lib/tsan/rtl/tsan_rtl.h

   1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file is a part of ThreadSanitizer (TSan), a race detector.
  11 //
  12 // Main internal TSan header file.
  13 //
  14 // Ground rules:
  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 (*).
  22 //
  23 //  (*) Except when inlining is critical for performance.
  24 //===----------------------------------------------------------------------===//
  25
  26 #ifndef TSAN_RTL_H
  27 #define TSAN_RTL_H
  28
  29 #include "sanitizer_common/sanitizer_allocator.h"
  30 #include "sanitizer_common/sanitizer_allocator_internal.h"
  31 #include "sanitizer_common/sanitizer_common.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
  45 #if SANITIZER_WORDSIZE != 64
  46 # error "ThreadSanitizer is supported only on 64-bit platforms"
  47 #endif
  48
  49 namespace __tsan {
  50
  51 // Descriptor of user's memory block.
  52 struct MBlock {
  53   /*
  54   u64 mtx : 1;  // must be first
  55   u64 lst : 44;
  56   u64 stk : 31;  // on word boundary
  57   u64 tid : kTidBits;
  58   u64 siz : 128 - 1 - 31 - 44 - kTidBits;  // 39
  59   */
  60   u64 raw[2];
  61
  62   void Init(uptr siz, u32 tid, u32 stk) {
  63     raw[0] = raw[1] = 0;
  64     raw[1] |= (u64)siz << ((1 + 44 + 31 + kTidBits) % 64);
  65     raw[1] |= (u64)tid << ((1 + 44 + 31) % 64);
  66     raw[0] |= (u64)stk << (1 + 44);
  67     raw[1] |= (u64)stk >> (64 - 44 - 1);
  68     DCHECK_EQ(Size(), siz);
  69     DCHECK_EQ(Tid(), tid);
  70     DCHECK_EQ(StackId(), stk);
  71   }
  72
  73   u32 Tid() const {
  74     return GetLsb(raw[1] >> ((1 + 44 + 31) % 64), kTidBits);
  75   }
  76
  77   uptr Size() const {
  78     return raw[1] >> ((1 + 31 + 44 + kTidBits) % 64);
  79   }
  80
  81   u32 StackId() const {
  82     return (raw[0] >> (1 + 44)) | GetLsb(raw[1] << (64 - 44 - 1), 31);
  83   }
  84
  85   SyncVar *ListHead() const {
  86     return (SyncVar*)(GetLsb(raw[0] >> 1, 44) << 3);
  87   }
  88
  89   void ListPush(SyncVar *v) {
  90     SyncVar *lst = ListHead();
  91     v->next = lst;
  92     u64 x = (u64)v ^ (u64)lst;
  93     x = (x >> 3) << 1;
  94     raw[0] ^= x;
  95     DCHECK_EQ(ListHead(), v);
  96   }
  97
  98   SyncVar *ListPop() {
  99     SyncVar *lst = ListHead();
 100     SyncVar *nxt = lst->next;
 101     lst->next = 0;
 102     u64 x = (u64)lst ^ (u64)nxt;
 103     x = (x >> 3) << 1;
 104     raw[0] ^= x;
 105     DCHECK_EQ(ListHead(), nxt);
 106     return lst;
 107   }
 108
 109   void ListReset() {
 110     SyncVar *lst = ListHead();
 111     u64 x = (u64)lst;
 112     x = (x >> 3) << 1;
 113     raw[0] ^= x;
 114     DCHECK_EQ(ListHead(), 0);
 115   }
 116
 117   void Lock();
 118   void Unlock();
 119   typedef GenericScopedLock<MBlock> ScopedLock;
 120 };
 121
 122 #ifndef TSAN_GO
 123 #if defined(TSAN_COMPAT_SHADOW) && TSAN_COMPAT_SHADOW
 124 const uptr kAllocatorSpace = 0x7d0000000000ULL;
 125 #else
 126 const uptr kAllocatorSpace = 0x7d0000000000ULL;
 127 #endif
 128 const uptr kAllocatorSize  =  0x10000000000ULL;  // 1T.
 129
 130 struct MapUnmapCallback;
 131 typedef SizeClassAllocator64<kAllocatorSpace, kAllocatorSize, sizeof(MBlock),
 132     DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
 133 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
 134 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
 135 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
 136     SecondaryAllocator> Allocator;
 137 Allocator *allocator();
 138 #endif
 139
 140 void TsanCheckFailed(const char *file, int line, const char *cond,
 141                      u64 v1, u64 v2);
 142
 143 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
 144
 145 // FastState (from most significant bit):
 146 //   ignore          : 1
 147 //   tid             : kTidBits
 148 //   epoch           : kClkBits
 149 //   unused          : -
 150 //   history_size    : 3
 151 class FastState {
 152  public:
 153   FastState(u64 tid, u64 epoch) {
 154     x_ = tid << kTidShift;
 155     x_ |= epoch << kClkShift;
 156     DCHECK_EQ(tid, this->tid());
 157     DCHECK_EQ(epoch, this->epoch());
 158     DCHECK_EQ(GetIgnoreBit(), false);
 159   }
 160
 161   explicit FastState(u64 x)
 162       : x_(x) {
 163   }
 164
 165   u64 raw() const {
 166     return x_;
 167   }
 168
 169   u64 tid() const {
 170     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
 171     return res;
 172   }
 173
 174   u64 TidWithIgnore() const {
 175     u64 res = x_ >> kTidShift;
 176     return res;
 177   }
 178
 179   u64 epoch() const {
 180     u64 res = (x_ << (kTidBits + 1)) >> (64 - kClkBits);
 181     return res;
 182   }
 183
 184   void IncrementEpoch() {
 185     u64 old_epoch = epoch();
 186     x_ += 1 << kClkShift;
 187     DCHECK_EQ(old_epoch + 1, epoch());
 188     (void)old_epoch;
 189   }
 190
 191   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 192   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 193   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 194
 195   void SetHistorySize(int hs) {
 196     CHECK_GE(hs, 0);
 197     CHECK_LE(hs, 7);
 198     x_ = (x_ & ~7) | hs;
 199   }
 200
 201   int GetHistorySize() const {
 202     return (int)(x_ & 7);
 203   }
 204
 205   void ClearHistorySize() {
 206     x_ &= ~7;
 207   }
 208
 209   u64 GetTracePos() const {
 210     const int hs = GetHistorySize();
 211     // When hs == 0, the trace consists of 2 parts.
 212     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 213     return epoch() & mask;
 214   }
 215
 216  private:
 217   friend class Shadow;
 218   static const int kTidShift = 64 - kTidBits - 1;
 219   static const int kClkShift = kTidShift - kClkBits;
 220   static const u64 kIgnoreBit = 1ull << 63;
 221   static const u64 kFreedBit = 1ull << 63;
 222   u64 x_;
 223 };
 224
 225 // Shadow (from most significant bit):
 226 //   freed           : 1
 227 //   tid             : kTidBits
 228 //   epoch           : kClkBits
 229 //   is_atomic       : 1
 230 //   is_read         : 1
 231 //   size_log        : 2
 232 //   addr0           : 3
 233 class Shadow : public FastState {
 234  public:
 235   explicit Shadow(u64 x)
 236       : FastState(x) {
 237   }
 238
 239   explicit Shadow(const FastState &s)
 240       : FastState(s.x_) {
 241     ClearHistorySize();
 242   }
 243
 244   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 245     DCHECK_EQ(x_ & 31, 0);
 246     DCHECK_LE(addr0, 7);
 247     DCHECK_LE(kAccessSizeLog, 3);
 248     x_ |= (kAccessSizeLog << 3) | addr0;
 249     DCHECK_EQ(kAccessSizeLog, size_log());
 250     DCHECK_EQ(addr0, this->addr0());
 251   }
 252
 253   void SetWrite(unsigned kAccessIsWrite) {
 254     DCHECK_EQ(x_ & kReadBit, 0);
 255     if (!kAccessIsWrite)
 256       x_ |= kReadBit;
 257     DCHECK_EQ(kAccessIsWrite, IsWrite());
 258   }
 259
 260   void SetAtomic(bool kIsAtomic) {
 261     DCHECK(!IsAtomic());
 262     if (kIsAtomic)
 263       x_ |= kAtomicBit;
 264     DCHECK_EQ(IsAtomic(), kIsAtomic);
 265   }
 266
 267   bool IsAtomic() const {
 268     return x_ & kAtomicBit;
 269   }
 270
 271   bool IsZero() const {
 272     return x_ == 0;
 273   }
 274
 275   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 276     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 277     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 278     return shifted_xor == 0;
 279   }
 280
 281   static inline bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 282     u64 masked_xor = (s1.x_ ^ s2.x_) & 31;
 283     return masked_xor == 0;
 284   }
 285
 286   static inline bool TwoRangesIntersect(Shadow s1, Shadow s2,
 287       unsigned kS2AccessSize) {
 288     bool res = false;
 289     u64 diff = s1.addr0() - s2.addr0();
 290     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
 291       // if (s1.addr0() + size1) > s2.addr0()) return true;
 292       if (s1.size() > -diff)  res = true;
 293     } else {
 294       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 295       if (kS2AccessSize > diff) res = true;
 296     }
 297     DCHECK_EQ(res, TwoRangesIntersectSLOW(s1, s2));
 298     DCHECK_EQ(res, TwoRangesIntersectSLOW(s2, s1));
 299     return res;
 300   }
 301
 302   // The idea behind the offset is as follows.
 303   // Consider that we have 8 bool's contained within a single 8-byte block
 304   // (mapped to a single shadow "cell"). Now consider that we write to the bools
 305   // from a single thread (which we consider the common case).
 306   // W/o offsetting each access will have to scan 4 shadow values at average
 307   // to find the corresponding shadow value for the bool.
 308   // With offsetting we start scanning shadow with the offset so that
 309   // each access hits necessary shadow straight off (at least in an expected
 310   // optimistic case).
 311   // This logic works seamlessly for any layout of user data. For example,
 312   // if user data is {int, short, char, char}, then accesses to the int are
 313   // offsetted to 0, short - 4, 1st char - 6, 2nd char - 7. Hopefully, accesses
 314   // from a single thread won't need to scan all 8 shadow values.
 315   unsigned ComputeSearchOffset() {
 316     return x_ & 7;
 317   }
 318   u64 addr0() const { return x_ & 7; }
 319   u64 size() const { return 1ull << size_log(); }
 320   bool IsWrite() const { return !IsRead(); }
 321   bool IsRead() const { return x_ & kReadBit; }
 322
 323   // The idea behind the freed bit is as follows.
 324   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 325   // values with tid/epoch related to the free and the freed bit set.
 326   // During memory accesses processing the freed bit is considered
 327   // as msb of tid. So any access races with shadow with freed bit set
 328   // (it is as if write from a thread with which we never synchronized before).
 329   // This allows us to detect accesses to freed memory w/o additional
 330   // overheads in memory access processing and at the same time restore
 331   // tid/epoch of free.
 332   void MarkAsFreed() {
 333      x_ |= kFreedBit;
 334   }
 335
 336   bool IsFreed() const {
 337     return x_ & kFreedBit;
 338   }
 339
 340   bool GetFreedAndReset() {
 341     bool res = x_ & kFreedBit;
 342     x_ &= ~kFreedBit;
 343     return res;
 344   }
 345
 346   bool IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 347     // analyzes 5-th bit (is_read) and 6-th bit (is_atomic)
 348     bool v = x_ & u64(((kIsWrite ^ 1) << kReadShift)
 349         | (kIsAtomic << kAtomicShift));
 350     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 351     return v;
 352   }
 353
 354   bool IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 355     bool v = ((x_ >> kReadShift) & 3)
 356         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 357     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 358         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 359     return v;
 360   }
 361
 362   bool IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 363     bool v = ((x_ >> kReadShift) & 3)
 364         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 365     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 366         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 367     return v;
 368   }
 369
 370  private:
 371   static const u64 kReadShift   = 5;
 372   static const u64 kReadBit     = 1ull << kReadShift;
 373   static const u64 kAtomicShift = 6;
 374   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 375
 376   u64 size_log() const { return (x_ >> 3) & 3; }
 377
 378   static bool TwoRangesIntersectSLOW(const Shadow s1, const Shadow s2) {
 379     if (s1.addr0() == s2.addr0()) return true;
 380     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 381       return true;
 382     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 383       return true;
 384     return false;
 385   }
 386 };
 387
 388 struct SignalContext;
 389
 390 struct JmpBuf {
 391   uptr sp;
 392   uptr mangled_sp;
 393   uptr *shadow_stack_pos;
 394 };
 395
 396 // This struct is stored in TLS.
 397 struct ThreadState {
 398   FastState fast_state;
 399   // Synch epoch represents the threads's epoch before the last synchronization
 400   // action. It allows to reduce number of shadow state updates.
 401   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 402   // if we are processing write to X from the same thread at epoch=200,
 403   // we do nothing, because both writes happen in the same 'synch epoch'.
 404   // That is, if another memory access does not race with the former write,
 405   // it does not race with the latter as well.
 406   // QUESTION: can we can squeeze this into ThreadState::Fast?
 407   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 408   // taken by epoch between synchs.
 409   // This way we can save one load from tls.
 410   u64 fast_synch_epoch;
 411   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 412   // We do not distinguish beteween ignoring reads and writes
 413   // for better performance.
 414   int ignore_reads_and_writes;
 415   int ignore_sync;
 416   // C/C++ uses fixed size shadow stack embed into Trace.
 417   // Go uses malloc-allocated shadow stack with dynamic size.
 418   uptr *shadow_stack;
 419   uptr *shadow_stack_end;
 420   uptr *shadow_stack_pos;
 421   u64 *racy_shadow_addr;
 422   u64 racy_state[2];
 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 in_ignored_lib;
 436   bool is_alive;
 437   bool is_freeing;
 438   bool is_vptr_access;
 439   const uptr stk_addr;
 440   const uptr stk_size;
 441   const uptr tls_addr;
 442   const uptr tls_size;
 443
 444   DeadlockDetector deadlock_detector;
 445
 446   bool in_signal_handler;
 447   SignalContext *signal_ctx;
 448
 449 #ifndef TSAN_GO
 450   u32 last_sleep_stack_id;
 451   ThreadClock last_sleep_clock;
 452 #endif
 453
 454   // Set in regions of runtime that must be signal-safe and fork-safe.
 455   // If set, malloc must not be called.
 456   int nomalloc;
 457
 458   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 459                        uptr stk_addr, uptr stk_size,
 460                        uptr tls_addr, uptr tls_size);
 461 };
 462
 463 Context *CTX();
 464
 465 #ifndef TSAN_GO
 466 extern THREADLOCAL char cur_thread_placeholder[];
 467 INLINE ThreadState *cur_thread() {
 468   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
 469 }
 470 #endif
 471
 472 class ThreadContext : public ThreadContextBase {
 473  public:
 474   explicit ThreadContext(int tid);
 475   ~ThreadContext();
 476   ThreadState *thr;
 477 #ifdef TSAN_GO
 478   StackTrace creation_stack;
 479 #else
 480   u32 creation_stack_id;
 481 #endif
 482   SyncClock sync;
 483   // Epoch at which the thread had started.
 484   // If we see an event from the thread stamped by an older epoch,
 485   // the event is from a dead thread that shared tid with this thread.
 486   u64 epoch0;
 487   u64 epoch1;
 488
 489   // Override superclass callbacks.
 490   void OnDead();
 491   void OnJoined(void *arg);
 492   void OnFinished();
 493   void OnStarted(void *arg);
 494   void OnCreated(void *arg);
 495   void OnReset();
 496 };
 497
 498 struct RacyStacks {
 499   MD5Hash hash[2];
 500   bool operator==(const RacyStacks &other) const {
 501     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 502       return true;
 503     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 504       return true;
 505     return false;
 506   }
 507 };
 508
 509 struct RacyAddress {
 510   uptr addr_min;
 511   uptr addr_max;
 512 };
 513
 514 struct FiredSuppression {
 515   ReportType type;
 516   uptr pc;
 517   Suppression *supp;
 518 };
 519
 520 struct Context {
 521   Context();
 522
 523   bool initialized;
 524
 525   SyncTab synctab;
 526
 527   Mutex report_mtx;
 528   int nreported;
 529   int nmissed_expected;
 530   atomic_uint64_t last_symbolize_time_ns;
 531
 532   ThreadRegistry *thread_registry;
 533
 534   Vector<RacyStacks> racy_stacks;
 535   Vector<RacyAddress> racy_addresses;
 536   // Number of fired suppressions may be large enough.
 537   InternalMmapVector<FiredSuppression> fired_suppressions;
 538
 539   Flags flags;
 540
 541   u64 stat[StatCnt];
 542   u64 int_alloc_cnt[MBlockTypeCount];
 543   u64 int_alloc_siz[MBlockTypeCount];
 544 };
 545
 546 class ScopedInRtl {
 547  public:
 548   ScopedInRtl();
 549   ~ScopedInRtl();
 550  private:
 551   ThreadState*thr_;
 552   int in_rtl_;
 553   int errno_;
 554 };
 555
 556 class ScopedReport {
 557  public:
 558   explicit ScopedReport(ReportType typ);
 559   ~ScopedReport();
 560
 561   void AddStack(const StackTrace *stack);
 562   void AddMemoryAccess(uptr addr, Shadow s, const StackTrace *stack,
 563                        const MutexSet *mset);
 564   void AddThread(const ThreadContext *tctx);
 565   void AddMutex(const SyncVar *s);
 566   void AddLocation(uptr addr, uptr size);
 567   void AddSleep(u32 stack_id);
 568   void SetCount(int count);
 569
 570   const ReportDesc *GetReport() const;
 571
 572  private:
 573   Context *ctx_;
 574   ReportDesc *rep_;
 575
 576   void AddMutex(u64 id);
 577
 578   ScopedReport(const ScopedReport&);
 579   void operator = (const ScopedReport&);
 580 };
 581
 582 void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset);
 583
 584 void StatAggregate(u64 *dst, u64 *src);
 585 void StatOutput(u64 *stat);
 586 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 587   if (kCollectStats)
 588     thr->stat[typ] += n;
 589 }
 590 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 591   if (kCollectStats)
 592     thr->stat[typ] = n;
 593 }
 594
 595 void MapShadow(uptr addr, uptr size);
 596 void MapThreadTrace(uptr addr, uptr size);
 597 void DontNeedShadowFor(uptr addr, uptr size);
 598 void InitializeShadowMemory();
 599 void InitializeInterceptors();
 600 void InitializeLibIgnore();
 601 void InitializeDynamicAnnotations();
 602
 603 void ReportRace(ThreadState *thr);
 604 bool OutputReport(Context *ctx,
 605                   const ScopedReport &srep,
 606                   const ReportStack *suppress_stack1 = 0,
 607                   const ReportStack *suppress_stack2 = 0,
 608                   const ReportLocation *suppress_loc = 0);
 609 bool IsFiredSuppression(Context *ctx,
 610                         const ScopedReport &srep,
 611                         const StackTrace &trace);
 612 bool IsExpectedReport(uptr addr, uptr size);
 613 void PrintMatchedBenignRaces();
 614 bool FrameIsInternal(const ReportStack *frame);
 615 ReportStack *SkipTsanInternalFrames(ReportStack *ent);
 616
 617 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 618 # define DPrintf Printf
 619 #else
 620 # define DPrintf(...)
 621 #endif
 622
 623 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 624 # define DPrintf2 Printf
 625 #else
 626 # define DPrintf2(...)
 627 #endif
 628
 629 u32 CurrentStackId(ThreadState *thr, uptr pc);
 630 void PrintCurrentStack(ThreadState *thr, uptr pc);
 631 void PrintCurrentStackSlow();  // uses libunwind
 632
 633 void Initialize(ThreadState *thr);
 634 int Finalize(ThreadState *thr);
 635
 636 SyncVar* GetJavaSync(ThreadState *thr, uptr pc, uptr addr,
 637                      bool write_lock, bool create);
 638 SyncVar* GetAndRemoveJavaSync(ThreadState *thr, uptr pc, uptr addr);
 639
 640 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 641     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 642 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 643     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 644     u64 *shadow_mem, Shadow cur);
 645 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 646     uptr size, bool is_write);
 647 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 648     uptr size, uptr step, bool is_write);
 649 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 650     int size, bool kAccessIsWrite, bool kIsAtomic);
 651
 652 const int kSizeLog1 = 0;
 653 const int kSizeLog2 = 1;
 654 const int kSizeLog4 = 2;
 655 const int kSizeLog8 = 3;
 656
 657 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 658                                      uptr addr, int kAccessSizeLog) {
 659   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 660 }
 661
 662 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 663                                       uptr addr, int kAccessSizeLog) {
 664   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 665 }
 666
 667 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 668                                            uptr addr, int kAccessSizeLog) {
 669   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 670 }
 671
 672 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 673                                             uptr addr, int kAccessSizeLog) {
 674   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 675 }
 676
 677 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 678 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 679 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 680
 681 void ThreadIgnoreBegin(ThreadState *thr);
 682 void ThreadIgnoreEnd(ThreadState *thr);
 683 void ThreadIgnoreSyncBegin(ThreadState *thr);
 684 void ThreadIgnoreSyncEnd(ThreadState *thr);
 685
 686 void FuncEntry(ThreadState *thr, uptr pc);
 687 void FuncExit(ThreadState *thr);
 688
 689 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 690 void ThreadStart(ThreadState *thr, int tid, uptr os_id);
 691 void ThreadFinish(ThreadState *thr);
 692 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
 693 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 694 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 695 void ThreadFinalize(ThreadState *thr);
 696 void ThreadSetName(ThreadState *thr, const char *name);
 697 int ThreadCount(ThreadState *thr);
 698 void ProcessPendingSignals(ThreadState *thr);
 699
 700 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
 701                  bool rw, bool recursive, bool linker_init);
 702 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
 703 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1);
 704 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
 705 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr);
 706 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 707 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 708
 709 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 710 void AcquireGlobal(ThreadState *thr, uptr pc);
 711 void Release(ThreadState *thr, uptr pc, uptr addr);
 712 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 713 void AfterSleep(ThreadState *thr, uptr pc);
 714 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 715 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 716 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
 717 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 718
 719 // The hacky call uses custom calling convention and an assembly thunk.
 720 // It is considerably faster that a normal call for the caller
 721 // if it is not executed (it is intended for slow paths from hot functions).
 722 // The trick is that the call preserves all registers and the compiler
 723 // does not treat it as a call.
 724 // If it does not work for you, use normal call.
 725 #if TSAN_DEBUG == 0
 726 // The caller may not create the stack frame for itself at all,
 727 // so we create a reserve stack frame for it (1024b must be enough).
 728 #define HACKY_CALL(f) \
 729   __asm__ __volatile__("sub $1024, %%rsp;" \
 730                        ".cfi_adjust_cfa_offset 1024;" \
 731                        ".hidden " #f "_thunk;" \
 732                        "call " #f "_thunk;" \
 733                        "add $1024, %%rsp;" \
 734                        ".cfi_adjust_cfa_offset -1024;" \
 735                        ::: "memory", "cc");
 736 #else
 737 #define HACKY_CALL(f) f()
 738 #endif
 739
 740 void TraceSwitch(ThreadState *thr);
 741 uptr TraceTopPC(ThreadState *thr);
 742 uptr TraceSize();
 743 uptr TraceParts();
 744 Trace *ThreadTrace(int tid);
 745
 746 extern "C" void __tsan_trace_switch();
 747 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 748                                         EventType typ, u64 addr) {
 749   DCHECK_GE((int)typ, 0);
 750   DCHECK_LE((int)typ, 7);
 751   DCHECK_EQ(GetLsb(addr, 61), addr);
 752   StatInc(thr, StatEvents);
 753   u64 pos = fs.GetTracePos();
 754   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 755 #ifndef TSAN_GO
 756     HACKY_CALL(__tsan_trace_switch);
 757 #else
 758     TraceSwitch(thr);
 759 #endif
 760   }
 761   Event *trace = (Event*)GetThreadTrace(fs.tid());
 762   Event *evp = &trace[pos];
 763   Event ev = (u64)addr | ((u64)typ << 61);
 764   *evp = ev;
 765 }
 766
 767 }  // namespace __tsan
 768
 769 #endif  // TSAN_RTL_H