libsanitizer/tsan/tsan_rtl.h

   1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
   2 //
   3 // This file is distributed under the University of Illinois Open Source
   4 // License. See LICENSE.TXT for details.
   5 //
   6 //===----------------------------------------------------------------------===//
   7 //
   8 // This file is a part of ThreadSanitizer (TSan), a race detector.
   9 //
  10 // Main internal TSan header file.
  11 //
  12 // Ground rules:
  13 //   - C++ run-time should not be used (static CTORs, RTTI, exceptions, static
  14 //     function-scope locals)
  15 //   - All functions/classes/etc reside in namespace __tsan, except for those
  16 //     declared in tsan_interface.h.
  17 //   - Platform-specific files should be used instead of ifdefs (*).
  18 //   - No system headers included in header files (*).
  19 //   - Platform specific headres included only into platform-specific files (*).
  20 //
  21 //  (*) Except when inlining is critical for performance.
  22 //===----------------------------------------------------------------------===//
  23
  24 #ifndef TSAN_RTL_H
  25 #define TSAN_RTL_H
  26
  27 #include "sanitizer_common/sanitizer_allocator.h"
  28 #include "sanitizer_common/sanitizer_allocator_internal.h"
  29 #include "sanitizer_common/sanitizer_asm.h"
  30 #include "sanitizer_common/sanitizer_common.h"
  31 #include "sanitizer_common/sanitizer_deadlock_detector_interface.h"
  32 #include "sanitizer_common/sanitizer_libignore.h"
  33 #include "sanitizer_common/sanitizer_suppressions.h"
  34 #include "sanitizer_common/sanitizer_thread_registry.h"
  35 #include "tsan_clock.h"
  36 #include "tsan_defs.h"
  37 #include "tsan_flags.h"
  38 #include "tsan_sync.h"
  39 #include "tsan_trace.h"
  40 #include "tsan_vector.h"
  41 #include "tsan_report.h"
  42 #include "tsan_platform.h"
  43 #include "tsan_mutexset.h"
  44 #include "tsan_ignoreset.h"
  45 #include "tsan_stack_trace.h"
  46
  47 #if SANITIZER_WORDSIZE != 64
  48 # error "ThreadSanitizer is supported only on 64-bit platforms"
  49 #endif
  50
  51 namespace __tsan {
  52
  53 #if !SANITIZER_GO
  54 struct MapUnmapCallback;
  55 #if defined(__mips64) || defined(__aarch64__) || defined(__powerpc__)
  56 static const uptr kAllocatorRegionSizeLog = 20;
  57 static const uptr kAllocatorNumRegions =
  58     SANITIZER_MMAP_RANGE_SIZE >> kAllocatorRegionSizeLog;
  59 typedef TwoLevelByteMap<(kAllocatorNumRegions >> 12), 1 << 12,
  60     MapUnmapCallback> ByteMap;
  61 struct AP32 {
  62   static const uptr kSpaceBeg = 0;
  63   static const u64 kSpaceSize = SANITIZER_MMAP_RANGE_SIZE;
  64   static const uptr kMetadataSize = 0;
  65   typedef __sanitizer::CompactSizeClassMap SizeClassMap;
  66   static const uptr kRegionSizeLog = kAllocatorRegionSizeLog;
  67   typedef __tsan::ByteMap ByteMap;
  68   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  69   static const uptr kFlags = 0;
  70 };
  71 typedef SizeClassAllocator32<AP32> PrimaryAllocator;
  72 #else
  73 struct AP64 {  // Allocator64 parameters. Deliberately using a short name.
  74   static const uptr kSpaceBeg = Mapping::kHeapMemBeg;
  75   static const uptr kSpaceSize = Mapping::kHeapMemEnd - Mapping::kHeapMemBeg;
  76   static const uptr kMetadataSize = 0;
  77   typedef DefaultSizeClassMap SizeClassMap;
  78   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  79   static const uptr kFlags = 0;
  80 };
  81 typedef SizeClassAllocator64<AP64> PrimaryAllocator;
  82 #endif
  83 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
  84 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
  85 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
  86     SecondaryAllocator> Allocator;
  87 Allocator *allocator();
  88 #endif
  89
  90 void TsanCheckFailed(const char *file, int line, const char *cond,
  91                      u64 v1, u64 v2);
  92
  93 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
  94
  95 // FastState (from most significant bit):
  96 //   ignore          : 1
  97 //   tid             : kTidBits
  98 //   unused          : -
  99 //   history_size    : 3
 100 //   epoch           : kClkBits
 101 class FastState {
 102  public:
 103   FastState(u64 tid, u64 epoch) {
 104     x_ = tid << kTidShift;
 105     x_ |= epoch;
 106     DCHECK_EQ(tid, this->tid());
 107     DCHECK_EQ(epoch, this->epoch());
 108     DCHECK_EQ(GetIgnoreBit(), false);
 109   }
 110
 111   explicit FastState(u64 x)
 112       : x_(x) {
 113   }
 114
 115   u64 raw() const {
 116     return x_;
 117   }
 118
 119   u64 tid() const {
 120     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
 121     return res;
 122   }
 123
 124   u64 TidWithIgnore() const {
 125     u64 res = x_ >> kTidShift;
 126     return res;
 127   }
 128
 129   u64 epoch() const {
 130     u64 res = x_ & ((1ull << kClkBits) - 1);
 131     return res;
 132   }
 133
 134   void IncrementEpoch() {
 135     u64 old_epoch = epoch();
 136     x_ += 1;
 137     DCHECK_EQ(old_epoch + 1, epoch());
 138     (void)old_epoch;
 139   }
 140
 141   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 142   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 143   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 144
 145   void SetHistorySize(int hs) {
 146     CHECK_GE(hs, 0);
 147     CHECK_LE(hs, 7);
 148     x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
 149   }
 150
 151   ALWAYS_INLINE
 152   int GetHistorySize() const {
 153     return (int)((x_ >> kHistoryShift) & kHistoryMask);
 154   }
 155
 156   void ClearHistorySize() {
 157     SetHistorySize(0);
 158   }
 159
 160   ALWAYS_INLINE
 161   u64 GetTracePos() const {
 162     const int hs = GetHistorySize();
 163     // When hs == 0, the trace consists of 2 parts.
 164     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 165     return epoch() & mask;
 166   }
 167
 168  private:
 169   friend class Shadow;
 170   static const int kTidShift = 64 - kTidBits - 1;
 171   static const u64 kIgnoreBit = 1ull << 63;
 172   static const u64 kFreedBit = 1ull << 63;
 173   static const u64 kHistoryShift = kClkBits;
 174   static const u64 kHistoryMask = 7;
 175   u64 x_;
 176 };
 177
 178 // Shadow (from most significant bit):
 179 //   freed           : 1
 180 //   tid             : kTidBits
 181 //   is_atomic       : 1
 182 //   is_read         : 1
 183 //   size_log        : 2
 184 //   addr0           : 3
 185 //   epoch           : kClkBits
 186 class Shadow : public FastState {
 187  public:
 188   explicit Shadow(u64 x)
 189       : FastState(x) {
 190   }
 191
 192   explicit Shadow(const FastState &s)
 193       : FastState(s.x_) {
 194     ClearHistorySize();
 195   }
 196
 197   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 198     DCHECK_EQ((x_ >> kClkBits) & 31, 0);
 199     DCHECK_LE(addr0, 7);
 200     DCHECK_LE(kAccessSizeLog, 3);
 201     x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
 202     DCHECK_EQ(kAccessSizeLog, size_log());
 203     DCHECK_EQ(addr0, this->addr0());
 204   }
 205
 206   void SetWrite(unsigned kAccessIsWrite) {
 207     DCHECK_EQ(x_ & kReadBit, 0);
 208     if (!kAccessIsWrite)
 209       x_ |= kReadBit;
 210     DCHECK_EQ(kAccessIsWrite, IsWrite());
 211   }
 212
 213   void SetAtomic(bool kIsAtomic) {
 214     DCHECK(!IsAtomic());
 215     if (kIsAtomic)
 216       x_ |= kAtomicBit;
 217     DCHECK_EQ(IsAtomic(), kIsAtomic);
 218   }
 219
 220   bool IsAtomic() const {
 221     return x_ & kAtomicBit;
 222   }
 223
 224   bool IsZero() const {
 225     return x_ == 0;
 226   }
 227
 228   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 229     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 230     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 231     return shifted_xor == 0;
 232   }
 233
 234   static ALWAYS_INLINE
 235   bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 236     u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
 237     return masked_xor == 0;
 238   }
 239
 240   static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
 241       unsigned kS2AccessSize) {
 242     bool res = false;
 243     u64 diff = s1.addr0() - s2.addr0();
 244     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
 245       // if (s1.addr0() + size1) > s2.addr0()) return true;
 246       if (s1.size() > -diff)
 247         res = true;
 248     } else {
 249       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 250       if (kS2AccessSize > diff)
 251         res = true;
 252     }
 253     DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
 254     DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
 255     return res;
 256   }
 257
 258   u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
 259   u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
 260   bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
 261   bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
 262
 263   // The idea behind the freed bit is as follows.
 264   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 265   // values with tid/epoch related to the free and the freed bit set.
 266   // During memory accesses processing the freed bit is considered
 267   // as msb of tid. So any access races with shadow with freed bit set
 268   // (it is as if write from a thread with which we never synchronized before).
 269   // This allows us to detect accesses to freed memory w/o additional
 270   // overheads in memory access processing and at the same time restore
 271   // tid/epoch of free.
 272   void MarkAsFreed() {
 273      x_ |= kFreedBit;
 274   }
 275
 276   bool IsFreed() const {
 277     return x_ & kFreedBit;
 278   }
 279
 280   bool GetFreedAndReset() {
 281     bool res = x_ & kFreedBit;
 282     x_ &= ~kFreedBit;
 283     return res;
 284   }
 285
 286   bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 287     bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
 288         | (u64(kIsAtomic) << kAtomicShift));
 289     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 290     return v;
 291   }
 292
 293   bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 294     bool v = ((x_ >> kReadShift) & 3)
 295         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 296     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 297         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 298     return v;
 299   }
 300
 301   bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 302     bool v = ((x_ >> kReadShift) & 3)
 303         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 304     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 305         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 306     return v;
 307   }
 308
 309  private:
 310   static const u64 kReadShift   = 5 + kClkBits;
 311   static const u64 kReadBit     = 1ull << kReadShift;
 312   static const u64 kAtomicShift = 6 + kClkBits;
 313   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 314
 315   u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
 316
 317   static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
 318     if (s1.addr0() == s2.addr0()) return true;
 319     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 320       return true;
 321     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 322       return true;
 323     return false;
 324   }
 325 };
 326
 327 struct ThreadSignalContext;
 328
 329 struct JmpBuf {
 330   uptr sp;
 331   uptr mangled_sp;
 332   int int_signal_send;
 333   bool in_blocking_func;
 334   uptr in_signal_handler;
 335   uptr *shadow_stack_pos;
 336 };
 337
 338 // A Processor represents a physical thread, or a P for Go.
 339 // It is used to store internal resources like allocate cache, and does not
 340 // participate in race-detection logic (invisible to end user).
 341 // In C++ it is tied to an OS thread just like ThreadState, however ideally
 342 // it should be tied to a CPU (this way we will have fewer allocator caches).
 343 // In Go it is tied to a P, so there are significantly fewer Processor's than
 344 // ThreadState's (which are tied to Gs).
 345 // A ThreadState must be wired with a Processor to handle events.
 346 struct Processor {
 347   ThreadState *thr; // currently wired thread, or nullptr
 348 #if !SANITIZER_GO
 349   AllocatorCache alloc_cache;
 350   InternalAllocatorCache internal_alloc_cache;
 351 #endif
 352   DenseSlabAllocCache block_cache;
 353   DenseSlabAllocCache sync_cache;
 354   DenseSlabAllocCache clock_cache;
 355   DDPhysicalThread *dd_pt;
 356 };
 357
 358 #if !SANITIZER_GO
 359 // ScopedGlobalProcessor temporary setups a global processor for the current
 360 // thread, if it does not have one. Intended for interceptors that can run
 361 // at the very thread end, when we already destroyed the thread processor.
 362 struct ScopedGlobalProcessor {
 363   ScopedGlobalProcessor();
 364   ~ScopedGlobalProcessor();
 365 };
 366 #endif
 367
 368 // This struct is stored in TLS.
 369 struct ThreadState {
 370   FastState fast_state;
 371   // Synch epoch represents the threads's epoch before the last synchronization
 372   // action. It allows to reduce number of shadow state updates.
 373   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 374   // if we are processing write to X from the same thread at epoch=200,
 375   // we do nothing, because both writes happen in the same 'synch epoch'.
 376   // That is, if another memory access does not race with the former write,
 377   // it does not race with the latter as well.
 378   // QUESTION: can we can squeeze this into ThreadState::Fast?
 379   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 380   // taken by epoch between synchs.
 381   // This way we can save one load from tls.
 382   u64 fast_synch_epoch;
 383   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 384   // We do not distinguish beteween ignoring reads and writes
 385   // for better performance.
 386   int ignore_reads_and_writes;
 387   int ignore_sync;
 388   int suppress_reports;
 389   // Go does not support ignores.
 390 #if !SANITIZER_GO
 391   IgnoreSet mop_ignore_set;
 392   IgnoreSet sync_ignore_set;
 393 #endif
 394   // C/C++ uses fixed size shadow stack embed into Trace.
 395   // Go uses malloc-allocated shadow stack with dynamic size.
 396   uptr *shadow_stack;
 397   uptr *shadow_stack_end;
 398   uptr *shadow_stack_pos;
 399   u64 *racy_shadow_addr;
 400   u64 racy_state[2];
 401   MutexSet mset;
 402   ThreadClock clock;
 403 #if !SANITIZER_GO
 404   Vector<JmpBuf> jmp_bufs;
 405   int ignore_interceptors;
 406 #endif
 407 #if TSAN_COLLECT_STATS
 408   u64 stat[StatCnt];
 409 #endif
 410   const int tid;
 411   const int unique_id;
 412   bool in_symbolizer;
 413   bool in_ignored_lib;
 414   bool is_inited;
 415   bool is_dead;
 416   bool is_freeing;
 417   bool is_vptr_access;
 418   const uptr stk_addr;
 419   const uptr stk_size;
 420   const uptr tls_addr;
 421   const uptr tls_size;
 422   ThreadContext *tctx;
 423
 424 #if SANITIZER_DEBUG && !SANITIZER_GO
 425   InternalDeadlockDetector internal_deadlock_detector;
 426 #endif
 427   DDLogicalThread *dd_lt;
 428
 429   // Current wired Processor, or nullptr. Required to handle any events.
 430   Processor *proc1;
 431 #if !SANITIZER_GO
 432   Processor *proc() { return proc1; }
 433 #else
 434   Processor *proc();
 435 #endif
 436
 437   atomic_uintptr_t in_signal_handler;
 438   ThreadSignalContext *signal_ctx;
 439
 440 #if !SANITIZER_GO
 441   u32 last_sleep_stack_id;
 442   ThreadClock last_sleep_clock;
 443 #endif
 444
 445   // Set in regions of runtime that must be signal-safe and fork-safe.
 446   // If set, malloc must not be called.
 447   int nomalloc;
 448
 449   const ReportDesc *current_report;
 450
 451   explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 452                        unsigned reuse_count,
 453                        uptr stk_addr, uptr stk_size,
 454                        uptr tls_addr, uptr tls_size);
 455 };
 456
 457 #if !SANITIZER_GO
 458 #if SANITIZER_MAC || SANITIZER_ANDROID
 459 ThreadState *cur_thread();
 460 void cur_thread_finalize();
 461 #else
 462 __attribute__((tls_model("initial-exec")))
 463 extern THREADLOCAL char cur_thread_placeholder[];
 464 INLINE ThreadState *cur_thread() {
 465   return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
 466 }
 467 INLINE void cur_thread_finalize() { }
 468 #endif  // SANITIZER_MAC || SANITIZER_ANDROID
 469 #endif  // SANITIZER_GO
 470
 471 class ThreadContext : public ThreadContextBase {
 472  public:
 473   explicit ThreadContext(int tid);
 474   ~ThreadContext();
 475   ThreadState *thr;
 476   u32 creation_stack_id;
 477   SyncClock sync;
 478   // Epoch at which the thread had started.
 479   // If we see an event from the thread stamped by an older epoch,
 480   // the event is from a dead thread that shared tid with this thread.
 481   u64 epoch0;
 482   u64 epoch1;
 483
 484   // Override superclass callbacks.
 485   void OnDead() override;
 486   void OnJoined(void *arg) override;
 487   void OnFinished() override;
 488   void OnStarted(void *arg) override;
 489   void OnCreated(void *arg) override;
 490   void OnReset() override;
 491   void OnDetached(void *arg) override;
 492 };
 493
 494 struct RacyStacks {
 495   MD5Hash hash[2];
 496   bool operator==(const RacyStacks &other) const {
 497     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 498       return true;
 499     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 500       return true;
 501     return false;
 502   }
 503 };
 504
 505 struct RacyAddress {
 506   uptr addr_min;
 507   uptr addr_max;
 508 };
 509
 510 struct FiredSuppression {
 511   ReportType type;
 512   uptr pc_or_addr;
 513   Suppression *supp;
 514 };
 515
 516 struct Context {
 517   Context();
 518
 519   bool initialized;
 520   bool after_multithreaded_fork;
 521
 522   MetaMap metamap;
 523
 524   Mutex report_mtx;
 525   int nreported;
 526   int nmissed_expected;
 527   atomic_uint64_t last_symbolize_time_ns;
 528
 529   void *background_thread;
 530   atomic_uint32_t stop_background_thread;
 531
 532   ThreadRegistry *thread_registry;
 533
 534   Mutex racy_mtx;
 535   Vector<RacyStacks> racy_stacks;
 536   Vector<RacyAddress> racy_addresses;
 537   // Number of fired suppressions may be large enough.
 538   Mutex fired_suppressions_mtx;
 539   InternalMmapVector<FiredSuppression> fired_suppressions;
 540   DDetector *dd;
 541
 542   ClockAlloc clock_alloc;
 543
 544   Flags flags;
 545
 546   u64 stat[StatCnt];
 547   u64 int_alloc_cnt[MBlockTypeCount];
 548   u64 int_alloc_siz[MBlockTypeCount];
 549 };
 550
 551 extern Context *ctx;  // The one and the only global runtime context.
 552
 553 ALWAYS_INLINE Flags *flags() {
 554   return &ctx->flags;
 555 }
 556
 557 struct ScopedIgnoreInterceptors {
 558   ScopedIgnoreInterceptors() {
 559 #if !SANITIZER_GO
 560     cur_thread()->ignore_interceptors++;
 561 #endif
 562   }
 563
 564   ~ScopedIgnoreInterceptors() {
 565 #if !SANITIZER_GO
 566     cur_thread()->ignore_interceptors--;
 567 #endif
 568   }
 569 };
 570
 571 const char *GetObjectTypeFromTag(uptr tag);
 572 const char *GetReportHeaderFromTag(uptr tag);
 573 uptr TagFromShadowStackFrame(uptr pc);
 574
 575 class ScopedReport {
 576  public:
 577   explicit ScopedReport(ReportType typ, uptr tag = kExternalTagNone);
 578   ~ScopedReport();
 579
 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  private:
 595   ReportDesc *rep_;
 596   // Symbolizer makes lots of intercepted calls. If we try to process them,
 597   // at best it will cause deadlocks on internal mutexes.
 598   ScopedIgnoreInterceptors ignore_interceptors_;
 599
 600   void AddDeadMutex(u64 id);
 601
 602   ScopedReport(const ScopedReport&);
 603   void operator = (const ScopedReport&);
 604 };
 605
 606 ThreadContext *IsThreadStackOrTls(uptr addr, bool *is_stack);
 607 void RestoreStack(int tid, const u64 epoch, VarSizeStackTrace *stk,
 608                   MutexSet *mset, uptr *tag = nullptr);
 609
 610 // The stack could look like:
 611 //   <start> | <main> | <foo> | tag | <bar>
 612 // This will extract the tag and keep:
 613 //   <start> | <main> | <foo> | <bar>
 614 template<typename StackTraceTy>
 615 void ExtractTagFromStack(StackTraceTy *stack, uptr *tag = nullptr) {
 616   if (stack->size < 2) return;
 617   uptr possible_tag_pc = stack->trace[stack->size - 2];
 618   uptr possible_tag = TagFromShadowStackFrame(possible_tag_pc);
 619   if (possible_tag == kExternalTagNone) return;
 620   stack->trace_buffer[stack->size - 2] = stack->trace_buffer[stack->size - 1];
 621   stack->size -= 1;
 622   if (tag) *tag = possible_tag;
 623 }
 624
 625 template<typename StackTraceTy>
 626 void ObtainCurrentStack(ThreadState *thr, uptr toppc, StackTraceTy *stack,
 627                         uptr *tag = nullptr) {
 628   uptr size = thr->shadow_stack_pos - thr->shadow_stack;
 629   uptr start = 0;
 630   if (size + !!toppc > kStackTraceMax) {
 631     start = size + !!toppc - kStackTraceMax;
 632     size = kStackTraceMax - !!toppc;
 633   }
 634   stack->Init(&thr->shadow_stack[start], size, toppc);
 635   ExtractTagFromStack(stack, tag);
 636 }
 637
 638
 639 #if TSAN_COLLECT_STATS
 640 void StatAggregate(u64 *dst, u64 *src);
 641 void StatOutput(u64 *stat);
 642 #endif
 643
 644 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 645 #if TSAN_COLLECT_STATS
 646   thr->stat[typ] += n;
 647 #endif
 648 }
 649 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 650 #if TSAN_COLLECT_STATS
 651   thr->stat[typ] = n;
 652 #endif
 653 }
 654
 655 void MapShadow(uptr addr, uptr size);
 656 void MapThreadTrace(uptr addr, uptr size, const char *name);
 657 void DontNeedShadowFor(uptr addr, uptr size);
 658 void InitializeShadowMemory();
 659 void InitializeInterceptors();
 660 void InitializeLibIgnore();
 661 void InitializeDynamicAnnotations();
 662
 663 void ForkBefore(ThreadState *thr, uptr pc);
 664 void ForkParentAfter(ThreadState *thr, uptr pc);
 665 void ForkChildAfter(ThreadState *thr, uptr pc);
 666
 667 void ReportRace(ThreadState *thr);
 668 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
 669 bool IsFiredSuppression(Context *ctx, ReportType type, StackTrace trace);
 670 bool IsExpectedReport(uptr addr, uptr size);
 671 void PrintMatchedBenignRaces();
 672
 673 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 674 # define DPrintf Printf
 675 #else
 676 # define DPrintf(...)
 677 #endif
 678
 679 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 680 # define DPrintf2 Printf
 681 #else
 682 # define DPrintf2(...)
 683 #endif
 684
 685 u32 CurrentStackId(ThreadState *thr, uptr pc);
 686 ReportStack *SymbolizeStackId(u32 stack_id);
 687 void PrintCurrentStack(ThreadState *thr, uptr pc);
 688 void PrintCurrentStackSlow(uptr pc);  // uses libunwind
 689
 690 void Initialize(ThreadState *thr);
 691 int Finalize(ThreadState *thr);
 692
 693 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
 694 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
 695
 696 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 697     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 698 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 699     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 700     u64 *shadow_mem, Shadow cur);
 701 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 702     uptr size, bool is_write);
 703 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 704     uptr size, uptr step, bool is_write);
 705 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 706     int size, bool kAccessIsWrite, bool kIsAtomic);
 707
 708 const int kSizeLog1 = 0;
 709 const int kSizeLog2 = 1;
 710 const int kSizeLog4 = 2;
 711 const int kSizeLog8 = 3;
 712
 713 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 714                                      uptr addr, int kAccessSizeLog) {
 715   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 716 }
 717
 718 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 719                                       uptr addr, int kAccessSizeLog) {
 720   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 721 }
 722
 723 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 724                                            uptr addr, int kAccessSizeLog) {
 725   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 726 }
 727
 728 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 729                                             uptr addr, int kAccessSizeLog) {
 730   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 731 }
 732
 733 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 734 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 735 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 736
 737 void ThreadIgnoreBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 738 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
 739 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 740 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
 741
 742 void FuncEntry(ThreadState *thr, uptr pc);
 743 void FuncExit(ThreadState *thr);
 744
 745 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 746 void ThreadStart(ThreadState *thr, int tid, tid_t os_id, bool workerthread);
 747 void ThreadFinish(ThreadState *thr);
 748 int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
 749 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 750 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 751 void ThreadFinalize(ThreadState *thr);
 752 void ThreadSetName(ThreadState *thr, const char *name);
 753 int ThreadCount(ThreadState *thr);
 754 void ProcessPendingSignals(ThreadState *thr);
 755
 756 Processor *ProcCreate();
 757 void ProcDestroy(Processor *proc);
 758 void ProcWire(Processor *proc, ThreadState *thr);
 759 void ProcUnwire(Processor *proc, ThreadState *thr);
 760
 761 // Note: the parameter is called flagz, because flags is already taken
 762 // by the global function that returns flags.
 763 void MutexCreate(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 764 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 765 void MutexPreLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 766 void MutexPostLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0,
 767     int rec = 1);
 768 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 769 void MutexPreReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 770 void MutexPostReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 771 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 772 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 773 void MutexRepair(ThreadState *thr, uptr pc, uptr addr);  // call on EOWNERDEAD
 774 void MutexInvalidAccess(ThreadState *thr, uptr pc, uptr addr);
 775
 776 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 777 // AcquireGlobal synchronizes the current thread with all other threads.
 778 // In terms of happens-before relation, it draws a HB edge from all threads
 779 // (where they happen to execute right now) to the current thread. We use it to
 780 // handle Go finalizers. Namely, finalizer goroutine executes AcquireGlobal
 781 // right before executing finalizers. This provides a coarse, but simple
 782 // approximation of the actual required synchronization.
 783 void AcquireGlobal(ThreadState *thr, uptr pc);
 784 void Release(ThreadState *thr, uptr pc, uptr addr);
 785 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 786 void AfterSleep(ThreadState *thr, uptr pc);
 787 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 788 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 789 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
 790 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 791
 792 // The hacky call uses custom calling convention and an assembly thunk.
 793 // It is considerably faster that a normal call for the caller
 794 // if it is not executed (it is intended for slow paths from hot functions).
 795 // The trick is that the call preserves all registers and the compiler
 796 // does not treat it as a call.
 797 // If it does not work for you, use normal call.
 798 #if !SANITIZER_DEBUG && defined(__x86_64__) && !SANITIZER_MAC
 799 // The caller may not create the stack frame for itself at all,
 800 // so we create a reserve stack frame for it (1024b must be enough).
 801 #define HACKY_CALL(f) \
 802   __asm__ __volatile__("sub $1024, %%rsp;" \
 803                        CFI_INL_ADJUST_CFA_OFFSET(1024) \
 804                        ".hidden " #f "_thunk;" \
 805                        "call " #f "_thunk;" \
 806                        "add $1024, %%rsp;" \
 807                        CFI_INL_ADJUST_CFA_OFFSET(-1024) \
 808                        ::: "memory", "cc");
 809 #else
 810 #define HACKY_CALL(f) f()
 811 #endif
 812
 813 void TraceSwitch(ThreadState *thr);
 814 uptr TraceTopPC(ThreadState *thr);
 815 uptr TraceSize();
 816 uptr TraceParts();
 817 Trace *ThreadTrace(int tid);
 818
 819 extern "C" void __tsan_trace_switch();
 820 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 821                                         EventType typ, u64 addr) {
 822   if (!kCollectHistory)
 823     return;
 824   DCHECK_GE((int)typ, 0);
 825   DCHECK_LE((int)typ, 7);
 826   DCHECK_EQ(GetLsb(addr, kEventPCBits), addr);
 827   StatInc(thr, StatEvents);
 828   u64 pos = fs.GetTracePos();
 829   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 830 #if !SANITIZER_GO
 831     HACKY_CALL(__tsan_trace_switch);
 832 #else
 833     TraceSwitch(thr);
 834 #endif
 835   }
 836   Event *trace = (Event*)GetThreadTrace(fs.tid());
 837   Event *evp = &trace[pos];
 838   Event ev = (u64)addr | ((u64)typ << kEventPCBits);
 839   *evp = ev;
 840 }
 841
 842 #if !SANITIZER_GO
 843 uptr ALWAYS_INLINE HeapEnd() {
 844   return HeapMemEnd() + PrimaryAllocator::AdditionalSize();
 845 }
 846 #endif
 847
 848 }  // namespace __tsan
 849
 850 #endif  // TSAN_RTL_H