1 /* 2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "classfile/vmSymbols.hpp" 26 #include "gc/shared/oopStorage.hpp" 27 #include "gc/shared/oopStorageSet.hpp" 28 #include "jfr/jfrEvents.hpp" 29 #include "jfr/support/jfrThreadId.hpp" 30 #include "logging/log.hpp" 31 #include "logging/logStream.hpp" 32 #include "memory/allocation.inline.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "oops/markWord.hpp" 35 #include "oops/oop.inline.hpp" 36 #include "oops/oopHandle.inline.hpp" 37 #include "oops/weakHandle.inline.hpp" 38 #include "prims/jvmtiDeferredUpdates.hpp" 39 #include "prims/jvmtiExport.hpp" 40 #include "runtime/atomicAccess.hpp" 41 #include "runtime/continuationWrapper.inline.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/lightweightSynchronizer.hpp" 47 #include "runtime/mutexLocker.hpp" 48 #include "runtime/objectMonitor.inline.hpp" 49 #include "runtime/orderAccess.hpp" 50 #include "runtime/osThread.hpp" 51 #include "runtime/safefetch.hpp" 52 #include "runtime/safepointMechanism.inline.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "runtime/threads.hpp" 55 #include "services/threadService.hpp" 56 #include "utilities/debug.hpp" 57 #include "utilities/dtrace.hpp" 58 #include "utilities/globalCounter.inline.hpp" 59 #include "utilities/globalDefinitions.hpp" 60 #include "utilities/macros.hpp" 61 #include "utilities/preserveException.hpp" 62 #if INCLUDE_JFR 63 #include "jfr/support/jfrFlush.hpp" 64 #endif 65 66 #ifdef DTRACE_ENABLED 67 68 // Only bother with this argument setup if dtrace is available 69 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 72 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 73 char* bytes = nullptr; \ 74 int len = 0; \ 75 jlong jtid = SharedRuntime::get_java_tid(thread); \ 76 Symbol* klassname = obj->klass()->name(); \ 77 if (klassname != nullptr) { \ 78 bytes = (char*)klassname->bytes(); \ 79 len = klassname->utf8_length(); \ 80 } 81 82 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 83 { \ 84 if (DTraceMonitorProbes) { \ 85 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 86 HOTSPOT_MONITOR_WAIT(jtid, \ 87 (monitor), bytes, len, (millis)); \ 88 } \ 89 } 90 91 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 92 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 93 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 94 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 95 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 96 97 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 98 { \ 99 if (DTraceMonitorProbes) { \ 100 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 101 HOTSPOT_MONITOR_##probe(jtid, \ 102 (uintptr_t)(monitor), bytes, len); \ 103 } \ 104 } 105 106 #else // ndef DTRACE_ENABLED 107 108 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 109 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 110 111 #endif // ndef DTRACE_ENABLED 112 113 DEBUG_ONLY(static volatile bool InitDone = false;) 114 115 OopStorage* ObjectMonitor::_oop_storage = nullptr; 116 117 OopHandle ObjectMonitor::_vthread_list_head; 118 ParkEvent* ObjectMonitor::_vthread_unparker_ParkEvent = nullptr; 119 120 // ----------------------------------------------------------------------------- 121 // Theory of operations -- Monitors lists, thread residency, etc: 122 // 123 // * A thread acquires ownership of a monitor by successfully 124 // CAS()ing the _owner field from NO_OWNER/DEFLATER_MARKER to 125 // its owner_id (return value from owner_id_from()). 126 // 127 // * Invariant: A thread appears on at most one monitor list -- 128 // entry_list or wait_set -- at any one time. 129 // 130 // * Contending threads "push" themselves onto the entry_list with CAS 131 // and then spin/park. 132 // If the thread is a virtual thread it will first attempt to 133 // unmount itself. The virtual thread will first try to freeze 134 // all frames in the heap. If the operation fails it will just 135 // follow the regular path for platform threads. If the operation 136 // succeeds, it will push itself onto the entry_list with CAS and then 137 // return back to Java to continue the unmount logic. 138 // 139 // * After a contending thread eventually acquires the lock it must 140 // dequeue itself from the entry_list. 141 // 142 // * The exiting thread identifies and unparks an "heir presumptive" 143 // tentative successor thread on the entry_list. In case the successor 144 // is an unmounted virtual thread, the exiting thread will first try 145 // to add it to the list of vthreads waiting to be unblocked, and on 146 // success it will unpark the special unblocker thread instead, which 147 // will be in charge of submitting the vthread back to the scheduler 148 // queue. Critically, the exiting thread doesn't unlink the successor 149 // thread from the entry_list. After having been unparked/re-scheduled, 150 // the wakee will recontend for ownership of the monitor. The successor 151 // (wakee) will either acquire the lock or re-park/unmount itself. 152 // 153 // Succession is provided for by a policy of competitive handoff. 154 // The exiting thread does _not_ grant or pass ownership to the 155 // successor thread. (This is also referred to as "handoff succession"). 156 // Instead the exiting thread releases ownership and possibly wakes 157 // a successor, so the successor can (re)compete for ownership of the lock. 158 // 159 // * The entry_list forms a queue of threads stalled trying to acquire 160 // the lock. Within the entry_list the next pointers always form a 161 // consistent singly linked list. At unlock-time when the unlocking 162 // thread notices that the tail of the entry_list is not known, we 163 // convert the singly linked entry_list into a doubly linked list by 164 // assigning the prev pointers and the entry_list_tail pointer. 165 // 166 // Example: 167 // 168 // The first contending thread that "pushed" itself onto entry_list, 169 // will be the last thread in the list. Each newly pushed thread in 170 // entry_list will be linked through its next pointer, and have its 171 // prev pointer set to null. Thus pushing six threads A-F (in that 172 // order) onto entry_list, will form a singly linked list, see 1) 173 // below. 174 // 175 // 1) entry_list ->F->E->D->C->B->A->null 176 // entry_list_tail ->null 177 // 178 // Since the successor is chosen in FIFO order, the exiting thread 179 // needs to find the tail of the entry_list. This is done by walking 180 // from the entry_list head. While walking the list we also assign 181 // the prev pointers of each thread, essentially forming a doubly 182 // linked list, see 2) below. 183 // 184 // 2) entry_list ->F<=>E<=>D<=>C<=>B<=>A->null 185 // entry_list_tail ----------------------^ 186 // 187 // Once we have formed a doubly linked list it's easy to find the 188 // successor (A), wake it up, have it remove itself, and update the 189 // tail pointer, as seen in and 3) below. 190 // 191 // 3) entry_list ->F<=>E<=>D<=>C<=>B->null 192 // entry_list_tail ------------------^ 193 // 194 // At any time new threads can add themselves to the entry_list, see 195 // 4) below. 196 // 197 // 4) entry_list ->I->H->G->F<=>E<=>D->null 198 // entry_list_tail -------------------^ 199 // 200 // At some point in time the thread (F) that wants to remove itself 201 // from the end of the list, will not have any prev pointer, see 5) 202 // below. 203 // 204 // 5) entry_list ->I->H->G->F->null 205 // entry_list_tail -----------^ 206 // 207 // To resolve this we just start walking from the entry_list head 208 // again, forming a new doubly linked list, before removing the 209 // thread (F), see 6) and 7) below. 210 // 211 // 6) entry_list ->I<=>H<=>G<=>F->null 212 // entry_list_tail --------------^ 213 // 214 // 7) entry_list ->I<=>H<=>G->null 215 // entry_list_tail ----------^ 216 // 217 // * The monitor itself protects all of the operations on the 218 // entry_list except for the CAS of a new arrival to the head. Only 219 // the monitor owner can read or write the prev links (e.g. to 220 // remove itself) or update the tail. 221 // 222 // * The monitor entry list operations avoid locks, but strictly speaking 223 // they're not lock-free. Enter is lock-free, exit is not. 224 // For a description of 'Methods and apparatus providing non-blocking access 225 // to a resource,' see U.S. Pat. No. 7844973. 226 // 227 // * The entry_list can have multiple concurrent "pushers" but only 228 // one concurrent detaching thread. There is no ABA-problem with 229 // this usage of CAS. 230 // 231 // * As long as the entry_list_tail is known the odds are good that we 232 // should be able to dequeue after acquisition (in the ::enter() 233 // epilogue) in constant-time. This is good since a key desideratum 234 // is to minimize queue & monitor metadata manipulation that occurs 235 // while holding the monitor lock -- that is, we want to minimize 236 // monitor lock holds times. Note that even a small amount of fixed 237 // spinning will greatly reduce the # of enqueue-dequeue operations 238 // on entry_list. That is, spinning relieves contention on the 239 // "inner" locks and monitor metadata. 240 // 241 // Insert and delete operations may not operate in constant-time if 242 // we have interference because some other thread is adding or 243 // removing the head element of entry_list or if we need to convert 244 // the singly linked entry_list into a doubly linked list to find the 245 // tail. 246 // 247 // * The monitor synchronization subsystem avoids the use of native 248 // synchronization primitives except for the narrow platform-specific 249 // park-unpark abstraction. See the comments in os_posix.cpp regarding 250 // the semantics of park-unpark. Put another way, this monitor implementation 251 // depends only on atomic operations and park-unpark. 252 // 253 // * Waiting threads reside on the wait_set list -- wait() puts 254 // the caller onto the wait_set. 255 // 256 // * notify() or notifyAll() simply transfers threads from the wait_set 257 // to the entry_list. Subsequent exit() operations will 258 // unpark/re-schedule the notifyee. Unparking/re-scheduling a 259 // notifyee in notify() is inefficient - it's likely the notifyee 260 // would simply impale itself on the lock held by the notifier. 261 262 // Check that object() and set_object() are called from the right context: 263 static void check_object_context() { 264 #ifdef ASSERT 265 Thread* self = Thread::current(); 266 if (self->is_Java_thread()) { 267 // Mostly called from JavaThreads so sanity check the thread state. 268 JavaThread* jt = JavaThread::cast(self); 269 switch (jt->thread_state()) { 270 case _thread_in_vm: // the usual case 271 case _thread_in_Java: // during deopt 272 break; 273 default: 274 fatal("called from an unsafe thread state"); 275 } 276 assert(jt->is_active_Java_thread(), "must be active JavaThread"); 277 } else { 278 // However, ThreadService::get_current_contended_monitor() 279 // can call here via the VMThread so sanity check it. 280 assert(self->is_VM_thread(), "must be"); 281 } 282 #endif // ASSERT 283 } 284 285 ObjectMonitor::ObjectMonitor(oop object) : 286 _metadata(0), 287 _object(_oop_storage, object), 288 _owner(NO_OWNER), 289 _previous_owner_tid(0), 290 _next_om(nullptr), 291 _recursions(0), 292 _entry_list(nullptr), 293 _entry_list_tail(nullptr), 294 _succ(NO_OWNER), 295 _SpinDuration(ObjectMonitor::Knob_SpinLimit), 296 _contentions(0), 297 _unmounted_vthreads(0), 298 _wait_set(nullptr), 299 _waiters(0), 300 _wait_set_lock(0) 301 { } 302 303 ObjectMonitor::~ObjectMonitor() { 304 _object.release(_oop_storage); 305 _object_strong.release(JavaThread::thread_oop_storage()); 306 } 307 308 oop ObjectMonitor::object() const { 309 check_object_context(); 310 return _object.resolve(); 311 } 312 313 // Keep object protected during ObjectLocker preemption. 314 void ObjectMonitor::set_object_strong() { 315 check_object_context(); 316 if (_object_strong.is_empty()) { 317 if (Thread::TrySpinAcquire(&_object_strong_lock)) { 318 if (_object_strong.is_empty()) { 319 assert(_object.resolve() != nullptr, ""); 320 _object_strong = OopHandle(JavaThread::thread_oop_storage(), _object.resolve()); 321 } 322 Thread::SpinRelease(&_object_strong_lock); 323 } 324 } 325 } 326 327 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) { 328 if (current->is_suspended()) { 329 _om->_recursions = 0; 330 _om->clear_successor(); 331 // Don't need a full fence after clearing successor here because of the call to exit(). 332 _om->exit(current, false /* not_suspended */); 333 _om_exited = true; 334 335 current->set_current_pending_monitor(_om); 336 } 337 } 338 339 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) { 340 if (current->is_suspended()) { 341 if (_om->has_successor(current)) { 342 _om->clear_successor(); 343 OrderAccess::fence(); // always do a full fence when successor is cleared 344 } 345 } 346 } 347 348 #define assert_mark_word_consistency() \ 349 assert(UseObjectMonitorTable || object()->mark() == markWord::encode(this), \ 350 "object mark must match encoded this: mark=" INTPTR_FORMAT \ 351 ", encoded this=" INTPTR_FORMAT, object()->mark().value(), \ 352 markWord::encode(this).value()); 353 354 // ----------------------------------------------------------------------------- 355 // Enter support 356 357 bool ObjectMonitor::enter_is_async_deflating() { 358 if (is_being_async_deflated()) { 359 if (!UseObjectMonitorTable) { 360 const oop l_object = object(); 361 if (l_object != nullptr) { 362 // Attempt to restore the header/dmw to the object's header so that 363 // we only retry once if the deflater thread happens to be slow. 364 install_displaced_markword_in_object(l_object); 365 } 366 } 367 return true; 368 } 369 370 return false; 371 } 372 373 bool ObjectMonitor::try_lock_with_contention_mark(JavaThread* locking_thread, ObjectMonitorContentionMark& contention_mark) { 374 assert(contention_mark._monitor == this, "must be"); 375 assert(!is_being_async_deflated(), "must be"); 376 377 int64_t prev_owner = try_set_owner_from(NO_OWNER, locking_thread); 378 bool success = false; 379 380 if (prev_owner == NO_OWNER) { 381 assert(_recursions == 0, "invariant"); 382 success = true; 383 } else if (prev_owner == owner_id_from(locking_thread)) { 384 _recursions++; 385 success = true; 386 } else if (prev_owner == DEFLATER_MARKER) { 387 // Racing with deflation. 388 prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread); 389 if (prev_owner == DEFLATER_MARKER) { 390 // We successfully cancelled the in-progress async deflation by 391 // changing owner from DEFLATER_MARKER to current. We now extend 392 // the lifetime of the contention_mark (e.g. contentions++) here 393 // to prevent the deflater thread from winning the last part of 394 // the 2-part async deflation protocol after the regular 395 // decrement occurs when the contention_mark goes out of 396 // scope. ObjectMonitor::deflate_monitor() which is called by 397 // the deflater thread will decrement contentions after it 398 // recognizes that the async deflation was cancelled. 399 contention_mark.extend(); 400 success = true; 401 } else if (prev_owner == NO_OWNER) { 402 // At this point we cannot race with deflation as we have both incremented 403 // contentions, seen contention > 0 and seen a DEFLATER_MARKER. 404 // success will only be false if this races with something other than 405 // deflation. 406 prev_owner = try_set_owner_from(NO_OWNER, locking_thread); 407 success = prev_owner == NO_OWNER; 408 } 409 } 410 assert(!success || has_owner(locking_thread), "must be"); 411 412 return success; 413 } 414 415 void ObjectMonitor::enter_for_with_contention_mark(JavaThread* locking_thread, ObjectMonitorContentionMark& contention_mark) { 416 // Used by LightweightSynchronizer::inflate_and_enter in deoptimization path to enter for another thread. 417 // The monitor is private to or already owned by locking_thread which must be suspended. 418 // So this code may only contend with deflation. 419 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 420 bool success = try_lock_with_contention_mark(locking_thread, contention_mark); 421 422 assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT 423 ", this=" INTPTR_FORMAT "{owner=" INT64_FORMAT "}", 424 p2i(locking_thread), p2i(this), owner_raw()); 425 } 426 427 bool ObjectMonitor::enter_for(JavaThread* locking_thread) { 428 // Used by ObjectSynchronizer::enter_for() to enter for another thread. 429 // The monitor is private to or already owned by locking_thread which must be suspended. 430 // So this code may only contend with deflation. 431 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be"); 432 433 // Block out deflation as soon as possible. 434 ObjectMonitorContentionMark contention_mark(this); 435 436 // Check for deflation. 437 if (enter_is_async_deflating()) { 438 return false; 439 } 440 441 bool success = try_lock_with_contention_mark(locking_thread, contention_mark); 442 443 assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT 444 ", this=" INTPTR_FORMAT "{owner=" INT64_FORMAT "}", 445 p2i(locking_thread), p2i(this), owner_raw()); 446 assert(has_owner(locking_thread), "must be"); 447 return true; 448 } 449 450 bool ObjectMonitor::try_enter(JavaThread* current, bool check_for_recursion) { 451 // TryLock avoids the CAS and handles deflation. 452 TryLockResult r = try_lock(current); 453 if (r == TryLockResult::Success) { 454 assert(_recursions == 0, "invariant"); 455 return true; 456 } 457 458 // If called from SharedRuntime::monitor_exit_helper(), we know that 459 // this thread doesn't already own the lock. 460 if (!check_for_recursion) { 461 return false; 462 } 463 464 if (r == TryLockResult::HasOwner && has_owner(current)) { 465 _recursions++; 466 return true; 467 } 468 469 return false; 470 } 471 472 bool ObjectMonitor::spin_enter(JavaThread* current) { 473 assert(current == JavaThread::current(), "must be"); 474 475 // Check for recursion. 476 if (try_enter(current)) { 477 return true; 478 } 479 480 // Check for deflation. 481 if (enter_is_async_deflating()) { 482 return false; 483 } 484 485 // We've encountered genuine contention. 486 487 // Do one round of spinning. 488 // Note that if we acquire the monitor from an initial spin 489 // we forgo posting JVMTI events and firing DTRACE probes. 490 if (try_spin(current)) { 491 assert(has_owner(current), "must be current: owner=" INT64_FORMAT, owner_raw()); 492 assert(_recursions == 0, "must be 0: recursions=%zd", _recursions); 493 assert_mark_word_consistency(); 494 return true; 495 } 496 497 return false; 498 } 499 500 bool ObjectMonitor::enter(JavaThread* current) { 501 assert(current == JavaThread::current(), "must be"); 502 503 if (spin_enter(current)) { 504 return true; 505 } 506 507 assert(!has_owner(current), "invariant"); 508 assert(!has_successor(current), "invariant"); 509 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 510 assert(current->thread_state() != _thread_blocked, "invariant"); 511 512 // Keep is_being_async_deflated stable across the rest of enter 513 ObjectMonitorContentionMark contention_mark(this); 514 515 // Check for deflation. 516 if (enter_is_async_deflating()) { 517 return false; 518 } 519 520 // At this point this ObjectMonitor cannot be deflated, finish contended enter 521 enter_with_contention_mark(current, contention_mark); 522 return true; 523 } 524 525 void ObjectMonitor::notify_contended_enter(JavaThread* current) { 526 current->set_current_pending_monitor(this); 527 528 DTRACE_MONITOR_PROBE(contended__enter, this, object(), current); 529 if (JvmtiExport::should_post_monitor_contended_enter()) { 530 JvmtiExport::post_monitor_contended_enter(current, this); 531 532 // The current thread does not yet own the monitor and does not 533 // yet appear on any queues that would get it made the successor. 534 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 535 // handler cannot accidentally consume an unpark() meant for the 536 // ParkEvent associated with this ObjectMonitor. 537 } 538 } 539 540 void ObjectMonitor::enter_with_contention_mark(JavaThread* current, ObjectMonitorContentionMark &cm) { 541 assert(current == JavaThread::current(), "must be"); 542 assert(!has_owner(current), "must be"); 543 assert(cm._monitor == this, "must be"); 544 assert(!is_being_async_deflated(), "must be"); 545 546 JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);) 547 EventJavaMonitorEnter enter_event; 548 if (enter_event.is_started()) { 549 enter_event.set_monitorClass(object()->klass()); 550 // Set an address that is 'unique enough', such that events close in 551 // time and with the same address are likely (but not guaranteed) to 552 // belong to the same object. 553 enter_event.set_address((uintptr_t)this); 554 } 555 EventVirtualThreadPinned vthread_pinned_event; 556 557 freeze_result result; 558 559 assert(current->current_pending_monitor() == nullptr, "invariant"); 560 561 ContinuationEntry* ce = current->last_continuation(); 562 bool is_virtual = ce != nullptr && ce->is_virtual_thread(); 563 if (is_virtual) { 564 notify_contended_enter(current); 565 result = Continuation::try_preempt(current, ce->cont_oop(current)); 566 if (result == freeze_ok) { 567 bool acquired = vthread_monitor_enter(current); 568 if (acquired) { 569 // We actually acquired the monitor while trying to add the vthread to the 570 // _entry_list so cancel preemption. We will still go through the preempt stub 571 // but instead of unmounting we will call thaw to continue execution. 572 current->set_preemption_cancelled(true); 573 if (JvmtiExport::should_post_monitor_contended_entered()) { 574 // We are going to call thaw again after this and finish the VMTS 575 // transition so no need to do it here. We will post the event there. 576 current->set_contended_entered_monitor(this); 577 } 578 } 579 current->set_current_pending_monitor(nullptr); 580 DEBUG_ONLY(int state = java_lang_VirtualThread::state(current->vthread())); 581 assert((acquired && current->preemption_cancelled() && state == java_lang_VirtualThread::RUNNING) || 582 (!acquired && !current->preemption_cancelled() && state == java_lang_VirtualThread::BLOCKING), "invariant"); 583 return; 584 } 585 } 586 587 { 588 // Change java thread status to indicate blocked on monitor enter. 589 JavaThreadBlockedOnMonitorEnterState jtbmes(current, this); 590 591 if (!is_virtual) { // already notified contended_enter for virtual 592 notify_contended_enter(current); 593 } 594 OSThreadContendState osts(current->osthread()); 595 596 assert(current->thread_state() == _thread_in_vm, "invariant"); 597 598 for (;;) { 599 ExitOnSuspend eos(this); 600 { 601 ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */); 602 enter_internal(current); 603 current->set_current_pending_monitor(nullptr); 604 // We can go to a safepoint at the end of this block. If we 605 // do a thread dump during that safepoint, then this thread will show 606 // as having "-locked" the monitor, but the OS and java.lang.Thread 607 // states will still report that the thread is blocked trying to 608 // acquire it. 609 // If there is a suspend request, ExitOnSuspend will exit the OM 610 // and set the OM as pending. 611 } 612 if (!eos.exited()) { 613 // ExitOnSuspend did not exit the OM 614 assert(has_owner(current), "invariant"); 615 break; 616 } 617 } 618 619 // We've just gotten past the enter-check-for-suspend dance and we now own 620 // the monitor free and clear. 621 } 622 623 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions()); 624 625 // Must either set _recursions = 0 or ASSERT _recursions == 0. 626 assert(_recursions == 0, "invariant"); 627 assert(has_owner(current), "invariant"); 628 assert(!has_successor(current), "invariant"); 629 assert_mark_word_consistency(); 630 631 // The thread -- now the owner -- is back in vm mode. 632 // Report the glorious news via TI,DTrace and jvmstat. 633 // The probe effect is non-trivial. All the reportage occurs 634 // while we hold the monitor, increasing the length of the critical 635 // section. Amdahl's parallel speedup law comes vividly into play. 636 // 637 // Another option might be to aggregate the events (thread local or 638 // per-monitor aggregation) and defer reporting until a more opportune 639 // time -- such as next time some thread encounters contention but has 640 // yet to acquire the lock. While spinning that thread could 641 // spinning we could increment JVMStat counters, etc. 642 643 DTRACE_MONITOR_PROBE(contended__entered, this, object(), current); 644 if (JvmtiExport::should_post_monitor_contended_entered()) { 645 JvmtiExport::post_monitor_contended_entered(current, this); 646 647 // The current thread already owns the monitor and is not going to 648 // call park() for the remainder of the monitor enter protocol. So 649 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 650 // event handler consumed an unpark() issued by the thread that 651 // just exited the monitor. 652 } 653 if (enter_event.should_commit()) { 654 enter_event.set_previousOwner(_previous_owner_tid); 655 enter_event.commit(); 656 } 657 658 if (current->current_waiting_monitor() == nullptr) { 659 ContinuationEntry* ce = current->last_continuation(); 660 if (ce != nullptr && ce->is_virtual_thread()) { 661 current->post_vthread_pinned_event(&vthread_pinned_event, "Contended monitor enter", result); 662 } 663 } 664 } 665 666 // Caveat: try_lock() is not necessarily serializing if it returns failure. 667 // Callers must compensate as needed. 668 669 ObjectMonitor::TryLockResult ObjectMonitor::try_lock(JavaThread* current) { 670 int64_t own = owner_raw(); 671 int64_t first_own = own; 672 673 for (;;) { 674 if (own == DEFLATER_MARKER) { 675 // Block out deflation as soon as possible. 676 ObjectMonitorContentionMark contention_mark(this); 677 678 // Check for deflation. 679 if (enter_is_async_deflating()) { 680 // Treat deflation as interference. 681 return TryLockResult::Interference; 682 } 683 if (try_lock_with_contention_mark(current, contention_mark)) { 684 assert(_recursions == 0, "invariant"); 685 return TryLockResult::Success; 686 } else { 687 // Deflation won or change of owner; dont spin 688 break; 689 } 690 } else if (own == NO_OWNER) { 691 int64_t prev_own = try_set_owner_from(NO_OWNER, current); 692 if (prev_own == NO_OWNER) { 693 assert(_recursions == 0, "invariant"); 694 return TryLockResult::Success; 695 } else { 696 // The lock had been free momentarily, but we lost the race to the lock. 697 own = prev_own; 698 } 699 } else { 700 // Retry doesn't make as much sense because the lock was just acquired. 701 break; 702 } 703 } 704 return first_own == own ? TryLockResult::HasOwner : TryLockResult::Interference; 705 } 706 707 // Push "current" onto the head of the _entry_list. Once on _entry_list, 708 // current stays on-queue until it acquires the lock. 709 void ObjectMonitor::add_to_entry_list(JavaThread* current, ObjectWaiter* node) { 710 node->_prev = nullptr; 711 node->TState = ObjectWaiter::TS_ENTER; 712 713 for (;;) { 714 ObjectWaiter* head = AtomicAccess::load(&_entry_list); 715 node->_next = head; 716 if (AtomicAccess::cmpxchg(&_entry_list, head, node) == head) { 717 return; 718 } 719 } 720 } 721 722 // Push "current" onto the head of the entry_list. 723 // If the _entry_list was changed during our push operation, we try to 724 // lock the monitor. Returns true if we locked the monitor, and false 725 // if we added current to _entry_list. Once on _entry_list, current 726 // stays on-queue until it acquires the lock. 727 bool ObjectMonitor::try_lock_or_add_to_entry_list(JavaThread* current, ObjectWaiter* node) { 728 assert(node->TState == ObjectWaiter::TS_RUN, ""); 729 node->_prev = nullptr; 730 node->TState = ObjectWaiter::TS_ENTER; 731 732 for (;;) { 733 ObjectWaiter* head = AtomicAccess::load(&_entry_list); 734 node->_next = head; 735 if (AtomicAccess::cmpxchg(&_entry_list, head, node) == head) { 736 return false; 737 } 738 739 // Interference - the CAS failed because _entry_list changed. Before 740 // retrying the CAS retry taking the lock as it may now be free. 741 if (try_lock(current) == TryLockResult::Success) { 742 assert(!has_successor(current), "invariant"); 743 assert(has_owner(current), "invariant"); 744 node->TState = ObjectWaiter::TS_RUN; 745 return true; 746 } 747 } 748 } 749 750 static void post_monitor_deflate_event(EventJavaMonitorDeflate* event, 751 const oop obj) { 752 assert(event != nullptr, "invariant"); 753 if (obj == nullptr) { 754 // Accept the case when obj was already garbage-collected. 755 // Emit the event anyway, but without details. 756 event->set_monitorClass(nullptr); 757 event->set_address(0); 758 } else { 759 const Klass* monitor_klass = obj->klass(); 760 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) { 761 return; 762 } 763 event->set_monitorClass(monitor_klass); 764 event->set_address((uintptr_t)(void*)obj); 765 } 766 event->commit(); 767 } 768 769 // Deflate the specified ObjectMonitor if not in-use. Returns true if it 770 // was deflated and false otherwise. 771 // 772 // The async deflation protocol sets owner to DEFLATER_MARKER and 773 // makes contentions negative as signals to contending threads that 774 // an async deflation is in progress. There are a number of checks 775 // as part of the protocol to make sure that the calling thread has 776 // not lost the race to a contending thread. 777 // 778 // The ObjectMonitor has been successfully async deflated when: 779 // (contentions < 0) 780 // Contending threads that see that condition know to retry their operation. 781 // 782 bool ObjectMonitor::deflate_monitor(Thread* current) { 783 if (is_busy()) { 784 // Easy checks are first - the ObjectMonitor is busy so no deflation. 785 return false; 786 } 787 788 EventJavaMonitorDeflate event; 789 790 const oop obj = object_peek(); 791 792 if (obj == nullptr) { 793 // If the object died, we can recycle the monitor without racing with 794 // Java threads. The GC already broke the association with the object. 795 set_owner_from_raw(NO_OWNER, DEFLATER_MARKER); 796 assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions()); 797 _contentions = INT_MIN; // minimum negative int 798 } else { 799 // Attempt async deflation protocol. 800 801 // Set a null owner to DEFLATER_MARKER to force any contending thread 802 // through the slow path. This is just the first part of the async 803 // deflation dance. 804 if (try_set_owner_from_raw(NO_OWNER, DEFLATER_MARKER) != NO_OWNER) { 805 // The owner field is no longer null so we lost the race since the 806 // ObjectMonitor is now busy. 807 return false; 808 } 809 810 if (contentions() > 0 || _waiters != 0) { 811 // Another thread has raced to enter the ObjectMonitor after 812 // is_busy() above or has already entered and waited on 813 // it which makes it busy so no deflation. Restore owner to 814 // null if it is still DEFLATER_MARKER. 815 if (try_set_owner_from_raw(DEFLATER_MARKER, NO_OWNER) != DEFLATER_MARKER) { 816 // Deferred decrement for the JT enter_internal() that cancelled the async deflation. 817 add_to_contentions(-1); 818 } 819 return false; 820 } 821 822 // Make a zero contentions field negative to force any contending threads 823 // to retry. This is the second part of the async deflation dance. 824 if (AtomicAccess::cmpxchg(&_contentions, 0, INT_MIN) != 0) { 825 // Contentions was no longer 0 so we lost the race since the 826 // ObjectMonitor is now busy. Restore owner to null if it is 827 // still DEFLATER_MARKER: 828 if (try_set_owner_from_raw(DEFLATER_MARKER, NO_OWNER) != DEFLATER_MARKER) { 829 // Deferred decrement for the JT enter_internal() that cancelled the async deflation. 830 add_to_contentions(-1); 831 } 832 return false; 833 } 834 } 835 836 // Sanity checks for the races: 837 guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker"); 838 guarantee(contentions() < 0, "must be negative: contentions=%d", 839 contentions()); 840 guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters); 841 ObjectWaiter* w = AtomicAccess::load(&_entry_list); 842 guarantee(w == nullptr, 843 "must be no entering threads: entry_list=" INTPTR_FORMAT, 844 p2i(w)); 845 846 if (obj != nullptr) { 847 if (log_is_enabled(Trace, monitorinflation)) { 848 ResourceMark rm; 849 log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT 850 ", mark=" INTPTR_FORMAT ", type='%s'", 851 p2i(obj), obj->mark().value(), 852 obj->klass()->external_name()); 853 } 854 } 855 856 if (UseObjectMonitorTable) { 857 LightweightSynchronizer::deflate_monitor(current, obj, this); 858 } else if (obj != nullptr) { 859 // Install the old mark word if nobody else has already done it. 860 install_displaced_markword_in_object(obj); 861 } 862 863 if (event.should_commit()) { 864 post_monitor_deflate_event(&event, obj); 865 } 866 867 // We leave owner == DEFLATER_MARKER and contentions < 0 868 // to force any racing threads to retry. 869 return true; // Success, ObjectMonitor has been deflated. 870 } 871 872 // Install the displaced mark word (dmw) of a deflating ObjectMonitor 873 // into the header of the object associated with the monitor. This 874 // idempotent method is called by a thread that is deflating a 875 // monitor and by other threads that have detected a race with the 876 // deflation process. 877 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) { 878 assert(!UseObjectMonitorTable, "ObjectMonitorTable has no dmw"); 879 // This function must only be called when (owner == DEFLATER_MARKER 880 // && contentions <= 0), but we can't guarantee that here because 881 // those values could change when the ObjectMonitor gets moved from 882 // the global free list to a per-thread free list. 883 884 guarantee(obj != nullptr, "must be non-null"); 885 886 // Separate loads in is_being_async_deflated(), which is almost always 887 // called before this function, from the load of dmw/header below. 888 889 // _contentions and dmw/header may get written by different threads. 890 // Make sure to observe them in the same order when having several observers. 891 OrderAccess::loadload_for_IRIW(); 892 893 const oop l_object = object_peek(); 894 if (l_object == nullptr) { 895 // ObjectMonitor's object ref has already been cleared by async 896 // deflation or GC so we're done here. 897 return; 898 } 899 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj=" 900 INTPTR_FORMAT, p2i(l_object), p2i(obj)); 901 902 markWord dmw = header(); 903 // The dmw has to be neutral (not null, not locked and not marked). 904 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value()); 905 906 // Install displaced mark word if the object's header still points 907 // to this ObjectMonitor. More than one racing caller to this function 908 // can rarely reach this point, but only one can win. 909 markWord res = obj->cas_set_mark(dmw, markWord::encode(this)); 910 if (res != markWord::encode(this)) { 911 // This should be rare so log at the Info level when it happens. 912 log_info(monitorinflation)("install_displaced_markword_in_object: " 913 "failed cas_set_mark: new_mark=" INTPTR_FORMAT 914 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT, 915 dmw.value(), markWord::encode(this).value(), 916 res.value()); 917 } 918 919 // Note: It does not matter which thread restored the header/dmw 920 // into the object's header. The thread deflating the monitor just 921 // wanted the object's header restored and it is. The threads that 922 // detected a race with the deflation process also wanted the 923 // object's header restored before they retry their operation and 924 // because it is restored they will only retry once. 925 } 926 927 // Convert the fields used by is_busy() to a string that can be 928 // used for diagnostic output. 929 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) { 930 ss->print("is_busy: waiters=%d" 931 ", contentions=%d" 932 ", owner=" INT64_FORMAT 933 ", entry_list=" PTR_FORMAT, 934 _waiters, 935 (contentions() > 0 ? contentions() : 0), 936 owner_is_DEFLATER_MARKER() 937 // We report null instead of DEFLATER_MARKER here because is_busy() 938 // ignores DEFLATER_MARKER values. 939 ? NO_OWNER 940 : owner_raw(), 941 p2i(_entry_list)); 942 return ss->base(); 943 } 944 945 void ObjectMonitor::enter_internal(JavaThread* current) { 946 assert(current->thread_state() == _thread_blocked, "invariant"); 947 948 // Try the lock - TATAS 949 if (try_lock(current) == TryLockResult::Success) { 950 assert(!has_successor(current), "invariant"); 951 assert(has_owner(current), "invariant"); 952 return; 953 } 954 955 assert(InitDone, "Unexpectedly not initialized"); 956 957 // We try one round of spinning *before* enqueueing current. 958 // 959 // If the _owner is ready but OFFPROC we could use a YieldTo() 960 // operation to donate the remainder of this thread's quantum 961 // to the owner. This has subtle but beneficial affinity 962 // effects. 963 964 if (try_spin(current)) { 965 assert(has_owner(current), "invariant"); 966 assert(!has_successor(current), "invariant"); 967 return; 968 } 969 970 // The Spin failed -- Enqueue and park the thread ... 971 assert(!has_successor(current), "invariant"); 972 assert(!has_owner(current), "invariant"); 973 974 // Enqueue "current" on ObjectMonitor's _entry_list. 975 // 976 // Node acts as a proxy for current. 977 // As an aside, if were to ever rewrite the synchronization code mostly 978 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 979 // Java objects. This would avoid awkward lifecycle and liveness issues, 980 // as well as eliminate a subset of ABA issues. 981 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 982 983 ObjectWaiter node(current); 984 current->_ParkEvent->reset(); 985 986 if (try_lock_or_add_to_entry_list(current, &node)) { 987 return; // We got the lock. 988 } 989 // This thread is now added to the _entry_list. 990 991 // The lock might have been released while this thread was occupied queueing 992 // itself onto _entry_list. To close the race and avoid "stranding" and 993 // progress-liveness failure we must resample-retry _owner before parking. 994 // Note the Dekker/Lamport duality: ST _entry_list; MEMBAR; LD Owner. 995 // In this case the ST-MEMBAR is accomplished with CAS(). 996 // 997 // TODO: Defer all thread state transitions until park-time. 998 // Since state transitions are heavy and inefficient we'd like 999 // to defer the state transitions until absolutely necessary, 1000 // and in doing so avoid some transitions ... 1001 1002 // If there are unmounted virtual threads in the _entry_list do a timed-park 1003 // instead to alleviate some deadlocks cases where one of them is picked as 1004 // the successor but cannot run due to having run out of carriers. This can 1005 // happen, for example, if this is a pinned virtual thread currently loading 1006 // or initializining a class, and all other carriers have a pinned vthread 1007 // waiting for said class to be loaded/initialized. 1008 // Read counter *after* adding this thread to the _entry_list. 1009 // Adding to _entry_list uses Atomic::cmpxchg() which already provides 1010 // a fence that prevents this load from floating up previous store. 1011 bool do_timed_parked = has_unmounted_vthreads(); 1012 static int MAX_RECHECK_INTERVAL = 1000; 1013 int recheck_interval = 1; 1014 1015 for (;;) { 1016 1017 if (try_lock(current) == TryLockResult::Success) { 1018 break; 1019 } 1020 assert(!has_owner(current), "invariant"); 1021 1022 // park self 1023 if (do_timed_parked) { 1024 current->_ParkEvent->park((jlong) recheck_interval); 1025 // Increase the recheck_interval, but clamp the value. 1026 recheck_interval *= 8; 1027 if (recheck_interval > MAX_RECHECK_INTERVAL) { 1028 recheck_interval = MAX_RECHECK_INTERVAL; 1029 } 1030 } else { 1031 current->_ParkEvent->park(); 1032 } 1033 1034 if (try_lock(current) == TryLockResult::Success) { 1035 break; 1036 } 1037 1038 // The lock is still contested. 1039 1040 // Assuming this is not a spurious wakeup we'll normally find _succ == current. 1041 // We can defer clearing _succ until after the spin completes 1042 // try_spin() must tolerate being called with _succ == current. 1043 // Try yet another round of adaptive spinning. 1044 if (try_spin(current)) { 1045 break; 1046 } 1047 1048 // We can find that we were unpark()ed and redesignated _succ while 1049 // we were spinning. That's harmless. If we iterate and call park(), 1050 // park() will consume the event and return immediately and we'll 1051 // just spin again. This pattern can repeat, leaving _succ to simply 1052 // spin on a CPU. 1053 1054 if (has_successor(current)) clear_successor(); 1055 1056 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 1057 OrderAccess::fence(); 1058 } 1059 1060 // Egress : 1061 // Current has acquired the lock -- Unlink current from the _entry_list. 1062 unlink_after_acquire(current, &node); 1063 if (has_successor(current)) { 1064 clear_successor(); 1065 // Note that we don't need to do OrderAccess::fence() after clearing 1066 // _succ here, since we own the lock. 1067 } 1068 1069 // We've acquired ownership with CAS(). 1070 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 1071 // But since the CAS() this thread may have also stored into _succ 1072 // or entry_list. These meta-data updates must be visible __before 1073 // this thread subsequently drops the lock. 1074 // Consider what could occur if we didn't enforce this constraint -- 1075 // STs to monitor meta-data and user-data could reorder with (become 1076 // visible after) the ST in exit that drops ownership of the lock. 1077 // Some other thread could then acquire the lock, but observe inconsistent 1078 // or old monitor meta-data and heap data. That violates the JMM. 1079 // To that end, the exit() operation must have at least STST|LDST 1080 // "release" barrier semantics. Specifically, there must be at least a 1081 // STST|LDST barrier in exit() before the ST of null into _owner that drops 1082 // the lock. The barrier ensures that changes to monitor meta-data and data 1083 // protected by the lock will be visible before we release the lock, and 1084 // therefore before some other thread (CPU) has a chance to acquire the lock. 1085 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 1086 // 1087 // Critically, any prior STs to _succ or entry_list must be visible before 1088 // the ST of null into _owner in the *subsequent* (following) corresponding 1089 // monitorexit. 1090 1091 return; 1092 } 1093 1094 // reenter_internal() is a specialized inline form of the latter half of the 1095 // contended slow-path from enter_internal(). We use reenter_internal() only for 1096 // monitor reentry in wait(). 1097 // 1098 // In the future we should reconcile enter_internal() and reenter_internal(). 1099 1100 void ObjectMonitor::reenter_internal(JavaThread* current, ObjectWaiter* currentNode) { 1101 assert(current != nullptr, "invariant"); 1102 assert(current->thread_state() != _thread_blocked, "invariant"); 1103 assert(currentNode != nullptr, "invariant"); 1104 assert(currentNode->_thread == current, "invariant"); 1105 assert(_waiters > 0, "invariant"); 1106 assert_mark_word_consistency(); 1107 1108 // If there are unmounted virtual threads in the _entry_list do a timed-park 1109 // instead to alleviate some deadlocks cases where one of them is picked as 1110 // the successor but cannot run due to having run out of carriers. This can 1111 // happen, for example, if this is a pinned virtual thread (or plain carrier) 1112 // waiting for a class to be initialized. 1113 bool do_timed_parked = has_unmounted_vthreads(); 1114 static int MAX_RECHECK_INTERVAL = 1000; 1115 int recheck_interval = 1; 1116 1117 for (;;) { 1118 ObjectWaiter::TStates v = currentNode->TState; 1119 guarantee(v == ObjectWaiter::TS_ENTER, "invariant"); 1120 assert(!has_owner(current), "invariant"); 1121 1122 // This thread has been notified so try to reacquire the lock. 1123 if (try_lock(current) == TryLockResult::Success) { 1124 break; 1125 } 1126 1127 // If that fails, spin again. Note that spin count may be zero so the above TryLock 1128 // is necessary. 1129 if (try_spin(current)) { 1130 break; 1131 } 1132 1133 { 1134 OSThreadContendState osts(current->osthread()); 1135 1136 assert(current->thread_state() == _thread_in_vm, "invariant"); 1137 1138 { 1139 ClearSuccOnSuspend csos(this); 1140 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1141 if (do_timed_parked) { 1142 current->_ParkEvent->park((jlong) recheck_interval); 1143 // Increase the recheck_interval, but clamp the value. 1144 recheck_interval *= 8; 1145 if (recheck_interval > MAX_RECHECK_INTERVAL) { 1146 recheck_interval = MAX_RECHECK_INTERVAL; 1147 } 1148 } else { 1149 current->_ParkEvent->park(); 1150 } 1151 } 1152 } 1153 1154 // Try again, but just so we distinguish between futile wakeups and 1155 // successful wakeups. The following test isn't algorithmically 1156 // necessary, but it helps us maintain sensible statistics. 1157 if (try_lock(current) == TryLockResult::Success) { 1158 break; 1159 } 1160 1161 // The lock is still contested. 1162 1163 // Assuming this is not a spurious wakeup we'll normally 1164 // find that _succ == current. 1165 if (has_successor(current)) clear_successor(); 1166 1167 // Invariant: after clearing _succ a contending thread 1168 // *must* retry _owner before parking. 1169 OrderAccess::fence(); 1170 1171 // See comment in notify_internal 1172 do_timed_parked |= currentNode->_do_timed_park; 1173 } 1174 1175 // Current has acquired the lock -- Unlink current from the _entry_list. 1176 assert(has_owner(current), "invariant"); 1177 assert_mark_word_consistency(); 1178 unlink_after_acquire(current, currentNode); 1179 if (has_successor(current)) clear_successor(); 1180 assert(!has_successor(current), "invariant"); 1181 currentNode->TState = ObjectWaiter::TS_RUN; 1182 OrderAccess::fence(); // see comments at the end of enter_internal() 1183 } 1184 1185 // This method is called from two places: 1186 // - On monitorenter contention with a null waiter. 1187 // - After Object.wait() times out or the target is interrupted to reenter the 1188 // monitor, with the existing waiter. 1189 // For the Object.wait() case we do not delete the ObjectWaiter in case we 1190 // succesfully acquire the monitor since we are going to need it on return. 1191 bool ObjectMonitor::vthread_monitor_enter(JavaThread* current, ObjectWaiter* waiter) { 1192 if (try_lock(current) == TryLockResult::Success) { 1193 assert(has_owner(current), "invariant"); 1194 assert(!has_successor(current), "invariant"); 1195 return true; 1196 } 1197 1198 oop vthread = current->vthread(); 1199 ObjectWaiter* node = waiter != nullptr ? waiter : new ObjectWaiter(vthread, this); 1200 1201 // Increment counter *before* adding the vthread to the _entry_list. 1202 // Adding to _entry_list uses Atomic::cmpxchg() which already provides 1203 // a fence that prevents reordering of the stores. 1204 inc_unmounted_vthreads(); 1205 1206 if (try_lock_or_add_to_entry_list(current, node)) { 1207 // We got the lock. 1208 if (waiter == nullptr) delete node; // for Object.wait() don't delete yet 1209 dec_unmounted_vthreads(); 1210 return true; 1211 } 1212 // This thread is now added to the entry_list. 1213 1214 // We have to try once more since owner could have exited monitor and checked 1215 // _entry_list before we added the node to the queue. 1216 if (try_lock(current) == TryLockResult::Success) { 1217 assert(has_owner(current), "invariant"); 1218 unlink_after_acquire(current, node); 1219 if (has_successor(current)) clear_successor(); 1220 if (waiter == nullptr) delete node; // for Object.wait() don't delete yet 1221 dec_unmounted_vthreads(); 1222 return true; 1223 } 1224 1225 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1226 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1227 1228 // We didn't succeed in acquiring the monitor so increment _contentions and 1229 // save ObjectWaiter* in the vthread since we will need it when resuming execution. 1230 add_to_contentions(1); 1231 java_lang_VirtualThread::set_objectWaiter(vthread, node); 1232 return false; 1233 } 1234 1235 // Called from thaw code to resume the monitor operation that caused the vthread 1236 // to be unmounted. Method returns true if the monitor is successfully acquired, 1237 // which marks the end of the monitor operation, otherwise it returns false. 1238 bool ObjectMonitor::resume_operation(JavaThread* current, ObjectWaiter* node, ContinuationWrapper& cont) { 1239 assert(java_lang_VirtualThread::state(current->vthread()) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 1240 assert(!has_owner(current), ""); 1241 1242 if (node->is_wait() && !node->at_reenter()) { 1243 bool acquired_monitor = vthread_wait_reenter(current, node, cont); 1244 if (acquired_monitor) return true; 1245 } 1246 1247 // Retry acquiring monitor... 1248 1249 int state = node->TState; 1250 guarantee(state == ObjectWaiter::TS_ENTER, "invariant"); 1251 1252 if (try_lock(current) == TryLockResult::Success) { 1253 vthread_epilog(current, node); 1254 return true; 1255 } 1256 1257 oop vthread = current->vthread(); 1258 if (has_successor(current)) clear_successor(); 1259 1260 // Invariant: after clearing _succ a thread *must* retry acquiring the monitor. 1261 OrderAccess::fence(); 1262 1263 if (try_lock(current) == TryLockResult::Success) { 1264 vthread_epilog(current, node); 1265 return true; 1266 } 1267 1268 // We will return to Continuation.run() and unmount so set the right state. 1269 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::BLOCKING); 1270 1271 return false; 1272 } 1273 1274 void ObjectMonitor::vthread_epilog(JavaThread* current, ObjectWaiter* node) { 1275 assert(has_owner(current), "invariant"); 1276 add_to_contentions(-1); 1277 dec_unmounted_vthreads(); 1278 1279 if (has_successor(current)) clear_successor(); 1280 1281 guarantee(_recursions == 0, "invariant"); 1282 1283 if (node->is_wait()) { 1284 _recursions = node->_recursions; // restore the old recursion count 1285 _waiters--; // decrement the number of waiters 1286 1287 if (node->_interrupted) { 1288 // We will throw at thaw end after finishing the mount transition. 1289 current->set_pending_interrupted_exception(true); 1290 } 1291 } 1292 1293 unlink_after_acquire(current, node); 1294 delete node; 1295 1296 // Clear the ObjectWaiter* from the vthread. 1297 java_lang_VirtualThread::set_objectWaiter(current->vthread(), nullptr); 1298 1299 if (JvmtiExport::should_post_monitor_contended_entered()) { 1300 // We are going to call thaw again after this and finish the VMTS 1301 // transition so no need to do it here. We will post the event there. 1302 current->set_contended_entered_monitor(this); 1303 } 1304 } 1305 1306 // Convert entry_list into a doubly linked list by assigning the prev 1307 // pointers and the entry_list_tail pointer (if needed). Within the 1308 // entry_list the next pointers always form a consistent singly linked 1309 // list. When this function is called, the entry_list will be either 1310 // singly linked, or starting as singly linked (at the head), but 1311 // ending as doubly linked (at the tail). 1312 void ObjectMonitor::entry_list_build_dll(JavaThread* current) { 1313 assert(has_owner(current), "invariant"); 1314 ObjectWaiter* prev = nullptr; 1315 // Need acquire here to match the implicit release of the cmpxchg 1316 // that updated entry_list, so we can access w->prev(). 1317 ObjectWaiter* w = AtomicAccess::load_acquire(&_entry_list); 1318 assert(w != nullptr, "should only be called when entry list is not empty"); 1319 while (w != nullptr) { 1320 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1321 assert(w->prev() == nullptr || w->prev() == prev, "invariant"); 1322 if (w->prev() != nullptr) { 1323 break; 1324 } 1325 w->_prev = prev; 1326 prev = w; 1327 w = w->next(); 1328 } 1329 if (w == nullptr) { 1330 // We converted the entire entry_list from a singly linked list 1331 // into a doubly linked list. Now we just need to set the tail 1332 // pointer. 1333 assert(prev != nullptr && prev->next() == nullptr, "invariant"); 1334 assert(_entry_list_tail == nullptr || _entry_list_tail == prev, "invariant"); 1335 _entry_list_tail = prev; 1336 } else { 1337 #ifdef ASSERT 1338 // We stopped iterating through the _entry_list when we found a 1339 // node that had its prev pointer set. I.e. we converted the first 1340 // part of the entry_list from a singly linked list into a doubly 1341 // linked list. Now we just want to make sure the rest of the list 1342 // is doubly linked. But first we check that we have a tail 1343 // pointer, because if the end of the entry_list is doubly linked 1344 // and we don't have the tail pointer, something is broken. 1345 assert(_entry_list_tail != nullptr, "invariant"); 1346 while (w != nullptr) { 1347 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1348 assert(w->prev() == prev, "invariant"); 1349 prev = w; 1350 w = w->next(); 1351 } 1352 assert(_entry_list_tail == prev, "invariant"); 1353 #endif 1354 } 1355 } 1356 1357 // Return the tail of the _entry_list. If the tail is currently not 1358 // known, it can be found by first calling entry_list_build_dll(). 1359 ObjectWaiter* ObjectMonitor::entry_list_tail(JavaThread* current) { 1360 assert(has_owner(current), "invariant"); 1361 ObjectWaiter* w = _entry_list_tail; 1362 if (w != nullptr) { 1363 return w; 1364 } 1365 entry_list_build_dll(current); 1366 w = _entry_list_tail; 1367 assert(w != nullptr, "invariant"); 1368 return w; 1369 } 1370 1371 // By convention we unlink a contending thread from _entry_list 1372 // immediately after the thread acquires the lock in ::enter(). 1373 // The head of _entry_list is volatile but the interior is stable. 1374 // In addition, current.TState is stable. 1375 1376 void ObjectMonitor::unlink_after_acquire(JavaThread* current, ObjectWaiter* currentNode) { 1377 assert(has_owner(current), "invariant"); 1378 assert((!currentNode->is_vthread() && currentNode->thread() == current) || 1379 (currentNode->is_vthread() && currentNode->vthread() == current->vthread()), "invariant"); 1380 1381 // Check if we are unlinking the last element in the _entry_list. 1382 // This is by far the most common case. 1383 if (currentNode->next() == nullptr) { 1384 assert(_entry_list_tail == nullptr || _entry_list_tail == currentNode, "invariant"); 1385 1386 ObjectWaiter* w = AtomicAccess::load(&_entry_list); 1387 if (w == currentNode) { 1388 // The currentNode is the only element in _entry_list. 1389 if (AtomicAccess::cmpxchg(&_entry_list, w, (ObjectWaiter*)nullptr) == w) { 1390 _entry_list_tail = nullptr; 1391 currentNode->set_bad_pointers(); 1392 return; 1393 } 1394 // The CAS above can fail from interference IFF a contending 1395 // thread "pushed" itself onto entry_list. So fall-through to 1396 // building the doubly linked list. 1397 assert(currentNode->prev() == nullptr, "invariant"); 1398 } 1399 if (currentNode->prev() == nullptr) { 1400 // Build the doubly linked list to get hold of 1401 // currentNode->prev(). 1402 entry_list_build_dll(current); 1403 assert(currentNode->prev() != nullptr, "must be"); 1404 assert(_entry_list_tail == currentNode, "must be"); 1405 } 1406 // The currentNode is the last element in _entry_list and we know 1407 // which element is the previous one. 1408 assert(_entry_list != currentNode, "invariant"); 1409 _entry_list_tail = currentNode->prev(); 1410 _entry_list_tail->_next = nullptr; 1411 currentNode->set_bad_pointers(); 1412 return; 1413 } 1414 1415 // If we get here it means the current thread enqueued itself on the 1416 // _entry_list but was then able to "steal" the lock before the 1417 // chosen successor was able to. Consequently currentNode must be an 1418 // interior node in the _entry_list, or the head. 1419 assert(currentNode->next() != nullptr, "invariant"); 1420 assert(currentNode != _entry_list_tail, "invariant"); 1421 1422 // Check if we are in the singly linked portion of the 1423 // _entry_list. If we are the head then we try to remove ourselves, 1424 // else we convert to the doubly linked list. 1425 if (currentNode->prev() == nullptr) { 1426 ObjectWaiter* w = AtomicAccess::load(&_entry_list); 1427 1428 assert(w != nullptr, "invariant"); 1429 if (w == currentNode) { 1430 ObjectWaiter* next = currentNode->next(); 1431 // currentNode is at the head of _entry_list. 1432 if (AtomicAccess::cmpxchg(&_entry_list, w, next) == w) { 1433 // The CAS above sucsessfully unlinked currentNode from the 1434 // head of the _entry_list. 1435 assert(_entry_list != w, "invariant"); 1436 next->_prev = nullptr; 1437 currentNode->set_bad_pointers(); 1438 return; 1439 } else { 1440 // The CAS above can fail from interference IFF a contending 1441 // thread "pushed" itself onto _entry_list, in which case 1442 // currentNode must now be in the interior of the 1443 // list. Fall-through to building the doubly linked list. 1444 assert(_entry_list != currentNode, "invariant"); 1445 } 1446 } 1447 // Build the doubly linked list to get hold of currentNode->prev(). 1448 entry_list_build_dll(current); 1449 assert(currentNode->prev() != nullptr, "must be"); 1450 } 1451 1452 // We now know we are unlinking currentNode from the interior of a 1453 // doubly linked list. 1454 assert(currentNode->next() != nullptr, ""); 1455 assert(currentNode->prev() != nullptr, ""); 1456 assert(currentNode != _entry_list, ""); 1457 assert(currentNode != _entry_list_tail, ""); 1458 1459 ObjectWaiter* nxt = currentNode->next(); 1460 ObjectWaiter* prv = currentNode->prev(); 1461 assert(nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 1462 assert(prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 1463 1464 nxt->_prev = prv; 1465 prv->_next = nxt; 1466 currentNode->set_bad_pointers(); 1467 } 1468 1469 // ----------------------------------------------------------------------------- 1470 // Exit support 1471 // 1472 // exit() 1473 // ~~~~~~ 1474 // Note that the collector can't reclaim the objectMonitor or deflate 1475 // the object out from underneath the thread calling ::exit() as the 1476 // thread calling ::exit() never transitions to a stable state. 1477 // This inhibits GC, which in turn inhibits asynchronous (and 1478 // inopportune) reclamation of "this". 1479 // 1480 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 1481 // There's one exception to the claim above, however. enter_internal() can call 1482 // exit() to drop a lock if the acquirer has been externally suspended. 1483 // In that case exit() is called with _thread_state == _thread_blocked, 1484 // but the monitor's _contentions field is > 0, which inhibits reclamation. 1485 // 1486 // This is the exit part of the locking protocol, often implemented in 1487 // C2_MacroAssembler::fast_unlock() 1488 // 1489 // 1. A release barrier ensures that changes to monitor meta-data 1490 // (_succ, _entry_list) and data protected by the lock will be 1491 // visible before we release the lock. 1492 // 2. Release the lock by clearing the owner. 1493 // 3. A storeload MEMBAR is needed between releasing the owner and 1494 // subsequently reading meta-data to safely determine if the lock is 1495 // contended (step 4) without an elected successor (step 5). 1496 // 4. If _entry_list is null, we are done, since there is no 1497 // other thread waiting on the lock to wake up. I.e. there is no 1498 // contention. 1499 // 5. If there is a successor (_succ is non-null), we are done. The 1500 // responsibility for guaranteeing progress-liveness has now implicitly 1501 // been moved from the exiting thread to the successor. 1502 // 6. There are waiters in the entry list (_entry_list is non-null), 1503 // but there is no successor (_succ is null), so we need to 1504 // wake up (unpark) a waiting thread to avoid stranding. 1505 // 1506 // Note that since only the current lock owner can manipulate the 1507 // _entry_list (except for pushing new threads to the head), we need to 1508 // reacquire the lock before we can wake up (unpark) a waiting thread. 1509 // 1510 // The CAS() in enter provides for safety and exclusion, while the 1511 // MEMBAR in exit provides for progress and avoids stranding. 1512 // 1513 // There is also the risk of a futile wake-up. If we drop the lock 1514 // another thread can reacquire the lock immediately, and we can 1515 // then wake a thread unnecessarily. This is benign, and we've 1516 // structured the code so the windows are short and the frequency 1517 // of such futile wakups is low. 1518 1519 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) { 1520 if (!has_owner(current)) { 1521 // Apparent unbalanced locking ... 1522 // Naively we'd like to throw IllegalMonitorStateException. 1523 // As a practical matter we can neither allocate nor throw an 1524 // exception as ::exit() can be called from leaf routines. 1525 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties. 1526 // Upon deeper reflection, however, in a properly run JVM the only 1527 // way we should encounter this situation is in the presence of 1528 // unbalanced JNI locking. TODO: CheckJNICalls. 1529 // See also: CR4414101 1530 #ifdef ASSERT 1531 LogStreamHandle(Error, monitorinflation) lsh; 1532 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT 1533 " is exiting an ObjectMonitor it does not own.", p2i(current)); 1534 lsh.print_cr("The imbalance is possibly caused by JNI locking."); 1535 print_debug_style_on(&lsh); 1536 assert(false, "Non-balanced monitor enter/exit!"); 1537 #endif 1538 return; 1539 } 1540 1541 if (_recursions != 0) { 1542 _recursions--; // this is simple recursive enter 1543 return; 1544 } 1545 1546 #if INCLUDE_JFR 1547 // get the owner's thread id for the MonitorEnter event 1548 // if it is enabled and the thread isn't suspended 1549 if (not_suspended && EventJavaMonitorEnter::is_enabled()) { 1550 _previous_owner_tid = JFR_THREAD_ID(current); 1551 } 1552 #endif 1553 1554 for (;;) { 1555 // If there is a successor we should release the lock as soon as 1556 // possible, so that the successor can acquire the lock. If there is 1557 // no successor, we might need to wake up a waiting thread. 1558 if (!has_successor()) { 1559 ObjectWaiter* w = AtomicAccess::load(&_entry_list); 1560 if (w != nullptr) { 1561 // Other threads are blocked trying to acquire the lock and 1562 // there is no successor, so it appears that an heir- 1563 // presumptive (successor) must be made ready. Since threads 1564 // are woken up in FIFO order, we need to find the tail of the 1565 // entry_list. 1566 w = entry_list_tail(current); 1567 // I'd like to write: guarantee (w->_thread != current). 1568 // But in practice an exiting thread may find itself on the entry_list. 1569 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1570 // then calls exit(). Exit release the lock by setting O._owner to null. 1571 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The 1572 // notify() operation moves T1 from O's waitset to O's entry_list. T2 then 1573 // release the lock "O". T1 resumes immediately after the ST of null into 1574 // _owner, above. T1 notices that the entry_list is populated, so it 1575 // reacquires the lock and then finds itself on the entry_list. 1576 // Given all that, we have to tolerate the circumstance where "w" is 1577 // associated with current. 1578 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1579 exit_epilog(current, w); 1580 return; 1581 } 1582 } 1583 1584 // Drop the lock. 1585 // release semantics: prior loads and stores from within the critical section 1586 // must not float (reorder) past the following store that drops the lock. 1587 // Uses a storeload to separate release_store(owner) from the 1588 // successor check. The try_set_owner_from() below uses cmpxchg() so 1589 // we get the fence down there. 1590 release_clear_owner(current); 1591 OrderAccess::storeload(); 1592 1593 // Normally the exiting thread is responsible for ensuring succession, 1594 // but if this thread observes other successors are ready or other 1595 // entering threads are spinning after it has stored null into _owner 1596 // then it can exit without waking a successor. The existence of 1597 // spinners or ready successors guarantees proper succession (liveness). 1598 // Responsibility passes to the ready or running successors. The exiting 1599 // thread delegates the duty. More precisely, if a successor already 1600 // exists this thread is absolved of the responsibility of waking 1601 // (unparking) one. 1602 1603 // The _succ variable is critical to reducing futile wakeup frequency. 1604 // _succ identifies the "heir presumptive" thread that has been made 1605 // ready (unparked) but that has not yet run. We need only one such 1606 // successor thread to guarantee progress. 1607 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 1608 // section 3.3 "Futile Wakeup Throttling" for details. 1609 // 1610 // Note that spinners in Enter() also set _succ non-null. 1611 // In the current implementation spinners opportunistically set 1612 // _succ so that exiting threads might avoid waking a successor. 1613 // Which means that the exiting thread could exit immediately without 1614 // waking a successor, if it observes a successor after it has dropped 1615 // the lock. Note that the dropped lock needs to become visible to the 1616 // spinner. 1617 1618 if (_entry_list == nullptr || has_successor()) { 1619 return; 1620 } 1621 1622 // Only the current lock owner can manipulate the entry_list 1623 // (except for pushing new threads to the head), therefore we need 1624 // to reacquire the lock. If we fail to reacquire the lock the 1625 // responsibility for ensuring succession falls to the new owner. 1626 1627 if (try_lock(current) != TryLockResult::Success) { 1628 // Some other thread acquired the lock (or the monitor was 1629 // deflated). Either way we are done. 1630 return; 1631 } 1632 1633 guarantee(has_owner(current), "invariant"); 1634 } 1635 } 1636 1637 void ObjectMonitor::exit_epilog(JavaThread* current, ObjectWaiter* Wakee) { 1638 assert(has_owner(current), "invariant"); 1639 1640 // Exit protocol: 1641 // 1. ST _succ = wakee 1642 // 2. membar #loadstore|#storestore; 1643 // 2. ST _owner = nullptr 1644 // 3. unpark(wakee) 1645 1646 oop vthread = nullptr; 1647 ParkEvent * Trigger; 1648 if (!Wakee->is_vthread()) { 1649 JavaThread* t = Wakee->thread(); 1650 assert(t != nullptr, ""); 1651 Trigger = t->_ParkEvent; 1652 set_successor(t); 1653 } else { 1654 vthread = Wakee->vthread(); 1655 assert(vthread != nullptr, ""); 1656 Trigger = ObjectMonitor::vthread_unparker_ParkEvent(); 1657 set_successor(vthread); 1658 } 1659 1660 // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again. 1661 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1662 // out-of-scope (non-extant). 1663 Wakee = nullptr; 1664 1665 // Drop the lock. 1666 // Uses a fence to separate release_store(owner) from the LD in unpark(). 1667 release_clear_owner(current); 1668 OrderAccess::fence(); 1669 1670 DTRACE_MONITOR_PROBE(contended__exit, this, object(), current); 1671 1672 if (vthread == nullptr) { 1673 // Platform thread case. 1674 Trigger->unpark(); 1675 } else if (java_lang_VirtualThread::set_onWaitingList(vthread, vthread_list_head())) { 1676 // Virtual thread case. 1677 Trigger->unpark(); 1678 } 1679 } 1680 1681 // Exits the monitor returning recursion count. _owner should 1682 // be set to current's owner_id, i.e. no ANONYMOUS_OWNER allowed. 1683 intx ObjectMonitor::complete_exit(JavaThread* current) { 1684 assert(InitDone, "Unexpectedly not initialized"); 1685 guarantee(has_owner(current), "complete_exit not owner"); 1686 1687 intx save = _recursions; // record the old recursion count 1688 _recursions = 0; // set the recursion level to be 0 1689 exit(current); // exit the monitor 1690 guarantee(!has_owner(current), "invariant"); 1691 return save; 1692 } 1693 1694 // Checks that the current THREAD owns this monitor and causes an 1695 // immediate return if it doesn't. We don't use the CHECK macro 1696 // because we want the IMSE to be the only exception that is thrown 1697 // from the call site when false is returned. Any other pending 1698 // exception is ignored. 1699 #define CHECK_OWNER() \ 1700 do { \ 1701 if (!check_owner(THREAD)) { \ 1702 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \ 1703 return; \ 1704 } \ 1705 } while (false) 1706 1707 // Returns true if the specified thread owns the ObjectMonitor. 1708 // Otherwise returns false and throws IllegalMonitorStateException 1709 // (IMSE). If there is a pending exception and the specified thread 1710 // is not the owner, that exception will be replaced by the IMSE. 1711 bool ObjectMonitor::check_owner(TRAPS) { 1712 JavaThread* current = THREAD; 1713 int64_t cur = owner_raw(); 1714 if (cur == owner_id_from(current)) { 1715 return true; 1716 } 1717 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(), 1718 "current thread is not owner", false); 1719 } 1720 1721 static void post_monitor_wait_event(EventJavaMonitorWait* event, 1722 ObjectMonitor* monitor, 1723 uint64_t notifier_tid, 1724 jlong timeout, 1725 bool timedout) { 1726 assert(event != nullptr, "invariant"); 1727 assert(monitor != nullptr, "invariant"); 1728 const Klass* monitor_klass = monitor->object()->klass(); 1729 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) { 1730 return; 1731 } 1732 event->set_monitorClass(monitor_klass); 1733 event->set_timeout(timeout); 1734 // Set an address that is 'unique enough', such that events close in 1735 // time and with the same address are likely (but not guaranteed) to 1736 // belong to the same object. 1737 event->set_address((uintptr_t)monitor); 1738 event->set_notifier(notifier_tid); 1739 event->set_timedOut(timedout); 1740 event->commit(); 1741 } 1742 1743 static void vthread_monitor_waited_event(JavaThread* current, ObjectWaiter* node, ContinuationWrapper& cont, EventJavaMonitorWait* event, jboolean timed_out) { 1744 // Since we might safepoint set the anchor so that the stack can we walked. 1745 assert(current->last_continuation() != nullptr, ""); 1746 JavaFrameAnchor* anchor = current->frame_anchor(); 1747 anchor->set_last_Java_sp(current->last_continuation()->entry_sp()); 1748 anchor->set_last_Java_pc(current->last_continuation()->entry_pc()); 1749 1750 ContinuationWrapper::SafepointOp so(current, cont); 1751 1752 JRT_BLOCK 1753 if (event->should_commit()) { 1754 long timeout = java_lang_VirtualThread::timeout(current->vthread()); 1755 post_monitor_wait_event(event, node->_monitor, node->_notifier_tid, timeout, timed_out); 1756 } 1757 if (JvmtiExport::should_post_monitor_waited()) { 1758 // We mark this call in case of an upcall to Java while posting the event. 1759 // If somebody walks the stack in that case, processing the enterSpecial 1760 // frame should not include processing callee arguments since there is no 1761 // actual callee (see nmethod::preserve_callee_argument_oops()). 1762 ThreadOnMonitorWaitedEvent tmwe(current); 1763 JvmtiExport::vthread_post_monitor_waited(current, node->_monitor, timed_out); 1764 } 1765 JRT_BLOCK_END 1766 current->frame_anchor()->clear(); 1767 } 1768 1769 // ----------------------------------------------------------------------------- 1770 // Wait/Notify/NotifyAll 1771 // 1772 // Note: a subset of changes to ObjectMonitor::wait() 1773 // will need to be replicated in complete_exit 1774 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1775 JavaThread* current = THREAD; 1776 1777 assert(InitDone, "Unexpectedly not initialized"); 1778 1779 CHECK_OWNER(); // Throws IMSE if not owner. 1780 1781 EventJavaMonitorWait wait_event; 1782 EventVirtualThreadPinned vthread_pinned_event; 1783 1784 // check for a pending interrupt 1785 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 1786 JavaThreadInObjectWaitState jtiows(current, millis != 0, interruptible); 1787 1788 if (JvmtiExport::should_post_monitor_wait()) { 1789 JvmtiExport::post_monitor_wait(current, object(), millis); 1790 } 1791 // post monitor waited event. Note that this is past-tense, we are done waiting. 1792 if (JvmtiExport::should_post_monitor_waited()) { 1793 // Note: 'false' parameter is passed here because the 1794 // wait was not timed out due to thread interrupt. 1795 JvmtiExport::post_monitor_waited(current, this, false); 1796 1797 // In this short circuit of the monitor wait protocol, the 1798 // current thread never drops ownership of the monitor and 1799 // never gets added to the wait queue so the current thread 1800 // cannot be made the successor. This means that the 1801 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1802 // consume an unpark() meant for the ParkEvent associated with 1803 // this ObjectMonitor. 1804 } 1805 if (wait_event.should_commit()) { 1806 post_monitor_wait_event(&wait_event, this, 0, millis, false); 1807 } 1808 THROW(vmSymbols::java_lang_InterruptedException()); 1809 return; 1810 } 1811 1812 freeze_result result; 1813 ContinuationEntry* ce = current->last_continuation(); 1814 bool is_virtual = ce != nullptr && ce->is_virtual_thread(); 1815 if (is_virtual) { 1816 if (interruptible && JvmtiExport::should_post_monitor_wait()) { 1817 JvmtiExport::post_monitor_wait(current, object(), millis); 1818 } 1819 current->set_current_waiting_monitor(this); 1820 result = Continuation::try_preempt(current, ce->cont_oop(current)); 1821 if (result == freeze_ok) { 1822 vthread_wait(current, millis, interruptible); 1823 current->set_current_waiting_monitor(nullptr); 1824 return; 1825 } 1826 } 1827 // The jtiows does nothing for non-interruptible. 1828 JavaThreadInObjectWaitState jtiows(current, millis != 0, interruptible); 1829 1830 if (!is_virtual) { // it was already set for virtual thread 1831 if (interruptible && JvmtiExport::should_post_monitor_wait()) { 1832 JvmtiExport::post_monitor_wait(current, object(), millis); 1833 1834 // The current thread already owns the monitor and it has not yet 1835 // been added to the wait queue so the current thread cannot be 1836 // made the successor. This means that the JVMTI_EVENT_MONITOR_WAIT 1837 // event handler cannot accidentally consume an unpark() meant for 1838 // the ParkEvent associated with this ObjectMonitor. 1839 } 1840 current->set_current_waiting_monitor(this); 1841 } 1842 // create a node to be put into the queue 1843 // Critically, after we reset() the event but prior to park(), we must check 1844 // for a pending interrupt. 1845 ObjectWaiter node(current); 1846 node.TState = ObjectWaiter::TS_WAIT; 1847 current->_ParkEvent->reset(); 1848 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1849 1850 // Enter the waiting queue, which is a circular doubly linked list in this case 1851 // but it could be a priority queue or any data structure. 1852 // _wait_set_lock protects the wait queue. Normally the wait queue is accessed only 1853 // by the owner of the monitor *except* in the case where park() 1854 // returns because of a timeout of interrupt. Contention is exceptionally rare 1855 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1856 1857 Thread::SpinAcquire(&_wait_set_lock); 1858 add_waiter(&node); 1859 Thread::SpinRelease(&_wait_set_lock); 1860 1861 intx save = _recursions; // record the old recursion count 1862 _waiters++; // increment the number of waiters 1863 _recursions = 0; // set the recursion level to be 1 1864 exit(current); // exit the monitor 1865 guarantee(!has_owner(current), "invariant"); 1866 1867 // The thread is on the wait_set list - now park() it. 1868 // On MP systems it's conceivable that a brief spin before we park 1869 // could be profitable. 1870 // 1871 // TODO-FIXME: change the following logic to a loop of the form 1872 // while (!timeout && !interrupted && _notified == 0) park() 1873 1874 int ret = OS_OK; 1875 int WasNotified = 0; 1876 1877 // Need to check interrupt state whilst still _thread_in_vm 1878 bool interrupted = interruptible && current->is_interrupted(false); 1879 1880 { // State transition wrappers 1881 OSThread* osthread = current->osthread(); 1882 OSThreadWaitState osts(osthread, true); 1883 1884 assert(current->thread_state() == _thread_in_vm, "invariant"); 1885 1886 { 1887 ClearSuccOnSuspend csos(this); 1888 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */); 1889 if (interrupted || HAS_PENDING_EXCEPTION) { 1890 // Intentionally empty 1891 } else if (!node._notified) { 1892 if (millis <= 0) { 1893 current->_ParkEvent->park(); 1894 } else { 1895 ret = current->_ParkEvent->park(millis); 1896 } 1897 } 1898 } 1899 1900 // Node may be on the wait_set, or on the entry_list, or in transition 1901 // from the wait_set to the entry_list. 1902 // See if we need to remove Node from the wait_set. 1903 // We use double-checked locking to avoid grabbing _wait_set_lock 1904 // if the thread is not on the wait queue. 1905 // 1906 // Note that we don't need a fence before the fetch of TState. 1907 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1908 // written by the is thread. (perhaps the fetch might even be satisfied 1909 // by a look-aside into the processor's own store buffer, although given 1910 // the length of the code path between the prior ST and this load that's 1911 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1912 // then we'll acquire the lock and then re-fetch a fresh TState value. 1913 // That is, we fail toward safety. 1914 1915 if (node.TState == ObjectWaiter::TS_WAIT) { 1916 Thread::SpinAcquire(&_wait_set_lock); 1917 if (node.TState == ObjectWaiter::TS_WAIT) { 1918 dequeue_specific_waiter(&node); // unlink from wait_set 1919 assert(!node._notified, "invariant"); 1920 node.TState = ObjectWaiter::TS_RUN; 1921 } 1922 Thread::SpinRelease(&_wait_set_lock); 1923 } 1924 1925 // The thread is now either on off-list (TS_RUN), 1926 // or on the entry_list (TS_ENTER). 1927 // The Node's TState variable is stable from the perspective of this thread. 1928 // No other threads will asynchronously modify TState. 1929 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1930 OrderAccess::loadload(); 1931 if (has_successor(current)) clear_successor(); 1932 WasNotified = node._notified; 1933 1934 // Reentry phase -- reacquire the monitor. 1935 // re-enter contended monitor after object.wait(). 1936 // retain OBJECT_WAIT state until re-enter successfully completes 1937 // Thread state is thread_in_vm and oop access is again safe, 1938 // although the raw address of the object may have changed. 1939 // (Don't cache naked oops over safepoints, of course). 1940 1941 // post monitor waited event. Note that this is past-tense, we are done waiting. 1942 if (JvmtiExport::should_post_monitor_waited()) { 1943 JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT); 1944 1945 if (node._notified && has_successor(current)) { 1946 // In this part of the monitor wait-notify-reenter protocol it 1947 // is possible (and normal) for another thread to do a fastpath 1948 // monitor enter-exit while this thread is still trying to get 1949 // to the reenter portion of the protocol. 1950 // 1951 // The ObjectMonitor was notified and the current thread is 1952 // the successor which also means that an unpark() has already 1953 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1954 // consume the unpark() that was done when the successor was 1955 // set because the same ParkEvent is shared between Java 1956 // monitors and JVM/TI RawMonitors (for now). 1957 // 1958 // We redo the unpark() to ensure forward progress, i.e., we 1959 // don't want all pending threads hanging (parked) with none 1960 // entering the unlocked monitor. 1961 current->_ParkEvent->unpark(); 1962 } 1963 } 1964 1965 if (wait_event.should_commit()) { 1966 post_monitor_wait_event(&wait_event, this, node._notifier_tid, millis, ret == OS_TIMEOUT); 1967 } 1968 1969 OrderAccess::fence(); 1970 1971 assert(!has_owner(current), "invariant"); 1972 ObjectWaiter::TStates v = node.TState; 1973 if (v == ObjectWaiter::TS_RUN) { 1974 // We use the NoPreemptMark for the very rare case where the previous 1975 // preempt attempt failed due to OOM. The preempt on monitor contention 1976 // could succeed but we can't unmount now. 1977 NoPreemptMark npm(current); 1978 enter(current); 1979 } else { 1980 guarantee(v == ObjectWaiter::TS_ENTER, "invariant"); 1981 reenter_internal(current, &node); 1982 node.wait_reenter_end(this); 1983 } 1984 1985 // current has reacquired the lock. 1986 // Lifecycle - the node representing current must not appear on any queues. 1987 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1988 // want residual elements associated with this thread left on any lists. 1989 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1990 assert(has_owner(current), "invariant"); 1991 assert(!has_successor(current), "invariant"); 1992 } // OSThreadWaitState() 1993 1994 current->set_current_waiting_monitor(nullptr); 1995 1996 guarantee(_recursions == 0, "invariant"); 1997 int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current); 1998 _recursions = save // restore the old recursion count 1999 + relock_count; // increased by the deferred relock count 2000 current->inc_held_monitor_count(relock_count); // Deopt never entered these counts. 2001 _waiters--; // decrement the number of waiters 2002 2003 // Verify a few postconditions 2004 assert(has_owner(current), "invariant"); 2005 assert(!has_successor(current), "invariant"); 2006 assert_mark_word_consistency(); 2007 2008 if (ce != nullptr && ce->is_virtual_thread()) { 2009 current->post_vthread_pinned_event(&vthread_pinned_event, "Object.wait", result); 2010 } 2011 2012 // check if the notification happened 2013 if (!WasNotified) { 2014 // no, it could be timeout or Thread.interrupt() or both 2015 // check for interrupt event, otherwise it is timeout 2016 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) { 2017 THROW(vmSymbols::java_lang_InterruptedException()); 2018 } 2019 } 2020 2021 // NOTE: Spurious wake up will be consider as timeout. 2022 // Monitor notify has precedence over thread interrupt. 2023 } 2024 2025 // Consider: 2026 // If the lock is cool (entry_list == null && succ == null) and we're on an MP system 2027 // then instead of transferring a thread from the wait_set to the entry_list 2028 // we might just dequeue a thread from the wait_set and directly unpark() it. 2029 2030 bool ObjectMonitor::notify_internal(JavaThread* current) { 2031 bool did_notify = false; 2032 Thread::SpinAcquire(&_wait_set_lock); 2033 ObjectWaiter* iterator = dequeue_waiter(); 2034 if (iterator != nullptr) { 2035 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 2036 guarantee(!iterator->_notified, "invariant"); 2037 2038 if (iterator->is_vthread()) { 2039 oop vthread = iterator->vthread(); 2040 java_lang_VirtualThread::set_notified(vthread, true); 2041 int old_state = java_lang_VirtualThread::state(vthread); 2042 // If state is not WAIT/TIMED_WAIT then target could still be on 2043 // unmount transition, or wait could have already timed-out or target 2044 // could have been interrupted. In the first case, the target itself 2045 // will set the state to BLOCKED at the end of the unmount transition. 2046 // In the other cases the target would have been already unblocked so 2047 // there is nothing to do. 2048 if (old_state == java_lang_VirtualThread::WAIT || 2049 old_state == java_lang_VirtualThread::TIMED_WAIT) { 2050 java_lang_VirtualThread::cmpxchg_state(vthread, old_state, java_lang_VirtualThread::BLOCKED); 2051 } 2052 // Increment counter *before* adding the vthread to the _entry_list. 2053 // Adding to _entry_list uses Atomic::cmpxchg() which already provides 2054 // a fence that prevents reordering of the stores. 2055 inc_unmounted_vthreads(); 2056 } 2057 2058 iterator->_notified = true; 2059 iterator->_notifier_tid = JFR_THREAD_ID(current); 2060 did_notify = true; 2061 add_to_entry_list(current, iterator); 2062 2063 // _wait_set_lock protects the wait queue, not the entry_list. We could 2064 // move the add-to-entry_list operation, above, outside the critical section 2065 // protected by _wait_set_lock. In practice that's not useful. With the 2066 // exception of wait() timeouts and interrupts the monitor owner 2067 // is the only thread that grabs _wait_set_lock. There's almost no contention 2068 // on _wait_set_lock so it's not profitable to reduce the length of the 2069 // critical section. 2070 2071 if (!iterator->is_vthread()) { 2072 iterator->wait_reenter_begin(this); 2073 2074 // Read counter *after* adding the thread to the _entry_list. 2075 // Adding to _entry_list uses Atomic::cmpxchg() which already provides 2076 // a fence that prevents this load from floating up previous store. 2077 if (has_unmounted_vthreads()) { 2078 // Wake up the thread to alleviate some deadlocks cases where the successor 2079 // that will be picked up when this thread releases the monitor is an unmounted 2080 // virtual thread that cannot run due to having run out of carriers. Upon waking 2081 // up, the thread will call reenter_internal() which will use timed-park in case 2082 // there is contention and there are still vthreads in the _entry_list. 2083 // If the target was interrupted or the wait timed-out at the same time, it could 2084 // have reached reenter_internal and read a false value of has_unmounted_vthreads() 2085 // before we added it to the _entry_list above. To fix that, we set _do_timed_park 2086 // which will be read by the target on the next loop iteration in reenter_internal. 2087 iterator->_do_timed_park = true; 2088 JavaThread* t = iterator->thread(); 2089 t->_ParkEvent->unpark(); 2090 } 2091 } 2092 } 2093 Thread::SpinRelease(&_wait_set_lock); 2094 return did_notify; 2095 } 2096 2097 static void post_monitor_notify_event(EventJavaMonitorNotify* event, 2098 ObjectMonitor* monitor, 2099 int notified_count) { 2100 assert(event != nullptr, "invariant"); 2101 assert(monitor != nullptr, "invariant"); 2102 const Klass* monitor_klass = monitor->object()->klass(); 2103 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) { 2104 return; 2105 } 2106 event->set_monitorClass(monitor_klass); 2107 // Set an address that is 'unique enough', such that events close in 2108 // time and with the same address are likely (but not guaranteed) to 2109 // belong to the same object. 2110 event->set_address((uintptr_t)monitor); 2111 event->set_notifiedCount(notified_count); 2112 event->commit(); 2113 } 2114 2115 // Consider: a not-uncommon synchronization bug is to use notify() when 2116 // notifyAll() is more appropriate, potentially resulting in stranded 2117 // threads; this is one example of a lost wakeup. A useful diagnostic 2118 // option is to force all notify() operations to behave as notifyAll(). 2119 // 2120 // Note: We can also detect many such problems with a "minimum wait". 2121 // When the "minimum wait" is set to a small non-zero timeout value 2122 // and the program does not hang whereas it did absent "minimum wait", 2123 // that suggests a lost wakeup bug. 2124 2125 void ObjectMonitor::notify(TRAPS) { 2126 JavaThread* current = THREAD; 2127 CHECK_OWNER(); // Throws IMSE if not owner. 2128 if (_wait_set == nullptr) { 2129 return; 2130 } 2131 2132 quick_notify(current); 2133 } 2134 2135 void ObjectMonitor::quick_notify(JavaThread* current) { 2136 assert(has_owner(current), "Precondition"); 2137 2138 EventJavaMonitorNotify event; 2139 DTRACE_MONITOR_PROBE(notify, this, object(), current); 2140 int tally = notify_internal(current) ? 1 : 0; 2141 2142 if ((tally > 0) && event.should_commit()) { 2143 post_monitor_notify_event(&event, this, /* notified_count = */ tally); 2144 } 2145 } 2146 2147 // notifyAll() transfers the waiters one-at-a-time from the waitset to 2148 // the entry_list. If the waitset is "ABCD" (where A was added first 2149 // and D last) and the entry_list is ->X->Y->Z. After a notifyAll() 2150 // the waitset will be empty and the entry_list will be 2151 // ->D->C->B->A->X->Y->Z, and the next choosen successor will be Z. 2152 2153 void ObjectMonitor::notifyAll(TRAPS) { 2154 JavaThread* current = THREAD; 2155 CHECK_OWNER(); // Throws IMSE if not owner. 2156 if (_wait_set == nullptr) { 2157 return; 2158 } 2159 2160 quick_notifyAll(current); 2161 } 2162 2163 void ObjectMonitor::quick_notifyAll(JavaThread* current) { 2164 assert(has_owner(current), "Precondition"); 2165 2166 EventJavaMonitorNotify event; 2167 DTRACE_MONITOR_PROBE(notifyAll, this, object(), current); 2168 int tally = 0; 2169 while (_wait_set != nullptr) { 2170 if (notify_internal(current)) { 2171 tally++; 2172 } 2173 } 2174 2175 if ((tally > 0) && event.should_commit()) { 2176 post_monitor_notify_event(&event, this, /* notified_count = */ tally); 2177 } 2178 } 2179 2180 void ObjectMonitor::vthread_wait(JavaThread* current, jlong millis, bool interruptible) { 2181 oop vthread = current->vthread(); 2182 ObjectWaiter* node = new ObjectWaiter(vthread, this); 2183 node->_is_wait = true; 2184 node->_interruptible = interruptible; 2185 node->TState = ObjectWaiter::TS_WAIT; 2186 java_lang_VirtualThread::set_notified(vthread, false); // Reset notified flag 2187 java_lang_VirtualThread::set_interruptible_wait(vthread, interruptible); 2188 2189 // Enter the waiting queue, which is a circular doubly linked list in this case 2190 // but it could be a priority queue or any data structure. 2191 // _wait_set_lock protects the wait queue. Normally the wait queue is accessed only 2192 // by the owner of the monitor *except* in the case where park() 2193 // returns because of a timeout or interrupt. Contention is exceptionally rare 2194 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 2195 2196 Thread::SpinAcquire(&_wait_set_lock); 2197 add_waiter(node); 2198 Thread::SpinRelease(&_wait_set_lock); 2199 2200 node->_recursions = _recursions; // record the old recursion count 2201 _recursions = 0; // set the recursion level to be 0 2202 _waiters++; // increment the number of waiters 2203 exit(current); // exit the monitor 2204 guarantee(!has_owner(current), "invariant"); 2205 2206 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread"); 2207 java_lang_VirtualThread::set_state(vthread, millis == 0 ? java_lang_VirtualThread::WAITING : java_lang_VirtualThread::TIMED_WAITING); 2208 java_lang_VirtualThread::set_timeout(vthread, millis); 2209 2210 // Save the ObjectWaiter* in the vthread since we will need it when resuming execution. 2211 java_lang_VirtualThread::set_objectWaiter(vthread, node); 2212 } 2213 2214 bool ObjectMonitor::vthread_wait_reenter(JavaThread* current, ObjectWaiter* node, ContinuationWrapper& cont) { 2215 // The first time we run after being preempted on Object.wait() we 2216 // need to check if we were interrupted or the wait timed-out, and 2217 // in that case remove ourselves from the _wait_set queue. 2218 if (node->TState == ObjectWaiter::TS_WAIT) { 2219 Thread::SpinAcquire(&_wait_set_lock); 2220 if (node->TState == ObjectWaiter::TS_WAIT) { 2221 dequeue_specific_waiter(node); // unlink from wait_set 2222 assert(!node->_notified, "invariant"); 2223 node->TState = ObjectWaiter::TS_RUN; 2224 } 2225 Thread::SpinRelease(&_wait_set_lock); 2226 } 2227 2228 // If this was an interrupted case, set the _interrupted boolean so that 2229 // once we re-acquire the monitor we know if we need to throw IE or not. 2230 ObjectWaiter::TStates state = node->TState; 2231 bool was_notified = state == ObjectWaiter::TS_ENTER; 2232 assert(was_notified || state == ObjectWaiter::TS_RUN, ""); 2233 node->_interrupted = node->_interruptible && !was_notified && current->is_interrupted(false); 2234 2235 // Post JFR and JVMTI events. If non-interruptible we are in 2236 // ObjectLocker case so we don't post anything. 2237 EventJavaMonitorWait wait_event; 2238 if (node->_interruptible && (wait_event.should_commit() || JvmtiExport::should_post_monitor_waited())) { 2239 vthread_monitor_waited_event(current, node, cont, &wait_event, !was_notified && !node->_interrupted); 2240 } 2241 2242 // Mark that we are at reenter so that we don't call this method again. 2243 node->_at_reenter = true; 2244 2245 if (!was_notified) { 2246 bool acquired = vthread_monitor_enter(current, node); 2247 if (acquired) { 2248 guarantee(_recursions == 0, "invariant"); 2249 _recursions = node->_recursions; // restore the old recursion count 2250 _waiters--; // decrement the number of waiters 2251 2252 if (node->_interrupted) { 2253 // We will throw at thaw end after finishing the mount transition. 2254 current->set_pending_interrupted_exception(true); 2255 } 2256 2257 delete node; 2258 // Clear the ObjectWaiter* from the vthread. 2259 java_lang_VirtualThread::set_objectWaiter(current->vthread(), nullptr); 2260 return true; 2261 } 2262 } else { 2263 // Already moved to _entry_list by notifier, so just add to contentions. 2264 add_to_contentions(1); 2265 } 2266 return false; 2267 } 2268 2269 // ----------------------------------------------------------------------------- 2270 // Adaptive Spinning Support 2271 // 2272 // Adaptive spin-then-block - rational spinning 2273 // 2274 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS 2275 // algorithm. 2276 // 2277 // Broadly, we can fix the spin frequency -- that is, the % of contended lock 2278 // acquisition attempts where we opt to spin -- at 100% and vary the spin count 2279 // (duration) or we can fix the count at approximately the duration of 2280 // a context switch and vary the frequency. Of course we could also 2281 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 2282 // For a description of 'Adaptive spin-then-block mutual exclusion in 2283 // multi-threaded processing,' see U.S. Pat. No. 8046758. 2284 // 2285 // This implementation varies the duration "D", where D varies with 2286 // the success rate of recent spin attempts. (D is capped at approximately 2287 // length of a round-trip context switch). The success rate for recent 2288 // spin attempts is a good predictor of the success rate of future spin 2289 // attempts. The mechanism adapts automatically to varying critical 2290 // section length (lock modality), system load and degree of parallelism. 2291 // D is maintained per-monitor in _SpinDuration and is initialized 2292 // optimistically. Spin frequency is fixed at 100%. 2293 // 2294 // Note that _SpinDuration is volatile, but we update it without locks 2295 // or atomics. The code is designed so that _SpinDuration stays within 2296 // a reasonable range even in the presence of races. The arithmetic 2297 // operations on _SpinDuration are closed over the domain of legal values, 2298 // so at worst a race will install and older but still legal value. 2299 // At the very worst this introduces some apparent non-determinism. 2300 // We might spin when we shouldn't or vice-versa, but since the spin 2301 // count are relatively short, even in the worst case, the effect is harmless. 2302 // 2303 // Care must be taken that a low "D" value does not become an 2304 // an absorbing state. Transient spinning failures -- when spinning 2305 // is overall profitable -- should not cause the system to converge 2306 // on low "D" values. We want spinning to be stable and predictable 2307 // and fairly responsive to change and at the same time we don't want 2308 // it to oscillate, become metastable, be "too" non-deterministic, 2309 // or converge on or enter undesirable stable absorbing states. 2310 // 2311 // We implement a feedback-based control system -- using past behavior 2312 // to predict future behavior. We face two issues: (a) if the 2313 // input signal is random then the spin predictor won't provide optimal 2314 // results, and (b) if the signal frequency is too high then the control 2315 // system, which has some natural response lag, will "chase" the signal. 2316 // (b) can arise from multimodal lock hold times. Transient preemption 2317 // can also result in apparent bimodal lock hold times. 2318 // Although sub-optimal, neither condition is particularly harmful, as 2319 // in the worst-case we'll spin when we shouldn't or vice-versa. 2320 // The maximum spin duration is rather short so the failure modes aren't bad. 2321 // To be conservative, I've tuned the gain in system to bias toward 2322 // _not spinning. Relatedly, the system can sometimes enter a mode where it 2323 // "rings" or oscillates between spinning and not spinning. This happens 2324 // when spinning is just on the cusp of profitability, however, so the 2325 // situation is not dire. The state is benign -- there's no need to add 2326 // hysteresis control to damp the transition rate between spinning and 2327 // not spinning. 2328 2329 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool 2330 2331 static int Knob_Bonus = 100; // spin success bonus 2332 static int Knob_Penalty = 200; // spin failure penalty 2333 static int Knob_Poverty = 1000; 2334 static int Knob_FixedSpin = 0; 2335 static int Knob_PreSpin = 10; // 20-100 likely better, but it's not better in my testing. 2336 2337 inline static int adjust_up(int spin_duration) { 2338 int x = spin_duration; 2339 if (x < ObjectMonitor::Knob_SpinLimit) { 2340 if (x < Knob_Poverty) { 2341 x = Knob_Poverty; 2342 } 2343 return x + Knob_Bonus; 2344 } else { 2345 return spin_duration; 2346 } 2347 } 2348 2349 inline static int adjust_down(int spin_duration) { 2350 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2351 // AIMD is globally stable. 2352 int x = spin_duration; 2353 if (x > 0) { 2354 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2355 // This is globally sample and tends to damp the response. 2356 x -= Knob_Penalty; 2357 if (x < 0) { x = 0; } 2358 return x; 2359 } else { 2360 return spin_duration; 2361 } 2362 } 2363 2364 bool ObjectMonitor::short_fixed_spin(JavaThread* current, int spin_count, bool adapt) { 2365 for (int ctr = 0; ctr < spin_count; ctr++) { 2366 TryLockResult status = try_lock(current); 2367 if (status == TryLockResult::Success) { 2368 if (adapt) { 2369 _SpinDuration = adjust_up(_SpinDuration); 2370 } 2371 return true; 2372 } else if (status == TryLockResult::Interference) { 2373 break; 2374 } 2375 SpinPause(); 2376 } 2377 return false; 2378 } 2379 2380 // Spinning: Fixed frequency (100%), vary duration 2381 bool ObjectMonitor::try_spin(JavaThread* current) { 2382 2383 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 2384 int knob_fixed_spin = Knob_FixedSpin; // 0 (don't spin: default), 2000 good test 2385 if (knob_fixed_spin > 0) { 2386 return short_fixed_spin(current, knob_fixed_spin, false); 2387 } 2388 2389 // Admission control - verify preconditions for spinning 2390 // 2391 // We always spin a little bit, just to prevent _SpinDuration == 0 from 2392 // becoming an absorbing state. Put another way, we spin briefly to 2393 // sample, just in case the system load, parallelism, contention, or lock 2394 // modality changed. 2395 2396 int knob_pre_spin = Knob_PreSpin; // 10 (default), 100, 1000 or 2000 2397 if (short_fixed_spin(current, knob_pre_spin, true)) { 2398 return true; 2399 } 2400 2401 // 2402 // Consider the following alternative: 2403 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2404 // spin attempt. "Periodically" might mean after a tally of 2405 // the # of failed spin attempts (or iterations) reaches some threshold. 2406 // This takes us into the realm of 1-out-of-N spinning, where we 2407 // hold the duration constant but vary the frequency. 2408 2409 int ctr = _SpinDuration; 2410 if (ctr <= 0) return false; 2411 2412 // We're good to spin ... spin ingress. 2413 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2414 // when preparing to LD...CAS _owner, etc and the CAS is likely 2415 // to succeed. 2416 if (!has_successor()) { 2417 set_successor(current); 2418 } 2419 int64_t prv = NO_OWNER; 2420 2421 // There are three ways to exit the following loop: 2422 // 1. A successful spin where this thread has acquired the lock. 2423 // 2. Spin failure with prejudice 2424 // 3. Spin failure without prejudice 2425 2426 while (--ctr >= 0) { 2427 2428 // Periodic polling -- Check for pending GC 2429 // Threads may spin while they're unsafe. 2430 // We don't want spinning threads to delay the JVM from reaching 2431 // a stop-the-world safepoint or to steal cycles from GC. 2432 // If we detect a pending safepoint we abort in order that 2433 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2434 // this thread, if safe, doesn't steal cycles from GC. 2435 // This is in keeping with the "no loitering in runtime" rule. 2436 // We periodically check to see if there's a safepoint pending. 2437 if ((ctr & 0xFF) == 0) { 2438 // Can't call SafepointMechanism::should_process() since that 2439 // might update the poll values and we could be in a thread_blocked 2440 // state here which is not allowed so just check the poll. 2441 if (SafepointMechanism::local_poll_armed(current)) { 2442 break; 2443 } 2444 SpinPause(); 2445 } 2446 2447 // Probe _owner with TATAS 2448 // If this thread observes the monitor transition or flicker 2449 // from locked to unlocked to locked, then the odds that this 2450 // thread will acquire the lock in this spin attempt go down 2451 // considerably. The same argument applies if the CAS fails 2452 // or if we observe _owner change from one non-null value to 2453 // another non-null value. In such cases we might abort 2454 // the spin without prejudice or apply a "penalty" to the 2455 // spin count-down variable "ctr", reducing it by 100, say. 2456 2457 int64_t ox = owner_raw(); 2458 if (ox == NO_OWNER) { 2459 ox = try_set_owner_from(NO_OWNER, current); 2460 if (ox == NO_OWNER) { 2461 // The CAS succeeded -- this thread acquired ownership 2462 // Take care of some bookkeeping to exit spin state. 2463 if (has_successor(current)) { 2464 clear_successor(); 2465 } 2466 2467 // Increase _SpinDuration : 2468 // The spin was successful (profitable) so we tend toward 2469 // longer spin attempts in the future. 2470 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2471 // If we acquired the lock early in the spin cycle it 2472 // makes sense to increase _SpinDuration proportionally. 2473 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2474 _SpinDuration = adjust_up(_SpinDuration); 2475 return true; 2476 } 2477 2478 // The CAS failed ... we can take any of the following actions: 2479 // * penalize: ctr -= CASPenalty 2480 // * exit spin with prejudice -- abort without adapting spinner 2481 // * exit spin without prejudice. 2482 // * Since CAS is high-latency, retry again immediately. 2483 break; 2484 } 2485 2486 // Did lock ownership change hands ? 2487 if (ox != prv && prv != NO_OWNER) { 2488 break; 2489 } 2490 prv = ox; 2491 2492 if (!has_successor()) { 2493 set_successor(current); 2494 } 2495 } 2496 2497 // Spin failed with prejudice -- reduce _SpinDuration. 2498 if (ctr < 0) { 2499 _SpinDuration = adjust_down(_SpinDuration); 2500 } 2501 2502 if (has_successor(current)) { 2503 clear_successor(); 2504 // Invariant: after setting succ=null a contending thread 2505 // must recheck-retry _owner before parking. This usually happens 2506 // in the normal usage of try_spin(), but it's safest 2507 // to make try_spin() as foolproof as possible. 2508 OrderAccess::fence(); 2509 if (try_lock(current) == TryLockResult::Success) { 2510 return true; 2511 } 2512 } 2513 2514 return false; 2515 } 2516 2517 2518 // ----------------------------------------------------------------------------- 2519 // wait_set management ... 2520 2521 ObjectWaiter::ObjectWaiter(JavaThread* current) { 2522 _next = nullptr; 2523 _prev = nullptr; 2524 _thread = current; 2525 _monitor = nullptr; 2526 _notifier_tid = 0; 2527 _recursions = 0; 2528 TState = TS_RUN; 2529 _notified = false; 2530 _is_wait = false; 2531 _at_reenter = false; 2532 _interrupted = false; 2533 _do_timed_park = false; 2534 _active = false; 2535 } 2536 2537 ObjectWaiter::ObjectWaiter(oop vthread, ObjectMonitor* mon) : ObjectWaiter(nullptr) { 2538 assert(oopDesc::is_oop(vthread), ""); 2539 _vthread = OopHandle(JavaThread::thread_oop_storage(), vthread); 2540 _monitor = mon; 2541 } 2542 2543 ObjectWaiter::~ObjectWaiter() { 2544 if (is_vthread()) { 2545 assert(vthread() != nullptr, ""); 2546 _vthread.release(JavaThread::thread_oop_storage()); 2547 } 2548 } 2549 2550 oop ObjectWaiter::vthread() const { 2551 return _vthread.resolve(); 2552 } 2553 2554 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) { 2555 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon); 2556 } 2557 2558 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) { 2559 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active); 2560 } 2561 2562 inline void ObjectMonitor::add_waiter(ObjectWaiter* node) { 2563 assert(node != nullptr, "should not add null node"); 2564 assert(node->_prev == nullptr, "node already in list"); 2565 assert(node->_next == nullptr, "node already in list"); 2566 // put node at end of queue (circular doubly linked list) 2567 if (_wait_set == nullptr) { 2568 _wait_set = node; 2569 node->_prev = node; 2570 node->_next = node; 2571 } else { 2572 ObjectWaiter* head = _wait_set; 2573 ObjectWaiter* tail = head->_prev; 2574 assert(tail->_next == head, "invariant check"); 2575 tail->_next = node; 2576 head->_prev = node; 2577 node->_next = head; 2578 node->_prev = tail; 2579 } 2580 } 2581 2582 inline ObjectWaiter* ObjectMonitor::dequeue_waiter() { 2583 // dequeue the very first waiter 2584 ObjectWaiter* waiter = _wait_set; 2585 if (waiter) { 2586 dequeue_specific_waiter(waiter); 2587 } 2588 return waiter; 2589 } 2590 2591 inline void ObjectMonitor::dequeue_specific_waiter(ObjectWaiter* node) { 2592 assert(node != nullptr, "should not dequeue nullptr node"); 2593 assert(node->_prev != nullptr, "node already removed from list"); 2594 assert(node->_next != nullptr, "node already removed from list"); 2595 // when the waiter has woken up because of interrupt, 2596 // timeout or other spurious wake-up, dequeue the 2597 // waiter from waiting list 2598 ObjectWaiter* next = node->_next; 2599 if (next == node) { 2600 assert(node->_prev == node, "invariant check"); 2601 _wait_set = nullptr; 2602 } else { 2603 ObjectWaiter* prev = node->_prev; 2604 assert(prev->_next == node, "invariant check"); 2605 assert(next->_prev == node, "invariant check"); 2606 next->_prev = prev; 2607 prev->_next = next; 2608 if (_wait_set == node) { 2609 _wait_set = next; 2610 } 2611 } 2612 node->_next = nullptr; 2613 node->_prev = nullptr; 2614 } 2615 2616 // ----------------------------------------------------------------------------- 2617 2618 // One-shot global initialization for the sync subsystem. 2619 // We could also defer initialization and initialize on-demand 2620 // the first time we call ObjectSynchronizer::inflate(). 2621 // Initialization would be protected - like so many things - by 2622 // the MonitorCache_lock. 2623 2624 void ObjectMonitor::Initialize() { 2625 assert(!InitDone, "invariant"); 2626 2627 if (!os::is_MP()) { 2628 Knob_SpinLimit = 0; 2629 Knob_PreSpin = 0; 2630 Knob_FixedSpin = -1; 2631 } 2632 2633 _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer); 2634 2635 DEBUG_ONLY(InitDone = true;) 2636 } 2637 2638 // We can't call this during Initialize() because BarrierSet needs to be set. 2639 void ObjectMonitor::Initialize2() { 2640 _vthread_list_head = OopHandle(JavaThread::thread_oop_storage(), nullptr); 2641 _vthread_unparker_ParkEvent = ParkEvent::Allocate(nullptr); 2642 } 2643 2644 void ObjectMonitor::print_on(outputStream* st) const { 2645 // The minimal things to print for markWord printing, more can be added for debugging and logging. 2646 st->print("{contentions=0x%08x,waiters=0x%08x" 2647 ",recursions=%zd,owner=" INT64_FORMAT "}", 2648 contentions(), waiters(), recursions(), 2649 owner_raw()); 2650 } 2651 void ObjectMonitor::print() const { print_on(tty); } 2652 2653 #ifdef ASSERT 2654 // Print the ObjectMonitor like a debugger would: 2655 // 2656 // (ObjectMonitor) 0x00007fdfb6012e40 = { 2657 // _metadata = 0x0000000000000001 2658 // _object = 0x000000070ff45fd0 2659 // _pad_buf0 = { 2660 // [0] = '\0' 2661 // ... 2662 // [43] = '\0' 2663 // } 2664 // _owner = 0x0000000000000000 2665 // _previous_owner_tid = 0 2666 // _pad_buf1 = { 2667 // [0] = '\0' 2668 // ... 2669 // [47] = '\0' 2670 // } 2671 // _next_om = 0x0000000000000000 2672 // _recursions = 0 2673 // _entry_list = 0x0000000000000000 2674 // _entry_list_tail = 0x0000000000000000 2675 // _succ = 0x0000000000000000 2676 // _SpinDuration = 5000 2677 // _contentions = 0 2678 // _wait_set = 0x0000700009756248 2679 // _waiters = 1 2680 // _wait_set_lock = 0 2681 // } 2682 // 2683 void ObjectMonitor::print_debug_style_on(outputStream* st) const { 2684 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this)); 2685 st->print_cr(" _metadata = " INTPTR_FORMAT, _metadata); 2686 st->print_cr(" _object = " INTPTR_FORMAT, p2i(object_peek())); 2687 st->print_cr(" _pad_buf0 = {"); 2688 st->print_cr(" [0] = '\\0'"); 2689 st->print_cr(" ..."); 2690 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1); 2691 st->print_cr(" }"); 2692 st->print_cr(" _owner = " INT64_FORMAT, owner_raw()); 2693 st->print_cr(" _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid); 2694 st->print_cr(" _pad_buf1 = {"); 2695 st->print_cr(" [0] = '\\0'"); 2696 st->print_cr(" ..."); 2697 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1); 2698 st->print_cr(" }"); 2699 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om())); 2700 st->print_cr(" _recursions = %zd", _recursions); 2701 st->print_cr(" _entry_list = " INTPTR_FORMAT, p2i(_entry_list)); 2702 st->print_cr(" _entry_list_tail = " INTPTR_FORMAT, p2i(_entry_list_tail)); 2703 st->print_cr(" _succ = " INT64_FORMAT, successor()); 2704 st->print_cr(" _SpinDuration = %d", _SpinDuration); 2705 st->print_cr(" _contentions = %d", contentions()); 2706 st->print_cr(" _unmounted_vthreads = " INT64_FORMAT, _unmounted_vthreads); 2707 st->print_cr(" _wait_set = " INTPTR_FORMAT, p2i(_wait_set)); 2708 st->print_cr(" _waiters = %d", _waiters); 2709 st->print_cr(" _wait_set_lock = %d", _wait_set_lock); 2710 st->print_cr("}"); 2711 } 2712 #endif