1 /* 2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "classfile/vmSymbols.hpp" 26 #include "gc/shared/collectedHeap.hpp" 27 #include "jfr/jfrEvents.hpp" 28 #include "logging/log.hpp" 29 #include "logging/logStream.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/padded.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "memory/universe.hpp" 34 #include "oops/markWord.hpp" 35 #include "oops/oop.inline.hpp" 36 #include "runtime/atomicAccess.hpp" 37 #include "runtime/basicLock.inline.hpp" 38 #include "runtime/frame.inline.hpp" 39 #include "runtime/globals.hpp" 40 #include "runtime/handles.inline.hpp" 41 #include "runtime/handshake.hpp" 42 #include "runtime/interfaceSupport.inline.hpp" 43 #include "runtime/javaThread.hpp" 44 #include "runtime/lightweightSynchronizer.hpp" 45 #include "runtime/lockStack.inline.hpp" 46 #include "runtime/mutexLocker.hpp" 47 #include "runtime/objectMonitor.inline.hpp" 48 #include "runtime/os.inline.hpp" 49 #include "runtime/osThread.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.inline.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/globalCounter.inline.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 = AtomicAccess::load(&_head); 74 m->set_next_om(head); 75 } while (AtomicAccess::cmpxchg(&_head, head, m) != head); 76 77 size_t count = AtomicAccess::add(&_count, 1u, memory_order_relaxed); 78 size_t old_max; 79 do { 80 old_max = AtomicAccess::load(&_max); 81 if (count <= old_max) { 82 break; 83 } 84 } while (AtomicAccess::cmpxchg(&_max, old_max, count, memory_order_relaxed) != old_max); 85 } 86 87 size_t MonitorList::count() const { 88 return AtomicAccess::load(&_count); 89 } 90 91 size_t MonitorList::max() const { 92 return AtomicAccess::load(&_max); 93 } 94 95 class ObjectMonitorDeflationSafepointer : public StackObj { 96 JavaThread* const _current; 97 ObjectMonitorDeflationLogging* const _log; 98 99 public: 100 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log) 101 : _current(current), _log(log) {} 102 103 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter); 104 }; 105 106 // Walk the in-use list and unlink deflated ObjectMonitors. 107 // Returns the number of unlinked ObjectMonitors. 108 size_t MonitorList::unlink_deflated(size_t deflated_count, 109 GrowableArray<ObjectMonitor*>* unlinked_list, 110 ObjectMonitorDeflationSafepointer* safepointer) { 111 size_t unlinked_count = 0; 112 ObjectMonitor* prev = nullptr; 113 ObjectMonitor* m = AtomicAccess::load_acquire(&_head); 114 115 while (m != nullptr) { 116 if (m->is_being_async_deflated()) { 117 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can 118 // modify the list once per batch. The batch starts at "m". 119 size_t unlinked_batch = 0; 120 ObjectMonitor* next = m; 121 // Look for at most MonitorUnlinkBatch monitors, or the number of 122 // deflated and not unlinked monitors, whatever comes first. 123 assert(deflated_count >= unlinked_count, "Sanity: underflow"); 124 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch); 125 do { 126 ObjectMonitor* next_next = next->next_om(); 127 unlinked_batch++; 128 unlinked_list->append(next); 129 next = next_next; 130 if (unlinked_batch >= unlinked_batch_limit) { 131 // Reached the max batch, so bail out of the gathering loop. 132 break; 133 } 134 if (prev == nullptr && AtomicAccess::load(&_head) != m) { 135 // Current batch used to be at head, but it is not at head anymore. 136 // Bail out and figure out where we currently are. This avoids long 137 // walks searching for new prev during unlink under heavy list inserts. 138 break; 139 } 140 } while (next != nullptr && next->is_being_async_deflated()); 141 142 // Unlink the found batch. 143 if (prev == nullptr) { 144 // The current batch is the first batch, so there is a chance that it starts at head. 145 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head. 146 ObjectMonitor* prev_head = AtomicAccess::cmpxchg(&_head, m, next); 147 if (prev_head != m) { 148 // Something must have updated the head. Figure out the actual prev for this batch. 149 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) { 150 prev = n; 151 } 152 assert(prev != nullptr, "Should have found the prev for the current batch"); 153 prev->set_next_om(next); 154 } 155 } else { 156 // The current batch is preceded by another batch. This guarantees the current batch 157 // does not start at head. Unlink the entire current batch without updating the head. 158 assert(AtomicAccess::load(&_head) != m, "Sanity"); 159 prev->set_next_om(next); 160 } 161 162 unlinked_count += unlinked_batch; 163 if (unlinked_count >= deflated_count) { 164 // Reached the max so bail out of the searching loop. 165 // There should be no more deflated monitors left. 166 break; 167 } 168 m = next; 169 } else { 170 prev = m; 171 m = m->next_om(); 172 } 173 174 // Must check for a safepoint/handshake and honor it. 175 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count); 176 } 177 178 #ifdef ASSERT 179 // Invariant: the code above should unlink all deflated monitors. 180 // The code that runs after this unlinking does not expect deflated monitors. 181 // Notably, attempting to deflate the already deflated monitor would break. 182 { 183 ObjectMonitor* m = AtomicAccess::load_acquire(&_head); 184 while (m != nullptr) { 185 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked"); 186 m = m->next_om(); 187 } 188 } 189 #endif 190 191 AtomicAccess::sub(&_count, unlinked_count); 192 return unlinked_count; 193 } 194 195 MonitorList::Iterator MonitorList::iterator() const { 196 return Iterator(AtomicAccess::load_acquire(&_head)); 197 } 198 199 ObjectMonitor* MonitorList::Iterator::next() { 200 ObjectMonitor* current = _current; 201 _current = current->next_om(); 202 return current; 203 } 204 205 // The "core" versions of monitor enter and exit reside in this file. 206 // The interpreter and compilers contain specialized transliterated 207 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp 208 // fast_lock(...) for instance. If you make changes here, make sure to modify the 209 // interpreter, and both C1 and C2 fast-path inline locking code emission. 210 // 211 // ----------------------------------------------------------------------------- 212 213 #ifdef DTRACE_ENABLED 214 215 // Only bother with this argument setup if dtrace is available 216 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 217 218 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 219 char* bytes = nullptr; \ 220 int len = 0; \ 221 jlong jtid = SharedRuntime::get_java_tid(thread); \ 222 Symbol* klassname = obj->klass()->name(); \ 223 if (klassname != nullptr) { \ 224 bytes = (char*)klassname->bytes(); \ 225 len = klassname->utf8_length(); \ 226 } 227 228 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 229 { \ 230 if (DTraceMonitorProbes) { \ 231 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 232 HOTSPOT_MONITOR_WAIT(jtid, \ 233 (uintptr_t)(monitor), bytes, len, (millis)); \ 234 } \ 235 } 236 237 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 238 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 239 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 240 241 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 242 { \ 243 if (DTraceMonitorProbes) { \ 244 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 245 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 246 (uintptr_t)(monitor), bytes, len); \ 247 } \ 248 } 249 250 #else // ndef DTRACE_ENABLED 251 252 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 253 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 254 255 #endif // ndef DTRACE_ENABLED 256 257 // This exists only as a workaround of dtrace bug 6254741 258 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) { 259 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 260 return 0; 261 } 262 263 static constexpr size_t inflation_lock_count() { 264 return 256; 265 } 266 267 // Static storage for an array of PlatformMutex. 268 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)]; 269 270 static inline PlatformMutex* inflation_lock(size_t index) { 271 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]); 272 } 273 274 void ObjectSynchronizer::initialize() { 275 for (size_t i = 0; i < inflation_lock_count(); i++) { 276 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex(); 277 } 278 // Start the ceiling with the estimate for one thread. 279 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate); 280 281 // Start the timer for deflations, so it does not trigger immediately. 282 _last_async_deflation_time_ns = os::javaTimeNanos(); 283 284 LightweightSynchronizer::initialize(); 285 } 286 287 MonitorList ObjectSynchronizer::_in_use_list; 288 // monitors_used_above_threshold() policy is as follows: 289 // 290 // The ratio of the current _in_use_list count to the ceiling is used 291 // to determine if we are above MonitorUsedDeflationThreshold and need 292 // to do an async monitor deflation cycle. The ceiling is increased by 293 // AvgMonitorsPerThreadEstimate when a thread is added to the system 294 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is 295 // removed from the system. 296 // 297 // Note: If the _in_use_list max exceeds the ceiling, then 298 // monitors_used_above_threshold() will use the in_use_list max instead 299 // of the thread count derived ceiling because we have used more 300 // ObjectMonitors than the estimated average. 301 // 302 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax 303 // no-progress async monitor deflation cycles in a row, then the ceiling 304 // is adjusted upwards by monitors_used_above_threshold(). 305 // 306 // Start the ceiling with the estimate for one thread in initialize() 307 // which is called after cmd line options are processed. 308 static size_t _in_use_list_ceiling = 0; 309 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 310 bool volatile ObjectSynchronizer::_is_final_audit = false; 311 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 312 static uintx _no_progress_cnt = 0; 313 static bool _no_progress_skip_increment = false; 314 315 // =====================> Quick functions 316 317 // The quick_* forms are special fast-path variants used to improve 318 // performance. In the simplest case, a "quick_*" implementation could 319 // simply return false, in which case the caller will perform the necessary 320 // state transitions and call the slow-path form. 321 // The fast-path is designed to handle frequently arising cases in an efficient 322 // manner and is just a degenerate "optimistic" variant of the slow-path. 323 // returns true -- to indicate the call was satisfied. 324 // returns false -- to indicate the call needs the services of the slow-path. 325 // A no-loitering ordinance is in effect for code in the quick_* family 326 // operators: safepoints or indefinite blocking (blocking that might span a 327 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 328 // entry. 329 // 330 // Consider: An interesting optimization is to have the JIT recognize the 331 // following common idiom: 332 // synchronized (someobj) { .... ; notify(); } 333 // That is, we find a notify() or notifyAll() call that immediately precedes 334 // the monitorexit operation. In that case the JIT could fuse the operations 335 // into a single notifyAndExit() runtime primitive. 336 337 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) { 338 assert(current->thread_state() == _thread_in_Java, "invariant"); 339 NoSafepointVerifier nsv; 340 if (obj == nullptr) return false; // slow-path for invalid obj 341 const markWord mark = obj->mark(); 342 343 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) { 344 // Degenerate notify 345 // fast-locked by caller so by definition the implied waitset is empty. 346 return true; 347 } 348 349 if (mark.has_monitor()) { 350 ObjectMonitor* const mon = read_monitor(current, obj, mark); 351 if (mon == nullptr) { 352 // Racing with inflation/deflation go slow path 353 return false; 354 } 355 assert(mon->object() == oop(obj), "invariant"); 356 if (!mon->has_owner(current)) return false; // slow-path for IMS exception 357 358 if (mon->first_waiter() != nullptr) { 359 // We have one or more waiters. Since this is an inflated monitor 360 // that we own, we quickly notify them here and now, avoiding the slow-path. 361 if (all) { 362 mon->quick_notifyAll(current); 363 } else { 364 mon->quick_notify(current); 365 } 366 } 367 return true; 368 } 369 370 // other IMS exception states take the slow-path 371 return false; 372 } 373 374 // Handle notifications when synchronizing on value based classes 375 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) { 376 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 377 frame last_frame = locking_thread->last_frame(); 378 bool bcp_was_adjusted = false; 379 // Don't decrement bcp if it points to the frame's first instruction. This happens when 380 // handle_sync_on_value_based_class() is called because of a synchronized method. There 381 // is no actual monitorenter instruction in the byte code in this case. 382 if (last_frame.is_interpreted_frame() && 383 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) { 384 // adjust bcp to point back to monitorenter so that we print the correct line numbers 385 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1); 386 bcp_was_adjusted = true; 387 } 388 389 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) { 390 ResourceMark rm; 391 stringStream ss; 392 locking_thread->print_active_stack_on(&ss); 393 char* base = (char*)strstr(ss.base(), "at"); 394 char* newline = (char*)strchr(ss.base(), '\n'); 395 if (newline != nullptr) { 396 *newline = '\0'; 397 } 398 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base); 399 } else { 400 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses"); 401 ResourceMark rm; 402 Log(valuebasedclasses) vblog; 403 404 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name()); 405 if (locking_thread->has_last_Java_frame()) { 406 LogStream info_stream(vblog.info()); 407 locking_thread->print_active_stack_on(&info_stream); 408 } else { 409 vblog.info("Cannot find the last Java frame"); 410 } 411 412 EventSyncOnValueBasedClass event; 413 if (event.should_commit()) { 414 event.set_valueBasedClass(obj->klass()); 415 event.commit(); 416 } 417 } 418 419 if (bcp_was_adjusted) { 420 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1); 421 } 422 } 423 424 // ----------------------------------------------------------------------------- 425 // Monitor Enter/Exit 426 427 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) { 428 // When called with locking_thread != Thread::current() some mechanism must synchronize 429 // the locking_thread with respect to the current thread. Currently only used when 430 // deoptimizing and re-locking locks. See Deoptimization::relock_objects 431 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 432 return LightweightSynchronizer::enter_for(obj, lock, locking_thread); 433 } 434 435 // ----------------------------------------------------------------------------- 436 // JNI locks on java objects 437 // NOTE: must use heavy weight monitor to handle jni monitor enter 438 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) { 439 // Top native frames in the stack will not be seen if we attempt 440 // preemption, since we start walking from the last Java anchor. 441 NoPreemptMark npm(current); 442 443 if (obj->klass()->is_value_based()) { 444 handle_sync_on_value_based_class(obj, current); 445 } 446 447 // the current locking is from JNI instead of Java code 448 current->set_current_pending_monitor_is_from_java(false); 449 // An async deflation can race after the inflate() call and before 450 // enter() can make the ObjectMonitor busy. enter() returns false if 451 // we have lost the race to async deflation and we simply try again. 452 while (true) { 453 BasicLock lock; 454 if (LightweightSynchronizer::inflate_and_enter(obj(), &lock, inflate_cause_jni_enter, current, current) != nullptr) { 455 current->inc_held_monitor_count(1, true); 456 break; 457 } 458 } 459 current->set_current_pending_monitor_is_from_java(true); 460 } 461 462 // NOTE: must use heavy weight monitor to handle jni monitor exit 463 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) { 464 JavaThread* current = THREAD; 465 466 ObjectMonitor* monitor; 467 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK); 468 // If this thread has locked the object, exit the monitor. We 469 // intentionally do not use CHECK on check_owner because we must exit the 470 // monitor even if an exception was already pending. 471 if (monitor->check_owner(THREAD)) { 472 monitor->exit(current); 473 current->dec_held_monitor_count(1, true); 474 } 475 } 476 477 // ----------------------------------------------------------------------------- 478 // Internal VM locks on java objects 479 // standard constructor, allows locking failures 480 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) { 481 _thread = thread; 482 _thread->check_for_valid_safepoint_state(); 483 _obj = obj; 484 485 if (_obj() != nullptr) { 486 ObjectSynchronizer::enter(_obj, &_lock, _thread); 487 } 488 } 489 490 ObjectLocker::~ObjectLocker() { 491 if (_obj() != nullptr) { 492 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 493 } 494 } 495 496 497 // ----------------------------------------------------------------------------- 498 // Wait/Notify/NotifyAll 499 // NOTE: must use heavy weight monitor to handle wait() 500 501 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 502 JavaThread* current = THREAD; 503 if (millis < 0) { 504 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 505 } 506 507 ObjectMonitor* monitor; 508 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0); 509 510 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis); 511 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code 512 513 // This dummy call is in place to get around dtrace bug 6254741. Once 514 // that's fixed we can uncomment the following line, remove the call 515 // and change this function back into a "void" func. 516 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 517 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 518 return ret_code; 519 } 520 521 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) { 522 if (millis < 0) { 523 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 524 } 525 526 ObjectMonitor* monitor; 527 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK); 528 monitor->wait(millis, false, THREAD); 529 } 530 531 532 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 533 JavaThread* current = THREAD; 534 535 markWord mark = obj->mark(); 536 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 537 // Not inflated so there can't be any waiters to notify. 538 return; 539 } 540 ObjectMonitor* monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 541 monitor->notify(CHECK); 542 } 543 544 // NOTE: see comment of notify() 545 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 546 JavaThread* current = THREAD; 547 548 markWord mark = obj->mark(); 549 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) { 550 // Not inflated so there can't be any waiters to notify. 551 return; 552 } 553 554 ObjectMonitor* monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK); 555 monitor->notifyAll(CHECK); 556 } 557 558 // ----------------------------------------------------------------------------- 559 // Hash Code handling 560 561 struct SharedGlobals { 562 char _pad_prefix[OM_CACHE_LINE_SIZE]; 563 // This is a highly shared mostly-read variable. 564 // To avoid false-sharing it needs to be the sole occupant of a cache line. 565 volatile int stw_random; 566 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 567 // Hot RW variable -- Sequester to avoid false-sharing 568 volatile int hc_sequence; 569 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 570 }; 571 572 static SharedGlobals GVars; 573 574 // hashCode() generation : 575 // 576 // Possibilities: 577 // * MD5Digest of {obj,stw_random} 578 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 579 // * A DES- or AES-style SBox[] mechanism 580 // * One of the Phi-based schemes, such as: 581 // 2654435761 = 2^32 * Phi (golden ratio) 582 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 583 // * A variation of Marsaglia's shift-xor RNG scheme. 584 // * (obj ^ stw_random) is appealing, but can result 585 // in undesirable regularity in the hashCode values of adjacent objects 586 // (objects allocated back-to-back, in particular). This could potentially 587 // result in hashtable collisions and reduced hashtable efficiency. 588 // There are simple ways to "diffuse" the middle address bits over the 589 // generated hashCode values: 590 591 static intptr_t get_next_hash(Thread* current, oop obj) { 592 intptr_t value = 0; 593 if (hashCode == 0) { 594 // This form uses global Park-Miller RNG. 595 // On MP system we'll have lots of RW access to a global, so the 596 // mechanism induces lots of coherency traffic. 597 value = os::random(); 598 } else if (hashCode == 1) { 599 // This variation has the property of being stable (idempotent) 600 // between STW operations. This can be useful in some of the 1-0 601 // synchronization schemes. 602 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 603 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 604 } else if (hashCode == 2) { 605 value = 1; // for sensitivity testing 606 } else if (hashCode == 3) { 607 value = ++GVars.hc_sequence; 608 } else if (hashCode == 4) { 609 value = cast_from_oop<intptr_t>(obj); 610 } else { 611 // Marsaglia's xor-shift scheme with thread-specific state 612 // This is probably the best overall implementation -- we'll 613 // likely make this the default in future releases. 614 unsigned t = current->_hashStateX; 615 t ^= (t << 11); 616 current->_hashStateX = current->_hashStateY; 617 current->_hashStateY = current->_hashStateZ; 618 current->_hashStateZ = current->_hashStateW; 619 unsigned v = current->_hashStateW; 620 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 621 current->_hashStateW = v; 622 value = v; 623 } 624 625 value &= markWord::hash_mask; 626 if (value == 0) value = 0xBAD; 627 assert(value != markWord::no_hash, "invariant"); 628 return value; 629 } 630 631 static intptr_t install_hash_code(Thread* current, oop obj) { 632 assert(UseObjectMonitorTable, "must be"); 633 634 markWord mark = obj->mark_acquire(); 635 for (;;) { 636 intptr_t hash = mark.hash(); 637 if (hash != 0) { 638 return hash; 639 } 640 641 hash = get_next_hash(current, obj); 642 const markWord old_mark = mark; 643 const markWord new_mark = old_mark.copy_set_hash(hash); 644 645 mark = obj->cas_set_mark(new_mark, old_mark); 646 if (old_mark == mark) { 647 return hash; 648 } 649 } 650 } 651 652 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) { 653 if (UseObjectMonitorTable) { 654 // Since the monitor isn't in the object header, the hash can simply be 655 // installed in the object header. 656 return install_hash_code(current, obj); 657 } 658 659 while (true) { 660 ObjectMonitor* monitor = nullptr; 661 markWord temp, test; 662 intptr_t hash; 663 markWord mark = obj->mark_acquire(); 664 if (mark.is_unlocked() || mark.is_fast_locked()) { 665 hash = mark.hash(); 666 if (hash != 0) { // if it has a hash, just return it 667 return hash; 668 } 669 hash = get_next_hash(current, obj); // get a new hash 670 temp = mark.copy_set_hash(hash); // merge the hash into header 671 // try to install the hash 672 test = obj->cas_set_mark(temp, mark); 673 if (test == mark) { // if the hash was installed, return it 674 return hash; 675 } 676 // CAS failed, retry 677 continue; 678 679 // Failed to install the hash. It could be that another thread 680 // installed the hash just before our attempt or inflation has 681 // occurred or... so we fall thru to inflate the monitor for 682 // stability and then install the hash. 683 } else if (mark.has_monitor()) { 684 monitor = mark.monitor(); 685 temp = monitor->header(); 686 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 687 hash = temp.hash(); 688 if (hash != 0) { 689 // It has a hash. 690 691 // Separate load of dmw/header above from the loads in 692 // is_being_async_deflated(). 693 694 // dmw/header and _contentions may get written by different threads. 695 // Make sure to observe them in the same order when having several observers. 696 OrderAccess::loadload_for_IRIW(); 697 698 if (monitor->is_being_async_deflated()) { 699 // But we can't safely use the hash if we detect that async 700 // deflation has occurred. So we attempt to restore the 701 // header/dmw to the object's header so that we only retry 702 // once if the deflater thread happens to be slow. 703 monitor->install_displaced_markword_in_object(obj); 704 continue; 705 } 706 return hash; 707 } 708 // Fall thru so we only have one place that installs the hash in 709 // the ObjectMonitor. 710 } 711 712 // NOTE: an async deflation can race after we get the monitor and 713 // before we can update the ObjectMonitor's header with the hash 714 // value below. 715 assert(mark.has_monitor(), "must be"); 716 monitor = mark.monitor(); 717 718 // Load ObjectMonitor's header/dmw field and see if it has a hash. 719 mark = monitor->header(); 720 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 721 hash = mark.hash(); 722 if (hash == 0) { // if it does not have a hash 723 hash = get_next_hash(current, obj); // get a new hash 724 temp = mark.copy_set_hash(hash) ; // merge the hash into header 725 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 726 uintptr_t v = AtomicAccess::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value()); 727 test = markWord(v); 728 if (test != mark) { 729 // The attempt to update the ObjectMonitor's header/dmw field 730 // did not work. This can happen if another thread managed to 731 // merge in the hash just before our cmpxchg(). 732 // If we add any new usages of the header/dmw field, this code 733 // will need to be updated. 734 hash = test.hash(); 735 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 736 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 737 } 738 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) { 739 // If we detect that async deflation has occurred, then we 740 // attempt to restore the header/dmw to the object's header 741 // so that we only retry once if the deflater thread happens 742 // to be slow. 743 monitor->install_displaced_markword_in_object(obj); 744 continue; 745 } 746 } 747 // We finally get the hash. 748 return hash; 749 } 750 } 751 752 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current, 753 Handle h_obj) { 754 assert(current == JavaThread::current(), "Can only be called on current thread"); 755 oop obj = h_obj(); 756 757 markWord mark = obj->mark_acquire(); 758 759 if (mark.is_fast_locked()) { 760 // fast-locking case, see if lock is in current's lock stack 761 return current->lock_stack().contains(h_obj()); 762 } 763 764 while (mark.has_monitor()) { 765 ObjectMonitor* monitor = read_monitor(current, obj, mark); 766 if (monitor != nullptr) { 767 return monitor->is_entered(current) != 0; 768 } 769 // Racing with inflation/deflation, retry 770 mark = obj->mark_acquire(); 771 772 if (mark.is_fast_locked()) { 773 // Some other thread fast_locked, current could not have held the lock 774 return false; 775 } 776 } 777 778 // Unlocked case, header in place 779 assert(mark.is_unlocked(), "sanity check"); 780 return false; 781 } 782 783 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 784 oop obj = h_obj(); 785 markWord mark = obj->mark_acquire(); 786 787 if (mark.is_fast_locked()) { 788 // fast-locked so get owner from the object. 789 // owning_thread_from_object() may also return null here: 790 return Threads::owning_thread_from_object(t_list, h_obj()); 791 } 792 793 while (mark.has_monitor()) { 794 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark); 795 if (monitor != nullptr) { 796 return Threads::owning_thread_from_monitor(t_list, monitor); 797 } 798 // Racing with inflation/deflation, retry 799 mark = obj->mark_acquire(); 800 801 if (mark.is_fast_locked()) { 802 // Some other thread fast_locked 803 return Threads::owning_thread_from_object(t_list, h_obj()); 804 } 805 } 806 807 // Unlocked case, header in place 808 // Cannot have assertion since this object may have been 809 // locked by another thread when reaching here. 810 // assert(mark.is_unlocked(), "sanity check"); 811 812 return nullptr; 813 } 814 815 // Visitors ... 816 817 // Iterate over all ObjectMonitors. 818 template <typename Function> 819 void ObjectSynchronizer::monitors_iterate(Function function) { 820 MonitorList::Iterator iter = _in_use_list.iterator(); 821 while (iter.has_next()) { 822 ObjectMonitor* monitor = iter.next(); 823 function(monitor); 824 } 825 } 826 827 // Iterate ObjectMonitors owned by any thread and where the owner `filter` 828 // returns true. 829 template <typename OwnerFilter> 830 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) { 831 monitors_iterate([&](ObjectMonitor* monitor) { 832 // This function is only called at a safepoint or when the 833 // target thread is suspended or when the target thread is 834 // operating on itself. The current closures in use today are 835 // only interested in an owned ObjectMonitor and ownership 836 // cannot be dropped under the calling contexts so the 837 // ObjectMonitor cannot be async deflated. 838 if (monitor->has_owner() && filter(monitor)) { 839 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating"); 840 841 closure->do_monitor(monitor); 842 } 843 }); 844 } 845 846 // Iterate ObjectMonitors where the owner == thread; this does NOT include 847 // ObjectMonitors where owner is set to a stack-lock address in thread. 848 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) { 849 int64_t key = ObjectMonitor::owner_id_from(thread); 850 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 851 return owned_monitors_iterate_filtered(closure, thread_filter); 852 } 853 854 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) { 855 int64_t key = ObjectMonitor::owner_id_from(vthread); 856 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; }; 857 return owned_monitors_iterate_filtered(closure, thread_filter); 858 } 859 860 // Iterate ObjectMonitors owned by any thread. 861 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) { 862 auto all_filter = [&](ObjectMonitor* monitor) { return true; }; 863 return owned_monitors_iterate_filtered(closure, all_filter); 864 } 865 866 static bool monitors_used_above_threshold(MonitorList* list) { 867 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy 868 return false; 869 } 870 size_t monitors_used = list->count(); 871 if (monitors_used == 0) { // empty list is easy 872 return false; 873 } 874 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling(); 875 // Make sure that we use a ceiling value that is not lower than 876 // previous, not lower than the recorded max used by the system, and 877 // not lower than the current number of monitors in use (which can 878 // race ahead of max). The result is guaranteed > 0. 879 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used); 880 881 // Check if our monitor usage is above the threshold: 882 size_t monitor_usage = (monitors_used * 100LL) / ceiling; 883 if (int(monitor_usage) > MonitorUsedDeflationThreshold) { 884 // Deflate monitors if over the threshold percentage, unless no 885 // progress on previous deflations. 886 bool is_above_threshold = true; 887 888 // Check if it's time to adjust the in_use_list_ceiling up, due 889 // to too many async deflation attempts without any progress. 890 if (NoAsyncDeflationProgressMax != 0 && 891 _no_progress_cnt >= NoAsyncDeflationProgressMax) { 892 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0; 893 size_t delta = (size_t)(ceiling * remainder) + 1; 894 size_t new_ceiling = (ceiling > SIZE_MAX - delta) 895 ? SIZE_MAX // Overflow, let's clamp new_ceiling. 896 : ceiling + delta; 897 898 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling); 899 log_info(monitorinflation)("Too many deflations without progress; " 900 "bumping in_use_list_ceiling from %zu" 901 " to %zu", old_ceiling, new_ceiling); 902 _no_progress_cnt = 0; 903 ceiling = new_ceiling; 904 905 // Check if our monitor usage is still above the threshold: 906 monitor_usage = (monitors_used * 100LL) / ceiling; 907 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold; 908 } 909 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu" 910 ", monitor_usage=%zu, threshold=%d", 911 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold); 912 return is_above_threshold; 913 } 914 915 return false; 916 } 917 918 size_t ObjectSynchronizer::in_use_list_count() { 919 return _in_use_list.count(); 920 } 921 922 size_t ObjectSynchronizer::in_use_list_max() { 923 return _in_use_list.max(); 924 } 925 926 size_t ObjectSynchronizer::in_use_list_ceiling() { 927 return _in_use_list_ceiling; 928 } 929 930 void ObjectSynchronizer::dec_in_use_list_ceiling() { 931 AtomicAccess::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 932 } 933 934 void ObjectSynchronizer::inc_in_use_list_ceiling() { 935 AtomicAccess::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate); 936 } 937 938 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) { 939 _in_use_list_ceiling = new_value; 940 } 941 942 bool ObjectSynchronizer::is_async_deflation_needed() { 943 if (is_async_deflation_requested()) { 944 // Async deflation request. 945 log_info(monitorinflation)("Async deflation needed: explicit request"); 946 return true; 947 } 948 949 jlong time_since_last = time_since_last_async_deflation_ms(); 950 951 if (AsyncDeflationInterval > 0 && 952 time_since_last > AsyncDeflationInterval && 953 monitors_used_above_threshold(&_in_use_list)) { 954 // It's been longer than our specified deflate interval and there 955 // are too many monitors in use. We don't deflate more frequently 956 // than AsyncDeflationInterval (unless is_async_deflation_requested) 957 // in order to not swamp the MonitorDeflationThread. 958 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold"); 959 return true; 960 } 961 962 if (GuaranteedAsyncDeflationInterval > 0 && 963 time_since_last > GuaranteedAsyncDeflationInterval) { 964 // It's been longer than our specified guaranteed deflate interval. 965 // We need to clean up the used monitors even if the threshold is 966 // not reached, to keep the memory utilization at bay when many threads 967 // touched many monitors. 968 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) " 969 "is greater than time since last deflation (" JLONG_FORMAT " ms)", 970 GuaranteedAsyncDeflationInterval, time_since_last); 971 972 // If this deflation has no progress, then it should not affect the no-progress 973 // tracking, otherwise threshold heuristics would think it was triggered, experienced 974 // no progress, and needs to backoff more aggressively. In this "no progress" case, 975 // the generic code would bump the no-progress counter, and we compensate for that 976 // by telling it to skip the update. 977 // 978 // If this deflation has progress, then it should let non-progress tracking 979 // know about this, otherwise the threshold heuristics would kick in, potentially 980 // experience no-progress due to aggressive cleanup by this deflation, and think 981 // it is still in no-progress stride. In this "progress" case, the generic code would 982 // zero the counter, and we allow it to happen. 983 _no_progress_skip_increment = true; 984 985 return true; 986 } 987 988 return false; 989 } 990 991 void ObjectSynchronizer::request_deflate_idle_monitors() { 992 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag); 993 set_is_async_deflation_requested(true); 994 ml.notify_all(); 995 } 996 997 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() { 998 JavaThread* current = JavaThread::current(); 999 bool ret_code = false; 1000 1001 jlong last_time = last_async_deflation_time_ns(); 1002 1003 request_deflate_idle_monitors(); 1004 1005 const int N_CHECKS = 5; 1006 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds 1007 if (last_async_deflation_time_ns() > last_time) { 1008 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i); 1009 ret_code = true; 1010 break; 1011 } 1012 { 1013 // JavaThread has to honor the blocking protocol. 1014 ThreadBlockInVM tbivm(current); 1015 os::naked_short_sleep(999); // sleep for almost 1 second 1016 } 1017 } 1018 if (!ret_code) { 1019 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS); 1020 } 1021 1022 return ret_code; 1023 } 1024 1025 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1026 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS); 1027 } 1028 1029 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle 1030 // ObjectMonitors. Returns the number of deflated ObjectMonitors. 1031 // 1032 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) { 1033 MonitorList::Iterator iter = _in_use_list.iterator(); 1034 size_t deflated_count = 0; 1035 Thread* current = Thread::current(); 1036 1037 while (iter.has_next()) { 1038 if (deflated_count >= (size_t)MonitorDeflationMax) { 1039 break; 1040 } 1041 ObjectMonitor* mid = iter.next(); 1042 if (mid->deflate_monitor(current)) { 1043 deflated_count++; 1044 } 1045 1046 // Must check for a safepoint/handshake and honor it. 1047 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count); 1048 } 1049 1050 return deflated_count; 1051 } 1052 1053 class DeflationHandshakeClosure : public HandshakeClosure { 1054 public: 1055 DeflationHandshakeClosure() : HandshakeClosure("DeflationHandshakeClosure") {} 1056 1057 void do_thread(Thread* thread) { 1058 log_trace(monitorinflation)("DeflationHandshakeClosure::do_thread: thread=" 1059 INTPTR_FORMAT, p2i(thread)); 1060 if (thread->is_Java_thread()) { 1061 // Clear OM cache 1062 JavaThread* jt = JavaThread::cast(thread); 1063 jt->om_clear_monitor_cache(); 1064 } 1065 } 1066 }; 1067 1068 class VM_RendezvousGCThreads : public VM_Operation { 1069 public: 1070 bool evaluate_at_safepoint() const override { return false; } 1071 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; } 1072 void doit() override { 1073 Universe::heap()->safepoint_synchronize_begin(); 1074 Universe::heap()->safepoint_synchronize_end(); 1075 }; 1076 }; 1077 1078 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list, 1079 ObjectMonitorDeflationSafepointer* safepointer) { 1080 NativeHeapTrimmer::SuspendMark sm("monitor deletion"); 1081 size_t deleted_count = 0; 1082 for (ObjectMonitor* monitor: *delete_list) { 1083 delete monitor; 1084 deleted_count++; 1085 // A JavaThread must check for a safepoint/handshake and honor it. 1086 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count); 1087 } 1088 return deleted_count; 1089 } 1090 1091 class ObjectMonitorDeflationLogging: public StackObj { 1092 LogStreamHandle(Debug, monitorinflation) _debug; 1093 LogStreamHandle(Info, monitorinflation) _info; 1094 LogStream* _stream; 1095 elapsedTimer _timer; 1096 1097 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); } 1098 size_t count() const { return ObjectSynchronizer::in_use_list_count(); } 1099 size_t max() const { return ObjectSynchronizer::in_use_list_max(); } 1100 1101 public: 1102 ObjectMonitorDeflationLogging() 1103 : _debug(), _info(), _stream(nullptr) { 1104 if (_debug.is_enabled()) { 1105 _stream = &_debug; 1106 } else if (_info.is_enabled()) { 1107 _stream = &_info; 1108 } 1109 } 1110 1111 void begin() { 1112 if (_stream != nullptr) { 1113 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1114 ceiling(), count(), max()); 1115 _timer.start(); 1116 } 1117 } 1118 1119 void before_handshake(size_t unlinked_count) { 1120 if (_stream != nullptr) { 1121 _timer.stop(); 1122 _stream->print_cr("before handshaking: unlinked_count=%zu" 1123 ", in_use_list stats: ceiling=%zu, count=" 1124 "%zu, max=%zu", 1125 unlinked_count, ceiling(), count(), max()); 1126 } 1127 } 1128 1129 void after_handshake() { 1130 if (_stream != nullptr) { 1131 _stream->print_cr("after handshaking: in_use_list stats: ceiling=" 1132 "%zu, count=%zu, max=%zu", 1133 ceiling(), count(), max()); 1134 _timer.start(); 1135 } 1136 } 1137 1138 void end(size_t deflated_count, size_t unlinked_count) { 1139 if (_stream != nullptr) { 1140 _timer.stop(); 1141 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) { 1142 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs", 1143 deflated_count, unlinked_count, _timer.seconds()); 1144 } 1145 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu", 1146 ceiling(), count(), max()); 1147 } 1148 } 1149 1150 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) { 1151 if (_stream != nullptr) { 1152 _timer.stop(); 1153 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling=" 1154 "%zu, count=%zu, max=%zu", 1155 op_name, cnt_name, cnt, ceiling(), count(), max()); 1156 } 1157 } 1158 1159 void after_block_for_safepoint(const char* op_name) { 1160 if (_stream != nullptr) { 1161 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu" 1162 ", count=%zu, max=%zu", op_name, 1163 ceiling(), count(), max()); 1164 _timer.start(); 1165 } 1166 } 1167 }; 1168 1169 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) { 1170 if (!SafepointMechanism::should_process(_current)) { 1171 return; 1172 } 1173 1174 // A safepoint/handshake has started. 1175 _log->before_block_for_safepoint(op_name, count_name, counter); 1176 1177 { 1178 // Honor block request. 1179 ThreadBlockInVM tbivm(_current); 1180 } 1181 1182 _log->after_block_for_safepoint(op_name); 1183 } 1184 1185 // This function is called by the MonitorDeflationThread to deflate 1186 // ObjectMonitors. 1187 size_t ObjectSynchronizer::deflate_idle_monitors() { 1188 JavaThread* current = JavaThread::current(); 1189 assert(current->is_monitor_deflation_thread(), "The only monitor deflater"); 1190 1191 // The async deflation request has been processed. 1192 _last_async_deflation_time_ns = os::javaTimeNanos(); 1193 set_is_async_deflation_requested(false); 1194 1195 ObjectMonitorDeflationLogging log; 1196 ObjectMonitorDeflationSafepointer safepointer(current, &log); 1197 1198 log.begin(); 1199 1200 // Deflate some idle ObjectMonitors. 1201 size_t deflated_count = deflate_monitor_list(&safepointer); 1202 1203 // Unlink the deflated ObjectMonitors from the in-use list. 1204 size_t unlinked_count = 0; 1205 size_t deleted_count = 0; 1206 if (deflated_count > 0) { 1207 ResourceMark rm(current); 1208 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count); 1209 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer); 1210 1211 #ifdef ASSERT 1212 if (UseObjectMonitorTable) { 1213 for (ObjectMonitor* monitor : delete_list) { 1214 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed"); 1215 } 1216 } 1217 #endif 1218 1219 log.before_handshake(unlinked_count); 1220 1221 // A JavaThread needs to handshake in order to safely free the 1222 // ObjectMonitors that were deflated in this cycle. 1223 DeflationHandshakeClosure dhc; 1224 Handshake::execute(&dhc); 1225 // Also, we sync and desync GC threads around the handshake, so that they can 1226 // safely read the mark-word and look-through to the object-monitor, without 1227 // being afraid that the object-monitor is going away. 1228 VM_RendezvousGCThreads sync_gc; 1229 VMThread::execute(&sync_gc); 1230 1231 log.after_handshake(); 1232 1233 // After the handshake, safely free the ObjectMonitors that were 1234 // deflated and unlinked in this cycle. 1235 1236 // Delete the unlinked ObjectMonitors. 1237 deleted_count = delete_monitors(&delete_list, &safepointer); 1238 assert(unlinked_count == deleted_count, "must be"); 1239 } 1240 1241 log.end(deflated_count, unlinked_count); 1242 1243 GVars.stw_random = os::random(); 1244 1245 if (deflated_count != 0) { 1246 _no_progress_cnt = 0; 1247 } else if (_no_progress_skip_increment) { 1248 _no_progress_skip_increment = false; 1249 } else { 1250 _no_progress_cnt++; 1251 } 1252 1253 return deflated_count; 1254 } 1255 1256 // Monitor cleanup on JavaThread::exit 1257 1258 // Iterate through monitor cache and attempt to release thread's monitors 1259 class ReleaseJavaMonitorsClosure: public MonitorClosure { 1260 private: 1261 JavaThread* _thread; 1262 1263 public: 1264 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {} 1265 void do_monitor(ObjectMonitor* mid) { 1266 intx rec = mid->complete_exit(_thread); 1267 _thread->dec_held_monitor_count(rec + 1); 1268 } 1269 }; 1270 1271 // Release all inflated monitors owned by current thread. Lightweight monitors are 1272 // ignored. This is meant to be called during JNI thread detach which assumes 1273 // all remaining monitors are heavyweight. All exceptions are swallowed. 1274 // Scanning the extant monitor list can be time consuming. 1275 // A simple optimization is to add a per-thread flag that indicates a thread 1276 // called jni_monitorenter() during its lifetime. 1277 // 1278 // Instead of NoSafepointVerifier it might be cheaper to 1279 // use an idiom of the form: 1280 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 1281 // <code that must not run at safepoint> 1282 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 1283 // Since the tests are extremely cheap we could leave them enabled 1284 // for normal product builds. 1285 1286 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) { 1287 assert(current == JavaThread::current(), "must be current Java thread"); 1288 NoSafepointVerifier nsv; 1289 ReleaseJavaMonitorsClosure rjmc(current); 1290 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current); 1291 assert(!current->has_pending_exception(), "Should not be possible"); 1292 current->clear_pending_exception(); 1293 assert(current->held_monitor_count() == 0, "Should not be possible"); 1294 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count. 1295 current->clear_jni_monitor_count(); 1296 } 1297 1298 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 1299 switch (cause) { 1300 case inflate_cause_vm_internal: return "VM Internal"; 1301 case inflate_cause_monitor_enter: return "Monitor Enter"; 1302 case inflate_cause_wait: return "Monitor Wait"; 1303 case inflate_cause_notify: return "Monitor Notify"; 1304 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 1305 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 1306 default: 1307 ShouldNotReachHere(); 1308 } 1309 return "Unknown"; 1310 } 1311 1312 //------------------------------------------------------------------------------ 1313 // Debugging code 1314 1315 u_char* ObjectSynchronizer::get_gvars_addr() { 1316 return (u_char*)&GVars; 1317 } 1318 1319 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 1320 return (u_char*)&GVars.hc_sequence; 1321 } 1322 1323 size_t ObjectSynchronizer::get_gvars_size() { 1324 return sizeof(SharedGlobals); 1325 } 1326 1327 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 1328 return (u_char*)&GVars.stw_random; 1329 } 1330 1331 // Do the final audit and print of ObjectMonitor stats; must be done 1332 // by the VMThread at VM exit time. 1333 void ObjectSynchronizer::do_final_audit_and_print_stats() { 1334 assert(Thread::current()->is_VM_thread(), "sanity check"); 1335 1336 if (is_final_audit()) { // Only do the audit once. 1337 return; 1338 } 1339 set_is_final_audit(); 1340 log_info(monitorinflation)("Starting the final audit."); 1341 1342 if (log_is_enabled(Info, monitorinflation)) { 1343 LogStreamHandle(Info, monitorinflation) ls; 1344 audit_and_print_stats(&ls, true /* on_exit */); 1345 } 1346 } 1347 1348 // This function can be called by the MonitorDeflationThread or it can be called when 1349 // we are trying to exit the VM. The list walker functions can run in parallel with 1350 // the other list operations. 1351 // Calls to this function can be added in various places as a debugging 1352 // aid. 1353 // 1354 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) { 1355 int error_cnt = 0; 1356 1357 ls->print_cr("Checking in_use_list:"); 1358 chk_in_use_list(ls, &error_cnt); 1359 1360 if (error_cnt == 0) { 1361 ls->print_cr("No errors found in in_use_list checks."); 1362 } else { 1363 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt); 1364 } 1365 1366 // When exiting, only log the interesting entries at the Info level. 1367 // When called at intervals by the MonitorDeflationThread, log output 1368 // at the Trace level since there can be a lot of it. 1369 if (!on_exit && log_is_enabled(Trace, monitorinflation)) { 1370 LogStreamHandle(Trace, monitorinflation) ls_tr; 1371 log_in_use_monitor_details(&ls_tr, true /* log_all */); 1372 } else if (on_exit) { 1373 log_in_use_monitor_details(ls, false /* log_all */); 1374 } 1375 1376 ls->flush(); 1377 1378 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 1379 } 1380 1381 // Check the in_use_list; log the results of the checks. 1382 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) { 1383 size_t l_in_use_count = _in_use_list.count(); 1384 size_t l_in_use_max = _in_use_list.max(); 1385 out->print_cr("count=%zu, max=%zu", l_in_use_count, 1386 l_in_use_max); 1387 1388 size_t ck_in_use_count = 0; 1389 MonitorList::Iterator iter = _in_use_list.iterator(); 1390 while (iter.has_next()) { 1391 ObjectMonitor* mid = iter.next(); 1392 chk_in_use_entry(mid, out, error_cnt_p); 1393 ck_in_use_count++; 1394 } 1395 1396 if (l_in_use_count == ck_in_use_count) { 1397 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu", 1398 l_in_use_count, ck_in_use_count); 1399 } else { 1400 out->print_cr("WARNING: in_use_count=%zu is not equal to " 1401 "ck_in_use_count=%zu", l_in_use_count, 1402 ck_in_use_count); 1403 } 1404 1405 size_t ck_in_use_max = _in_use_list.max(); 1406 if (l_in_use_max == ck_in_use_max) { 1407 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu", 1408 l_in_use_max, ck_in_use_max); 1409 } else { 1410 out->print_cr("WARNING: in_use_max=%zu is not equal to " 1411 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max); 1412 } 1413 } 1414 1415 // Check an in-use monitor entry; log any errors. 1416 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out, 1417 int* error_cnt_p) { 1418 if (n->owner_is_DEFLATER_MARKER()) { 1419 // This could happen when monitor deflation blocks for a safepoint. 1420 return; 1421 } 1422 1423 1424 if (n->metadata() == 0) { 1425 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must " 1426 "have non-null _metadata (header/hash) field.", p2i(n)); 1427 *error_cnt_p = *error_cnt_p + 1; 1428 } 1429 1430 const oop obj = n->object_peek(); 1431 if (obj == nullptr) { 1432 return; 1433 } 1434 1435 const markWord mark = obj->mark(); 1436 if (!mark.has_monitor()) { 1437 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 1438 "object does not think it has a monitor: obj=" 1439 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 1440 p2i(obj), mark.value()); 1441 *error_cnt_p = *error_cnt_p + 1; 1442 return; 1443 } 1444 1445 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark); 1446 if (n != obj_mon) { 1447 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's " 1448 "object does not refer to the same monitor: obj=" 1449 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 1450 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 1451 *error_cnt_p = *error_cnt_p + 1; 1452 } 1453 } 1454 1455 // Log details about ObjectMonitors on the in_use_list. The 'BHL' 1456 // flags indicate why the entry is in-use, 'object' and 'object type' 1457 // indicate the associated object and its type. 1458 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) { 1459 if (_in_use_list.count() > 0) { 1460 stringStream ss; 1461 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)"); 1462 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 1463 out->print_cr("%18s %s %18s %18s", 1464 "monitor", "BHL", "object", "object type"); 1465 out->print_cr("================== === ================== =================="); 1466 1467 auto is_interesting = [&](ObjectMonitor* monitor) { 1468 return log_all || monitor->has_owner() || monitor->is_busy(); 1469 }; 1470 1471 monitors_iterate([&](ObjectMonitor* monitor) { 1472 if (is_interesting(monitor)) { 1473 const oop obj = monitor->object_peek(); 1474 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash(); 1475 ResourceMark rm; 1476 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor), 1477 monitor->is_busy(), hash != 0, monitor->has_owner(), 1478 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name()); 1479 if (monitor->is_busy()) { 1480 out->print(" (%s)", monitor->is_busy_to_string(&ss)); 1481 ss.reset(); 1482 } 1483 out->cr(); 1484 } 1485 }); 1486 } 1487 1488 out->flush(); 1489 }