libsanitizer/tsan/tsan_rtl.h

   1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This file is a part of ThreadSanitizer (TSan), a race detector.
  10 //
  11 // Main internal TSan header file.
  12 //
  13 // Ground rules:
  14 //   - C++ run-time should not be used (static CTORs, RTTI, exceptions, static
  15 //     function-scope locals)
  16 //   - All functions/classes/etc reside in namespace __tsan, except for those
  17 //     declared in tsan_interface.h.
  18 //   - Platform-specific files should be used instead of ifdefs (*).
  19 //   - No system headers included in header files (*).
  20 //   - Platform specific headres included only into platform-specific files (*).
  21 //
  22 //  (*) Except when inlining is critical for performance.
  23 //===----------------------------------------------------------------------===//
  24
  25 #ifndef TSAN_RTL_H
  26 #define TSAN_RTL_H
  27
  28 #include "sanitizer_common/sanitizer_allocator.h"
  29 #include "sanitizer_common/sanitizer_allocator_internal.h"
  30 #include "sanitizer_common/sanitizer_asm.h"
  31 #include "sanitizer_common/sanitizer_common.h"
  32 #include "sanitizer_common/sanitizer_deadlock_detector_interface.h"
  33 #include "sanitizer_common/sanitizer_libignore.h"
  34 #include "sanitizer_common/sanitizer_suppressions.h"
  35 #include "sanitizer_common/sanitizer_thread_registry.h"
  36 #include "sanitizer_common/sanitizer_vector.h"
  37 #include "tsan_clock.h"
  38 #include "tsan_defs.h"
  39 #include "tsan_flags.h"
  40 #include "tsan_mman.h"
  41 #include "tsan_sync.h"
  42 #include "tsan_trace.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 #if !SANITIZER_GO
  56 struct MapUnmapCallback;
  57 #if defined(__mips64) || defined(__aarch64__) || defined(__powerpc__)
  58
  59 struct AP32 {
  60   static const uptr kSpaceBeg = 0;
  61   static const u64 kSpaceSize = SANITIZER_MMAP_RANGE_SIZE;
  62   static const uptr kMetadataSize = 0;
  63   typedef __sanitizer::CompactSizeClassMap SizeClassMap;
  64   static const uptr kRegionSizeLog = 20;
  65   using AddressSpaceView = LocalAddressSpaceView;
  66   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  67   static const uptr kFlags = 0;
  68 };
  69 typedef SizeClassAllocator32<AP32> PrimaryAllocator;
  70 #else
  71 struct AP64 {  // Allocator64 parameters. Deliberately using a short name.
  72   static const uptr kSpaceBeg = Mapping::kHeapMemBeg;
  73   static const uptr kSpaceSize = Mapping::kHeapMemEnd - Mapping::kHeapMemBeg;
  74   static const uptr kMetadataSize = 0;
  75   typedef DefaultSizeClassMap SizeClassMap;
  76   typedef __tsan::MapUnmapCallback MapUnmapCallback;
  77   static const uptr kFlags = 0;
  78   using AddressSpaceView = LocalAddressSpaceView;
  79 };
  80 typedef SizeClassAllocator64<AP64> PrimaryAllocator;
  81 #endif
  82 typedef CombinedAllocator<PrimaryAllocator> Allocator;
  83 typedef Allocator::AllocatorCache AllocatorCache;
  84 Allocator *allocator();
  85 #endif
  86
  87 const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
  88
  89 // FastState (from most significant bit):
  90 //   ignore          : 1
  91 //   tid             : kTidBits
  92 //   unused          : -
  93 //   history_size    : 3
  94 //   epoch           : kClkBits
  95 class FastState {
  96  public:
  97   FastState(u64 tid, u64 epoch) {
  98     x_ = tid << kTidShift;
  99     x_ |= epoch;
 100     DCHECK_EQ(tid, this->tid());
 101     DCHECK_EQ(epoch, this->epoch());
 102     DCHECK_EQ(GetIgnoreBit(), false);
 103   }
 104
 105   explicit FastState(u64 x)
 106       : x_(x) {
 107   }
 108
 109   u64 raw() const {
 110     return x_;
 111   }
 112
 113   u64 tid() const {
 114     u64 res = (x_ & ~kIgnoreBit) >> kTidShift;
 115     return res;
 116   }
 117
 118   u64 TidWithIgnore() const {
 119     u64 res = x_ >> kTidShift;
 120     return res;
 121   }
 122
 123   u64 epoch() const {
 124     u64 res = x_ & ((1ull << kClkBits) - 1);
 125     return res;
 126   }
 127
 128   void IncrementEpoch() {
 129     u64 old_epoch = epoch();
 130     x_ += 1;
 131     DCHECK_EQ(old_epoch + 1, epoch());
 132     (void)old_epoch;
 133   }
 134
 135   void SetIgnoreBit() { x_ |= kIgnoreBit; }
 136   void ClearIgnoreBit() { x_ &= ~kIgnoreBit; }
 137   bool GetIgnoreBit() const { return (s64)x_ < 0; }
 138
 139   void SetHistorySize(int hs) {
 140     CHECK_GE(hs, 0);
 141     CHECK_LE(hs, 7);
 142     x_ = (x_ & ~(kHistoryMask << kHistoryShift)) | (u64(hs) << kHistoryShift);
 143   }
 144
 145   ALWAYS_INLINE
 146   int GetHistorySize() const {
 147     return (int)((x_ >> kHistoryShift) & kHistoryMask);
 148   }
 149
 150   void ClearHistorySize() {
 151     SetHistorySize(0);
 152   }
 153
 154   ALWAYS_INLINE
 155   u64 GetTracePos() const {
 156     const int hs = GetHistorySize();
 157     // When hs == 0, the trace consists of 2 parts.
 158     const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1;
 159     return epoch() & mask;
 160   }
 161
 162  private:
 163   friend class Shadow;
 164   static const int kTidShift = 64 - kTidBits - 1;
 165   static const u64 kIgnoreBit = 1ull << 63;
 166   static const u64 kFreedBit = 1ull << 63;
 167   static const u64 kHistoryShift = kClkBits;
 168   static const u64 kHistoryMask = 7;
 169   u64 x_;
 170 };
 171
 172 // Shadow (from most significant bit):
 173 //   freed           : 1
 174 //   tid             : kTidBits
 175 //   is_atomic       : 1
 176 //   is_read         : 1
 177 //   size_log        : 2
 178 //   addr0           : 3
 179 //   epoch           : kClkBits
 180 class Shadow : public FastState {
 181  public:
 182   explicit Shadow(u64 x)
 183       : FastState(x) {
 184   }
 185
 186   explicit Shadow(const FastState &s)
 187       : FastState(s.x_) {
 188     ClearHistorySize();
 189   }
 190
 191   void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
 192     DCHECK_EQ((x_ >> kClkBits) & 31, 0);
 193     DCHECK_LE(addr0, 7);
 194     DCHECK_LE(kAccessSizeLog, 3);
 195     x_ |= ((kAccessSizeLog << 3) | addr0) << kClkBits;
 196     DCHECK_EQ(kAccessSizeLog, size_log());
 197     DCHECK_EQ(addr0, this->addr0());
 198   }
 199
 200   void SetWrite(unsigned kAccessIsWrite) {
 201     DCHECK_EQ(x_ & kReadBit, 0);
 202     if (!kAccessIsWrite)
 203       x_ |= kReadBit;
 204     DCHECK_EQ(kAccessIsWrite, IsWrite());
 205   }
 206
 207   void SetAtomic(bool kIsAtomic) {
 208     DCHECK(!IsAtomic());
 209     if (kIsAtomic)
 210       x_ |= kAtomicBit;
 211     DCHECK_EQ(IsAtomic(), kIsAtomic);
 212   }
 213
 214   bool IsAtomic() const {
 215     return x_ & kAtomicBit;
 216   }
 217
 218   bool IsZero() const {
 219     return x_ == 0;
 220   }
 221
 222   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
 223     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
 224     DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore());
 225     return shifted_xor == 0;
 226   }
 227
 228   static ALWAYS_INLINE
 229   bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) {
 230     u64 masked_xor = ((s1.x_ ^ s2.x_) >> kClkBits) & 31;
 231     return masked_xor == 0;
 232   }
 233
 234   static ALWAYS_INLINE bool TwoRangesIntersect(Shadow s1, Shadow s2,
 235       unsigned kS2AccessSize) {
 236     bool res = false;
 237     u64 diff = s1.addr0() - s2.addr0();
 238     if ((s64)diff < 0) {  // s1.addr0 < s2.addr0
 239       // if (s1.addr0() + size1) > s2.addr0()) return true;
 240       if (s1.size() > -diff)
 241         res = true;
 242     } else {
 243       // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
 244       if (kS2AccessSize > diff)
 245         res = true;
 246     }
 247     DCHECK_EQ(res, TwoRangesIntersectSlow(s1, s2));
 248     DCHECK_EQ(res, TwoRangesIntersectSlow(s2, s1));
 249     return res;
 250   }
 251
 252   u64 ALWAYS_INLINE addr0() const { return (x_ >> kClkBits) & 7; }
 253   u64 ALWAYS_INLINE size() const { return 1ull << size_log(); }
 254   bool ALWAYS_INLINE IsWrite() const { return !IsRead(); }
 255   bool ALWAYS_INLINE IsRead() const { return x_ & kReadBit; }
 256
 257   // The idea behind the freed bit is as follows.
 258   // When the memory is freed (or otherwise unaccessible) we write to the shadow
 259   // values with tid/epoch related to the free and the freed bit set.
 260   // During memory accesses processing the freed bit is considered
 261   // as msb of tid. So any access races with shadow with freed bit set
 262   // (it is as if write from a thread with which we never synchronized before).
 263   // This allows us to detect accesses to freed memory w/o additional
 264   // overheads in memory access processing and at the same time restore
 265   // tid/epoch of free.
 266   void MarkAsFreed() {
 267      x_ |= kFreedBit;
 268   }
 269
 270   bool IsFreed() const {
 271     return x_ & kFreedBit;
 272   }
 273
 274   bool GetFreedAndReset() {
 275     bool res = x_ & kFreedBit;
 276     x_ &= ~kFreedBit;
 277     return res;
 278   }
 279
 280   bool ALWAYS_INLINE IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
 281     bool v = x_ & ((u64(kIsWrite ^ 1) << kReadShift)
 282         | (u64(kIsAtomic) << kAtomicShift));
 283     DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
 284     return v;
 285   }
 286
 287   bool ALWAYS_INLINE IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
 288     bool v = ((x_ >> kReadShift) & 3)
 289         <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 290     DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
 291         (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
 292     return v;
 293   }
 294
 295   bool ALWAYS_INLINE IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
 296     bool v = ((x_ >> kReadShift) & 3)
 297         >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
 298     DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
 299         (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
 300     return v;
 301   }
 302
 303  private:
 304   static const u64 kReadShift   = 5 + kClkBits;
 305   static const u64 kReadBit     = 1ull << kReadShift;
 306   static const u64 kAtomicShift = 6 + kClkBits;
 307   static const u64 kAtomicBit   = 1ull << kAtomicShift;
 308
 309   u64 size_log() const { return (x_ >> (3 + kClkBits)) & 3; }
 310
 311   static bool TwoRangesIntersectSlow(const Shadow s1, const Shadow s2) {
 312     if (s1.addr0() == s2.addr0()) return true;
 313     if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
 314       return true;
 315     if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
 316       return true;
 317     return false;
 318   }
 319 };
 320
 321 struct ThreadSignalContext;
 322
 323 struct JmpBuf {
 324   uptr sp;
 325   int int_signal_send;
 326   bool in_blocking_func;
 327   uptr in_signal_handler;
 328   uptr *shadow_stack_pos;
 329 };
 330
 331 // A Processor represents a physical thread, or a P for Go.
 332 // It is used to store internal resources like allocate cache, and does not
 333 // participate in race-detection logic (invisible to end user).
 334 // In C++ it is tied to an OS thread just like ThreadState, however ideally
 335 // it should be tied to a CPU (this way we will have fewer allocator caches).
 336 // In Go it is tied to a P, so there are significantly fewer Processor's than
 337 // ThreadState's (which are tied to Gs).
 338 // A ThreadState must be wired with a Processor to handle events.
 339 struct Processor {
 340   ThreadState *thr; // currently wired thread, or nullptr
 341 #if !SANITIZER_GO
 342   AllocatorCache alloc_cache;
 343   InternalAllocatorCache internal_alloc_cache;
 344 #endif
 345   DenseSlabAllocCache block_cache;
 346   DenseSlabAllocCache sync_cache;
 347   DenseSlabAllocCache clock_cache;
 348   DDPhysicalThread *dd_pt;
 349 };
 350
 351 #if !SANITIZER_GO
 352 // ScopedGlobalProcessor temporary setups a global processor for the current
 353 // thread, if it does not have one. Intended for interceptors that can run
 354 // at the very thread end, when we already destroyed the thread processor.
 355 struct ScopedGlobalProcessor {
 356   ScopedGlobalProcessor();
 357   ~ScopedGlobalProcessor();
 358 };
 359 #endif
 360
 361 // This struct is stored in TLS.
 362 struct ThreadState {
 363   FastState fast_state;
 364   // Synch epoch represents the threads's epoch before the last synchronization
 365   // action. It allows to reduce number of shadow state updates.
 366   // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
 367   // if we are processing write to X from the same thread at epoch=200,
 368   // we do nothing, because both writes happen in the same 'synch epoch'.
 369   // That is, if another memory access does not race with the former write,
 370   // it does not race with the latter as well.
 371   // QUESTION: can we can squeeze this into ThreadState::Fast?
 372   // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
 373   // taken by epoch between synchs.
 374   // This way we can save one load from tls.
 375   u64 fast_synch_epoch;
 376   // Technically `current` should be a separate THREADLOCAL variable;
 377   // but it is placed here in order to share cache line with previous fields.
 378   ThreadState* current;
 379   // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
 380   // We do not distinguish beteween ignoring reads and writes
 381   // for better performance.
 382   int ignore_reads_and_writes;
 383   int ignore_sync;
 384   int suppress_reports;
 385   // Go does not support ignores.
 386 #if !SANITIZER_GO
 387   IgnoreSet mop_ignore_set;
 388   IgnoreSet sync_ignore_set;
 389 #endif
 390   // C/C++ uses fixed size shadow stack embed into Trace.
 391   // Go uses malloc-allocated shadow stack with dynamic size.
 392   uptr *shadow_stack;
 393   uptr *shadow_stack_end;
 394   uptr *shadow_stack_pos;
 395   u64 *racy_shadow_addr;
 396   u64 racy_state[2];
 397   MutexSet mset;
 398   ThreadClock clock;
 399 #if !SANITIZER_GO
 400   Vector<JmpBuf> jmp_bufs;
 401   int ignore_interceptors;
 402 #endif
 403 #if TSAN_COLLECT_STATS
 404   u64 stat[StatCnt];
 405 #endif
 406   const u32 tid;
 407   const int unique_id;
 408   bool in_symbolizer;
 409   bool in_ignored_lib;
 410   bool is_inited;
 411   bool is_dead;
 412   bool is_freeing;
 413   bool is_vptr_access;
 414   const uptr stk_addr;
 415   const uptr stk_size;
 416   const uptr tls_addr;
 417   const uptr tls_size;
 418   ThreadContext *tctx;
 419
 420 #if SANITIZER_DEBUG && !SANITIZER_GO
 421   InternalDeadlockDetector internal_deadlock_detector;
 422 #endif
 423   DDLogicalThread *dd_lt;
 424
 425   // Current wired Processor, or nullptr. Required to handle any events.
 426   Processor *proc1;
 427 #if !SANITIZER_GO
 428   Processor *proc() { return proc1; }
 429 #else
 430   Processor *proc();
 431 #endif
 432
 433   atomic_uintptr_t in_signal_handler;
 434   ThreadSignalContext *signal_ctx;
 435
 436 #if !SANITIZER_GO
 437   u32 last_sleep_stack_id;
 438   ThreadClock last_sleep_clock;
 439 #endif
 440
 441   // Set in regions of runtime that must be signal-safe and fork-safe.
 442   // If set, malloc must not be called.
 443   int nomalloc;
 444
 445   const ReportDesc *current_report;
 446
 447   explicit ThreadState(Context *ctx, u32 tid, int unique_id, u64 epoch,
 448                        unsigned reuse_count, uptr stk_addr, uptr stk_size,
 449                        uptr tls_addr, uptr tls_size);
 450 };
 451
 452 #if !SANITIZER_GO
 453 #if SANITIZER_MAC || SANITIZER_ANDROID
 454 ThreadState *cur_thread();
 455 void set_cur_thread(ThreadState *thr);
 456 void cur_thread_finalize();
 457 inline void cur_thread_init() { }
 458 #else
 459 __attribute__((tls_model("initial-exec")))
 460 extern THREADLOCAL char cur_thread_placeholder[];
 461 inline ThreadState *cur_thread() {
 462   return reinterpret_cast<ThreadState *>(cur_thread_placeholder)->current;
 463 }
 464 inline void cur_thread_init() {
 465   ThreadState *thr = reinterpret_cast<ThreadState *>(cur_thread_placeholder);
 466   if (UNLIKELY(!thr->current))
 467     thr->current = thr;
 468 }
 469 inline void set_cur_thread(ThreadState *thr) {
 470   reinterpret_cast<ThreadState *>(cur_thread_placeholder)->current = thr;
 471 }
 472 inline void cur_thread_finalize() { }
 473 #endif  // SANITIZER_MAC || SANITIZER_ANDROID
 474 #endif  // SANITIZER_GO
 475
 476 class ThreadContext final : public ThreadContextBase {
 477  public:
 478   explicit ThreadContext(int tid);
 479   ~ThreadContext();
 480   ThreadState *thr;
 481   u32 creation_stack_id;
 482   SyncClock sync;
 483   // Epoch at which the thread had started.
 484   // If we see an event from the thread stamped by an older epoch,
 485   // the event is from a dead thread that shared tid with this thread.
 486   u64 epoch0;
 487   u64 epoch1;
 488
 489   // Override superclass callbacks.
 490   void OnDead() override;
 491   void OnJoined(void *arg) override;
 492   void OnFinished() override;
 493   void OnStarted(void *arg) override;
 494   void OnCreated(void *arg) override;
 495   void OnReset() override;
 496   void OnDetached(void *arg) override;
 497 };
 498
 499 struct RacyStacks {
 500   MD5Hash hash[2];
 501   bool operator==(const RacyStacks &other) const {
 502     if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
 503       return true;
 504     if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
 505       return true;
 506     return false;
 507   }
 508 };
 509
 510 struct RacyAddress {
 511   uptr addr_min;
 512   uptr addr_max;
 513 };
 514
 515 struct FiredSuppression {
 516   ReportType type;
 517   uptr pc_or_addr;
 518   Suppression *supp;
 519 };
 520
 521 struct Context {
 522   Context();
 523
 524   bool initialized;
 525 #if !SANITIZER_GO
 526   bool after_multithreaded_fork;
 527 #endif
 528
 529   MetaMap metamap;
 530
 531   Mutex report_mtx;
 532   int nreported;
 533   int nmissed_expected;
 534   atomic_uint64_t last_symbolize_time_ns;
 535
 536   void *background_thread;
 537   atomic_uint32_t stop_background_thread;
 538
 539   ThreadRegistry *thread_registry;
 540
 541   Mutex racy_mtx;
 542   Vector<RacyStacks> racy_stacks;
 543   Vector<RacyAddress> racy_addresses;
 544   // Number of fired suppressions may be large enough.
 545   Mutex fired_suppressions_mtx;
 546   InternalMmapVector<FiredSuppression> fired_suppressions;
 547   DDetector *dd;
 548
 549   ClockAlloc clock_alloc;
 550
 551   Flags flags;
 552
 553   u64 stat[StatCnt];
 554   u64 int_alloc_cnt[MBlockTypeCount];
 555   u64 int_alloc_siz[MBlockTypeCount];
 556 };
 557
 558 extern Context *ctx;  // The one and the only global runtime context.
 559
 560 ALWAYS_INLINE Flags *flags() {
 561   return &ctx->flags;
 562 }
 563
 564 struct ScopedIgnoreInterceptors {
 565   ScopedIgnoreInterceptors() {
 566 #if !SANITIZER_GO
 567     cur_thread()->ignore_interceptors++;
 568 #endif
 569   }
 570
 571   ~ScopedIgnoreInterceptors() {
 572 #if !SANITIZER_GO
 573     cur_thread()->ignore_interceptors--;
 574 #endif
 575   }
 576 };
 577
 578 const char *GetObjectTypeFromTag(uptr tag);
 579 const char *GetReportHeaderFromTag(uptr tag);
 580 uptr TagFromShadowStackFrame(uptr pc);
 581
 582 class ScopedReportBase {
 583  public:
 584   void AddMemoryAccess(uptr addr, uptr external_tag, Shadow s, StackTrace stack,
 585                        const MutexSet *mset);
 586   void AddStack(StackTrace stack, bool suppressable = false);
 587   void AddThread(const ThreadContext *tctx, bool suppressable = false);
 588   void AddThread(int unique_tid, bool suppressable = false);
 589   void AddUniqueTid(int unique_tid);
 590   void AddMutex(const SyncVar *s);
 591   u64 AddMutex(u64 id);
 592   void AddLocation(uptr addr, uptr size);
 593   void AddSleep(u32 stack_id);
 594   void SetCount(int count);
 595
 596   const ReportDesc *GetReport() const;
 597
 598  protected:
 599   ScopedReportBase(ReportType typ, uptr tag);
 600   ~ScopedReportBase();
 601
 602  private:
 603   ReportDesc *rep_;
 604   // Symbolizer makes lots of intercepted calls. If we try to process them,
 605   // at best it will cause deadlocks on internal mutexes.
 606   ScopedIgnoreInterceptors ignore_interceptors_;
 607
 608   void AddDeadMutex(u64 id);
 609
 610   ScopedReportBase(const ScopedReportBase &) = delete;
 611   void operator=(const ScopedReportBase &) = delete;
 612 };
 613
 614 class ScopedReport : public ScopedReportBase {
 615  public:
 616   explicit ScopedReport(ReportType typ, uptr tag = kExternalTagNone);
 617   ~ScopedReport();
 618
 619  private:
 620   ScopedErrorReportLock lock_;
 621 };
 622
 623 bool ShouldReport(ThreadState *thr, ReportType typ);
 624 ThreadContext *IsThreadStackOrTls(uptr addr, bool *is_stack);
 625 void RestoreStack(int tid, const u64 epoch, VarSizeStackTrace *stk,
 626                   MutexSet *mset, uptr *tag = nullptr);
 627
 628 // The stack could look like:
 629 //   <start> | <main> | <foo> | tag | <bar>
 630 // This will extract the tag and keep:
 631 //   <start> | <main> | <foo> | <bar>
 632 template<typename StackTraceTy>
 633 void ExtractTagFromStack(StackTraceTy *stack, uptr *tag = nullptr) {
 634   if (stack->size < 2) return;
 635   uptr possible_tag_pc = stack->trace[stack->size - 2];
 636   uptr possible_tag = TagFromShadowStackFrame(possible_tag_pc);
 637   if (possible_tag == kExternalTagNone) return;
 638   stack->trace_buffer[stack->size - 2] = stack->trace_buffer[stack->size - 1];
 639   stack->size -= 1;
 640   if (tag) *tag = possible_tag;
 641 }
 642
 643 template<typename StackTraceTy>
 644 void ObtainCurrentStack(ThreadState *thr, uptr toppc, StackTraceTy *stack,
 645                         uptr *tag = nullptr) {
 646   uptr size = thr->shadow_stack_pos - thr->shadow_stack;
 647   uptr start = 0;
 648   if (size + !!toppc > kStackTraceMax) {
 649     start = size + !!toppc - kStackTraceMax;
 650     size = kStackTraceMax - !!toppc;
 651   }
 652   stack->Init(&thr->shadow_stack[start], size, toppc);
 653   ExtractTagFromStack(stack, tag);
 654 }
 655
 656 #define GET_STACK_TRACE_FATAL(thr, pc) \
 657   VarSizeStackTrace stack; \
 658   ObtainCurrentStack(thr, pc, &stack); \
 659   stack.ReverseOrder();
 660
 661 #if TSAN_COLLECT_STATS
 662 void StatAggregate(u64 *dst, u64 *src);
 663 void StatOutput(u64 *stat);
 664 #endif
 665
 666 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
 667 #if TSAN_COLLECT_STATS
 668   thr->stat[typ] += n;
 669 #endif
 670 }
 671 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
 672 #if TSAN_COLLECT_STATS
 673   thr->stat[typ] = n;
 674 #endif
 675 }
 676
 677 void MapShadow(uptr addr, uptr size);
 678 void MapThreadTrace(uptr addr, uptr size, const char *name);
 679 void DontNeedShadowFor(uptr addr, uptr size);
 680 void UnmapShadow(ThreadState *thr, uptr addr, uptr size);
 681 void InitializeShadowMemory();
 682 void InitializeInterceptors();
 683 void InitializeLibIgnore();
 684 void InitializeDynamicAnnotations();
 685
 686 void ForkBefore(ThreadState *thr, uptr pc);
 687 void ForkParentAfter(ThreadState *thr, uptr pc);
 688 void ForkChildAfter(ThreadState *thr, uptr pc);
 689
 690 void ReportRace(ThreadState *thr);
 691 bool OutputReport(ThreadState *thr, const ScopedReport &srep);
 692 bool IsFiredSuppression(Context *ctx, ReportType type, StackTrace trace);
 693 bool IsExpectedReport(uptr addr, uptr size);
 694 void PrintMatchedBenignRaces();
 695
 696 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 697 # define DPrintf Printf
 698 #else
 699 # define DPrintf(...)
 700 #endif
 701
 702 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
 703 # define DPrintf2 Printf
 704 #else
 705 # define DPrintf2(...)
 706 #endif
 707
 708 u32 CurrentStackId(ThreadState *thr, uptr pc);
 709 ReportStack *SymbolizeStackId(u32 stack_id);
 710 void PrintCurrentStack(ThreadState *thr, uptr pc);
 711 void PrintCurrentStackSlow(uptr pc);  // uses libunwind
 712
 713 void Initialize(ThreadState *thr);
 714 void MaybeSpawnBackgroundThread();
 715 int Finalize(ThreadState *thr);
 716
 717 void OnUserAlloc(ThreadState *thr, uptr pc, uptr p, uptr sz, bool write);
 718 void OnUserFree(ThreadState *thr, uptr pc, uptr p, bool write);
 719
 720 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 721     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 722 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 723     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 724     u64 *shadow_mem, Shadow cur);
 725 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 726     uptr size, bool is_write);
 727 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
 728     uptr size, uptr step, bool is_write);
 729 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 730     int size, bool kAccessIsWrite, bool kIsAtomic);
 731
 732 const int kSizeLog1 = 0;
 733 const int kSizeLog2 = 1;
 734 const int kSizeLog4 = 2;
 735 const int kSizeLog8 = 3;
 736
 737 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
 738                                      uptr addr, int kAccessSizeLog) {
 739   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
 740 }
 741
 742 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
 743                                       uptr addr, int kAccessSizeLog) {
 744   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
 745 }
 746
 747 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
 748                                            uptr addr, int kAccessSizeLog) {
 749   MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
 750 }
 751
 752 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
 753                                             uptr addr, int kAccessSizeLog) {
 754   MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
 755 }
 756
 757 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 758 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 759 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
 760 void MemoryRangeImitateWriteOrResetRange(ThreadState *thr, uptr pc, uptr addr,
 761                                          uptr size);
 762
 763 void ThreadIgnoreBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 764 void ThreadIgnoreEnd(ThreadState *thr, uptr pc);
 765 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack = true);
 766 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc);
 767
 768 void FuncEntry(ThreadState *thr, uptr pc);
 769 void FuncExit(ThreadState *thr);
 770
 771 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
 772 void ThreadStart(ThreadState *thr, int tid, tid_t os_id,
 773                  ThreadType thread_type);
 774 void ThreadFinish(ThreadState *thr);
 775 int ThreadConsumeTid(ThreadState *thr, uptr pc, uptr uid);
 776 void ThreadJoin(ThreadState *thr, uptr pc, int tid);
 777 void ThreadDetach(ThreadState *thr, uptr pc, int tid);
 778 void ThreadFinalize(ThreadState *thr);
 779 void ThreadSetName(ThreadState *thr, const char *name);
 780 int ThreadCount(ThreadState *thr);
 781 void ProcessPendingSignals(ThreadState *thr);
 782 void ThreadNotJoined(ThreadState *thr, uptr pc, int tid, uptr uid);
 783
 784 Processor *ProcCreate();
 785 void ProcDestroy(Processor *proc);
 786 void ProcWire(Processor *proc, ThreadState *thr);
 787 void ProcUnwire(Processor *proc, ThreadState *thr);
 788
 789 // Note: the parameter is called flagz, because flags is already taken
 790 // by the global function that returns flags.
 791 void MutexCreate(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 792 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 793 void MutexPreLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 794 void MutexPostLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0,
 795     int rec = 1);
 796 int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 797 void MutexPreReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 798 void MutexPostReadLock(ThreadState *thr, uptr pc, uptr addr, u32 flagz = 0);
 799 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 800 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
 801 void MutexRepair(ThreadState *thr, uptr pc, uptr addr);  // call on EOWNERDEAD
 802 void MutexInvalidAccess(ThreadState *thr, uptr pc, uptr addr);
 803
 804 void Acquire(ThreadState *thr, uptr pc, uptr addr);
 805 // AcquireGlobal synchronizes the current thread with all other threads.
 806 // In terms of happens-before relation, it draws a HB edge from all threads
 807 // (where they happen to execute right now) to the current thread. We use it to
 808 // handle Go finalizers. Namely, finalizer goroutine executes AcquireGlobal
 809 // right before executing finalizers. This provides a coarse, but simple
 810 // approximation of the actual required synchronization.
 811 void AcquireGlobal(ThreadState *thr, uptr pc);
 812 void Release(ThreadState *thr, uptr pc, uptr addr);
 813 void ReleaseStoreAcquire(ThreadState *thr, uptr pc, uptr addr);
 814 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr);
 815 void AfterSleep(ThreadState *thr, uptr pc);
 816 void AcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 817 void ReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 818 void ReleaseStoreAcquireImpl(ThreadState *thr, uptr pc, SyncClock *c);
 819 void ReleaseStoreImpl(ThreadState *thr, uptr pc, SyncClock *c);
 820 void AcquireReleaseImpl(ThreadState *thr, uptr pc, SyncClock *c);
 821
 822 // The hacky call uses custom calling convention and an assembly thunk.
 823 // It is considerably faster that a normal call for the caller
 824 // if it is not executed (it is intended for slow paths from hot functions).
 825 // The trick is that the call preserves all registers and the compiler
 826 // does not treat it as a call.
 827 // If it does not work for you, use normal call.
 828 #if !SANITIZER_DEBUG && defined(__x86_64__) && !SANITIZER_MAC
 829 // The caller may not create the stack frame for itself at all,
 830 // so we create a reserve stack frame for it (1024b must be enough).
 831 #define HACKY_CALL(f) \
 832   __asm__ __volatile__("sub $1024, %%rsp;" \
 833                        CFI_INL_ADJUST_CFA_OFFSET(1024) \
 834                        ".hidden " #f "_thunk;" \
 835                        "call " #f "_thunk;" \
 836                        "add $1024, %%rsp;" \
 837                        CFI_INL_ADJUST_CFA_OFFSET(-1024) \
 838                        ::: "memory", "cc");
 839 #else
 840 #define HACKY_CALL(f) f()
 841 #endif
 842
 843 void TraceSwitch(ThreadState *thr);
 844 uptr TraceTopPC(ThreadState *thr);
 845 uptr TraceSize();
 846 uptr TraceParts();
 847 Trace *ThreadTrace(int tid);
 848
 849 extern "C" void __tsan_trace_switch();
 850 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
 851                                         EventType typ, u64 addr) {
 852   if (!kCollectHistory)
 853     return;
 854   DCHECK_GE((int)typ, 0);
 855   DCHECK_LE((int)typ, 7);
 856   DCHECK_EQ(GetLsb(addr, kEventPCBits), addr);
 857   StatInc(thr, StatEvents);
 858   u64 pos = fs.GetTracePos();
 859   if (UNLIKELY((pos % kTracePartSize) == 0)) {
 860 #if !SANITIZER_GO
 861     HACKY_CALL(__tsan_trace_switch);
 862 #else
 863     TraceSwitch(thr);
 864 #endif
 865   }
 866   Event *trace = (Event*)GetThreadTrace(fs.tid());
 867   Event *evp = &trace[pos];
 868   Event ev = (u64)addr | ((u64)typ << kEventPCBits);
 869   *evp = ev;
 870 }
 871
 872 #if !SANITIZER_GO
 873 uptr ALWAYS_INLINE HeapEnd() {
 874   return HeapMemEnd() + PrimaryAllocator::AdditionalSize();
 875 }
 876 #endif
 877
 878 ThreadState *FiberCreate(ThreadState *thr, uptr pc, unsigned flags);
 879 void FiberDestroy(ThreadState *thr, uptr pc, ThreadState *fiber);
 880 void FiberSwitch(ThreadState *thr, uptr pc, ThreadState *fiber, unsigned flags);
 881
 882 // These need to match __tsan_switch_to_fiber_* flags defined in
 883 // tsan_interface.h. See documentation there as well.
 884 enum FiberSwitchFlags {
 885   FiberSwitchFlagNoSync = 1 << 0, // __tsan_switch_to_fiber_no_sync
 886 };
 887
 888 }  // namespace __tsan
 889
 890 #endif  // TSAN_RTL_H