Home | History | Annotate | Download | only in rtl
      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_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"
     48 
     49 #if SANITIZER_WORDSIZE != 64
     50 # error "ThreadSanitizer is supported only on 64-bit platforms"
     51 #endif
     52 
     53 namespace __tsan {
     54 
     55 #ifndef SANITIZER_GO
     56 struct MapUnmapCallback;
     57 #ifdef __mips64
     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;
     68 #else
     69 typedef SizeClassAllocator64<kHeapMemBeg, kHeapMemEnd - kHeapMemBeg, 0,
     70     DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
     71 #endif
     72 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
     73 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
     74 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
     75     SecondaryAllocator> Allocator;
     76 Allocator *allocator();
     77 #endif
     78 
     79 void TsanCheckFailed(const char *file, int line, const char *cond,
     80                      u64 v1, u64 v2);
     81 
     82 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
     83 
     84 // FastState (from most significant bit):
     85 //   ignore          : 1
     86 //   tid             : kTidBits
     87 //   unused          : -
     88 //   history_size    : 3
     89 //   epoch           : kClkBits
     90 class FastState {
     91  public:
     92   FastState(u64 tid, u64 epoch) {
     93     x_ = tid << kTidShift;
     94     x_ |= epoch;
     95     DCHECK_EQ(tid, this->tid());
     96     DCHECK_EQ(epoch, this->epoch());
     97     DCHECK_EQ(GetIgnoreBit(), false);
     98   }
     99 
    100   explicit FastState(u64 x)
    101       : x_(x) {
    102   }
    103 
    104   u64 raw() const {
    105     return x_;
    106   }
    107 
    108   u64 tid() const {
    109     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
    110     return res;
    111   }
    112 
    113   u64 TidWithIgnore() const {
    114     u64 res = x_ >> kTidShift;
    115     return res;
    116   }
    117 
    118   u64 epoch() const {
    119     u64 res = x_ & ((1ull << kClkBits) - 1);
    120     return res;
    121   }
    122 
    123   void IncrementEpoch() {
    124     u64 old_epoch = epoch();
    125     x_ += 1;
    126     DCHECK_EQ(old_epoch + 1, epoch());
    127     (void)old_epoch;
    128   }
    129 
    130   void SetIgnoreBit() { x_ |= kIgnoreBit; }
    131   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
    132   bool GetIgnoreBit() const { return (s64)x_ < 0; }
    133 
    134   void SetHistorySize(int hs) {
    135     CHECK_GE(hs, 0);
    136     CHECK_LE(hs, 7);
    137     x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
    138   }
    139 
    140   ALWAYS_INLINE
    141   int GetHistorySize() const {
    142     return (int)((x_ >> kHistoryShift) & kHistoryMask);
    143   }
    144 
    145   void ClearHistorySize() {
    146     SetHistorySize(0);
    147   }
    148 
    149   ALWAYS_INLINE
    150   u64 GetTracePos() const {
    151     const int hs = GetHistorySize();
    152     // When hs == 0, the trace consists of 2 parts.
    153     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
    154     return epoch() & mask;
    155   }
    156 
    157  private:
    158   friend class Shadow;
    159   static const int kTidShift = 64 - kTidBits - 1;
    160   static const u64 kIgnoreBit = 1ull << 63;
    161   static const u64 kFreedBit = 1ull << 63;
    162   static const u64 kHistoryShift = kClkBits;
    163   static const u64 kHistoryMask = 7;
    164   u64 x_;
    165 };
    166 
    167 // Shadow (from most significant bit):
    168 //   freed           : 1
    169 //   tid             : kTidBits
    170 //   is_atomic       : 1
    171 //   is_read         : 1
    172 //   size_log        : 2
    173 //   addr0           : 3
    174 //   epoch           : kClkBits
    175 class Shadow : public FastState {
    176  public:
    177   explicit Shadow(u64 x)
    178       : FastState(x) {
    179   }
    180 
    181   explicit Shadow(const FastState &s)
    182       : FastState(s.x_) {
    183     ClearHistorySize();
    184   }
    185 
    186   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
    187     DCHECK_EQ((x_ >> kClkBits) & 31, 0);
    188     DCHECK_LE(addr0, 7);
    189     DCHECK_LE(kAccessSizeLog, 3);
    190     x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
    191     DCHECK_EQ(kAccessSizeLog, size_log());
    192     DCHECK_EQ(addr0, this->addr0());
    193   }
    194 
    195   void SetWrite(unsigned kAccessIsWrite) {
    196     DCHECK_EQ(x_ & kReadBit, 0);
    197     if (!kAccessIsWrite)
    198       x_ |= kReadBit;
    199     DCHECK_EQ(kAccessIsWrite, IsWrite());
    200   }
    201 
    202   void SetAtomic(bool kIsAtomic) {
    203     DCHECK(!IsAtomic());
    204     if (kIsAtomic)
    205       x_ |= kAtomicBit;
    206     DCHECK_EQ(IsAtomic(), kIsAtomic);
    207   }
    208 
    209   bool IsAtomic() const {
    210     return x_ & kAtomicBit;
    211   }
    212 
    213   bool IsZero() const {
    214     return x_ == 0;
    215   }
    216 
    217   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
    218     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
    219     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
    220     return shifted_xor == 0;
    221   }
    222 
    223   static ALWAYS_INLINE
    224   bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
    225     u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
    226     return masked_xor == 0;
    227   }
    228 
    229   static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
    230       unsigned kS2AccessSize) {
    231     bool res = false;
    232     u64 diff = s1.addr0() - s2.addr0();
    233     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
    234       // if (s1.addr0() + size1) > s2.addr0()) return true;
    235       if (s1.size() > -diff)
    236         res = true;
    237     } else {
    238       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
    239       if (kS2AccessSize > diff)
    240         res = true;
    241     }
    242     DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
    243     DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
    244     return res;
    245   }
    246 
    247   u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
    248   u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
    249   bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
    250   bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
    251 
    252   // The idea behind the freed bit is as follows.
    253   // When the memory is freed (or otherwise unaccessible) we write to the shadow
    254   // values with tid/epoch related to the free and the freed bit set.
    255   // During memory accesses processing the freed bit is considered
    256   // as msb of tid. So any access races with shadow with freed bit set
    257   // (it is as if write from a thread with which we never synchronized before).
    258   // This allows us to detect accesses to freed memory w/o additional
    259   // overheads in memory access processing and at the same time restore
    260   // tid/epoch of free.
    261   void MarkAsFreed() {
    262      x_ |= kFreedBit;
    263   }
    264 
    265   bool IsFreed() const {
    266     return x_ & kFreedBit;
    267   }
    268 
    269   bool GetFreedAndReset() {
    270     bool res = x_ & kFreedBit;
    271     x_ &= ~kFreedBit;
    272     return res;
    273   }
    274 
    275   bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
    276     bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
    277         | (u64(kIsAtomic) << kAtomicShift));
    278     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
    279     return v;
    280   }
    281 
    282   bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
    283     bool v = ((x_ >> kReadShift) & 3)
    284         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
    285     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
    286         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
    287     return v;
    288   }
    289 
    290   bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
    291     bool v = ((x_ >> kReadShift) & 3)
    292         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
    293     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
    294         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
    295     return v;
    296   }
    297 
    298  private:
    299   static const u64 kReadShift   = 5 + kClkBits;
    300   static const u64 kReadBit     = 1ull << kReadShift;
    301   static const u64 kAtomicShift = 6 + kClkBits;
    302   static const u64 kAtomicBit   = 1ull << kAtomicShift;
    303 
    304   u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
    305 
    306   static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
    307     if (s1.addr0() == s2.addr0()) return true;
    308     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
    309       return true;
    310     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
    311       return true;
    312     return false;
    313   }
    314 };
    315 
    316 struct ThreadSignalContext;
    317 
    318 struct JmpBuf {
    319   uptr sp;
    320   uptr mangled_sp;
    321   int int_signal_send;
    322   bool in_blocking_func;
    323   uptr in_signal_handler;
    324   uptr *shadow_stack_pos;
    325 };
    326 
    327 // This struct is stored in TLS.
    328 struct ThreadState {
    329   FastState fast_state;
    330   // Synch epoch represents the threads's epoch before the last synchronization
    331   // action. It allows to reduce number of shadow state updates.
    332   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
    333   // if we are processing write to X from the same thread at epoch=200,
    334   // we do nothing, because both writes happen in the same 'synch epoch'.
    335   // That is, if another memory access does not race with the former write,
    336   // it does not race with the latter as well.
    337   // QUESTION: can we can squeeze this into ThreadState::Fast?
    338   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
    339   // taken by epoch between synchs.
    340   // This way we can save one load from tls.
    341   u64 fast_synch_epoch;
    342   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
    343   // We do not distinguish beteween ignoring reads and writes
    344   // for better performance.
    345   int ignore_reads_and_writes;
    346   int ignore_sync;
    347   // Go does not support ignores.
    348 #ifndef SANITIZER_GO
    349   IgnoreSet mop_ignore_set;
    350   IgnoreSet sync_ignore_set;
    351 #endif
    352   // C/C++ uses fixed size shadow stack embed into Trace.
    353   // Go uses malloc-allocated shadow stack with dynamic size.
    354   uptr *shadow_stack;
    355   uptr *shadow_stack_end;
    356   uptr *shadow_stack_pos;
    357   u64 *racy_shadow_addr;
    358   u64 racy_state[2];
    359   MutexSet mset;
    360   ThreadClock clock;
    361 #ifndef SANITIZER_GO
    362   AllocatorCache alloc_cache;
    363   InternalAllocatorCache internal_alloc_cache;
    364   Vector<JmpBuf> jmp_bufs;
    365   int ignore_interceptors;
    366 #endif
    367 #if TSAN_COLLECT_STATS
    368   u64 stat[StatCnt];
    369 #endif
    370   const int tid;
    371   const int unique_id;
    372   bool in_symbolizer;
    373   bool in_ignored_lib;
    374   bool is_inited;
    375   bool is_dead;
    376   bool is_freeing;
    377   bool is_vptr_access;
    378   const uptr stk_addr;
    379   const uptr stk_size;
    380   const uptr tls_addr;
    381   const uptr tls_size;
    382   ThreadContext *tctx;
    383 
    384 #if SANITIZER_DEBUG && !SANITIZER_GO
    385   InternalDeadlockDetector internal_deadlock_detector;
    386 #endif
    387   DDPhysicalThread *dd_pt;
    388   DDLogicalThread *dd_lt;
    389 
    390   atomic_uintptr_t in_signal_handler;
    391   ThreadSignalContext *signal_ctx;
    392 
    393   DenseSlabAllocCache block_cache;
    394   DenseSlabAllocCache sync_cache;
    395   DenseSlabAllocCache clock_cache;
    396 
    397 #ifndef SANITIZER_GO
    398   u32 last_sleep_stack_id;
    399   ThreadClock last_sleep_clock;
    400 #endif
    401 
    402   // Set in regions of runtime that must be signal-safe and fork-safe.
    403   // If set, malloc must not be called.
    404   int nomalloc;
    405 
    406   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
    407                        unsigned reuse_count,
    408                        uptr stk_addr, uptr stk_size,
    409                        uptr tls_addr, uptr tls_size);
    410 };
    411 
    412 #ifndef SANITIZER_GO
    413 __attribute__((tls_model("initial-exec")))
    414 extern THREADLOCAL char cur_thread_placeholder[];
    415 INLINE ThreadState *cur_thread() {
    416   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
    417 }
    418 #endif
    419 
    420 class ThreadContext : public ThreadContextBase {
    421  public:
    422   explicit ThreadContext(int tid);
    423   ~ThreadContext();
    424   ThreadState *thr;
    425   u32 creation_stack_id;
    426   SyncClock sync;
    427   // Epoch at which the thread had started.
    428   // If we see an event from the thread stamped by an older epoch,
    429   // the event is from a dead thread that shared tid with this thread.
    430   u64 epoch0;
    431   u64 epoch1;
    432 
    433   // Override superclass callbacks.
    434   void OnDead() override;
    435   void OnJoined(void *arg) override;
    436   void OnFinished() override;
    437   void OnStarted(void *arg) override;
    438   void OnCreated(void *arg) override;
    439   void OnReset() override;
    440   void OnDetached(void *arg) override;
    441 };
    442 
    443 struct RacyStacks {
    444   MD5Hash hash[2];
    445   bool operator==(const RacyStacks &other) const {
    446     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
    447       return true;
    448     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
    449       return true;
    450     return false;
    451   }
    452 };
    453 
    454 struct RacyAddress {
    455   uptr addr_min;
    456   uptr addr_max;
    457 };
    458 
    459 struct FiredSuppression {
    460   ReportType type;
    461   uptr pc;
    462   Suppression *supp;
    463 };
    464 
    465 struct Context {
    466   Context();
    467 
    468   bool initialized;
    469   bool after_multithreaded_fork;
    470 
    471   MetaMap metamap;
    472 
    473   Mutex report_mtx;
    474   int nreported;
    475   int nmissed_expected;
    476   atomic_uint64_t last_symbolize_time_ns;
    477 
    478   void *background_thread;
    479   atomic_uint32_t stop_background_thread;
    480 
    481   ThreadRegistry *thread_registry;
    482 
    483   Vector<RacyStacks> racy_stacks;
    484   Vector<RacyAddress> racy_addresses;
    485   // Number of fired suppressions may be large enough.
    486   InternalMmapVector<FiredSuppression> fired_suppressions;
    487   DDetector *dd;
    488 
    489   ClockAlloc clock_alloc;
    490 
    491   Flags flags;
    492 
    493   u64 stat[StatCnt];
    494   u64 int_alloc_cnt[MBlockTypeCount];
    495   u64 int_alloc_siz[MBlockTypeCount];
    496 };
    497 
    498 extern Context *ctx;  // The one and the only global runtime context.
    499 
    500 struct ScopedIgnoreInterceptors {
    501   ScopedIgnoreInterceptors() {
    502 #ifndef SANITIZER_GO
    503     cur_thread()->ignore_interceptors++;
    504 #endif
    505   }
    506 
    507   ~ScopedIgnoreInterceptors() {
    508 #ifndef SANITIZER_GO
    509     cur_thread()->ignore_interceptors--;
    510 #endif
    511   }
    512 };
    513 
    514 class ScopedReport {
    515  public:
    516   explicit ScopedReport(ReportType typ);
    517   ~ScopedReport();
    518 
    519   void AddMemoryAccess(uptr addr, Shadow s, StackTrace stack,
    520                        const MutexSet *mset);
    521   void AddStack(StackTrace stack, bool suppressable = false);
    522   void AddThread(const ThreadContext *tctx, bool suppressable = false);
    523   void AddThread(int unique_tid, bool suppressable = false);
    524   void AddUniqueTid(int unique_tid);
    525   void AddMutex(const SyncVar *s);
    526   u64 AddMutex(u64 id);
    527   void AddLocation(uptr addr, uptr size);
    528   void AddSleep(u32 stack_id);
    529   void SetCount(int count);
    530 
    531   const ReportDesc *GetReport() const;
    532 
    533  private:
    534   ReportDesc *rep_;
    535   // Symbolizer makes lots of intercepted calls. If we try to process them,
    536   // at best it will cause deadlocks on internal mutexes.
    537   ScopedIgnoreInterceptors ignore_interceptors_;
    538 
    539   void AddDeadMutex(u64 id);
    540 
    541   ScopedReport(const ScopedReport&);
    542   void operator = (const ScopedReport&);
    543 };
    544 
    545 void RestoreStack(int tid, const u64 epoch, VarSizeStackTrace *stk,
    546                   MutexSet *mset);
    547 
    548 template<typename StackTraceTy>
    549 void ObtainCurrentStack(ThreadState *thr, uptr toppc, StackTraceTy *stack) {
    550   uptr size = thr->shadow_stack_pos - thr->shadow_stack;
    551   uptr start = 0;
    552   if (size + !!toppc > kStackTraceMax) {
    553     start = size + !!toppc - kStackTraceMax;
    554     size = kStackTraceMax - !!toppc;
    555   }
    556   stack->Init(&thr->shadow_stack[start], size, toppc);
    557 }
    558 
    559 
    560 #if TSAN_COLLECT_STATS
    561 void StatAggregate(u64 *dst, u64 *src);
    562 void StatOutput(u64 *stat);
    563 #endif
    564 
    565 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
    566 #if TSAN_COLLECT_STATS
    567   thr->stat[typ] += n;
    568 #endif
    569 }
    570 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
    571 #if TSAN_COLLECT_STATS
    572   thr->stat[typ] = n;
    573 #endif
    574 }
    575 
    576 void MapShadow(uptr addr, uptr size);
    577 void MapThreadTrace(uptr addr, uptr size);
    578 void DontNeedShadowFor(uptr addr, uptr size);
    579 void InitializeShadowMemory();
    580 void InitializeInterceptors();
    581 void InitializeLibIgnore();
    582 void InitializeDynamicAnnotations();
    583 
    584 void ForkBefore(ThreadState *thr, uptr pc);
    585 void ForkParentAfter(ThreadState *thr, uptr pc);
    586 void ForkChildAfter(ThreadState *thr, uptr pc);
    587 
    588 void ReportRace(ThreadState *thr);
    589 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
    590 bool IsFiredSuppression(Context *ctx, const ScopedReport &srep,
    591                         StackTrace trace);
    592 bool IsExpectedReport(uptr addr, uptr size);
    593 void PrintMatchedBenignRaces();
    594 
    595 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
    596 # define DPrintf Printf
    597 #else
    598 # define DPrintf(...)
    599 #endif
    600 
    601 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
    602 # define DPrintf2 Printf
    603 #else
    604 # define DPrintf2(...)
    605 #endif
    606 
    607 u32 CurrentStackId(ThreadState *thr, uptr pc);
    608 ReportStack *SymbolizeStackId(u32 stack_id);
    609 void PrintCurrentStack(ThreadState *thr, uptr pc);
    610 void PrintCurrentStackSlow(uptr pc);  // uses libunwind
    611 
    612 void Initialize(ThreadState *thr);
    613 int Finalize(ThreadState *thr);
    614 
    615 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
    616 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
    617 
    618 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
    619     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
    620 void MemoryAccessImpl(ThreadState *thr, uptr addr,
    621     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
    622     u64 *shadow_mem, Shadow cur);
    623 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
    624     uptr size, bool is_write);
    625 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
    626     uptr size, uptr step, bool is_write);
    627 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
    628     int size, bool kAccessIsWrite, bool kIsAtomic);
    629 
    630 const int kSizeLog1 = 0;
    631 const int kSizeLog2 = 1;
    632 const int kSizeLog4 = 2;
    633 const int kSizeLog8 = 3;
    634 
    635 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
    636                                      uptr addr, int kAccessSizeLog) {
    637   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
    638 }
    639 
    640 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
    641                                       uptr addr, int kAccessSizeLog) {
    642   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
    643 }
    644 
    645 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
    646                                            uptr addr, int kAccessSizeLog) {
    647   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
    648 }
    649 
    650 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
    651                                             uptr addr, int kAccessSizeLog) {
    652   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
    653 }
    654 
    655 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
    656 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
    657 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
    658 
    659 void ThreadIgnoreBegin(ThreadState *thr, uptr pc);
    660 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
    661 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc);
    662 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
    663 
    664 void FuncEntry(ThreadState *thr, uptr pc);
    665 void FuncExit(ThreadState *thr);
    666 
    667 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
    668 void ThreadStart(ThreadState *thr, int tid, uptr os_id);
    669 void ThreadFinish(ThreadState *thr);
    670 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
    671 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
    672 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
    673 void ThreadFinalize(ThreadState *thr);
    674 void ThreadSetName(ThreadState *thr, const char *name);
    675 int ThreadCount(ThreadState *thr);
    676 void ProcessPendingSignals(ThreadState *thr);
    677 
    678 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
    679                  bool rw, bool recursive, bool linker_init);
    680 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
    681 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1,
    682                bool try_lock = false);
    683 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
    684 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr, bool try_lock = false);
    685 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
    686 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
    687 void MutexRepair(ThreadState *thr, uptr pc, uptr addr);  // call on EOWNERDEAD
    688 
    689 void Acquire(ThreadState *thr, uptr pc, uptr addr);
    690 // AcquireGlobal synchronizes the current thread with all other threads.
    691 // In terms of happens-before relation, it draws a HB edge from all threads
    692 // (where they happen to execute right now) to the current thread. We use it to
    693 // handle Go finalizers. Namely, finalizer goroutine executes AcquireGlobal
    694 // right before executing finalizers. This provides a coarse, but simple
    695 // approximation of the actual required synchronization.
    696 void AcquireGlobal(ThreadState *thr, uptr pc);
    697 void Release(ThreadState *thr, uptr pc, uptr addr);
    698 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
    699 void AfterSleep(ThreadState *thr, uptr pc);
    700 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
    701 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
    702 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
    703 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
    704 
    705 // The hacky call uses custom calling convention and an assembly thunk.
    706 // It is considerably faster that a normal call for the caller
    707 // if it is not executed (it is intended for slow paths from hot functions).
    708 // The trick is that the call preserves all registers and the compiler
    709 // does not treat it as a call.
    710 // If it does not work for you, use normal call.
    711 #if !SANITIZER_DEBUG && defined(__x86_64__)
    712 // The caller may not create the stack frame for itself at all,
    713 // so we create a reserve stack frame for it (1024b must be enough).
    714 #define HACKY_CALL(f) \
    715   __asm__ __volatile__("sub $1024, %%rsp;" \
    716                        CFI_INL_ADJUST_CFA_OFFSET(1024) \
    717                        ".hidden " #f "_thunk;" \
    718                        "call " #f "_thunk;" \
    719                        "add $1024, %%rsp;" \
    720                        CFI_INL_ADJUST_CFA_OFFSET(-1024) \
    721                        ::: "memory", "cc");
    722 #else
    723 #define HACKY_CALL(f) f()
    724 #endif
    725 
    726 void TraceSwitch(ThreadState *thr);
    727 uptr TraceTopPC(ThreadState *thr);
    728 uptr TraceSize();
    729 uptr TraceParts();
    730 Trace *ThreadTrace(int tid);
    731 
    732 extern "C" void __tsan_trace_switch();
    733 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
    734                                         EventType typ, u64 addr) {
    735   if (!kCollectHistory)
    736     return;
    737   DCHECK_GE((int)typ, 0);
    738   DCHECK_LE((int)typ, 7);
    739   DCHECK_EQ(GetLsb(addr, 61), addr);
    740   StatInc(thr, StatEvents);
    741   u64 pos = fs.GetTracePos();
    742   if (UNLIKELY((pos % kTracePartSize) == 0)) {
    743 #ifndef SANITIZER_GO
    744     HACKY_CALL(__tsan_trace_switch);
    745 #else
    746     TraceSwitch(thr);
    747 #endif
    748   }
    749   Event *trace = (Event*)GetThreadTrace(fs.tid());
    750   Event *evp = &trace[pos];
    751   Event ev = (u64)addr | ((u64)typ << 61);
    752   *evp = ev;
    753 }
    754 
    755 #ifndef SANITIZER_GO
    756 uptr ALWAYS_INLINE HeapEnd() {
    757 #if SANITIZER_CAN_USE_ALLOCATOR64
    758   return kHeapMemEnd + PrimaryAllocator::AdditionalSize();
    759 #else
    760   return kHeapMemEnd;
    761 #endif
    762 }
    763 #endif
    764 
    765 }  // namespace __tsan
    766 
    767 #endif  // TSAN_RTL_H
    768