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, ¶m);
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