1 /*
  2  * Copyright (c) 1997, 2024, 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 "precompiled.hpp"
 27 #include "cds/cdsConfig.hpp"
 28 #include "classfile/javaClasses.hpp"
 29 #include "classfile/javaThreadStatus.hpp"
 30 #include "gc/shared/barrierSet.hpp"
 31 #include "jfr/jfrEvents.hpp"
 32 #include "jvm.h"
 33 #include "jvmtifiles/jvmtiEnv.hpp"
 34 #include "logging/log.hpp"
 35 #include "memory/allocation.inline.hpp"
 36 #include "memory/iterator.hpp"
 37 #include "memory/resourceArea.hpp"
 38 #include "nmt/memTracker.hpp"
 39 #include "oops/oop.inline.hpp"
 40 #include "runtime/atomic.hpp"
 41 #include "runtime/handles.inline.hpp"
 42 #include "runtime/javaThread.inline.hpp"
 43 #include "runtime/nonJavaThread.hpp"
 44 #include "runtime/orderAccess.hpp"
 45 #include "runtime/osThread.hpp"
 46 #include "runtime/safepoint.hpp"
 47 #include "runtime/safepointMechanism.inline.hpp"
 48 #include "runtime/thread.inline.hpp"
 49 #include "runtime/threadSMR.inline.hpp"
 50 #include "utilities/macros.hpp"
 51 #include "utilities/spinYield.hpp"
 52 #if INCLUDE_JFR
 53 #include "jfr/jfr.hpp"
 54 #endif
 55 
 56 #ifndef USE_LIBRARY_BASED_TLS_ONLY
 57 // Current thread is maintained as a thread-local variable
 58 THREAD_LOCAL Thread* Thread::_thr_current = nullptr;
 59 #endif
 60 
 61 // ======= Thread ========
 62 // Base class for all threads: VMThread, WatcherThread, ConcurrentMarkSweepThread,
 63 // JavaThread
 64 
 65 DEBUG_ONLY(Thread* Thread::_starting_thread = nullptr;)
 66 
 67 Thread::Thread(MemTag mem_tag) {
 68 
 69   DEBUG_ONLY(_run_state = PRE_CALL_RUN;)
 70 
 71   // stack and get_thread
 72   set_stack_base(nullptr);
 73   set_stack_size(0);
 74   set_lgrp_id(-1);
 75   DEBUG_ONLY(clear_suspendible_thread();)
 76   DEBUG_ONLY(clear_indirectly_suspendible_thread();)
 77   DEBUG_ONLY(clear_indirectly_safepoint_thread();)
 78 
 79   // allocated data structures
 80   set_osthread(nullptr);
 81   set_resource_area(new (mem_tag) ResourceArea(mem_tag));
 82   DEBUG_ONLY(_current_resource_mark = nullptr;)
 83   set_handle_area(new (mem_tag) HandleArea(mem_tag, nullptr));
 84   set_metadata_handles(new (mtClass) GrowableArray<Metadata*>(30, mtClass));
 85   set_last_handle_mark(nullptr);
 86 
 87   // Initial value of zero ==> never claimed.
 88   _threads_do_token = 0;
 89   _threads_hazard_ptr = nullptr;
 90   _threads_list_ptr = nullptr;
 91   _nested_threads_hazard_ptr_cnt = 0;
 92   _rcu_counter = 0;
 93 
 94   // the handle mark links itself to last_handle_mark
 95   new HandleMark(this);
 96 
 97   // plain initialization
 98   debug_only(_owned_locks = nullptr;)
 99   NOT_PRODUCT(_skip_gcalot = false;)
100   _jvmti_env_iteration_count = 0;
101   set_allocated_bytes(0);
102   _current_pending_raw_monitor = nullptr;
103   _vm_error_callbacks = nullptr;
104 
105   // thread-specific hashCode stream generator state - Marsaglia shift-xor form
106   // If we are dumping, keep ihashes constant. Note that during dumping we only
107   // ever run one java thread, and no other thread should generate ihashes either,
108   // so using a constant seed should work fine.
109   _hashStateX = CDSConfig::is_dumping_static_archive() ? 0x12345678 : os::random();
110   _hashStateY = 842502087;
111   _hashStateZ = 0x8767;    // (int)(3579807591LL & 0xffff) ;
112   _hashStateW = 273326509;
113 
114   // Many of the following fields are effectively final - immutable
115   // Note that nascent threads can't use the Native Monitor-Mutex
116   // construct until the _MutexEvent is initialized ...
117   // CONSIDER: instead of using a fixed set of purpose-dedicated ParkEvents
118   // we might instead use a stack of ParkEvents that we could provision on-demand.
119   // The stack would act as a cache to avoid calls to ParkEvent::Allocate()
120   // and ::Release()
121   _ParkEvent   = ParkEvent::Allocate(this);
122 
123 #ifdef CHECK_UNHANDLED_OOPS
124   if (CheckUnhandledOops) {
125     _unhandled_oops = new UnhandledOops(this);
126   }
127 #endif // CHECK_UNHANDLED_OOPS
128 
129   // Notify the barrier set that a thread is being created. The initial
130   // thread is created before the barrier set is available.  The call to
131   // BarrierSet::on_thread_create() for this thread is therefore deferred
132   // to BarrierSet::set_barrier_set().
133   BarrierSet* const barrier_set = BarrierSet::barrier_set();
134   if (barrier_set != nullptr) {
135     barrier_set->on_thread_create(this);
136   } else {
137     // Only the main thread should be created before the barrier set
138     // and that happens just before Thread::current is set. No other thread
139     // can attach as the VM is not created yet, so they can't execute this code.
140     // If the main thread creates other threads before the barrier set that is an error.
141     assert(Thread::current_or_null() == nullptr, "creating thread before barrier set");
142   }
143 
144   MACOS_AARCH64_ONLY(DEBUG_ONLY(_wx_init = false));
145 }
146 
147 #ifdef ASSERT
148 address Thread::stack_base() const {
149   // Note: can't report Thread::name() here as that can require a ResourceMark which we
150   // can't use because this gets called too early in the thread initialization.
151   assert(_stack_base != nullptr, "Stack base not yet set for thread id:%d (0 if not set)",
152          osthread() != nullptr ? osthread()->thread_id() : 0);
153   return _stack_base;
154 }
155 #endif
156 
157 void Thread::initialize_tlab() {
158   if (UseTLAB) {
159     tlab().initialize();
160   }
161 }
162 
163 void Thread::initialize_thread_current() {
164 #ifndef USE_LIBRARY_BASED_TLS_ONLY
165   assert(_thr_current == nullptr, "Thread::current already initialized");
166   _thr_current = this;
167 #endif
168   assert(ThreadLocalStorage::thread() == nullptr, "ThreadLocalStorage::thread already initialized");
169   ThreadLocalStorage::set_thread(this);
170   assert(Thread::current() == ThreadLocalStorage::thread(), "TLS mismatch!");
171 }
172 
173 void Thread::clear_thread_current() {
174   assert(Thread::current() == ThreadLocalStorage::thread(), "TLS mismatch!");
175 #ifndef USE_LIBRARY_BASED_TLS_ONLY
176   _thr_current = nullptr;
177 #endif
178   ThreadLocalStorage::set_thread(nullptr);
179 }
180 
181 void Thread::record_stack_base_and_size() {
182   // Note: at this point, Thread object is not yet initialized. Do not rely on
183   // any members being initialized. Do not rely on Thread::current() being set.
184   // If possible, refrain from doing anything which may crash or assert since
185   // quite probably those crash dumps will be useless.
186   address base;
187   size_t size;
188   os::current_stack_base_and_size(&base, &size);
189   set_stack_base(base);
190   set_stack_size(size);
191 
192   // Set stack limits after thread is initialized.
193   if (is_Java_thread()) {
194     JavaThread::cast(this)->stack_overflow_state()->initialize(stack_base(), stack_end());
195   }
196 }
197 
198 void Thread::register_thread_stack_with_NMT() {
199   MemTracker::record_thread_stack(stack_end(), stack_size());
200 }
201 
202 void Thread::unregister_thread_stack_with_NMT() {
203   MemTracker::release_thread_stack(stack_end(), stack_size());
204 }
205 
206 void Thread::call_run() {
207   DEBUG_ONLY(_run_state = CALL_RUN;)
208 
209   // At this point, Thread object should be fully initialized and
210   // Thread::current() should be set.
211 
212   assert(Thread::current_or_null() != nullptr, "current thread is unset");
213   assert(Thread::current_or_null() == this, "current thread is wrong");
214 
215   // Perform common initialization actions
216 
217   MACOS_AARCH64_ONLY(this->init_wx());
218 
219   register_thread_stack_with_NMT();
220 
221   JFR_ONLY(Jfr::on_thread_start(this);)
222 
223   log_debug(os, thread)("Thread " UINTX_FORMAT " stack dimensions: "
224     PTR_FORMAT "-" PTR_FORMAT " (" SIZE_FORMAT "k).",
225     os::current_thread_id(), p2i(stack_end()),
226     p2i(stack_base()), stack_size()/1024);
227 
228   // Perform <ChildClass> initialization actions
229   DEBUG_ONLY(_run_state = PRE_RUN;)
230   this->pre_run();
231 
232   // Invoke <ChildClass>::run()
233   DEBUG_ONLY(_run_state = RUN;)
234   this->run();
235   // Returned from <ChildClass>::run(). Thread finished.
236 
237   // Perform common tear-down actions
238 
239   assert(Thread::current_or_null() != nullptr, "current thread is unset");
240   assert(Thread::current_or_null() == this, "current thread is wrong");
241 
242   // Perform <ChildClass> tear-down actions
243   DEBUG_ONLY(_run_state = POST_RUN;)
244   this->post_run();
245 
246   // Note: at this point the thread object may already have deleted itself,
247   // so from here on do not dereference *this*. Not all thread types currently
248   // delete themselves when they terminate. But no thread should ever be deleted
249   // asynchronously with respect to its termination - that is what _run_state can
250   // be used to check.
251 
252   assert(Thread::current_or_null() == nullptr, "current thread still present");
253 }
254 
255 Thread::~Thread() {
256 
257   // Attached threads will remain in PRE_CALL_RUN, as will threads that don't actually
258   // get started due to errors etc. Any active thread should at least reach post_run
259   // before it is deleted (usually in post_run()).
260   assert(_run_state == PRE_CALL_RUN ||
261          _run_state == POST_RUN, "Active Thread deleted before post_run(): "
262          "_run_state=%d", (int)_run_state);
263 
264   // Notify the barrier set that a thread is being destroyed. Note that a barrier
265   // set might not be available if we encountered errors during bootstrapping.
266   BarrierSet* const barrier_set = BarrierSet::barrier_set();
267   if (barrier_set != nullptr) {
268     barrier_set->on_thread_destroy(this);
269   }
270 
271   // deallocate data structures
272   delete resource_area();
273   // since the handle marks are using the handle area, we have to deallocated the root
274   // handle mark before deallocating the thread's handle area,
275   assert(last_handle_mark() != nullptr, "check we have an element");
276   delete last_handle_mark();
277   assert(last_handle_mark() == nullptr, "check we have reached the end");
278 
279   ParkEvent::Release(_ParkEvent);
280   // Set to null as a termination indicator for has_terminated().
281   Atomic::store(&_ParkEvent, (ParkEvent*)nullptr);
282 
283   delete handle_area();
284   delete metadata_handles();
285 
286   // osthread() can be null, if creation of thread failed.
287   if (osthread() != nullptr) os::free_thread(osthread());
288 
289   // Clear Thread::current if thread is deleting itself and it has not
290   // already been done. This must be done before the memory is deallocated.
291   // Needed to ensure JNI correctly detects non-attached threads.
292   if (this == Thread::current_or_null()) {
293     Thread::clear_thread_current();
294   }
295 
296   CHECK_UNHANDLED_OOPS_ONLY(if (CheckUnhandledOops) delete unhandled_oops();)
297 }
298 
299 #ifdef ASSERT
300 // A JavaThread is considered dangling if it not handshake-safe with respect to
301 // the current thread, it is not on a ThreadsList, or not at safepoint.
302 void Thread::check_for_dangling_thread_pointer(Thread *thread) {
303   assert(!thread->is_Java_thread() ||
304          JavaThread::cast(thread)->is_handshake_safe_for(Thread::current()) ||
305          !JavaThread::cast(thread)->on_thread_list() ||
306          SafepointSynchronize::is_at_safepoint() ||
307          ThreadsSMRSupport::is_a_protected_JavaThread_with_lock(JavaThread::cast(thread)),
308          "possibility of dangling Thread pointer");
309 }
310 #endif
311 
312 // Is the target JavaThread protected by the calling Thread or by some other
313 // mechanism?
314 //
315 bool Thread::is_JavaThread_protected(const JavaThread* target) {
316   Thread* current_thread = Thread::current();
317 
318   // Do the simplest check first:
319   if (SafepointSynchronize::is_at_safepoint()) {
320     // The target is protected since JavaThreads cannot exit
321     // while we're at a safepoint.
322     return true;
323   }
324 
325   // If the target hasn't been started yet then it is trivially
326   // "protected". We assume the caller is the thread that will do
327   // the starting.
328   if (target->osthread() == nullptr || target->osthread()->get_state() <= INITIALIZED) {
329     return true;
330   }
331 
332   // Now make the simple checks based on who the caller is:
333   if (current_thread == target || Threads_lock->owner() == current_thread) {
334     // Target JavaThread is self or calling thread owns the Threads_lock.
335     // Second check is the same as Threads_lock->owner_is_self(),
336     // but we already have the current thread so check directly.
337     return true;
338   }
339 
340   // Check the ThreadsLists associated with the calling thread (if any)
341   // to see if one of them protects the target JavaThread:
342   if (is_JavaThread_protected_by_TLH(target)) {
343     return true;
344   }
345 
346   // Use this debug code with -XX:+UseNewCode to diagnose locations that
347   // are missing a ThreadsListHandle or other protection mechanism:
348   // guarantee(!UseNewCode, "current_thread=" INTPTR_FORMAT " is not protecting target="
349   //           INTPTR_FORMAT, p2i(current_thread), p2i(target));
350 
351   // Note: Since 'target' isn't protected by a TLH, the call to
352   // target->is_handshake_safe_for() may crash, but we have debug bits so
353   // we'll be able to figure out what protection mechanism is missing.
354   assert(target->is_handshake_safe_for(current_thread), "JavaThread=" INTPTR_FORMAT
355          " is not protected and not handshake safe.", p2i(target));
356 
357   // The target JavaThread is not protected so it is not safe to query:
358   return false;
359 }
360 
361 // Is the target JavaThread protected by a ThreadsListHandle (TLH) associated
362 // with the calling Thread?
363 //
364 bool Thread::is_JavaThread_protected_by_TLH(const JavaThread* target) {
365   Thread* current_thread = Thread::current();
366 
367   // Check the ThreadsLists associated with the calling thread (if any)
368   // to see if one of them protects the target JavaThread:
369   for (SafeThreadsListPtr* stlp = current_thread->_threads_list_ptr;
370        stlp != nullptr; stlp = stlp->previous()) {
371     if (stlp->list()->includes(target)) {
372       // The target JavaThread is protected by this ThreadsList:
373       return true;
374     }
375   }
376 
377   // The target JavaThread is not protected by a TLH so it is not safe to query:
378   return false;
379 }
380 
381 void Thread::set_priority(Thread* thread, ThreadPriority priority) {
382   debug_only(check_for_dangling_thread_pointer(thread);)
383   // Can return an error!
384   (void)os::set_priority(thread, priority);
385 }
386 
387 
388 void Thread::start(Thread* thread) {
389   // Start is different from resume in that its safety is guaranteed by context or
390   // being called from a Java method synchronized on the Thread object.
391   if (thread->is_Java_thread()) {
392     // Initialize the thread state to RUNNABLE before starting this thread.
393     // Can not set it after the thread started because we do not know the
394     // exact thread state at that time. It could be in MONITOR_WAIT or
395     // in SLEEPING or some other state.
396     java_lang_Thread::set_thread_status(JavaThread::cast(thread)->threadObj(),
397                                         JavaThreadStatus::RUNNABLE);
398   }
399   os::start_thread(thread);
400 }
401 
402 // GC Support
403 bool Thread::claim_par_threads_do(uintx claim_token) {
404   uintx token = _threads_do_token;
405   if (token != claim_token) {
406     uintx res = Atomic::cmpxchg(&_threads_do_token, token, claim_token);
407     if (res == token) {
408       return true;
409     }
410     guarantee(res == claim_token, "invariant");
411   }
412   return false;
413 }
414 
415 void Thread::oops_do_no_frames(OopClosure* f, NMethodClosure* cf) {
416   // Do oop for ThreadShadow
417   f->do_oop((oop*)&_pending_exception);
418   handle_area()->oops_do(f);
419 }
420 
421 // If the caller is a NamedThread, then remember, in the current scope,
422 // the given JavaThread in its _processed_thread field.
423 class RememberProcessedThread: public StackObj {
424   NamedThread* _cur_thr;
425 public:
426   RememberProcessedThread(Thread* thread) {
427     Thread* self = Thread::current();
428     if (self->is_Named_thread()) {
429       _cur_thr = (NamedThread *)self;
430       assert(_cur_thr->processed_thread() == nullptr, "nesting not supported");
431       _cur_thr->set_processed_thread(thread);
432     } else {
433       _cur_thr = nullptr;
434     }
435   }
436 
437   ~RememberProcessedThread() {
438     if (_cur_thr) {
439       assert(_cur_thr->processed_thread() != nullptr, "nesting not supported");
440       _cur_thr->set_processed_thread(nullptr);
441     }
442   }
443 };
444 
445 void Thread::oops_do(OopClosure* f, NMethodClosure* cf) {
446   // Record JavaThread to GC thread
447   RememberProcessedThread rpt(this);
448   oops_do_no_frames(f, cf);
449   oops_do_frames(f, cf);
450 }
451 
452 void Thread::metadata_handles_do(void f(Metadata*)) {
453   // Only walk the Handles in Thread.
454   if (metadata_handles() != nullptr) {
455     for (int i = 0; i< metadata_handles()->length(); i++) {
456       f(metadata_handles()->at(i));
457     }
458   }
459 }
460 
461 void Thread::print_on(outputStream* st, bool print_extended_info) const {
462   // get_priority assumes osthread initialized
463   if (osthread() != nullptr) {
464     int os_prio;
465     if (os::get_native_priority(this, &os_prio) == OS_OK) {
466       st->print("os_prio=%d ", os_prio);
467     }
468 
469     st->print("cpu=%.2fms ",
470               (double)os::thread_cpu_time(const_cast<Thread*>(this), true) / 1000000.0
471               );
472     st->print("elapsed=%.2fs ",
473               (double)_statistical_info.getElapsedTime() / 1000.0
474               );
475     if (is_Java_thread() && (PrintExtendedThreadInfo || print_extended_info)) {
476       size_t allocated_bytes = (size_t) const_cast<Thread*>(this)->cooked_allocated_bytes();
477       st->print("allocated=" SIZE_FORMAT "%s ",
478                 byte_size_in_proper_unit(allocated_bytes),
479                 proper_unit_for_byte_size(allocated_bytes)
480                 );
481       st->print("defined_classes=" INT64_FORMAT " ", _statistical_info.getDefineClassCount());
482     }
483 
484     st->print("tid=" INTPTR_FORMAT " ", p2i(this));
485     if (!is_Java_thread() || !JavaThread::cast(this)->is_vthread_mounted()) {
486       osthread()->print_on(st);
487     }
488   }
489   ThreadsSMRSupport::print_info_on(this, st);
490   st->print(" ");
491   debug_only(if (WizardMode) print_owned_locks_on(st);)
492 }
493 
494 void Thread::print() const { print_on(tty); }
495 
496 // Thread::print_on_error() is called by fatal error handler. Don't use
497 // any lock or allocate memory.
498 void Thread::print_on_error(outputStream* st, char* buf, int buflen) const {
499   assert(!(is_Compiler_thread() || is_Java_thread()), "Can't call name() here if it allocates");
500 
501   st->print("%s \"%s\"", type_name(), name());
502 
503   OSThread* os_thr = osthread();
504   if (os_thr != nullptr) {
505     st->fill_to(67);
506     if (os_thr->get_state() != ZOMBIE) {
507       // Use raw field members for stack base/size as this could be
508       // called before a thread has run enough to initialize them.
509       st->print(" [id=%d, stack(" PTR_FORMAT "," PTR_FORMAT ") (" PROPERFMT ")]",
510                 osthread()->thread_id(), p2i(_stack_base - _stack_size), p2i(_stack_base),
511                 PROPERFMTARGS(_stack_size));
512     } else {
513       st->print(" terminated");
514     }
515   } else {
516     st->print(" unknown state (no osThread)");
517   }
518   ThreadsSMRSupport::print_info_on(this, st);
519 }
520 
521 void Thread::print_value_on(outputStream* st) const {
522   if (is_Named_thread()) {
523     st->print(" \"%s\" ", name());
524   }
525   st->print(INTPTR_FORMAT, p2i(this));   // print address
526 }
527 
528 #ifdef ASSERT
529 void Thread::print_owned_locks_on(outputStream* st) const {
530   Mutex* cur = _owned_locks;
531   if (cur == nullptr) {
532     st->print(" (no locks) ");
533   } else {
534     st->print_cr(" Locks owned:");
535     while (cur) {
536       cur->print_on(st);
537       cur = cur->next();
538     }
539   }
540 }
541 #endif // ASSERT
542 
543 bool Thread::set_as_starting_thread() {
544   assert(_starting_thread == nullptr, "already initialized: "
545          "_starting_thread=" INTPTR_FORMAT, p2i(_starting_thread));
546   // NOTE: this must be called inside the main thread.
547   DEBUG_ONLY(_starting_thread = this;)
548   return os::create_main_thread(JavaThread::cast(this));
549 }
550 
551 // Ad-hoc mutual exclusion primitives: SpinLock
552 //
553 // We employ SpinLocks _only for low-contention, fixed-length
554 // short-duration critical sections where we're concerned
555 // about native mutex_t or HotSpot Mutex:: latency.
556 //
557 // TODO-FIXME: ListLock should be of type SpinLock.
558 // We should make this a 1st-class type, integrated into the lock
559 // hierarchy as leaf-locks.  Critically, the SpinLock structure
560 // should have sufficient padding to avoid false-sharing and excessive
561 // cache-coherency traffic.
562 
563 
564 typedef volatile int SpinLockT;
565 
566 void Thread::SpinAcquire(volatile int * adr, const char * LockName) {
567   if (Atomic::cmpxchg(adr, 0, 1) == 0) {
568     return;   // normal fast-path return
569   }
570 
571   // Slow-path : We've encountered contention -- Spin/Yield/Block strategy.
572   int ctr = 0;
573   int Yields = 0;
574   for (;;) {
575     while (*adr != 0) {
576       ++ctr;
577       if ((ctr & 0xFFF) == 0 || !os::is_MP()) {
578         if (Yields > 5) {
579           os::naked_short_sleep(1);
580         } else {
581           os::naked_yield();
582           ++Yields;
583         }
584       } else {
585         SpinPause();
586       }
587     }
588     if (Atomic::cmpxchg(adr, 0, 1) == 0) return;
589   }
590 }
591 
592 void Thread::SpinRelease(volatile int * adr) {
593   assert(*adr != 0, "invariant");
594   OrderAccess::fence();      // guarantee at least release consistency.
595   // Roach-motel semantics.
596   // It's safe if subsequent LDs and STs float "up" into the critical section,
597   // but prior LDs and STs within the critical section can't be allowed
598   // to reorder or float past the ST that releases the lock.
599   // Loads and stores in the critical section - which appear in program
600   // order before the store that releases the lock - must also appear
601   // before the store that releases the lock in memory visibility order.
602   // Conceptually we need a #loadstore|#storestore "release" MEMBAR before
603   // the ST of 0 into the lock-word which releases the lock, so fence
604   // more than covers this on all platforms.
605   *adr = 0;
606 }