xref: /aosp_15_r20/external/angle/third_party/abseil-cpp/absl/synchronization/mutex.cc (revision 8975f5c5ed3d1c378011245431ada316dfb6f244)
1 // Copyright 2017 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #include "absl/synchronization/mutex.h"
16 
17 #ifdef _WIN32
18 #include <windows.h>
19 #ifdef ERROR
20 #undef ERROR
21 #endif
22 #else
23 #include <fcntl.h>
24 #include <pthread.h>
25 #include <sched.h>
26 #include <sys/time.h>
27 #endif
28 
29 #include <assert.h>
30 #include <errno.h>
31 #include <stdio.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include <time.h>
35 
36 #include <algorithm>
37 #include <atomic>
38 #include <cstddef>
39 #include <cstdlib>
40 #include <cstring>
41 #include <thread>  // NOLINT(build/c++11)
42 
43 #include "absl/base/attributes.h"
44 #include "absl/base/call_once.h"
45 #include "absl/base/config.h"
46 #include "absl/base/dynamic_annotations.h"
47 #include "absl/base/internal/atomic_hook.h"
48 #include "absl/base/internal/cycleclock.h"
49 #include "absl/base/internal/hide_ptr.h"
50 #include "absl/base/internal/low_level_alloc.h"
51 #include "absl/base/internal/raw_logging.h"
52 #include "absl/base/internal/spinlock.h"
53 #include "absl/base/internal/sysinfo.h"
54 #include "absl/base/internal/thread_identity.h"
55 #include "absl/base/internal/tsan_mutex_interface.h"
56 #include "absl/base/optimization.h"
57 #include "absl/debugging/stacktrace.h"
58 #include "absl/debugging/symbolize.h"
59 #include "absl/synchronization/internal/graphcycles.h"
60 #include "absl/synchronization/internal/per_thread_sem.h"
61 #include "absl/time/time.h"
62 
63 using absl::base_internal::CurrentThreadIdentityIfPresent;
64 using absl::base_internal::CycleClock;
65 using absl::base_internal::PerThreadSynch;
66 using absl::base_internal::SchedulingGuard;
67 using absl::base_internal::ThreadIdentity;
68 using absl::synchronization_internal::GetOrCreateCurrentThreadIdentity;
69 using absl::synchronization_internal::GraphCycles;
70 using absl::synchronization_internal::GraphId;
71 using absl::synchronization_internal::InvalidGraphId;
72 using absl::synchronization_internal::KernelTimeout;
73 using absl::synchronization_internal::PerThreadSem;
74 
75 extern "C" {
ABSL_INTERNAL_C_SYMBOL(AbslInternalMutexYield)76 ABSL_ATTRIBUTE_WEAK void ABSL_INTERNAL_C_SYMBOL(AbslInternalMutexYield)() {
77   std::this_thread::yield();
78 }
79 }  // extern "C"
80 
81 namespace absl {
82 ABSL_NAMESPACE_BEGIN
83 
84 namespace {
85 
86 #if defined(ABSL_HAVE_THREAD_SANITIZER)
87 constexpr OnDeadlockCycle kDeadlockDetectionDefault = OnDeadlockCycle::kIgnore;
88 #else
89 constexpr OnDeadlockCycle kDeadlockDetectionDefault = OnDeadlockCycle::kAbort;
90 #endif
91 
92 ABSL_CONST_INIT std::atomic<OnDeadlockCycle> synch_deadlock_detection(
93     kDeadlockDetectionDefault);
94 ABSL_CONST_INIT std::atomic<bool> synch_check_invariants(false);
95 
96 ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES
97 absl::base_internal::AtomicHook<void (*)(int64_t wait_cycles)>
98     submit_profile_data;
99 ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES absl::base_internal::AtomicHook<void (*)(
100     const char* msg, const void* obj, int64_t wait_cycles)>
101     mutex_tracer;
102 ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES
103 absl::base_internal::AtomicHook<void (*)(const char* msg, const void* cv)>
104     cond_var_tracer;
105 
106 }  // namespace
107 
108 static inline bool EvalConditionAnnotated(const Condition* cond, Mutex* mu,
109                                           bool locking, bool trylock,
110                                           bool read_lock);
111 
RegisterMutexProfiler(void (* fn)(int64_t wait_cycles))112 void RegisterMutexProfiler(void (*fn)(int64_t wait_cycles)) {
113   submit_profile_data.Store(fn);
114 }
115 
RegisterMutexTracer(void (* fn)(const char * msg,const void * obj,int64_t wait_cycles))116 void RegisterMutexTracer(void (*fn)(const char* msg, const void* obj,
117                                     int64_t wait_cycles)) {
118   mutex_tracer.Store(fn);
119 }
120 
RegisterCondVarTracer(void (* fn)(const char * msg,const void * cv))121 void RegisterCondVarTracer(void (*fn)(const char* msg, const void* cv)) {
122   cond_var_tracer.Store(fn);
123 }
124 
125 namespace {
126 // Represents the strategy for spin and yield.
127 // See the comment in GetMutexGlobals() for more information.
128 enum DelayMode { AGGRESSIVE, GENTLE };
129 
130 struct ABSL_CACHELINE_ALIGNED MutexGlobals {
131   absl::once_flag once;
132   // Note: this variable is initialized separately in Mutex::LockSlow,
133   // so that Mutex::Lock does not have a stack frame in optimized build.
134   std::atomic<int> spinloop_iterations{0};
135   int32_t mutex_sleep_spins[2] = {};
136   absl::Duration mutex_sleep_time;
137 };
138 
139 ABSL_CONST_INIT static MutexGlobals globals;
140 
MeasureTimeToYield()141 absl::Duration MeasureTimeToYield() {
142   absl::Time before = absl::Now();
143   ABSL_INTERNAL_C_SYMBOL(AbslInternalMutexYield)();
144   return absl::Now() - before;
145 }
146 
GetMutexGlobals()147 const MutexGlobals& GetMutexGlobals() {
148   absl::base_internal::LowLevelCallOnce(&globals.once, [&]() {
149     if (absl::base_internal::NumCPUs() > 1) {
150       // If the mode is aggressive then spin many times before yielding.
151       // If the mode is gentle then spin only a few times before yielding.
152       // Aggressive spinning is used to ensure that an Unlock() call,
153       // which must get the spin lock for any thread to make progress gets it
154       // without undue delay.
155       globals.mutex_sleep_spins[AGGRESSIVE] = 5000;
156       globals.mutex_sleep_spins[GENTLE] = 250;
157       globals.mutex_sleep_time = absl::Microseconds(10);
158     } else {
159       // If this a uniprocessor, only yield/sleep. Real-time threads are often
160       // unable to yield, so the sleep time needs to be long enough to keep
161       // the calling thread asleep until scheduling happens.
162       globals.mutex_sleep_spins[AGGRESSIVE] = 0;
163       globals.mutex_sleep_spins[GENTLE] = 0;
164       globals.mutex_sleep_time = MeasureTimeToYield() * 5;
165       globals.mutex_sleep_time =
166           std::min(globals.mutex_sleep_time, absl::Milliseconds(1));
167       globals.mutex_sleep_time =
168           std::max(globals.mutex_sleep_time, absl::Microseconds(10));
169     }
170   });
171   return globals;
172 }
173 }  // namespace
174 
175 namespace synchronization_internal {
176 // Returns the Mutex delay on iteration `c` depending on the given `mode`.
177 // The returned value should be used as `c` for the next call to `MutexDelay`.
MutexDelay(int32_t c,int mode)178 int MutexDelay(int32_t c, int mode) {
179   const int32_t limit = GetMutexGlobals().mutex_sleep_spins[mode];
180   const absl::Duration sleep_time = GetMutexGlobals().mutex_sleep_time;
181   if (c < limit) {
182     // Spin.
183     c++;
184   } else {
185     SchedulingGuard::ScopedEnable enable_rescheduling;
186     ABSL_TSAN_MUTEX_PRE_DIVERT(nullptr, 0);
187     if (c == limit) {
188       // Yield once.
189       ABSL_INTERNAL_C_SYMBOL(AbslInternalMutexYield)();
190       c++;
191     } else {
192       // Then wait.
193       absl::SleepFor(sleep_time);
194       c = 0;
195     }
196     ABSL_TSAN_MUTEX_POST_DIVERT(nullptr, 0);
197   }
198   return c;
199 }
200 }  // namespace synchronization_internal
201 
202 // --------------------------Generic atomic ops
203 // Ensure that "(*pv & bits) == bits" by doing an atomic update of "*pv" to
204 // "*pv | bits" if necessary.  Wait until (*pv & wait_until_clear)==0
205 // before making any change.
206 // Returns true if bits were previously unset and set by the call.
207 // This is used to set flags in mutex and condition variable words.
AtomicSetBits(std::atomic<intptr_t> * pv,intptr_t bits,intptr_t wait_until_clear)208 static bool AtomicSetBits(std::atomic<intptr_t>* pv, intptr_t bits,
209                           intptr_t wait_until_clear) {
210   for (;;) {
211     intptr_t v = pv->load(std::memory_order_relaxed);
212     if ((v & bits) == bits) {
213       return false;
214     }
215     if ((v & wait_until_clear) != 0) {
216       continue;
217     }
218     if (pv->compare_exchange_weak(v, v | bits, std::memory_order_release,
219                                   std::memory_order_relaxed)) {
220       return true;
221     }
222   }
223 }
224 
225 //------------------------------------------------------------------
226 
227 // Data for doing deadlock detection.
228 ABSL_CONST_INIT static absl::base_internal::SpinLock deadlock_graph_mu(
229     absl::kConstInit, base_internal::SCHEDULE_KERNEL_ONLY);
230 
231 // Graph used to detect deadlocks.
232 ABSL_CONST_INIT static GraphCycles* deadlock_graph
233     ABSL_GUARDED_BY(deadlock_graph_mu) ABSL_PT_GUARDED_BY(deadlock_graph_mu);
234 
235 //------------------------------------------------------------------
236 // An event mechanism for debugging mutex use.
237 // It also allows mutexes to be given names for those who can't handle
238 // addresses, and instead like to give their data structures names like
239 // "Henry", "Fido", or "Rupert IV, King of Yondavia".
240 
241 namespace {  // to prevent name pollution
242 enum {       // Mutex and CondVar events passed as "ev" to PostSynchEvent
243              // Mutex events
244   SYNCH_EV_TRYLOCK_SUCCESS,
245   SYNCH_EV_TRYLOCK_FAILED,
246   SYNCH_EV_READERTRYLOCK_SUCCESS,
247   SYNCH_EV_READERTRYLOCK_FAILED,
248   SYNCH_EV_LOCK,
249   SYNCH_EV_LOCK_RETURNING,
250   SYNCH_EV_READERLOCK,
251   SYNCH_EV_READERLOCK_RETURNING,
252   SYNCH_EV_UNLOCK,
253   SYNCH_EV_READERUNLOCK,
254 
255   // CondVar events
256   SYNCH_EV_WAIT,
257   SYNCH_EV_WAIT_RETURNING,
258   SYNCH_EV_SIGNAL,
259   SYNCH_EV_SIGNALALL,
260 };
261 
262 enum {                    // Event flags
263   SYNCH_F_R = 0x01,       // reader event
264   SYNCH_F_LCK = 0x02,     // PostSynchEvent called with mutex held
265   SYNCH_F_TRY = 0x04,     // TryLock or ReaderTryLock
266   SYNCH_F_UNLOCK = 0x08,  // Unlock or ReaderUnlock
267 
268   SYNCH_F_LCK_W = SYNCH_F_LCK,
269   SYNCH_F_LCK_R = SYNCH_F_LCK | SYNCH_F_R,
270 };
271 }  // anonymous namespace
272 
273 // Properties of the events.
274 static const struct {
275   int flags;
276   const char* msg;
277 } event_properties[] = {
278     {SYNCH_F_LCK_W | SYNCH_F_TRY, "TryLock succeeded "},
279     {0, "TryLock failed "},
280     {SYNCH_F_LCK_R | SYNCH_F_TRY, "ReaderTryLock succeeded "},
281     {0, "ReaderTryLock failed "},
282     {0, "Lock blocking "},
283     {SYNCH_F_LCK_W, "Lock returning "},
284     {0, "ReaderLock blocking "},
285     {SYNCH_F_LCK_R, "ReaderLock returning "},
286     {SYNCH_F_LCK_W | SYNCH_F_UNLOCK, "Unlock "},
287     {SYNCH_F_LCK_R | SYNCH_F_UNLOCK, "ReaderUnlock "},
288     {0, "Wait on "},
289     {0, "Wait unblocked "},
290     {0, "Signal on "},
291     {0, "SignalAll on "},
292 };
293 
294 ABSL_CONST_INIT static absl::base_internal::SpinLock synch_event_mu(
295     absl::kConstInit, base_internal::SCHEDULE_KERNEL_ONLY);
296 
297 // Hash table size; should be prime > 2.
298 // Can't be too small, as it's used for deadlock detection information.
299 static constexpr uint32_t kNSynchEvent = 1031;
300 
301 static struct SynchEvent {  // this is a trivial hash table for the events
302   // struct is freed when refcount reaches 0
303   int refcount ABSL_GUARDED_BY(synch_event_mu);
304 
305   // buckets have linear, 0-terminated  chains
306   SynchEvent* next ABSL_GUARDED_BY(synch_event_mu);
307 
308   // Constant after initialization
309   uintptr_t masked_addr;  // object at this address is called "name"
310 
311   // No explicit synchronization used.  Instead we assume that the
312   // client who enables/disables invariants/logging on a Mutex does so
313   // while the Mutex is not being concurrently accessed by others.
314   void (*invariant)(void* arg);  // called on each event
315   void* arg;                     // first arg to (*invariant)()
316   bool log;                      // logging turned on
317 
318   // Constant after initialization
319   char name[1];  // actually longer---NUL-terminated string
320 }* synch_event[kNSynchEvent] ABSL_GUARDED_BY(synch_event_mu);
321 
322 // Ensure that the object at "addr" has a SynchEvent struct associated with it,
323 // set "bits" in the word there (waiting until lockbit is clear before doing
324 // so), and return a refcounted reference that will remain valid until
325 // UnrefSynchEvent() is called.  If a new SynchEvent is allocated,
326 // the string name is copied into it.
327 // When used with a mutex, the caller should also ensure that kMuEvent
328 // is set in the mutex word, and similarly for condition variables and kCVEvent.
EnsureSynchEvent(std::atomic<intptr_t> * addr,const char * name,intptr_t bits,intptr_t lockbit)329 static SynchEvent* EnsureSynchEvent(std::atomic<intptr_t>* addr,
330                                     const char* name, intptr_t bits,
331                                     intptr_t lockbit) {
332   uint32_t h = reinterpret_cast<uintptr_t>(addr) % kNSynchEvent;
333   synch_event_mu.Lock();
334   // When a Mutex/CondVar is destroyed, we don't remove the associated
335   // SynchEvent to keep destructors empty in release builds for performance
336   // reasons. If the current call is the first to set bits (kMuEvent/kCVEvent),
337   // we don't look up the existing even because (if it exists, it must be for
338   // the previous Mutex/CondVar that existed at the same address).
339   // The leaking events must not be a problem for tests, which should create
340   // bounded amount of events. And debug logging is not supposed to be enabled
341   // in production. However, if it's accidentally enabled, or briefly enabled
342   // for some debugging, we don't want to crash the program. Instead we drop
343   // all events, if we accumulated too many of them. Size of a single event
344   // is ~48 bytes, so 100K events is ~5 MB.
345   // Additionally we could delete the old event for the same address,
346   // but it would require a better hashmap (if we accumulate too many events,
347   // linked lists will grow and traversing them will be very slow).
348   constexpr size_t kMaxSynchEventCount = 100 << 10;
349   // Total number of live synch events.
350   static size_t synch_event_count ABSL_GUARDED_BY(synch_event_mu);
351   if (++synch_event_count > kMaxSynchEventCount) {
352     synch_event_count = 0;
353     ABSL_RAW_LOG(ERROR,
354                  "Accumulated %zu Mutex debug objects. If you see this"
355                  " in production, it may mean that the production code"
356                  " accidentally calls "
357                  "Mutex/CondVar::EnableDebugLog/EnableInvariantDebugging.",
358                  kMaxSynchEventCount);
359     for (auto*& head : synch_event) {
360       for (auto* e = head; e != nullptr;) {
361         SynchEvent* next = e->next;
362         if (--(e->refcount) == 0) {
363           base_internal::LowLevelAlloc::Free(e);
364         }
365         e = next;
366       }
367       head = nullptr;
368     }
369   }
370   SynchEvent* e = nullptr;
371   if (!AtomicSetBits(addr, bits, lockbit)) {
372     for (e = synch_event[h];
373          e != nullptr && e->masked_addr != base_internal::HidePtr(addr);
374          e = e->next) {
375     }
376   }
377   if (e == nullptr) {  // no SynchEvent struct found; make one.
378     if (name == nullptr) {
379       name = "";
380     }
381     size_t l = strlen(name);
382     e = reinterpret_cast<SynchEvent*>(
383         base_internal::LowLevelAlloc::Alloc(sizeof(*e) + l));
384     e->refcount = 2;  // one for return value, one for linked list
385     e->masked_addr = base_internal::HidePtr(addr);
386     e->invariant = nullptr;
387     e->arg = nullptr;
388     e->log = false;
389     strcpy(e->name, name);  // NOLINT(runtime/printf)
390     e->next = synch_event[h];
391     synch_event[h] = e;
392   } else {
393     e->refcount++;  // for return value
394   }
395   synch_event_mu.Unlock();
396   return e;
397 }
398 
399 // Decrement the reference count of *e, or do nothing if e==null.
UnrefSynchEvent(SynchEvent * e)400 static void UnrefSynchEvent(SynchEvent* e) {
401   if (e != nullptr) {
402     synch_event_mu.Lock();
403     bool del = (--(e->refcount) == 0);
404     synch_event_mu.Unlock();
405     if (del) {
406       base_internal::LowLevelAlloc::Free(e);
407     }
408   }
409 }
410 
411 // Return a refcounted reference to the SynchEvent of the object at address
412 // "addr", if any.  The pointer returned is valid until the UnrefSynchEvent() is
413 // called.
GetSynchEvent(const void * addr)414 static SynchEvent* GetSynchEvent(const void* addr) {
415   uint32_t h = reinterpret_cast<uintptr_t>(addr) % kNSynchEvent;
416   SynchEvent* e;
417   synch_event_mu.Lock();
418   for (e = synch_event[h];
419        e != nullptr && e->masked_addr != base_internal::HidePtr(addr);
420        e = e->next) {
421   }
422   if (e != nullptr) {
423     e->refcount++;
424   }
425   synch_event_mu.Unlock();
426   return e;
427 }
428 
429 // Called when an event "ev" occurs on a Mutex of CondVar "obj"
430 // if event recording is on
PostSynchEvent(void * obj,int ev)431 static void PostSynchEvent(void* obj, int ev) {
432   SynchEvent* e = GetSynchEvent(obj);
433   // logging is on if event recording is on and either there's no event struct,
434   // or it explicitly says to log
435   if (e == nullptr || e->log) {
436     void* pcs[40];
437     int n = absl::GetStackTrace(pcs, ABSL_ARRAYSIZE(pcs), 1);
438     // A buffer with enough space for the ASCII for all the PCs, even on a
439     // 64-bit machine.
440     char buffer[ABSL_ARRAYSIZE(pcs) * 24];
441     int pos = snprintf(buffer, sizeof(buffer), " @");
442     for (int i = 0; i != n; i++) {
443       int b = snprintf(&buffer[pos], sizeof(buffer) - static_cast<size_t>(pos),
444                        " %p", pcs[i]);
445       if (b < 0 ||
446           static_cast<size_t>(b) >= sizeof(buffer) - static_cast<size_t>(pos)) {
447         break;
448       }
449       pos += b;
450     }
451     ABSL_RAW_LOG(INFO, "%s%p %s %s", event_properties[ev].msg, obj,
452                  (e == nullptr ? "" : e->name), buffer);
453   }
454   const int flags = event_properties[ev].flags;
455   if ((flags & SYNCH_F_LCK) != 0 && e != nullptr && e->invariant != nullptr) {
456     // Calling the invariant as is causes problems under ThreadSanitizer.
457     // We are currently inside of Mutex Lock/Unlock and are ignoring all
458     // memory accesses and synchronization. If the invariant transitively
459     // synchronizes something else and we ignore the synchronization, we will
460     // get false positive race reports later.
461     // Reuse EvalConditionAnnotated to properly call into user code.
462     struct local {
463       static bool pred(SynchEvent* ev) {
464         (*ev->invariant)(ev->arg);
465         return false;
466       }
467     };
468     Condition cond(&local::pred, e);
469     Mutex* mu = static_cast<Mutex*>(obj);
470     const bool locking = (flags & SYNCH_F_UNLOCK) == 0;
471     const bool trylock = (flags & SYNCH_F_TRY) != 0;
472     const bool read_lock = (flags & SYNCH_F_R) != 0;
473     EvalConditionAnnotated(&cond, mu, locking, trylock, read_lock);
474   }
475   UnrefSynchEvent(e);
476 }
477 
478 //------------------------------------------------------------------
479 
480 // The SynchWaitParams struct encapsulates the way in which a thread is waiting:
481 // whether it has a timeout, the condition, exclusive/shared, and whether a
482 // condition variable wait has an associated Mutex (as opposed to another
483 // type of lock).  It also points to the PerThreadSynch struct of its thread.
484 // cv_word tells Enqueue() to enqueue on a CondVar using CondVarEnqueue().
485 //
486 // This structure is held on the stack rather than directly in
487 // PerThreadSynch because a thread can be waiting on multiple Mutexes if,
488 // while waiting on one Mutex, the implementation calls a client callback
489 // (such as a Condition function) that acquires another Mutex. We don't
490 // strictly need to allow this, but programmers become confused if we do not
491 // allow them to use functions such a LOG() within Condition functions.  The
492 // PerThreadSynch struct points at the most recent SynchWaitParams struct when
493 // the thread is on a Mutex's waiter queue.
494 struct SynchWaitParams {
SynchWaitParamsabsl::SynchWaitParams495   SynchWaitParams(Mutex::MuHow how_arg, const Condition* cond_arg,
496                   KernelTimeout timeout_arg, Mutex* cvmu_arg,
497                   PerThreadSynch* thread_arg,
498                   std::atomic<intptr_t>* cv_word_arg)
499       : how(how_arg),
500         cond(cond_arg),
501         timeout(timeout_arg),
502         cvmu(cvmu_arg),
503         thread(thread_arg),
504         cv_word(cv_word_arg),
505         contention_start_cycles(CycleClock::Now()),
506         should_submit_contention_data(false) {}
507 
508   const Mutex::MuHow how;  // How this thread needs to wait.
509   const Condition* cond;   // The condition that this thread is waiting for.
510                            // In Mutex, this field is set to zero if a timeout
511                            // expires.
512   KernelTimeout timeout;  // timeout expiry---absolute time
513                           // In Mutex, this field is set to zero if a timeout
514                           // expires.
515   Mutex* const cvmu;      // used for transfer from cond var to mutex
516   PerThreadSynch* const thread;  // thread that is waiting
517 
518   // If not null, thread should be enqueued on the CondVar whose state
519   // word is cv_word instead of queueing normally on the Mutex.
520   std::atomic<intptr_t>* cv_word;
521 
522   int64_t contention_start_cycles;  // Time (in cycles) when this thread started
523                                     // to contend for the mutex.
524   bool should_submit_contention_data;
525 };
526 
527 struct SynchLocksHeld {
528   int n;          // number of valid entries in locks[]
529   bool overflow;  // true iff we overflowed the array at some point
530   struct {
531     Mutex* mu;      // lock acquired
532     int32_t count;  // times acquired
533     GraphId id;     // deadlock_graph id of acquired lock
534   } locks[40];
535   // If a thread overfills the array during deadlock detection, we
536   // continue, discarding information as needed.  If no overflow has
537   // taken place, we can provide more error checking, such as
538   // detecting when a thread releases a lock it does not hold.
539 };
540 
541 // A sentinel value in lists that is not 0.
542 // A 0 value is used to mean "not on a list".
543 static PerThreadSynch* const kPerThreadSynchNull =
544     reinterpret_cast<PerThreadSynch*>(1);
545 
LocksHeldAlloc()546 static SynchLocksHeld* LocksHeldAlloc() {
547   SynchLocksHeld* ret = reinterpret_cast<SynchLocksHeld*>(
548       base_internal::LowLevelAlloc::Alloc(sizeof(SynchLocksHeld)));
549   ret->n = 0;
550   ret->overflow = false;
551   return ret;
552 }
553 
554 // Return the PerThreadSynch-struct for this thread.
Synch_GetPerThread()555 static PerThreadSynch* Synch_GetPerThread() {
556   ThreadIdentity* identity = GetOrCreateCurrentThreadIdentity();
557   return &identity->per_thread_synch;
558 }
559 
Synch_GetPerThreadAnnotated(Mutex * mu)560 static PerThreadSynch* Synch_GetPerThreadAnnotated(Mutex* mu) {
561   if (mu) {
562     ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
563   }
564   PerThreadSynch* w = Synch_GetPerThread();
565   if (mu) {
566     ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
567   }
568   return w;
569 }
570 
Synch_GetAllLocks()571 static SynchLocksHeld* Synch_GetAllLocks() {
572   PerThreadSynch* s = Synch_GetPerThread();
573   if (s->all_locks == nullptr) {
574     s->all_locks = LocksHeldAlloc();  // Freed by ReclaimThreadIdentity.
575   }
576   return s->all_locks;
577 }
578 
579 // Post on "w"'s associated PerThreadSem.
IncrementSynchSem(Mutex * mu,PerThreadSynch * w)580 void Mutex::IncrementSynchSem(Mutex* mu, PerThreadSynch* w) {
581   static_cast<void>(mu);  // Prevent unused param warning in non-TSAN builds.
582   ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
583   // We miss synchronization around passing PerThreadSynch between threads
584   // since it happens inside of the Mutex code, so we need to ignore all
585   // accesses to the object.
586   ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN();
587   PerThreadSem::Post(w->thread_identity());
588   ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_END();
589   ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
590 }
591 
592 // Wait on "w"'s associated PerThreadSem; returns false if timeout expired.
DecrementSynchSem(Mutex * mu,PerThreadSynch * w,KernelTimeout t)593 bool Mutex::DecrementSynchSem(Mutex* mu, PerThreadSynch* w, KernelTimeout t) {
594   static_cast<void>(mu);  // Prevent unused param warning in non-TSAN builds.
595   ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
596   assert(w == Synch_GetPerThread());
597   static_cast<void>(w);
598   bool res = PerThreadSem::Wait(t);
599   ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
600   return res;
601 }
602 
603 // We're in a fatal signal handler that hopes to use Mutex and to get
604 // lucky by not deadlocking.  We try to improve its chances of success
605 // by effectively disabling some of the consistency checks.  This will
606 // prevent certain ABSL_RAW_CHECK() statements from being triggered when
607 // re-rentry is detected.  The ABSL_RAW_CHECK() statements are those in the
608 // Mutex code checking that the "waitp" field has not been reused.
InternalAttemptToUseMutexInFatalSignalHandler()609 void Mutex::InternalAttemptToUseMutexInFatalSignalHandler() {
610   // Fix the per-thread state only if it exists.
611   ThreadIdentity* identity = CurrentThreadIdentityIfPresent();
612   if (identity != nullptr) {
613     identity->per_thread_synch.suppress_fatal_errors = true;
614   }
615   // Don't do deadlock detection when we are already failing.
616   synch_deadlock_detection.store(OnDeadlockCycle::kIgnore,
617                                  std::memory_order_release);
618 }
619 
620 // --------------------------Mutexes
621 
622 // In the layout below, the msb of the bottom byte is currently unused.  Also,
623 // the following constraints were considered in choosing the layout:
624 //  o Both the debug allocator's "uninitialized" and "freed" patterns (0xab and
625 //    0xcd) are illegal: reader and writer lock both held.
626 //  o kMuWriter and kMuEvent should exceed kMuDesig and kMuWait, to enable the
627 //    bit-twiddling trick in Mutex::Unlock().
628 //  o kMuWriter / kMuReader == kMuWrWait / kMuWait,
629 //    to enable the bit-twiddling trick in CheckForMutexCorruption().
630 static const intptr_t kMuReader = 0x0001L;  // a reader holds the lock
631 // There's a designated waker.
632 // INVARIANT1:  there's a thread that was blocked on the mutex, is
633 // no longer, yet has not yet acquired the mutex.  If there's a
634 // designated waker, all threads can avoid taking the slow path in
635 // unlock because the designated waker will subsequently acquire
636 // the lock and wake someone.  To maintain INVARIANT1 the bit is
637 // set when a thread is unblocked(INV1a), and threads that were
638 // unblocked reset the bit when they either acquire or re-block (INV1b).
639 static const intptr_t kMuDesig = 0x0002L;
640 static const intptr_t kMuWait = 0x0004L;    // threads are waiting
641 static const intptr_t kMuWriter = 0x0008L;  // a writer holds the lock
642 static const intptr_t kMuEvent = 0x0010L;   // record this mutex's events
643 // Runnable writer is waiting for a reader.
644 // If set, new readers will not lock the mutex to avoid writer starvation.
645 // Note: if a reader has higher priority than the writer, it will still lock
646 // the mutex ahead of the waiting writer, but in a very inefficient manner:
647 // the reader will first queue itself and block, but then the last unlocking
648 // reader will wake it.
649 static const intptr_t kMuWrWait = 0x0020L;
650 static const intptr_t kMuSpin = 0x0040L;  // spinlock protects wait list
651 static const intptr_t kMuLow = 0x00ffL;   // mask all mutex bits
652 static const intptr_t kMuHigh = ~kMuLow;  // mask pointer/reader count
653 
654 static_assert((0xab & (kMuWriter | kMuReader)) == (kMuWriter | kMuReader),
655               "The debug allocator's uninitialized pattern (0xab) must be an "
656               "invalid mutex state");
657 static_assert((0xcd & (kMuWriter | kMuReader)) == (kMuWriter | kMuReader),
658               "The debug allocator's freed pattern (0xcd) must be an invalid "
659               "mutex state");
660 
661 // Hack to make constant values available to gdb pretty printer
662 enum {
663   kGdbMuSpin = kMuSpin,
664   kGdbMuEvent = kMuEvent,
665   kGdbMuWait = kMuWait,
666   kGdbMuWriter = kMuWriter,
667   kGdbMuDesig = kMuDesig,
668   kGdbMuWrWait = kMuWrWait,
669   kGdbMuReader = kMuReader,
670   kGdbMuLow = kMuLow,
671 };
672 
673 // kMuWrWait implies kMuWait.
674 // kMuReader and kMuWriter are mutually exclusive.
675 // If kMuReader is zero, there are no readers.
676 // Otherwise, if kMuWait is zero, the high order bits contain a count of the
677 // number of readers.  Otherwise, the reader count is held in
678 // PerThreadSynch::readers of the most recently queued waiter, again in the
679 // bits above kMuLow.
680 static const intptr_t kMuOne = 0x0100;  // a count of one reader
681 
682 // flags passed to Enqueue and LockSlow{,WithTimeout,Loop}
683 static const int kMuHasBlocked = 0x01;  // already blocked (MUST == 1)
684 static const int kMuIsCond = 0x02;      // conditional waiter (CV or Condition)
685 static const int kMuIsFer = 0x04;       // wait morphing from a CondVar
686 
687 static_assert(PerThreadSynch::kAlignment > kMuLow,
688               "PerThreadSynch::kAlignment must be greater than kMuLow");
689 
690 // This struct contains various bitmasks to be used in
691 // acquiring and releasing a mutex in a particular mode.
692 struct MuHowS {
693   // if all the bits in fast_need_zero are zero, the lock can be acquired by
694   // adding fast_add and oring fast_or.  The bit kMuDesig should be reset iff
695   // this is the designated waker.
696   intptr_t fast_need_zero;
697   intptr_t fast_or;
698   intptr_t fast_add;
699 
700   intptr_t slow_need_zero;  // fast_need_zero with events (e.g. logging)
701 
702   intptr_t slow_inc_need_zero;  // if all the bits in slow_inc_need_zero are
703                                 // zero a reader can acquire a read share by
704                                 // setting the reader bit and incrementing
705                                 // the reader count (in last waiter since
706                                 // we're now slow-path).  kMuWrWait be may
707                                 // be ignored if we already waited once.
708 };
709 
710 static const MuHowS kSharedS = {
711     // shared or read lock
712     kMuWriter | kMuWait | kMuEvent,   // fast_need_zero
713     kMuReader,                        // fast_or
714     kMuOne,                           // fast_add
715     kMuWriter | kMuWait,              // slow_need_zero
716     kMuSpin | kMuWriter | kMuWrWait,  // slow_inc_need_zero
717 };
718 static const MuHowS kExclusiveS = {
719     // exclusive or write lock
720     kMuWriter | kMuReader | kMuEvent,  // fast_need_zero
721     kMuWriter,                         // fast_or
722     0,                                 // fast_add
723     kMuWriter | kMuReader,             // slow_need_zero
724     ~static_cast<intptr_t>(0),         // slow_inc_need_zero
725 };
726 static const Mutex::MuHow kShared = &kSharedS;        // shared lock
727 static const Mutex::MuHow kExclusive = &kExclusiveS;  // exclusive lock
728 
729 #ifdef NDEBUG
730 static constexpr bool kDebugMode = false;
731 #else
732 static constexpr bool kDebugMode = true;
733 #endif
734 
735 #ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
TsanFlags(Mutex::MuHow how)736 static unsigned TsanFlags(Mutex::MuHow how) {
737   return how == kShared ? __tsan_mutex_read_lock : 0;
738 }
739 #endif
740 
741 #if defined(__APPLE__) || defined(ABSL_BUILD_DLL)
742 // When building a dll symbol export lists may reference the destructor
743 // and want it to be an exported symbol rather than an inline function.
744 // Some apple builds also do dynamic library build but don't say it explicitly.
~Mutex()745 Mutex::~Mutex() { Dtor(); }
746 #endif
747 
748 #if !defined(NDEBUG) || defined(ABSL_HAVE_THREAD_SANITIZER)
Dtor()749 void Mutex::Dtor() {
750   if (kDebugMode) {
751     this->ForgetDeadlockInfo();
752   }
753   ABSL_TSAN_MUTEX_DESTROY(this, __tsan_mutex_not_static);
754 }
755 #endif
756 
EnableDebugLog(const char * name)757 void Mutex::EnableDebugLog(const char* name) {
758   // Need to disable writes here and in EnableInvariantDebugging to prevent
759   // false race reports on SynchEvent objects. TSan ignores synchronization
760   // on synch_event_mu in Lock/Unlock/etc methods due to mutex annotations,
761   // but it sees few accesses to SynchEvent in EvalConditionAnnotated.
762   // If we don't ignore accesses here, it can result in false races
763   // between EvalConditionAnnotated and SynchEvent reuse in EnsureSynchEvent.
764   ABSL_ANNOTATE_IGNORE_WRITES_BEGIN();
765   SynchEvent* e = EnsureSynchEvent(&this->mu_, name, kMuEvent, kMuSpin);
766   e->log = true;
767   UnrefSynchEvent(e);
768   // This prevents "error: undefined symbol: absl::Mutex::~Mutex()"
769   // in a release build (NDEBUG defined) when a test does "#undef NDEBUG"
770   // to use assert macro. In such case, the test does not get the dtor
771   // definition because it's supposed to be outline when NDEBUG is not defined,
772   // and this source file does not define one either because NDEBUG is defined.
773   // Since it's not possible to take address of a destructor, we move the
774   // actual destructor code into the separate Dtor function and force the
775   // compiler to emit this function even if it's inline by taking its address.
776   ABSL_ATTRIBUTE_UNUSED volatile auto dtor = &Mutex::Dtor;
777   ABSL_ANNOTATE_IGNORE_WRITES_END();
778 }
779 
EnableMutexInvariantDebugging(bool enabled)780 void EnableMutexInvariantDebugging(bool enabled) {
781   synch_check_invariants.store(enabled, std::memory_order_release);
782 }
783 
EnableInvariantDebugging(void (* invariant)(void *),void * arg)784 void Mutex::EnableInvariantDebugging(void (*invariant)(void*), void* arg) {
785   ABSL_ANNOTATE_IGNORE_WRITES_BEGIN();
786   if (synch_check_invariants.load(std::memory_order_acquire) &&
787       invariant != nullptr) {
788     SynchEvent* e = EnsureSynchEvent(&this->mu_, nullptr, kMuEvent, kMuSpin);
789     e->invariant = invariant;
790     e->arg = arg;
791     UnrefSynchEvent(e);
792   }
793   ABSL_ANNOTATE_IGNORE_WRITES_END();
794 }
795 
SetMutexDeadlockDetectionMode(OnDeadlockCycle mode)796 void SetMutexDeadlockDetectionMode(OnDeadlockCycle mode) {
797   synch_deadlock_detection.store(mode, std::memory_order_release);
798 }
799 
800 // Return true iff threads x and y are part of the same equivalence
801 // class of waiters. An equivalence class is defined as the set of
802 // waiters with the same condition, type of lock, and thread priority.
803 //
804 // Requires that x and y be waiting on the same Mutex queue.
MuEquivalentWaiter(PerThreadSynch * x,PerThreadSynch * y)805 static bool MuEquivalentWaiter(PerThreadSynch* x, PerThreadSynch* y) {
806   return x->waitp->how == y->waitp->how && x->priority == y->priority &&
807          Condition::GuaranteedEqual(x->waitp->cond, y->waitp->cond);
808 }
809 
810 // Given the contents of a mutex word containing a PerThreadSynch pointer,
811 // return the pointer.
GetPerThreadSynch(intptr_t v)812 static inline PerThreadSynch* GetPerThreadSynch(intptr_t v) {
813   return reinterpret_cast<PerThreadSynch*>(v & kMuHigh);
814 }
815 
816 // The next several routines maintain the per-thread next and skip fields
817 // used in the Mutex waiter queue.
818 // The queue is a circular singly-linked list, of which the "head" is the
819 // last element, and head->next if the first element.
820 // The skip field has the invariant:
821 //   For thread x, x->skip is one of:
822 //     - invalid (iff x is not in a Mutex wait queue),
823 //     - null, or
824 //     - a pointer to a distinct thread waiting later in the same Mutex queue
825 //       such that all threads in [x, x->skip] have the same condition, priority
826 //       and lock type (MuEquivalentWaiter() is true for all pairs in [x,
827 //       x->skip]).
828 // In addition, if x->skip is  valid, (x->may_skip || x->skip == null)
829 //
830 // By the spec of MuEquivalentWaiter(), it is not necessary when removing the
831 // first runnable thread y from the front a Mutex queue to adjust the skip
832 // field of another thread x because if x->skip==y, x->skip must (have) become
833 // invalid before y is removed.  The function TryRemove can remove a specified
834 // thread from an arbitrary position in the queue whether runnable or not, so
835 // it fixes up skip fields that would otherwise be left dangling.
836 // The statement
837 //     if (x->may_skip && MuEquivalentWaiter(x, x->next)) { x->skip = x->next; }
838 // maintains the invariant provided x is not the last waiter in a Mutex queue
839 // The statement
840 //          if (x->skip != null) { x->skip = x->skip->skip; }
841 // maintains the invariant.
842 
843 // Returns the last thread y in a mutex waiter queue such that all threads in
844 // [x, y] inclusive share the same condition.  Sets skip fields of some threads
845 // in that range to optimize future evaluation of Skip() on x values in
846 // the range.  Requires thread x is in a mutex waiter queue.
847 // The locking is unusual.  Skip() is called under these conditions:
848 //   - spinlock is held in call from Enqueue(), with maybe_unlocking == false
849 //   - Mutex is held in call from UnlockSlow() by last unlocker, with
850 //     maybe_unlocking == true
851 //   - both Mutex and spinlock are held in call from DequeueAllWakeable() (from
852 //     UnlockSlow()) and TryRemove()
853 // These cases are mutually exclusive, so Skip() never runs concurrently
854 // with itself on the same Mutex.   The skip chain is used in these other places
855 // that cannot occur concurrently:
856 //   - FixSkip() (from TryRemove()) - spinlock and Mutex are held)
857 //   - Dequeue() (with spinlock and Mutex held)
858 //   - UnlockSlow() (with spinlock and Mutex held)
859 // A more complex case is Enqueue()
860 //   - Enqueue() (with spinlock held and maybe_unlocking == false)
861 //               This is the first case in which Skip is called, above.
862 //   - Enqueue() (without spinlock held; but queue is empty and being freshly
863 //                formed)
864 //   - Enqueue() (with spinlock held and maybe_unlocking == true)
865 // The first case has mutual exclusion, and the second isolation through
866 // working on an otherwise unreachable data structure.
867 // In the last case, Enqueue() is required to change no skip/next pointers
868 // except those in the added node and the former "head" node.  This implies
869 // that the new node is added after head, and so must be the new head or the
870 // new front of the queue.
Skip(PerThreadSynch * x)871 static PerThreadSynch* Skip(PerThreadSynch* x) {
872   PerThreadSynch* x0 = nullptr;
873   PerThreadSynch* x1 = x;
874   PerThreadSynch* x2 = x->skip;
875   if (x2 != nullptr) {
876     // Each iteration attempts to advance sequence (x0,x1,x2) to next sequence
877     // such that   x1 == x0->skip && x2 == x1->skip
878     while ((x0 = x1, x1 = x2, x2 = x2->skip) != nullptr) {
879       x0->skip = x2;  // short-circuit skip from x0 to x2
880     }
881     x->skip = x1;  // short-circuit skip from x to result
882   }
883   return x1;
884 }
885 
886 // "ancestor" appears before "to_be_removed" in the same Mutex waiter queue.
887 // The latter is going to be removed out of order, because of a timeout.
888 // Check whether "ancestor" has a skip field pointing to "to_be_removed",
889 // and fix it if it does.
FixSkip(PerThreadSynch * ancestor,PerThreadSynch * to_be_removed)890 static void FixSkip(PerThreadSynch* ancestor, PerThreadSynch* to_be_removed) {
891   if (ancestor->skip == to_be_removed) {  // ancestor->skip left dangling
892     if (to_be_removed->skip != nullptr) {
893       ancestor->skip = to_be_removed->skip;  // can skip past to_be_removed
894     } else if (ancestor->next != to_be_removed) {  // they are not adjacent
895       ancestor->skip = ancestor->next;             // can skip one past ancestor
896     } else {
897       ancestor->skip = nullptr;  // can't skip at all
898     }
899   }
900 }
901 
902 static void CondVarEnqueue(SynchWaitParams* waitp);
903 
904 // Enqueue thread "waitp->thread" on a waiter queue.
905 // Called with mutex spinlock held if head != nullptr
906 // If head==nullptr and waitp->cv_word==nullptr, then Enqueue() is
907 // idempotent; it alters no state associated with the existing (empty)
908 // queue.
909 //
910 // If waitp->cv_word == nullptr, queue the thread at either the front or
911 // the end (according to its priority) of the circular mutex waiter queue whose
912 // head is "head", and return the new head.  mu is the previous mutex state,
913 // which contains the reader count (perhaps adjusted for the operation in
914 // progress) if the list was empty and a read lock held, and the holder hint if
915 // the list was empty and a write lock held.  (flags & kMuIsCond) indicates
916 // whether this thread was transferred from a CondVar or is waiting for a
917 // non-trivial condition.  In this case, Enqueue() never returns nullptr
918 //
919 // If waitp->cv_word != nullptr, CondVarEnqueue() is called, and "head" is
920 // returned. This mechanism is used by CondVar to queue a thread on the
921 // condition variable queue instead of the mutex queue in implementing Wait().
922 // In this case, Enqueue() can return nullptr (if head==nullptr).
Enqueue(PerThreadSynch * head,SynchWaitParams * waitp,intptr_t mu,int flags)923 static PerThreadSynch* Enqueue(PerThreadSynch* head, SynchWaitParams* waitp,
924                                intptr_t mu, int flags) {
925   // If we have been given a cv_word, call CondVarEnqueue() and return
926   // the previous head of the Mutex waiter queue.
927   if (waitp->cv_word != nullptr) {
928     CondVarEnqueue(waitp);
929     return head;
930   }
931 
932   PerThreadSynch* s = waitp->thread;
933   ABSL_RAW_CHECK(
934       s->waitp == nullptr ||    // normal case
935           s->waitp == waitp ||  // Fer()---transfer from condition variable
936           s->suppress_fatal_errors,
937       "detected illegal recursion into Mutex code");
938   s->waitp = waitp;
939   s->skip = nullptr;   // maintain skip invariant (see above)
940   s->may_skip = true;  // always true on entering queue
941   s->wake = false;     // not being woken
942   s->cond_waiter = ((flags & kMuIsCond) != 0);
943 #ifdef ABSL_HAVE_PTHREAD_GETSCHEDPARAM
944   if ((flags & kMuIsFer) == 0) {
945     assert(s == Synch_GetPerThread());
946     int64_t now_cycles = CycleClock::Now();
947     if (s->next_priority_read_cycles < now_cycles) {
948       // Every so often, update our idea of the thread's priority.
949       // pthread_getschedparam() is 5% of the block/wakeup time;
950       // CycleClock::Now() is 0.5%.
951       int policy;
952       struct sched_param param;
953       const int err = pthread_getschedparam(pthread_self(), &policy, &param);
954       if (err != 0) {
955         ABSL_RAW_LOG(ERROR, "pthread_getschedparam failed: %d", err);
956       } else {
957         s->priority = param.sched_priority;
958         s->next_priority_read_cycles =
959             now_cycles + static_cast<int64_t>(CycleClock::Frequency());
960       }
961     }
962   }
963 #endif
964   if (head == nullptr) {         // s is the only waiter
965     s->next = s;                 // it's the only entry in the cycle
966     s->readers = mu;             // reader count is from mu word
967     s->maybe_unlocking = false;  // no one is searching an empty list
968     head = s;                    // s is new head
969   } else {
970     PerThreadSynch* enqueue_after = nullptr;  // we'll put s after this element
971 #ifdef ABSL_HAVE_PTHREAD_GETSCHEDPARAM
972     if (s->priority > head->priority) {  // s's priority is above head's
973       // try to put s in priority-fifo order, or failing that at the front.
974       if (!head->maybe_unlocking) {
975         // No unlocker can be scanning the queue, so we can insert into the
976         // middle of the queue.
977         //
978         // Within a skip chain, all waiters have the same priority, so we can
979         // skip forward through the chains until we find one with a lower
980         // priority than the waiter to be enqueued.
981         PerThreadSynch* advance_to = head;  // next value of enqueue_after
982         do {
983           enqueue_after = advance_to;
984           // (side-effect: optimizes skip chain)
985           advance_to = Skip(enqueue_after->next);
986         } while (s->priority <= advance_to->priority);
987         // termination guaranteed because s->priority > head->priority
988         // and head is the end of a skip chain
989       } else if (waitp->how == kExclusive && waitp->cond == nullptr) {
990         // An unlocker could be scanning the queue, but we know it will recheck
991         // the queue front for writers that have no condition, which is what s
992         // is, so an insert at front is safe.
993         enqueue_after = head;  // add after head, at front
994       }
995     }
996 #endif
997     if (enqueue_after != nullptr) {
998       s->next = enqueue_after->next;
999       enqueue_after->next = s;
1000 
1001       // enqueue_after can be: head, Skip(...), or cur.
1002       // The first two imply enqueue_after->skip == nullptr, and
1003       // the last is used only if MuEquivalentWaiter(s, cur).
1004       // We require this because clearing enqueue_after->skip
1005       // is impossible; enqueue_after's predecessors might also
1006       // incorrectly skip over s if we were to allow other
1007       // insertion points.
1008       ABSL_RAW_CHECK(enqueue_after->skip == nullptr ||
1009                          MuEquivalentWaiter(enqueue_after, s),
1010                      "Mutex Enqueue failure");
1011 
1012       if (enqueue_after != head && enqueue_after->may_skip &&
1013           MuEquivalentWaiter(enqueue_after, enqueue_after->next)) {
1014         // enqueue_after can skip to its new successor, s
1015         enqueue_after->skip = enqueue_after->next;
1016       }
1017       if (MuEquivalentWaiter(s, s->next)) {  // s->may_skip is known to be true
1018         s->skip = s->next;                   // s may skip to its successor
1019       }
1020     } else if ((flags & kMuHasBlocked) &&
1021                (s->priority >= head->next->priority) &&
1022                (!head->maybe_unlocking ||
1023                 (waitp->how == kExclusive &&
1024                  Condition::GuaranteedEqual(waitp->cond, nullptr)))) {
1025       // This thread has already waited, then was woken, then failed to acquire
1026       // the mutex and now tries to requeue. Try to requeue it at head,
1027       // otherwise it can suffer bad latency (wait whole queue several times).
1028       // However, we need to be conservative. First, we need to ensure that we
1029       // respect priorities. Then, we need to be careful to not break wait
1030       // queue invariants: we require either that unlocker is not scanning
1031       // the queue or that the current thread is a writer with no condition
1032       // (unlocker will recheck the queue for such waiters).
1033       s->next = head->next;
1034       head->next = s;
1035       if (MuEquivalentWaiter(s, s->next)) {  // s->may_skip is known to be true
1036         s->skip = s->next;                   // s may skip to its successor
1037       }
1038     } else {  // enqueue not done any other way, so
1039               // we're inserting s at the back
1040       // s will become new head; copy data from head into it
1041       s->next = head->next;  // add s after head
1042       head->next = s;
1043       s->readers = head->readers;  // reader count is from previous head
1044       s->maybe_unlocking = head->maybe_unlocking;  // same for unlock hint
1045       if (head->may_skip && MuEquivalentWaiter(head, s)) {
1046         // head now has successor; may skip
1047         head->skip = s;
1048       }
1049       head = s;  // s is new head
1050     }
1051   }
1052   s->state.store(PerThreadSynch::kQueued, std::memory_order_relaxed);
1053   return head;
1054 }
1055 
1056 // Dequeue the successor pw->next of thread pw from the Mutex waiter queue
1057 // whose last element is head.  The new head element is returned, or null
1058 // if the list is made empty.
1059 // Dequeue is called with both spinlock and Mutex held.
Dequeue(PerThreadSynch * head,PerThreadSynch * pw)1060 static PerThreadSynch* Dequeue(PerThreadSynch* head, PerThreadSynch* pw) {
1061   PerThreadSynch* w = pw->next;
1062   pw->next = w->next;                 // snip w out of list
1063   if (head == w) {                    // we removed the head
1064     head = (pw == w) ? nullptr : pw;  // either emptied list, or pw is new head
1065   } else if (pw != head && MuEquivalentWaiter(pw, pw->next)) {
1066     // pw can skip to its new successor
1067     if (pw->next->skip !=
1068         nullptr) {  // either skip to its successors skip target
1069       pw->skip = pw->next->skip;
1070     } else {  // or to pw's successor
1071       pw->skip = pw->next;
1072     }
1073   }
1074   return head;
1075 }
1076 
1077 // Traverse the elements [ pw->next, h] of the circular list whose last element
1078 // is head.
1079 // Remove all elements with wake==true and place them in the
1080 // singly-linked list wake_list in the order found.   Assumes that
1081 // there is only one such element if the element has how == kExclusive.
1082 // Return the new head.
DequeueAllWakeable(PerThreadSynch * head,PerThreadSynch * pw,PerThreadSynch ** wake_tail)1083 static PerThreadSynch* DequeueAllWakeable(PerThreadSynch* head,
1084                                           PerThreadSynch* pw,
1085                                           PerThreadSynch** wake_tail) {
1086   PerThreadSynch* orig_h = head;
1087   PerThreadSynch* w = pw->next;
1088   bool skipped = false;
1089   do {
1090     if (w->wake) {  // remove this element
1091       ABSL_RAW_CHECK(pw->skip == nullptr, "bad skip in DequeueAllWakeable");
1092       // we're removing pw's successor so either pw->skip is zero or we should
1093       // already have removed pw since if pw->skip!=null, pw has the same
1094       // condition as w.
1095       head = Dequeue(head, pw);
1096       w->next = *wake_tail;               // keep list terminated
1097       *wake_tail = w;                     // add w to wake_list;
1098       wake_tail = &w->next;               // next addition to end
1099       if (w->waitp->how == kExclusive) {  // wake at most 1 writer
1100         break;
1101       }
1102     } else {         // not waking this one; skip
1103       pw = Skip(w);  // skip as much as possible
1104       skipped = true;
1105     }
1106     w = pw->next;
1107     // We want to stop processing after we've considered the original head,
1108     // orig_h.  We can't test for w==orig_h in the loop because w may skip over
1109     // it; we are guaranteed only that w's predecessor will not skip over
1110     // orig_h.  When we've considered orig_h, either we've processed it and
1111     // removed it (so orig_h != head), or we considered it and skipped it (so
1112     // skipped==true && pw == head because skipping from head always skips by
1113     // just one, leaving pw pointing at head).  So we want to
1114     // continue the loop with the negation of that expression.
1115   } while (orig_h == head && (pw != head || !skipped));
1116   return head;
1117 }
1118 
1119 // Try to remove thread s from the list of waiters on this mutex.
1120 // Does nothing if s is not on the waiter list.
TryRemove(PerThreadSynch * s)1121 void Mutex::TryRemove(PerThreadSynch* s) {
1122   SchedulingGuard::ScopedDisable disable_rescheduling;
1123   intptr_t v = mu_.load(std::memory_order_relaxed);
1124   // acquire spinlock & lock
1125   if ((v & (kMuWait | kMuSpin | kMuWriter | kMuReader)) == kMuWait &&
1126       mu_.compare_exchange_strong(v, v | kMuSpin | kMuWriter,
1127                                   std::memory_order_acquire,
1128                                   std::memory_order_relaxed)) {
1129     PerThreadSynch* h = GetPerThreadSynch(v);
1130     if (h != nullptr) {
1131       PerThreadSynch* pw = h;  // pw is w's predecessor
1132       PerThreadSynch* w;
1133       if ((w = pw->next) != s) {  // search for thread,
1134         do {                      // processing at least one element
1135           // If the current element isn't equivalent to the waiter to be
1136           // removed, we can skip the entire chain.
1137           if (!MuEquivalentWaiter(s, w)) {
1138             pw = Skip(w);  // so skip all that won't match
1139             // we don't have to worry about dangling skip fields
1140             // in the threads we skipped; none can point to s
1141             // because they are in a different equivalence class.
1142           } else {          // seeking same condition
1143             FixSkip(w, s);  // fix up any skip pointer from w to s
1144             pw = w;
1145           }
1146           // don't search further if we found the thread, or we're about to
1147           // process the first thread again.
1148         } while ((w = pw->next) != s && pw != h);
1149       }
1150       if (w == s) {  // found thread; remove it
1151         // pw->skip may be non-zero here; the loop above ensured that
1152         // no ancestor of s can skip to s, so removal is safe anyway.
1153         h = Dequeue(h, pw);
1154         s->next = nullptr;
1155         s->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
1156       }
1157     }
1158     intptr_t nv;
1159     do {  // release spinlock and lock
1160       v = mu_.load(std::memory_order_relaxed);
1161       nv = v & (kMuDesig | kMuEvent);
1162       if (h != nullptr) {
1163         nv |= kMuWait | reinterpret_cast<intptr_t>(h);
1164         h->readers = 0;              // we hold writer lock
1165         h->maybe_unlocking = false;  // finished unlocking
1166       }
1167     } while (!mu_.compare_exchange_weak(v, nv, std::memory_order_release,
1168                                         std::memory_order_relaxed));
1169   }
1170 }
1171 
1172 // Wait until thread "s", which must be the current thread, is removed from the
1173 // this mutex's waiter queue.  If "s->waitp->timeout" has a timeout, wake up
1174 // if the wait extends past the absolute time specified, even if "s" is still
1175 // on the mutex queue.  In this case, remove "s" from the queue and return
1176 // true, otherwise return false.
Block(PerThreadSynch * s)1177 void Mutex::Block(PerThreadSynch* s) {
1178   while (s->state.load(std::memory_order_acquire) == PerThreadSynch::kQueued) {
1179     if (!DecrementSynchSem(this, s, s->waitp->timeout)) {
1180       // After a timeout, we go into a spin loop until we remove ourselves
1181       // from the queue, or someone else removes us.  We can't be sure to be
1182       // able to remove ourselves in a single lock acquisition because this
1183       // mutex may be held, and the holder has the right to read the centre
1184       // of the waiter queue without holding the spinlock.
1185       this->TryRemove(s);
1186       int c = 0;
1187       while (s->next != nullptr) {
1188         c = synchronization_internal::MutexDelay(c, GENTLE);
1189         this->TryRemove(s);
1190       }
1191       if (kDebugMode) {
1192         // This ensures that we test the case that TryRemove() is called when s
1193         // is not on the queue.
1194         this->TryRemove(s);
1195       }
1196       s->waitp->timeout = KernelTimeout::Never();  // timeout is satisfied
1197       s->waitp->cond = nullptr;  // condition no longer relevant for wakeups
1198     }
1199   }
1200   ABSL_RAW_CHECK(s->waitp != nullptr || s->suppress_fatal_errors,
1201                  "detected illegal recursion in Mutex code");
1202   s->waitp = nullptr;
1203 }
1204 
1205 // Wake thread w, and return the next thread in the list.
Wakeup(PerThreadSynch * w)1206 PerThreadSynch* Mutex::Wakeup(PerThreadSynch* w) {
1207   PerThreadSynch* next = w->next;
1208   w->next = nullptr;
1209   w->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
1210   IncrementSynchSem(this, w);
1211 
1212   return next;
1213 }
1214 
GetGraphIdLocked(Mutex * mu)1215 static GraphId GetGraphIdLocked(Mutex* mu)
1216     ABSL_EXCLUSIVE_LOCKS_REQUIRED(deadlock_graph_mu) {
1217   if (!deadlock_graph) {  // (re)create the deadlock graph.
1218     deadlock_graph =
1219         new (base_internal::LowLevelAlloc::Alloc(sizeof(*deadlock_graph)))
1220             GraphCycles;
1221   }
1222   return deadlock_graph->GetId(mu);
1223 }
1224 
GetGraphId(Mutex * mu)1225 static GraphId GetGraphId(Mutex* mu) ABSL_LOCKS_EXCLUDED(deadlock_graph_mu) {
1226   deadlock_graph_mu.Lock();
1227   GraphId id = GetGraphIdLocked(mu);
1228   deadlock_graph_mu.Unlock();
1229   return id;
1230 }
1231 
1232 // Record a lock acquisition.  This is used in debug mode for deadlock
1233 // detection.  The held_locks pointer points to the relevant data
1234 // structure for each case.
LockEnter(Mutex * mu,GraphId id,SynchLocksHeld * held_locks)1235 static void LockEnter(Mutex* mu, GraphId id, SynchLocksHeld* held_locks) {
1236   int n = held_locks->n;
1237   int i = 0;
1238   while (i != n && held_locks->locks[i].id != id) {
1239     i++;
1240   }
1241   if (i == n) {
1242     if (n == ABSL_ARRAYSIZE(held_locks->locks)) {
1243       held_locks->overflow = true;  // lost some data
1244     } else {                        // we have room for lock
1245       held_locks->locks[i].mu = mu;
1246       held_locks->locks[i].count = 1;
1247       held_locks->locks[i].id = id;
1248       held_locks->n = n + 1;
1249     }
1250   } else {
1251     held_locks->locks[i].count++;
1252   }
1253 }
1254 
1255 // Record a lock release.  Each call to LockEnter(mu, id, x) should be
1256 // eventually followed by a call to LockLeave(mu, id, x) by the same thread.
1257 // It does not process the event if is not needed when deadlock detection is
1258 // disabled.
LockLeave(Mutex * mu,GraphId id,SynchLocksHeld * held_locks)1259 static void LockLeave(Mutex* mu, GraphId id, SynchLocksHeld* held_locks) {
1260   int n = held_locks->n;
1261   int i = 0;
1262   while (i != n && held_locks->locks[i].id != id) {
1263     i++;
1264   }
1265   if (i == n) {
1266     if (!held_locks->overflow) {
1267       // The deadlock id may have been reassigned after ForgetDeadlockInfo,
1268       // but in that case mu should still be present.
1269       i = 0;
1270       while (i != n && held_locks->locks[i].mu != mu) {
1271         i++;
1272       }
1273       if (i == n) {  // mu missing means releasing unheld lock
1274         SynchEvent* mu_events = GetSynchEvent(mu);
1275         ABSL_RAW_LOG(FATAL,
1276                      "thread releasing lock it does not hold: %p %s; "
1277                      ,
1278                      static_cast<void*>(mu),
1279                      mu_events == nullptr ? "" : mu_events->name);
1280       }
1281     }
1282   } else if (held_locks->locks[i].count == 1) {
1283     held_locks->n = n - 1;
1284     held_locks->locks[i] = held_locks->locks[n - 1];
1285     held_locks->locks[n - 1].id = InvalidGraphId();
1286     held_locks->locks[n - 1].mu =
1287         nullptr;  // clear mu to please the leak detector.
1288   } else {
1289     assert(held_locks->locks[i].count > 0);
1290     held_locks->locks[i].count--;
1291   }
1292 }
1293 
1294 // Call LockEnter() if in debug mode and deadlock detection is enabled.
DebugOnlyLockEnter(Mutex * mu)1295 static inline void DebugOnlyLockEnter(Mutex* mu) {
1296   if (kDebugMode) {
1297     if (synch_deadlock_detection.load(std::memory_order_acquire) !=
1298         OnDeadlockCycle::kIgnore) {
1299       LockEnter(mu, GetGraphId(mu), Synch_GetAllLocks());
1300     }
1301   }
1302 }
1303 
1304 // Call LockEnter() if in debug mode and deadlock detection is enabled.
DebugOnlyLockEnter(Mutex * mu,GraphId id)1305 static inline void DebugOnlyLockEnter(Mutex* mu, GraphId id) {
1306   if (kDebugMode) {
1307     if (synch_deadlock_detection.load(std::memory_order_acquire) !=
1308         OnDeadlockCycle::kIgnore) {
1309       LockEnter(mu, id, Synch_GetAllLocks());
1310     }
1311   }
1312 }
1313 
1314 // Call LockLeave() if in debug mode and deadlock detection is enabled.
DebugOnlyLockLeave(Mutex * mu)1315 static inline void DebugOnlyLockLeave(Mutex* mu) {
1316   if (kDebugMode) {
1317     if (synch_deadlock_detection.load(std::memory_order_acquire) !=
1318         OnDeadlockCycle::kIgnore) {
1319       LockLeave(mu, GetGraphId(mu), Synch_GetAllLocks());
1320     }
1321   }
1322 }
1323 
StackString(void ** pcs,int n,char * buf,int maxlen,bool symbolize)1324 static char* StackString(void** pcs, int n, char* buf, int maxlen,
1325                          bool symbolize) {
1326   static constexpr int kSymLen = 200;
1327   char sym[kSymLen];
1328   int len = 0;
1329   for (int i = 0; i != n; i++) {
1330     if (len >= maxlen)
1331       return buf;
1332     size_t count = static_cast<size_t>(maxlen - len);
1333     if (symbolize) {
1334       if (!absl::Symbolize(pcs[i], sym, kSymLen)) {
1335         sym[0] = '\0';
1336       }
1337       snprintf(buf + len, count, "%s\t@ %p %s\n", (i == 0 ? "\n" : ""), pcs[i],
1338                sym);
1339     } else {
1340       snprintf(buf + len, count, " %p", pcs[i]);
1341     }
1342     len += strlen(&buf[len]);
1343   }
1344   return buf;
1345 }
1346 
CurrentStackString(char * buf,int maxlen,bool symbolize)1347 static char* CurrentStackString(char* buf, int maxlen, bool symbolize) {
1348   void* pcs[40];
1349   return StackString(pcs, absl::GetStackTrace(pcs, ABSL_ARRAYSIZE(pcs), 2), buf,
1350                      maxlen, symbolize);
1351 }
1352 
1353 namespace {
1354 enum {
1355   kMaxDeadlockPathLen = 10
1356 };  // maximum length of a deadlock cycle;
1357     // a path this long would be remarkable
1358 // Buffers required to report a deadlock.
1359 // We do not allocate them on stack to avoid large stack frame.
1360 struct DeadlockReportBuffers {
1361   char buf[6100];
1362   GraphId path[kMaxDeadlockPathLen];
1363 };
1364 
1365 struct ScopedDeadlockReportBuffers {
ScopedDeadlockReportBuffersabsl::__anon2a3dcf5e0a11::ScopedDeadlockReportBuffers1366   ScopedDeadlockReportBuffers() {
1367     b = reinterpret_cast<DeadlockReportBuffers*>(
1368         base_internal::LowLevelAlloc::Alloc(sizeof(*b)));
1369   }
~ScopedDeadlockReportBuffersabsl::__anon2a3dcf5e0a11::ScopedDeadlockReportBuffers1370   ~ScopedDeadlockReportBuffers() { base_internal::LowLevelAlloc::Free(b); }
1371   DeadlockReportBuffers* b;
1372 };
1373 
1374 // Helper to pass to GraphCycles::UpdateStackTrace.
GetStack(void ** stack,int max_depth)1375 int GetStack(void** stack, int max_depth) {
1376   return absl::GetStackTrace(stack, max_depth, 3);
1377 }
1378 }  // anonymous namespace
1379 
1380 // Called in debug mode when a thread is about to acquire a lock in a way that
1381 // may block.
DeadlockCheck(Mutex * mu)1382 static GraphId DeadlockCheck(Mutex* mu) {
1383   if (synch_deadlock_detection.load(std::memory_order_acquire) ==
1384       OnDeadlockCycle::kIgnore) {
1385     return InvalidGraphId();
1386   }
1387 
1388   SynchLocksHeld* all_locks = Synch_GetAllLocks();
1389 
1390   absl::base_internal::SpinLockHolder lock(&deadlock_graph_mu);
1391   const GraphId mu_id = GetGraphIdLocked(mu);
1392 
1393   if (all_locks->n == 0) {
1394     // There are no other locks held. Return now so that we don't need to
1395     // call GetSynchEvent(). This way we do not record the stack trace
1396     // for this Mutex. It's ok, since if this Mutex is involved in a deadlock,
1397     // it can't always be the first lock acquired by a thread.
1398     return mu_id;
1399   }
1400 
1401   // We prefer to keep stack traces that show a thread holding and acquiring
1402   // as many locks as possible.  This increases the chances that a given edge
1403   // in the acquires-before graph will be represented in the stack traces
1404   // recorded for the locks.
1405   deadlock_graph->UpdateStackTrace(mu_id, all_locks->n + 1, GetStack);
1406 
1407   // For each other mutex already held by this thread:
1408   for (int i = 0; i != all_locks->n; i++) {
1409     const GraphId other_node_id = all_locks->locks[i].id;
1410     const Mutex* other =
1411         static_cast<const Mutex*>(deadlock_graph->Ptr(other_node_id));
1412     if (other == nullptr) {
1413       // Ignore stale lock
1414       continue;
1415     }
1416 
1417     // Add the acquired-before edge to the graph.
1418     if (!deadlock_graph->InsertEdge(other_node_id, mu_id)) {
1419       ScopedDeadlockReportBuffers scoped_buffers;
1420       DeadlockReportBuffers* b = scoped_buffers.b;
1421       static int number_of_reported_deadlocks = 0;
1422       number_of_reported_deadlocks++;
1423       // Symbolize only 2 first deadlock report to avoid huge slowdowns.
1424       bool symbolize = number_of_reported_deadlocks <= 2;
1425       ABSL_RAW_LOG(ERROR, "Potential Mutex deadlock: %s",
1426                    CurrentStackString(b->buf, sizeof (b->buf), symbolize));
1427       size_t len = 0;
1428       for (int j = 0; j != all_locks->n; j++) {
1429         void* pr = deadlock_graph->Ptr(all_locks->locks[j].id);
1430         if (pr != nullptr) {
1431           snprintf(b->buf + len, sizeof(b->buf) - len, " %p", pr);
1432           len += strlen(&b->buf[len]);
1433         }
1434       }
1435       ABSL_RAW_LOG(ERROR,
1436                    "Acquiring absl::Mutex %p while holding %s; a cycle in the "
1437                    "historical lock ordering graph has been observed",
1438                    static_cast<void*>(mu), b->buf);
1439       ABSL_RAW_LOG(ERROR, "Cycle: ");
1440       int path_len = deadlock_graph->FindPath(mu_id, other_node_id,
1441                                               ABSL_ARRAYSIZE(b->path), b->path);
1442       for (int j = 0; j != path_len && j != ABSL_ARRAYSIZE(b->path); j++) {
1443         GraphId id = b->path[j];
1444         Mutex* path_mu = static_cast<Mutex*>(deadlock_graph->Ptr(id));
1445         if (path_mu == nullptr) continue;
1446         void** stack;
1447         int depth = deadlock_graph->GetStackTrace(id, &stack);
1448         snprintf(b->buf, sizeof(b->buf),
1449                  "mutex@%p stack: ", static_cast<void*>(path_mu));
1450         StackString(stack, depth, b->buf + strlen(b->buf),
1451                     static_cast<int>(sizeof(b->buf) - strlen(b->buf)),
1452                     symbolize);
1453         ABSL_RAW_LOG(ERROR, "%s", b->buf);
1454       }
1455       if (path_len > static_cast<int>(ABSL_ARRAYSIZE(b->path))) {
1456         ABSL_RAW_LOG(ERROR, "(long cycle; list truncated)");
1457       }
1458       if (synch_deadlock_detection.load(std::memory_order_acquire) ==
1459           OnDeadlockCycle::kAbort) {
1460         deadlock_graph_mu.Unlock();  // avoid deadlock in fatal sighandler
1461         ABSL_RAW_LOG(FATAL, "dying due to potential deadlock");
1462         return mu_id;
1463       }
1464       break;  // report at most one potential deadlock per acquisition
1465     }
1466   }
1467 
1468   return mu_id;
1469 }
1470 
1471 // Invoke DeadlockCheck() iff we're in debug mode and
1472 // deadlock checking has been enabled.
DebugOnlyDeadlockCheck(Mutex * mu)1473 static inline GraphId DebugOnlyDeadlockCheck(Mutex* mu) {
1474   if (kDebugMode && synch_deadlock_detection.load(std::memory_order_acquire) !=
1475                         OnDeadlockCycle::kIgnore) {
1476     return DeadlockCheck(mu);
1477   } else {
1478     return InvalidGraphId();
1479   }
1480 }
1481 
ForgetDeadlockInfo()1482 void Mutex::ForgetDeadlockInfo() {
1483   if (kDebugMode && synch_deadlock_detection.load(std::memory_order_acquire) !=
1484                         OnDeadlockCycle::kIgnore) {
1485     deadlock_graph_mu.Lock();
1486     if (deadlock_graph != nullptr) {
1487       deadlock_graph->RemoveNode(this);
1488     }
1489     deadlock_graph_mu.Unlock();
1490   }
1491 }
1492 
AssertNotHeld() const1493 void Mutex::AssertNotHeld() const {
1494   // We have the data to allow this check only if in debug mode and deadlock
1495   // detection is enabled.
1496   if (kDebugMode &&
1497       (mu_.load(std::memory_order_relaxed) & (kMuWriter | kMuReader)) != 0 &&
1498       synch_deadlock_detection.load(std::memory_order_acquire) !=
1499           OnDeadlockCycle::kIgnore) {
1500     GraphId id = GetGraphId(const_cast<Mutex*>(this));
1501     SynchLocksHeld* locks = Synch_GetAllLocks();
1502     for (int i = 0; i != locks->n; i++) {
1503       if (locks->locks[i].id == id) {
1504         SynchEvent* mu_events = GetSynchEvent(this);
1505         ABSL_RAW_LOG(FATAL, "thread should not hold mutex %p %s",
1506                      static_cast<const void*>(this),
1507                      (mu_events == nullptr ? "" : mu_events->name));
1508       }
1509     }
1510   }
1511 }
1512 
1513 // Attempt to acquire *mu, and return whether successful.  The implementation
1514 // may spin for a short while if the lock cannot be acquired immediately.
TryAcquireWithSpinning(std::atomic<intptr_t> * mu)1515 static bool TryAcquireWithSpinning(std::atomic<intptr_t>* mu) {
1516   int c = globals.spinloop_iterations.load(std::memory_order_relaxed);
1517   do {  // do/while somewhat faster on AMD
1518     intptr_t v = mu->load(std::memory_order_relaxed);
1519     if ((v & (kMuReader | kMuEvent)) != 0) {
1520       return false;                       // a reader or tracing -> give up
1521     } else if (((v & kMuWriter) == 0) &&  // no holder -> try to acquire
1522                mu->compare_exchange_strong(v, kMuWriter | v,
1523                                            std::memory_order_acquire,
1524                                            std::memory_order_relaxed)) {
1525       return true;
1526     }
1527   } while (--c > 0);
1528   return false;
1529 }
1530 
Lock()1531 void Mutex::Lock() {
1532   ABSL_TSAN_MUTEX_PRE_LOCK(this, 0);
1533   GraphId id = DebugOnlyDeadlockCheck(this);
1534   intptr_t v = mu_.load(std::memory_order_relaxed);
1535   // try fast acquire, then spin loop
1536   if (ABSL_PREDICT_FALSE((v & (kMuWriter | kMuReader | kMuEvent)) != 0) ||
1537       ABSL_PREDICT_FALSE(!mu_.compare_exchange_strong(
1538           v, kMuWriter | v, std::memory_order_acquire,
1539           std::memory_order_relaxed))) {
1540     // try spin acquire, then slow loop
1541     if (ABSL_PREDICT_FALSE(!TryAcquireWithSpinning(&this->mu_))) {
1542       this->LockSlow(kExclusive, nullptr, 0);
1543     }
1544   }
1545   DebugOnlyLockEnter(this, id);
1546   ABSL_TSAN_MUTEX_POST_LOCK(this, 0, 0);
1547 }
1548 
ReaderLock()1549 void Mutex::ReaderLock() {
1550   ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_read_lock);
1551   GraphId id = DebugOnlyDeadlockCheck(this);
1552   intptr_t v = mu_.load(std::memory_order_relaxed);
1553   for (;;) {
1554     // If there are non-readers holding the lock, use the slow loop.
1555     if (ABSL_PREDICT_FALSE(v & (kMuWriter | kMuWait | kMuEvent)) != 0) {
1556       this->LockSlow(kShared, nullptr, 0);
1557       break;
1558     }
1559     // We can avoid the loop and only use the CAS when the lock is free or
1560     // only held by readers.
1561     if (ABSL_PREDICT_TRUE(mu_.compare_exchange_weak(
1562             v, (kMuReader | v) + kMuOne, std::memory_order_acquire,
1563             std::memory_order_relaxed))) {
1564       break;
1565     }
1566   }
1567   DebugOnlyLockEnter(this, id);
1568   ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_read_lock, 0);
1569 }
1570 
LockWhenCommon(const Condition & cond,synchronization_internal::KernelTimeout t,bool write)1571 bool Mutex::LockWhenCommon(const Condition& cond,
1572                            synchronization_internal::KernelTimeout t,
1573                            bool write) {
1574   MuHow how = write ? kExclusive : kShared;
1575   ABSL_TSAN_MUTEX_PRE_LOCK(this, TsanFlags(how));
1576   GraphId id = DebugOnlyDeadlockCheck(this);
1577   bool res = LockSlowWithDeadline(how, &cond, t, 0);
1578   DebugOnlyLockEnter(this, id);
1579   ABSL_TSAN_MUTEX_POST_LOCK(this, TsanFlags(how), 0);
1580   return res;
1581 }
1582 
AwaitCommon(const Condition & cond,KernelTimeout t)1583 bool Mutex::AwaitCommon(const Condition& cond, KernelTimeout t) {
1584   if (kDebugMode) {
1585     this->AssertReaderHeld();
1586   }
1587   if (cond.Eval()) {  // condition already true; nothing to do
1588     return true;
1589   }
1590   MuHow how =
1591       (mu_.load(std::memory_order_relaxed) & kMuWriter) ? kExclusive : kShared;
1592   ABSL_TSAN_MUTEX_PRE_UNLOCK(this, TsanFlags(how));
1593   SynchWaitParams waitp(how, &cond, t, nullptr /*no cvmu*/,
1594                         Synch_GetPerThreadAnnotated(this),
1595                         nullptr /*no cv_word*/);
1596   this->UnlockSlow(&waitp);
1597   this->Block(waitp.thread);
1598   ABSL_TSAN_MUTEX_POST_UNLOCK(this, TsanFlags(how));
1599   ABSL_TSAN_MUTEX_PRE_LOCK(this, TsanFlags(how));
1600   this->LockSlowLoop(&waitp, kMuHasBlocked | kMuIsCond);
1601   bool res = waitp.cond != nullptr ||  // => cond known true from LockSlowLoop
1602              EvalConditionAnnotated(&cond, this, true, false, how == kShared);
1603   ABSL_TSAN_MUTEX_POST_LOCK(this, TsanFlags(how), 0);
1604   ABSL_RAW_CHECK(res || t.has_timeout(),
1605                  "condition untrue on return from Await");
1606   return res;
1607 }
1608 
TryLock()1609 bool Mutex::TryLock() {
1610   ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_try_lock);
1611   intptr_t v = mu_.load(std::memory_order_relaxed);
1612   // Try fast acquire.
1613   if (ABSL_PREDICT_TRUE((v & (kMuWriter | kMuReader | kMuEvent)) == 0)) {
1614     if (ABSL_PREDICT_TRUE(mu_.compare_exchange_strong(
1615             v, kMuWriter | v, std::memory_order_acquire,
1616             std::memory_order_relaxed))) {
1617       DebugOnlyLockEnter(this);
1618       ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_try_lock, 0);
1619       return true;
1620     }
1621   } else if (ABSL_PREDICT_FALSE((v & kMuEvent) != 0)) {
1622     // We're recording events.
1623     return TryLockSlow();
1624   }
1625   ABSL_TSAN_MUTEX_POST_LOCK(
1626       this, __tsan_mutex_try_lock | __tsan_mutex_try_lock_failed, 0);
1627   return false;
1628 }
1629 
TryLockSlow()1630 ABSL_ATTRIBUTE_NOINLINE bool Mutex::TryLockSlow() {
1631   intptr_t v = mu_.load(std::memory_order_relaxed);
1632   if ((v & kExclusive->slow_need_zero) == 0 &&  // try fast acquire
1633       mu_.compare_exchange_strong(
1634           v, (kExclusive->fast_or | v) + kExclusive->fast_add,
1635           std::memory_order_acquire, std::memory_order_relaxed)) {
1636     DebugOnlyLockEnter(this);
1637     PostSynchEvent(this, SYNCH_EV_TRYLOCK_SUCCESS);
1638     ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_try_lock, 0);
1639     return true;
1640   }
1641   PostSynchEvent(this, SYNCH_EV_TRYLOCK_FAILED);
1642   ABSL_TSAN_MUTEX_POST_LOCK(
1643       this, __tsan_mutex_try_lock | __tsan_mutex_try_lock_failed, 0);
1644   return false;
1645 }
1646 
ReaderTryLock()1647 bool Mutex::ReaderTryLock() {
1648   ABSL_TSAN_MUTEX_PRE_LOCK(this,
1649                            __tsan_mutex_read_lock | __tsan_mutex_try_lock);
1650   intptr_t v = mu_.load(std::memory_order_relaxed);
1651   // Clang tends to unroll the loop when compiling with optimization.
1652   // But in this case it just unnecessary increases code size.
1653   // If CAS is failing due to contention, the jump cost is negligible.
1654 #if defined(__clang__)
1655 #pragma nounroll
1656 #endif
1657   // The while-loops (here and below) iterate only if the mutex word keeps
1658   // changing (typically because the reader count changes) under the CAS.
1659   // We limit the number of attempts to avoid having to think about livelock.
1660   for (int loop_limit = 5; loop_limit != 0; loop_limit--) {
1661     if (ABSL_PREDICT_FALSE((v & (kMuWriter | kMuWait | kMuEvent)) != 0)) {
1662       break;
1663     }
1664     if (ABSL_PREDICT_TRUE(mu_.compare_exchange_strong(
1665             v, (kMuReader | v) + kMuOne, std::memory_order_acquire,
1666             std::memory_order_relaxed))) {
1667       DebugOnlyLockEnter(this);
1668       ABSL_TSAN_MUTEX_POST_LOCK(
1669           this, __tsan_mutex_read_lock | __tsan_mutex_try_lock, 0);
1670       return true;
1671     }
1672   }
1673   if (ABSL_PREDICT_TRUE((v & kMuEvent) == 0)) {
1674     ABSL_TSAN_MUTEX_POST_LOCK(this,
1675                               __tsan_mutex_read_lock | __tsan_mutex_try_lock |
1676                                   __tsan_mutex_try_lock_failed,
1677                               0);
1678     return false;
1679   }
1680   // we're recording events
1681   return ReaderTryLockSlow();
1682 }
1683 
ReaderTryLockSlow()1684 ABSL_ATTRIBUTE_NOINLINE bool Mutex::ReaderTryLockSlow() {
1685   intptr_t v = mu_.load(std::memory_order_relaxed);
1686 #if defined(__clang__)
1687 #pragma nounroll
1688 #endif
1689   for (int loop_limit = 5; loop_limit != 0; loop_limit--) {
1690     if ((v & kShared->slow_need_zero) == 0 &&
1691         mu_.compare_exchange_strong(v, (kMuReader | v) + kMuOne,
1692                                     std::memory_order_acquire,
1693                                     std::memory_order_relaxed)) {
1694       DebugOnlyLockEnter(this);
1695       PostSynchEvent(this, SYNCH_EV_READERTRYLOCK_SUCCESS);
1696       ABSL_TSAN_MUTEX_POST_LOCK(
1697           this, __tsan_mutex_read_lock | __tsan_mutex_try_lock, 0);
1698       return true;
1699     }
1700   }
1701   PostSynchEvent(this, SYNCH_EV_READERTRYLOCK_FAILED);
1702   ABSL_TSAN_MUTEX_POST_LOCK(this,
1703                             __tsan_mutex_read_lock | __tsan_mutex_try_lock |
1704                                 __tsan_mutex_try_lock_failed,
1705                             0);
1706   return false;
1707 }
1708 
Unlock()1709 void Mutex::Unlock() {
1710   ABSL_TSAN_MUTEX_PRE_UNLOCK(this, 0);
1711   DebugOnlyLockLeave(this);
1712   intptr_t v = mu_.load(std::memory_order_relaxed);
1713 
1714   if (kDebugMode && ((v & (kMuWriter | kMuReader)) != kMuWriter)) {
1715     ABSL_RAW_LOG(FATAL, "Mutex unlocked when destroyed or not locked: v=0x%x",
1716                  static_cast<unsigned>(v));
1717   }
1718 
1719   // should_try_cas is whether we'll try a compare-and-swap immediately.
1720   // NOTE: optimized out when kDebugMode is false.
1721   bool should_try_cas = ((v & (kMuEvent | kMuWriter)) == kMuWriter &&
1722                          (v & (kMuWait | kMuDesig)) != kMuWait);
1723 
1724   // But, we can use an alternate computation of it, that compilers
1725   // currently don't find on their own.  When that changes, this function
1726   // can be simplified.
1727   //
1728   // should_try_cas is true iff the bits satisfy the following conditions:
1729   //
1730   //                   Ev Wr Wa De
1731   // equal to           0  1
1732   // and not equal to         1  0
1733   //
1734   // after xoring by    0  1  0  1,  this is equivalent to:
1735   //
1736   // equal to           0  0
1737   // and not equal to         1  1,  which is the same as:
1738   //
1739   // smaller than       0  0  1  1
1740   static_assert(kMuEvent > kMuWait, "Needed for should_try_cas_fast");
1741   static_assert(kMuEvent > kMuDesig, "Needed for should_try_cas_fast");
1742   static_assert(kMuWriter > kMuWait, "Needed for should_try_cas_fast");
1743   static_assert(kMuWriter > kMuDesig, "Needed for should_try_cas_fast");
1744 
1745   bool should_try_cas_fast =
1746       ((v ^ (kMuWriter | kMuDesig)) &
1747        (kMuEvent | kMuWriter | kMuWait | kMuDesig)) < (kMuWait | kMuDesig);
1748 
1749   if (kDebugMode && should_try_cas != should_try_cas_fast) {
1750     // We would usually use PRIdPTR here, but is not correctly implemented
1751     // within the android toolchain.
1752     ABSL_RAW_LOG(FATAL, "internal logic error %llx %llx %llx\n",
1753                  static_cast<long long>(v),
1754                  static_cast<long long>(should_try_cas),
1755                  static_cast<long long>(should_try_cas_fast));
1756   }
1757   if (should_try_cas_fast &&
1758       mu_.compare_exchange_strong(v, v & ~(kMuWrWait | kMuWriter),
1759                                   std::memory_order_release,
1760                                   std::memory_order_relaxed)) {
1761     // fast writer release (writer with no waiters or with designated waker)
1762   } else {
1763     this->UnlockSlow(nullptr /*no waitp*/);  // take slow path
1764   }
1765   ABSL_TSAN_MUTEX_POST_UNLOCK(this, 0);
1766 }
1767 
1768 // Requires v to represent a reader-locked state.
ExactlyOneReader(intptr_t v)1769 static bool ExactlyOneReader(intptr_t v) {
1770   assert((v & (kMuWriter | kMuReader)) == kMuReader);
1771   assert((v & kMuHigh) != 0);
1772   // The more straightforward "(v & kMuHigh) == kMuOne" also works, but
1773   // on some architectures the following generates slightly smaller code.
1774   // It may be faster too.
1775   constexpr intptr_t kMuMultipleWaitersMask = kMuHigh ^ kMuOne;
1776   return (v & kMuMultipleWaitersMask) == 0;
1777 }
1778 
ReaderUnlock()1779 void Mutex::ReaderUnlock() {
1780   ABSL_TSAN_MUTEX_PRE_UNLOCK(this, __tsan_mutex_read_lock);
1781   DebugOnlyLockLeave(this);
1782   intptr_t v = mu_.load(std::memory_order_relaxed);
1783   assert((v & (kMuWriter | kMuReader)) == kMuReader);
1784   for (;;) {
1785     if (ABSL_PREDICT_FALSE((v & (kMuReader | kMuWait | kMuEvent)) !=
1786                            kMuReader)) {
1787       this->UnlockSlow(nullptr /*no waitp*/);  // take slow path
1788       break;
1789     }
1790     // fast reader release (reader with no waiters)
1791     intptr_t clear = ExactlyOneReader(v) ? kMuReader | kMuOne : kMuOne;
1792     if (ABSL_PREDICT_TRUE(
1793             mu_.compare_exchange_strong(v, v - clear, std::memory_order_release,
1794                                         std::memory_order_relaxed))) {
1795       break;
1796     }
1797   }
1798   ABSL_TSAN_MUTEX_POST_UNLOCK(this, __tsan_mutex_read_lock);
1799 }
1800 
1801 // Clears the designated waker flag in the mutex if this thread has blocked, and
1802 // therefore may be the designated waker.
ClearDesignatedWakerMask(int flag)1803 static intptr_t ClearDesignatedWakerMask(int flag) {
1804   assert(flag >= 0);
1805   assert(flag <= 1);
1806   switch (flag) {
1807     case 0:  // not blocked
1808       return ~static_cast<intptr_t>(0);
1809     case 1:  // blocked; turn off the designated waker bit
1810       return ~static_cast<intptr_t>(kMuDesig);
1811   }
1812   ABSL_UNREACHABLE();
1813 }
1814 
1815 // Conditionally ignores the existence of waiting writers if a reader that has
1816 // already blocked once wakes up.
IgnoreWaitingWritersMask(int flag)1817 static intptr_t IgnoreWaitingWritersMask(int flag) {
1818   assert(flag >= 0);
1819   assert(flag <= 1);
1820   switch (flag) {
1821     case 0:  // not blocked
1822       return ~static_cast<intptr_t>(0);
1823     case 1:  // blocked; pretend there are no waiting writers
1824       return ~static_cast<intptr_t>(kMuWrWait);
1825   }
1826   ABSL_UNREACHABLE();
1827 }
1828 
1829 // Internal version of LockWhen().  See LockSlowWithDeadline()
LockSlow(MuHow how,const Condition * cond,int flags)1830 ABSL_ATTRIBUTE_NOINLINE void Mutex::LockSlow(MuHow how, const Condition* cond,
1831                                              int flags) {
1832   // Note: we specifically initialize spinloop_iterations after the first use
1833   // in TryAcquireWithSpinning so that Lock function does not have any non-tail
1834   // calls and consequently a stack frame. It's fine to have spinloop_iterations
1835   // uninitialized (meaning no spinning) in all initial uncontended Lock calls
1836   // and in the first contended call. After that we will have
1837   // spinloop_iterations properly initialized.
1838   if (ABSL_PREDICT_FALSE(
1839           globals.spinloop_iterations.load(std::memory_order_relaxed) == 0)) {
1840     if (absl::base_internal::NumCPUs() > 1) {
1841       // If this is multiprocessor, allow spinning.
1842       globals.spinloop_iterations.store(1500, std::memory_order_relaxed);
1843     } else {
1844       // If this a uniprocessor, only yield/sleep.
1845       globals.spinloop_iterations.store(-1, std::memory_order_relaxed);
1846     }
1847   }
1848   ABSL_RAW_CHECK(
1849       this->LockSlowWithDeadline(how, cond, KernelTimeout::Never(), flags),
1850       "condition untrue on return from LockSlow");
1851 }
1852 
1853 // Compute cond->Eval() and tell race detectors that we do it under mutex mu.
EvalConditionAnnotated(const Condition * cond,Mutex * mu,bool locking,bool trylock,bool read_lock)1854 static inline bool EvalConditionAnnotated(const Condition* cond, Mutex* mu,
1855                                           bool locking, bool trylock,
1856                                           bool read_lock) {
1857   // Delicate annotation dance.
1858   // We are currently inside of read/write lock/unlock operation.
1859   // All memory accesses are ignored inside of mutex operations + for unlock
1860   // operation tsan considers that we've already released the mutex.
1861   bool res = false;
1862 #ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
1863   const uint32_t flags = read_lock ? __tsan_mutex_read_lock : 0;
1864   const uint32_t tryflags = flags | (trylock ? __tsan_mutex_try_lock : 0);
1865 #endif
1866   if (locking) {
1867     // For lock we pretend that we have finished the operation,
1868     // evaluate the predicate, then unlock the mutex and start locking it again
1869     // to match the annotation at the end of outer lock operation.
1870     // Note: we can't simply do POST_LOCK, Eval, PRE_LOCK, because then tsan
1871     // will think the lock acquisition is recursive which will trigger
1872     // deadlock detector.
1873     ABSL_TSAN_MUTEX_POST_LOCK(mu, tryflags, 0);
1874     res = cond->Eval();
1875     // There is no "try" version of Unlock, so use flags instead of tryflags.
1876     ABSL_TSAN_MUTEX_PRE_UNLOCK(mu, flags);
1877     ABSL_TSAN_MUTEX_POST_UNLOCK(mu, flags);
1878     ABSL_TSAN_MUTEX_PRE_LOCK(mu, tryflags);
1879   } else {
1880     // Similarly, for unlock we pretend that we have unlocked the mutex,
1881     // lock the mutex, evaluate the predicate, and start unlocking it again
1882     // to match the annotation at the end of outer unlock operation.
1883     ABSL_TSAN_MUTEX_POST_UNLOCK(mu, flags);
1884     ABSL_TSAN_MUTEX_PRE_LOCK(mu, flags);
1885     ABSL_TSAN_MUTEX_POST_LOCK(mu, flags, 0);
1886     res = cond->Eval();
1887     ABSL_TSAN_MUTEX_PRE_UNLOCK(mu, flags);
1888   }
1889   // Prevent unused param warnings in non-TSAN builds.
1890   static_cast<void>(mu);
1891   static_cast<void>(trylock);
1892   static_cast<void>(read_lock);
1893   return res;
1894 }
1895 
1896 // Compute cond->Eval() hiding it from race detectors.
1897 // We are hiding it because inside of UnlockSlow we can evaluate a predicate
1898 // that was just added by a concurrent Lock operation; Lock adds the predicate
1899 // to the internal Mutex list without actually acquiring the Mutex
1900 // (it only acquires the internal spinlock, which is rightfully invisible for
1901 // tsan). As the result there is no tsan-visible synchronization between the
1902 // addition and this thread. So if we would enable race detection here,
1903 // it would race with the predicate initialization.
EvalConditionIgnored(Mutex * mu,const Condition * cond)1904 static inline bool EvalConditionIgnored(Mutex* mu, const Condition* cond) {
1905   // Memory accesses are already ignored inside of lock/unlock operations,
1906   // but synchronization operations are also ignored. When we evaluate the
1907   // predicate we must ignore only memory accesses but not synchronization,
1908   // because missed synchronization can lead to false reports later.
1909   // So we "divert" (which un-ignores both memory accesses and synchronization)
1910   // and then separately turn on ignores of memory accesses.
1911   ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
1912   ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN();
1913   bool res = cond->Eval();
1914   ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_END();
1915   ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
1916   static_cast<void>(mu);  // Prevent unused param warning in non-TSAN builds.
1917   return res;
1918 }
1919 
1920 // Internal equivalent of *LockWhenWithDeadline(), where
1921 //   "t" represents the absolute timeout; !t.has_timeout() means "forever".
1922 //   "how" is "kShared" (for ReaderLockWhen) or "kExclusive" (for LockWhen)
1923 // In flags, bits are ored together:
1924 // - kMuHasBlocked indicates that the client has already blocked on the call so
1925 //   the designated waker bit must be cleared and waiting writers should not
1926 //   obstruct this call
1927 // - kMuIsCond indicates that this is a conditional acquire (condition variable,
1928 //   Await,  LockWhen) so contention profiling should be suppressed.
LockSlowWithDeadline(MuHow how,const Condition * cond,KernelTimeout t,int flags)1929 bool Mutex::LockSlowWithDeadline(MuHow how, const Condition* cond,
1930                                  KernelTimeout t, int flags) {
1931   intptr_t v = mu_.load(std::memory_order_relaxed);
1932   bool unlock = false;
1933   if ((v & how->fast_need_zero) == 0 &&  // try fast acquire
1934       mu_.compare_exchange_strong(
1935           v,
1936           (how->fast_or |
1937            (v & ClearDesignatedWakerMask(flags & kMuHasBlocked))) +
1938               how->fast_add,
1939           std::memory_order_acquire, std::memory_order_relaxed)) {
1940     if (cond == nullptr ||
1941         EvalConditionAnnotated(cond, this, true, false, how == kShared)) {
1942       return true;
1943     }
1944     unlock = true;
1945   }
1946   SynchWaitParams waitp(how, cond, t, nullptr /*no cvmu*/,
1947                         Synch_GetPerThreadAnnotated(this),
1948                         nullptr /*no cv_word*/);
1949   if (cond != nullptr) {
1950     flags |= kMuIsCond;
1951   }
1952   if (unlock) {
1953     this->UnlockSlow(&waitp);
1954     this->Block(waitp.thread);
1955     flags |= kMuHasBlocked;
1956   }
1957   this->LockSlowLoop(&waitp, flags);
1958   return waitp.cond != nullptr ||  // => cond known true from LockSlowLoop
1959          cond == nullptr ||
1960          EvalConditionAnnotated(cond, this, true, false, how == kShared);
1961 }
1962 
1963 // RAW_CHECK_FMT() takes a condition, a printf-style format string, and
1964 // the printf-style argument list.   The format string must be a literal.
1965 // Arguments after the first are not evaluated unless the condition is true.
1966 #define RAW_CHECK_FMT(cond, ...)                                   \
1967   do {                                                             \
1968     if (ABSL_PREDICT_FALSE(!(cond))) {                             \
1969       ABSL_RAW_LOG(FATAL, "Check " #cond " failed: " __VA_ARGS__); \
1970     }                                                              \
1971   } while (0)
1972 
CheckForMutexCorruption(intptr_t v,const char * label)1973 static void CheckForMutexCorruption(intptr_t v, const char* label) {
1974   // Test for either of two situations that should not occur in v:
1975   //   kMuWriter and kMuReader
1976   //   kMuWrWait and !kMuWait
1977   const uintptr_t w = static_cast<uintptr_t>(v ^ kMuWait);
1978   // By flipping that bit, we can now test for:
1979   //   kMuWriter and kMuReader in w
1980   //   kMuWrWait and kMuWait in w
1981   // We've chosen these two pairs of values to be so that they will overlap,
1982   // respectively, when the word is left shifted by three.  This allows us to
1983   // save a branch in the common (correct) case of them not being coincident.
1984   static_assert(kMuReader << 3 == kMuWriter, "must match");
1985   static_assert(kMuWait << 3 == kMuWrWait, "must match");
1986   if (ABSL_PREDICT_TRUE((w & (w << 3) & (kMuWriter | kMuWrWait)) == 0)) return;
1987   RAW_CHECK_FMT((v & (kMuWriter | kMuReader)) != (kMuWriter | kMuReader),
1988                 "%s: Mutex corrupt: both reader and writer lock held: %p",
1989                 label, reinterpret_cast<void*>(v));
1990   RAW_CHECK_FMT((v & (kMuWait | kMuWrWait)) != kMuWrWait,
1991                 "%s: Mutex corrupt: waiting writer with no waiters: %p", label,
1992                 reinterpret_cast<void*>(v));
1993   assert(false);
1994 }
1995 
LockSlowLoop(SynchWaitParams * waitp,int flags)1996 void Mutex::LockSlowLoop(SynchWaitParams* waitp, int flags) {
1997   SchedulingGuard::ScopedDisable disable_rescheduling;
1998   int c = 0;
1999   intptr_t v = mu_.load(std::memory_order_relaxed);
2000   if ((v & kMuEvent) != 0) {
2001     PostSynchEvent(
2002         this, waitp->how == kExclusive ? SYNCH_EV_LOCK : SYNCH_EV_READERLOCK);
2003   }
2004   ABSL_RAW_CHECK(
2005       waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
2006       "detected illegal recursion into Mutex code");
2007   for (;;) {
2008     v = mu_.load(std::memory_order_relaxed);
2009     CheckForMutexCorruption(v, "Lock");
2010     if ((v & waitp->how->slow_need_zero) == 0) {
2011       if (mu_.compare_exchange_strong(
2012               v,
2013               (waitp->how->fast_or |
2014                (v & ClearDesignatedWakerMask(flags & kMuHasBlocked))) +
2015                   waitp->how->fast_add,
2016               std::memory_order_acquire, std::memory_order_relaxed)) {
2017         if (waitp->cond == nullptr ||
2018             EvalConditionAnnotated(waitp->cond, this, true, false,
2019                                    waitp->how == kShared)) {
2020           break;  // we timed out, or condition true, so return
2021         }
2022         this->UnlockSlow(waitp);  // got lock but condition false
2023         this->Block(waitp->thread);
2024         flags |= kMuHasBlocked;
2025         c = 0;
2026       }
2027     } else {  // need to access waiter list
2028       bool dowait = false;
2029       if ((v & (kMuSpin | kMuWait)) == 0) {  // no waiters
2030         // This thread tries to become the one and only waiter.
2031         PerThreadSynch* new_h = Enqueue(nullptr, waitp, v, flags);
2032         intptr_t nv =
2033             (v & ClearDesignatedWakerMask(flags & kMuHasBlocked) & kMuLow) |
2034             kMuWait;
2035         ABSL_RAW_CHECK(new_h != nullptr, "Enqueue to empty list failed");
2036         if (waitp->how == kExclusive && (v & kMuReader) != 0) {
2037           nv |= kMuWrWait;
2038         }
2039         if (mu_.compare_exchange_strong(
2040                 v, reinterpret_cast<intptr_t>(new_h) | nv,
2041                 std::memory_order_release, std::memory_order_relaxed)) {
2042           dowait = true;
2043         } else {  // attempted Enqueue() failed
2044           // zero out the waitp field set by Enqueue()
2045           waitp->thread->waitp = nullptr;
2046         }
2047       } else if ((v & waitp->how->slow_inc_need_zero &
2048                   IgnoreWaitingWritersMask(flags & kMuHasBlocked)) == 0) {
2049         // This is a reader that needs to increment the reader count,
2050         // but the count is currently held in the last waiter.
2051         if (mu_.compare_exchange_strong(
2052                 v,
2053                 (v & ClearDesignatedWakerMask(flags & kMuHasBlocked)) |
2054                     kMuSpin | kMuReader,
2055                 std::memory_order_acquire, std::memory_order_relaxed)) {
2056           PerThreadSynch* h = GetPerThreadSynch(v);
2057           h->readers += kMuOne;  // inc reader count in waiter
2058           do {                   // release spinlock
2059             v = mu_.load(std::memory_order_relaxed);
2060           } while (!mu_.compare_exchange_weak(v, (v & ~kMuSpin) | kMuReader,
2061                                               std::memory_order_release,
2062                                               std::memory_order_relaxed));
2063           if (waitp->cond == nullptr ||
2064               EvalConditionAnnotated(waitp->cond, this, true, false,
2065                                      waitp->how == kShared)) {
2066             break;  // we timed out, or condition true, so return
2067           }
2068           this->UnlockSlow(waitp);  // got lock but condition false
2069           this->Block(waitp->thread);
2070           flags |= kMuHasBlocked;
2071           c = 0;
2072         }
2073       } else if ((v & kMuSpin) == 0 &&  // attempt to queue ourselves
2074                  mu_.compare_exchange_strong(
2075                      v,
2076                      (v & ClearDesignatedWakerMask(flags & kMuHasBlocked)) |
2077                          kMuSpin | kMuWait,
2078                      std::memory_order_acquire, std::memory_order_relaxed)) {
2079         PerThreadSynch* h = GetPerThreadSynch(v);
2080         PerThreadSynch* new_h = Enqueue(h, waitp, v, flags);
2081         intptr_t wr_wait = 0;
2082         ABSL_RAW_CHECK(new_h != nullptr, "Enqueue to list failed");
2083         if (waitp->how == kExclusive && (v & kMuReader) != 0) {
2084           wr_wait = kMuWrWait;  // give priority to a waiting writer
2085         }
2086         do {  // release spinlock
2087           v = mu_.load(std::memory_order_relaxed);
2088         } while (!mu_.compare_exchange_weak(
2089             v,
2090             (v & (kMuLow & ~kMuSpin)) | kMuWait | wr_wait |
2091                 reinterpret_cast<intptr_t>(new_h),
2092             std::memory_order_release, std::memory_order_relaxed));
2093         dowait = true;
2094       }
2095       if (dowait) {
2096         this->Block(waitp->thread);  // wait until removed from list or timeout
2097         flags |= kMuHasBlocked;
2098         c = 0;
2099       }
2100     }
2101     ABSL_RAW_CHECK(
2102         waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
2103         "detected illegal recursion into Mutex code");
2104     // delay, then try again
2105     c = synchronization_internal::MutexDelay(c, GENTLE);
2106   }
2107   ABSL_RAW_CHECK(
2108       waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
2109       "detected illegal recursion into Mutex code");
2110   if ((v & kMuEvent) != 0) {
2111     PostSynchEvent(this, waitp->how == kExclusive
2112                              ? SYNCH_EV_LOCK_RETURNING
2113                              : SYNCH_EV_READERLOCK_RETURNING);
2114   }
2115 }
2116 
2117 // Unlock this mutex, which is held by the current thread.
2118 // If waitp is non-zero, it must be the wait parameters for the current thread
2119 // which holds the lock but is not runnable because its condition is false
2120 // or it is in the process of blocking on a condition variable; it must requeue
2121 // itself on the mutex/condvar to wait for its condition to become true.
UnlockSlow(SynchWaitParams * waitp)2122 ABSL_ATTRIBUTE_NOINLINE void Mutex::UnlockSlow(SynchWaitParams* waitp) {
2123   SchedulingGuard::ScopedDisable disable_rescheduling;
2124   intptr_t v = mu_.load(std::memory_order_relaxed);
2125   this->AssertReaderHeld();
2126   CheckForMutexCorruption(v, "Unlock");
2127   if ((v & kMuEvent) != 0) {
2128     PostSynchEvent(
2129         this, (v & kMuWriter) != 0 ? SYNCH_EV_UNLOCK : SYNCH_EV_READERUNLOCK);
2130   }
2131   int c = 0;
2132   // the waiter under consideration to wake, or zero
2133   PerThreadSynch* w = nullptr;
2134   // the predecessor to w or zero
2135   PerThreadSynch* pw = nullptr;
2136   // head of the list searched previously, or zero
2137   PerThreadSynch* old_h = nullptr;
2138   // a condition that's known to be false.
2139   PerThreadSynch* wake_list = kPerThreadSynchNull;  // list of threads to wake
2140   intptr_t wr_wait = 0;  // set to kMuWrWait if we wake a reader and a
2141                          // later writer could have acquired the lock
2142                          // (starvation avoidance)
2143   ABSL_RAW_CHECK(waitp == nullptr || waitp->thread->waitp == nullptr ||
2144                      waitp->thread->suppress_fatal_errors,
2145                  "detected illegal recursion into Mutex code");
2146   // This loop finds threads wake_list to wakeup if any, and removes them from
2147   // the list of waiters.  In addition, it places waitp.thread on the queue of
2148   // waiters if waitp is non-zero.
2149   for (;;) {
2150     v = mu_.load(std::memory_order_relaxed);
2151     if ((v & kMuWriter) != 0 && (v & (kMuWait | kMuDesig)) != kMuWait &&
2152         waitp == nullptr) {
2153       // fast writer release (writer with no waiters or with designated waker)
2154       if (mu_.compare_exchange_strong(v, v & ~(kMuWrWait | kMuWriter),
2155                                       std::memory_order_release,
2156                                       std::memory_order_relaxed)) {
2157         return;
2158       }
2159     } else if ((v & (kMuReader | kMuWait)) == kMuReader && waitp == nullptr) {
2160       // fast reader release (reader with no waiters)
2161       intptr_t clear = ExactlyOneReader(v) ? kMuReader | kMuOne : kMuOne;
2162       if (mu_.compare_exchange_strong(v, v - clear, std::memory_order_release,
2163                                       std::memory_order_relaxed)) {
2164         return;
2165       }
2166     } else if ((v & kMuSpin) == 0 &&  // attempt to get spinlock
2167                mu_.compare_exchange_strong(v, v | kMuSpin,
2168                                            std::memory_order_acquire,
2169                                            std::memory_order_relaxed)) {
2170       if ((v & kMuWait) == 0) {  // no one to wake
2171         intptr_t nv;
2172         bool do_enqueue = true;  // always Enqueue() the first time
2173         ABSL_RAW_CHECK(waitp != nullptr,
2174                        "UnlockSlow is confused");  // about to sleep
2175         do {  // must loop to release spinlock as reader count may change
2176           v = mu_.load(std::memory_order_relaxed);
2177           // decrement reader count if there are readers
2178           intptr_t new_readers = (v >= kMuOne) ? v - kMuOne : v;
2179           PerThreadSynch* new_h = nullptr;
2180           if (do_enqueue) {
2181             // If we are enqueuing on a CondVar (waitp->cv_word != nullptr) then
2182             // we must not retry here.  The initial attempt will always have
2183             // succeeded, further attempts would enqueue us against *this due to
2184             // Fer() handling.
2185             do_enqueue = (waitp->cv_word == nullptr);
2186             new_h = Enqueue(nullptr, waitp, new_readers, kMuIsCond);
2187           }
2188           intptr_t clear = kMuWrWait | kMuWriter;  // by default clear write bit
2189           if ((v & kMuWriter) == 0 && ExactlyOneReader(v)) {  // last reader
2190             clear = kMuWrWait | kMuReader;                    // clear read bit
2191           }
2192           nv = (v & kMuLow & ~clear & ~kMuSpin);
2193           if (new_h != nullptr) {
2194             nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
2195           } else {  // new_h could be nullptr if we queued ourselves on a
2196                     // CondVar
2197             // In that case, we must place the reader count back in the mutex
2198             // word, as Enqueue() did not store it in the new waiter.
2199             nv |= new_readers & kMuHigh;
2200           }
2201           // release spinlock & our lock; retry if reader-count changed
2202           // (writer count cannot change since we hold lock)
2203         } while (!mu_.compare_exchange_weak(v, nv, std::memory_order_release,
2204                                             std::memory_order_relaxed));
2205         break;
2206       }
2207 
2208       // There are waiters.
2209       // Set h to the head of the circular waiter list.
2210       PerThreadSynch* h = GetPerThreadSynch(v);
2211       if ((v & kMuReader) != 0 && (h->readers & kMuHigh) > kMuOne) {
2212         // a reader but not the last
2213         h->readers -= kMuOne;    // release our lock
2214         intptr_t nv = v;         // normally just release spinlock
2215         if (waitp != nullptr) {  // but waitp!=nullptr => must queue ourselves
2216           PerThreadSynch* new_h = Enqueue(h, waitp, v, kMuIsCond);
2217           ABSL_RAW_CHECK(new_h != nullptr,
2218                          "waiters disappeared during Enqueue()!");
2219           nv &= kMuLow;
2220           nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
2221         }
2222         mu_.store(nv, std::memory_order_release);  // release spinlock
2223         // can release with a store because there were waiters
2224         break;
2225       }
2226 
2227       // Either we didn't search before, or we marked the queue
2228       // as "maybe_unlocking" and no one else should have changed it.
2229       ABSL_RAW_CHECK(old_h == nullptr || h->maybe_unlocking,
2230                      "Mutex queue changed beneath us");
2231 
2232       // The lock is becoming free, and there's a waiter
2233       if (old_h != nullptr &&
2234           !old_h->may_skip) {    // we used old_h as a terminator
2235         old_h->may_skip = true;  // allow old_h to skip once more
2236         ABSL_RAW_CHECK(old_h->skip == nullptr, "illegal skip from head");
2237         if (h != old_h && MuEquivalentWaiter(old_h, old_h->next)) {
2238           old_h->skip = old_h->next;  // old_h not head & can skip to successor
2239         }
2240       }
2241       if (h->next->waitp->how == kExclusive &&
2242           h->next->waitp->cond == nullptr) {
2243         // easy case: writer with no condition; no need to search
2244         pw = h;  // wake w, the successor of h (=pw)
2245         w = h->next;
2246         w->wake = true;
2247         // We are waking up a writer.  This writer may be racing against
2248         // an already awake reader for the lock.  We want the
2249         // writer to usually win this race,
2250         // because if it doesn't, we can potentially keep taking a reader
2251         // perpetually and writers will starve.  Worse than
2252         // that, this can also starve other readers if kMuWrWait gets set
2253         // later.
2254         wr_wait = kMuWrWait;
2255       } else if (w != nullptr && (w->waitp->how == kExclusive || h == old_h)) {
2256         // we found a waiter w to wake on a previous iteration and either it's
2257         // a writer, or we've searched the entire list so we have all the
2258         // readers.
2259         if (pw == nullptr) {  // if w's predecessor is unknown, it must be h
2260           pw = h;
2261         }
2262       } else {
2263         // At this point we don't know all the waiters to wake, and the first
2264         // waiter has a condition or is a reader.  We avoid searching over
2265         // waiters we've searched on previous iterations by starting at
2266         // old_h if it's set.  If old_h==h, there's no one to wakeup at all.
2267         if (old_h == h) {  // we've searched before, and nothing's new
2268                            // so there's no one to wake.
2269           intptr_t nv = (v & ~(kMuReader | kMuWriter | kMuWrWait));
2270           h->readers = 0;
2271           h->maybe_unlocking = false;  // finished unlocking
2272           if (waitp != nullptr) {      // we must queue ourselves and sleep
2273             PerThreadSynch* new_h = Enqueue(h, waitp, v, kMuIsCond);
2274             nv &= kMuLow;
2275             if (new_h != nullptr) {
2276               nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
2277             }  // else new_h could be nullptr if we queued ourselves on a
2278                // CondVar
2279           }
2280           // release spinlock & lock
2281           // can release with a store because there were waiters
2282           mu_.store(nv, std::memory_order_release);
2283           break;
2284         }
2285 
2286         // set up to walk the list
2287         PerThreadSynch* w_walk;   // current waiter during list walk
2288         PerThreadSynch* pw_walk;  // previous waiter during list walk
2289         if (old_h != nullptr) {  // we've searched up to old_h before
2290           pw_walk = old_h;
2291           w_walk = old_h->next;
2292         } else {  // no prior search, start at beginning
2293           pw_walk =
2294               nullptr;  // h->next's predecessor may change; don't record it
2295           w_walk = h->next;
2296         }
2297 
2298         h->may_skip = false;  // ensure we never skip past h in future searches
2299                               // even if other waiters are queued after it.
2300         ABSL_RAW_CHECK(h->skip == nullptr, "illegal skip from head");
2301 
2302         h->maybe_unlocking = true;  // we're about to scan the waiter list
2303                                     // without the spinlock held.
2304                                     // Enqueue must be conservative about
2305                                     // priority queuing.
2306 
2307         // We must release the spinlock to evaluate the conditions.
2308         mu_.store(v, std::memory_order_release);  // release just spinlock
2309         // can release with a store because there were waiters
2310 
2311         // h is the last waiter queued, and w_walk the first unsearched waiter.
2312         // Without the spinlock, the locations mu_ and h->next may now change
2313         // underneath us, but since we hold the lock itself, the only legal
2314         // change is to add waiters between h and w_walk.  Therefore, it's safe
2315         // to walk the path from w_walk to h inclusive. (TryRemove() can remove
2316         // a waiter anywhere, but it acquires both the spinlock and the Mutex)
2317 
2318         old_h = h;  // remember we searched to here
2319 
2320         // Walk the path upto and including h looking for waiters we can wake.
2321         while (pw_walk != h) {
2322           w_walk->wake = false;
2323           if (w_walk->waitp->cond ==
2324                   nullptr ||  // no condition => vacuously true OR
2325                               // this thread's condition is true
2326               EvalConditionIgnored(this, w_walk->waitp->cond)) {
2327             if (w == nullptr) {
2328               w_walk->wake = true;  // can wake this waiter
2329               w = w_walk;
2330               pw = pw_walk;
2331               if (w_walk->waitp->how == kExclusive) {
2332                 wr_wait = kMuWrWait;
2333                 break;  // bail if waking this writer
2334               }
2335             } else if (w_walk->waitp->how == kShared) {  // wake if a reader
2336               w_walk->wake = true;
2337             } else {  // writer with true condition
2338               wr_wait = kMuWrWait;
2339             }
2340           }
2341           if (w_walk->wake) {  // we're waking reader w_walk
2342             pw_walk = w_walk;  // don't skip similar waiters
2343           } else {             // not waking; skip as much as possible
2344             pw_walk = Skip(w_walk);
2345           }
2346           // If pw_walk == h, then load of pw_walk->next can race with
2347           // concurrent write in Enqueue(). However, at the same time
2348           // we do not need to do the load, because we will bail out
2349           // from the loop anyway.
2350           if (pw_walk != h) {
2351             w_walk = pw_walk->next;
2352           }
2353         }
2354 
2355         continue;  // restart for(;;)-loop to wakeup w or to find more waiters
2356       }
2357       ABSL_RAW_CHECK(pw->next == w, "pw not w's predecessor");
2358       // The first (and perhaps only) waiter we've chosen to wake is w, whose
2359       // predecessor is pw.  If w is a reader, we must wake all the other
2360       // waiters with wake==true as well.  We may also need to queue
2361       // ourselves if waitp != null.  The spinlock and the lock are still
2362       // held.
2363 
2364       // This traverses the list in [ pw->next, h ], where h is the head,
2365       // removing all elements with wake==true and placing them in the
2366       // singly-linked list wake_list.  Returns the new head.
2367       h = DequeueAllWakeable(h, pw, &wake_list);
2368 
2369       intptr_t nv = (v & kMuEvent) | kMuDesig;
2370       // assume no waiters left,
2371       // set kMuDesig for INV1a
2372 
2373       if (waitp != nullptr) {  // we must queue ourselves and sleep
2374         h = Enqueue(h, waitp, v, kMuIsCond);
2375         // h is new last waiter; could be null if we queued ourselves on a
2376         // CondVar
2377       }
2378 
2379       ABSL_RAW_CHECK(wake_list != kPerThreadSynchNull,
2380                      "unexpected empty wake list");
2381 
2382       if (h != nullptr) {  // there are waiters left
2383         h->readers = 0;
2384         h->maybe_unlocking = false;  // finished unlocking
2385         nv |= wr_wait | kMuWait | reinterpret_cast<intptr_t>(h);
2386       }
2387 
2388       // release both spinlock & lock
2389       // can release with a store because there were waiters
2390       mu_.store(nv, std::memory_order_release);
2391       break;  // out of for(;;)-loop
2392     }
2393     // aggressive here; no one can proceed till we do
2394     c = synchronization_internal::MutexDelay(c, AGGRESSIVE);
2395   }  // end of for(;;)-loop
2396 
2397   if (wake_list != kPerThreadSynchNull) {
2398     int64_t total_wait_cycles = 0;
2399     int64_t max_wait_cycles = 0;
2400     int64_t now = CycleClock::Now();
2401     do {
2402       // Profile lock contention events only if the waiter was trying to acquire
2403       // the lock, not waiting on a condition variable or Condition.
2404       if (!wake_list->cond_waiter) {
2405         int64_t cycles_waited =
2406             (now - wake_list->waitp->contention_start_cycles);
2407         total_wait_cycles += cycles_waited;
2408         if (max_wait_cycles == 0) max_wait_cycles = cycles_waited;
2409         wake_list->waitp->contention_start_cycles = now;
2410         wake_list->waitp->should_submit_contention_data = true;
2411       }
2412       wake_list = Wakeup(wake_list);  // wake waiters
2413     } while (wake_list != kPerThreadSynchNull);
2414     if (total_wait_cycles > 0) {
2415       mutex_tracer("slow release", this, total_wait_cycles);
2416       ABSL_TSAN_MUTEX_PRE_DIVERT(this, 0);
2417       submit_profile_data(total_wait_cycles);
2418       ABSL_TSAN_MUTEX_POST_DIVERT(this, 0);
2419     }
2420   }
2421 }
2422 
2423 // Used by CondVar implementation to reacquire mutex after waking from
2424 // condition variable.  This routine is used instead of Lock() because the
2425 // waiting thread may have been moved from the condition variable queue to the
2426 // mutex queue without a wakeup, by Trans().  In that case, when the thread is
2427 // finally woken, the woken thread will believe it has been woken from the
2428 // condition variable (i.e. its PC will be in when in the CondVar code), when
2429 // in fact it has just been woken from the mutex.  Thus, it must enter the slow
2430 // path of the mutex in the same state as if it had just woken from the mutex.
2431 // That is, it must ensure to clear kMuDesig (INV1b).
Trans(MuHow how)2432 void Mutex::Trans(MuHow how) {
2433   this->LockSlow(how, nullptr, kMuHasBlocked | kMuIsCond);
2434 }
2435 
2436 // Used by CondVar implementation to effectively wake thread w from the
2437 // condition variable.  If this mutex is free, we simply wake the thread.
2438 // It will later acquire the mutex with high probability.  Otherwise, we
2439 // enqueue thread w on this mutex.
Fer(PerThreadSynch * w)2440 void Mutex::Fer(PerThreadSynch* w) {
2441   SchedulingGuard::ScopedDisable disable_rescheduling;
2442   int c = 0;
2443   ABSL_RAW_CHECK(w->waitp->cond == nullptr,
2444                  "Mutex::Fer while waiting on Condition");
2445   ABSL_RAW_CHECK(w->waitp->cv_word == nullptr,
2446                  "Mutex::Fer with pending CondVar queueing");
2447   // The CondVar timeout is not relevant for the Mutex wait.
2448   w->waitp->timeout = {};
2449   for (;;) {
2450     intptr_t v = mu_.load(std::memory_order_relaxed);
2451     // Note: must not queue if the mutex is unlocked (nobody will wake it).
2452     // For example, we can have only kMuWait (conditional) or maybe
2453     // kMuWait|kMuWrWait.
2454     // conflicting != 0 implies that the waking thread cannot currently take
2455     // the mutex, which in turn implies that someone else has it and can wake
2456     // us if we queue.
2457     const intptr_t conflicting =
2458         kMuWriter | (w->waitp->how == kShared ? 0 : kMuReader);
2459     if ((v & conflicting) == 0) {
2460       w->next = nullptr;
2461       w->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
2462       IncrementSynchSem(this, w);
2463       return;
2464     } else {
2465       if ((v & (kMuSpin | kMuWait)) == 0) {  // no waiters
2466         // This thread tries to become the one and only waiter.
2467         PerThreadSynch* new_h =
2468             Enqueue(nullptr, w->waitp, v, kMuIsCond | kMuIsFer);
2469         ABSL_RAW_CHECK(new_h != nullptr,
2470                        "Enqueue failed");  // we must queue ourselves
2471         if (mu_.compare_exchange_strong(
2472                 v, reinterpret_cast<intptr_t>(new_h) | (v & kMuLow) | kMuWait,
2473                 std::memory_order_release, std::memory_order_relaxed)) {
2474           return;
2475         }
2476       } else if ((v & kMuSpin) == 0 &&
2477                  mu_.compare_exchange_strong(v, v | kMuSpin | kMuWait)) {
2478         PerThreadSynch* h = GetPerThreadSynch(v);
2479         PerThreadSynch* new_h = Enqueue(h, w->waitp, v, kMuIsCond | kMuIsFer);
2480         ABSL_RAW_CHECK(new_h != nullptr,
2481                        "Enqueue failed");  // we must queue ourselves
2482         do {
2483           v = mu_.load(std::memory_order_relaxed);
2484         } while (!mu_.compare_exchange_weak(
2485             v,
2486             (v & kMuLow & ~kMuSpin) | kMuWait |
2487                 reinterpret_cast<intptr_t>(new_h),
2488             std::memory_order_release, std::memory_order_relaxed));
2489         return;
2490       }
2491     }
2492     c = synchronization_internal::MutexDelay(c, GENTLE);
2493   }
2494 }
2495 
AssertHeld() const2496 void Mutex::AssertHeld() const {
2497   if ((mu_.load(std::memory_order_relaxed) & kMuWriter) == 0) {
2498     SynchEvent* e = GetSynchEvent(this);
2499     ABSL_RAW_LOG(FATAL, "thread should hold write lock on Mutex %p %s",
2500                  static_cast<const void*>(this), (e == nullptr ? "" : e->name));
2501   }
2502 }
2503 
AssertReaderHeld() const2504 void Mutex::AssertReaderHeld() const {
2505   if ((mu_.load(std::memory_order_relaxed) & (kMuReader | kMuWriter)) == 0) {
2506     SynchEvent* e = GetSynchEvent(this);
2507     ABSL_RAW_LOG(FATAL,
2508                  "thread should hold at least a read lock on Mutex %p %s",
2509                  static_cast<const void*>(this), (e == nullptr ? "" : e->name));
2510   }
2511 }
2512 
2513 // -------------------------------- condition variables
2514 static const intptr_t kCvSpin = 0x0001L;   // spinlock protects waiter list
2515 static const intptr_t kCvEvent = 0x0002L;  // record events
2516 
2517 static const intptr_t kCvLow = 0x0003L;  // low order bits of CV
2518 
2519 // Hack to make constant values available to gdb pretty printer
2520 enum {
2521   kGdbCvSpin = kCvSpin,
2522   kGdbCvEvent = kCvEvent,
2523   kGdbCvLow = kCvLow,
2524 };
2525 
2526 static_assert(PerThreadSynch::kAlignment > kCvLow,
2527               "PerThreadSynch::kAlignment must be greater than kCvLow");
2528 
EnableDebugLog(const char * name)2529 void CondVar::EnableDebugLog(const char* name) {
2530   SynchEvent* e = EnsureSynchEvent(&this->cv_, name, kCvEvent, kCvSpin);
2531   e->log = true;
2532   UnrefSynchEvent(e);
2533 }
2534 
2535 // Remove thread s from the list of waiters on this condition variable.
Remove(PerThreadSynch * s)2536 void CondVar::Remove(PerThreadSynch* s) {
2537   SchedulingGuard::ScopedDisable disable_rescheduling;
2538   intptr_t v;
2539   int c = 0;
2540   for (v = cv_.load(std::memory_order_relaxed);;
2541        v = cv_.load(std::memory_order_relaxed)) {
2542     if ((v & kCvSpin) == 0 &&  // attempt to acquire spinlock
2543         cv_.compare_exchange_strong(v, v | kCvSpin, std::memory_order_acquire,
2544                                     std::memory_order_relaxed)) {
2545       PerThreadSynch* h = reinterpret_cast<PerThreadSynch*>(v & ~kCvLow);
2546       if (h != nullptr) {
2547         PerThreadSynch* w = h;
2548         while (w->next != s && w->next != h) {  // search for thread
2549           w = w->next;
2550         }
2551         if (w->next == s) {  // found thread; remove it
2552           w->next = s->next;
2553           if (h == s) {
2554             h = (w == s) ? nullptr : w;
2555           }
2556           s->next = nullptr;
2557           s->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
2558         }
2559       }
2560       // release spinlock
2561       cv_.store((v & kCvEvent) | reinterpret_cast<intptr_t>(h),
2562                 std::memory_order_release);
2563       return;
2564     } else {
2565       // try again after a delay
2566       c = synchronization_internal::MutexDelay(c, GENTLE);
2567     }
2568   }
2569 }
2570 
2571 // Queue thread waitp->thread on condition variable word cv_word using
2572 // wait parameters waitp.
2573 // We split this into a separate routine, rather than simply doing it as part
2574 // of WaitCommon().  If we were to queue ourselves on the condition variable
2575 // before calling Mutex::UnlockSlow(), the Mutex code might be re-entered (via
2576 // the logging code, or via a Condition function) and might potentially attempt
2577 // to block this thread.  That would be a problem if the thread were already on
2578 // a condition variable waiter queue.  Thus, we use the waitp->cv_word to tell
2579 // the unlock code to call CondVarEnqueue() to queue the thread on the condition
2580 // variable queue just before the mutex is to be unlocked, and (most
2581 // importantly) after any call to an external routine that might re-enter the
2582 // mutex code.
CondVarEnqueue(SynchWaitParams * waitp)2583 static void CondVarEnqueue(SynchWaitParams* waitp) {
2584   // This thread might be transferred to the Mutex queue by Fer() when
2585   // we are woken.  To make sure that is what happens, Enqueue() doesn't
2586   // call CondVarEnqueue() again but instead uses its normal code.  We
2587   // must do this before we queue ourselves so that cv_word will be null
2588   // when seen by the dequeuer, who may wish immediately to requeue
2589   // this thread on another queue.
2590   std::atomic<intptr_t>* cv_word = waitp->cv_word;
2591   waitp->cv_word = nullptr;
2592 
2593   intptr_t v = cv_word->load(std::memory_order_relaxed);
2594   int c = 0;
2595   while ((v & kCvSpin) != 0 ||  // acquire spinlock
2596          !cv_word->compare_exchange_weak(v, v | kCvSpin,
2597                                          std::memory_order_acquire,
2598                                          std::memory_order_relaxed)) {
2599     c = synchronization_internal::MutexDelay(c, GENTLE);
2600     v = cv_word->load(std::memory_order_relaxed);
2601   }
2602   ABSL_RAW_CHECK(waitp->thread->waitp == nullptr, "waiting when shouldn't be");
2603   waitp->thread->waitp = waitp;  // prepare ourselves for waiting
2604   PerThreadSynch* h = reinterpret_cast<PerThreadSynch*>(v & ~kCvLow);
2605   if (h == nullptr) {  // add this thread to waiter list
2606     waitp->thread->next = waitp->thread;
2607   } else {
2608     waitp->thread->next = h->next;
2609     h->next = waitp->thread;
2610   }
2611   waitp->thread->state.store(PerThreadSynch::kQueued,
2612                              std::memory_order_relaxed);
2613   cv_word->store((v & kCvEvent) | reinterpret_cast<intptr_t>(waitp->thread),
2614                  std::memory_order_release);
2615 }
2616 
WaitCommon(Mutex * mutex,KernelTimeout t)2617 bool CondVar::WaitCommon(Mutex* mutex, KernelTimeout t) {
2618   bool rc = false;  // return value; true iff we timed-out
2619 
2620   intptr_t mutex_v = mutex->mu_.load(std::memory_order_relaxed);
2621   Mutex::MuHow mutex_how = ((mutex_v & kMuWriter) != 0) ? kExclusive : kShared;
2622   ABSL_TSAN_MUTEX_PRE_UNLOCK(mutex, TsanFlags(mutex_how));
2623 
2624   // maybe trace this call
2625   intptr_t v = cv_.load(std::memory_order_relaxed);
2626   cond_var_tracer("Wait", this);
2627   if ((v & kCvEvent) != 0) {
2628     PostSynchEvent(this, SYNCH_EV_WAIT);
2629   }
2630 
2631   // Release mu and wait on condition variable.
2632   SynchWaitParams waitp(mutex_how, nullptr, t, mutex,
2633                         Synch_GetPerThreadAnnotated(mutex), &cv_);
2634   // UnlockSlow() will call CondVarEnqueue() just before releasing the
2635   // Mutex, thus queuing this thread on the condition variable.  See
2636   // CondVarEnqueue() for the reasons.
2637   mutex->UnlockSlow(&waitp);
2638 
2639   // wait for signal
2640   while (waitp.thread->state.load(std::memory_order_acquire) ==
2641          PerThreadSynch::kQueued) {
2642     if (!Mutex::DecrementSynchSem(mutex, waitp.thread, t)) {
2643       // DecrementSynchSem returned due to timeout.
2644       // Now we will either (1) remove ourselves from the wait list in Remove
2645       // below, in which case Remove will set thread.state = kAvailable and
2646       // we will not call DecrementSynchSem again; or (2) Signal/SignalAll
2647       // has removed us concurrently and is calling Wakeup, which will set
2648       // thread.state = kAvailable and post to the semaphore.
2649       // It's important to reset the timeout for the case (2) because otherwise
2650       // we can live-lock in this loop since DecrementSynchSem will always
2651       // return immediately due to timeout, but Signal/SignalAll is not
2652       // necessary set thread.state = kAvailable yet (and is not scheduled
2653       // due to thread priorities or other scheduler artifacts).
2654       // Note this could also be resolved if Signal/SignalAll would set
2655       // thread.state = kAvailable while holding the wait list spin lock.
2656       // But this can't be easily done for SignalAll since it grabs the whole
2657       // wait list with a single compare-exchange and does not really grab
2658       // the spin lock.
2659       t = KernelTimeout::Never();
2660       this->Remove(waitp.thread);
2661       rc = true;
2662     }
2663   }
2664 
2665   ABSL_RAW_CHECK(waitp.thread->waitp != nullptr, "not waiting when should be");
2666   waitp.thread->waitp = nullptr;  // cleanup
2667 
2668   // maybe trace this call
2669   cond_var_tracer("Unwait", this);
2670   if ((v & kCvEvent) != 0) {
2671     PostSynchEvent(this, SYNCH_EV_WAIT_RETURNING);
2672   }
2673 
2674   // From synchronization point of view Wait is unlock of the mutex followed
2675   // by lock of the mutex. We've annotated start of unlock in the beginning
2676   // of the function. Now, finish unlock and annotate lock of the mutex.
2677   // (Trans is effectively lock).
2678   ABSL_TSAN_MUTEX_POST_UNLOCK(mutex, TsanFlags(mutex_how));
2679   ABSL_TSAN_MUTEX_PRE_LOCK(mutex, TsanFlags(mutex_how));
2680   mutex->Trans(mutex_how);  // Reacquire mutex
2681   ABSL_TSAN_MUTEX_POST_LOCK(mutex, TsanFlags(mutex_how), 0);
2682   return rc;
2683 }
2684 
Signal()2685 void CondVar::Signal() {
2686   SchedulingGuard::ScopedDisable disable_rescheduling;
2687   ABSL_TSAN_MUTEX_PRE_SIGNAL(nullptr, 0);
2688   intptr_t v;
2689   int c = 0;
2690   for (v = cv_.load(std::memory_order_relaxed); v != 0;
2691        v = cv_.load(std::memory_order_relaxed)) {
2692     if ((v & kCvSpin) == 0 &&  // attempt to acquire spinlock
2693         cv_.compare_exchange_strong(v, v | kCvSpin, std::memory_order_acquire,
2694                                     std::memory_order_relaxed)) {
2695       PerThreadSynch* h = reinterpret_cast<PerThreadSynch*>(v & ~kCvLow);
2696       PerThreadSynch* w = nullptr;
2697       if (h != nullptr) {  // remove first waiter
2698         w = h->next;
2699         if (w == h) {
2700           h = nullptr;
2701         } else {
2702           h->next = w->next;
2703         }
2704       }
2705       // release spinlock
2706       cv_.store((v & kCvEvent) | reinterpret_cast<intptr_t>(h),
2707                 std::memory_order_release);
2708       if (w != nullptr) {
2709         w->waitp->cvmu->Fer(w);  // wake waiter, if there was one
2710         cond_var_tracer("Signal wakeup", this);
2711       }
2712       if ((v & kCvEvent) != 0) {
2713         PostSynchEvent(this, SYNCH_EV_SIGNAL);
2714       }
2715       ABSL_TSAN_MUTEX_POST_SIGNAL(nullptr, 0);
2716       return;
2717     } else {
2718       c = synchronization_internal::MutexDelay(c, GENTLE);
2719     }
2720   }
2721   ABSL_TSAN_MUTEX_POST_SIGNAL(nullptr, 0);
2722 }
2723 
SignalAll()2724 void CondVar::SignalAll() {
2725   ABSL_TSAN_MUTEX_PRE_SIGNAL(nullptr, 0);
2726   intptr_t v;
2727   int c = 0;
2728   for (v = cv_.load(std::memory_order_relaxed); v != 0;
2729        v = cv_.load(std::memory_order_relaxed)) {
2730     // empty the list if spinlock free
2731     // We do this by simply setting the list to empty using
2732     // compare and swap.   We then have the entire list in our hands,
2733     // which cannot be changing since we grabbed it while no one
2734     // held the lock.
2735     if ((v & kCvSpin) == 0 &&
2736         cv_.compare_exchange_strong(v, v & kCvEvent, std::memory_order_acquire,
2737                                     std::memory_order_relaxed)) {
2738       PerThreadSynch* h = reinterpret_cast<PerThreadSynch*>(v & ~kCvLow);
2739       if (h != nullptr) {
2740         PerThreadSynch* w;
2741         PerThreadSynch* n = h->next;
2742         do {  // for every thread, wake it up
2743           w = n;
2744           n = n->next;
2745           w->waitp->cvmu->Fer(w);
2746         } while (w != h);
2747         cond_var_tracer("SignalAll wakeup", this);
2748       }
2749       if ((v & kCvEvent) != 0) {
2750         PostSynchEvent(this, SYNCH_EV_SIGNALALL);
2751       }
2752       ABSL_TSAN_MUTEX_POST_SIGNAL(nullptr, 0);
2753       return;
2754     } else {
2755       // try again after a delay
2756       c = synchronization_internal::MutexDelay(c, GENTLE);
2757     }
2758   }
2759   ABSL_TSAN_MUTEX_POST_SIGNAL(nullptr, 0);
2760 }
2761 
Release()2762 void ReleasableMutexLock::Release() {
2763   ABSL_RAW_CHECK(this->mu_ != nullptr,
2764                  "ReleasableMutexLock::Release may only be called once");
2765   this->mu_->Unlock();
2766   this->mu_ = nullptr;
2767 }
2768 
2769 #ifdef ABSL_HAVE_THREAD_SANITIZER
2770 extern "C" void __tsan_read1(void* addr);
2771 #else
2772 #define __tsan_read1(addr)  // do nothing if TSan not enabled
2773 #endif
2774 
2775 // A function that just returns its argument, dereferenced
Dereference(void * arg)2776 static bool Dereference(void* arg) {
2777   // ThreadSanitizer does not instrument this file for memory accesses.
2778   // This function dereferences a user variable that can participate
2779   // in a data race, so we need to manually tell TSan about this memory access.
2780   __tsan_read1(arg);
2781   return *(static_cast<bool*>(arg));
2782 }
2783 
2784 ABSL_CONST_INIT const Condition Condition::kTrue;
2785 
Condition(bool (* func)(void *),void * arg)2786 Condition::Condition(bool (*func)(void*), void* arg)
2787     : eval_(&CallVoidPtrFunction), arg_(arg) {
2788   static_assert(sizeof(&func) <= sizeof(callback_),
2789                 "An overlarge function pointer passed to Condition.");
2790   StoreCallback(func);
2791 }
2792 
CallVoidPtrFunction(const Condition * c)2793 bool Condition::CallVoidPtrFunction(const Condition* c) {
2794   using FunctionPointer = bool (*)(void*);
2795   FunctionPointer function_pointer;
2796   std::memcpy(&function_pointer, c->callback_, sizeof(function_pointer));
2797   return (*function_pointer)(c->arg_);
2798 }
2799 
Condition(const bool * cond)2800 Condition::Condition(const bool* cond)
2801     : eval_(CallVoidPtrFunction),
2802       // const_cast is safe since Dereference does not modify arg
2803       arg_(const_cast<bool*>(cond)) {
2804   using FunctionPointer = bool (*)(void*);
2805   const FunctionPointer dereference = Dereference;
2806   StoreCallback(dereference);
2807 }
2808 
Eval() const2809 bool Condition::Eval() const { return (*this->eval_)(this); }
2810 
GuaranteedEqual(const Condition * a,const Condition * b)2811 bool Condition::GuaranteedEqual(const Condition* a, const Condition* b) {
2812   if (a == nullptr || b == nullptr) {
2813     return a == b;
2814   }
2815   // Check equality of the representative fields.
2816   return a->eval_ == b->eval_ && a->arg_ == b->arg_ &&
2817          !memcmp(a->callback_, b->callback_, sizeof(a->callback_));
2818 }
2819 
2820 ABSL_NAMESPACE_END
2821 }  // namespace absl
2822