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/oopStorage.hpp" 28 #include "gc/shared/oopStorageSet.hpp" 29 #include "jfr/jfrEvents.hpp" 30 #include "jfr/support/jfrThreadId.hpp" 31 #include "logging/log.hpp" 32 #include "logging/logStream.hpp" 33 #include "memory/allocation.inline.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/markWord.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "oops/oopHandle.inline.hpp" 38 #include "oops/weakHandle.inline.hpp" 39 #include "prims/jvmtiDeferredUpdates.hpp" 40 #include "prims/jvmtiExport.hpp" 41 #include "runtime/atomic.hpp" 42 #include "runtime/globals.hpp" 43 #include "runtime/handles.inline.hpp" 44 #include "runtime/interfaceSupport.inline.hpp" 45 #include "runtime/javaThread.inline.hpp" 46 #include "runtime/mutexLocker.hpp" 47 #include "runtime/objectMonitor.hpp" 48 #include "runtime/objectMonitor.inline.hpp" 49 #include "runtime/orderAccess.hpp" 50 #include "runtime/osThread.hpp" 51 #include "runtime/perfData.hpp" 52 #include "runtime/safefetch.hpp" 53 #include "runtime/safepointMechanism.inline.hpp" 54 #include "runtime/sharedRuntime.hpp" 55 #include "runtime/threads.hpp" 56 #include "services/threadService.hpp" 57 #include "utilities/dtrace.hpp" 58 #include "utilities/globalDefinitions.hpp" 59 #include "utilities/macros.hpp" 60 #include "utilities/preserveException.hpp" 61 #if INCLUDE_JFR 62 #include "jfr/support/jfrFlush.hpp" 63 #endif 64 65 #ifdef DTRACE_ENABLED 66 67 // Only bother with this argument setup if dtrace is available 68 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 69 70 71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 72 char* bytes = nullptr; \ 73 int len = 0; \ 74 jlong jtid = SharedRuntime::get_java_tid(thread); \ 75 Symbol* klassname = obj->klass()->name(); \ 76 if (klassname != nullptr) { \ 77 bytes = (char*)klassname->bytes(); \ 78 len = klassname->utf8_length(); \ 79 } 80 81 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 82 { \ 83 if (DTraceMonitorProbes) { \ 84 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 85 HOTSPOT_MONITOR_WAIT(jtid, \ 86 (monitor), bytes, len, (millis)); \ 87 } \ 88 } 89 90 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 91 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 92 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 93 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 94 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 95 96 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 97 { \ 98 if (DTraceMonitorProbes) { \ 99 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 100 HOTSPOT_MONITOR_##probe(jtid, \ 101 (uintptr_t)(monitor), bytes, len); \ 102 } \ 103 } 104 105 #else // ndef DTRACE_ENABLED 106 107 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 108 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 109 110 #endif // ndef DTRACE_ENABLED 111 112 DEBUG_ONLY(static volatile bool InitDone = false;) 113 114 OopStorage* ObjectMonitor::_oop_storage = nullptr; 115 116 OopHandle ObjectMonitor::_vthread_cxq_head; 117 ParkEvent* ObjectMonitor::_vthread_unparker_ParkEvent = nullptr; 118 119 static void post_virtual_thread_pinned_event(JavaThread* current, const char* reason) { 120 EventVirtualThreadPinned e; 121 if (e.should_commit()) { 122 e.set_pinnedReason(reason); 123 e.set_carrierThread(JFR_JVM_THREAD_ID(current)); 124 e.commit(); 125 } 126 } 127 128 // ----------------------------------------------------------------------------- 129 // Theory of operations -- Monitors lists, thread residency, etc: 130 // 131 // * A thread acquires ownership of a monitor by successfully 132 // CAS()ing the _owner field from null to non-null. 133 // 134 // * Invariant: A thread appears on at most one monitor list -- 135 // cxq, EntryList or WaitSet -- at any one time. 136 // 137 // * Contending threads "push" themselves onto the cxq with CAS 138 // and then spin/park. 139 // 140 // * After a contending thread eventually acquires the lock it must 141 // dequeue itself from either the EntryList or the cxq. 142 // 143 // * The exiting thread identifies and unparks an "heir presumptive" 144 // tentative successor thread on the EntryList. Critically, the 145 // exiting thread doesn't unlink the successor thread from the EntryList. 146 // After having been unparked, the wakee will recontend for ownership of 147 // the monitor. The successor (wakee) will either acquire the lock or 148 // re-park itself. 149 // 150 // Succession is provided for by a policy of competitive handoff. 151 // The exiting thread does _not_ grant or pass ownership to the 152 // successor thread. (This is also referred to as "handoff" succession"). 153 // Instead the exiting thread releases ownership and possibly wakes 154 // a successor, so the successor can (re)compete for ownership of the lock. 155 // If the EntryList is empty but the cxq is populated the exiting 156 // thread will drain the cxq into the EntryList. It does so by 157 // by detaching the cxq (installing null with CAS) and folding 158 // the threads from the cxq into the EntryList. The EntryList is 159 // doubly linked, while the cxq is singly linked because of the 160 // CAS-based "push" used to enqueue recently arrived threads (RATs). 161 // 162 // * Concurrency invariants: 163 // 164 // -- only the monitor owner may access or mutate the EntryList. 165 // The mutex property of the monitor itself protects the EntryList 166 // from concurrent interference. 167 // -- Only the monitor owner may detach the cxq. 168 // 169 // * The monitor entry list operations avoid locks, but strictly speaking 170 // they're not lock-free. Enter is lock-free, exit is not. 171 // For a description of 'Methods and apparatus providing non-blocking access 172 // to a resource,' see U.S. Pat. No. 7844973. 173 // 174 // * The cxq can have multiple concurrent "pushers" but only one concurrent 175 // detaching thread. This mechanism is immune from the ABA corruption. 176 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 177 // 178 // * Taken together, the cxq and the EntryList constitute or form a 179 // single logical queue of threads stalled trying to acquire the lock. 180 // We use two distinct lists to improve the odds of a constant-time 181 // dequeue operation after acquisition (in the ::enter() epilogue) and 182 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 183 // A key desideratum is to minimize queue & monitor metadata manipulation 184 // that occurs while holding the monitor lock -- that is, we want to 185 // minimize monitor lock holds times. Note that even a small amount of 186 // fixed spinning will greatly reduce the # of enqueue-dequeue operations 187 // on EntryList|cxq. That is, spinning relieves contention on the "inner" 188 // locks and monitor metadata. 189 // 190 // Cxq points to the set of Recently Arrived Threads attempting entry. 191 // Because we push threads onto _cxq with CAS, the RATs must take the form of 192 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 193 // the unlocking thread notices that EntryList is null but _cxq is != null. 194 // 195 // The EntryList is ordered by the prevailing queue discipline and 196 // can be organized in any convenient fashion, such as a doubly-linked list or 197 // a circular doubly-linked list. Critically, we want insert and delete operations 198 // to operate in constant-time. If we need a priority queue then something akin 199 // to Solaris' sleepq would work nicely. Viz., 200 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 201 // Queue discipline is enforced at ::exit() time, when the unlocking thread 202 // drains the cxq into the EntryList, and orders or reorders the threads on the 203 // EntryList accordingly. 204 // 205 // Barring "lock barging", this mechanism provides fair cyclic ordering, 206 // somewhat similar to an elevator-scan. 207 // 208 // * The monitor synchronization subsystem avoids the use of native 209 // synchronization primitives except for the narrow platform-specific 210 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 211 // the semantics of park-unpark. Put another way, this monitor implementation 212 // depends only on atomic operations and park-unpark. The monitor subsystem 213 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 214 // underlying OS manages the READY<->RUN transitions. 215 // 216 // * Waiting threads reside on the WaitSet list -- wait() puts 217 // the caller onto the WaitSet. 218 // 219 // * notify() or notifyAll() simply transfers threads from the WaitSet to 220 // either the EntryList or cxq. Subsequent exit() operations will 221 // unpark the notifyee. Unparking a notifee in notify() is inefficient - 222 // it's likely the notifyee would simply impale itself on the lock held 223 // by the notifier. 224 // 225 // * An interesting alternative is to encode cxq as (List,LockByte) where 226 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 227 // variable, like _recursions, in the scheme. The threads or Events that form 228 // the list would have to be aligned in 256-byte addresses. A thread would 229 // try to acquire the lock or enqueue itself with CAS, but exiting threads 230 // could use a 1-0 protocol and simply STB to set the LockByte to 0. 231 // Note that is is *not* word-tearing, but it does presume that full-word 232 // CAS operations are coherent with intermix with STB operations. That's true 233 // on most common processors. 234 // 235 // * See also http://blogs.sun.com/dave 236 237 238 // Check that object() and set_object() are called from the right context: 239 static void check_object_context() { 240 #ifdef ASSERT 241 Thread* self = Thread::current(); 242 if (self->is_Java_thread()) { 243 // Mostly called from JavaThreads so sanity check the thread state. 244 JavaThread* jt = JavaThread::cast(self); 245 switch (jt->thread_state()) { 246 case _thread_in_vm: // the usual case 247 case _thread_in_Java: // during deopt 248 break; 249 default: 250 fatal("called from an unsafe thread state"); 251 } 252 assert(jt->is_active_Java_thread(), "must be active JavaThread"); 253 } else { 254 // However, ThreadService::get_current_contended_monitor() 255 // can call here via the VMThread so sanity check it. 256 assert(self->is_VM_thread(), "must be"); 257 } 258 #endif // ASSERT 259 } 260 261 ObjectMonitor::ObjectMonitor(oop object) : 262 _header(markWord::zero()), 263 _object(_oop_storage, object), 264 _owner(nullptr), 265 _stack_locker(nullptr), 266 _previous_owner_tid(0), 267 _next_om(nullptr), 268 _recursions(0), 269 _EntryList(nullptr), 270 _cxq(nullptr), 271 _succ(nullptr), 272 _Responsible(nullptr), 273 _SpinDuration(ObjectMonitor::Knob_SpinLimit), 274 _contentions(0), 275 _WaitSet(nullptr), 276 _waiters(0), 277 _WaitSetLock(0) 278 { } 279 280 ObjectMonitor::~ObjectMonitor() { 281 _object.release(_oop_storage); 282 } 283 284 oop ObjectMonitor::object() const { 285 check_object_context(); 286 return _object.resolve(); 287 } 288 289 oop ObjectMonitor::object_peek() const { 290 return _object.peek(); 291 } 292 293 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) { 294 if (current->is_suspended()) { 295 _om->_recursions = 0; 296 _om->_succ = nullptr; 297 // Don't need a full fence after clearing successor here because of the call to exit(). 298 _om->exit(current, false /* not_suspended */); 299 _om_exited = true; 300 301 current->set_current_pending_monitor(_om); 302 } 303 } 304 305 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) { 306 if (current->is_suspended()) { 307 if (_om->_succ == current) { 308 _om->_succ = nullptr; 309 OrderAccess::fence(); // always do a full fence when successor is cleared 310 } 311 } 312 } 313 314 // ----------------------------------------------------------------------------- 315 // Enter support 316 317 bool ObjectMonitor::enter_for(JavaThread* locking_thread) { 318 // Used by ObjectSynchronizer::enter_for to enter for another thread. 319 // The monitor is private to or already owned by locking_thread which must be suspended. 320 // So this code may only contend with deflation. 321 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 322 323 // Block out deflation as soon as possible. 324 add_to_contentions(1); 325 326 bool success = false; 327 if (!is_being_async_deflated()) { 328 void* prev_owner = try_set_owner_from(nullptr, locking_thread); 329 330 if (prev_owner == nullptr) { 331 assert(_recursions == 0, "invariant"); 332 success = true; 333 } else if (prev_owner == owner_for(locking_thread)) { 334 _recursions++; 335 success = true; 336 } else if (prev_owner == DEFLATER_MARKER) { 337 // Racing with deflation. 338 prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread); 339 if (prev_owner == DEFLATER_MARKER) { 340 // Cancelled deflation. Increment contentions as part of the deflation protocol. 341 add_to_contentions(1); 342 success = true; 343 } else if (prev_owner == nullptr) { 344 // At this point we cannot race with deflation as we have both incremented 345 // contentions, seen contention > 0 and seen a DEFLATER_MARKER. 346 // success will only be false if this races with something other than 347 // deflation. 348 prev_owner = try_set_owner_from(nullptr, locking_thread); 349 success = prev_owner == nullptr; 350 } 351 } 352 assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT 353 ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT, 354 p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner)); 355 } else { 356 // Async deflation is in progress and our contentions increment 357 // above lost the race to async deflation. Undo the work and 358 // force the caller to retry. 359 const oop l_object = object(); 360 if (l_object != nullptr) { 361 // Attempt to restore the header/dmw to the object's header so that 362 // we only retry once if the deflater thread happens to be slow. 363 install_displaced_markword_in_object(l_object); 364 } 365 } 366 367 add_to_contentions(-1); 368 369 assert(!success || is_owner(locking_thread), "must be"); 370 371 return success; 372 } 373 374 bool ObjectMonitor::enter(JavaThread* current) { 375 assert(current == JavaThread::current(), "must be"); 376 // The following code is ordered to check the most common cases first 377 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 378 379 void* cur = try_set_owner_from(nullptr, current); 380 if (cur == nullptr) { 381 assert(_recursions == 0, "invariant"); 382 return true; 383 } 384 385 if (cur == owner_for(current)) { 386 // TODO-FIXME: check for integer overflow! BUGID 6557169. 387 _recursions++; 388 return true; 389 } 390 391 // We've encountered genuine contention. 392 393 // Try one round of spinning *before* enqueueing current 394 // and before going through the awkward and expensive state 395 // transitions. The following spin is strictly optional ... 396 // Note that if we acquire the monitor from an initial spin 397 // we forgo posting JVMTI events and firing DTRACE probes. 398 if (TrySpin(current)) { 399 assert(owner_raw() == owner_for(current), "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw())); 400 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions); 401 assert(object()->mark() == markWord::encode(this), 402 "object mark must match encoded this: mark=" INTPTR_FORMAT 403 ", encoded this=" INTPTR_FORMAT, object()->mark().value(), 404 markWord::encode(this).value()); 405 return true; 406 } 407 408 assert(owner_raw() != owner_for(current), "invariant"); 409 assert(_succ != current, "invariant"); 410 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 411 assert(current->thread_state() != _thread_blocked, "invariant"); 412 413 // Keep track of contention for JVM/TI and M&M queries. 414 add_to_contentions(1); 415 if (is_being_async_deflated()) { 416 // Async deflation is in progress and our contentions increment 417 // above lost the race to async deflation. Undo the work and 418 // force the caller to retry. 419 const oop l_object = object(); 420 if (l_object != nullptr) { 421 // Attempt to restore the header/dmw to the object's header so that 422 // we only retry once if the deflater thread happens to be slow. 423 install_displaced_markword_in_object(l_object); 424 } 425 add_to_contentions(-1); 426 return false; 427 } 428 429 JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);) 430 EventJavaMonitorEnter event; 431 if (event.is_started()) { 432 event.set_monitorClass(object()->klass()); 433 // Set an address that is 'unique enough', such that events close in 434 // time and with the same address are likely (but not guaranteed) to 435 // belong to the same object. 436 event.set_address((uintptr_t)this); 437 } 438 439 { // Change java thread status to indicate blocked on monitor enter. 440 JavaThreadBlockedOnMonitorEnterState jtbmes(current, this); 441 442 assert(current->current_pending_monitor() == nullptr, "invariant"); 443 current->set_current_pending_monitor(this); 444 445 DTRACE_MONITOR_PROBE(contended__enter, this, object(), current); 446 if (JvmtiExport::should_post_monitor_contended_enter()) { 447 JvmtiExport::post_monitor_contended_enter(current, this); 448 449 // The current thread does not yet own the monitor and does not 450 // yet appear on any queues that would get it made the successor. 451 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 452 // handler cannot accidentally consume an unpark() meant for the 453 // ParkEvent associated with this ObjectMonitor. 454 } 455 456 #ifdef LOOM_MONITOR_SUPPORT 457 ContinuationEntry* ce = current->last_continuation(); 458 if (ce != nullptr && ce->is_virtual_thread() && current->is_on_monitorenter()) { 459 int result = Continuation::try_preempt(current, ce->cont_oop(current)); 460 if (result == freeze_ok) { 461 bool acquired = HandlePreemptedVThread(current); 462 DEBUG_ONLY(int state = java_lang_VirtualThread::state(current->vthread())); 463 assert((acquired && current->preemption_cancelled() && state == java_lang_VirtualThread::RUNNING) || 464 (!acquired && !current->preemption_cancelled() && state == java_lang_VirtualThread::BLOCKING), "invariant"); 465 return true; 466 } 467 if (result == freeze_pinned_native) { 468 post_virtual_thread_pinned_event(current, "Native frame or <clinit> on stack"); 469 } 470 } 471 #endif 472 473 OSThreadContendState osts(current->osthread()); 474 475 assert(current->thread_state() == _thread_in_vm, "invariant"); 476 477 for (;;) { 478 ExitOnSuspend eos(this); 479 { 480 ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */); 481 EnterI(current); 482 current->set_current_pending_monitor(nullptr); 483 // We can go to a safepoint at the end of this block. If we 484 // do a thread dump during that safepoint, then this thread will show 485 // as having "-locked" the monitor, but the OS and java.lang.Thread 486 // states will still report that the thread is blocked trying to 487 // acquire it. 488 // If there is a suspend request, ExitOnSuspend will exit the OM 489 // and set the OM as pending. 490 } 491 if (!eos.exited()) { 492 // ExitOnSuspend did not exit the OM 493 assert(owner_raw() == owner_for(current), "invariant"); 494 break; 495 } 496 } 497 498 // We've just gotten past the enter-check-for-suspend dance and we now own 499 // the monitor free and clear. 500 } 501 502 add_to_contentions(-1); 503 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions()); 504 505 // Must either set _recursions = 0 or ASSERT _recursions == 0. 506 assert(_recursions == 0, "invariant"); 507 assert(owner_raw() == owner_for(current), "invariant"); 508 assert(_succ != current, "invariant"); 509 assert(object()->mark() == markWord::encode(this), "invariant"); 510 511 // The thread -- now the owner -- is back in vm mode. 512 // Report the glorious news via TI,DTrace and jvmstat. 513 // The probe effect is non-trivial. All the reportage occurs 514 // while we hold the monitor, increasing the length of the critical 515 // section. Amdahl's parallel speedup law comes vividly into play. 516 // 517 // Another option might be to aggregate the events (thread local or 518 // per-monitor aggregation) and defer reporting until a more opportune 519 // time -- such as next time some thread encounters contention but has 520 // yet to acquire the lock. While spinning that thread could 521 // spinning we could increment JVMStat counters, etc. 522 523 DTRACE_MONITOR_PROBE(contended__entered, this, object(), current); 524 if (JvmtiExport::should_post_monitor_contended_entered()) { 525 JvmtiExport::post_monitor_contended_entered(current, this); 526 527 // The current thread already owns the monitor and is not going to 528 // call park() for the remainder of the monitor enter protocol. So 529 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 530 // event handler consumed an unpark() issued by the thread that 531 // just exited the monitor. 532 } 533 if (event.should_commit()) { 534 event.set_previousOwner(_previous_owner_tid); 535 event.commit(); 536 } 537 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 538 return true; 539 } 540 541 // Caveat: TryLock() is not necessarily serializing if it returns failure. 542 // Callers must compensate as needed. 543 544 ObjectMonitor::TryLockResult ObjectMonitor::TryLock(JavaThread* current) { 545 void* own = owner_raw(); 546 if (own != nullptr) return TryLockResult::HasOwner; 547 if (try_set_owner_from(nullptr, current) == nullptr) { 548 assert(_recursions == 0, "invariant"); 549 return TryLockResult::Success; 550 } 551 // The lock had been free momentarily, but we lost the race to the lock. 552 // Interference -- the CAS failed. 553 // We can either return -1 or retry. 554 // Retry doesn't make as much sense because the lock was just acquired. 555 return TryLockResult::Interference; 556 } 557 558 // Deflate the specified ObjectMonitor if not in-use. Returns true if it 559 // was deflated and false otherwise. 560 // 561 // The async deflation protocol sets owner to DEFLATER_MARKER and 562 // makes contentions negative as signals to contending threads that 563 // an async deflation is in progress. There are a number of checks 564 // as part of the protocol to make sure that the calling thread has 565 // not lost the race to a contending thread. 566 // 567 // The ObjectMonitor has been successfully async deflated when: 568 // (contentions < 0) 569 // Contending threads that see that condition know to retry their operation. 570 // 571 bool ObjectMonitor::deflate_monitor() { 572 if (is_busy()) { 573 // Easy checks are first - the ObjectMonitor is busy so no deflation. 574 return false; 575 } 576 577 const oop obj = object_peek(); 578 579 if (obj == nullptr) { 580 // If the object died, we can recycle the monitor without racing with 581 // Java threads. The GC already broke the association with the object. 582 set_owner_from_raw(nullptr, DEFLATER_MARKER); 583 assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions()); 584 _contentions = INT_MIN; // minimum negative int 585 } else { 586 // Attempt async deflation protocol. 587 588 // Set a null owner to DEFLATER_MARKER to force any contending thread 589 // through the slow path. This is just the first part of the async 590 // deflation dance. 591 if (try_set_owner_from_raw(nullptr, DEFLATER_MARKER) != nullptr) { 592 // The owner field is no longer null so we lost the race since the 593 // ObjectMonitor is now busy. 594 return false; 595 } 596 597 if (contentions() > 0 || _waiters != 0) { 598 // Another thread has raced to enter the ObjectMonitor after 599 // is_busy() above or has already entered and waited on 600 // it which makes it busy so no deflation. Restore owner to 601 // null if it is still DEFLATER_MARKER. 602 if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) { 603 // Deferred decrement for the JT EnterI() that cancelled the async deflation. 604 add_to_contentions(-1); 605 } 606 return false; 607 } 608 609 // Make a zero contentions field negative to force any contending threads 610 // to retry. This is the second part of the async deflation dance. 611 if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) { 612 // Contentions was no longer 0 so we lost the race since the 613 // ObjectMonitor is now busy. Restore owner to null if it is 614 // still DEFLATER_MARKER: 615 if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) { 616 // Deferred decrement for the JT EnterI() that cancelled the async deflation. 617 add_to_contentions(-1); 618 } 619 return false; 620 } 621 } 622 623 // Sanity checks for the races: 624 guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker"); 625 guarantee(contentions() < 0, "must be negative: contentions=%d", 626 contentions()); 627 guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters); 628 guarantee(_cxq == nullptr, "must be no contending threads: cxq=" 629 INTPTR_FORMAT, p2i(_cxq)); 630 guarantee(_EntryList == nullptr, 631 "must be no entering threads: EntryList=" INTPTR_FORMAT, 632 p2i(_EntryList)); 633 634 if (obj != nullptr) { 635 if (log_is_enabled(Trace, monitorinflation)) { 636 ResourceMark rm; 637 log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT 638 ", mark=" INTPTR_FORMAT ", type='%s'", 639 p2i(obj), obj->mark().value(), 640 obj->klass()->external_name()); 641 } 642 643 // Install the old mark word if nobody else has already done it. 644 install_displaced_markword_in_object(obj); 645 } 646 647 // We leave owner == DEFLATER_MARKER and contentions < 0 648 // to force any racing threads to retry. 649 return true; // Success, ObjectMonitor has been deflated. 650 } 651 652 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 653 // into the header of the object associated with the monitor. This 654 // idempotent method is called by a thread that is deflating a 655 // monitor and by other threads that have detected a race with the 656 // deflation process. 657 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 658 // This function must only be called when (owner == DEFLATER_MARKER 659 // && contentions <= 0), but we can't guarantee that here because 660 // those values could change when the ObjectMonitor gets moved from 661 // the global free list to a per-thread free list. 662 663 guarantee(obj != nullptr, "must be non-null"); 664 665 // Separate loads in is_being_async_deflated(), which is almost always 666 // called before this function, from the load of dmw/header below. 667 668 // _contentions and dmw/header may get written by different threads. 669 // Make sure to observe them in the same order when having several observers. 670 OrderAccess::loadload_for_IRIW(); 671 672 const oop l_object = object_peek(); 673 if (l_object == nullptr) { 674 // ObjectMonitor's object ref has already been cleared by async 675 // deflation or GC so we're done here. 676 return; 677 } 678 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 679 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 680 681 markWord dmw = header(); 682 // The dmw has to be neutral (not null, not locked and not marked). 683 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 684 685 // Install displaced mark word if the object's header still points 686 // to this ObjectMonitor. More than one racing caller to this function 687 // can rarely reach this point, but only one can win. 688 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 689 if (res != markWord::encode(this)) { 690 // This should be rare so log at the Info level when it happens. 691 log_info(monitorinflation)("install_displaced_markword_in_object: " 692 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 693 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 694 dmw.value(), markWord::encode(this).value(), 695 res.value()); 696 } 697 698 // Note: It does not matter which thread restored the header/dmw 699 // into the object's header. The thread deflating the monitor just 700 // wanted the object's header restored and it is. The threads that 701 // detected a race with the deflation process also wanted the 702 // object's header restored before they retry their operation and 703 // because it is restored they will only retry once. 704 } 705 706 // Convert the fields used by is_busy() to a string that can be 707 // used for diagnostic output. 708 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 709 ss->print("is_busy: waiters=%d" 710 ", contentions=%d" 711 ", owner=" INTPTR_FORMAT 712 ", cxq=" PTR_FORMAT 713 ", EntryList=" PTR_FORMAT, 714 _waiters, 715 (contentions() > 0 ? contentions() : 0), 716 owner_is_DEFLATER_MARKER() 717 // We report null instead of DEFLATER_MARKER here because is_busy() 718 // ignores DEFLATER_MARKER values. 719 ? p2i(nullptr) 720 : p2i(owner_raw()), 721 p2i(_cxq), 722 p2i(_EntryList)); 723 return ss->base(); 724 } 725 726 #define MAX_RECHECK_INTERVAL 1000 727 728 void ObjectMonitor::EnterI(JavaThread* current) { 729 assert(current->thread_state() == _thread_blocked, "invariant"); 730 731 // Try the lock - TATAS 732 if (TryLock(current) == TryLockResult::Success) { 733 assert(_succ != current, "invariant"); 734 assert(owner_raw() == owner_for(current), "invariant"); 735 assert(_Responsible != current, "invariant"); 736 return; 737 } 738 739 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 740 // Cancelled the in-progress async deflation by changing owner from 741 // DEFLATER_MARKER to current. As part of the contended enter protocol, 742 // contentions was incremented to a positive value before EnterI() 743 // was called and that prevents the deflater thread from winning the 744 // last part of the 2-part async deflation protocol. After EnterI() 745 // returns to enter(), contentions is decremented because the caller 746 // now owns the monitor. We bump contentions an extra time here to 747 // prevent the deflater thread from winning the last part of the 748 // 2-part async deflation protocol after the regular decrement 749 // occurs in enter(). The deflater thread will decrement contentions 750 // after it recognizes that the async deflation was cancelled. 751 add_to_contentions(1); 752 assert(_succ != current, "invariant"); 753 assert(_Responsible != current, "invariant"); 754 return; 755 } 756 757 assert(InitDone, "Unexpectedly not initialized"); 758 759 // We try one round of spinning *before* enqueueing current. 760 // 761 // If the _owner is ready but OFFPROC we could use a YieldTo() 762 // operation to donate the remainder of this thread's quantum 763 // to the owner. This has subtle but beneficial affinity 764 // effects. 765 766 if (TrySpin(current)) { 767 assert(owner_raw() == owner_for(current), "invariant"); 768 assert(_succ != current, "invariant"); 769 assert(_Responsible != current, "invariant"); 770 return; 771 } 772 773 // The Spin failed -- Enqueue and park the thread ... 774 assert(_succ != current, "invariant"); 775 assert(owner_raw() != owner_for(current), "invariant"); 776 assert(_Responsible != current, "invariant"); 777 778 // Enqueue "current" on ObjectMonitor's _cxq. 779 // 780 // Node acts as a proxy for current. 781 // As an aside, if were to ever rewrite the synchronization code mostly 782 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 783 // Java objects. This would avoid awkward lifecycle and liveness issues, 784 // as well as eliminate a subset of ABA issues. 785 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 786 787 ObjectWaiter node(current); 788 current->_ParkEvent->reset(); 789 node._prev = (ObjectWaiter*) 0xBAD; 790 node.TState = ObjectWaiter::TS_CXQ; 791 792 // Push "current" onto the front of the _cxq. 793 // Once on cxq/EntryList, current stays on-queue until it acquires the lock. 794 // Note that spinning tends to reduce the rate at which threads 795 // enqueue and dequeue on EntryList|cxq. 796 ObjectWaiter* nxt; 797 for (;;) { 798 node._next = nxt = _cxq; 799 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 800 801 // Interference - the CAS failed because _cxq changed. Just retry. 802 // As an optional optimization we retry the lock. 803 if (TryLock(current) == TryLockResult::Success) { 804 assert(_succ != current, "invariant"); 805 assert(owner_raw() == owner_for(current), "invariant"); 806 assert(_Responsible != current, "invariant"); 807 return; 808 } 809 } 810 811 // Check for cxq|EntryList edge transition to non-null. This indicates 812 // the onset of contention. While contention persists exiting threads 813 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 814 // operations revert to the faster 1-0 mode. This enter operation may interleave 815 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 816 // arrange for one of the contending thread to use a timed park() operations 817 // to detect and recover from the race. (Stranding is form of progress failure 818 // where the monitor is unlocked but all the contending threads remain parked). 819 // That is, at least one of the contended threads will periodically poll _owner. 820 // One of the contending threads will become the designated "Responsible" thread. 821 // The Responsible thread uses a timed park instead of a normal indefinite park 822 // operation -- it periodically wakes and checks for and recovers from potential 823 // strandings admitted by 1-0 exit operations. We need at most one Responsible 824 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 825 // be responsible for a monitor. 826 // 827 // Currently, one of the contended threads takes on the added role of "Responsible". 828 // A viable alternative would be to use a dedicated "stranding checker" thread 829 // that periodically iterated over all the threads (or active monitors) and unparked 830 // successors where there was risk of stranding. This would help eliminate the 831 // timer scalability issues we see on some platforms as we'd only have one thread 832 // -- the checker -- parked on a timer. 833 834 if (nxt == nullptr && _EntryList == nullptr) { 835 // Try to assume the role of responsible thread for the monitor. 836 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=current } 837 Atomic::replace_if_null(&_Responsible, current); 838 } 839 840 // The lock might have been released while this thread was occupied queueing 841 // itself onto _cxq. To close the race and avoid "stranding" and 842 // progress-liveness failure we must resample-retry _owner before parking. 843 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 844 // In this case the ST-MEMBAR is accomplished with CAS(). 845 // 846 // TODO: Defer all thread state transitions until park-time. 847 // Since state transitions are heavy and inefficient we'd like 848 // to defer the state transitions until absolutely necessary, 849 // and in doing so avoid some transitions ... 850 851 int nWakeups = 0; 852 int recheckInterval = 1; 853 bool do_timed_parked = false; 854 855 ContinuationEntry* ce = current->last_continuation(); 856 if (ce != nullptr && ce->is_virtual_thread()) { 857 do_timed_parked = true; 858 } 859 860 for (;;) { 861 862 if (TryLock(current) == TryLockResult::Success) { 863 break; 864 } 865 assert(owner_raw() != owner_for(current), "invariant"); 866 867 // park self 868 if (_Responsible == current || do_timed_parked) { 869 current->_ParkEvent->park((jlong) recheckInterval); 870 // Increase the recheckInterval, but clamp the value. 871 recheckInterval *= 8; 872 if (recheckInterval > MAX_RECHECK_INTERVAL) { 873 recheckInterval = MAX_RECHECK_INTERVAL; 874 } 875 } else { 876 current->_ParkEvent->park(); 877 } 878 879 if (TryLock(current) == TryLockResult::Success) { 880 break; 881 } 882 883 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 884 // Cancelled the in-progress async deflation by changing owner from 885 // DEFLATER_MARKER to current. As part of the contended enter protocol, 886 // contentions was incremented to a positive value before EnterI() 887 // was called and that prevents the deflater thread from winning the 888 // last part of the 2-part async deflation protocol. After EnterI() 889 // returns to enter(), contentions is decremented because the caller 890 // now owns the monitor. We bump contentions an extra time here to 891 // prevent the deflater thread from winning the last part of the 892 // 2-part async deflation protocol after the regular decrement 893 // occurs in enter(). The deflater thread will decrement contentions 894 // after it recognizes that the async deflation was cancelled. 895 add_to_contentions(1); 896 break; 897 } 898 899 // The lock is still contested. 900 // Keep a tally of the # of futile wakeups. 901 // Note that the counter is not protected by a lock or updated by atomics. 902 // That is by design - we trade "lossy" counters which are exposed to 903 // races during updates for a lower probe effect. 904 905 // This PerfData object can be used in parallel with a safepoint. 906 // See the work around in PerfDataManager::destroy(). 907 OM_PERFDATA_OP(FutileWakeups, inc()); 908 ++nWakeups; 909 910 // Assuming this is not a spurious wakeup we'll normally find _succ == current. 911 // We can defer clearing _succ until after the spin completes 912 // TrySpin() must tolerate being called with _succ == current. 913 // Try yet another round of adaptive spinning. 914 if (TrySpin(current)) { 915 break; 916 } 917 918 // We can find that we were unpark()ed and redesignated _succ while 919 // we were spinning. That's harmless. If we iterate and call park(), 920 // park() will consume the event and return immediately and we'll 921 // just spin again. This pattern can repeat, leaving _succ to simply 922 // spin on a CPU. 923 924 if (_succ == current) _succ = nullptr; 925 926 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 927 OrderAccess::fence(); 928 } 929 930 // Egress : 931 // current has acquired the lock -- Unlink current from the cxq or EntryList. 932 // Normally we'll find current on the EntryList . 933 // From the perspective of the lock owner (this thread), the 934 // EntryList is stable and cxq is prepend-only. 935 // The head of cxq is volatile but the interior is stable. 936 // In addition, current.TState is stable. 937 938 assert(owner_raw() == owner_for(current), "invariant"); 939 940 UnlinkAfterAcquire(current, &node); 941 if (_succ == current) _succ = nullptr; 942 943 assert(_succ != current, "invariant"); 944 if (_Responsible == current) { 945 _Responsible = nullptr; 946 OrderAccess::fence(); // Dekker pivot-point 947 948 // We may leave threads on cxq|EntryList without a designated 949 // "Responsible" thread. This is benign. When this thread subsequently 950 // exits the monitor it can "see" such preexisting "old" threads -- 951 // threads that arrived on the cxq|EntryList before the fence, above -- 952 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 953 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 954 // non-null and elect a new "Responsible" timer thread. 955 // 956 // This thread executes: 957 // ST Responsible=null; MEMBAR (in enter epilogue - here) 958 // LD cxq|EntryList (in subsequent exit) 959 // 960 // Entering threads in the slow/contended path execute: 961 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 962 // The (ST cxq; MEMBAR) is accomplished with CAS(). 963 // 964 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 965 // exit operation from floating above the ST Responsible=null. 966 } 967 968 // We've acquired ownership with CAS(). 969 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 970 // But since the CAS() this thread may have also stored into _succ, 971 // EntryList, cxq or Responsible. These meta-data updates must be 972 // visible __before this thread subsequently drops the lock. 973 // Consider what could occur if we didn't enforce this constraint -- 974 // STs to monitor meta-data and user-data could reorder with (become 975 // visible after) the ST in exit that drops ownership of the lock. 976 // Some other thread could then acquire the lock, but observe inconsistent 977 // or old monitor meta-data and heap data. That violates the JMM. 978 // To that end, the 1-0 exit() operation must have at least STST|LDST 979 // "release" barrier semantics. Specifically, there must be at least a 980 // STST|LDST barrier in exit() before the ST of null into _owner that drops 981 // the lock. The barrier ensures that changes to monitor meta-data and data 982 // protected by the lock will be visible before we release the lock, and 983 // therefore before some other thread (CPU) has a chance to acquire the lock. 984 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 985 // 986 // Critically, any prior STs to _succ or EntryList must be visible before 987 // the ST of null into _owner in the *subsequent* (following) corresponding 988 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 989 // execute a serializing instruction. 990 991 return; 992 } 993 994 bool ObjectMonitor::HandlePreemptedVThread(JavaThread* current) { 995 // Either because we acquire the lock below or because we will preempt the 996 // vthread clear the _current_pending_monitor field from the current JavaThread. 997 current->set_current_pending_monitor(nullptr); 998 999 // Try once more after freezing the continuation. 1000 if (TryLock(current) == TryLockResult::Success) { 1001 assert(owner_raw() == owner_for(current), "invariant"); 1002 assert(_succ != current, "invariant"); 1003 assert(_Responsible != current, "invariant"); 1004 current->set_preemption_cancelled(true); 1005 add_to_contentions(-1); 1006 return true; 1007 } 1008 1009 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 1010 // Cancelled the in-progress async deflation by changing owner from 1011 // DEFLATER_MARKER to current. As part of the contended enter protocol, 1012 // contentions was incremented to a positive value before this call to 1013 // HandlePreemptedVThread(). We avoid decrementing contentions to 1014 // prevent the deflater thread from winning the last part of the 1015 // 2-part async deflation protocol. The deflater thread will decrement 1016 // contentions after it recognizes that the async deflation was cancelled. 1017 assert(_succ != current, "invariant"); 1018 assert(_Responsible != current, "invariant"); 1019 current->set_preemption_cancelled(true); 1020 return true; 1021 } 1022 1023 oop vthread = current->vthread(); 1024 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1025 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1026 1027 ObjectWaiter* node = new ObjectWaiter(vthread); 1028 node->_prev = (ObjectWaiter*) 0xBAD; 1029 node->TState = ObjectWaiter::TS_CXQ; 1030 1031 // Push node associated with vthread onto the front of the _cxq. 1032 ObjectWaiter* nxt; 1033 for (;;) { 1034 node->_next = nxt = _cxq; 1035 if (Atomic::cmpxchg(&_cxq, nxt, node) == nxt) break; 1036 1037 // Interference - the CAS failed because _cxq changed. Just retry. 1038 // As an optional optimization we retry the lock. 1039 if (TryLock(current) == TryLockResult::Success) { 1040 assert(owner_raw() == owner_for(current), "invariant"); 1041 assert(_succ != current, "invariant"); 1042 assert(_Responsible != current, "invariant"); 1043 current->set_preemption_cancelled(true); 1044 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING); 1045 add_to_contentions(-1); 1046 delete node; 1047 return true; 1048 } 1049 } 1050 1051 // We have to try once more since owner could have exited monitor and checked 1052 // _cxq before we added the node to the queue. 1053 if (TryLock(current) == TryLockResult::Success) { 1054 assert(owner_raw() == owner_for(current), "invariant"); 1055 assert(_Responsible != current, "invariant"); 1056 current->set_preemption_cancelled(true); 1057 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::RUNNING); 1058 UnlinkAfterAcquire(current, node, vthread); 1059 delete node; 1060 if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr; 1061 add_to_contentions(-1); 1062 return true; 1063 } 1064 1065 if (nxt == nullptr && _EntryList == nullptr) { 1066 // The C2 unlock() fast path first checks if _cxq and _EntryList are empty and 1067 // if they are it just clears the _owner field. Since we always run the risk of 1068 // having that check happening before we added the node to _cxq and the release 1069 // of the monitor happening after the last TryLock attempt we need to do something 1070 // to avoid stranding. We set the _Responsible field which results in a timed-wait. 1071 if (Atomic::replace_if_null(&_Responsible, (JavaThread*)java_lang_Thread::thread_id(vthread))) { 1072 java_lang_VirtualThread::set_recheckInterval(vthread, 1); 1073 } 1074 } 1075 1076 return false; 1077 } 1078 1079 // ReenterI() is a specialized inline form of the latter half of the 1080 // contended slow-path from EnterI(). We use ReenterI() only for 1081 // monitor reentry in wait(). 1082 // 1083 // In the future we should reconcile EnterI() and ReenterI(). 1084 1085 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) { 1086 assert(current != nullptr, "invariant"); 1087 assert(currentNode != nullptr, "invariant"); 1088 assert(currentNode->_thread == current, "invariant"); 1089 assert(_waiters > 0, "invariant"); 1090 assert(object()->mark() == markWord::encode(this), "invariant"); 1091 1092 assert(current->thread_state() != _thread_blocked, "invariant"); 1093 1094 int nWakeups = 0; 1095 for (;;) { 1096 ObjectWaiter::TStates v = currentNode->TState; 1097 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1098 assert(owner_raw() != owner_for(current), "invariant"); 1099 1100 if (TrySpin(current)) { 1101 break; 1102 } 1103 1104 { 1105 OSThreadContendState osts(current->osthread()); 1106 1107 assert(current->thread_state() == _thread_in_vm, "invariant"); 1108 1109 { 1110 ClearSuccOnSuspend csos(this); 1111 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1112 current->_ParkEvent->park(); 1113 } 1114 } 1115 1116 // Try again, but just so we distinguish between futile wakeups and 1117 // successful wakeups. The following test isn't algorithmically 1118 // necessary, but it helps us maintain sensible statistics. 1119 if (TryLock(current) == TryLockResult::Success) { 1120 break; 1121 } 1122 1123 // The lock is still contested. 1124 // Keep a tally of the # of futile wakeups. 1125 // Note that the counter is not protected by a lock or updated by atomics. 1126 // That is by design - we trade "lossy" counters which are exposed to 1127 // races during updates for a lower probe effect. 1128 ++nWakeups; 1129 1130 // Assuming this is not a spurious wakeup we'll normally 1131 // find that _succ == current. 1132 if (_succ == current) _succ = nullptr; 1133 1134 // Invariant: after clearing _succ a contending thread 1135 // *must* retry _owner before parking. 1136 OrderAccess::fence(); 1137 1138 // This PerfData object can be used in parallel with a safepoint. 1139 // See the work around in PerfDataManager::destroy(). 1140 OM_PERFDATA_OP(FutileWakeups, inc()); 1141 } 1142 1143 // current has acquired the lock -- Unlink current from the cxq or EntryList . 1144 // Normally we'll find current on the EntryList. 1145 // Unlinking from the EntryList is constant-time and atomic-free. 1146 // From the perspective of the lock owner (this thread), the 1147 // EntryList is stable and cxq is prepend-only. 1148 // The head of cxq is volatile but the interior is stable. 1149 // In addition, current.TState is stable. 1150 1151 assert(owner_raw() == owner_for(current), "invariant"); 1152 assert(object()->mark() == markWord::encode(this), "invariant"); 1153 UnlinkAfterAcquire(current, currentNode); 1154 if (_succ == current) _succ = nullptr; 1155 assert(_succ != current, "invariant"); 1156 currentNode->TState = ObjectWaiter::TS_RUN; 1157 OrderAccess::fence(); // see comments at the end of EnterI() 1158 } 1159 1160 void ObjectMonitor::redo_enter(JavaThread* current) { 1161 assert(java_lang_VirtualThread::state(current->vthread()) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1162 assert(current->is_in_VTMS_transition(), "must be"); 1163 1164 if (TryLock(current) == TryLockResult::Success) { 1165 VThreadEpilog(current); 1166 return; 1167 } 1168 1169 oop vthread = current->vthread(); 1170 if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr; 1171 1172 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 1173 OrderAccess::fence(); 1174 1175 if (TryLock(current) == TryLockResult::Success) { 1176 assert(owner_raw() == owner_for(current), "invariant"); 1177 VThreadEpilog(current); 1178 return; 1179 } 1180 1181 // Fast preemption. The JT will read this variable on return to the 1182 // monitorenter_redo stub and will just remove enterSpecial frame 1183 // from the stack and return to Continuation.run() 1184 current->set_preempting(true); 1185 1186 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1187 if (_Responsible == (JavaThread*)java_lang_Thread::thread_id(vthread)) { 1188 int recheckInterval = java_lang_VirtualThread::recheckInterval(vthread); 1189 assert(recheckInterval >= 1 && recheckInterval <= 6, "invariant"); 1190 if (recheckInterval < 6) { 1191 recheckInterval++; 1192 java_lang_VirtualThread::set_recheckInterval(vthread, recheckInterval); 1193 } 1194 } else if (java_lang_VirtualThread::recheckInterval(vthread) > 0) { 1195 // No need to do timed park anymore 1196 java_lang_VirtualThread::set_recheckInterval(vthread, 0); 1197 } 1198 } 1199 1200 void ObjectMonitor::VThreadEpilog(JavaThread* current) { 1201 assert(owner_raw() == owner_for(current), "invariant"); 1202 add_to_contentions(-1); 1203 1204 oop vthread = current->vthread(); 1205 if (java_lang_VirtualThread::recheckInterval(vthread) > 0) { 1206 java_lang_VirtualThread::set_recheckInterval(vthread, 0); 1207 } 1208 int64_t threadid = java_lang_Thread::thread_id(vthread); 1209 if (_succ == (JavaThread*)threadid) _succ = nullptr; 1210 if (_Responsible == (JavaThread*)threadid) { 1211 _Responsible = nullptr; 1212 OrderAccess::fence(); // Dekker pivot-point 1213 } 1214 ObjectWaiter* node = LookupWaiter(threadid); 1215 UnlinkAfterAcquire(current, node, vthread); 1216 delete node; 1217 } 1218 1219 // By convention we unlink a contending thread from EntryList|cxq immediately 1220 // after the thread acquires the lock in ::enter(). Equally, we could defer 1221 // unlinking the thread until ::exit()-time. 1222 1223 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode, oop vthread) { 1224 assert(owner_raw() == owner_for(current), "invariant"); 1225 assert((currentNode->_thread == current) || (currentNode->_thread == nullptr && currentNode->vthread() == vthread), "invariant"); 1226 1227 if (currentNode->TState == ObjectWaiter::TS_ENTER) { 1228 // Normal case: remove current from the DLL EntryList . 1229 // This is a constant-time operation. 1230 ObjectWaiter* nxt = currentNode->_next; 1231 ObjectWaiter* prv = currentNode->_prev; 1232 if (nxt != nullptr) nxt->_prev = prv; 1233 if (prv != nullptr) prv->_next = nxt; 1234 if (currentNode == _EntryList) _EntryList = nxt; 1235 assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 1236 assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 1237 } else { 1238 assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 1239 // Inopportune interleaving -- current is still on the cxq. 1240 // This usually means the enqueue of self raced an exiting thread. 1241 // Normally we'll find current near the front of the cxq, so 1242 // dequeueing is typically fast. If needbe we can accelerate 1243 // this with some MCS/CHL-like bidirectional list hints and advisory 1244 // back-links so dequeueing from the interior will normally operate 1245 // in constant-time. 1246 // Dequeue current from either the head (with CAS) or from the interior 1247 // with a linear-time scan and normal non-atomic memory operations. 1248 // CONSIDER: if current is on the cxq then simply drain cxq into EntryList 1249 // and then unlink current from EntryList. We have to drain eventually, 1250 // so it might as well be now. 1251 1252 ObjectWaiter* v = _cxq; 1253 assert(v != nullptr, "invariant"); 1254 if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) { 1255 // The CAS above can fail from interference IFF a "RAT" arrived. 1256 // In that case current must be in the interior and can no longer be 1257 // at the head of cxq. 1258 if (v == currentNode) { 1259 assert(_cxq != v, "invariant"); 1260 v = _cxq; // CAS above failed - start scan at head of list 1261 } 1262 ObjectWaiter* p; 1263 ObjectWaiter* q = nullptr; 1264 for (p = v; p != nullptr && p != currentNode; p = p->_next) { 1265 q = p; 1266 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 1267 } 1268 assert(v != currentNode, "invariant"); 1269 assert(p == currentNode, "Node not found on cxq"); 1270 assert(p != _cxq, "invariant"); 1271 assert(q != nullptr, "invariant"); 1272 assert(q->_next == p, "invariant"); 1273 q->_next = p->_next; 1274 } 1275 } 1276 1277 #ifdef ASSERT 1278 // Diagnostic hygiene ... 1279 currentNode->_prev = (ObjectWaiter*) 0xBAD; 1280 currentNode->_next = (ObjectWaiter*) 0xBAD; 1281 currentNode->TState = ObjectWaiter::TS_RUN; 1282 #endif 1283 } 1284 1285 // Fix this. Save ObjectWaiter* when freezing. Or use hashtable. 1286 ObjectWaiter* ObjectMonitor::LookupWaiter(int64_t threadid) { 1287 ObjectWaiter* p; 1288 for (p = _EntryList; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {} 1289 if (p != nullptr) return p; 1290 for (p = _cxq; p != nullptr && (!p->is_vthread() || java_lang_Thread::thread_id(p->vthread()) != threadid); p = p->_next) {} 1291 assert(p != nullptr, "should be on either _cxq or _EntryList"); 1292 return p; 1293 } 1294 1295 // ----------------------------------------------------------------------------- 1296 // Exit support 1297 // 1298 // exit() 1299 // ~~~~~~ 1300 // Note that the collector can't reclaim the objectMonitor or deflate 1301 // the object out from underneath the thread calling ::exit() as the 1302 // thread calling ::exit() never transitions to a stable state. 1303 // This inhibits GC, which in turn inhibits asynchronous (and 1304 // inopportune) reclamation of "this". 1305 // 1306 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 1307 // There's one exception to the claim above, however. EnterI() can call 1308 // exit() to drop a lock if the acquirer has been externally suspended. 1309 // In that case exit() is called with _thread_state == _thread_blocked, 1310 // but the monitor's _contentions field is > 0, which inhibits reclamation. 1311 // 1312 // 1-0 exit 1313 // ~~~~~~~~ 1314 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 1315 // the fast-path operators have been optimized so the common ::exit() 1316 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 1317 // The code emitted by fast_unlock() elides the usual MEMBAR. This 1318 // greatly improves latency -- MEMBAR and CAS having considerable local 1319 // latency on modern processors -- but at the cost of "stranding". Absent the 1320 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 1321 // ::enter() path, resulting in the entering thread being stranding 1322 // and a progress-liveness failure. Stranding is extremely rare. 1323 // We use timers (timed park operations) & periodic polling to detect 1324 // and recover from stranding. Potentially stranded threads periodically 1325 // wake up and poll the lock. See the usage of the _Responsible variable. 1326 // 1327 // The CAS() in enter provides for safety and exclusion, while the CAS or 1328 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 1329 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 1330 // We detect and recover from stranding with timers. 1331 // 1332 // If a thread transiently strands it'll park until (a) another 1333 // thread acquires the lock and then drops the lock, at which time the 1334 // exiting thread will notice and unpark the stranded thread, or, (b) 1335 // the timer expires. If the lock is high traffic then the stranding latency 1336 // will be low due to (a). If the lock is low traffic then the odds of 1337 // stranding are lower, although the worst-case stranding latency 1338 // is longer. Critically, we don't want to put excessive load in the 1339 // platform's timer subsystem. We want to minimize both the timer injection 1340 // rate (timers created/sec) as well as the number of timers active at 1341 // any one time. (more precisely, we want to minimize timer-seconds, which is 1342 // the integral of the # of active timers at any instant over time). 1343 // Both impinge on OS scalability. Given that, at most one thread parked on 1344 // a monitor will use a timer. 1345 // 1346 // There is also the risk of a futile wake-up. If we drop the lock 1347 // another thread can reacquire the lock immediately, and we can 1348 // then wake a thread unnecessarily. This is benign, and we've 1349 // structured the code so the windows are short and the frequency 1350 // of such futile wakups is low. 1351 1352 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) { 1353 void* cur = owner_raw(); 1354 if (owner_for(current) != cur) { 1355 // Apparent unbalanced locking ... 1356 // Naively we'd like to throw IllegalMonitorStateException. 1357 // As a practical matter we can neither allocate nor throw an 1358 // exception as ::exit() can be called from leaf routines. 1359 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1360 // Upon deeper reflection, however, in a properly run JVM the only 1361 // way we should encounter this situation is in the presence of 1362 // unbalanced JNI locking. TODO: CheckJNICalls. 1363 // See also: CR4414101 1364 #ifdef ASSERT 1365 LogStreamHandle(Error, monitorinflation) lsh; 1366 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1367 " is exiting an ObjectMonitor it does not own.", p2i(current)); 1368 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 1369 print_debug_style_on(&lsh); 1370 assert(false, "Non-balanced monitor enter/exit!"); 1371 #endif 1372 return; 1373 } 1374 1375 if (_recursions != 0) { 1376 _recursions--; // this is simple recursive enter 1377 return; 1378 } 1379 1380 // Invariant: after setting Responsible=null an thread must execute 1381 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1382 _Responsible = nullptr; 1383 1384 #if INCLUDE_JFR 1385 // get the owner's thread id for the MonitorEnter event 1386 // if it is enabled and the thread isn't suspended 1387 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1388 _previous_owner_tid = JFR_THREAD_ID(current); 1389 } 1390 #endif 1391 1392 for (;;) { 1393 assert(owner_for(current) == owner_raw(), "invariant"); 1394 1395 // Drop the lock. 1396 // release semantics: prior loads and stores from within the critical section 1397 // must not float (reorder) past the following store that drops the lock. 1398 // Uses a storeload to separate release_store(owner) from the 1399 // successor check. The try_set_owner_from() below uses cmpxchg() so 1400 // we get the fence down there. 1401 release_clear_owner(current); 1402 OrderAccess::storeload(); 1403 1404 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) { 1405 return; 1406 } 1407 // Other threads are blocked trying to acquire the lock. 1408 1409 // Normally the exiting thread is responsible for ensuring succession, 1410 // but if other successors are ready or other entering threads are spinning 1411 // then this thread can simply store null into _owner and exit without 1412 // waking a successor. The existence of spinners or ready successors 1413 // guarantees proper succession (liveness). Responsibility passes to the 1414 // ready or running successors. The exiting thread delegates the duty. 1415 // More precisely, if a successor already exists this thread is absolved 1416 // of the responsibility of waking (unparking) one. 1417 // 1418 // The _succ variable is critical to reducing futile wakeup frequency. 1419 // _succ identifies the "heir presumptive" thread that has been made 1420 // ready (unparked) but that has not yet run. We need only one such 1421 // successor thread to guarantee progress. 1422 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1423 // section 3.3 "Futile Wakeup Throttling" for details. 1424 // 1425 // Note that spinners in Enter() also set _succ non-null. 1426 // In the current implementation spinners opportunistically set 1427 // _succ so that exiting threads might avoid waking a successor. 1428 // Another less appealing alternative would be for the exiting thread 1429 // to drop the lock and then spin briefly to see if a spinner managed 1430 // to acquire the lock. If so, the exiting thread could exit 1431 // immediately without waking a successor, otherwise the exiting 1432 // thread would need to dequeue and wake a successor. 1433 // (Note that we'd need to make the post-drop spin short, but no 1434 // shorter than the worst-case round-trip cache-line migration time. 1435 // The dropped lock needs to become visible to the spinner, and then 1436 // the acquisition of the lock by the spinner must become visible to 1437 // the exiting thread). 1438 1439 // It appears that an heir-presumptive (successor) must be made ready. 1440 // Only the current lock owner can manipulate the EntryList or 1441 // drain _cxq, so we need to reacquire the lock. If we fail 1442 // to reacquire the lock the responsibility for ensuring succession 1443 // falls to the new owner. 1444 // 1445 if (try_set_owner_from(nullptr, current) != nullptr) { 1446 return; 1447 } 1448 1449 guarantee(owner_raw() == owner_for(current), "invariant"); 1450 1451 ObjectWaiter* w = nullptr; 1452 1453 w = _EntryList; 1454 if (w != nullptr) { 1455 // I'd like to write: guarantee (w->_thread != current). 1456 // But in practice an exiting thread may find itself on the EntryList. 1457 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1458 // then calls exit(). Exit release the lock by setting O._owner to null. 1459 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1460 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1461 // release the lock "O". T2 resumes immediately after the ST of null into 1462 // _owner, above. T2 notices that the EntryList is populated, so it 1463 // reacquires the lock and then finds itself on the EntryList. 1464 // Given all that, we have to tolerate the circumstance where "w" is 1465 // associated with current. 1466 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1467 ExitEpilog(current, w); 1468 return; 1469 } 1470 1471 // If we find that both _cxq and EntryList are null then just 1472 // re-run the exit protocol from the top. 1473 w = _cxq; 1474 if (w == nullptr) continue; 1475 1476 // Drain _cxq into EntryList - bulk transfer. 1477 // First, detach _cxq. 1478 // The following loop is tantamount to: w = swap(&cxq, nullptr) 1479 for (;;) { 1480 assert(w != nullptr, "Invariant"); 1481 ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr); 1482 if (u == w) break; 1483 w = u; 1484 } 1485 1486 assert(w != nullptr, "invariant"); 1487 assert(_EntryList == nullptr, "invariant"); 1488 1489 // Convert the LIFO SLL anchored by _cxq into a DLL. 1490 // The list reorganization step operates in O(LENGTH(w)) time. 1491 // It's critical that this step operate quickly as 1492 // "current" still holds the outer-lock, restricting parallelism 1493 // and effectively lengthening the critical section. 1494 // Invariant: s chases t chases u. 1495 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1496 // we have faster access to the tail. 1497 1498 _EntryList = w; 1499 ObjectWaiter* q = nullptr; 1500 ObjectWaiter* p; 1501 for (p = w; p != nullptr; p = p->_next) { 1502 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1503 p->TState = ObjectWaiter::TS_ENTER; 1504 p->_prev = q; 1505 q = p; 1506 } 1507 1508 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr 1509 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1510 1511 // See if we can abdicate to a spinner instead of waking a thread. 1512 // A primary goal of the implementation is to reduce the 1513 // context-switch rate. 1514 if (_succ != nullptr) continue; 1515 1516 w = _EntryList; 1517 if (w != nullptr) { 1518 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1519 ExitEpilog(current, w); 1520 return; 1521 } 1522 } 1523 } 1524 1525 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) { 1526 assert(owner_raw() == owner_for(current), "invariant"); 1527 1528 // Exit protocol: 1529 // 1. ST _succ = wakee 1530 // 2. membar #loadstore|#storestore; 1531 // 2. ST _owner = nullptr 1532 // 3. unpark(wakee) 1533 1534 oop vthread = nullptr; 1535 if (Wakee->_thread != nullptr) { 1536 // Platform thread case 1537 _succ = Wakee->_thread; 1538 } else { 1539 assert(Wakee->vthread() != nullptr, "invariant"); 1540 vthread = Wakee->vthread(); 1541 _succ = (JavaThread*)java_lang_Thread::thread_id(vthread); 1542 } 1543 ParkEvent * Trigger = Wakee->_event; 1544 1545 // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again. 1546 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1547 // out-of-scope (non-extant). 1548 Wakee = nullptr; 1549 1550 // Drop the lock. 1551 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1552 release_clear_owner(current); 1553 OrderAccess::fence(); 1554 1555 DTRACE_MONITOR_PROBE(contended__exit, this, object(), current); 1556 1557 if (vthread == nullptr) { 1558 // Platform thread case 1559 Trigger->unpark(); 1560 } else if (java_lang_VirtualThread::set_onWaitingList(vthread, _vthread_cxq_head)) { 1561 Trigger->unpark(); 1562 } 1563 1564 // Maintain stats and report events to JVMTI 1565 OM_PERFDATA_OP(Parks, inc()); 1566 } 1567 1568 // complete_exit exits a lock returning recursion count 1569 // complete_exit requires an inflated monitor 1570 // The _owner field is not always the Thread addr even with an 1571 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1572 // thread due to contention. 1573 intx ObjectMonitor::complete_exit(JavaThread* current) { 1574 assert(InitDone, "Unexpectedly not initialized"); 1575 1576 void* cur = owner_raw(); 1577 if (owner_for(current) != cur) { 1578 if (LockingMode == LM_LEGACY && is_stack_locker(current)) { 1579 assert(_recursions == 0, "internal state error"); 1580 set_owner_from_BasicLock(current); // Convert from BasicLock* to Thread*. 1581 _recursions = 0; 1582 } 1583 } 1584 1585 guarantee(owner_for(current) == owner_raw(), "complete_exit not owner"); 1586 intx save = _recursions; // record the old recursion count 1587 _recursions = 0; // set the recursion level to be 0 1588 exit(current); // exit the monitor 1589 guarantee(owner_raw() != owner_for(current), "invariant"); 1590 return save; 1591 } 1592 1593 // Checks that the current THREAD owns this monitor and causes an 1594 // immediate return if it doesn't. We don't use the CHECK macro 1595 // because we want the IMSE to be the only exception that is thrown 1596 // from the call site when false is returned. Any other pending 1597 // exception is ignored. 1598 #define CHECK_OWNER() \ 1599 do { \ 1600 if (!check_owner(THREAD)) { \ 1601 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1602 return; \ 1603 } \ 1604 } while (false) 1605 1606 // Returns true if the specified thread owns the ObjectMonitor. 1607 // Otherwise returns false and throws IllegalMonitorStateException 1608 // (IMSE). If there is a pending exception and the specified thread 1609 // is not the owner, that exception will be replaced by the IMSE. 1610 bool ObjectMonitor::check_owner(TRAPS) { 1611 JavaThread* current = THREAD; 1612 void* cur = owner_raw(); 1613 if (cur == owner_for(current)) { 1614 return true; 1615 } 1616 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1617 "current thread is not owner", false); 1618 } 1619 1620 static inline bool is_excluded(const Klass* monitor_klass) { 1621 assert(monitor_klass != nullptr, "invariant"); 1622 NOT_JFR_RETURN_(false); 1623 JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name();) 1624 } 1625 1626 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1627 ObjectMonitor* monitor, 1628 uint64_t notifier_tid, 1629 jlong timeout, 1630 bool timedout) { 1631 assert(event != nullptr, "invariant"); 1632 assert(monitor != nullptr, "invariant"); 1633 const Klass* monitor_klass = monitor->object()->klass(); 1634 if (is_excluded(monitor_klass)) { 1635 return; 1636 } 1637 event->set_monitorClass(monitor_klass); 1638 event->set_timeout(timeout); 1639 // Set an address that is 'unique enough', such that events close in 1640 // time and with the same address are likely (but not guaranteed) to 1641 // belong to the same object. 1642 event->set_address((uintptr_t)monitor); 1643 event->set_notifier(notifier_tid); 1644 event->set_timedOut(timedout); 1645 event->commit(); 1646 } 1647 1648 // ----------------------------------------------------------------------------- 1649 // Wait/Notify/NotifyAll 1650 // 1651 // Note: a subset of changes to ObjectMonitor::wait() 1652 // will need to be replicated in complete_exit 1653 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1654 JavaThread* current = THREAD; 1655 1656 assert(InitDone, "Unexpectedly not initialized"); 1657 1658 CHECK_OWNER(); // Throws IMSE if not owner. 1659 1660 EventJavaMonitorWait event; 1661 1662 // check for a pending interrupt 1663 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1664 // post monitor waited event. Note that this is past-tense, we are done waiting. 1665 if (JvmtiExport::should_post_monitor_waited()) { 1666 // Note: 'false' parameter is passed here because the 1667 // wait was not timed out due to thread interrupt. 1668 JvmtiExport::post_monitor_waited(current, this, false); 1669 1670 // In this short circuit of the monitor wait protocol, the 1671 // current thread never drops ownership of the monitor and 1672 // never gets added to the wait queue so the current thread 1673 // cannot be made the successor. This means that the 1674 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1675 // consume an unpark() meant for the ParkEvent associated with 1676 // this ObjectMonitor. 1677 } 1678 if (event.should_commit()) { 1679 post_monitor_wait_event(&event, this, 0, millis, false); 1680 } 1681 THROW(vmSymbols::java_lang_InterruptedException()); 1682 return; 1683 } 1684 1685 ContinuationEntry* ce = current->last_continuation(); 1686 if (ce != nullptr && ce->is_virtual_thread()) { 1687 const Klass* monitor_klass = object()->klass(); 1688 if (!is_excluded(monitor_klass)) { 1689 ResourceMark rm; 1690 char reason[256]; 1691 jio_snprintf(reason, sizeof reason, "Object.wait on object of klass %s", monitor_klass->external_name()); 1692 post_virtual_thread_pinned_event(current, reason); 1693 } 1694 } 1695 1696 current->set_current_waiting_monitor(this); 1697 1698 // create a node to be put into the queue 1699 // Critically, after we reset() the event but prior to park(), we must check 1700 // for a pending interrupt. 1701 ObjectWaiter node(current); 1702 node.TState = ObjectWaiter::TS_WAIT; 1703 current->_ParkEvent->reset(); 1704 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1705 1706 // Enter the waiting queue, which is a circular doubly linked list in this case 1707 // but it could be a priority queue or any data structure. 1708 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1709 // by the owner of the monitor *except* in the case where park() 1710 // returns because of a timeout of interrupt. Contention is exceptionally rare 1711 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1712 1713 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1714 AddWaiter(&node); 1715 Thread::SpinRelease(&_WaitSetLock); 1716 1717 _Responsible = nullptr; 1718 1719 intx save = _recursions; // record the old recursion count 1720 _waiters++; // increment the number of waiters 1721 _recursions = 0; // set the recursion level to be 1 1722 exit(current); // exit the monitor 1723 guarantee(owner_raw() != owner_for(current), "invariant"); 1724 1725 // The thread is on the WaitSet list - now park() it. 1726 // On MP systems it's conceivable that a brief spin before we park 1727 // could be profitable. 1728 // 1729 // TODO-FIXME: change the following logic to a loop of the form 1730 // while (!timeout && !interrupted && _notified == 0) park() 1731 1732 int ret = OS_OK; 1733 int WasNotified = 0; 1734 1735 // Need to check interrupt state whilst still _thread_in_vm 1736 bool interrupted = interruptible && current->is_interrupted(false); 1737 1738 { // State transition wrappers 1739 OSThread* osthread = current->osthread(); 1740 OSThreadWaitState osts(osthread, true); 1741 1742 assert(current->thread_state() == _thread_in_vm, "invariant"); 1743 1744 { 1745 ClearSuccOnSuspend csos(this); 1746 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1747 if (interrupted || HAS_PENDING_EXCEPTION) { 1748 // Intentionally empty 1749 } else if (node._notified == 0) { 1750 if (millis <= 0) { 1751 current->_ParkEvent->park(); 1752 } else { 1753 ret = current->_ParkEvent->park(millis); 1754 } 1755 } 1756 } 1757 1758 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1759 // from the WaitSet to the EntryList. 1760 // See if we need to remove Node from the WaitSet. 1761 // We use double-checked locking to avoid grabbing _WaitSetLock 1762 // if the thread is not on the wait queue. 1763 // 1764 // Note that we don't need a fence before the fetch of TState. 1765 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1766 // written by the is thread. (perhaps the fetch might even be satisfied 1767 // by a look-aside into the processor's own store buffer, although given 1768 // the length of the code path between the prior ST and this load that's 1769 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1770 // then we'll acquire the lock and then re-fetch a fresh TState value. 1771 // That is, we fail toward safety. 1772 1773 if (node.TState == ObjectWaiter::TS_WAIT) { 1774 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1775 if (node.TState == ObjectWaiter::TS_WAIT) { 1776 DequeueSpecificWaiter(&node); // unlink from WaitSet 1777 assert(node._notified == 0, "invariant"); 1778 node.TState = ObjectWaiter::TS_RUN; 1779 } 1780 Thread::SpinRelease(&_WaitSetLock); 1781 } 1782 1783 // The thread is now either on off-list (TS_RUN), 1784 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1785 // The Node's TState variable is stable from the perspective of this thread. 1786 // No other threads will asynchronously modify TState. 1787 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1788 OrderAccess::loadload(); 1789 if (_succ == current) _succ = nullptr; 1790 WasNotified = node._notified; 1791 1792 // Reentry phase -- reacquire the monitor. 1793 // re-enter contended monitor after object.wait(). 1794 // retain OBJECT_WAIT state until re-enter successfully completes 1795 // Thread state is thread_in_vm and oop access is again safe, 1796 // although the raw address of the object may have changed. 1797 // (Don't cache naked oops over safepoints, of course). 1798 1799 // post monitor waited event. Note that this is past-tense, we are done waiting. 1800 if (JvmtiExport::should_post_monitor_waited()) { 1801 JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT); 1802 1803 if (node._notified != 0 && _succ == current) { 1804 // In this part of the monitor wait-notify-reenter protocol it 1805 // is possible (and normal) for another thread to do a fastpath 1806 // monitor enter-exit while this thread is still trying to get 1807 // to the reenter portion of the protocol. 1808 // 1809 // The ObjectMonitor was notified and the current thread is 1810 // the successor which also means that an unpark() has already 1811 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1812 // consume the unpark() that was done when the successor was 1813 // set because the same ParkEvent is shared between Java 1814 // monitors and JVM/TI RawMonitors (for now). 1815 // 1816 // We redo the unpark() to ensure forward progress, i.e., we 1817 // don't want all pending threads hanging (parked) with none 1818 // entering the unlocked monitor. 1819 node._event->unpark(); 1820 } 1821 } 1822 1823 if (event.should_commit()) { 1824 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1825 } 1826 1827 OrderAccess::fence(); 1828 1829 assert(owner_raw() != owner_for(current), "invariant"); 1830 ObjectWaiter::TStates v = node.TState; 1831 if (v == ObjectWaiter::TS_RUN) { 1832 enter(current); 1833 } else { 1834 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1835 ReenterI(current, &node); 1836 node.wait_reenter_end(this); 1837 } 1838 1839 // current has reacquired the lock. 1840 // Lifecycle - the node representing current must not appear on any queues. 1841 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1842 // want residual elements associated with this thread left on any lists. 1843 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1844 assert(owner_raw() == owner_for(current), "invariant"); 1845 assert(_succ != current, "invariant"); 1846 } // OSThreadWaitState() 1847 1848 current->set_current_waiting_monitor(nullptr); 1849 1850 guarantee(_recursions == 0, "invariant"); 1851 int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current); 1852 _recursions = save // restore the old recursion count 1853 + relock_count; // increased by the deferred relock count 1854 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count(relock_count);) // Deopt never entered these counts. 1855 _waiters--; // decrement the number of waiters 1856 1857 // Verify a few postconditions 1858 assert(owner_raw() == owner_for(current), "invariant"); 1859 assert(_succ != current, "invariant"); 1860 assert(object()->mark() == markWord::encode(this), "invariant"); 1861 1862 // check if the notification happened 1863 if (!WasNotified) { 1864 // no, it could be timeout or Thread.interrupt() or both 1865 // check for interrupt event, otherwise it is timeout 1866 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1867 THROW(vmSymbols::java_lang_InterruptedException()); 1868 } 1869 } 1870 1871 // NOTE: Spurious wake up will be consider as timeout. 1872 // Monitor notify has precedence over thread interrupt. 1873 } 1874 1875 1876 // Consider: 1877 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1878 // then instead of transferring a thread from the WaitSet to the EntryList 1879 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1880 1881 void ObjectMonitor::INotify(JavaThread* current) { 1882 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1883 ObjectWaiter* iterator = DequeueWaiter(); 1884 if (iterator != nullptr) { 1885 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1886 guarantee(iterator->_notified == 0, "invariant"); 1887 // Disposition - what might we do with iterator ? 1888 // a. add it directly to the EntryList - either tail (policy == 1) 1889 // or head (policy == 0). 1890 // b. push it onto the front of the _cxq (policy == 2). 1891 // For now we use (b). 1892 1893 iterator->TState = ObjectWaiter::TS_ENTER; 1894 1895 iterator->_notified = 1; 1896 iterator->_notifier_tid = JFR_THREAD_ID(current); 1897 1898 ObjectWaiter* list = _EntryList; 1899 if (list != nullptr) { 1900 assert(list->_prev == nullptr, "invariant"); 1901 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1902 assert(list != iterator, "invariant"); 1903 } 1904 1905 // prepend to cxq 1906 if (list == nullptr) { 1907 iterator->_next = iterator->_prev = nullptr; 1908 _EntryList = iterator; 1909 } else { 1910 iterator->TState = ObjectWaiter::TS_CXQ; 1911 for (;;) { 1912 ObjectWaiter* front = _cxq; 1913 iterator->_next = front; 1914 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1915 break; 1916 } 1917 } 1918 } 1919 1920 // _WaitSetLock protects the wait queue, not the EntryList. We could 1921 // move the add-to-EntryList operation, above, outside the critical section 1922 // protected by _WaitSetLock. In practice that's not useful. With the 1923 // exception of wait() timeouts and interrupts the monitor owner 1924 // is the only thread that grabs _WaitSetLock. There's almost no contention 1925 // on _WaitSetLock so it's not profitable to reduce the length of the 1926 // critical section. 1927 1928 iterator->wait_reenter_begin(this); 1929 } 1930 Thread::SpinRelease(&_WaitSetLock); 1931 } 1932 1933 // Consider: a not-uncommon synchronization bug is to use notify() when 1934 // notifyAll() is more appropriate, potentially resulting in stranded 1935 // threads; this is one example of a lost wakeup. A useful diagnostic 1936 // option is to force all notify() operations to behave as notifyAll(). 1937 // 1938 // Note: We can also detect many such problems with a "minimum wait". 1939 // When the "minimum wait" is set to a small non-zero timeout value 1940 // and the program does not hang whereas it did absent "minimum wait", 1941 // that suggests a lost wakeup bug. 1942 1943 void ObjectMonitor::notify(TRAPS) { 1944 JavaThread* current = THREAD; 1945 CHECK_OWNER(); // Throws IMSE if not owner. 1946 if (_WaitSet == nullptr) { 1947 return; 1948 } 1949 DTRACE_MONITOR_PROBE(notify, this, object(), current); 1950 INotify(current); 1951 OM_PERFDATA_OP(Notifications, inc(1)); 1952 } 1953 1954 1955 // The current implementation of notifyAll() transfers the waiters one-at-a-time 1956 // from the waitset to the EntryList. This could be done more efficiently with a 1957 // single bulk transfer but in practice it's not time-critical. Beware too, 1958 // that in prepend-mode we invert the order of the waiters. Let's say that the 1959 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 1960 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 1961 1962 void ObjectMonitor::notifyAll(TRAPS) { 1963 JavaThread* current = THREAD; 1964 CHECK_OWNER(); // Throws IMSE if not owner. 1965 if (_WaitSet == nullptr) { 1966 return; 1967 } 1968 1969 DTRACE_MONITOR_PROBE(notifyAll, this, object(), current); 1970 int tally = 0; 1971 while (_WaitSet != nullptr) { 1972 tally++; 1973 INotify(current); 1974 } 1975 1976 OM_PERFDATA_OP(Notifications, inc(tally)); 1977 } 1978 1979 // ----------------------------------------------------------------------------- 1980 // Adaptive Spinning Support 1981 // 1982 // Adaptive spin-then-block - rational spinning 1983 // 1984 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1985 // algorithm. On high order SMP systems it would be better to start with 1986 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1987 // a contending thread could enqueue itself on the cxq and then spin locally 1988 // on a thread-specific variable such as its ParkEvent._Event flag. 1989 // That's left as an exercise for the reader. Note that global spinning is 1990 // not problematic on Niagara, as the L2 cache serves the interconnect and 1991 // has both low latency and massive bandwidth. 1992 // 1993 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 1994 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 1995 // (duration) or we can fix the count at approximately the duration of 1996 // a context switch and vary the frequency. Of course we could also 1997 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1998 // For a description of 'Adaptive spin-then-block mutual exclusion in 1999 // multi-threaded processing,' see U.S. Pat. No. 8046758. 2000 // 2001 // This implementation varies the duration "D", where D varies with 2002 // the success rate of recent spin attempts. (D is capped at approximately 2003 // length of a round-trip context switch). The success rate for recent 2004 // spin attempts is a good predictor of the success rate of future spin 2005 // attempts. The mechanism adapts automatically to varying critical 2006 // section length (lock modality), system load and degree of parallelism. 2007 // D is maintained per-monitor in _SpinDuration and is initialized 2008 // optimistically. Spin frequency is fixed at 100%. 2009 // 2010 // Note that _SpinDuration is volatile, but we update it without locks 2011 // or atomics. The code is designed so that _SpinDuration stays within 2012 // a reasonable range even in the presence of races. The arithmetic 2013 // operations on _SpinDuration are closed over the domain of legal values, 2014 // so at worst a race will install and older but still legal value. 2015 // At the very worst this introduces some apparent non-determinism. 2016 // We might spin when we shouldn't or vice-versa, but since the spin 2017 // count are relatively short, even in the worst case, the effect is harmless. 2018 // 2019 // Care must be taken that a low "D" value does not become an 2020 // an absorbing state. Transient spinning failures -- when spinning 2021 // is overall profitable -- should not cause the system to converge 2022 // on low "D" values. We want spinning to be stable and predictable 2023 // and fairly responsive to change and at the same time we don't want 2024 // it to oscillate, become metastable, be "too" non-deterministic, 2025 // or converge on or enter undesirable stable absorbing states. 2026 // 2027 // We implement a feedback-based control system -- using past behavior 2028 // to predict future behavior. We face two issues: (a) if the 2029 // input signal is random then the spin predictor won't provide optimal 2030 // results, and (b) if the signal frequency is too high then the control 2031 // system, which has some natural response lag, will "chase" the signal. 2032 // (b) can arise from multimodal lock hold times. Transient preemption 2033 // can also result in apparent bimodal lock hold times. 2034 // Although sub-optimal, neither condition is particularly harmful, as 2035 // in the worst-case we'll spin when we shouldn't or vice-versa. 2036 // The maximum spin duration is rather short so the failure modes aren't bad. 2037 // To be conservative, I've tuned the gain in system to bias toward 2038 // _not spinning. Relatedly, the system can sometimes enter a mode where it 2039 // "rings" or oscillates between spinning and not spinning. This happens 2040 // when spinning is just on the cusp of profitability, however, so the 2041 // situation is not dire. The state is benign -- there's no need to add 2042 // hysteresis control to damp the transition rate between spinning and 2043 // not spinning. 2044 2045 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool 2046 2047 static int Knob_Bonus = 100; // spin success bonus 2048 static int Knob_Penalty = 200; // spin failure penalty 2049 static int Knob_Poverty = 1000; 2050 static int Knob_FixedSpin = 0; 2051 static int Knob_PreSpin = 10; // 20-100 likely better, but it's not better in my testing. 2052 2053 inline static int adjust_up(int spin_duration) { 2054 int x = spin_duration; 2055 if (x < ObjectMonitor::Knob_SpinLimit) { 2056 if (x < Knob_Poverty) { 2057 x = Knob_Poverty; 2058 } 2059 return x + Knob_Bonus; 2060 } else { 2061 return spin_duration; 2062 } 2063 } 2064 2065 inline static int adjust_down(int spin_duration) { 2066 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2067 // AIMD is globally stable. 2068 int x = spin_duration; 2069 if (x > 0) { 2070 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2071 // This is globally sample and tends to damp the response. 2072 x -= Knob_Penalty; 2073 if (x < 0) { x = 0; } 2074 return x; 2075 } else { 2076 return spin_duration; 2077 } 2078 } 2079 2080 bool ObjectMonitor::short_fixed_spin(JavaThread* current, int spin_count, bool adapt) { 2081 for (int ctr = 0; ctr < spin_count; ctr++) { 2082 TryLockResult status = TryLock(current); 2083 if (status == TryLockResult::Success) { 2084 if (adapt) { 2085 _SpinDuration = adjust_up(_SpinDuration); 2086 } 2087 return true; 2088 } else if (status == TryLockResult::Interference) { 2089 break; 2090 } 2091 SpinPause(); 2092 } 2093 return false; 2094 } 2095 2096 // Spinning: Fixed frequency (100%), vary duration 2097 bool ObjectMonitor::TrySpin(JavaThread* current) { 2098 2099 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2100 int knob_fixed_spin = Knob_FixedSpin; // 0 (don't spin: default), 2000 good test 2101 if (knob_fixed_spin > 0) { 2102 return short_fixed_spin(current, knob_fixed_spin, false); 2103 } 2104 2105 // Admission control - verify preconditions for spinning 2106 // 2107 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2108 // becoming an absorbing state. Put another way, we spin briefly to 2109 // sample, just in case the system load, parallelism, contention, or lock 2110 // modality changed. 2111 2112 int knob_pre_spin = Knob_PreSpin; // 10 (default), 100, 1000 or 2000 2113 if (short_fixed_spin(current, knob_pre_spin, true)) { 2114 return true; 2115 } 2116 2117 // 2118 // Consider the following alternative: 2119 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2120 // spin attempt. "Periodically" might mean after a tally of 2121 // the # of failed spin attempts (or iterations) reaches some threshold. 2122 // This takes us into the realm of 1-out-of-N spinning, where we 2123 // hold the duration constant but vary the frequency. 2124 2125 int ctr = _SpinDuration; 2126 if (ctr <= 0) return false; 2127 2128 // We're good to spin ... spin ingress. 2129 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2130 // when preparing to LD...CAS _owner, etc and the CAS is likely 2131 // to succeed. 2132 if (_succ == nullptr) { 2133 _succ = current; 2134 } 2135 void* prv = nullptr; 2136 2137 // There are three ways to exit the following loop: 2138 // 1. A successful spin where this thread has acquired the lock. 2139 // 2. Spin failure with prejudice 2140 // 3. Spin failure without prejudice 2141 2142 while (--ctr >= 0) { 2143 2144 // Periodic polling -- Check for pending GC 2145 // Threads may spin while they're unsafe. 2146 // We don't want spinning threads to delay the JVM from reaching 2147 // a stop-the-world safepoint or to steal cycles from GC. 2148 // If we detect a pending safepoint we abort in order that 2149 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2150 // this thread, if safe, doesn't steal cycles from GC. 2151 // This is in keeping with the "no loitering in runtime" rule. 2152 // We periodically check to see if there's a safepoint pending. 2153 if ((ctr & 0xFF) == 0) { 2154 // Can't call SafepointMechanism::should_process() since that 2155 // might update the poll values and we could be in a thread_blocked 2156 // state here which is not allowed so just check the poll. 2157 if (SafepointMechanism::local_poll_armed(current)) { 2158 break; 2159 } 2160 SpinPause(); 2161 } 2162 2163 // Probe _owner with TATAS 2164 // If this thread observes the monitor transition or flicker 2165 // from locked to unlocked to locked, then the odds that this 2166 // thread will acquire the lock in this spin attempt go down 2167 // considerably. The same argument applies if the CAS fails 2168 // or if we observe _owner change from one non-null value to 2169 // another non-null value. In such cases we might abort 2170 // the spin without prejudice or apply a "penalty" to the 2171 // spin count-down variable "ctr", reducing it by 100, say. 2172 2173 void* ox = owner_raw(); 2174 if (ox == nullptr) { 2175 ox = try_set_owner_from(nullptr, current); 2176 if (ox == nullptr) { 2177 // The CAS succeeded -- this thread acquired ownership 2178 // Take care of some bookkeeping to exit spin state. 2179 if (_succ == current) { 2180 _succ = nullptr; 2181 } 2182 2183 // Increase _SpinDuration : 2184 // The spin was successful (profitable) so we tend toward 2185 // longer spin attempts in the future. 2186 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2187 // If we acquired the lock early in the spin cycle it 2188 // makes sense to increase _SpinDuration proportionally. 2189 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2190 _SpinDuration = adjust_up(_SpinDuration); 2191 return true; 2192 } 2193 2194 // The CAS failed ... we can take any of the following actions: 2195 // * penalize: ctr -= CASPenalty 2196 // * exit spin with prejudice -- abort without adapting spinner 2197 // * exit spin without prejudice. 2198 // * Since CAS is high-latency, retry again immediately. 2199 break; 2200 } 2201 2202 // Did lock ownership change hands ? 2203 if (ox != prv && prv != nullptr) { 2204 break; 2205 } 2206 prv = ox; 2207 2208 if (_succ == nullptr) { 2209 _succ = current; 2210 } 2211 } 2212 2213 // Spin failed with prejudice -- reduce _SpinDuration. 2214 if (ctr < 0) { 2215 _SpinDuration = adjust_down(_SpinDuration); 2216 } 2217 2218 if (_succ == current) { 2219 _succ = nullptr; 2220 // Invariant: after setting succ=null a contending thread 2221 // must recheck-retry _owner before parking. This usually happens 2222 // in the normal usage of TrySpin(), but it's safest 2223 // to make TrySpin() as foolproof as possible. 2224 OrderAccess::fence(); 2225 if (TryLock(current) == TryLockResult::Success) { 2226 return true; 2227 } 2228 } 2229 2230 return false; 2231 } 2232 2233 2234 // ----------------------------------------------------------------------------- 2235 // WaitSet management ... 2236 2237 ObjectWaiter::ObjectWaiter(JavaThread* current) { 2238 _next = nullptr; 2239 _prev = nullptr; 2240 _notified = 0; 2241 _notifier_tid = 0; 2242 TState = TS_RUN; 2243 _thread = current; 2244 _event = _thread != nullptr ? _thread->_ParkEvent : ObjectMonitor::vthread_unparker_ParkEvent(); 2245 _active = false; 2246 assert(_event != nullptr, "invariant"); 2247 } 2248 2249 ObjectWaiter::ObjectWaiter(oop vthread) : ObjectWaiter((JavaThread*)nullptr) { 2250 _vthread = OopHandle(JavaThread::thread_oop_storage(), vthread); 2251 } 2252 2253 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2254 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon); 2255 } 2256 2257 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2258 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active); 2259 } 2260 2261 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2262 assert(node != nullptr, "should not add null node"); 2263 assert(node->_prev == nullptr, "node already in list"); 2264 assert(node->_next == nullptr, "node already in list"); 2265 // put node at end of queue (circular doubly linked list) 2266 if (_WaitSet == nullptr) { 2267 _WaitSet = node; 2268 node->_prev = node; 2269 node->_next = node; 2270 } else { 2271 ObjectWaiter* head = _WaitSet; 2272 ObjectWaiter* tail = head->_prev; 2273 assert(tail->_next == head, "invariant check"); 2274 tail->_next = node; 2275 head->_prev = node; 2276 node->_next = head; 2277 node->_prev = tail; 2278 } 2279 } 2280 2281 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2282 // dequeue the very first waiter 2283 ObjectWaiter* waiter = _WaitSet; 2284 if (waiter) { 2285 DequeueSpecificWaiter(waiter); 2286 } 2287 return waiter; 2288 } 2289 2290 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2291 assert(node != nullptr, "should not dequeue nullptr node"); 2292 assert(node->_prev != nullptr, "node already removed from list"); 2293 assert(node->_next != nullptr, "node already removed from list"); 2294 // when the waiter has woken up because of interrupt, 2295 // timeout or other spurious wake-up, dequeue the 2296 // waiter from waiting list 2297 ObjectWaiter* next = node->_next; 2298 if (next == node) { 2299 assert(node->_prev == node, "invariant check"); 2300 _WaitSet = nullptr; 2301 } else { 2302 ObjectWaiter* prev = node->_prev; 2303 assert(prev->_next == node, "invariant check"); 2304 assert(next->_prev == node, "invariant check"); 2305 next->_prev = prev; 2306 prev->_next = next; 2307 if (_WaitSet == node) { 2308 _WaitSet = next; 2309 } 2310 } 2311 node->_next = nullptr; 2312 node->_prev = nullptr; 2313 } 2314 2315 // ----------------------------------------------------------------------------- 2316 // PerfData support 2317 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = nullptr; 2318 PerfCounter * ObjectMonitor::_sync_FutileWakeups = nullptr; 2319 PerfCounter * ObjectMonitor::_sync_Parks = nullptr; 2320 PerfCounter * ObjectMonitor::_sync_Notifications = nullptr; 2321 PerfCounter * ObjectMonitor::_sync_Inflations = nullptr; 2322 PerfCounter * ObjectMonitor::_sync_Deflations = nullptr; 2323 PerfLongVariable * ObjectMonitor::_sync_MonExtant = nullptr; 2324 2325 // One-shot global initialization for the sync subsystem. 2326 // We could also defer initialization and initialize on-demand 2327 // the first time we call ObjectSynchronizer::inflate(). 2328 // Initialization would be protected - like so many things - by 2329 // the MonitorCache_lock. 2330 2331 void ObjectMonitor::Initialize() { 2332 assert(!InitDone, "invariant"); 2333 2334 if (!os::is_MP()) { 2335 Knob_SpinLimit = 0; 2336 Knob_PreSpin = 0; 2337 Knob_FixedSpin = -1; 2338 } 2339 2340 if (UsePerfData) { 2341 EXCEPTION_MARK; 2342 #define NEWPERFCOUNTER(n) \ 2343 { \ 2344 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2345 CHECK); \ 2346 } 2347 #define NEWPERFVARIABLE(n) \ 2348 { \ 2349 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2350 CHECK); \ 2351 } 2352 NEWPERFCOUNTER(_sync_Inflations); 2353 NEWPERFCOUNTER(_sync_Deflations); 2354 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2355 NEWPERFCOUNTER(_sync_FutileWakeups); 2356 NEWPERFCOUNTER(_sync_Parks); 2357 NEWPERFCOUNTER(_sync_Notifications); 2358 NEWPERFVARIABLE(_sync_MonExtant); 2359 #undef NEWPERFCOUNTER 2360 #undef NEWPERFVARIABLE 2361 } 2362 2363 _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer); 2364 2365 DEBUG_ONLY(InitDone = true;) 2366 } 2367 2368 void ObjectMonitor::Initialize2() { 2369 _vthread_cxq_head = OopHandle(JavaThread::thread_oop_storage(), nullptr); 2370 _vthread_unparker_ParkEvent = ParkEvent::Allocate(nullptr); 2371 } 2372 2373 void ObjectMonitor::print_on(outputStream* st) const { 2374 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2375 st->print("{contentions=0x%08x,waiters=0x%08x" 2376 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2377 contentions(), waiters(), recursions(), 2378 p2i(owner())); 2379 } 2380 void ObjectMonitor::print() const { print_on(tty); } 2381 2382 #ifdef ASSERT 2383 // Print the ObjectMonitor like a debugger would: 2384 // 2385 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2386 // _header = 0x0000000000000001 2387 // _object = 0x000000070ff45fd0 2388 // _pad_buf0 = { 2389 // [0] = '\0' 2390 // ... 2391 // [43] = '\0' 2392 // } 2393 // _owner = 0x0000000000000000 2394 // _previous_owner_tid = 0 2395 // _pad_buf1 = { 2396 // [0] = '\0' 2397 // ... 2398 // [47] = '\0' 2399 // } 2400 // _next_om = 0x0000000000000000 2401 // _recursions = 0 2402 // _EntryList = 0x0000000000000000 2403 // _cxq = 0x0000000000000000 2404 // _succ = 0x0000000000000000 2405 // _Responsible = 0x0000000000000000 2406 // _SpinDuration = 5000 2407 // _contentions = 0 2408 // _WaitSet = 0x0000700009756248 2409 // _waiters = 1 2410 // _WaitSetLock = 0 2411 // } 2412 // 2413 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2414 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2415 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2416 st->print_cr(" _object = " INTPTR_FORMAT, p2i(object_peek())); 2417 st->print_cr(" _pad_buf0 = {"); 2418 st->print_cr(" [0] = '\\0'"); 2419 st->print_cr(" ..."); 2420 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2421 st->print_cr(" }"); 2422 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(owner_raw())); 2423 st->print_cr(" _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid); 2424 st->print_cr(" _pad_buf1 = {"); 2425 st->print_cr(" [0] = '\\0'"); 2426 st->print_cr(" ..."); 2427 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2428 st->print_cr(" }"); 2429 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2430 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2431 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2432 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2433 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2434 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2435 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2436 st->print_cr(" _contentions = %d", contentions()); 2437 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2438 st->print_cr(" _waiters = %d", _waiters); 2439 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2440 st->print_cr("}"); 2441 } 2442 #endif