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 "sanitizer_common/sanitizer_vector.h"
  36 #include "tsan_clock.h"
  37 #include "tsan_defs.h"
  38 #include "tsan_flags.h"
  39 #include "tsan_mman.h"
  40 #include "tsan_sync.h"
  41 #include "tsan_trace.h"
  42 #include "tsan_report.h"
  43 #include "tsan_platform.h"
  44 #include "tsan_mutexset.h"
  45 #include "tsan_ignoreset.h"
  46 #include "tsan_stack_trace.h"
  47
  48 #if SANITIZER_WORDSIZE != 64
  49 # error "ThreadSanitizer is supported only on 64-bit platforms"
  50 #endif
  51
  52 namespace __tsan {
  53
  54 #if !SANITIZER_GO
  55 struct MapUnmapCallback;
  56 #if defined(__mips64) || defined(__aarch64__) || defined(__powerpc__)
  57 static const uptr kAllocatorRegionSizeLog = 20;
  58 static const uptr kAllocatorNumRegions =
  59     SANITIZER_MMAP_RANGE_SIZE >> kAllocatorRegionSizeLog;
  60 typedef TwoLevelByteMap<(kAllocatorNumRegions >> 12), 1 << 12,
  61     MapUnmapCallback> ByteMap;
  62 struct AP32 {
  63   static const uptr kSpaceBeg = 0;
  64   static const u64 kSpaceSize = SANITIZER_MMAP_RANGE_SIZE;
  65   static const uptr kMetadataSize = 0;
  66   typedef __sanitizer::CompactSizeClassMap SizeClassMap;
  67   static const uptr kRegionSizeLog = kAllocatorRegionSizeLog;
  68   typedef __tsan::ByteMap ByteMap;
  69   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  70   static const uptr kFlags = 0;
  71 };
  72 typedef SizeClassAllocator32<AP32> PrimaryAllocator;
  73 #else
  74 struct AP64 {  // Allocator64 parameters. Deliberately using a short name.
  75   static const uptr kSpaceBeg = Mapping::kHeapMemBeg;
  76   static const uptr kSpaceSize = Mapping::kHeapMemEnd - Mapping::kHeapMemBeg;
  77   static const uptr kMetadataSize = 0;
  78   typedef DefaultSizeClassMap SizeClassMap;
  79   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  80   static const uptr kFlags = 0;
  81 };
  82 typedef SizeClassAllocator64<AP64> PrimaryAllocator;
  83 #endif
  84 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
  85 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
  86 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
  87     SecondaryAllocator> Allocator;
  88 Allocator *allocator();
  89 #endif
  90
  91 void TsanCheckFailed(const char *file, int line, const char *cond,
  92                      u64 v1, u64 v2);
  93
  94 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
  95
  96 // FastState (from most significant bit):
  97 //   ignore          : 1
  98 //   tid             : kTidBits
  99 //   unused          : -
 100 //   history_size    : 3
 101 //   epoch           : kClkBits
 102 class FastState {
 103  public:
 104   FastState(u64 tid, u64 epoch) {
 105     x_ = tid << kTidShift;
 106     x_ |= epoch;
 107     DCHECK_EQ(tid, this->tid());
 108     DCHECK_EQ(epoch, this->epoch());
 109     DCHECK_EQ(GetIgnoreBit(), false);
 110   }
 111
 112   explicit FastState(u64 x)
 113       : x_(x) {
 114   }
 115
 116   u64 raw() const {
 117     return x_;
 118   }
 119
 120   u64 tid() const {
 121     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
 122     return res;
 123   }
 124
 125   u64 TidWithIgnore() const {
 126     u64 res = x_ >> kTidShift;
 127     return res;
 128   }
 129
 130   u64 epoch() const {
 131     u64 res = x_ & ((1ull << kClkBits) - 1);
 132     return res;
 133   }
 134
 135   void IncrementEpoch() {
 136     u64 old_epoch = epoch();
 137     x_ += 1;
 138     DCHECK_EQ(old_epoch + 1, epoch());
 139     (void)old_epoch;
 140   }
 141
 142   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 143   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 144   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 145
 146   void SetHistorySize(int hs) {
 147     CHECK_GE(hs, 0);
 148     CHECK_LE(hs, 7);
 149     x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
 150   }
 151
 152   ALWAYS_INLINE
 153   int GetHistorySize() const {
 154     return (int)((x_ >> kHistoryShift) & kHistoryMask);
 155   }
 156
 157   void ClearHistorySize() {
 158     SetHistorySize(0);
 159   }
 160
 161   ALWAYS_INLINE
 162   u64 GetTracePos() const {
 163     const int hs = GetHistorySize();
 164     // When hs == 0, the trace consists of 2 parts.
 165     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 166     return epoch() & mask;
 167   }
 168
 169  private:
 170   friend class Shadow;
 171   static const int kTidShift = 64 - kTidBits - 1;
 172   static const u64 kIgnoreBit = 1ull << 63;
 173   static const u64 kFreedBit = 1ull << 63;
 174   static const u64 kHistoryShift = kClkBits;
 175   static const u64 kHistoryMask = 7;
 176   u64 x_;
 177 };
 178
 179 // Shadow (from most significant bit):
 180 //   freed           : 1
 181 //   tid             : kTidBits
 182 //   is_atomic       : 1
 183 //   is_read         : 1
 184 //   size_log        : 2
 185 //   addr0           : 3
 186 //   epoch           : kClkBits
 187 class Shadow : public FastState {
 188  public:
 189   explicit Shadow(u64 x)
 190       : FastState(x) {
 191   }
 192
 193   explicit Shadow(const FastState &s)
 194       : FastState(s.x_) {
 195     ClearHistorySize();
 196   }
 197
 198   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 199     DCHECK_EQ((x_ >> kClkBits) & 31, 0);
 200     DCHECK_LE(addr0, 7);
 201     DCHECK_LE(kAccessSizeLog, 3);
 202     x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
 203     DCHECK_EQ(kAccessSizeLog, size_log());
 204     DCHECK_EQ(addr0, this->addr0());
 205   }
 206
 207   void SetWrite(unsigned kAccessIsWrite) {
 208     DCHECK_EQ(x_ & kReadBit, 0);
 209     if (!kAccessIsWrite)
 210       x_ |= kReadBit;
 211     DCHECK_EQ(kAccessIsWrite, IsWrite());
 212   }
 213
 214   void SetAtomic(bool kIsAtomic) {
 215     DCHECK(!IsAtomic());
 216     if (kIsAtomic)
 217       x_ |= kAtomicBit;
 218     DCHECK_EQ(IsAtomic(), kIsAtomic);
 219   }
 220
 221   bool IsAtomic() const {
 222     return x_ & kAtomicBit;
 223   }
 224
 225   bool IsZero() const {
 226     return x_ == 0;
 227   }
 228
 229   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 230     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 231     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 232     return shifted_xor == 0;
 233   }
 234
 235   static ALWAYS_INLINE
 236   bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 237     u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
 238     return masked_xor == 0;
 239   }
 240
 241   static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
 242       unsigned kS2AccessSize) {
 243     bool res = false;
 244     u64 diff = s1.addr0() - s2.addr0();
 245     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
 246       // if (s1.addr0() + size1) > s2.addr0()) return true;
 247       if (s1.size() > -diff)
 248         res = true;
 249     } else {
 250       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 251       if (kS2AccessSize > diff)
 252         res = true;
 253     }
 254     DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
 255     DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
 256     return res;
 257   }
 258
 259   u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
 260   u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
 261   bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
 262   bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
 263
 264   // The idea behind the freed bit is as follows.
 265   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 266   // values with tid/epoch related to the free and the freed bit set.
 267   // During memory accesses processing the freed bit is considered
 268   // as msb of tid. So any access races with shadow with freed bit set
 269   // (it is as if write from a thread with which we never synchronized before).
 270   // This allows us to detect accesses to freed memory w/o additional
 271   // overheads in memory access processing and at the same time restore
 272   // tid/epoch of free.
 273   void MarkAsFreed() {
 274      x_ |= kFreedBit;
 275   }
 276
 277   bool IsFreed() const {
 278     return x_ & kFreedBit;
 279   }
 280
 281   bool GetFreedAndReset() {
 282     bool res = x_ & kFreedBit;
 283     x_ &= ~kFreedBit;
 284     return res;
 285   }
 286
 287   bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 288     bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
 289         | (u64(kIsAtomic) << kAtomicShift));
 290     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 291     return v;
 292   }
 293
 294   bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 295     bool v = ((x_ >> kReadShift) & 3)
 296         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 297     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 298         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 299     return v;
 300   }
 301
 302   bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 303     bool v = ((x_ >> kReadShift) & 3)
 304         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 305     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 306         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 307     return v;
 308   }
 309
 310  private:
 311   static const u64 kReadShift   = 5 + kClkBits;
 312   static const u64 kReadBit     = 1ull << kReadShift;
 313   static const u64 kAtomicShift = 6 + kClkBits;
 314   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 315
 316   u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
 317
 318   static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
 319     if (s1.addr0() == s2.addr0()) return true;
 320     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 321       return true;
 322     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 323       return true;
 324     return false;
 325   }
 326 };
 327
 328 struct ThreadSignalContext;
 329
 330 struct JmpBuf {
 331   uptr sp;
 332   uptr mangled_sp;
 333   int int_signal_send;
 334   bool in_blocking_func;
 335   uptr in_signal_handler;
 336   uptr *shadow_stack_pos;
 337 };
 338
 339 // A Processor represents a physical thread, or a P for Go.
 340 // It is used to store internal resources like allocate cache, and does not
 341 // participate in race-detection logic (invisible to end user).
 342 // In C++ it is tied to an OS thread just like ThreadState, however ideally
 343 // it should be tied to a CPU (this way we will have fewer allocator caches).
 344 // In Go it is tied to a P, so there are significantly fewer Processor's than
 345 // ThreadState's (which are tied to Gs).
 346 // A ThreadState must be wired with a Processor to handle events.
 347 struct Processor {
 348   ThreadState *thr; // currently wired thread, or nullptr
 349 #if !SANITIZER_GO
 350   AllocatorCache alloc_cache;
 351   InternalAllocatorCache internal_alloc_cache;
 352 #endif
 353   DenseSlabAllocCache block_cache;
 354   DenseSlabAllocCache sync_cache;
 355   DenseSlabAllocCache clock_cache;
 356   DDPhysicalThread *dd_pt;
 357 };
 358
 359 #if !SANITIZER_GO
 360 // ScopedGlobalProcessor temporary setups a global processor for the current
 361 // thread, if it does not have one. Intended for interceptors that can run
 362 // at the very thread end, when we already destroyed the thread processor.
 363 struct ScopedGlobalProcessor {
 364   ScopedGlobalProcessor();
 365   ~ScopedGlobalProcessor();
 366 };
 367 #endif
 368
 369 // This struct is stored in TLS.
 370 struct ThreadState {
 371   FastState fast_state;
 372   // Synch epoch represents the threads's epoch before the last synchronization
 373   // action. It allows to reduce number of shadow state updates.
 374   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 375   // if we are processing write to X from the same thread at epoch=200,
 376   // we do nothing, because both writes happen in the same 'synch epoch'.
 377   // That is, if another memory access does not race with the former write,
 378   // it does not race with the latter as well.
 379   // QUESTION: can we can squeeze this into ThreadState::Fast?
 380   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 381   // taken by epoch between synchs.
 382   // This way we can save one load from tls.
 383   u64 fast_synch_epoch;
 384   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 385   // We do not distinguish beteween ignoring reads and writes
 386   // for better performance.
 387   int ignore_reads_and_writes;
 388   int ignore_sync;
 389   int suppress_reports;
 390   // Go does not support ignores.
 391 #if !SANITIZER_GO
 392   IgnoreSet mop_ignore_set;
 393   IgnoreSet sync_ignore_set;
 394 #endif
 395   // C/C++ uses fixed size shadow stack embed into Trace.
 396   // Go uses malloc-allocated shadow stack with dynamic size.
 397   uptr *shadow_stack;
 398   uptr *shadow_stack_end;
 399   uptr *shadow_stack_pos;
 400   u64 *racy_shadow_addr;
 401   u64 racy_state[2];
 402   MutexSet mset;
 403   ThreadClock clock;
 404 #if !SANITIZER_GO
 405   Vector<JmpBuf> jmp_bufs;
 406   int ignore_interceptors;
 407 #endif
 408 #if TSAN_COLLECT_STATS
 409   u64 stat[StatCnt];
 410 #endif
 411   const int tid;
 412   const int unique_id;
 413   bool in_symbolizer;
 414   bool in_ignored_lib;
 415   bool is_inited;
 416   bool is_dead;
 417   bool is_freeing;
 418   bool is_vptr_access;
 419   const uptr stk_addr;
 420   const uptr stk_size;
 421   const uptr tls_addr;
 422   const uptr tls_size;
 423   ThreadContext *tctx;
 424
 425 #if SANITIZER_DEBUG && !SANITIZER_GO
 426   InternalDeadlockDetector internal_deadlock_detector;
 427 #endif
 428   DDLogicalThread *dd_lt;
 429
 430   // Current wired Processor, or nullptr. Required to handle any events.
 431   Processor *proc1;
 432 #if !SANITIZER_GO
 433   Processor *proc() { return proc1; }
 434 #else
 435   Processor *proc();
 436 #endif
 437
 438   atomic_uintptr_t in_signal_handler;
 439   ThreadSignalContext *signal_ctx;
 440
 441 #if !SANITIZER_GO
 442   u32 last_sleep_stack_id;
 443   ThreadClock last_sleep_clock;
 444 #endif
 445
 446   // Set in regions of runtime that must be signal-safe and fork-safe.
 447   // If set, malloc must not be called.
 448   int nomalloc;
 449
 450   const ReportDesc *current_report;
 451
 452   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 453                        unsigned reuse_count,
 454                        uptr stk_addr, uptr stk_size,
 455                        uptr tls_addr, uptr tls_size);
 456 };
 457
 458 #if !SANITIZER_GO
 459 #if SANITIZER_MAC || SANITIZER_ANDROID
 460 ThreadState *cur_thread();
 461 void cur_thread_finalize();
 462 #else
 463 __attribute__((tls_model("initial-exec")))
 464 extern THREADLOCAL char cur_thread_placeholder[];
 465 INLINE ThreadState *cur_thread() {
 466   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
 467 }
 468 INLINE void cur_thread_finalize() { }
 469 #endif  // SANITIZER_MAC || SANITIZER_ANDROID
 470 #endif  // SANITIZER_GO
 471
 472 class ThreadContext : public ThreadContextBase {
 473  public:
 474   explicit ThreadContext(int tid);
 475   ~ThreadContext();
 476   ThreadState *thr;
 477   u32 creation_stack_id;
 478   SyncClock sync;
 479   // Epoch at which the thread had started.
 480   // If we see an event from the thread stamped by an older epoch,
 481   // the event is from a dead thread that shared tid with this thread.
 482   u64 epoch0;
 483   u64 epoch1;
 484
 485   // Override superclass callbacks.
 486   void OnDead() override;
 487   void OnJoined(void *arg) override;
 488   void OnFinished() override;
 489   void OnStarted(void *arg) override;
 490   void OnCreated(void *arg) override;
 491   void OnReset() override;
 492   void OnDetached(void *arg) override;
 493 };
 494
 495 struct RacyStacks {
 496   MD5Hash hash[2];
 497   bool operator==(const RacyStacks &other) const {
 498     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 499       return true;
 500     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 501       return true;
 502     return false;
 503   }
 504 };
 505
 506 struct RacyAddress {
 507   uptr addr_min;
 508   uptr addr_max;
 509 };
 510
 511 struct FiredSuppression {
 512   ReportType type;
 513   uptr pc_or_addr;
 514   Suppression *supp;
 515 };
 516
 517 struct Context {
 518   Context();
 519
 520   bool initialized;
 521 #if !SANITIZER_GO
 522   bool after_multithreaded_fork;
 523 #endif
 524
 525   MetaMap metamap;
 526
 527   Mutex report_mtx;
 528   int nreported;
 529   int nmissed_expected;
 530   atomic_uint64_t last_symbolize_time_ns;
 531
 532   void *background_thread;
 533   atomic_uint32_t stop_background_thread;
 534
 535   ThreadRegistry *thread_registry;
 536
 537   Mutex racy_mtx;
 538   Vector<RacyStacks> racy_stacks;
 539   Vector<RacyAddress> racy_addresses;
 540   // Number of fired suppressions may be large enough.
 541   Mutex fired_suppressions_mtx;
 542   InternalMmapVector<FiredSuppression> fired_suppressions;
 543   DDetector *dd;
 544
 545   ClockAlloc clock_alloc;
 546
 547   Flags flags;
 548
 549   u64 stat[StatCnt];
 550   u64 int_alloc_cnt[MBlockTypeCount];
 551   u64 int_alloc_siz[MBlockTypeCount];
 552 };
 553
 554 extern Context *ctx;  // The one and the only global runtime context.
 555
 556 ALWAYS_INLINE Flags *flags() {
 557   return &ctx->flags;
 558 }
 559
 560 struct ScopedIgnoreInterceptors {
 561   ScopedIgnoreInterceptors() {
 562 #if !SANITIZER_GO
 563     cur_thread()->ignore_interceptors++;
 564 #endif
 565   }
 566
 567   ~ScopedIgnoreInterceptors() {
 568 #if !SANITIZER_GO
 569     cur_thread()->ignore_interceptors--;
 570 #endif
 571   }
 572 };
 573
 574 const char *GetObjectTypeFromTag(uptr tag);
 575 const char *GetReportHeaderFromTag(uptr tag);
 576 uptr TagFromShadowStackFrame(uptr pc);
 577
 578 class ScopedReportBase {
 579  public:
 580   void AddMemoryAccess(uptr addr, uptr external_tag, Shadow s, StackTrace stack,
 581                        const MutexSet *mset);
 582   void AddStack(StackTrace stack, bool suppressable = false);
 583   void AddThread(const ThreadContext *tctx, bool suppressable = false);
 584   void AddThread(int unique_tid, bool suppressable = false);
 585   void AddUniqueTid(int unique_tid);
 586   void AddMutex(const SyncVar *s);
 587   u64 AddMutex(u64 id);
 588   void AddLocation(uptr addr, uptr size);
 589   void AddSleep(u32 stack_id);
 590   void SetCount(int count);
 591
 592   const ReportDesc *GetReport() const;
 593
 594  protected:
 595   ScopedReportBase(ReportType typ, uptr tag);
 596   ~ScopedReportBase();
 597
 598  private:
 599   ReportDesc *rep_;
 600   // Symbolizer makes lots of intercepted calls. If we try to process them,
 601   // at best it will cause deadlocks on internal mutexes.
 602   ScopedIgnoreInterceptors ignore_interceptors_;
 603
 604   void AddDeadMutex(u64 id);
 605
 606   ScopedReportBase(const ScopedReportBase &) = delete;
 607   void operator=(const ScopedReportBase &) = delete;
 608 };
 609
 610 class ScopedReport : public ScopedReportBase {
 611  public:
 612   explicit ScopedReport(ReportType typ, uptr tag = kExternalTagNone);
 613   ~ScopedReport();
 614
 615  private:
 616   ScopedErrorReportLock lock_;
 617 };
 618
 619 ThreadContext *IsThreadStackOrTls(uptr addr, bool *is_stack);
 620 void RestoreStack(int tid, const u64 epoch, VarSizeStackTrace *stk,
 621                   MutexSet *mset, uptr *tag = nullptr);
 622
 623 // The stack could look like:
 624 //   <start> | <main> | <foo> | tag | <bar>
 625 // This will extract the tag and keep:
 626 //   <start> | <main> | <foo> | <bar>
 627 template<typename StackTraceTy>
 628 void ExtractTagFromStack(StackTraceTy *stack, uptr *tag = nullptr) {
 629   if (stack->size < 2) return;
 630   uptr possible_tag_pc = stack->trace[stack->size - 2];
 631   uptr possible_tag = TagFromShadowStackFrame(possible_tag_pc);
 632   if (possible_tag == kExternalTagNone) return;
 633   stack->trace_buffer[stack->size - 2] = stack->trace_buffer[stack->size - 1];
 634   stack->size -= 1;
 635   if (tag) *tag = possible_tag;
 636 }
 637
 638 template<typename StackTraceTy>
 639 void ObtainCurrentStack(ThreadState *thr, uptr toppc, StackTraceTy *stack,
 640                         uptr *tag = nullptr) {
 641   uptr size = thr->shadow_stack_pos - thr->shadow_stack;
 642   uptr start = 0;
 643   if (size + !!toppc > kStackTraceMax) {
 644     start = size + !!toppc - kStackTraceMax;
 645     size = kStackTraceMax - !!toppc;
 646   }
 647   stack->Init(&thr->shadow_stack[start], size, toppc);
 648   ExtractTagFromStack(stack, tag);
 649 }
 650
 651 #define GET_STACK_TRACE_FATAL(thr, pc) \
 652   VarSizeStackTrace stack; \
 653   ObtainCurrentStack(thr, pc, &stack); \
 654   stack.ReverseOrder();
 655
 656 #if TSAN_COLLECT_STATS
 657 void StatAggregate(u64 *dst, u64 *src);
 658 void StatOutput(u64 *stat);
 659 #endif
 660
 661 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 662 #if TSAN_COLLECT_STATS
 663   thr->stat[typ] += n;
 664 #endif
 665 }
 666 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 667 #if TSAN_COLLECT_STATS
 668   thr->stat[typ] = n;
 669 #endif
 670 }
 671
 672 void MapShadow(uptr addr, uptr size);
 673 void MapThreadTrace(uptr addr, uptr size, const char *name);
 674 void DontNeedShadowFor(uptr addr, uptr size);
 675 void InitializeShadowMemory();
 676 void InitializeInterceptors();
 677 void InitializeLibIgnore();
 678 void InitializeDynamicAnnotations();
 679
 680 void ForkBefore(ThreadState *thr, uptr pc);
 681 void ForkParentAfter(ThreadState *thr, uptr pc);
 682 void ForkChildAfter(ThreadState *thr, uptr pc);
 683
 684 void ReportRace(ThreadState *thr);
 685 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
 686 bool IsFiredSuppression(Context *ctx, ReportType type, StackTrace trace);
 687 bool IsExpectedReport(uptr addr, uptr size);
 688 void PrintMatchedBenignRaces();
 689
 690 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 691 # define DPrintf Printf
 692 #else
 693 # define DPrintf(...)
 694 #endif
 695
 696 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 697 # define DPrintf2 Printf
 698 #else
 699 # define DPrintf2(...)
 700 #endif
 701
 702 u32 CurrentStackId(ThreadState *thr, uptr pc);
 703 ReportStack *SymbolizeStackId(u32 stack_id);
 704 void PrintCurrentStack(ThreadState *thr, uptr pc);
 705 void PrintCurrentStackSlow(uptr pc);  // uses libunwind
 706
 707 void Initialize(ThreadState *thr);
 708 void MaybeSpawnBackgroundThread();
 709 int Finalize(ThreadState *thr);
 710
 711 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
 712 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
 713
 714 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 715     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 716 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 717     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 718     u64 *shadow_mem, Shadow cur);
 719 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 720     uptr size, bool is_write);
 721 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 722     uptr size, uptr step, bool is_write);
 723 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 724     int size, bool kAccessIsWrite, bool kIsAtomic);
 725
 726 const int kSizeLog1 = 0;
 727 const int kSizeLog2 = 1;
 728 const int kSizeLog4 = 2;
 729 const int kSizeLog8 = 3;
 730
 731 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 732                                      uptr addr, int kAccessSizeLog) {
 733   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 734 }
 735
 736 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 737                                       uptr addr, int kAccessSizeLog) {
 738   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 739 }
 740
 741 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 742                                            uptr addr, int kAccessSizeLog) {
 743   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 744 }
 745
 746 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 747                                             uptr addr, int kAccessSizeLog) {
 748   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 749 }
 750
 751 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 752 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 753 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 754
 755 void ThreadIgnoreBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 756 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
 757 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 758 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
 759
 760 void FuncEntry(ThreadState *thr, uptr pc);
 761 void FuncExit(ThreadState *thr);
 762
 763 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 764 void ThreadStart(ThreadState *thr, int tid, tid_t os_id, bool workerthread);
 765 void ThreadFinish(ThreadState *thr);
 766 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
 767 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 768 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 769 void ThreadFinalize(ThreadState *thr);
 770 void ThreadSetName(ThreadState *thr, const char *name);
 771 int ThreadCount(ThreadState *thr);
 772 void ProcessPendingSignals(ThreadState *thr);
 773
 774 Processor *ProcCreate();
 775 void ProcDestroy(Processor *proc);
 776 void ProcWire(Processor *proc, ThreadState *thr);
 777 void ProcUnwire(Processor *proc, ThreadState *thr);
 778
 779 // Note: the parameter is called flagz, because flags is already taken
 780 // by the global function that returns flags.
 781 void MutexCreate(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 782 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 783 void MutexPreLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 784 void MutexPostLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0,
 785     int rec = 1);
 786 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 787 void MutexPreReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 788 void MutexPostReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 789 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 790 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 791 void MutexRepair(ThreadState *thr, uptr pc, uptr addr);  // call on EOWNERDEAD
 792 void MutexInvalidAccess(ThreadState *thr, uptr pc, uptr addr);
 793
 794 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 795 // AcquireGlobal synchronizes the current thread with all other threads.
 796 // In terms of happens-before relation, it draws a HB edge from all threads
 797 // (where they happen to execute right now) to the current thread. We use it to
 798 // handle Go finalizers. Namely, finalizer goroutine executes AcquireGlobal
 799 // right before executing finalizers. This provides a coarse, but simple
 800 // approximation of the actual required synchronization.
 801 void AcquireGlobal(ThreadState *thr, uptr pc);
 802 void Release(ThreadState *thr, uptr pc, uptr addr);
 803 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 804 void AfterSleep(ThreadState *thr, uptr pc);
 805 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 806 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 807 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
 808 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 809
 810 // The hacky call uses custom calling convention and an assembly thunk.
 811 // It is considerably faster that a normal call for the caller
 812 // if it is not executed (it is intended for slow paths from hot functions).
 813 // The trick is that the call preserves all registers and the compiler
 814 // does not treat it as a call.
 815 // If it does not work for you, use normal call.
 816 #if !SANITIZER_DEBUG && defined(__x86_64__) && !SANITIZER_MAC
 817 // The caller may not create the stack frame for itself at all,
 818 // so we create a reserve stack frame for it (1024b must be enough).
 819 #define HACKY_CALL(f) \
 820   __asm__ __volatile__("sub $1024, %%rsp;" \
 821                        CFI_INL_ADJUST_CFA_OFFSET(1024) \
 822                        ".hidden " #f "_thunk;" \
 823                        "call " #f "_thunk;" \
 824                        "add $1024, %%rsp;" \
 825                        CFI_INL_ADJUST_CFA_OFFSET(-1024) \
 826                        ::: "memory", "cc");
 827 #else
 828 #define HACKY_CALL(f) f()
 829 #endif
 830
 831 void TraceSwitch(ThreadState *thr);
 832 uptr TraceTopPC(ThreadState *thr);
 833 uptr TraceSize();
 834 uptr TraceParts();
 835 Trace *ThreadTrace(int tid);
 836
 837 extern "C" void __tsan_trace_switch();
 838 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 839                                         EventType typ, u64 addr) {
 840   if (!kCollectHistory)
 841     return;
 842   DCHECK_GE((int)typ, 0);
 843   DCHECK_LE((int)typ, 7);
 844   DCHECK_EQ(GetLsb(addr, kEventPCBits), addr);
 845   StatInc(thr, StatEvents);
 846   u64 pos = fs.GetTracePos();
 847   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 848 #if !SANITIZER_GO
 849     HACKY_CALL(__tsan_trace_switch);
 850 #else
 851     TraceSwitch(thr);
 852 #endif
 853   }
 854   Event *trace = (Event*)GetThreadTrace(fs.tid());
 855   Event *evp = &trace[pos];
 856   Event ev = (u64)addr | ((u64)typ << kEventPCBits);
 857   *evp = ev;
 858 }
 859
 860 #if !SANITIZER_GO
 861 uptr ALWAYS_INLINE HeapEnd() {
 862   return HeapMemEnd() + PrimaryAllocator::AdditionalSize();
 863 }
 864 #endif
 865
 866 }  // namespace __tsan
 867
 868 #endif  // TSAN_RTL_H