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_common.h"
  30 #include "sanitizer_common/sanitizer_suppressions.h"
  31 #include "sanitizer_common/sanitizer_thread_registry.h"
  32 #include "tsan_clock.h"
  33 #include "tsan_defs.h"
  34 #include "tsan_flags.h"
  35 #include "tsan_sync.h"
  36 #include "tsan_trace.h"
  37 #include "tsan_vector.h"
  38 #include "tsan_report.h"
  39 #include "tsan_platform.h"
  40 #include "tsan_mutexset.h"
  41
  42 #if SANITIZER_WORDSIZE != 64
  43 # error "ThreadSanitizer is supported only on 64-bit platforms"
  44 #endif
  45
  46 namespace __tsan {
  47
  48 // Descriptor of user's memory block.
  49 struct MBlock {
  50   /*
  51   u64 mtx : 1;  // must be first
  52   u64 lst : 44;
  53   u64 stk : 31;  // on word boundary
  54   u64 tid : kTidBits;
  55   u64 siz : 128 - 1 - 31 - 44 - kTidBits;  // 39
  56   */
  57   u64 raw[2];
  58
  59   void Init(uptr siz, u32 tid, u32 stk) {
  60     raw[0] = raw[1] = 0;
  61     raw[1] |= (u64)siz << ((1 + 44 + 31 + kTidBits) % 64);
  62     raw[1] |= (u64)tid << ((1 + 44 + 31) % 64);
  63     raw[0] |= (u64)stk << (1 + 44);
  64     raw[1] |= (u64)stk >> (64 - 44 - 1);
  65     DCHECK_EQ(Size(), siz);
  66     DCHECK_EQ(Tid(), tid);
  67     DCHECK_EQ(StackId(), stk);
  68   }
  69
  70   u32 Tid() const {
  71     return GetLsb(raw[1] >> ((1 + 44 + 31) % 64), kTidBits);
  72   }
  73
  74   uptr Size() const {
  75     return raw[1] >> ((1 + 31 + 44 + kTidBits) % 64);
  76   }
  77
  78   u32 StackId() const {
  79     return (raw[0] >> (1 + 44)) | GetLsb(raw[1] << (64 - 44 - 1), 31);
  80   }
  81
  82   SyncVar *ListHead() const {
  83     return (SyncVar*)(GetLsb(raw[0] >> 1, 44) << 3);
  84   }
  85
  86   void ListPush(SyncVar *v) {
  87     SyncVar *lst = ListHead();
  88     v->next = lst;
  89     u64 x = (u64)v ^ (u64)lst;
  90     x = (x >> 3) << 1;
  91     raw[0] ^= x;
  92     DCHECK_EQ(ListHead(), v);
  93   }
  94
  95   SyncVar *ListPop() {
  96     SyncVar *lst = ListHead();
  97     SyncVar *nxt = lst->next;
  98     lst->next = 0;
  99     u64 x = (u64)lst ^ (u64)nxt;
 100     x = (x >> 3) << 1;
 101     raw[0] ^= x;
 102     DCHECK_EQ(ListHead(), nxt);
 103     return lst;
 104   }
 105
 106   void ListReset() {
 107     SyncVar *lst = ListHead();
 108     u64 x = (u64)lst;
 109     x = (x >> 3) << 1;
 110     raw[0] ^= x;
 111     DCHECK_EQ(ListHead(), 0);
 112   }
 113
 114   void Lock();
 115   void Unlock();
 116   typedef GenericScopedLock<MBlock> ScopedLock;
 117 };
 118
 119 #ifndef TSAN_GO
 120 #if defined(TSAN_COMPAT_SHADOW) && TSAN_COMPAT_SHADOW
 121 const uptr kAllocatorSpace = 0x7d0000000000ULL;
 122 #else
 123 const uptr kAllocatorSpace = 0x7d0000000000ULL;
 124 #endif
 125 const uptr kAllocatorSize  =  0x10000000000ULL;  // 1T.
 126
 127 struct MapUnmapCallback;
 128 typedef SizeClassAllocator64<kAllocatorSpace, kAllocatorSize, sizeof(MBlock),
 129     DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
 130 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
 131 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
 132 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
 133     SecondaryAllocator> Allocator;
 134 Allocator *allocator();
 135 #endif
 136
 137 void TsanCheckFailed(const char *file, int line, const char *cond,
 138                      u64 v1, u64 v2);
 139
 140 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
 141
 142 // FastState (from most significant bit):
 143 //   ignore          : 1
 144 //   tid             : kTidBits
 145 //   epoch           : kClkBits
 146 //   unused          : -
 147 //   history_size    : 3
 148 class FastState {
 149  public:
 150   FastState(u64 tid, u64 epoch) {
 151     x_ = tid << kTidShift;
 152     x_ |= epoch << kClkShift;
 153     DCHECK_EQ(tid, this->tid());
 154     DCHECK_EQ(epoch, this->epoch());
 155     DCHECK_EQ(GetIgnoreBit(), false);
 156   }
 157
 158   explicit FastState(u64 x)
 159       : x_(x) {
 160   }
 161
 162   u64 raw() const {
 163     return x_;
 164   }
 165
 166   u64 tid() const {
 167     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
 168     return res;
 169   }
 170
 171   u64 TidWithIgnore() const {
 172     u64 res = x_ >> kTidShift;
 173     return res;
 174   }
 175
 176   u64 epoch() const {
 177     u64 res = (x_ << (kTidBits + 1)) >> (64 - kClkBits);
 178     return res;
 179   }
 180
 181   void IncrementEpoch() {
 182     u64 old_epoch = epoch();
 183     x_ += 1 << kClkShift;
 184     DCHECK_EQ(old_epoch + 1, epoch());
 185     (void)old_epoch;
 186   }
 187
 188   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 189   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 190   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 191
 192   void SetHistorySize(int hs) {
 193     CHECK_GE(hs, 0);
 194     CHECK_LE(hs, 7);
 195     x_ = (x_ & ~7) | hs;
 196   }
 197
 198   int GetHistorySize() const {
 199     return (int)(x_ & 7);
 200   }
 201
 202   void ClearHistorySize() {
 203     x_ &= ~7;
 204   }
 205
 206   u64 GetTracePos() const {
 207     const int hs = GetHistorySize();
 208     // When hs == 0, the trace consists of 2 parts.
 209     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 210     return epoch() & mask;
 211   }
 212
 213  private:
 214   friend class Shadow;
 215   static const int kTidShift = 64 - kTidBits - 1;
 216   static const int kClkShift = kTidShift - kClkBits;
 217   static const u64 kIgnoreBit = 1ull << 63;
 218   static const u64 kFreedBit = 1ull << 63;
 219   u64 x_;
 220 };
 221
 222 // Shadow (from most significant bit):
 223 //   freed           : 1
 224 //   tid             : kTidBits
 225 //   epoch           : kClkBits
 226 //   is_atomic       : 1
 227 //   is_read         : 1
 228 //   size_log        : 2
 229 //   addr0           : 3
 230 class Shadow : public FastState {
 231  public:
 232   explicit Shadow(u64 x)
 233       : FastState(x) {
 234   }
 235
 236   explicit Shadow(const FastState &s)
 237       : FastState(s.x_) {
 238     ClearHistorySize();
 239   }
 240
 241   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 242     DCHECK_EQ(x_ & 31, 0);
 243     DCHECK_LE(addr0, 7);
 244     DCHECK_LE(kAccessSizeLog, 3);
 245     x_ |= (kAccessSizeLog << 3) | addr0;
 246     DCHECK_EQ(kAccessSizeLog, size_log());
 247     DCHECK_EQ(addr0, this->addr0());
 248   }
 249
 250   void SetWrite(unsigned kAccessIsWrite) {
 251     DCHECK_EQ(x_ & kReadBit, 0);
 252     if (!kAccessIsWrite)
 253       x_ |= kReadBit;
 254     DCHECK_EQ(kAccessIsWrite, IsWrite());
 255   }
 256
 257   void SetAtomic(bool kIsAtomic) {
 258     DCHECK(!IsAtomic());
 259     if (kIsAtomic)
 260       x_ |= kAtomicBit;
 261     DCHECK_EQ(IsAtomic(), kIsAtomic);
 262   }
 263
 264   bool IsAtomic() const {
 265     return x_ & kAtomicBit;
 266   }
 267
 268   bool IsZero() const {
 269     return x_ == 0;
 270   }
 271
 272   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 273     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 274     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 275     return shifted_xor == 0;
 276   }
 277
 278   static inline bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 279     u64 masked_xor = (s1.x_ ^ s2.x_) & 31;
 280     return masked_xor == 0;
 281   }
 282
 283   static inline bool TwoRangesIntersect(Shadow s1, Shadow s2,
 284       unsigned kS2AccessSize) {
 285     bool res = false;
 286     u64 diff = s1.addr0() - s2.addr0();
 287     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
 288       // if (s1.addr0() + size1) > s2.addr0()) return true;
 289       if (s1.size() > -diff)  res = true;
 290     } else {
 291       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 292       if (kS2AccessSize > diff) res = true;
 293     }
 294     DCHECK_EQ(res, TwoRangesIntersectSLOW(s1, s2));
 295     DCHECK_EQ(res, TwoRangesIntersectSLOW(s2, s1));
 296     return res;
 297   }
 298
 299   // The idea behind the offset is as follows.
 300   // Consider that we have 8 bool's contained within a single 8-byte block
 301   // (mapped to a single shadow "cell"). Now consider that we write to the bools
 302   // from a single thread (which we consider the common case).
 303   // W/o offsetting each access will have to scan 4 shadow values at average
 304   // to find the corresponding shadow value for the bool.
 305   // With offsetting we start scanning shadow with the offset so that
 306   // each access hits necessary shadow straight off (at least in an expected
 307   // optimistic case).
 308   // This logic works seamlessly for any layout of user data. For example,
 309   // if user data is {int, short, char, char}, then accesses to the int are
 310   // offsetted to 0, short - 4, 1st char - 6, 2nd char - 7. Hopefully, accesses
 311   // from a single thread won't need to scan all 8 shadow values.
 312   unsigned ComputeSearchOffset() {
 313     return x_ & 7;
 314   }
 315   u64 addr0() const { return x_ & 7; }
 316   u64 size() const { return 1ull << size_log(); }
 317   bool IsWrite() const { return !IsRead(); }
 318   bool IsRead() const { return x_ & kReadBit; }
 319
 320   // The idea behind the freed bit is as follows.
 321   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 322   // values with tid/epoch related to the free and the freed bit set.
 323   // During memory accesses processing the freed bit is considered
 324   // as msb of tid. So any access races with shadow with freed bit set
 325   // (it is as if write from a thread with which we never synchronized before).
 326   // This allows us to detect accesses to freed memory w/o additional
 327   // overheads in memory access processing and at the same time restore
 328   // tid/epoch of free.
 329   void MarkAsFreed() {
 330      x_ |= kFreedBit;
 331   }
 332
 333   bool IsFreed() const {
 334     return x_ & kFreedBit;
 335   }
 336
 337   bool GetFreedAndReset() {
 338     bool res = x_ & kFreedBit;
 339     x_ &= ~kFreedBit;
 340     return res;
 341   }
 342
 343   bool IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 344     // analyzes 5-th bit (is_read) and 6-th bit (is_atomic)
 345     bool v = x_ & u64(((kIsWrite ^ 1) << kReadShift)
 346         | (kIsAtomic << kAtomicShift));
 347     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 348     return v;
 349   }
 350
 351   bool IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 352     bool v = ((x_ >> kReadShift) & 3)
 353         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 354     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 355         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 356     return v;
 357   }
 358
 359   bool IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 360     bool v = ((x_ >> kReadShift) & 3)
 361         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 362     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 363         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 364     return v;
 365   }
 366
 367  private:
 368   static const u64 kReadShift   = 5;
 369   static const u64 kReadBit     = 1ull << kReadShift;
 370   static const u64 kAtomicShift = 6;
 371   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 372
 373   u64 size_log() const { return (x_ >> 3) & 3; }
 374
 375   static bool TwoRangesIntersectSLOW(const Shadow s1, const Shadow s2) {
 376     if (s1.addr0() == s2.addr0()) return true;
 377     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 378       return true;
 379     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 380       return true;
 381     return false;
 382   }
 383 };
 384
 385 struct SignalContext;
 386
 387 struct JmpBuf {
 388   uptr sp;
 389   uptr mangled_sp;
 390   uptr *shadow_stack_pos;
 391 };
 392
 393 // This struct is stored in TLS.
 394 struct ThreadState {
 395   FastState fast_state;
 396   // Synch epoch represents the threads's epoch before the last synchronization
 397   // action. It allows to reduce number of shadow state updates.
 398   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 399   // if we are processing write to X from the same thread at epoch=200,
 400   // we do nothing, because both writes happen in the same 'synch epoch'.
 401   // That is, if another memory access does not race with the former write,
 402   // it does not race with the latter as well.
 403   // QUESTION: can we can squeeze this into ThreadState::Fast?
 404   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 405   // taken by epoch between synchs.
 406   // This way we can save one load from tls.
 407   u64 fast_synch_epoch;
 408   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 409   // We do not distinguish beteween ignoring reads and writes
 410   // for better performance.
 411   int ignore_reads_and_writes;
 412   uptr *shadow_stack_pos;
 413   u64 *racy_shadow_addr;
 414   u64 racy_state[2];
 415 #ifndef TSAN_GO
 416   // C/C++ uses embed shadow stack of fixed size.
 417   uptr shadow_stack[kShadowStackSize];
 418 #else
 419   // Go uses satellite shadow stack with dynamic size.
 420   uptr *shadow_stack;
 421   uptr *shadow_stack_end;
 422 #endif
 423   MutexSet mset;
 424   ThreadClock clock;
 425 #ifndef TSAN_GO
 426   AllocatorCache alloc_cache;
 427   InternalAllocatorCache internal_alloc_cache;
 428   Vector<JmpBuf> jmp_bufs;
 429 #endif
 430   u64 stat[StatCnt];
 431   const int tid;
 432   const int unique_id;
 433   int in_rtl;
 434   bool in_symbolizer;
 435   bool is_alive;
 436   bool is_freeing;
 437   bool is_vptr_access;
 438   const uptr stk_addr;
 439   const uptr stk_size;
 440   const uptr tls_addr;
 441   const uptr tls_size;
 442
 443   DeadlockDetector deadlock_detector;
 444
 445   bool in_signal_handler;
 446   SignalContext *signal_ctx;
 447
 448 #ifndef TSAN_GO
 449   u32 last_sleep_stack_id;
 450   ThreadClock last_sleep_clock;
 451 #endif
 452
 453   // Set in regions of runtime that must be signal-safe and fork-safe.
 454   // If set, malloc must not be called.
 455   int nomalloc;
 456
 457   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 458                        uptr stk_addr, uptr stk_size,
 459                        uptr tls_addr, uptr tls_size);
 460 };
 461
 462 Context *CTX();
 463
 464 #ifndef TSAN_GO
 465 extern THREADLOCAL char cur_thread_placeholder[];
 466 INLINE ThreadState *cur_thread() {
 467   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
 468 }
 469 #endif
 470
 471 class ThreadContext : public ThreadContextBase {
 472  public:
 473   explicit ThreadContext(int tid);
 474   ~ThreadContext();
 475   ThreadState *thr;
 476 #ifdef TSAN_GO
 477   StackTrace creation_stack;
 478 #else
 479   u32 creation_stack_id;
 480 #endif
 481   SyncClock sync;
 482   // Epoch at which the thread had started.
 483   // If we see an event from the thread stamped by an older epoch,
 484   // the event is from a dead thread that shared tid with this thread.
 485   u64 epoch0;
 486   u64 epoch1;
 487
 488   // Override superclass callbacks.
 489   void OnDead();
 490   void OnJoined(void *arg);
 491   void OnFinished();
 492   void OnStarted(void *arg);
 493   void OnCreated(void *arg);
 494   void OnReset();
 495 };
 496
 497 struct RacyStacks {
 498   MD5Hash hash[2];
 499   bool operator==(const RacyStacks &other) const {
 500     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 501       return true;
 502     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 503       return true;
 504     return false;
 505   }
 506 };
 507
 508 struct RacyAddress {
 509   uptr addr_min;
 510   uptr addr_max;
 511 };
 512
 513 struct FiredSuppression {
 514   ReportType type;
 515   uptr pc;
 516   Suppression *supp;
 517 };
 518
 519 struct Context {
 520   Context();
 521
 522   bool initialized;
 523
 524   SyncTab synctab;
 525
 526   Mutex report_mtx;
 527   int nreported;
 528   int nmissed_expected;
 529   atomic_uint64_t last_symbolize_time_ns;
 530
 531   ThreadRegistry *thread_registry;
 532
 533   Vector<RacyStacks> racy_stacks;
 534   Vector<RacyAddress> racy_addresses;
 535   // Number of fired suppressions may be large enough.
 536   InternalMmapVector<FiredSuppression> fired_suppressions;
 537
 538   Flags flags;
 539
 540   u64 stat[StatCnt];
 541   u64 int_alloc_cnt[MBlockTypeCount];
 542   u64 int_alloc_siz[MBlockTypeCount];
 543 };
 544
 545 class ScopedInRtl {
 546  public:
 547   ScopedInRtl();
 548   ~ScopedInRtl();
 549  private:
 550   ThreadState*thr_;
 551   int in_rtl_;
 552   int errno_;
 553 };
 554
 555 class ScopedReport {
 556  public:
 557   explicit ScopedReport(ReportType typ);
 558   ~ScopedReport();
 559
 560   void AddStack(const StackTrace *stack);
 561   void AddMemoryAccess(uptr addr, Shadow s, const StackTrace *stack,
 562                        const MutexSet *mset);
 563   void AddThread(const ThreadContext *tctx);
 564   void AddMutex(const SyncVar *s);
 565   void AddLocation(uptr addr, uptr size);
 566   void AddSleep(u32 stack_id);
 567   void SetCount(int count);
 568
 569   const ReportDesc *GetReport() const;
 570
 571  private:
 572   Context *ctx_;
 573   ReportDesc *rep_;
 574
 575   void AddMutex(u64 id);
 576
 577   ScopedReport(const ScopedReport&);
 578   void operator = (const ScopedReport&);
 579 };
 580
 581 void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset);
 582
 583 void StatAggregate(u64 *dst, u64 *src);
 584 void StatOutput(u64 *stat);
 585 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 586   if (kCollectStats)
 587     thr->stat[typ] += n;
 588 }
 589 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 590   if (kCollectStats)
 591     thr->stat[typ] = n;
 592 }
 593
 594 void MapShadow(uptr addr, uptr size);
 595 void MapThreadTrace(uptr addr, uptr size);
 596 void DontNeedShadowFor(uptr addr, uptr size);
 597 void InitializeShadowMemory();
 598 void InitializeInterceptors();
 599 void InitializeDynamicAnnotations();
 600
 601 void ReportRace(ThreadState *thr);
 602 bool OutputReport(Context *ctx,
 603                   const ScopedReport &srep,
 604                   const ReportStack *suppress_stack1 = 0,
 605                   const ReportStack *suppress_stack2 = 0,
 606                   const ReportLocation *suppress_loc = 0);
 607 bool IsFiredSuppression(Context *ctx,
 608                         const ScopedReport &srep,
 609                         const StackTrace &trace);
 610 bool IsExpectedReport(uptr addr, uptr size);
 611 void PrintMatchedBenignRaces();
 612 bool FrameIsInternal(const ReportStack *frame);
 613 ReportStack *SkipTsanInternalFrames(ReportStack *ent);
 614
 615 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 616 # define DPrintf Printf
 617 #else
 618 # define DPrintf(...)
 619 #endif
 620
 621 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 622 # define DPrintf2 Printf
 623 #else
 624 # define DPrintf2(...)
 625 #endif
 626
 627 u32 CurrentStackId(ThreadState *thr, uptr pc);
 628 void PrintCurrentStack(ThreadState *thr, uptr pc);
 629 void PrintCurrentStackSlow();  // uses libunwind
 630
 631 void Initialize(ThreadState *thr);
 632 int Finalize(ThreadState *thr);
 633
 634 SyncVar* GetJavaSync(ThreadState *thr, uptr pc, uptr addr,
 635                      bool write_lock, bool create);
 636 SyncVar* GetAndRemoveJavaSync(ThreadState *thr, uptr pc, uptr addr);
 637
 638 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 639     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 640 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 641     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 642     u64 *shadow_mem, Shadow cur);
 643 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 644     uptr size, bool is_write);
 645 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 646     uptr size, uptr step, bool is_write);
 647 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 648     int size, bool kAccessIsWrite, bool kIsAtomic);
 649
 650 const int kSizeLog1 = 0;
 651 const int kSizeLog2 = 1;
 652 const int kSizeLog4 = 2;
 653 const int kSizeLog8 = 3;
 654
 655 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 656                                      uptr addr, int kAccessSizeLog) {
 657   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 658 }
 659
 660 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 661                                       uptr addr, int kAccessSizeLog) {
 662   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 663 }
 664
 665 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 666                                            uptr addr, int kAccessSizeLog) {
 667   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 668 }
 669
 670 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 671                                             uptr addr, int kAccessSizeLog) {
 672   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 673 }
 674
 675 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 676 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 677 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 678 void ThreadIgnoreBegin(ThreadState *thr);
 679 void ThreadIgnoreEnd(ThreadState *thr);
 680
 681 void FuncEntry(ThreadState *thr, uptr pc);
 682 void FuncExit(ThreadState *thr);
 683
 684 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 685 void ThreadStart(ThreadState *thr, int tid, uptr os_id);
 686 void ThreadFinish(ThreadState *thr);
 687 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
 688 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 689 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 690 void ThreadFinalize(ThreadState *thr);
 691 void ThreadSetName(ThreadState *thr, const char *name);
 692 int ThreadCount(ThreadState *thr);
 693 void ProcessPendingSignals(ThreadState *thr);
 694
 695 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
 696                  bool rw, bool recursive, bool linker_init);
 697 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
 698 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1);
 699 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
 700 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr);
 701 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 702 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 703
 704 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 705 void AcquireGlobal(ThreadState *thr, uptr pc);
 706 void Release(ThreadState *thr, uptr pc, uptr addr);
 707 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 708 void AfterSleep(ThreadState *thr, uptr pc);
 709
 710 // The hacky call uses custom calling convention and an assembly thunk.
 711 // It is considerably faster that a normal call for the caller
 712 // if it is not executed (it is intended for slow paths from hot functions).
 713 // The trick is that the call preserves all registers and the compiler
 714 // does not treat it as a call.
 715 // If it does not work for you, use normal call.
 716 #if TSAN_DEBUG == 0
 717 // The caller may not create the stack frame for itself at all,
 718 // so we create a reserve stack frame for it (1024b must be enough).
 719 #define HACKY_CALL(f) \
 720   __asm__ __volatile__("sub $1024, %%rsp;" \
 721                        "/*.cfi_adjust_cfa_offset 1024;*/" \
 722                        ".hidden " #f "_thunk;" \
 723                        "call " #f "_thunk;" \
 724                        "add $1024, %%rsp;" \
 725                        "/*.cfi_adjust_cfa_offset -1024;*/" \
 726                        ::: "memory", "cc");
 727 #else
 728 #define HACKY_CALL(f) f()
 729 #endif
 730
 731 void TraceSwitch(ThreadState *thr);
 732 uptr TraceTopPC(ThreadState *thr);
 733 uptr TraceSize();
 734 uptr TraceParts();
 735 Trace *ThreadTrace(int tid);
 736
 737 extern "C" void __tsan_trace_switch();
 738 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 739                                         EventType typ, u64 addr) {
 740   DCHECK_GE((int)typ, 0);
 741   DCHECK_LE((int)typ, 7);
 742   DCHECK_EQ(GetLsb(addr, 61), addr);
 743   StatInc(thr, StatEvents);
 744   u64 pos = fs.GetTracePos();
 745   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 746 #ifndef TSAN_GO
 747     HACKY_CALL(__tsan_trace_switch);
 748 #else
 749     TraceSwitch(thr);
 750 #endif
 751   }
 752   Event *trace = (Event*)GetThreadTrace(fs.tid());
 753   Event *evp = &trace[pos];
 754   Event ev = (u64)addr | ((u64)typ << 61);
 755   *evp = ev;
 756 }
 757
 758 }  // namespace __tsan
 759
 760 #endif  // TSAN_RTL_H