1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a part of ThreadSanitizer (TSan), a race detector.
12 // Main internal TSan header file.
15 // - C++ run-time should not be used (static CTORs, RTTI, exceptions, static
16 // function-scope locals)
17 // - All functions/classes/etc reside in namespace __tsan, except for those
18 // declared in tsan_interface.h.
19 // - Platform-specific files should be used instead of ifdefs (*).
20 // - No system headers included in header files (*).
21 // - Platform specific headres included only into platform-specific files (*).
23 // (*) Except when inlining is critical for performance.
24 //===----------------------------------------------------------------------===//
29 #include "sanitizer_common/sanitizer_allocator.h"
30 #include "sanitizer_common/sanitizer_allocator_internal.h"
31 #include "sanitizer_common/sanitizer_asm.h"
32 #include "sanitizer_common/sanitizer_common.h"
33 #include "sanitizer_common/sanitizer_deadlock_detector_interface.h"
34 #include "sanitizer_common/sanitizer_libignore.h"
35 #include "sanitizer_common/sanitizer_suppressions.h"
36 #include "sanitizer_common/sanitizer_thread_registry.h"
37 #include "tsan_clock.h"
38 #include "tsan_defs.h"
39 #include "tsan_flags.h"
40 #include "tsan_sync.h"
41 #include "tsan_trace.h"
42 #include "tsan_vector.h"
43 #include "tsan_report.h"
44 #include "tsan_platform.h"
45 #include "tsan_mutexset.h"
46 #include "tsan_ignoreset.h"
47 #include "tsan_stack_trace.h"
49 #if SANITIZER_WORDSIZE != 64
50 # error "ThreadSanitizer is supported only on 64-bit platforms"
56 struct MapUnmapCallback;
57 #if defined(__mips64) || defined(__aarch64__) || defined(__powerpc__)
58 static const uptr kAllocatorSpace = 0;
59 static const uptr kAllocatorSize = SANITIZER_MMAP_RANGE_SIZE;
60 static const uptr kAllocatorRegionSizeLog = 20;
61 static const uptr kAllocatorNumRegions =
62 kAllocatorSize >> kAllocatorRegionSizeLog;
63 typedef TwoLevelByteMap<(kAllocatorNumRegions >> 12), 1 << 12,
64 MapUnmapCallback> ByteMap;
65 typedef SizeClassAllocator32<kAllocatorSpace, kAllocatorSize, 0,
66 CompactSizeClassMap, kAllocatorRegionSizeLog, ByteMap,
67 MapUnmapCallback> PrimaryAllocator;
69 typedef SizeClassAllocator64<Mapping::kHeapMemBeg,
70 Mapping::kHeapMemEnd - Mapping::kHeapMemBeg, 0,
71 DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
73 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
74 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
75 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
76 SecondaryAllocator> Allocator;
77 Allocator *allocator();
80 void TsanCheckFailed(const char *file, int line, const char *cond,
83 const u64 kShadowRodata = (u64)-1; // .rodata shadow marker
85 // FastState (from most significant bit):
93 FastState(u64 tid, u64 epoch) {
94 x_ = tid << kTidShift;
96 DCHECK_EQ(tid, this->tid());
97 DCHECK_EQ(epoch, this->epoch());
98 DCHECK_EQ(GetIgnoreBit(), false);
101 explicit FastState(u64 x)
110 u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
114 u64 TidWithIgnore() const {
115 u64 res = x_ >> kTidShift;
120 u64 res = x_ & ((1ull << kClkBits) - 1);
124 void IncrementEpoch() {
125 u64 old_epoch = epoch();
127 DCHECK_EQ(old_epoch + 1, epoch());
131 void SetIgnoreBit() { x_ |= kIgnoreBit; }
132 void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
133 bool GetIgnoreBit() const { return (s64)x_ < 0; }
135 void SetHistorySize(int hs) {
138 x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
142 int GetHistorySize() const {
143 return (int)((x_ >> kHistoryShift) & kHistoryMask);
146 void ClearHistorySize() {
151 u64 GetTracePos() const {
152 const int hs = GetHistorySize();
153 // When hs == 0, the trace consists of 2 parts.
154 const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
155 return epoch() & mask;
160 static const int kTidShift = 64 - kTidBits - 1;
161 static const u64 kIgnoreBit = 1ull << 63;
162 static const u64 kFreedBit = 1ull << 63;
163 static const u64 kHistoryShift = kClkBits;
164 static const u64 kHistoryMask = 7;
168 // Shadow (from most significant bit):
176 class Shadow : public FastState {
178 explicit Shadow(u64 x)
182 explicit Shadow(const FastState &s)
187 void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
188 DCHECK_EQ((x_ >> kClkBits) & 31, 0);
190 DCHECK_LE(kAccessSizeLog, 3);
191 x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
192 DCHECK_EQ(kAccessSizeLog, size_log());
193 DCHECK_EQ(addr0, this->addr0());
196 void SetWrite(unsigned kAccessIsWrite) {
197 DCHECK_EQ(x_ & kReadBit, 0);
200 DCHECK_EQ(kAccessIsWrite, IsWrite());
203 void SetAtomic(bool kIsAtomic) {
207 DCHECK_EQ(IsAtomic(), kIsAtomic);
210 bool IsAtomic() const {
211 return x_ & kAtomicBit;
214 bool IsZero() const {
218 static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
219 u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
220 DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
221 return shifted_xor == 0;
225 bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
226 u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
227 return masked_xor == 0;
230 static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
231 unsigned kS2AccessSize) {
233 u64 diff = s1.addr0() - s2.addr0();
234 if ((s64)diff < 0) { // s1.addr0 < s2.addr0 // NOLINT
235 // if (s1.addr0() + size1) > s2.addr0()) return true;
236 if (s1.size() > -diff)
239 // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
240 if (kS2AccessSize > diff)
243 DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
244 DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
248 u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
249 u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
250 bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
251 bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
253 // The idea behind the freed bit is as follows.
254 // When the memory is freed (or otherwise unaccessible) we write to the shadow
255 // values with tid/epoch related to the free and the freed bit set.
256 // During memory accesses processing the freed bit is considered
257 // as msb of tid. So any access races with shadow with freed bit set
258 // (it is as if write from a thread with which we never synchronized before).
259 // This allows us to detect accesses to freed memory w/o additional
260 // overheads in memory access processing and at the same time restore
261 // tid/epoch of free.
266 bool IsFreed() const {
267 return x_ & kFreedBit;
270 bool GetFreedAndReset() {
271 bool res = x_ & kFreedBit;
276 bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
277 bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
278 | (u64(kIsAtomic) << kAtomicShift));
279 DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
283 bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
284 bool v = ((x_ >> kReadShift) & 3)
285 <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
286 DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
287 (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
291 bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
292 bool v = ((x_ >> kReadShift) & 3)
293 >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
294 DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
295 (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
300 static const u64 kReadShift = 5 + kClkBits;
301 static const u64 kReadBit = 1ull << kReadShift;
302 static const u64 kAtomicShift = 6 + kClkBits;
303 static const u64 kAtomicBit = 1ull << kAtomicShift;
305 u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
307 static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
308 if (s1.addr0() == s2.addr0()) return true;
309 if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
311 if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
317 struct ThreadSignalContext;
323 bool in_blocking_func;
324 uptr in_signal_handler;
325 uptr *shadow_stack_pos;
328 // A Processor represents a physical thread, or a P for Go.
329 // It is used to store internal resources like allocate cache, and does not
330 // participate in race-detection logic (invisible to end user).
331 // In C++ it is tied to an OS thread just like ThreadState, however ideally
332 // it should be tied to a CPU (this way we will have fewer allocator caches).
333 // In Go it is tied to a P, so there are significantly fewer Processor's than
334 // ThreadState's (which are tied to Gs).
335 // A ThreadState must be wired with a Processor to handle events.
337 ThreadState *thr; // currently wired thread, or nullptr
339 AllocatorCache alloc_cache;
340 InternalAllocatorCache internal_alloc_cache;
342 DenseSlabAllocCache block_cache;
343 DenseSlabAllocCache sync_cache;
344 DenseSlabAllocCache clock_cache;
345 DDPhysicalThread *dd_pt;
349 // ScopedGlobalProcessor temporary setups a global processor for the current
350 // thread, if it does not have one. Intended for interceptors that can run
351 // at the very thread end, when we already destroyed the thread processor.
352 struct ScopedGlobalProcessor {
353 ScopedGlobalProcessor();
354 ~ScopedGlobalProcessor();
358 // This struct is stored in TLS.
360 FastState fast_state;
361 // Synch epoch represents the threads's epoch before the last synchronization
362 // action. It allows to reduce number of shadow state updates.
363 // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
364 // if we are processing write to X from the same thread at epoch=200,
365 // we do nothing, because both writes happen in the same 'synch epoch'.
366 // That is, if another memory access does not race with the former write,
367 // it does not race with the latter as well.
368 // QUESTION: can we can squeeze this into ThreadState::Fast?
369 // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
370 // taken by epoch between synchs.
371 // This way we can save one load from tls.
372 u64 fast_synch_epoch;
373 // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
374 // We do not distinguish beteween ignoring reads and writes
375 // for better performance.
376 int ignore_reads_and_writes;
378 // Go does not support ignores.
380 IgnoreSet mop_ignore_set;
381 IgnoreSet sync_ignore_set;
383 // C/C++ uses fixed size shadow stack embed into Trace.
384 // Go uses malloc-allocated shadow stack with dynamic size.
386 uptr *shadow_stack_end;
387 uptr *shadow_stack_pos;
388 u64 *racy_shadow_addr;
393 Vector<JmpBuf> jmp_bufs;
394 int ignore_interceptors;
396 #if TSAN_COLLECT_STATS
413 #if SANITIZER_DEBUG && !SANITIZER_GO
414 InternalDeadlockDetector internal_deadlock_detector;
416 DDLogicalThread *dd_lt;
418 // Current wired Processor, or nullptr. Required to handle any events.
421 Processor *proc() { return proc1; }
426 atomic_uintptr_t in_signal_handler;
427 ThreadSignalContext *signal_ctx;
430 u32 last_sleep_stack_id;
431 ThreadClock last_sleep_clock;
434 // Set in regions of runtime that must be signal-safe and fork-safe.
435 // If set, malloc must not be called.
438 const ReportDesc *current_report;
440 explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
441 unsigned reuse_count,
442 uptr stk_addr, uptr stk_size,
443 uptr tls_addr, uptr tls_size);
447 #if SANITIZER_MAC || SANITIZER_ANDROID
448 ThreadState *cur_thread();
449 void cur_thread_finalize();
451 __attribute__((tls_model("initial-exec")))
452 extern THREADLOCAL char cur_thread_placeholder[];
453 INLINE ThreadState *cur_thread() {
454 return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
456 INLINE void cur_thread_finalize() { }
457 #endif // SANITIZER_MAC || SANITIZER_ANDROID
458 #endif // SANITIZER_GO
460 class ThreadContext : public ThreadContextBase {
462 explicit ThreadContext(int tid);
465 u32 creation_stack_id;
467 // Epoch at which the thread had started.
468 // If we see an event from the thread stamped by an older epoch,
469 // the event is from a dead thread that shared tid with this thread.
473 // Override superclass callbacks.
474 void OnDead() override;
475 void OnJoined(void *arg) override;
476 void OnFinished() override;
477 void OnStarted(void *arg) override;
478 void OnCreated(void *arg) override;
479 void OnReset() override;
480 void OnDetached(void *arg) override;
485 bool operator==(const RacyStacks &other) const {
486 if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
488 if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
499 struct FiredSuppression {
509 bool after_multithreaded_fork;
515 int nmissed_expected;
516 atomic_uint64_t last_symbolize_time_ns;
518 void *background_thread;
519 atomic_uint32_t stop_background_thread;
521 ThreadRegistry *thread_registry;
524 Vector<RacyStacks> racy_stacks;
525 Vector<RacyAddress> racy_addresses;
526 // Number of fired suppressions may be large enough.
527 Mutex fired_suppressions_mtx;
528 InternalMmapVector<FiredSuppression> fired_suppressions;
531 ClockAlloc clock_alloc;
536 u64 int_alloc_cnt[MBlockTypeCount];
537 u64 int_alloc_siz[MBlockTypeCount];
540 extern Context *ctx; // The one and the only global runtime context.
542 struct ScopedIgnoreInterceptors {
543 ScopedIgnoreInterceptors() {
545 cur_thread()->ignore_interceptors++;
549 ~ScopedIgnoreInterceptors() {
551 cur_thread()->ignore_interceptors--;
558 explicit ScopedReport(ReportType typ);
561 void AddMemoryAccess(uptr addr, Shadow s, StackTrace stack,
562 const MutexSet *mset);
563 void AddStack(StackTrace stack, bool suppressable = false);
564 void AddThread(const ThreadContext *tctx, bool suppressable = false);
565 void AddThread(int unique_tid, bool suppressable = false);
566 void AddUniqueTid(int unique_tid);
567 void AddMutex(const SyncVar *s);
568 u64 AddMutex(u64 id);
569 void AddLocation(uptr addr, uptr size);
570 void AddSleep(u32 stack_id);
571 void SetCount(int count);
573 const ReportDesc *GetReport() const;
577 // Symbolizer makes lots of intercepted calls. If we try to process them,
578 // at best it will cause deadlocks on internal mutexes.
579 ScopedIgnoreInterceptors ignore_interceptors_;
581 void AddDeadMutex(u64 id);
583 ScopedReport(const ScopedReport&);
584 void operator = (const ScopedReport&);
587 void RestoreStack(int tid, const u64 epoch, VarSizeStackTrace *stk,
590 template<typename StackTraceTy>
591 void ObtainCurrentStack(ThreadState *thr, uptr toppc, StackTraceTy *stack) {
592 uptr size = thr->shadow_stack_pos - thr->shadow_stack;
594 if (size + !!toppc > kStackTraceMax) {
595 start = size + !!toppc - kStackTraceMax;
596 size = kStackTraceMax - !!toppc;
598 stack->Init(&thr->shadow_stack[start], size, toppc);
602 #if TSAN_COLLECT_STATS
603 void StatAggregate(u64 *dst, u64 *src);
604 void StatOutput(u64 *stat);
607 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
608 #if TSAN_COLLECT_STATS
612 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
613 #if TSAN_COLLECT_STATS
618 void MapShadow(uptr addr, uptr size);
619 void MapThreadTrace(uptr addr, uptr size, const char *name);
620 void DontNeedShadowFor(uptr addr, uptr size);
621 void InitializeShadowMemory();
622 void InitializeInterceptors();
623 void InitializeLibIgnore();
624 void InitializeDynamicAnnotations();
626 void ForkBefore(ThreadState *thr, uptr pc);
627 void ForkParentAfter(ThreadState *thr, uptr pc);
628 void ForkChildAfter(ThreadState *thr, uptr pc);
630 void ReportRace(ThreadState *thr);
631 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
632 bool IsFiredSuppression(Context *ctx, ReportType type, StackTrace trace);
633 bool IsExpectedReport(uptr addr, uptr size);
634 void PrintMatchedBenignRaces();
636 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
637 # define DPrintf Printf
639 # define DPrintf(...)
642 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
643 # define DPrintf2 Printf
645 # define DPrintf2(...)
648 u32 CurrentStackId(ThreadState *thr, uptr pc);
649 ReportStack *SymbolizeStackId(u32 stack_id);
650 void PrintCurrentStack(ThreadState *thr, uptr pc);
651 void PrintCurrentStackSlow(uptr pc); // uses libunwind
653 void Initialize(ThreadState *thr);
654 int Finalize(ThreadState *thr);
656 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
657 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
659 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
660 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
661 void MemoryAccessImpl(ThreadState *thr, uptr addr,
662 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
663 u64 *shadow_mem, Shadow cur);
664 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
665 uptr size, bool is_write);
666 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
667 uptr size, uptr step, bool is_write);
668 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
669 int size, bool kAccessIsWrite, bool kIsAtomic);
671 const int kSizeLog1 = 0;
672 const int kSizeLog2 = 1;
673 const int kSizeLog4 = 2;
674 const int kSizeLog8 = 3;
676 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
677 uptr addr, int kAccessSizeLog) {
678 MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
681 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
682 uptr addr, int kAccessSizeLog) {
683 MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
686 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
687 uptr addr, int kAccessSizeLog) {
688 MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
691 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
692 uptr addr, int kAccessSizeLog) {
693 MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
696 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
697 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
698 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
700 void ThreadIgnoreBegin(ThreadState *thr, uptr pc);
701 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
702 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc);
703 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
705 void FuncEntry(ThreadState *thr, uptr pc);
706 void FuncExit(ThreadState *thr);
708 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
709 void ThreadStart(ThreadState *thr, int tid, uptr os_id);
710 void ThreadFinish(ThreadState *thr);
711 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
712 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
713 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
714 void ThreadFinalize(ThreadState *thr);
715 void ThreadSetName(ThreadState *thr, const char *name);
716 int ThreadCount(ThreadState *thr);
717 void ProcessPendingSignals(ThreadState *thr);
719 Processor *ProcCreate();
720 void ProcDestroy(Processor *proc);
721 void ProcWire(Processor *proc, ThreadState *thr);
722 void ProcUnwire(Processor *proc, ThreadState *thr);
724 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
725 bool rw, bool recursive, bool linker_init);
726 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
727 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1,
728 bool try_lock = false);
729 int MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
730 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr, bool try_lock = false);
731 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
732 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
733 void MutexRepair(ThreadState *thr, uptr pc, uptr addr); // call on EOWNERDEAD
734 void MutexInvalidAccess(ThreadState *thr, uptr pc, uptr addr);
736 void Acquire(ThreadState *thr, uptr pc, uptr addr);
737 // AcquireGlobal synchronizes the current thread with all other threads.
738 // In terms of happens-before relation, it draws a HB edge from all threads
739 // (where they happen to execute right now) to the current thread. We use it to
740 // handle Go finalizers. Namely, finalizer goroutine executes AcquireGlobal
741 // right before executing finalizers. This provides a coarse, but simple
742 // approximation of the actual required synchronization.
743 void AcquireGlobal(ThreadState *thr, uptr pc);
744 void Release(ThreadState *thr, uptr pc, uptr addr);
745 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
746 void AfterSleep(ThreadState *thr, uptr pc);
747 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
748 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
749 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
750 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
752 // The hacky call uses custom calling convention and an assembly thunk.
753 // It is considerably faster that a normal call for the caller
754 // if it is not executed (it is intended for slow paths from hot functions).
755 // The trick is that the call preserves all registers and the compiler
756 // does not treat it as a call.
757 // If it does not work for you, use normal call.
758 #if !SANITIZER_DEBUG && defined(__x86_64__) && !SANITIZER_MAC
759 // The caller may not create the stack frame for itself at all,
760 // so we create a reserve stack frame for it (1024b must be enough).
761 #define HACKY_CALL(f) \
762 __asm__ __volatile__("sub $1024, %%rsp;" \
763 CFI_INL_ADJUST_CFA_OFFSET(1024) \
764 ".hidden " #f "_thunk;" \
765 "call " #f "_thunk;" \
766 "add $1024, %%rsp;" \
767 CFI_INL_ADJUST_CFA_OFFSET(-1024) \
770 #define HACKY_CALL(f) f()
773 void TraceSwitch(ThreadState *thr);
774 uptr TraceTopPC(ThreadState *thr);
777 Trace *ThreadTrace(int tid);
779 extern "C" void __tsan_trace_switch();
780 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
781 EventType typ, u64 addr) {
782 if (!kCollectHistory)
784 DCHECK_GE((int)typ, 0);
785 DCHECK_LE((int)typ, 7);
786 DCHECK_EQ(GetLsb(addr, 61), addr);
787 StatInc(thr, StatEvents);
788 u64 pos = fs.GetTracePos();
789 if (UNLIKELY((pos % kTracePartSize) == 0)) {
791 HACKY_CALL(__tsan_trace_switch);
796 Event *trace = (Event*)GetThreadTrace(fs.tid());
797 Event *evp = &trace[pos];
798 Event ev = (u64)addr | ((u64)typ << 61);
803 uptr ALWAYS_INLINE HeapEnd() {
804 return HeapMemEnd() + PrimaryAllocator::AdditionalSize();
808 } // namespace __tsan