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