadBlockInVM tbivm(_current); 1842 } 1843 1844 _log->after_block_for_safepoint(op_name); 1845 } 1846 1847 // This function is called by the MonitorDeflationThread to deflate 1848 // ObjectMonitors. 1849 size_t ObjectSynchronizer::deflate_idle_monitors() { 1850 JavaThread* current = JavaThread::current(); 1851 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1852 1853 // The async deflation request has been processed. 1854 _last_async_deflation_time_ns = os::javaTimeNanos(); 1855 set_is_async_deflation_requested(false); 1856 1857 ObjectMonitorDeflationLogging log; 1858 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1859 1860 log.begin(); 1861 1862 // Deflate some idle ObjectMonitors. 1863 size_t deflated_count = deflate_monitor_list(&safepointer); 1864 1865 // Unlink the deflated ObjectMonitors from the in-use list. 1866 size_t unlinked_count = 0; 1867 size_t deleted_count = 0; 1868 if (deflated_count > 0) { 1869 ResourceMark rm(current); 1870 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1871 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1872 1873 log.before_handshake(unlinked_count); 1874 1875 // A JavaThread needs to handshake in order to safely free the 1876 // ObjectMonitors that were deflated in this cycle. 1877 HandshakeForDeflation hfd_hc; 1878 Handshake::execute(&hfd_hc); 1879 // Also, we sync and desync GC threads around the handshake, so that they can 1880 // safely read the mark-word and look-through to the object-monitor, without 1881 // being afraid that the object-monitor is going away. 1882 VM_RendezvousGCThreads sync_gc; 1883 VMThread::execute(&sync_gc); 1884 1885 log.after_handshake(); 1886 1887 // After the handshake, safely free the ObjectMonitors that were 1888 // deflated and unlinked in this cycle. 1889 1890 // Delete the unlinked ObjectMonitors. 1891 deleted_count = delete_monitors(&delete_list, &safepointer); 1892 assert(unlinked_count == deleted_count, "must be"); 1893 } 1894 1895 log.end(deflated_count, unlinked_count); 1896 1897 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count())); 1898 OM_PERFDATA_OP(Deflations, inc(deflated_count)); 1899 1900 GVars.stw_random = os::random(); 1901 1902 if (deflated_count != 0) { 1903 _no_progress_cnt = 0; 1904 } else if (_no_progress_skip_increment) { 1905 _no_progress_skip_increment = false; 1906 } else { 1907 _no_progress_cnt++; 1908 } 1909 1910 return deflated_count; 1911 } 1912 1913 // Monitor cleanup on JavaThread::exit 1914 1915 // Iterate through monitor cache and attempt to release thread's monitors 1916 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1917 private: 1918 JavaThread* _thread; 1919 1920 public: 1921 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1922 void do_monitor(ObjectMonitor* mid) { 1923 intx rec = mid->complete_exit(_thread); 1924 _thread->dec_held_monitor_count(rec + 1); 1925 } 1926 }; 1927 1928 // Release all inflated monitors owned by current thread. Lightweight monitors are 1929 // ignored. This is meant to be called during JNI thread detach which assumes 1930 // all remaining monitors are heavyweight. All exceptions are swallowed. 1931 // Scanning the extant monitor list can be time consuming. 1932 // A simple optimization is to add a per-thread flag that indicates a thread 1933 // called jni_monitorenter() during its lifetime. 1934 // 1935 // Instead of NoSafepointVerifier it might be cheaper to 1936 // use an idiom of the form: 1937 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1938 // <code that must not run at safepoint> 1939 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1940 // Since the tests are extremely cheap we could leave them enabled 1941 // for normal product builds. 1942 1943 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1944 assert(current == JavaThread::current(), "must be current Java thread"); 1945 NoSafepointVerifier nsv; 1946 ReleaseJavaMonitorsClosure rjmc(current); 1947 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1948 assert(!current->has_pending_exception(), "Should not be possible"); 1949 current->clear_pending_exception(); 1950 assert(current->held_monitor_count() == 0, "Should not be possible"); 1951 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1952 current->clear_jni_monitor_count(); 1953 } 1954 1955 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1956 switch (cause) { 1957 case inflate_cause_vm_internal: return "VM Internal"; 1958 case inflate_cause_monitor_enter: return "Monitor Enter"; 1959 case inflate_cause_wait: return "Monitor Wait"; 1960 case inflate_cause_notify: return "Monitor Notify"; 1961 case inflate_cause_hash_code: return "Monitor Hash Code"; 1962 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1963 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1964 default: 1965 ShouldNotReachHere(); 1966 } 1967 return "Unknown"; 1968 } 1969 1970 //------------------------------------------------------------------------------ 1971 // Debugging code 1972 1973 u_char* ObjectSynchronizer::get_gvars_addr() { 1974 return (u_char*)&GVars; 1975 } 1976 1977 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1978 return (u_char*)&GVars.hc_sequence; 1979 } 1980 1981 size_t ObjectSynchronizer::get_gvars_size() { 1982 return sizeof(SharedGlobals); 1983 } 1984 1985 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1986 return (u_char*)&GVars.stw_random; 1987 } 1988 1989 // Do the final audit and print of ObjectMonitor stats; must be done 1990 // by the VMThread at VM exit time. 1991 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1992 assert(Thread::current()->is_VM_thread(), "sanity check"); 1993 1994 if (is_final_audit()) { // Only do the audit once. 1995 return; 1996 } 1997 set_is_final_audit(); 1998 log_info(monitorinflation)("Starting the final audit."); 1999 2000 if (log_is_enabled(Info, monitorinflation)) { 2001 // The other audit_and_print_stats() call is done at the Debug 2002 // level at a safepoint in SafepointSynchronize::do_cleanup_tasks. 2003 audit_and_print_stats(true /* on_exit */); 2004 } 2005 } 2006 2007 // This function can be called at a safepoint or it can be called when 2008 // we are trying to exit the VM. When we are trying to exit the VM, the 2009 // list walker functions can run in parallel with the other list 2010 // operations so spin-locking is used for safety. 2011 // 2012 // Calls to this function can be added in various places as a debugging 2013 // aid; pass 'true' for the 'on_exit' parameter to have in-use monitor 2014 // details logged at the Info level and 'false' for the 'on_exit' 2015 // parameter to have in-use monitor details logged at the Trace level. 2016 // 2017 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) { 2018 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant"); 2019 2020 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2021 LogStreamHandle(Info, monitorinflation) lsh_info; 2022 LogStreamHandle(Trace, monitorinflation) lsh_trace; 2023 LogStream* ls = nullptr; 2024 if (log_is_enabled(Trace, monitorinflation)) { 2025 ls = &lsh_trace; 2026 } else if (log_is_enabled(Debug, monitorinflation)) { 2027 ls = &lsh_debug; 2028 } else if (log_is_enabled(Info, monitorinflation)) { 2029 ls = &lsh_info; 2030 } 2031 assert(ls != nullptr, "sanity check"); 2032 2033 int error_cnt = 0; 2034 2035 ls->print_cr("Checking in_use_list:"); 2036 chk_in_use_list(ls, &error_cnt); 2037 2038 if (error_cnt == 0) { 2039 ls->print_cr("No errors found in in_use_list checks."); 2040 } else { 2041 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 2042 } 2043 2044 if ((on_exit && log_is_enabled(Info, monitorinflation)) || 2045 (!on_exit && log_is_enabled(Trace, monitorinflation))) { 2046 // When exiting this log output is at the Info level. When called 2047 // at a safepoint, this log output is at the Trace level since 2048 // there can be a lot of it. 2049 log_in_use_monitor_details(ls, !on_exit /* log_all */); 2050 } 2051 2052 ls->flush(); 2053 2054 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2055 } 2056 2057 // Check the in_use_list; log the results of the checks. 2058 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 2059 size_t l_in_use_count = _in_use_list.count(); 2060 size_t l_in_use_max = _in_use_list.max(); 2061 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count, 2062 l_in_use_max); 2063 2064 size_t ck_in_use_count = 0; 2065 MonitorList::Iterator iter = _in_use_list.iterator(); 2066 while (iter.has_next()) { 2067 ObjectMonitor* mid = iter.next(); 2068 chk_in_use_entry(mid, out, error_cnt_p); 2069 ck_in_use_count++; 2070 } 2071 2072 if (l_in_use_count == ck_in_use_count) { 2073 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count=" 2074 SIZE_FORMAT, l_in_use_count, ck_in_use_count); 2075 } else { 2076 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to " 2077 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count, 2078 ck_in_use_count); 2079 } 2080 2081 size_t ck_in_use_max = _in_use_list.max(); 2082 if (l_in_use_max == ck_in_use_max) { 2083 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max=" 2084 SIZE_FORMAT, l_in_use_max, ck_in_use_max); 2085 } else { 2086 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to " 2087 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max); 2088 } 2089 } 2090 2091 // Check an in-use monitor entry; log any errors. 2092 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 2093 int* error_cnt_p) { 2094 if (n->owner_is_DEFLATER_MARKER()) { 2095 // This could happen when monitor deflation blocks for a safepoint. 2096 return; 2097 } 2098 2099 if (n->header().value() == 0) { 2100 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 2101 "have non-null _header field.", p2i(n)); 2102 *error_cnt_p = *error_cnt_p + 1; 2103 } 2104 const oop obj = n->object_peek(); 2105 if (obj != nullptr) { 2106 const markWord mark = obj->mark(); 2107 if (!mark.has_monitor()) { 2108 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2109 "object does not think it has a monitor: obj=" 2110 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2111 p2i(obj), mark.value()); 2112 *error_cnt_p = *error_cnt_p + 1; 2113 } 2114 ObjectMonitor* const obj_mon = mark.monitor(); 2115 if (n != obj_mon) { 2116 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 2117 "object does not refer to the same monitor: obj=" 2118 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2119 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2120 *error_cnt_p = *error_cnt_p + 1; 2121 } 2122 } 2123 } 2124 2125 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 2126 // flags indicate why the entry is in-use, 'object' and 'object type' 2127 // indicate the associated object and its type. 2128 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 2129 if (_in_use_list.count() > 0) { 2130 stringStream ss; 2131 out->print_cr("In-use monitor info:"); 2132 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2133 out->print_cr("%18s %s %18s %18s", 2134 "monitor", "BHL", "object", "object type"); 2135 out->print_cr("================== === ================== =================="); 2136 2137 auto is_interesting = [&](ObjectMonitor* monitor) { 2138 return log_all || monitor->has_owner() || monitor->is_busy(); 2139 }; 2140 2141 monitors_iterate([&](ObjectMonitor* monitor) { 2142 if (is_interesting(monitor)) { 2143 const oop obj = monitor->object_peek(); 2144 const markWord mark = monitor->header(); 2145 ResourceMark rm; 2146 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 2147 monitor->is_busy(), mark.hash() != 0, monitor->owner() != nullptr, 2148 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 2149 if (monitor->is_busy()) { 2150 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 2151 ss.reset(); 2152 } 2153 out->cr(); 2154 } 2155 }); 2156 } 2157 2158 out->flush(); 2159 }