libsanitizer/tsan/tsan_rtl.h

   1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
   2 //
   3 // This file is distributed under the University of Illinois Open Source
   4 // License. See LICENSE.TXT for details.
   5 //
   6 //===----------------------------------------------------------------------===//
   7 //
   8 // This file is a part of ThreadSanitizer (TSan), a race detector.
   9 //
  10 // Main internal TSan header file.
  11 //
  12 // Ground rules:
  13 //   - C++ run-time should not be used (static CTORs, RTTI, exceptions, static
  14 //     function-scope locals)
  15 //   - All functions/classes/etc reside in namespace __tsan, except for those
  16 //     declared in tsan_interface.h.
  17 //   - Platform-specific files should be used instead of ifdefs (*).
  18 //   - No system headers included in header files (*).
  19 //   - Platform specific headres included only into platform-specific files (*).
  20 //
  21 //  (*) Except when inlining is critical for performance.
  22 //===----------------------------------------------------------------------===//
  23
  24 #ifndef TSAN_RTL_H
  25 #define TSAN_RTL_H
  26
  27 #include "sanitizer_common/sanitizer_allocator.h"
  28 #include "sanitizer_common/sanitizer_allocator_internal.h"
  29 #include "sanitizer_common/sanitizer_asm.h"
  30 #include "sanitizer_common/sanitizer_common.h"
  31 #include "sanitizer_common/sanitizer_deadlock_detector_interface.h"
  32 #include "sanitizer_common/sanitizer_libignore.h"
  33 #include "sanitizer_common/sanitizer_suppressions.h"
  34 #include "sanitizer_common/sanitizer_thread_registry.h"
  35 #include "tsan_clock.h"
  36 #include "tsan_defs.h"
  37 #include "tsan_flags.h"
  38 #include "tsan_sync.h"
  39 #include "tsan_trace.h"
  40 #include "tsan_vector.h"
  41 #include "tsan_report.h"
  42 #include "tsan_platform.h"
  43 #include "tsan_mutexset.h"
  44 #include "tsan_ignoreset.h"
  45 #include "tsan_stack_trace.h"
  46
  47 #if SANITIZER_WORDSIZE != 64
  48 # error "ThreadSanitizer is supported only on 64-bit platforms"
  49 #endif
  50
  51 namespace __tsan {
  52
  53 #ifndef TSAN_GO
  54 const uptr kAllocatorSpace = 0x7d0000000000ULL;
  55 const uptr kAllocatorSize  =  0x10000000000ULL;  // 1T.
  56
  57 struct MapUnmapCallback;
  58 typedef SizeClassAllocator64<kAllocatorSpace, kAllocatorSize, 0,
  59     DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
  60 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
  61 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
  62 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
  63     SecondaryAllocator> Allocator;
  64 Allocator *allocator();
  65 #endif
  66
  67 void TsanCheckFailed(const char *file, int line, const char *cond,
  68                      u64 v1, u64 v2);
  69
  70 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
  71
  72 // FastState (from most significant bit):
  73 //   ignore          : 1
  74 //   tid             : kTidBits
  75 //   unused          : -
  76 //   history_size    : 3
  77 //   epoch           : kClkBits
  78 class FastState {
  79  public:
  80   FastState(u64 tid, u64 epoch) {
  81     x_ = tid << kTidShift;
  82     x_ |= epoch;
  83     DCHECK_EQ(tid, this->tid());
  84     DCHECK_EQ(epoch, this->epoch());
  85     DCHECK_EQ(GetIgnoreBit(), false);
  86   }
  87
  88   explicit FastState(u64 x)
  89       : x_(x) {
  90   }
  91
  92   u64 raw() const {
  93     return x_;
  94   }
  95
  96   u64 tid() const {
  97     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
  98     return res;
  99   }
 100
 101   u64 TidWithIgnore() const {
 102     u64 res = x_ >> kTidShift;
 103     return res;
 104   }
 105
 106   u64 epoch() const {
 107     u64 res = x_ & ((1ull << kClkBits) - 1);
 108     return res;
 109   }
 110
 111   void IncrementEpoch() {
 112     u64 old_epoch = epoch();
 113     x_ += 1;
 114     DCHECK_EQ(old_epoch + 1, epoch());
 115     (void)old_epoch;
 116   }
 117
 118   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 119   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 120   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 121
 122   void SetHistorySize(int hs) {
 123     CHECK_GE(hs, 0);
 124     CHECK_LE(hs, 7);
 125     x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
 126   }
 127
 128   ALWAYS_INLINE
 129   int GetHistorySize() const {
 130     return (int)((x_ >> kHistoryShift) & kHistoryMask);
 131   }
 132
 133   void ClearHistorySize() {
 134     SetHistorySize(0);
 135   }
 136
 137   ALWAYS_INLINE
 138   u64 GetTracePos() const {
 139     const int hs = GetHistorySize();
 140     // When hs == 0, the trace consists of 2 parts.
 141     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 142     return epoch() & mask;
 143   }
 144
 145  private:
 146   friend class Shadow;
 147   static const int kTidShift = 64 - kTidBits - 1;
 148   static const u64 kIgnoreBit = 1ull << 63;
 149   static const u64 kFreedBit = 1ull << 63;
 150   static const u64 kHistoryShift = kClkBits;
 151   static const u64 kHistoryMask = 7;
 152   u64 x_;
 153 };
 154
 155 // Shadow (from most significant bit):
 156 //   freed           : 1
 157 //   tid             : kTidBits
 158 //   is_atomic       : 1
 159 //   is_read         : 1
 160 //   size_log        : 2
 161 //   addr0           : 3
 162 //   epoch           : kClkBits
 163 class Shadow : public FastState {
 164  public:
 165   explicit Shadow(u64 x)
 166       : FastState(x) {
 167   }
 168
 169   explicit Shadow(const FastState &s)
 170       : FastState(s.x_) {
 171     ClearHistorySize();
 172   }
 173
 174   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 175     DCHECK_EQ((x_ >> kClkBits) & 31, 0);
 176     DCHECK_LE(addr0, 7);
 177     DCHECK_LE(kAccessSizeLog, 3);
 178     x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
 179     DCHECK_EQ(kAccessSizeLog, size_log());
 180     DCHECK_EQ(addr0, this->addr0());
 181   }
 182
 183   void SetWrite(unsigned kAccessIsWrite) {
 184     DCHECK_EQ(x_ & kReadBit, 0);
 185     if (!kAccessIsWrite)
 186       x_ |= kReadBit;
 187     DCHECK_EQ(kAccessIsWrite, IsWrite());
 188   }
 189
 190   void SetAtomic(bool kIsAtomic) {
 191     DCHECK(!IsAtomic());
 192     if (kIsAtomic)
 193       x_ |= kAtomicBit;
 194     DCHECK_EQ(IsAtomic(), kIsAtomic);
 195   }
 196
 197   bool IsAtomic() const {
 198     return x_ & kAtomicBit;
 199   }
 200
 201   bool IsZero() const {
 202     return x_ == 0;
 203   }
 204
 205   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 206     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 207     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 208     return shifted_xor == 0;
 209   }
 210
 211   static ALWAYS_INLINE
 212   bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 213     u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
 214     return masked_xor == 0;
 215   }
 216
 217   static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
 218       unsigned kS2AccessSize) {
 219     bool res = false;
 220     u64 diff = s1.addr0() - s2.addr0();
 221     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
 222       // if (s1.addr0() + size1) > s2.addr0()) return true;
 223       if (s1.size() > -diff)
 224         res = true;
 225     } else {
 226       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 227       if (kS2AccessSize > diff)
 228         res = true;
 229     }
 230     DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
 231     DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
 232     return res;
 233   }
 234
 235   u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
 236   u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
 237   bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
 238   bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
 239
 240   // The idea behind the freed bit is as follows.
 241   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 242   // values with tid/epoch related to the free and the freed bit set.
 243   // During memory accesses processing the freed bit is considered
 244   // as msb of tid. So any access races with shadow with freed bit set
 245   // (it is as if write from a thread with which we never synchronized before).
 246   // This allows us to detect accesses to freed memory w/o additional
 247   // overheads in memory access processing and at the same time restore
 248   // tid/epoch of free.
 249   void MarkAsFreed() {
 250      x_ |= kFreedBit;
 251   }
 252
 253   bool IsFreed() const {
 254     return x_ & kFreedBit;
 255   }
 256
 257   bool GetFreedAndReset() {
 258     bool res = x_ & kFreedBit;
 259     x_ &= ~kFreedBit;
 260     return res;
 261   }
 262
 263   bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 264     bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
 265         | (u64(kIsAtomic) << kAtomicShift));
 266     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 267     return v;
 268   }
 269
 270   bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 271     bool v = ((x_ >> kReadShift) & 3)
 272         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 273     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 274         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 275     return v;
 276   }
 277
 278   bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 279     bool v = ((x_ >> kReadShift) & 3)
 280         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 281     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 282         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 283     return v;
 284   }
 285
 286  private:
 287   static const u64 kReadShift   = 5 + kClkBits;
 288   static const u64 kReadBit     = 1ull << kReadShift;
 289   static const u64 kAtomicShift = 6 + kClkBits;
 290   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 291
 292   u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
 293
 294   static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
 295     if (s1.addr0() == s2.addr0()) return true;
 296     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 297       return true;
 298     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 299       return true;
 300     return false;
 301   }
 302 };
 303
 304 struct SignalContext;
 305
 306 struct JmpBuf {
 307   uptr sp;
 308   uptr mangled_sp;
 309   int int_signal_send;
 310   bool in_blocking_func;
 311   uptr in_signal_handler;
 312   uptr *shadow_stack_pos;
 313 };
 314
 315 // This struct is stored in TLS.
 316 struct ThreadState {
 317   FastState fast_state;
 318   // Synch epoch represents the threads's epoch before the last synchronization
 319   // action. It allows to reduce number of shadow state updates.
 320   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 321   // if we are processing write to X from the same thread at epoch=200,
 322   // we do nothing, because both writes happen in the same 'synch epoch'.
 323   // That is, if another memory access does not race with the former write,
 324   // it does not race with the latter as well.
 325   // QUESTION: can we can squeeze this into ThreadState::Fast?
 326   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 327   // taken by epoch between synchs.
 328   // This way we can save one load from tls.
 329   u64 fast_synch_epoch;
 330   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 331   // We do not distinguish beteween ignoring reads and writes
 332   // for better performance.
 333   int ignore_reads_and_writes;
 334   int ignore_sync;
 335   // Go does not support ignores.
 336 #ifndef TSAN_GO
 337   IgnoreSet mop_ignore_set;
 338   IgnoreSet sync_ignore_set;
 339 #endif
 340   // C/C++ uses fixed size shadow stack embed into Trace.
 341   // Go uses malloc-allocated shadow stack with dynamic size.
 342   uptr *shadow_stack;
 343   uptr *shadow_stack_end;
 344   uptr *shadow_stack_pos;
 345   u64 *racy_shadow_addr;
 346   u64 racy_state[2];
 347   MutexSet mset;
 348   ThreadClock clock;
 349 #ifndef TSAN_GO
 350   AllocatorCache alloc_cache;
 351   InternalAllocatorCache internal_alloc_cache;
 352   Vector<JmpBuf> jmp_bufs;
 353   int ignore_interceptors;
 354 #endif
 355   u64 stat[StatCnt];
 356   const int tid;
 357   const int unique_id;
 358   bool in_symbolizer;
 359   bool in_ignored_lib;
 360   bool is_dead;
 361   bool is_freeing;
 362   bool is_vptr_access;
 363   const uptr stk_addr;
 364   const uptr stk_size;
 365   const uptr tls_addr;
 366   const uptr tls_size;
 367   ThreadContext *tctx;
 368
 369   InternalDeadlockDetector internal_deadlock_detector;
 370   DDPhysicalThread *dd_pt;
 371   DDLogicalThread *dd_lt;
 372
 373   atomic_uintptr_t in_signal_handler;
 374   SignalContext *signal_ctx;
 375
 376   DenseSlabAllocCache block_cache;
 377   DenseSlabAllocCache sync_cache;
 378   DenseSlabAllocCache clock_cache;
 379
 380 #ifndef TSAN_GO
 381   u32 last_sleep_stack_id;
 382   ThreadClock last_sleep_clock;
 383 #endif
 384
 385   // Set in regions of runtime that must be signal-safe and fork-safe.
 386   // If set, malloc must not be called.
 387   int nomalloc;
 388
 389   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 390                        unsigned reuse_count,
 391                        uptr stk_addr, uptr stk_size,
 392                        uptr tls_addr, uptr tls_size);
 393 };
 394
 395 #ifndef TSAN_GO
 396 __attribute__((tls_model("initial-exec")))
 397 extern THREADLOCAL char cur_thread_placeholder[];
 398 INLINE ThreadState *cur_thread() {
 399   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
 400 }
 401 #endif
 402
 403 class ThreadContext : public ThreadContextBase {
 404  public:
 405   explicit ThreadContext(int tid);
 406   ~ThreadContext();
 407   ThreadState *thr;
 408   u32 creation_stack_id;
 409   SyncClock sync;
 410   // Epoch at which the thread had started.
 411   // If we see an event from the thread stamped by an older epoch,
 412   // the event is from a dead thread that shared tid with this thread.
 413   u64 epoch0;
 414   u64 epoch1;
 415
 416   // Override superclass callbacks.
 417   void OnDead();
 418   void OnJoined(void *arg);
 419   void OnFinished();
 420   void OnStarted(void *arg);
 421   void OnCreated(void *arg);
 422   void OnReset();
 423   void OnDetached(void *arg);
 424 };
 425
 426 struct RacyStacks {
 427   MD5Hash hash[2];
 428   bool operator==(const RacyStacks &other) const {
 429     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 430       return true;
 431     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 432       return true;
 433     return false;
 434   }
 435 };
 436
 437 struct RacyAddress {
 438   uptr addr_min;
 439   uptr addr_max;
 440 };
 441
 442 struct FiredSuppression {
 443   ReportType type;
 444   uptr pc;
 445   Suppression *supp;
 446 };
 447
 448 struct Context {
 449   Context();
 450
 451   bool initialized;
 452   bool after_multithreaded_fork;
 453
 454   MetaMap metamap;
 455
 456   Mutex report_mtx;
 457   int nreported;
 458   int nmissed_expected;
 459   atomic_uint64_t last_symbolize_time_ns;
 460
 461   void *background_thread;
 462   atomic_uint32_t stop_background_thread;
 463
 464   ThreadRegistry *thread_registry;
 465
 466   Vector<RacyStacks> racy_stacks;
 467   Vector<RacyAddress> racy_addresses;
 468   // Number of fired suppressions may be large enough.
 469   InternalMmapVector<FiredSuppression> fired_suppressions;
 470   DDetector *dd;
 471
 472   ClockAlloc clock_alloc;
 473
 474   Flags flags;
 475
 476   u64 stat[StatCnt];
 477   u64 int_alloc_cnt[MBlockTypeCount];
 478   u64 int_alloc_siz[MBlockTypeCount];
 479 };
 480
 481 extern Context *ctx;  // The one and the only global runtime context.
 482
 483 struct ScopedIgnoreInterceptors {
 484   ScopedIgnoreInterceptors() {
 485 #ifndef TSAN_GO
 486     cur_thread()->ignore_interceptors++;
 487 #endif
 488   }
 489
 490   ~ScopedIgnoreInterceptors() {
 491 #ifndef TSAN_GO
 492     cur_thread()->ignore_interceptors--;
 493 #endif
 494   }
 495 };
 496
 497 class ScopedReport {
 498  public:
 499   explicit ScopedReport(ReportType typ);
 500   ~ScopedReport();
 501
 502   void AddMemoryAccess(uptr addr, Shadow s, const StackTrace *stack,
 503                        const MutexSet *mset);
 504   void AddStack(const StackTrace *stack, bool suppressable = false);
 505   void AddThread(const ThreadContext *tctx, bool suppressable = false);
 506   void AddThread(int unique_tid, bool suppressable = false);
 507   void AddUniqueTid(int unique_tid);
 508   void AddMutex(const SyncVar *s);
 509   u64 AddMutex(u64 id);
 510   void AddLocation(uptr addr, uptr size);
 511   void AddSleep(u32 stack_id);
 512   void SetCount(int count);
 513
 514   const ReportDesc *GetReport() const;
 515
 516  private:
 517   ReportDesc *rep_;
 518   // Symbolizer makes lots of intercepted calls. If we try to process them,
 519   // at best it will cause deadlocks on internal mutexes.
 520   ScopedIgnoreInterceptors ignore_interceptors_;
 521
 522   void AddDeadMutex(u64 id);
 523
 524   ScopedReport(const ScopedReport&);
 525   void operator = (const ScopedReport&);
 526 };
 527
 528 void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset);
 529
 530 void StatAggregate(u64 *dst, u64 *src);
 531 void StatOutput(u64 *stat);
 532 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 533   if (kCollectStats)
 534     thr->stat[typ] += n;
 535 }
 536 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 537   if (kCollectStats)
 538     thr->stat[typ] = n;
 539 }
 540
 541 void MapShadow(uptr addr, uptr size);
 542 void MapThreadTrace(uptr addr, uptr size);
 543 void DontNeedShadowFor(uptr addr, uptr size);
 544 void InitializeShadowMemory();
 545 void InitializeInterceptors();
 546 void InitializeLibIgnore();
 547 void InitializeDynamicAnnotations();
 548
 549 void ForkBefore(ThreadState *thr, uptr pc);
 550 void ForkParentAfter(ThreadState *thr, uptr pc);
 551 void ForkChildAfter(ThreadState *thr, uptr pc);
 552
 553 void ReportRace(ThreadState *thr);
 554 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
 555 bool IsFiredSuppression(Context *ctx,
 556                         const ScopedReport &srep,
 557                         const StackTrace &trace);
 558 bool IsExpectedReport(uptr addr, uptr size);
 559 void PrintMatchedBenignRaces();
 560 bool FrameIsInternal(const ReportStack *frame);
 561 ReportStack *SkipTsanInternalFrames(ReportStack *ent);
 562
 563 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 564 # define DPrintf Printf
 565 #else
 566 # define DPrintf(...)
 567 #endif
 568
 569 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 570 # define DPrintf2 Printf
 571 #else
 572 # define DPrintf2(...)
 573 #endif
 574
 575 u32 CurrentStackId(ThreadState *thr, uptr pc);
 576 ReportStack *SymbolizeStackId(u32 stack_id);
 577 void PrintCurrentStack(ThreadState *thr, uptr pc);
 578 void PrintCurrentStackSlow();  // uses libunwind
 579
 580 void Initialize(ThreadState *thr);
 581 int Finalize(ThreadState *thr);
 582
 583 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
 584 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
 585
 586 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 587     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 588 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 589     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 590     u64 *shadow_mem, Shadow cur);
 591 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 592     uptr size, bool is_write);
 593 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 594     uptr size, uptr step, bool is_write);
 595 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 596     int size, bool kAccessIsWrite, bool kIsAtomic);
 597
 598 const int kSizeLog1 = 0;
 599 const int kSizeLog2 = 1;
 600 const int kSizeLog4 = 2;
 601 const int kSizeLog8 = 3;
 602
 603 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 604                                      uptr addr, int kAccessSizeLog) {
 605   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 606 }
 607
 608 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 609                                       uptr addr, int kAccessSizeLog) {
 610   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 611 }
 612
 613 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 614                                            uptr addr, int kAccessSizeLog) {
 615   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 616 }
 617
 618 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 619                                             uptr addr, int kAccessSizeLog) {
 620   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 621 }
 622
 623 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 624 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 625 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 626
 627 void ThreadIgnoreBegin(ThreadState *thr, uptr pc);
 628 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
 629 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc);
 630 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
 631
 632 void FuncEntry(ThreadState *thr, uptr pc);
 633 void FuncExit(ThreadState *thr);
 634
 635 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 636 void ThreadStart(ThreadState *thr, int tid, uptr os_id);
 637 void ThreadFinish(ThreadState *thr);
 638 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
 639 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 640 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 641 void ThreadFinalize(ThreadState *thr);
 642 void ThreadSetName(ThreadState *thr, const char *name);
 643 int ThreadCount(ThreadState *thr);
 644 void ProcessPendingSignals(ThreadState *thr);
 645
 646 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
 647                  bool rw, bool recursive, bool linker_init);
 648 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
 649 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1,
 650                bool try_lock = false);
 651 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
 652 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr, bool try_lock = false);
 653 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 654 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 655 void MutexRepair(ThreadState *thr, uptr pc, uptr addr);  // call on EOWNERDEAD
 656
 657 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 658 void AcquireGlobal(ThreadState *thr, uptr pc);
 659 void Release(ThreadState *thr, uptr pc, uptr addr);
 660 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 661 void AfterSleep(ThreadState *thr, uptr pc);
 662 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 663 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 664 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
 665 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 666
 667 // The hacky call uses custom calling convention and an assembly thunk.
 668 // It is considerably faster that a normal call for the caller
 669 // if it is not executed (it is intended for slow paths from hot functions).
 670 // The trick is that the call preserves all registers and the compiler
 671 // does not treat it as a call.
 672 // If it does not work for you, use normal call.
 673 #if TSAN_DEBUG == 0
 674 // The caller may not create the stack frame for itself at all,
 675 // so we create a reserve stack frame for it (1024b must be enough).
 676 #define HACKY_CALL(f) \
 677   __asm__ __volatile__("sub $1024, %%rsp;" \
 678                        CFI_INL_ADJUST_CFA_OFFSET(1024) \
 679                        ".hidden " #f "_thunk;" \
 680                        "call " #f "_thunk;" \
 681                        "add $1024, %%rsp;" \
 682                        CFI_INL_ADJUST_CFA_OFFSET(-1024) \
 683                        ::: "memory", "cc");
 684 #else
 685 #define HACKY_CALL(f) f()
 686 #endif
 687
 688 void TraceSwitch(ThreadState *thr);
 689 uptr TraceTopPC(ThreadState *thr);
 690 uptr TraceSize();
 691 uptr TraceParts();
 692 Trace *ThreadTrace(int tid);
 693
 694 extern "C" void __tsan_trace_switch();
 695 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 696                                         EventType typ, u64 addr) {
 697   if (!kCollectHistory)
 698     return;
 699   DCHECK_GE((int)typ, 0);
 700   DCHECK_LE((int)typ, 7);
 701   DCHECK_EQ(GetLsb(addr, 61), addr);
 702   StatInc(thr, StatEvents);
 703   u64 pos = fs.GetTracePos();
 704   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 705 #ifndef TSAN_GO
 706     HACKY_CALL(__tsan_trace_switch);
 707 #else
 708     TraceSwitch(thr);
 709 #endif
 710   }
 711   Event *trace = (Event*)GetThreadTrace(fs.tid());
 712   Event *evp = &trace[pos];
 713   Event ev = (u64)addr | ((u64)typ << 61);
 714   *evp = ev;
 715 }
 716
 717 }  // namespace __tsan
 718
 719 #endif  // TSAN_RTL_H