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), freeze_on_monitorenter); 460 if (result == freeze_ok) { 461 bool acquired = VThreadMonitorEnter(current); 462 if (acquired) { 463 current->set_preemption_cancelled(true); 464 if (JvmtiExport::should_post_monitor_contended_entered()) { 465 // We are going to call thaw again after this and finish the VMTS 466 // transition so no need to do it here. We will post the event there. 467 current->set_contended_entered_monitor(this); 468 } 469 } 470 current->set_current_pending_monitor(nullptr); 471 DEBUG_ONLY(int state = java_lang_VirtualThread::state(current->vthread())); 472 assert((acquired && current->preemption_cancelled() && state == java_lang_VirtualThread::RUNNING) || 473 (!acquired && !current->preemption_cancelled() && state == java_lang_VirtualThread::BLOCKING), "invariant"); 474 return true; 475 } 476 if (result == freeze_pinned_native) { 477 post_virtual_thread_pinned_event(current, "Native frame or <clinit> on stack"); 478 } 479 } 480 #endif 481 482 OSThreadContendState osts(current->osthread()); 483 484 assert(current->thread_state() == _thread_in_vm, "invariant"); 485 486 for (;;) { 487 ExitOnSuspend eos(this); 488 { 489 ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */); 490 EnterI(current); 491 current->set_current_pending_monitor(nullptr); 492 // We can go to a safepoint at the end of this block. If we 493 // do a thread dump during that safepoint, then this thread will show 494 // as having "-locked" the monitor, but the OS and java.lang.Thread 495 // states will still report that the thread is blocked trying to 496 // acquire it. 497 // If there is a suspend request, ExitOnSuspend will exit the OM 498 // and set the OM as pending. 499 } 500 if (!eos.exited()) { 501 // ExitOnSuspend did not exit the OM 502 assert(owner_raw() == owner_for(current), "invariant"); 503 break; 504 } 505 } 506 507 // We've just gotten past the enter-check-for-suspend dance and we now own 508 // the monitor free and clear. 509 } 510 511 add_to_contentions(-1); 512 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions()); 513 514 // Must either set _recursions = 0 or ASSERT _recursions == 0. 515 assert(_recursions == 0, "invariant"); 516 assert(owner_raw() == owner_for(current), "invariant"); 517 assert(_succ != current, "invariant"); 518 assert(object()->mark() == markWord::encode(this), "invariant"); 519 520 // The thread -- now the owner -- is back in vm mode. 521 // Report the glorious news via TI,DTrace and jvmstat. 522 // The probe effect is non-trivial. All the reportage occurs 523 // while we hold the monitor, increasing the length of the critical 524 // section. Amdahl's parallel speedup law comes vividly into play. 525 // 526 // Another option might be to aggregate the events (thread local or 527 // per-monitor aggregation) and defer reporting until a more opportune 528 // time -- such as next time some thread encounters contention but has 529 // yet to acquire the lock. While spinning that thread could 530 // spinning we could increment JVMStat counters, etc. 531 532 DTRACE_MONITOR_PROBE(contended__entered, this, object(), current); 533 if (JvmtiExport::should_post_monitor_contended_entered()) { 534 JvmtiExport::post_monitor_contended_entered(current, this); 535 536 // The current thread already owns the monitor and is not going to 537 // call park() for the remainder of the monitor enter protocol. So 538 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 539 // event handler consumed an unpark() issued by the thread that 540 // just exited the monitor. 541 } 542 if (event.should_commit()) { 543 event.set_previousOwner(_previous_owner_tid); 544 event.commit(); 545 } 546 OM_PERFDATA_OP(ContendedLockAttempts, inc()); 547 return true; 548 } 549 550 // Caveat: TryLock() is not necessarily serializing if it returns failure. 551 // Callers must compensate as needed. 552 553 ObjectMonitor::TryLockResult ObjectMonitor::TryLock(JavaThread* current) { 554 void* own = owner_raw(); 555 if (own != nullptr) return TryLockResult::HasOwner; 556 if (try_set_owner_from(nullptr, current) == nullptr) { 557 assert(_recursions == 0, "invariant"); 558 return TryLockResult::Success; 559 } 560 // The lock had been free momentarily, but we lost the race to the lock. 561 // Interference -- the CAS failed. 562 // We can either return -1 or retry. 563 // Retry doesn't make as much sense because the lock was just acquired. 564 return TryLockResult::Interference; 565 } 566 567 // Deflate the specified ObjectMonitor if not in-use. Returns true if it 568 // was deflated and false otherwise. 569 // 570 // The async deflation protocol sets owner to DEFLATER_MARKER and 571 // makes contentions negative as signals to contending threads that 572 // an async deflation is in progress. There are a number of checks 573 // as part of the protocol to make sure that the calling thread has 574 // not lost the race to a contending thread. 575 // 576 // The ObjectMonitor has been successfully async deflated when: 577 // (contentions < 0) 578 // Contending threads that see that condition know to retry their operation. 579 // 580 bool ObjectMonitor::deflate_monitor() { 581 if (is_busy()) { 582 // Easy checks are first - the ObjectMonitor is busy so no deflation. 583 return false; 584 } 585 586 const oop obj = object_peek(); 587 588 if (obj == nullptr) { 589 // If the object died, we can recycle the monitor without racing with 590 // Java threads. The GC already broke the association with the object. 591 set_owner_from_raw(nullptr, DEFLATER_MARKER); 592 assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions()); 593 _contentions = INT_MIN; // minimum negative int 594 } else { 595 // Attempt async deflation protocol. 596 597 // Set a null owner to DEFLATER_MARKER to force any contending thread 598 // through the slow path. This is just the first part of the async 599 // deflation dance. 600 if (try_set_owner_from_raw(nullptr, DEFLATER_MARKER) != nullptr) { 601 // The owner field is no longer null so we lost the race since the 602 // ObjectMonitor is now busy. 603 return false; 604 } 605 606 if (contentions() > 0 || _waiters != 0) { 607 // Another thread has raced to enter the ObjectMonitor after 608 // is_busy() above or has already entered and waited on 609 // it which makes it busy so no deflation. Restore owner to 610 // null if it is still DEFLATER_MARKER. 611 if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) { 612 // Deferred decrement for the JT EnterI() that cancelled the async deflation. 613 add_to_contentions(-1); 614 } 615 return false; 616 } 617 618 // Make a zero contentions field negative to force any contending threads 619 // to retry. This is the second part of the async deflation dance. 620 if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) { 621 // Contentions was no longer 0 so we lost the race since the 622 // ObjectMonitor is now busy. Restore owner to null if it is 623 // still DEFLATER_MARKER: 624 if (try_set_owner_from_raw(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) { 625 // Deferred decrement for the JT EnterI() that cancelled the async deflation. 626 add_to_contentions(-1); 627 } 628 return false; 629 } 630 } 631 632 // Sanity checks for the races: 633 guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker"); 634 guarantee(contentions() < 0, "must be negative: contentions=%d", 635 contentions()); 636 guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters); 637 guarantee(_cxq == nullptr, "must be no contending threads: cxq=" 638 INTPTR_FORMAT, p2i(_cxq)); 639 guarantee(_EntryList == nullptr, 640 "must be no entering threads: EntryList=" INTPTR_FORMAT, 641 p2i(_EntryList)); 642 643 if (obj != nullptr) { 644 if (log_is_enabled(Trace, monitorinflation)) { 645 ResourceMark rm; 646 log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT 647 ", mark=" INTPTR_FORMAT ", type='%s'", 648 p2i(obj), obj->mark().value(), 649 obj->klass()->external_name()); 650 } 651 652 // Install the old mark word if nobody else has already done it. 653 install_displaced_markword_in_object(obj); 654 } 655 656 // We leave owner == DEFLATER_MARKER and contentions < 0 657 // to force any racing threads to retry. 658 return true; // Success, ObjectMonitor has been deflated. 659 } 660 661 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 662 // into the header of the object associated with the monitor. This 663 // idempotent method is called by a thread that is deflating a 664 // monitor and by other threads that have detected a race with the 665 // deflation process. 666 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 667 // This function must only be called when (owner == DEFLATER_MARKER 668 // && contentions <= 0), but we can't guarantee that here because 669 // those values could change when the ObjectMonitor gets moved from 670 // the global free list to a per-thread free list. 671 672 guarantee(obj != nullptr, "must be non-null"); 673 674 // Separate loads in is_being_async_deflated(), which is almost always 675 // called before this function, from the load of dmw/header below. 676 677 // _contentions and dmw/header may get written by different threads. 678 // Make sure to observe them in the same order when having several observers. 679 OrderAccess::loadload_for_IRIW(); 680 681 const oop l_object = object_peek(); 682 if (l_object == nullptr) { 683 // ObjectMonitor's object ref has already been cleared by async 684 // deflation or GC so we're done here. 685 return; 686 } 687 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 688 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 689 690 markWord dmw = header(); 691 // The dmw has to be neutral (not null, not locked and not marked). 692 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 693 694 // Install displaced mark word if the object's header still points 695 // to this ObjectMonitor. More than one racing caller to this function 696 // can rarely reach this point, but only one can win. 697 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 698 if (res != markWord::encode(this)) { 699 // This should be rare so log at the Info level when it happens. 700 log_info(monitorinflation)("install_displaced_markword_in_object: " 701 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 702 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 703 dmw.value(), markWord::encode(this).value(), 704 res.value()); 705 } 706 707 // Note: It does not matter which thread restored the header/dmw 708 // into the object's header. The thread deflating the monitor just 709 // wanted the object's header restored and it is. The threads that 710 // detected a race with the deflation process also wanted the 711 // object's header restored before they retry their operation and 712 // because it is restored they will only retry once. 713 } 714 715 // Convert the fields used by is_busy() to a string that can be 716 // used for diagnostic output. 717 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 718 ss->print("is_busy: waiters=%d" 719 ", contentions=%d" 720 ", owner=" INTPTR_FORMAT 721 ", cxq=" PTR_FORMAT 722 ", EntryList=" PTR_FORMAT, 723 _waiters, 724 (contentions() > 0 ? contentions() : 0), 725 owner_is_DEFLATER_MARKER() 726 // We report null instead of DEFLATER_MARKER here because is_busy() 727 // ignores DEFLATER_MARKER values. 728 ? p2i(nullptr) 729 : p2i(owner_raw()), 730 p2i(_cxq), 731 p2i(_EntryList)); 732 return ss->base(); 733 } 734 735 #define MAX_RECHECK_INTERVAL 1000 736 737 void ObjectMonitor::EnterI(JavaThread* current) { 738 assert(current->thread_state() == _thread_blocked, "invariant"); 739 740 // Try the lock - TATAS 741 if (TryLock(current) == TryLockResult::Success) { 742 assert(_succ != current, "invariant"); 743 assert(owner_raw() == owner_for(current), "invariant"); 744 assert(_Responsible != current, "invariant"); 745 return; 746 } 747 748 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 749 // Cancelled the in-progress async deflation by changing owner from 750 // DEFLATER_MARKER to current. As part of the contended enter protocol, 751 // contentions was incremented to a positive value before EnterI() 752 // was called and that prevents the deflater thread from winning the 753 // last part of the 2-part async deflation protocol. After EnterI() 754 // returns to enter(), contentions is decremented because the caller 755 // now owns the monitor. We bump contentions an extra time here to 756 // prevent the deflater thread from winning the last part of the 757 // 2-part async deflation protocol after the regular decrement 758 // occurs in enter(). The deflater thread will decrement contentions 759 // after it recognizes that the async deflation was cancelled. 760 add_to_contentions(1); 761 assert(_succ != current, "invariant"); 762 assert(_Responsible != current, "invariant"); 763 return; 764 } 765 766 assert(InitDone, "Unexpectedly not initialized"); 767 768 // We try one round of spinning *before* enqueueing current. 769 // 770 // If the _owner is ready but OFFPROC we could use a YieldTo() 771 // operation to donate the remainder of this thread's quantum 772 // to the owner. This has subtle but beneficial affinity 773 // effects. 774 775 if (TrySpin(current)) { 776 assert(owner_raw() == owner_for(current), "invariant"); 777 assert(_succ != current, "invariant"); 778 assert(_Responsible != current, "invariant"); 779 return; 780 } 781 782 // The Spin failed -- Enqueue and park the thread ... 783 assert(_succ != current, "invariant"); 784 assert(owner_raw() != owner_for(current), "invariant"); 785 assert(_Responsible != current, "invariant"); 786 787 // Enqueue "current" on ObjectMonitor's _cxq. 788 // 789 // Node acts as a proxy for current. 790 // As an aside, if were to ever rewrite the synchronization code mostly 791 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 792 // Java objects. This would avoid awkward lifecycle and liveness issues, 793 // as well as eliminate a subset of ABA issues. 794 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 795 796 ObjectWaiter node(current); 797 current->_ParkEvent->reset(); 798 node._prev = (ObjectWaiter*) 0xBAD; 799 node.TState = ObjectWaiter::TS_CXQ; 800 801 // Push "current" onto the front of the _cxq. 802 // Once on cxq/EntryList, current stays on-queue until it acquires the lock. 803 // Note that spinning tends to reduce the rate at which threads 804 // enqueue and dequeue on EntryList|cxq. 805 ObjectWaiter* nxt; 806 for (;;) { 807 node._next = nxt = _cxq; 808 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break; 809 810 // Interference - the CAS failed because _cxq changed. Just retry. 811 // As an optional optimization we retry the lock. 812 if (TryLock(current) == TryLockResult::Success) { 813 assert(_succ != current, "invariant"); 814 assert(owner_raw() == owner_for(current), "invariant"); 815 assert(_Responsible != current, "invariant"); 816 return; 817 } 818 } 819 820 // Check for cxq|EntryList edge transition to non-null. This indicates 821 // the onset of contention. While contention persists exiting threads 822 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 823 // operations revert to the faster 1-0 mode. This enter operation may interleave 824 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 825 // arrange for one of the contending thread to use a timed park() operations 826 // to detect and recover from the race. (Stranding is form of progress failure 827 // where the monitor is unlocked but all the contending threads remain parked). 828 // That is, at least one of the contended threads will periodically poll _owner. 829 // One of the contending threads will become the designated "Responsible" thread. 830 // The Responsible thread uses a timed park instead of a normal indefinite park 831 // operation -- it periodically wakes and checks for and recovers from potential 832 // strandings admitted by 1-0 exit operations. We need at most one Responsible 833 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 834 // be responsible for a monitor. 835 // 836 // Currently, one of the contended threads takes on the added role of "Responsible". 837 // A viable alternative would be to use a dedicated "stranding checker" thread 838 // that periodically iterated over all the threads (or active monitors) and unparked 839 // successors where there was risk of stranding. This would help eliminate the 840 // timer scalability issues we see on some platforms as we'd only have one thread 841 // -- the checker -- parked on a timer. 842 843 if (nxt == nullptr && _EntryList == nullptr) { 844 // Try to assume the role of responsible thread for the monitor. 845 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=current } 846 Atomic::replace_if_null(&_Responsible, current); 847 } 848 849 // The lock might have been released while this thread was occupied queueing 850 // itself onto _cxq. To close the race and avoid "stranding" and 851 // progress-liveness failure we must resample-retry _owner before parking. 852 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 853 // In this case the ST-MEMBAR is accomplished with CAS(). 854 // 855 // TODO: Defer all thread state transitions until park-time. 856 // Since state transitions are heavy and inefficient we'd like 857 // to defer the state transitions until absolutely necessary, 858 // and in doing so avoid some transitions ... 859 860 int recheckInterval = 1; 861 bool do_timed_parked = false; 862 863 ContinuationEntry* ce = current->last_continuation(); 864 if (ce != nullptr && ce->is_virtual_thread()) { 865 do_timed_parked = true; 866 } 867 868 for (;;) { 869 870 if (TryLock(current) == TryLockResult::Success) { 871 break; 872 } 873 assert(owner_raw() != owner_for(current), "invariant"); 874 875 // park self 876 if (_Responsible == current || do_timed_parked) { 877 current->_ParkEvent->park((jlong) recheckInterval); 878 // Increase the recheckInterval, but clamp the value. 879 recheckInterval *= 8; 880 if (recheckInterval > MAX_RECHECK_INTERVAL) { 881 recheckInterval = MAX_RECHECK_INTERVAL; 882 } 883 } else { 884 current->_ParkEvent->park(); 885 } 886 887 if (TryLock(current) == TryLockResult::Success) { 888 break; 889 } 890 891 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 892 // Cancelled the in-progress async deflation by changing owner from 893 // DEFLATER_MARKER to current. As part of the contended enter protocol, 894 // contentions was incremented to a positive value before EnterI() 895 // was called and that prevents the deflater thread from winning the 896 // last part of the 2-part async deflation protocol. After EnterI() 897 // returns to enter(), contentions is decremented because the caller 898 // now owns the monitor. We bump contentions an extra time here to 899 // prevent the deflater thread from winning the last part of the 900 // 2-part async deflation protocol after the regular decrement 901 // occurs in enter(). The deflater thread will decrement contentions 902 // after it recognizes that the async deflation was cancelled. 903 add_to_contentions(1); 904 break; 905 } 906 907 // The lock is still contested. 908 909 // Keep a tally of the # of futile wakeups. 910 // Note that the counter is not protected by a lock or updated by atomics. 911 // That is by design - we trade "lossy" counters which are exposed to 912 // races during updates for a lower probe effect. 913 // This PerfData object can be used in parallel with a safepoint. 914 // See the work around in PerfDataManager::destroy(). 915 OM_PERFDATA_OP(FutileWakeups, inc()); 916 917 // Assuming this is not a spurious wakeup we'll normally find _succ == current. 918 // We can defer clearing _succ until after the spin completes 919 // TrySpin() must tolerate being called with _succ == current. 920 // Try yet another round of adaptive spinning. 921 if (TrySpin(current)) { 922 break; 923 } 924 925 // We can find that we were unpark()ed and redesignated _succ while 926 // we were spinning. That's harmless. If we iterate and call park(), 927 // park() will consume the event and return immediately and we'll 928 // just spin again. This pattern can repeat, leaving _succ to simply 929 // spin on a CPU. 930 931 if (_succ == current) _succ = nullptr; 932 933 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 934 OrderAccess::fence(); 935 } 936 937 // Egress : 938 // current has acquired the lock -- Unlink current from the cxq or EntryList. 939 // Normally we'll find current on the EntryList . 940 // From the perspective of the lock owner (this thread), the 941 // EntryList is stable and cxq is prepend-only. 942 // The head of cxq is volatile but the interior is stable. 943 // In addition, current.TState is stable. 944 945 assert(owner_raw() == owner_for(current), "invariant"); 946 947 UnlinkAfterAcquire(current, &node); 948 if (_succ == current) _succ = nullptr; 949 950 assert(_succ != current, "invariant"); 951 if (_Responsible == current) { 952 _Responsible = nullptr; 953 OrderAccess::fence(); // Dekker pivot-point 954 955 // We may leave threads on cxq|EntryList without a designated 956 // "Responsible" thread. This is benign. When this thread subsequently 957 // exits the monitor it can "see" such preexisting "old" threads -- 958 // threads that arrived on the cxq|EntryList before the fence, above -- 959 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 960 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 961 // non-null and elect a new "Responsible" timer thread. 962 // 963 // This thread executes: 964 // ST Responsible=null; MEMBAR (in enter epilogue - here) 965 // LD cxq|EntryList (in subsequent exit) 966 // 967 // Entering threads in the slow/contended path execute: 968 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 969 // The (ST cxq; MEMBAR) is accomplished with CAS(). 970 // 971 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 972 // exit operation from floating above the ST Responsible=null. 973 } 974 975 // We've acquired ownership with CAS(). 976 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 977 // But since the CAS() this thread may have also stored into _succ, 978 // EntryList, cxq or Responsible. These meta-data updates must be 979 // visible __before this thread subsequently drops the lock. 980 // Consider what could occur if we didn't enforce this constraint -- 981 // STs to monitor meta-data and user-data could reorder with (become 982 // visible after) the ST in exit that drops ownership of the lock. 983 // Some other thread could then acquire the lock, but observe inconsistent 984 // or old monitor meta-data and heap data. That violates the JMM. 985 // To that end, the 1-0 exit() operation must have at least STST|LDST 986 // "release" barrier semantics. Specifically, there must be at least a 987 // STST|LDST barrier in exit() before the ST of null into _owner that drops 988 // the lock. The barrier ensures that changes to monitor meta-data and data 989 // protected by the lock will be visible before we release the lock, and 990 // therefore before some other thread (CPU) has a chance to acquire the lock. 991 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 992 // 993 // Critically, any prior STs to _succ or EntryList must be visible before 994 // the ST of null into _owner in the *subsequent* (following) corresponding 995 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 996 // execute a serializing instruction. 997 998 return; 999 } 1000 1001 // ReenterI() is a specialized inline form of the latter half of the 1002 // contended slow-path from EnterI(). We use ReenterI() only for 1003 // monitor reentry in wait(). 1004 // 1005 // In the future we should reconcile EnterI() and ReenterI(). 1006 1007 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) { 1008 assert(current != nullptr, "invariant"); 1009 assert(currentNode != nullptr, "invariant"); 1010 assert(currentNode->_thread == current, "invariant"); 1011 assert(_waiters > 0, "invariant"); 1012 assert(object()->mark() == markWord::encode(this), "invariant"); 1013 1014 assert(current->thread_state() != _thread_blocked, "invariant"); 1015 1016 for (;;) { 1017 uint8_t v = currentNode->TState; 1018 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1019 assert(owner_raw() != owner_for(current), "invariant"); 1020 1021 // This thread has been notified so try to reacquire the lock. 1022 if (TryLock(current) == TryLockResult::Success) { 1023 break; 1024 } 1025 1026 // If that fails, spin again. Note that spin count may be zero so the above TryLock 1027 // is necessary. 1028 if (TrySpin(current)) { 1029 break; 1030 } 1031 1032 { 1033 OSThreadContendState osts(current->osthread()); 1034 1035 assert(current->thread_state() == _thread_in_vm, "invariant"); 1036 1037 { 1038 ClearSuccOnSuspend csos(this); 1039 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1040 current->_ParkEvent->park(); 1041 } 1042 } 1043 1044 // Try again, but just so we distinguish between futile wakeups and 1045 // successful wakeups. The following test isn't algorithmically 1046 // necessary, but it helps us maintain sensible statistics. 1047 if (TryLock(current) == TryLockResult::Success) { 1048 break; 1049 } 1050 1051 // The lock is still contested. 1052 1053 // Assuming this is not a spurious wakeup we'll normally 1054 // find that _succ == current. 1055 if (_succ == current) _succ = nullptr; 1056 1057 // Invariant: after clearing _succ a contending thread 1058 // *must* retry _owner before parking. 1059 OrderAccess::fence(); 1060 1061 // Keep a tally of the # of futile wakeups. 1062 // Note that the counter is not protected by a lock or updated by atomics. 1063 // That is by design - we trade "lossy" counters which are exposed to 1064 // races during updates for a lower probe effect. 1065 // This PerfData object can be used in parallel with a safepoint. 1066 // See the work around in PerfDataManager::destroy(). 1067 OM_PERFDATA_OP(FutileWakeups, inc()); 1068 } 1069 1070 // current has acquired the lock -- Unlink current from the cxq or EntryList . 1071 // Normally we'll find current on the EntryList. 1072 // Unlinking from the EntryList is constant-time and atomic-free. 1073 // From the perspective of the lock owner (this thread), the 1074 // EntryList is stable and cxq is prepend-only. 1075 // The head of cxq is volatile but the interior is stable. 1076 // In addition, current.TState is stable. 1077 1078 assert(owner_raw() == owner_for(current), "invariant"); 1079 assert(object()->mark() == markWord::encode(this), "invariant"); 1080 UnlinkAfterAcquire(current, currentNode); 1081 if (_succ == current) _succ = nullptr; 1082 assert(_succ != current, "invariant"); 1083 currentNode->TState = ObjectWaiter::TS_RUN; 1084 OrderAccess::fence(); // see comments at the end of EnterI() 1085 } 1086 1087 // This method is called from two places: 1088 // - On monitorenter contention with a null waiter. 1089 // - After Object.wait() times out or the target is interrupted to reenter the 1090 // monitor, with the existing waiter. 1091 // For the Object.wait() case we do not delete the ObjectWaiter in case we 1092 // succesfully acquire the monitor since we are going to need it on return. 1093 bool ObjectMonitor::VThreadMonitorEnter(JavaThread* current, ObjectWaiter* waiter) { 1094 if (TryLock(current) == TryLockResult::Success) { 1095 assert(owner_raw() == owner_for(current), "invariant"); 1096 assert(_succ != current, "invariant"); 1097 assert(_Responsible != current, "invariant"); 1098 add_to_contentions(-1); 1099 return true; 1100 } 1101 1102 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) { 1103 // Cancelled the in-progress async deflation by changing owner from 1104 // DEFLATER_MARKER to current. As part of the contended enter protocol, 1105 // contentions was incremented to a positive value before this call to 1106 // VThreadMonitorEnter(). We avoid decrementing contentions to prevent 1107 // the deflater thread from winning the last part of the 2-part async 1108 // deflation protocol. The deflater thread will decrement contentions 1109 // after it recognizes that the async deflation was cancelled. 1110 assert(_succ != current, "invariant"); 1111 assert(_Responsible != current, "invariant"); 1112 assert(waiter != nullptr, "monitor currently in used marked for deflation??"); 1113 return true; 1114 } 1115 1116 oop vthread = current->vthread(); 1117 ObjectWaiter* node = waiter != nullptr ? waiter : new ObjectWaiter(vthread, this); 1118 node->_prev = (ObjectWaiter*) 0xBAD; 1119 node->TState = ObjectWaiter::TS_CXQ; 1120 1121 // Push node associated with vthread onto the front of the _cxq. 1122 ObjectWaiter* nxt; 1123 for (;;) { 1124 node->_next = nxt = _cxq; 1125 if (Atomic::cmpxchg(&_cxq, nxt, node) == nxt) break; 1126 1127 // Interference - the CAS failed because _cxq changed. Just retry. 1128 // As an optional optimization we retry the lock. 1129 if (TryLock(current) == TryLockResult::Success) { 1130 assert(owner_raw() == owner_for(current), "invariant"); 1131 assert(_succ != current, "invariant"); 1132 assert(_Responsible != current, "invariant"); 1133 add_to_contentions(-1); 1134 if (waiter == nullptr) delete node; // for Object.wait() don't delete yet 1135 return true; 1136 } 1137 } 1138 1139 // We have to try once more since owner could have exited monitor and checked 1140 // _cxq before we added the node to the queue. 1141 if (TryLock(current) == TryLockResult::Success) { 1142 assert(owner_raw() == owner_for(current), "invariant"); 1143 assert(_Responsible != current, "invariant"); 1144 UnlinkAfterAcquire(current, node); 1145 if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr; 1146 add_to_contentions(-1); 1147 if (waiter == nullptr) delete node; // for Object.wait() don't delete yet 1148 return true; 1149 } 1150 1151 if (nxt == nullptr && _EntryList == nullptr) { 1152 // The C2 unlock() fast path first checks if _cxq and _EntryList are empty and 1153 // if they are it just clears the _owner field. Since we always run the risk of 1154 // having that check happening before we added the node to _cxq and the release 1155 // of the monitor happening after the last TryLock attempt we need to do something 1156 // to avoid stranding. We set the _Responsible field which results in a timed-wait. 1157 if (Atomic::replace_if_null(&_Responsible, (JavaThread*)java_lang_Thread::thread_id(vthread))) { 1158 java_lang_VirtualThread::set_recheckInterval(vthread, 1); 1159 } 1160 } 1161 1162 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1163 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1164 1165 // We didn't succeed in acquiring the monitor so save ObjectWaiter* 1166 // in the chunk since we will need it when resuming execution. 1167 oop cont = java_lang_VirtualThread::continuation(vthread); 1168 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(cont); 1169 chunk->set_object_waiter(node); 1170 return false; 1171 } 1172 1173 void ObjectMonitor::resume_operation(JavaThread* current, ObjectWaiter* node) { 1174 assert(java_lang_VirtualThread::state(current->vthread()) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1175 assert(current->is_in_VTMS_transition(), "must be"); 1176 1177 if (node->is_wait() && !node->at_reenter()) { 1178 bool notified = VThreadWaitReenter(current, node); 1179 if (!notified) return; 1180 // Notified case. We were already added to CXQ or TS_ENTER 1181 // by the notifier so just try to reenter the monitor. 1182 } 1183 1184 // Retry acquiring monitor... 1185 1186 int state = node->TState; 1187 guarantee(state == ObjectWaiter::TS_ENTER || state == ObjectWaiter::TS_CXQ, "invariant"); 1188 1189 if (TryLock(current) == TryLockResult::Success) { 1190 VThreadEpilog(current, node); 1191 return; 1192 } 1193 1194 oop vthread = current->vthread(); 1195 if (_succ == (JavaThread*)java_lang_Thread::thread_id(vthread)) _succ = nullptr; 1196 1197 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 1198 OrderAccess::fence(); 1199 1200 if (TryLock(current) == TryLockResult::Success) { 1201 VThreadEpilog(current, node); 1202 return; 1203 } 1204 1205 // Update recheck interval in case we are the _Responsible. 1206 if (_Responsible == (JavaThread*)java_lang_Thread::thread_id(vthread)) { 1207 int recheckInterval = java_lang_VirtualThread::recheckInterval(vthread); 1208 assert(recheckInterval >= 1 && recheckInterval <= 6, "invariant"); 1209 if (recheckInterval < 6) { 1210 recheckInterval++; 1211 java_lang_VirtualThread::set_recheckInterval(vthread, recheckInterval); 1212 } 1213 } else if (java_lang_VirtualThread::recheckInterval(vthread) > 0) { 1214 // No need to do timed park anymore 1215 java_lang_VirtualThread::set_recheckInterval(vthread, 0); 1216 } 1217 1218 // The JT will read this variable on return to the resume_monitor_operation stub 1219 // and will unmount (enterSpecial frame removed and return to Continuation.run()). 1220 current->set_preempting(true); 1221 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1222 } 1223 1224 void ObjectMonitor::VThreadEpilog(JavaThread* current, ObjectWaiter* node) { 1225 assert(owner_raw() == owner_for(current), "invariant"); 1226 add_to_contentions(-1); 1227 1228 oop vthread = current->vthread(); 1229 if (java_lang_VirtualThread::recheckInterval(vthread) > 0) { 1230 java_lang_VirtualThread::set_recheckInterval(vthread, 0); 1231 } 1232 int64_t threadid = java_lang_Thread::thread_id(vthread); 1233 if (_succ == (JavaThread*)threadid) _succ = nullptr; 1234 if (_Responsible == (JavaThread*)threadid) { 1235 _Responsible = nullptr; 1236 OrderAccess::fence(); // Dekker pivot-point 1237 } 1238 1239 guarantee(_recursions == 0, "invariant"); 1240 1241 if (node->is_wait()) { 1242 _recursions = node->_recursions; // restore the old recursion count 1243 _waiters--; // decrement the number of waiters 1244 1245 if (node->_interrupted) { 1246 // We will throw at thaw end after finishing the mount transition. 1247 current->set_pending_interrupted_exception(true); 1248 } 1249 } 1250 1251 assert(node->TState == ObjectWaiter::TS_ENTER || node->TState == ObjectWaiter::TS_CXQ, ""); 1252 UnlinkAfterAcquire(current, node); 1253 delete node; 1254 1255 oop cont = java_lang_VirtualThread::continuation(vthread); 1256 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(cont); 1257 chunk->set_object_waiter(nullptr); 1258 1259 if (JvmtiExport::should_post_monitor_contended_entered()) { 1260 // We are going to call thaw again after this and finish the VMTS 1261 // transition so no need to do it here. We will post the event there. 1262 current->set_contended_entered_monitor(this); 1263 } 1264 } 1265 1266 // By convention we unlink a contending thread from EntryList|cxq immediately 1267 // after the thread acquires the lock in ::enter(). Equally, we could defer 1268 // unlinking the thread until ::exit()-time. 1269 1270 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) { 1271 assert(owner_raw() == owner_for(current), "invariant"); 1272 assert((!currentNode->is_vthread() && currentNode->thread() == current) || 1273 (currentNode->is_vthread() && currentNode->vthread() == current->vthread()), "invariant"); 1274 1275 if (currentNode->TState == ObjectWaiter::TS_ENTER) { 1276 // Normal case: remove current from the DLL EntryList . 1277 // This is a constant-time operation. 1278 ObjectWaiter* nxt = currentNode->_next; 1279 ObjectWaiter* prv = currentNode->_prev; 1280 if (nxt != nullptr) nxt->_prev = prv; 1281 if (prv != nullptr) prv->_next = nxt; 1282 if (currentNode == _EntryList) _EntryList = nxt; 1283 assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 1284 assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 1285 } else { 1286 assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 1287 // Inopportune interleaving -- current is still on the cxq. 1288 // This usually means the enqueue of self raced an exiting thread. 1289 // Normally we'll find current near the front of the cxq, so 1290 // dequeueing is typically fast. If needbe we can accelerate 1291 // this with some MCS/CHL-like bidirectional list hints and advisory 1292 // back-links so dequeueing from the interior will normally operate 1293 // in constant-time. 1294 // Dequeue current from either the head (with CAS) or from the interior 1295 // with a linear-time scan and normal non-atomic memory operations. 1296 // CONSIDER: if current is on the cxq then simply drain cxq into EntryList 1297 // and then unlink current from EntryList. We have to drain eventually, 1298 // so it might as well be now. 1299 1300 ObjectWaiter* v = _cxq; 1301 assert(v != nullptr, "invariant"); 1302 if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) { 1303 // The CAS above can fail from interference IFF a "RAT" arrived. 1304 // In that case current must be in the interior and can no longer be 1305 // at the head of cxq. 1306 if (v == currentNode) { 1307 assert(_cxq != v, "invariant"); 1308 v = _cxq; // CAS above failed - start scan at head of list 1309 } 1310 ObjectWaiter* p; 1311 ObjectWaiter* q = nullptr; 1312 for (p = v; p != nullptr && p != currentNode; p = p->_next) { 1313 q = p; 1314 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 1315 } 1316 assert(v != currentNode, "invariant"); 1317 assert(p == currentNode, "Node not found on cxq"); 1318 assert(p != _cxq, "invariant"); 1319 assert(q != nullptr, "invariant"); 1320 assert(q->_next == p, "invariant"); 1321 q->_next = p->_next; 1322 } 1323 } 1324 1325 #ifdef ASSERT 1326 // Diagnostic hygiene ... 1327 currentNode->_prev = (ObjectWaiter*) 0xBAD; 1328 currentNode->_next = (ObjectWaiter*) 0xBAD; 1329 currentNode->TState = ObjectWaiter::TS_RUN; 1330 #endif 1331 } 1332 1333 // ----------------------------------------------------------------------------- 1334 // Exit support 1335 // 1336 // exit() 1337 // ~~~~~~ 1338 // Note that the collector can't reclaim the objectMonitor or deflate 1339 // the object out from underneath the thread calling ::exit() as the 1340 // thread calling ::exit() never transitions to a stable state. 1341 // This inhibits GC, which in turn inhibits asynchronous (and 1342 // inopportune) reclamation of "this". 1343 // 1344 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 1345 // There's one exception to the claim above, however. EnterI() can call 1346 // exit() to drop a lock if the acquirer has been externally suspended. 1347 // In that case exit() is called with _thread_state == _thread_blocked, 1348 // but the monitor's _contentions field is > 0, which inhibits reclamation. 1349 // 1350 // 1-0 exit 1351 // ~~~~~~~~ 1352 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 1353 // the fast-path operators have been optimized so the common ::exit() 1354 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock(). 1355 // The code emitted by fast_unlock() elides the usual MEMBAR. This 1356 // greatly improves latency -- MEMBAR and CAS having considerable local 1357 // latency on modern processors -- but at the cost of "stranding". Absent the 1358 // MEMBAR, a thread in fast_unlock() can race a thread in the slow 1359 // ::enter() path, resulting in the entering thread being stranding 1360 // and a progress-liveness failure. Stranding is extremely rare. 1361 // We use timers (timed park operations) & periodic polling to detect 1362 // and recover from stranding. Potentially stranded threads periodically 1363 // wake up and poll the lock. See the usage of the _Responsible variable. 1364 // 1365 // The CAS() in enter provides for safety and exclusion, while the CAS or 1366 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 1367 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding. 1368 // We detect and recover from stranding with timers. 1369 // 1370 // If a thread transiently strands it'll park until (a) another 1371 // thread acquires the lock and then drops the lock, at which time the 1372 // exiting thread will notice and unpark the stranded thread, or, (b) 1373 // the timer expires. If the lock is high traffic then the stranding latency 1374 // will be low due to (a). If the lock is low traffic then the odds of 1375 // stranding are lower, although the worst-case stranding latency 1376 // is longer. Critically, we don't want to put excessive load in the 1377 // platform's timer subsystem. We want to minimize both the timer injection 1378 // rate (timers created/sec) as well as the number of timers active at 1379 // any one time. (more precisely, we want to minimize timer-seconds, which is 1380 // the integral of the # of active timers at any instant over time). 1381 // Both impinge on OS scalability. Given that, at most one thread parked on 1382 // a monitor will use a timer. 1383 // 1384 // There is also the risk of a futile wake-up. If we drop the lock 1385 // another thread can reacquire the lock immediately, and we can 1386 // then wake a thread unnecessarily. This is benign, and we've 1387 // structured the code so the windows are short and the frequency 1388 // of such futile wakups is low. 1389 1390 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) { 1391 void* cur = owner_raw(); 1392 if (owner_for(current) != cur) { 1393 // Apparent unbalanced locking ... 1394 // Naively we'd like to throw IllegalMonitorStateException. 1395 // As a practical matter we can neither allocate nor throw an 1396 // exception as ::exit() can be called from leaf routines. 1397 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1398 // Upon deeper reflection, however, in a properly run JVM the only 1399 // way we should encounter this situation is in the presence of 1400 // unbalanced JNI locking. TODO: CheckJNICalls. 1401 // See also: CR4414101 1402 #ifdef ASSERT 1403 LogStreamHandle(Error, monitorinflation) lsh; 1404 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1405 " is exiting an ObjectMonitor it does not own.", p2i(current)); 1406 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 1407 print_debug_style_on(&lsh); 1408 assert(false, "Non-balanced monitor enter/exit!"); 1409 #endif 1410 return; 1411 } 1412 1413 if (_recursions != 0) { 1414 _recursions--; // this is simple recursive enter 1415 return; 1416 } 1417 1418 // Invariant: after setting Responsible=null an thread must execute 1419 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 1420 _Responsible = nullptr; 1421 1422 #if INCLUDE_JFR 1423 // get the owner's thread id for the MonitorEnter event 1424 // if it is enabled and the thread isn't suspended 1425 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1426 _previous_owner_tid = JFR_THREAD_ID(current); 1427 } 1428 #endif 1429 1430 for (;;) { 1431 assert(owner_for(current) == owner_raw(), "invariant"); 1432 1433 // Drop the lock. 1434 // release semantics: prior loads and stores from within the critical section 1435 // must not float (reorder) past the following store that drops the lock. 1436 // Uses a storeload to separate release_store(owner) from the 1437 // successor check. The try_set_owner_from() below uses cmpxchg() so 1438 // we get the fence down there. 1439 release_clear_owner(current); 1440 OrderAccess::storeload(); 1441 1442 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) { 1443 return; 1444 } 1445 // Other threads are blocked trying to acquire the lock. 1446 1447 // Normally the exiting thread is responsible for ensuring succession, 1448 // but if other successors are ready or other entering threads are spinning 1449 // then this thread can simply store null into _owner and exit without 1450 // waking a successor. The existence of spinners or ready successors 1451 // guarantees proper succession (liveness). Responsibility passes to the 1452 // ready or running successors. The exiting thread delegates the duty. 1453 // More precisely, if a successor already exists this thread is absolved 1454 // of the responsibility of waking (unparking) one. 1455 // 1456 // The _succ variable is critical to reducing futile wakeup frequency. 1457 // _succ identifies the "heir presumptive" thread that has been made 1458 // ready (unparked) but that has not yet run. We need only one such 1459 // successor thread to guarantee progress. 1460 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1461 // section 3.3 "Futile Wakeup Throttling" for details. 1462 // 1463 // Note that spinners in Enter() also set _succ non-null. 1464 // In the current implementation spinners opportunistically set 1465 // _succ so that exiting threads might avoid waking a successor. 1466 // Another less appealing alternative would be for the exiting thread 1467 // to drop the lock and then spin briefly to see if a spinner managed 1468 // to acquire the lock. If so, the exiting thread could exit 1469 // immediately without waking a successor, otherwise the exiting 1470 // thread would need to dequeue and wake a successor. 1471 // (Note that we'd need to make the post-drop spin short, but no 1472 // shorter than the worst-case round-trip cache-line migration time. 1473 // The dropped lock needs to become visible to the spinner, and then 1474 // the acquisition of the lock by the spinner must become visible to 1475 // the exiting thread). 1476 1477 // It appears that an heir-presumptive (successor) must be made ready. 1478 // Only the current lock owner can manipulate the EntryList or 1479 // drain _cxq, so we need to reacquire the lock. If we fail 1480 // to reacquire the lock the responsibility for ensuring succession 1481 // falls to the new owner. 1482 // 1483 if (try_set_owner_from(nullptr, current) != nullptr) { 1484 return; 1485 } 1486 1487 guarantee(owner_raw() == owner_for(current), "invariant"); 1488 1489 ObjectWaiter* w = nullptr; 1490 1491 w = _EntryList; 1492 if (w != nullptr) { 1493 // I'd like to write: guarantee (w->_thread != current). 1494 // But in practice an exiting thread may find itself on the EntryList. 1495 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1496 // then calls exit(). Exit release the lock by setting O._owner to null. 1497 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1498 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1499 // release the lock "O". T2 resumes immediately after the ST of null into 1500 // _owner, above. T2 notices that the EntryList is populated, so it 1501 // reacquires the lock and then finds itself on the EntryList. 1502 // Given all that, we have to tolerate the circumstance where "w" is 1503 // associated with current. 1504 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1505 ExitEpilog(current, w); 1506 return; 1507 } 1508 1509 // If we find that both _cxq and EntryList are null then just 1510 // re-run the exit protocol from the top. 1511 w = _cxq; 1512 if (w == nullptr) continue; 1513 1514 // Drain _cxq into EntryList - bulk transfer. 1515 // First, detach _cxq. 1516 // The following loop is tantamount to: w = swap(&cxq, nullptr) 1517 for (;;) { 1518 assert(w != nullptr, "Invariant"); 1519 ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr); 1520 if (u == w) break; 1521 w = u; 1522 } 1523 1524 assert(w != nullptr, "invariant"); 1525 assert(_EntryList == nullptr, "invariant"); 1526 1527 // Convert the LIFO SLL anchored by _cxq into a DLL. 1528 // The list reorganization step operates in O(LENGTH(w)) time. 1529 // It's critical that this step operate quickly as 1530 // "current" still holds the outer-lock, restricting parallelism 1531 // and effectively lengthening the critical section. 1532 // Invariant: s chases t chases u. 1533 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1534 // we have faster access to the tail. 1535 1536 _EntryList = w; 1537 ObjectWaiter* q = nullptr; 1538 ObjectWaiter* p; 1539 for (p = w; p != nullptr; p = p->_next) { 1540 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1541 p->TState = ObjectWaiter::TS_ENTER; 1542 p->_prev = q; 1543 q = p; 1544 } 1545 1546 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr 1547 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1548 1549 // See if we can abdicate to a spinner instead of waking a thread. 1550 // A primary goal of the implementation is to reduce the 1551 // context-switch rate. 1552 if (_succ != nullptr) continue; 1553 1554 w = _EntryList; 1555 if (w != nullptr) { 1556 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1557 ExitEpilog(current, w); 1558 return; 1559 } 1560 } 1561 } 1562 1563 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) { 1564 assert(owner_raw() == owner_for(current), "invariant"); 1565 1566 // Exit protocol: 1567 // 1. ST _succ = wakee 1568 // 2. membar #loadstore|#storestore; 1569 // 2. ST _owner = nullptr 1570 // 3. unpark(wakee) 1571 1572 oop vthread = nullptr; 1573 ParkEvent * Trigger; 1574 if (!Wakee->is_vthread()) { 1575 JavaThread* t = Wakee->thread(); 1576 assert(t != nullptr, ""); 1577 Trigger = t->_ParkEvent; 1578 _succ = t; 1579 } else { 1580 vthread = Wakee->vthread(); 1581 assert(vthread != nullptr, ""); 1582 Trigger = ObjectMonitor::vthread_unparker_ParkEvent(); 1583 _succ = (JavaThread*)java_lang_Thread::thread_id(vthread); 1584 } 1585 1586 // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again. 1587 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1588 // out-of-scope (non-extant). 1589 Wakee = nullptr; 1590 1591 // Drop the lock. 1592 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1593 release_clear_owner(current); 1594 OrderAccess::fence(); 1595 1596 DTRACE_MONITOR_PROBE(contended__exit, this, object(), current); 1597 1598 if (vthread == nullptr) { 1599 // Platform thread case 1600 Trigger->unpark(); 1601 } else if (java_lang_VirtualThread::set_onWaitingList(vthread, _vthread_cxq_head)) { 1602 Trigger->unpark(); 1603 } 1604 1605 // Maintain stats and report events to JVMTI 1606 OM_PERFDATA_OP(Parks, inc()); 1607 } 1608 1609 // complete_exit exits a lock returning recursion count 1610 // complete_exit requires an inflated monitor 1611 // The _owner field is not always the Thread addr even with an 1612 // inflated monitor, e.g. the monitor can be inflated by a non-owning 1613 // thread due to contention. 1614 intx ObjectMonitor::complete_exit(JavaThread* current) { 1615 assert(InitDone, "Unexpectedly not initialized"); 1616 1617 void* cur = owner_raw(); 1618 if (owner_for(current) != cur) { 1619 if (LockingMode == LM_LEGACY && is_stack_locker(current)) { 1620 assert(_recursions == 0, "internal state error"); 1621 set_owner_from_BasicLock(current); // Convert from BasicLock* to Thread*. 1622 _recursions = 0; 1623 } 1624 } 1625 1626 guarantee(owner_for(current) == owner_raw(), "complete_exit not owner"); 1627 intx save = _recursions; // record the old recursion count 1628 _recursions = 0; // set the recursion level to be 0 1629 exit(current); // exit the monitor 1630 guarantee(owner_raw() != owner_for(current), "invariant"); 1631 return save; 1632 } 1633 1634 // Checks that the current THREAD owns this monitor and causes an 1635 // immediate return if it doesn't. We don't use the CHECK macro 1636 // because we want the IMSE to be the only exception that is thrown 1637 // from the call site when false is returned. Any other pending 1638 // exception is ignored. 1639 #define CHECK_OWNER() \ 1640 do { \ 1641 if (!check_owner(THREAD)) { \ 1642 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1643 return; \ 1644 } \ 1645 } while (false) 1646 1647 // Returns true if the specified thread owns the ObjectMonitor. 1648 // Otherwise returns false and throws IllegalMonitorStateException 1649 // (IMSE). If there is a pending exception and the specified thread 1650 // is not the owner, that exception will be replaced by the IMSE. 1651 bool ObjectMonitor::check_owner(TRAPS) { 1652 JavaThread* current = THREAD; 1653 void* cur = owner_raw(); 1654 if (cur == owner_for(current)) { 1655 return true; 1656 } 1657 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1658 "current thread is not owner", false); 1659 } 1660 1661 static inline bool is_excluded(const Klass* monitor_klass) { 1662 assert(monitor_klass != nullptr, "invariant"); 1663 NOT_JFR_RETURN_(false); 1664 JFR_ONLY(return vmSymbols::jdk_jfr_internal_management_HiddenWait() == monitor_klass->name();) 1665 } 1666 1667 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1668 ObjectMonitor* monitor, 1669 uint64_t notifier_tid, 1670 jlong timeout, 1671 bool timedout) { 1672 assert(event != nullptr, "invariant"); 1673 assert(monitor != nullptr, "invariant"); 1674 const Klass* monitor_klass = monitor->object()->klass(); 1675 if (is_excluded(monitor_klass)) { 1676 return; 1677 } 1678 event->set_monitorClass(monitor_klass); 1679 event->set_timeout(timeout); 1680 // Set an address that is 'unique enough', such that events close in 1681 // time and with the same address are likely (but not guaranteed) to 1682 // belong to the same object. 1683 event->set_address((uintptr_t)monitor); 1684 event->set_notifier(notifier_tid); 1685 event->set_timedOut(timedout); 1686 event->commit(); 1687 } 1688 1689 static void vthread_monitor_waited_event(JavaThread *current, ObjectWaiter* node, EventJavaMonitorWait* event, jboolean timed_out) { 1690 // Since we might safepoint set the anchor so that the stack can we walked. 1691 assert(current->last_continuation() != nullptr, ""); 1692 JavaFrameAnchor* anchor = current->frame_anchor(); 1693 anchor->set_last_Java_sp(current->last_continuation()->entry_sp()); 1694 anchor->set_last_Java_pc(current->last_continuation()->entry_pc()); 1695 1696 JRT_BLOCK 1697 if (event->should_commit()) { 1698 long timeout = java_lang_VirtualThread::waitTimeout(current->vthread()); 1699 post_monitor_wait_event(event, node->_monitor, node->_notifier_tid, timeout, timed_out); 1700 } 1701 if (JvmtiExport::should_post_monitor_waited()) { 1702 JvmtiExport::vthread_post_monitor_waited(current, node->_monitor, timed_out); 1703 } 1704 JRT_BLOCK_END 1705 current->frame_anchor()->clear(); 1706 } 1707 1708 // ----------------------------------------------------------------------------- 1709 // Wait/Notify/NotifyAll 1710 // 1711 // Note: a subset of changes to ObjectMonitor::wait() 1712 // will need to be replicated in complete_exit 1713 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1714 JavaThread* current = THREAD; 1715 1716 assert(InitDone, "Unexpectedly not initialized"); 1717 1718 CHECK_OWNER(); // Throws IMSE if not owner. 1719 1720 EventJavaMonitorWait event; 1721 1722 // check for a pending interrupt 1723 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1724 // post monitor waited event. Note that this is past-tense, we are done waiting. 1725 if (JvmtiExport::should_post_monitor_waited()) { 1726 // Note: 'false' parameter is passed here because the 1727 // wait was not timed out due to thread interrupt. 1728 JvmtiExport::post_monitor_waited(current, this, false); 1729 1730 // In this short circuit of the monitor wait protocol, the 1731 // current thread never drops ownership of the monitor and 1732 // never gets added to the wait queue so the current thread 1733 // cannot be made the successor. This means that the 1734 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1735 // consume an unpark() meant for the ParkEvent associated with 1736 // this ObjectMonitor. 1737 } 1738 if (event.should_commit()) { 1739 post_monitor_wait_event(&event, this, 0, millis, false); 1740 } 1741 THROW(vmSymbols::java_lang_InterruptedException()); 1742 return; 1743 } 1744 1745 current->set_current_waiting_monitor(this); 1746 1747 #ifdef LOOM_MONITOR_SUPPORT 1748 ContinuationEntry* ce = current->last_continuation(); 1749 if (interruptible && ce != nullptr && ce->is_virtual_thread()) { 1750 int result = Continuation::try_preempt(current, ce->cont_oop(current), freeze_on_wait); 1751 if (result == freeze_ok) { 1752 VThreadWait(current, millis); 1753 current->set_current_waiting_monitor(nullptr); 1754 return; 1755 } 1756 if (result == freeze_pinned_native) { 1757 const Klass* monitor_klass = object()->klass(); 1758 if (!is_excluded(monitor_klass)) { 1759 post_virtual_thread_pinned_event(current, "Native frame or <clinit> on stack"); 1760 } 1761 } 1762 } 1763 #endif 1764 1765 // create a node to be put into the queue 1766 // Critically, after we reset() the event but prior to park(), we must check 1767 // for a pending interrupt. 1768 ObjectWaiter node(current); 1769 node.TState = ObjectWaiter::TS_WAIT; 1770 current->_ParkEvent->reset(); 1771 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1772 1773 // Enter the waiting queue, which is a circular doubly linked list in this case 1774 // but it could be a priority queue or any data structure. 1775 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1776 // by the owner of the monitor *except* in the case where park() 1777 // returns because of a timeout of interrupt. Contention is exceptionally rare 1778 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1779 1780 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1781 AddWaiter(&node); 1782 Thread::SpinRelease(&_WaitSetLock); 1783 1784 _Responsible = nullptr; 1785 1786 intx save = _recursions; // record the old recursion count 1787 _waiters++; // increment the number of waiters 1788 _recursions = 0; // set the recursion level to be 1 1789 exit(current); // exit the monitor 1790 guarantee(owner_raw() != owner_for(current), "invariant"); 1791 1792 // The thread is on the WaitSet list - now park() it. 1793 // On MP systems it's conceivable that a brief spin before we park 1794 // could be profitable. 1795 // 1796 // TODO-FIXME: change the following logic to a loop of the form 1797 // while (!timeout && !interrupted && _notified == 0) park() 1798 1799 int ret = OS_OK; 1800 int WasNotified = 0; 1801 1802 // Need to check interrupt state whilst still _thread_in_vm 1803 bool interrupted = interruptible && current->is_interrupted(false); 1804 1805 { // State transition wrappers 1806 OSThread* osthread = current->osthread(); 1807 OSThreadWaitState osts(osthread, true); 1808 1809 assert(current->thread_state() == _thread_in_vm, "invariant"); 1810 1811 { 1812 ClearSuccOnSuspend csos(this); 1813 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1814 if (interrupted || HAS_PENDING_EXCEPTION) { 1815 // Intentionally empty 1816 } else if (!node._notified) { 1817 if (millis <= 0) { 1818 current->_ParkEvent->park(); 1819 } else { 1820 ret = current->_ParkEvent->park(millis); 1821 } 1822 } 1823 } 1824 1825 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1826 // from the WaitSet to the EntryList. 1827 // See if we need to remove Node from the WaitSet. 1828 // We use double-checked locking to avoid grabbing _WaitSetLock 1829 // if the thread is not on the wait queue. 1830 // 1831 // Note that we don't need a fence before the fetch of TState. 1832 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1833 // written by the is thread. (perhaps the fetch might even be satisfied 1834 // by a look-aside into the processor's own store buffer, although given 1835 // the length of the code path between the prior ST and this load that's 1836 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1837 // then we'll acquire the lock and then re-fetch a fresh TState value. 1838 // That is, we fail toward safety. 1839 1840 if (node.TState == ObjectWaiter::TS_WAIT) { 1841 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1842 if (node.TState == ObjectWaiter::TS_WAIT) { 1843 DequeueSpecificWaiter(&node); // unlink from WaitSet 1844 assert(!node._notified, "invariant"); 1845 node.TState = ObjectWaiter::TS_RUN; 1846 } 1847 Thread::SpinRelease(&_WaitSetLock); 1848 } 1849 1850 // The thread is now either on off-list (TS_RUN), 1851 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1852 // The Node's TState variable is stable from the perspective of this thread. 1853 // No other threads will asynchronously modify TState. 1854 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1855 OrderAccess::loadload(); 1856 if (_succ == current) _succ = nullptr; 1857 WasNotified = node._notified; 1858 1859 // Reentry phase -- reacquire the monitor. 1860 // re-enter contended monitor after object.wait(). 1861 // retain OBJECT_WAIT state until re-enter successfully completes 1862 // Thread state is thread_in_vm and oop access is again safe, 1863 // although the raw address of the object may have changed. 1864 // (Don't cache naked oops over safepoints, of course). 1865 1866 // post monitor waited event. Note that this is past-tense, we are done waiting. 1867 if (JvmtiExport::should_post_monitor_waited()) { 1868 JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT); 1869 1870 if (node._notified && _succ == current) { 1871 // In this part of the monitor wait-notify-reenter protocol it 1872 // is possible (and normal) for another thread to do a fastpath 1873 // monitor enter-exit while this thread is still trying to get 1874 // to the reenter portion of the protocol. 1875 // 1876 // The ObjectMonitor was notified and the current thread is 1877 // the successor which also means that an unpark() has already 1878 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1879 // consume the unpark() that was done when the successor was 1880 // set because the same ParkEvent is shared between Java 1881 // monitors and JVM/TI RawMonitors (for now). 1882 // 1883 // We redo the unpark() to ensure forward progress, i.e., we 1884 // don't want all pending threads hanging (parked) with none 1885 // entering the unlocked monitor. 1886 current->_ParkEvent->unpark(); 1887 } 1888 } 1889 1890 if (event.should_commit()) { 1891 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1892 } 1893 1894 OrderAccess::fence(); 1895 1896 assert(owner_raw() != owner_for(current), "invariant"); 1897 ObjectWaiter::TStates v = node.TState; 1898 if (v == ObjectWaiter::TS_RUN) { 1899 enter(current); 1900 } else { 1901 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1902 ReenterI(current, &node); 1903 node.wait_reenter_end(this); 1904 } 1905 1906 // current has reacquired the lock. 1907 // Lifecycle - the node representing current must not appear on any queues. 1908 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1909 // want residual elements associated with this thread left on any lists. 1910 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1911 assert(owner_raw() == owner_for(current), "invariant"); 1912 assert(_succ != current, "invariant"); 1913 } // OSThreadWaitState() 1914 1915 current->set_current_waiting_monitor(nullptr); 1916 1917 guarantee(_recursions == 0, "invariant"); 1918 int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current); 1919 _recursions = save // restore the old recursion count 1920 + relock_count; // increased by the deferred relock count 1921 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count(relock_count);) // Deopt never entered these counts. 1922 _waiters--; // decrement the number of waiters 1923 1924 // Verify a few postconditions 1925 assert(owner_raw() == owner_for(current), "invariant"); 1926 assert(_succ != current, "invariant"); 1927 assert(object()->mark() == markWord::encode(this), "invariant"); 1928 1929 // check if the notification happened 1930 if (!WasNotified) { 1931 // no, it could be timeout or Thread.interrupt() or both 1932 // check for interrupt event, otherwise it is timeout 1933 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1934 THROW(vmSymbols::java_lang_InterruptedException()); 1935 } 1936 } 1937 1938 // NOTE: Spurious wake up will be consider as timeout. 1939 // Monitor notify has precedence over thread interrupt. 1940 } 1941 1942 // Consider: 1943 // If the lock is cool (cxq == null && succ == null) and we're on an MP system 1944 // then instead of transferring a thread from the WaitSet to the EntryList 1945 // we might just dequeue a thread from the WaitSet and directly unpark() it. 1946 1947 void ObjectMonitor::INotify(JavaThread* current) { 1948 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1949 ObjectWaiter* iterator = DequeueWaiter(); 1950 if (iterator != nullptr) { 1951 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1952 guarantee(!iterator->_notified, "invariant"); 1953 // Disposition - what might we do with iterator ? 1954 // a. add it directly to the EntryList - either tail (policy == 1) 1955 // or head (policy == 0). 1956 // b. push it onto the front of the _cxq (policy == 2). 1957 // For now we use (b). 1958 1959 if (iterator->is_vthread()) { 1960 oop vthread = iterator->vthread(); 1961 java_lang_VirtualThread::set_notified(vthread, true); 1962 int old_state = java_lang_VirtualThread::state(vthread); 1963 // If state is not WAIT/TIMED_WAIT then target could still be on 1964 // unmount transition, or wait could have already timed-out or target 1965 // could have been interrupted. In the first case, the target itself 1966 // will set the state to BLOCKED at the end of the unmount transition. 1967 // In the other cases the target would have been already unblocked so 1968 // there is nothing to do. 1969 if (old_state == java_lang_VirtualThread::WAIT || 1970 old_state == java_lang_VirtualThread::TIMED_WAIT) { 1971 java_lang_VirtualThread::cmpxchg_state(vthread, old_state, java_lang_VirtualThread::BLOCKED); 1972 } 1973 } 1974 1975 iterator->TState = ObjectWaiter::TS_ENTER; 1976 1977 iterator->_notified = true; 1978 iterator->_notifier_tid = JFR_THREAD_ID(current); 1979 1980 ObjectWaiter* list = _EntryList; 1981 if (list != nullptr) { 1982 assert(list->_prev == nullptr, "invariant"); 1983 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant"); 1984 assert(list != iterator, "invariant"); 1985 } 1986 1987 // prepend to cxq 1988 if (list == nullptr) { 1989 iterator->_next = iterator->_prev = nullptr; 1990 _EntryList = iterator; 1991 } else { 1992 iterator->TState = ObjectWaiter::TS_CXQ; 1993 for (;;) { 1994 ObjectWaiter* front = _cxq; 1995 iterator->_next = front; 1996 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) { 1997 break; 1998 } 1999 } 2000 } 2001 2002 // _WaitSetLock protects the wait queue, not the EntryList. We could 2003 // move the add-to-EntryList operation, above, outside the critical section 2004 // protected by _WaitSetLock. In practice that's not useful. With the 2005 // exception of wait() timeouts and interrupts the monitor owner 2006 // is the only thread that grabs _WaitSetLock. There's almost no contention 2007 // on _WaitSetLock so it's not profitable to reduce the length of the 2008 // critical section. 2009 if (!iterator->is_vthread()) { 2010 iterator->wait_reenter_begin(this); 2011 } 2012 } 2013 Thread::SpinRelease(&_WaitSetLock); 2014 } 2015 2016 // Consider: a not-uncommon synchronization bug is to use notify() when 2017 // notifyAll() is more appropriate, potentially resulting in stranded 2018 // threads; this is one example of a lost wakeup. A useful diagnostic 2019 // option is to force all notify() operations to behave as notifyAll(). 2020 // 2021 // Note: We can also detect many such problems with a "minimum wait". 2022 // When the "minimum wait" is set to a small non-zero timeout value 2023 // and the program does not hang whereas it did absent "minimum wait", 2024 // that suggests a lost wakeup bug. 2025 2026 void ObjectMonitor::notify(TRAPS) { 2027 JavaThread* current = THREAD; 2028 CHECK_OWNER(); // Throws IMSE if not owner. 2029 if (_WaitSet == nullptr) { 2030 return; 2031 } 2032 DTRACE_MONITOR_PROBE(notify, this, object(), current); 2033 INotify(current); 2034 OM_PERFDATA_OP(Notifications, inc(1)); 2035 } 2036 2037 2038 // The current implementation of notifyAll() transfers the waiters one-at-a-time 2039 // from the waitset to the EntryList. This could be done more efficiently with a 2040 // single bulk transfer but in practice it's not time-critical. Beware too, 2041 // that in prepend-mode we invert the order of the waiters. Let's say that the 2042 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend 2043 // mode the waitset will be empty and the EntryList will be "DCBAXYZ". 2044 2045 void ObjectMonitor::notifyAll(TRAPS) { 2046 JavaThread* current = THREAD; 2047 CHECK_OWNER(); // Throws IMSE if not owner. 2048 if (_WaitSet == nullptr) { 2049 return; 2050 } 2051 2052 DTRACE_MONITOR_PROBE(notifyAll, this, object(), current); 2053 int tally = 0; 2054 while (_WaitSet != nullptr) { 2055 tally++; 2056 INotify(current); 2057 } 2058 2059 OM_PERFDATA_OP(Notifications, inc(tally)); 2060 } 2061 2062 void ObjectMonitor::VThreadWait(JavaThread* current, jlong millis) { 2063 oop vthread = current->vthread(); 2064 ObjectWaiter* node = new ObjectWaiter(vthread, this); 2065 node->_is_wait = true; 2066 node->TState = ObjectWaiter::TS_WAIT; 2067 java_lang_VirtualThread::set_notified(vthread, false); // Reset notified flag 2068 2069 // Enter the waiting queue, which is a circular doubly linked list in this case 2070 // but it could be a priority queue or any data structure. 2071 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 2072 // by the owner of the monitor *except* in the case where park() 2073 // returns because of a timeout or interrupt. Contention is exceptionally rare 2074 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 2075 2076 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 2077 AddWaiter(node); 2078 Thread::SpinRelease(&_WaitSetLock); 2079 2080 _Responsible = nullptr; 2081 2082 node->_recursions = _recursions; // record the old recursion count 2083 _recursions = 0; // set the recursion level to be 1 2084 _waiters++; // increment the number of waiters 2085 exit(current); // exit the monitor 2086 guarantee(owner_raw() != owner_for(current), "invariant"); 2087 2088 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 2089 java_lang_VirtualThread::set_state(vthread, millis == 0 ? java_lang_VirtualThread::WAITING : java_lang_VirtualThread::TIMED_WAITING); 2090 java_lang_VirtualThread::set_waitTimeout(vthread, millis); 2091 2092 // Save the ObjectWaiter* in the chunk since we will need it 2093 // when resuming execution. 2094 oop cont = java_lang_VirtualThread::continuation(vthread); 2095 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(cont); 2096 chunk->set_object_waiter(node); 2097 } 2098 2099 bool ObjectMonitor::VThreadWaitReenter(JavaThread* current, ObjectWaiter* node) { 2100 // First time we run after being preempted on Object.wait(). 2101 // We need to check if we were interrupted or wait() timed-out 2102 // and in that case remove ourselves from the _WaitSet queue. 2103 if (node->TState == ObjectWaiter::TS_WAIT) { 2104 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 2105 if (node->TState == ObjectWaiter::TS_WAIT) { 2106 DequeueSpecificWaiter(node); // unlink from WaitSet 2107 assert(!node->_notified, "invariant"); 2108 node->TState = ObjectWaiter::TS_RUN; 2109 } 2110 Thread::SpinRelease(&_WaitSetLock); 2111 } 2112 2113 ObjectWaiter::TStates state = node->TState; 2114 bool was_notified = state == ObjectWaiter::TS_ENTER || state == ObjectWaiter::TS_CXQ; 2115 assert(was_notified || state == ObjectWaiter::TS_RUN, ""); 2116 2117 // save it so that once we re-acquire the monitor we know if we need to throw IE. 2118 node->_interrupted = !was_notified && current->is_interrupted(false); 2119 2120 EventJavaMonitorWait event; 2121 if (event.should_commit() || JvmtiExport::should_post_monitor_waited()) { 2122 vthread_monitor_waited_event(current, node, &event, !was_notified && !node->_interrupted); 2123 } 2124 2125 node->_at_reenter = true; 2126 add_to_contentions(1); 2127 assert(owner_raw() != owner_for(current), "invariant"); 2128 2129 if (!was_notified) { 2130 bool acquired = VThreadMonitorEnter(current, node); 2131 if (acquired) { 2132 guarantee(_recursions == 0, "invariant"); 2133 _recursions = node->_recursions; // restore the old recursion count 2134 _waiters--; // decrement the number of waiters 2135 2136 if (node->_interrupted) { 2137 // We will throw at thaw end after finishing the mount transition. 2138 current->set_pending_interrupted_exception(true); 2139 } 2140 2141 delete node; 2142 oop cont = java_lang_VirtualThread::continuation(current->vthread()); 2143 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(cont); 2144 chunk->set_object_waiter(nullptr); 2145 } else { 2146 // The JT will read this variable on return to the resume_monitor_operation stub 2147 // and will unmount (enterSpecial frame removed and return to Continuation.run()). 2148 current->set_preempting(true); 2149 } 2150 } 2151 return was_notified; 2152 } 2153 2154 // ----------------------------------------------------------------------------- 2155 // Adaptive Spinning Support 2156 // 2157 // Adaptive spin-then-block - rational spinning 2158 // 2159 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 2160 // algorithm. On high order SMP systems it would be better to start with 2161 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 2162 // a contending thread could enqueue itself on the cxq and then spin locally 2163 // on a thread-specific variable such as its ParkEvent._Event flag. 2164 // That's left as an exercise for the reader. Note that global spinning is 2165 // not problematic on Niagara, as the L2 cache serves the interconnect and 2166 // has both low latency and massive bandwidth. 2167 // 2168 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 2169 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 2170 // (duration) or we can fix the count at approximately the duration of 2171 // a context switch and vary the frequency. Of course we could also 2172 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 2173 // For a description of 'Adaptive spin-then-block mutual exclusion in 2174 // multi-threaded processing,' see U.S. Pat. No. 8046758. 2175 // 2176 // This implementation varies the duration "D", where D varies with 2177 // the success rate of recent spin attempts. (D is capped at approximately 2178 // length of a round-trip context switch). The success rate for recent 2179 // spin attempts is a good predictor of the success rate of future spin 2180 // attempts. The mechanism adapts automatically to varying critical 2181 // section length (lock modality), system load and degree of parallelism. 2182 // D is maintained per-monitor in _SpinDuration and is initialized 2183 // optimistically. Spin frequency is fixed at 100%. 2184 // 2185 // Note that _SpinDuration is volatile, but we update it without locks 2186 // or atomics. The code is designed so that _SpinDuration stays within 2187 // a reasonable range even in the presence of races. The arithmetic 2188 // operations on _SpinDuration are closed over the domain of legal values, 2189 // so at worst a race will install and older but still legal value. 2190 // At the very worst this introduces some apparent non-determinism. 2191 // We might spin when we shouldn't or vice-versa, but since the spin 2192 // count are relatively short, even in the worst case, the effect is harmless. 2193 // 2194 // Care must be taken that a low "D" value does not become an 2195 // an absorbing state. Transient spinning failures -- when spinning 2196 // is overall profitable -- should not cause the system to converge 2197 // on low "D" values. We want spinning to be stable and predictable 2198 // and fairly responsive to change and at the same time we don't want 2199 // it to oscillate, become metastable, be "too" non-deterministic, 2200 // or converge on or enter undesirable stable absorbing states. 2201 // 2202 // We implement a feedback-based control system -- using past behavior 2203 // to predict future behavior. We face two issues: (a) if the 2204 // input signal is random then the spin predictor won't provide optimal 2205 // results, and (b) if the signal frequency is too high then the control 2206 // system, which has some natural response lag, will "chase" the signal. 2207 // (b) can arise from multimodal lock hold times. Transient preemption 2208 // can also result in apparent bimodal lock hold times. 2209 // Although sub-optimal, neither condition is particularly harmful, as 2210 // in the worst-case we'll spin when we shouldn't or vice-versa. 2211 // The maximum spin duration is rather short so the failure modes aren't bad. 2212 // To be conservative, I've tuned the gain in system to bias toward 2213 // _not spinning. Relatedly, the system can sometimes enter a mode where it 2214 // "rings" or oscillates between spinning and not spinning. This happens 2215 // when spinning is just on the cusp of profitability, however, so the 2216 // situation is not dire. The state is benign -- there's no need to add 2217 // hysteresis control to damp the transition rate between spinning and 2218 // not spinning. 2219 2220 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool 2221 2222 static int Knob_Bonus = 100; // spin success bonus 2223 static int Knob_Penalty = 200; // spin failure penalty 2224 static int Knob_Poverty = 1000; 2225 static int Knob_FixedSpin = 0; 2226 static int Knob_PreSpin = 10; // 20-100 likely better, but it's not better in my testing. 2227 2228 inline static int adjust_up(int spin_duration) { 2229 int x = spin_duration; 2230 if (x < ObjectMonitor::Knob_SpinLimit) { 2231 if (x < Knob_Poverty) { 2232 x = Knob_Poverty; 2233 } 2234 return x + Knob_Bonus; 2235 } else { 2236 return spin_duration; 2237 } 2238 } 2239 2240 inline static int adjust_down(int spin_duration) { 2241 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2242 // AIMD is globally stable. 2243 int x = spin_duration; 2244 if (x > 0) { 2245 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2246 // This is globally sample and tends to damp the response. 2247 x -= Knob_Penalty; 2248 if (x < 0) { x = 0; } 2249 return x; 2250 } else { 2251 return spin_duration; 2252 } 2253 } 2254 2255 bool ObjectMonitor::short_fixed_spin(JavaThread* current, int spin_count, bool adapt) { 2256 for (int ctr = 0; ctr < spin_count; ctr++) { 2257 TryLockResult status = TryLock(current); 2258 if (status == TryLockResult::Success) { 2259 if (adapt) { 2260 _SpinDuration = adjust_up(_SpinDuration); 2261 } 2262 return true; 2263 } else if (status == TryLockResult::Interference) { 2264 break; 2265 } 2266 SpinPause(); 2267 } 2268 return false; 2269 } 2270 2271 // Spinning: Fixed frequency (100%), vary duration 2272 bool ObjectMonitor::TrySpin(JavaThread* current) { 2273 2274 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2275 int knob_fixed_spin = Knob_FixedSpin; // 0 (don't spin: default), 2000 good test 2276 if (knob_fixed_spin > 0) { 2277 return short_fixed_spin(current, knob_fixed_spin, false); 2278 } 2279 2280 // Admission control - verify preconditions for spinning 2281 // 2282 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2283 // becoming an absorbing state. Put another way, we spin briefly to 2284 // sample, just in case the system load, parallelism, contention, or lock 2285 // modality changed. 2286 2287 int knob_pre_spin = Knob_PreSpin; // 10 (default), 100, 1000 or 2000 2288 if (short_fixed_spin(current, knob_pre_spin, true)) { 2289 return true; 2290 } 2291 2292 // 2293 // Consider the following alternative: 2294 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2295 // spin attempt. "Periodically" might mean after a tally of 2296 // the # of failed spin attempts (or iterations) reaches some threshold. 2297 // This takes us into the realm of 1-out-of-N spinning, where we 2298 // hold the duration constant but vary the frequency. 2299 2300 int ctr = _SpinDuration; 2301 if (ctr <= 0) return false; 2302 2303 // We're good to spin ... spin ingress. 2304 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2305 // when preparing to LD...CAS _owner, etc and the CAS is likely 2306 // to succeed. 2307 if (_succ == nullptr) { 2308 _succ = current; 2309 } 2310 void* prv = nullptr; 2311 2312 // There are three ways to exit the following loop: 2313 // 1. A successful spin where this thread has acquired the lock. 2314 // 2. Spin failure with prejudice 2315 // 3. Spin failure without prejudice 2316 2317 while (--ctr >= 0) { 2318 2319 // Periodic polling -- Check for pending GC 2320 // Threads may spin while they're unsafe. 2321 // We don't want spinning threads to delay the JVM from reaching 2322 // a stop-the-world safepoint or to steal cycles from GC. 2323 // If we detect a pending safepoint we abort in order that 2324 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2325 // this thread, if safe, doesn't steal cycles from GC. 2326 // This is in keeping with the "no loitering in runtime" rule. 2327 // We periodically check to see if there's a safepoint pending. 2328 if ((ctr & 0xFF) == 0) { 2329 // Can't call SafepointMechanism::should_process() since that 2330 // might update the poll values and we could be in a thread_blocked 2331 // state here which is not allowed so just check the poll. 2332 if (SafepointMechanism::local_poll_armed(current)) { 2333 break; 2334 } 2335 SpinPause(); 2336 } 2337 2338 // Probe _owner with TATAS 2339 // If this thread observes the monitor transition or flicker 2340 // from locked to unlocked to locked, then the odds that this 2341 // thread will acquire the lock in this spin attempt go down 2342 // considerably. The same argument applies if the CAS fails 2343 // or if we observe _owner change from one non-null value to 2344 // another non-null value. In such cases we might abort 2345 // the spin without prejudice or apply a "penalty" to the 2346 // spin count-down variable "ctr", reducing it by 100, say. 2347 2348 void* ox = owner_raw(); 2349 if (ox == nullptr) { 2350 ox = try_set_owner_from(nullptr, current); 2351 if (ox == nullptr) { 2352 // The CAS succeeded -- this thread acquired ownership 2353 // Take care of some bookkeeping to exit spin state. 2354 if (_succ == current) { 2355 _succ = nullptr; 2356 } 2357 2358 // Increase _SpinDuration : 2359 // The spin was successful (profitable) so we tend toward 2360 // longer spin attempts in the future. 2361 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2362 // If we acquired the lock early in the spin cycle it 2363 // makes sense to increase _SpinDuration proportionally. 2364 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2365 _SpinDuration = adjust_up(_SpinDuration); 2366 return true; 2367 } 2368 2369 // The CAS failed ... we can take any of the following actions: 2370 // * penalize: ctr -= CASPenalty 2371 // * exit spin with prejudice -- abort without adapting spinner 2372 // * exit spin without prejudice. 2373 // * Since CAS is high-latency, retry again immediately. 2374 break; 2375 } 2376 2377 // Did lock ownership change hands ? 2378 if (ox != prv && prv != nullptr) { 2379 break; 2380 } 2381 prv = ox; 2382 2383 if (_succ == nullptr) { 2384 _succ = current; 2385 } 2386 } 2387 2388 // Spin failed with prejudice -- reduce _SpinDuration. 2389 if (ctr < 0) { 2390 _SpinDuration = adjust_down(_SpinDuration); 2391 } 2392 2393 if (_succ == current) { 2394 _succ = nullptr; 2395 // Invariant: after setting succ=null a contending thread 2396 // must recheck-retry _owner before parking. This usually happens 2397 // in the normal usage of TrySpin(), but it's safest 2398 // to make TrySpin() as foolproof as possible. 2399 OrderAccess::fence(); 2400 if (TryLock(current) == TryLockResult::Success) { 2401 return true; 2402 } 2403 } 2404 2405 return false; 2406 } 2407 2408 2409 // ----------------------------------------------------------------------------- 2410 // WaitSet management ... 2411 2412 ObjectWaiter::ObjectWaiter(JavaThread* current) { 2413 _next = nullptr; 2414 _prev = nullptr; 2415 _thread = current; 2416 _monitor = nullptr; 2417 _notifier_tid = 0; 2418 _recursions = 0; 2419 TState = TS_RUN; 2420 _notified = false; 2421 _is_wait = false; 2422 _at_reenter = false; 2423 _interrupted = false; 2424 _active = false; 2425 } 2426 2427 ObjectWaiter::ObjectWaiter(oop vthread, ObjectMonitor* mon) : ObjectWaiter(nullptr) { 2428 assert(oopDesc::is_oop(vthread), ""); 2429 _vthread = OopHandle(JavaThread::thread_oop_storage(), vthread); 2430 _monitor = mon; 2431 } 2432 2433 ObjectWaiter::~ObjectWaiter() { 2434 if (is_vthread()) { 2435 assert(vthread() != nullptr, ""); 2436 _vthread.release(JavaThread::thread_oop_storage()); 2437 } 2438 } 2439 2440 oop ObjectWaiter::vthread() { 2441 return _vthread.resolve(); 2442 } 2443 2444 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2445 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon); 2446 } 2447 2448 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2449 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active); 2450 } 2451 2452 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2453 assert(node != nullptr, "should not add null node"); 2454 assert(node->_prev == nullptr, "node already in list"); 2455 assert(node->_next == nullptr, "node already in list"); 2456 // put node at end of queue (circular doubly linked list) 2457 if (_WaitSet == nullptr) { 2458 _WaitSet = node; 2459 node->_prev = node; 2460 node->_next = node; 2461 } else { 2462 ObjectWaiter* head = _WaitSet; 2463 ObjectWaiter* tail = head->_prev; 2464 assert(tail->_next == head, "invariant check"); 2465 tail->_next = node; 2466 head->_prev = node; 2467 node->_next = head; 2468 node->_prev = tail; 2469 } 2470 } 2471 2472 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2473 // dequeue the very first waiter 2474 ObjectWaiter* waiter = _WaitSet; 2475 if (waiter) { 2476 DequeueSpecificWaiter(waiter); 2477 } 2478 return waiter; 2479 } 2480 2481 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2482 assert(node != nullptr, "should not dequeue nullptr node"); 2483 assert(node->_prev != nullptr, "node already removed from list"); 2484 assert(node->_next != nullptr, "node already removed from list"); 2485 // when the waiter has woken up because of interrupt, 2486 // timeout or other spurious wake-up, dequeue the 2487 // waiter from waiting list 2488 ObjectWaiter* next = node->_next; 2489 if (next == node) { 2490 assert(node->_prev == node, "invariant check"); 2491 _WaitSet = nullptr; 2492 } else { 2493 ObjectWaiter* prev = node->_prev; 2494 assert(prev->_next == node, "invariant check"); 2495 assert(next->_prev == node, "invariant check"); 2496 next->_prev = prev; 2497 prev->_next = next; 2498 if (_WaitSet == node) { 2499 _WaitSet = next; 2500 } 2501 } 2502 node->_next = nullptr; 2503 node->_prev = nullptr; 2504 } 2505 2506 // ----------------------------------------------------------------------------- 2507 // PerfData support 2508 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = nullptr; 2509 PerfCounter * ObjectMonitor::_sync_FutileWakeups = nullptr; 2510 PerfCounter * ObjectMonitor::_sync_Parks = nullptr; 2511 PerfCounter * ObjectMonitor::_sync_Notifications = nullptr; 2512 PerfCounter * ObjectMonitor::_sync_Inflations = nullptr; 2513 PerfCounter * ObjectMonitor::_sync_Deflations = nullptr; 2514 PerfLongVariable * ObjectMonitor::_sync_MonExtant = nullptr; 2515 2516 // One-shot global initialization for the sync subsystem. 2517 // We could also defer initialization and initialize on-demand 2518 // the first time we call ObjectSynchronizer::inflate(). 2519 // Initialization would be protected - like so many things - by 2520 // the MonitorCache_lock. 2521 2522 void ObjectMonitor::Initialize() { 2523 assert(!InitDone, "invariant"); 2524 2525 if (!os::is_MP()) { 2526 Knob_SpinLimit = 0; 2527 Knob_PreSpin = 0; 2528 Knob_FixedSpin = -1; 2529 } 2530 2531 if (UsePerfData) { 2532 EXCEPTION_MARK; 2533 #define NEWPERFCOUNTER(n) \ 2534 { \ 2535 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \ 2536 CHECK); \ 2537 } 2538 #define NEWPERFVARIABLE(n) \ 2539 { \ 2540 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \ 2541 CHECK); \ 2542 } 2543 NEWPERFCOUNTER(_sync_Inflations); 2544 NEWPERFCOUNTER(_sync_Deflations); 2545 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2546 NEWPERFCOUNTER(_sync_FutileWakeups); 2547 NEWPERFCOUNTER(_sync_Parks); 2548 NEWPERFCOUNTER(_sync_Notifications); 2549 NEWPERFVARIABLE(_sync_MonExtant); 2550 #undef NEWPERFCOUNTER 2551 #undef NEWPERFVARIABLE 2552 } 2553 2554 _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer); 2555 2556 DEBUG_ONLY(InitDone = true;) 2557 } 2558 2559 void ObjectMonitor::Initialize2() { 2560 _vthread_cxq_head = OopHandle(JavaThread::thread_oop_storage(), nullptr); 2561 _vthread_unparker_ParkEvent = ParkEvent::Allocate(nullptr); 2562 } 2563 2564 void ObjectMonitor::print_on(outputStream* st) const { 2565 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2566 st->print("{contentions=0x%08x,waiters=0x%08x" 2567 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}", 2568 contentions(), waiters(), recursions(), 2569 p2i(owner())); 2570 } 2571 void ObjectMonitor::print() const { print_on(tty); } 2572 2573 #ifdef ASSERT 2574 // Print the ObjectMonitor like a debugger would: 2575 // 2576 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2577 // _header = 0x0000000000000001 2578 // _object = 0x000000070ff45fd0 2579 // _pad_buf0 = { 2580 // [0] = '\0' 2581 // ... 2582 // [43] = '\0' 2583 // } 2584 // _owner = 0x0000000000000000 2585 // _previous_owner_tid = 0 2586 // _pad_buf1 = { 2587 // [0] = '\0' 2588 // ... 2589 // [47] = '\0' 2590 // } 2591 // _next_om = 0x0000000000000000 2592 // _recursions = 0 2593 // _EntryList = 0x0000000000000000 2594 // _cxq = 0x0000000000000000 2595 // _succ = 0x0000000000000000 2596 // _Responsible = 0x0000000000000000 2597 // _SpinDuration = 5000 2598 // _contentions = 0 2599 // _WaitSet = 0x0000700009756248 2600 // _waiters = 1 2601 // _WaitSetLock = 0 2602 // } 2603 // 2604 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2605 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2606 st->print_cr(" _header = " INTPTR_FORMAT, header().value()); 2607 st->print_cr(" _object = " INTPTR_FORMAT, p2i(object_peek())); 2608 st->print_cr(" _pad_buf0 = {"); 2609 st->print_cr(" [0] = '\\0'"); 2610 st->print_cr(" ..."); 2611 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2612 st->print_cr(" }"); 2613 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(owner_raw())); 2614 st->print_cr(" _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid); 2615 st->print_cr(" _pad_buf1 = {"); 2616 st->print_cr(" [0] = '\\0'"); 2617 st->print_cr(" ..."); 2618 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2619 st->print_cr(" }"); 2620 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2621 st->print_cr(" _recursions = " INTX_FORMAT, _recursions); 2622 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList)); 2623 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq)); 2624 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ)); 2625 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible)); 2626 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2627 st->print_cr(" _contentions = %d", contentions()); 2628 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet)); 2629 st->print_cr(" _waiters = %d", _waiters); 2630 st->print_cr(" _WaitSetLock = %d", _WaitSetLock); 2631 st->print_cr("}"); 2632 } 2633 #endif --- EOF ---