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