1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2021, Azul Systems, Inc. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "cds/cdsConfig.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "classfile/javaThreadStatus.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "jfr/jfrEvents.hpp"
31 #include "jvm.h"
32 #include "jvmtifiles/jvmtiEnv.hpp"
33 #include "logging/log.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "memory/iterator.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "nmt/memTracker.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "runtime/atomicAccess.hpp"
40 #include "runtime/handles.inline.hpp"
41 #include "runtime/javaThread.inline.hpp"
42 #include "runtime/nonJavaThread.hpp"
43 #include "runtime/orderAccess.hpp"
44 #include "runtime/osThread.hpp"
45 #include "runtime/safepoint.hpp"
46 #include "runtime/safepointMechanism.inline.hpp"
47 #include "runtime/thread.inline.hpp"
48 #include "runtime/threadSMR.inline.hpp"
49 #include "utilities/macros.hpp"
50 #include "utilities/spinYield.hpp"
51 #if INCLUDE_JFR
52 #include "jfr/jfr.hpp"
53 #endif
54
55 THREAD_LOCAL Thread* Thread::_thr_current = nullptr;
56
57 // ======= Thread ========
58 // Base class for all threads: VMThread, WatcherThread, ConcurrentMarkSweepThread,
59 // JavaThread
60
61 Thread::Thread(MemTag mem_tag) {
62
63 DEBUG_ONLY(_run_state = PRE_CALL_RUN;)
64
65 // stack and get_thread
66 set_stack_base(nullptr);
67 set_stack_size(0);
68 _lgrp_id = -1;
69 DEBUG_ONLY(clear_suspendible_thread();)
70 DEBUG_ONLY(clear_indirectly_suspendible_thread();)
71 DEBUG_ONLY(clear_indirectly_safepoint_thread();)
72
73 // allocated data structures
74 set_osthread(nullptr);
75 set_resource_area(new (mem_tag) ResourceArea(mem_tag));
76 DEBUG_ONLY(_current_resource_mark = nullptr;)
77 set_handle_area(new (mem_tag) HandleArea(mem_tag));
78 set_metadata_handles(new (mtClass) GrowableArray<Metadata*>(30, mtClass));
79 set_last_handle_mark(nullptr);
80
81 // Initial value of zero ==> never claimed.
82 _threads_do_token = 0;
83 _threads_hazard_ptr = nullptr;
84 _threads_list_ptr = nullptr;
85 _nested_threads_hazard_ptr_cnt = 0;
86 _rcu_counter = 0;
87
88 // the handle mark links itself to last_handle_mark
89 new HandleMark(this);
90
91 // plain initialization
92 DEBUG_ONLY(_owned_locks = nullptr;)
93 NOT_PRODUCT(_skip_gcalot = false;)
94 _jvmti_env_iteration_count = 0;
95 set_allocated_bytes(0);
96 _current_pending_raw_monitor = nullptr;
97 _vm_error_callbacks = nullptr;
98
99 // thread-specific hashCode stream generator state - Marsaglia shift-xor form
100 // If we are dumping, keep ihashes constant. Note that during dumping we only
101 // ever run one java thread, and no other thread should generate ihashes either,
102 // so using a constant seed should work fine.
103 _hashStateX = CDSConfig::is_dumping_static_archive() ? 0x12345678 : os::random();
104 _hashStateY = 842502087;
105 _hashStateZ = 0x8767; // (int)(3579807591LL & 0xffff) ;
106 _hashStateW = 273326509;
107
108 // Many of the following fields are effectively final - immutable
109 // Note that nascent threads can't use the Native Monitor-Mutex
110 // construct until the _MutexEvent is initialized ...
111 // CONSIDER: instead of using a fixed set of purpose-dedicated ParkEvents
112 // we might instead use a stack of ParkEvents that we could provision on-demand.
113 // The stack would act as a cache to avoid calls to ParkEvent::Allocate()
114 // and ::Release()
115 _ParkEvent = ParkEvent::Allocate(this);
116
117 #ifdef CHECK_UNHANDLED_OOPS
118 if (CheckUnhandledOops) {
119 _unhandled_oops = new UnhandledOops(this);
120 }
121 #endif // CHECK_UNHANDLED_OOPS
122
123 // Notify the barrier set that a thread is being created. The initial
124 // thread is created before the barrier set is available. The call to
125 // BarrierSet::on_thread_create() for this thread is therefore deferred
126 // to BarrierSet::set_barrier_set().
127 BarrierSet* const barrier_set = BarrierSet::barrier_set();
128 if (barrier_set != nullptr) {
129 barrier_set->on_thread_create(this);
130 } else {
131 // Only the main thread should be created before the barrier set
132 // and that happens just before Thread::current is set. No other thread
133 // can attach as the VM is not created yet, so they can't execute this code.
134 // If the main thread creates other threads before the barrier set that is an error.
135 assert(Thread::current_or_null() == nullptr, "creating thread before barrier set");
136 }
137
138 MACOS_AARCH64_ONLY(DEBUG_ONLY(_wx_init = false));
139 }
140
141 #ifdef ASSERT
142 address Thread::stack_base() const {
143 // Note: can't report Thread::name() here as that can require a ResourceMark which we
144 // can't use because this gets called too early in the thread initialization.
145 assert(_stack_base != nullptr, "Stack base not yet set for thread id:%d (0 if not set)",
146 osthread() != nullptr ? osthread()->thread_id() : 0);
147 return _stack_base;
148 }
149 #endif
150
151 void Thread::initialize_tlab() {
152 if (UseTLAB) {
153 tlab().initialize();
154 }
155 }
156
157 void Thread::retire_tlab(ThreadLocalAllocStats* stats) {
158 // Sampling and serviceability support
159 if (tlab().end() != nullptr) {
160 incr_allocated_bytes(tlab().used_bytes());
161 heap_sampler().retire_tlab(tlab().top());
162 }
163
164 // Retire the TLAB
165 tlab().retire(stats);
166 }
167
168 void Thread::fill_tlab(HeapWord* start, size_t pre_reserved, size_t new_size) {
169 // Thread allocation sampling support
170 heap_sampler().set_tlab_top_at_sample_start(start);
171
172 // Fill the TLAB
173 tlab().fill(start, start + pre_reserved, new_size);
174 }
175
176 void Thread::initialize_thread_current() {
177 assert(_thr_current == nullptr, "Thread::current already initialized");
178 _thr_current = this;
179 assert(ThreadLocalStorage::thread() == nullptr, "ThreadLocalStorage::thread already initialized");
180 ThreadLocalStorage::set_thread(this);
181 assert(Thread::current() == ThreadLocalStorage::thread(), "TLS mismatch!");
182 }
183
184 void Thread::clear_thread_current() {
185 assert(Thread::current() == ThreadLocalStorage::thread(), "TLS mismatch!");
186 _thr_current = nullptr;
187 ThreadLocalStorage::set_thread(nullptr);
188 }
189
190 void Thread::record_stack_base_and_size() {
191 // Note: at this point, Thread object is not yet initialized. Do not rely on
192 // any members being initialized. Do not rely on Thread::current() being set.
193 // If possible, refrain from doing anything which may crash or assert since
194 // quite probably those crash dumps will be useless.
195 address base;
196 size_t size;
197 os::current_stack_base_and_size(&base, &size);
198 set_stack_base(base);
199 set_stack_size(size);
200
201 // Set stack limits after thread is initialized.
202 if (is_Java_thread()) {
203 JavaThread::cast(this)->stack_overflow_state()->initialize(stack_base(), stack_end());
204 }
205 }
206
207 void Thread::register_thread_stack_with_NMT() {
208 MemTracker::record_thread_stack(stack_end(), stack_size());
209 }
210
211 void Thread::unregister_thread_stack_with_NMT() {
212 MemTracker::release_thread_stack(stack_end(), stack_size());
213 }
214
215 void Thread::call_run() {
216 DEBUG_ONLY(_run_state = CALL_RUN;)
217
218 // At this point, Thread object should be fully initialized and
219 // Thread::current() should be set.
220
221 assert(Thread::current_or_null() != nullptr, "current thread is unset");
222 assert(Thread::current_or_null() == this, "current thread is wrong");
223
224 // Perform common initialization actions
225
226 MACOS_AARCH64_ONLY(this->init_wx());
227
228 register_thread_stack_with_NMT();
229
230 JFR_ONLY(Jfr::on_thread_start(this);)
231
232 log_debug(os, thread)("Thread %zu stack dimensions: "
233 PTR_FORMAT "-" PTR_FORMAT " (%zuk).",
234 os::current_thread_id(), p2i(stack_end()),
235 p2i(stack_base()), stack_size()/1024);
236
237 // Perform <ChildClass> initialization actions
238 DEBUG_ONLY(_run_state = PRE_RUN;)
239 this->pre_run();
240
241 // Invoke <ChildClass>::run()
242 DEBUG_ONLY(_run_state = RUN;)
243 this->run();
244 // Returned from <ChildClass>::run(). Thread finished.
245
246 // Perform common tear-down actions
247
248 assert(Thread::current_or_null() != nullptr, "current thread is unset");
249 assert(Thread::current_or_null() == this, "current thread is wrong");
250
251 // Perform <ChildClass> tear-down actions
252 DEBUG_ONLY(_run_state = POST_RUN;)
253 this->post_run();
254
255 // Note: at this point the thread object may already have deleted itself,
256 // so from here on do not dereference *this*. Not all thread types currently
257 // delete themselves when they terminate. But no thread should ever be deleted
258 // asynchronously with respect to its termination - that is what _run_state can
259 // be used to check.
260
261 // Logically we should do this->unregister_thread_stack_with_NMT() here, but we
262 // had to move that into post_run() because of the `this` deletion issue.
263
264 assert(Thread::current_or_null() == nullptr, "current thread still present");
265 }
266
267 Thread::~Thread() {
268
269 // Attached threads will remain in PRE_CALL_RUN, as will threads that don't actually
270 // get started due to errors etc. Any active thread should at least reach post_run
271 // before it is deleted (usually in post_run()).
272 assert(_run_state == PRE_CALL_RUN ||
273 _run_state == POST_RUN, "Active Thread deleted before post_run(): "
274 "_run_state=%d", (int)_run_state);
275
276 // Notify the barrier set that a thread is being destroyed. Note that a barrier
277 // set might not be available if we encountered errors during bootstrapping.
278 BarrierSet* const barrier_set = BarrierSet::barrier_set();
279 if (barrier_set != nullptr) {
280 barrier_set->on_thread_destroy(this);
281 }
282
283 // deallocate data structures
284 delete resource_area();
285 // since the handle marks are using the handle area, we have to deallocated the root
286 // handle mark before deallocating the thread's handle area,
287 assert(last_handle_mark() != nullptr, "check we have an element");
288 delete last_handle_mark();
289 assert(last_handle_mark() == nullptr, "check we have reached the end");
290
291 ParkEvent::Release(_ParkEvent);
292 // Set to null as a termination indicator for has_terminated().
293 AtomicAccess::store(&_ParkEvent, (ParkEvent*)nullptr);
294
295 delete handle_area();
296 delete metadata_handles();
297
298 // osthread() can be null, if creation of thread failed.
299 if (osthread() != nullptr) os::free_thread(osthread());
300
301 // Clear Thread::current if thread is deleting itself and it has not
302 // already been done. This must be done before the memory is deallocated.
303 // Needed to ensure JNI correctly detects non-attached threads.
304 if (this == Thread::current_or_null()) {
305 Thread::clear_thread_current();
306 }
307
308 CHECK_UNHANDLED_OOPS_ONLY(if (CheckUnhandledOops) delete unhandled_oops();)
309 }
310
311 #ifdef ASSERT
312 // A JavaThread is considered dangling if it not handshake-safe with respect to
313 // the current thread, it is not on a ThreadsList, or not at safepoint.
314 void Thread::check_for_dangling_thread_pointer(Thread *thread) {
315 assert(!thread->is_Java_thread() ||
316 JavaThread::cast(thread)->is_handshake_safe_for(Thread::current()) ||
317 !JavaThread::cast(thread)->on_thread_list() ||
318 SafepointSynchronize::is_at_safepoint() ||
319 ThreadsSMRSupport::is_a_protected_JavaThread_with_lock(JavaThread::cast(thread)),
320 "possibility of dangling Thread pointer");
321 }
322 #endif
323
324 // Is the target JavaThread protected by the calling Thread or by some other
325 // mechanism?
326 //
327 bool Thread::is_JavaThread_protected(const JavaThread* target) {
328 Thread* current_thread = Thread::current();
329
330 // Do the simplest check first:
331 if (SafepointSynchronize::is_at_safepoint()) {
332 // The target is protected since JavaThreads cannot exit
333 // while we're at a safepoint.
334 return true;
335 }
336
337 // If the target hasn't been started yet then it is trivially
338 // "protected". We assume the caller is the thread that will do
339 // the starting.
340 if (target->osthread() == nullptr || target->osthread()->get_state() <= INITIALIZED) {
341 return true;
342 }
343
344 // Now make the simple checks based on who the caller is:
345 if (current_thread == target || Threads_lock->owner() == current_thread) {
346 // Target JavaThread is self or calling thread owns the Threads_lock.
347 // Second check is the same as Threads_lock->owner_is_self(),
348 // but we already have the current thread so check directly.
349 return true;
350 }
351
352 // Check the ThreadsLists associated with the calling thread (if any)
353 // to see if one of them protects the target JavaThread:
354 if (is_JavaThread_protected_by_TLH(target)) {
355 return true;
356 }
357
358 // Use this debug code with -XX:+UseNewCode to diagnose locations that
359 // are missing a ThreadsListHandle or other protection mechanism:
360 // guarantee(!UseNewCode, "current_thread=" INTPTR_FORMAT " is not protecting target="
361 // INTPTR_FORMAT, p2i(current_thread), p2i(target));
362
363 // Note: Since 'target' isn't protected by a TLH, the call to
364 // target->is_handshake_safe_for() may crash, but we have debug bits so
365 // we'll be able to figure out what protection mechanism is missing.
366 assert(target->is_handshake_safe_for(current_thread), "JavaThread=" INTPTR_FORMAT
367 " is not protected and not handshake safe.", p2i(target));
368
369 // The target JavaThread is not protected so it is not safe to query:
370 return false;
371 }
372
373 // Is the target JavaThread protected by a ThreadsListHandle (TLH) associated
374 // with the calling Thread?
375 //
376 bool Thread::is_JavaThread_protected_by_TLH(const JavaThread* target) {
377 Thread* current_thread = Thread::current();
378
379 // Check the ThreadsLists associated with the calling thread (if any)
380 // to see if one of them protects the target JavaThread:
381 for (SafeThreadsListPtr* stlp = current_thread->_threads_list_ptr;
382 stlp != nullptr; stlp = stlp->previous()) {
383 if (stlp->list()->includes(target)) {
384 // The target JavaThread is protected by this ThreadsList:
385 return true;
386 }
387 }
388
389 // The target JavaThread is not protected by a TLH so it is not safe to query:
390 return false;
391 }
392
393 void Thread::set_priority(Thread* thread, ThreadPriority priority) {
394 DEBUG_ONLY(check_for_dangling_thread_pointer(thread);)
395 // Can return an error!
396 (void)os::set_priority(thread, priority);
397 }
398
399
400 void Thread::start(Thread* thread) {
401 // Start is different from resume in that its safety is guaranteed by context or
402 // being called from a Java method synchronized on the Thread object.
403 if (thread->is_Java_thread()) {
404 // Initialize the thread state to RUNNABLE before starting this thread.
405 // Can not set it after the thread started because we do not know the
406 // exact thread state at that time. It could be in MONITOR_WAIT or
407 // in SLEEPING or some other state.
408 java_lang_Thread::set_thread_status(JavaThread::cast(thread)->threadObj(),
409 JavaThreadStatus::RUNNABLE);
410 }
411 os::start_thread(thread);
412 }
413
414 // GC Support
415 bool Thread::claim_par_threads_do(uintx claim_token) {
416 uintx token = _threads_do_token;
417 if (token != claim_token) {
418 uintx res = AtomicAccess::cmpxchg(&_threads_do_token, token, claim_token);
419 if (res == token) {
420 return true;
421 }
422 guarantee(res == claim_token, "invariant");
423 }
424 return false;
425 }
426
427 void Thread::oops_do_no_frames(OopClosure* f, NMethodClosure* cf) {
428 // Do oop for ThreadShadow
429 f->do_oop((oop*)&_pending_exception);
430 handle_area()->oops_do(f);
431 }
432
433 // If the caller is a NamedThread, then remember, in the current scope,
434 // the given JavaThread in its _processed_thread field.
435 class RememberProcessedThread: public StackObj {
436 NamedThread* _cur_thr;
437 public:
438 RememberProcessedThread(Thread* thread) {
439 Thread* self = Thread::current();
440 if (self->is_Named_thread()) {
441 _cur_thr = (NamedThread *)self;
442 assert(_cur_thr->processed_thread() == nullptr, "nesting not supported");
443 _cur_thr->set_processed_thread(thread);
444 } else {
445 _cur_thr = nullptr;
446 }
447 }
448
449 ~RememberProcessedThread() {
450 if (_cur_thr) {
451 assert(_cur_thr->processed_thread() != nullptr, "nesting not supported");
452 _cur_thr->set_processed_thread(nullptr);
453 }
454 }
455 };
456
457 void Thread::oops_do(OopClosure* f, NMethodClosure* cf) {
458 // Record JavaThread to GC thread
459 RememberProcessedThread rpt(this);
460 oops_do_no_frames(f, cf);
461 oops_do_frames(f, cf);
462 }
463
464 void Thread::metadata_handles_do(void f(Metadata*)) {
465 // Only walk the Handles in Thread.
466 if (metadata_handles() != nullptr) {
467 for (int i = 0; i< metadata_handles()->length(); i++) {
468 f(metadata_handles()->at(i));
469 }
470 }
471 }
472
473 void Thread::print_on(outputStream* st, bool print_extended_info) const {
474 // get_priority assumes osthread initialized
475 if (osthread() != nullptr) {
476 int os_prio;
477 if (os::get_native_priority(this, &os_prio) == OS_OK) {
478 st->print("os_prio=%d ", os_prio);
479 }
480
481 st->print("cpu=%.2fms ",
482 (double)os::thread_cpu_time(const_cast<Thread*>(this), true) / 1000000.0
483 );
484 st->print("elapsed=%.2fs ",
485 (double)_statistical_info.getElapsedTime() / 1000.0
486 );
487 if (is_Java_thread() && (PrintExtendedThreadInfo || print_extended_info)) {
488 size_t allocated_bytes = (size_t) const_cast<Thread*>(this)->cooked_allocated_bytes();
489 st->print("allocated=%zu%s ",
490 byte_size_in_proper_unit(allocated_bytes),
491 proper_unit_for_byte_size(allocated_bytes)
492 );
493 st->print("defined_classes=" INT64_FORMAT " ", _statistical_info.getDefineClassCount());
494 }
495
496 st->print("tid=" INTPTR_FORMAT " ", p2i(this));
497 if (!is_Java_thread() || !JavaThread::cast(this)->is_vthread_mounted()) {
498 osthread()->print_on(st);
499 }
500 }
501 ThreadsSMRSupport::print_info_on(this, st);
502 st->print(" ");
503 DEBUG_ONLY(if (WizardMode) print_owned_locks_on(st);)
504 }
505
506 void Thread::print() const { print_on(tty); }
507
508 // Thread::print_on_error() is called by fatal error handler. Don't use
509 // any lock or allocate memory.
510 void Thread::print_on_error(outputStream* st, char* buf, int buflen) const {
511 assert(!(is_Compiler_thread() || is_Java_thread()), "Can't call name() here if it allocates");
512
513 st->print("%s \"%s\"", type_name(), name());
514
515 OSThread* os_thr = osthread();
516 if (os_thr != nullptr) {
517 st->fill_to(67);
518 if (os_thr->get_state() != ZOMBIE) {
519 // Use raw field members for stack base/size as this could be
520 // called before a thread has run enough to initialize them.
521 st->print(" [id=%d, stack(" PTR_FORMAT "," PTR_FORMAT ") (" PROPERFMT ")]",
522 osthread()->thread_id(), p2i(_stack_base - _stack_size), p2i(_stack_base),
523 PROPERFMTARGS(_stack_size));
524 } else {
525 st->print(" terminated");
526 }
527 } else {
528 st->print(" unknown state (no osThread)");
529 }
530 ThreadsSMRSupport::print_info_on(this, st);
531 }
532
533 void Thread::print_value_on(outputStream* st) const {
534 if (is_Named_thread()) {
535 st->print(" \"%s\" ", name());
536 }
537 st->print(INTPTR_FORMAT, p2i(this)); // print address
538 }
539
540 #ifdef ASSERT
541 void Thread::print_owned_locks_on(outputStream* st) const {
542 Mutex* cur = _owned_locks;
543 if (cur == nullptr) {
544 st->print(" (no locks) ");
545 } else {
546 st->print_cr(" Locks owned:");
547 while (cur) {
548 cur->print_on(st);
549 cur = cur->next();
550 }
551 }
552 }
553
554 Thread* Thread::_starting_thread = nullptr;
555
556 bool Thread::is_starting_thread(const Thread* t) {
557 assert(_starting_thread != nullptr, "invariant");
558 return t == _starting_thread;
559 }
560 #endif // ASSERT
561
562 bool Thread::set_as_starting_thread(JavaThread* jt) {
563 assert(jt != nullptr, "invariant");
564 assert(_starting_thread == nullptr, "already initialized: "
565 "_starting_thread=" INTPTR_FORMAT, p2i(_starting_thread));
566 // NOTE: this must be called from Threads::create_vm().
567 DEBUG_ONLY(_starting_thread = jt;)
568 return os::create_main_thread(jt);
569 }
570
571 // Ad-hoc mutual exclusion primitive: spin lock
572 //
573 // We employ a spin lock _only for low-contention, fixed-length
574 // short-duration critical sections where we're concerned
575 // about native mutex_t or HotSpot Mutex:: latency.
576
577 bool Thread::TrySpinAcquire(volatile int * adr) {
578 return AtomicAccess::cmpxchg(adr, 0, 1) == 0;
579 }
580
581 void Thread::SpinAcquire(volatile int * adr) {
582 if (AtomicAccess::cmpxchg(adr, 0, 1) == 0) {
583 return; // normal fast-path return
584 }
585
586 // Slow-path : We've encountered contention -- Spin/Yield/Block strategy.
587 int ctr = 0;
588 int Yields = 0;
589 for (;;) {
590 while (*adr != 0) {
591 ++ctr;
592 if ((ctr & 0xFFF) == 0 || !os::is_MP()) {
593 if (Yields > 5) {
594 os::naked_short_sleep(1);
595 } else {
596 os::naked_yield();
597 ++Yields;
598 }
599 } else {
600 SpinPause();
601 }
602 }
603 if (AtomicAccess::cmpxchg(adr, 0, 1) == 0) return;
604 }
605 }
606
607 void Thread::SpinRelease(volatile int * adr) {
608 assert(*adr != 0, "invariant");
609 // Roach-motel semantics.
610 // It's safe if subsequent LDs and STs float "up" into the critical section,
611 // but prior LDs and STs within the critical section can't be allowed
612 // to reorder or float past the ST that releases the lock.
613 // Loads and stores in the critical section - which appear in program
614 // order before the store that releases the lock - must also appear
615 // before the store that releases the lock in memory visibility order.
616 // So we need a #loadstore|#storestore "release" memory barrier before
617 // the ST of 0 into the lock-word which releases the lock.
618 AtomicAccess::release_store(adr, 0);
619 }