1 //===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file is a part of ThreadSanitizer (TSan), a race detector. 11 // 12 // Main internal TSan header file. 13 // 14 // Ground rules: 15 // - C++ run-time should not be used (static CTORs, RTTI, exceptions, static 16 // function-scope locals) 17 // - All functions/classes/etc reside in namespace __tsan, except for those 18 // declared in tsan_interface.h. 19 // - Platform-specific files should be used instead of ifdefs (*). 20 // - No system headers included in header files (*). 21 // - Platform specific headres included only into platform-specific files (*). 22 // 23 // (*) Except when inlining is critical for performance. 24 //===----------------------------------------------------------------------===// 25 26 #ifndef TSAN_RTL_H 27 #define TSAN_RTL_H 28 29 #include "sanitizer_common/sanitizer_allocator.h" 30 #include "sanitizer_common/sanitizer_allocator_internal.h" 31 #include "sanitizer_common/sanitizer_common.h" 32 #include "sanitizer_common/sanitizer_suppressions.h" 33 #include "sanitizer_common/sanitizer_thread_registry.h" 34 #include "tsan_clock.h" 35 #include "tsan_defs.h" 36 #include "tsan_flags.h" 37 #include "tsan_sync.h" 38 #include "tsan_trace.h" 39 #include "tsan_vector.h" 40 #include "tsan_report.h" 41 #include "tsan_platform.h" 42 #include "tsan_mutexset.h" 43 44 #if SANITIZER_WORDSIZE != 64 45 # error "ThreadSanitizer is supported only on 64-bit platforms" 46 #endif 47 48 namespace __tsan { 49 50 // Descriptor of user's memory block. 51 struct MBlock { 52 /* 53 u64 mtx : 1; // must be first 54 u64 lst : 44; 55 u64 stk : 31; // on word boundary 56 u64 tid : kTidBits; 57 u64 siz : 128 - 1 - 31 - 44 - kTidBits; // 39 58 */ 59 u64 raw[2]; 60 61 void Init(uptr siz, u32 tid, u32 stk) { 62 raw[0] = raw[1] = 0; 63 raw[1] |= (u64)siz << ((1 + 44 + 31 + kTidBits) % 64); 64 raw[1] |= (u64)tid << ((1 + 44 + 31) % 64); 65 raw[0] |= (u64)stk << (1 + 44); 66 raw[1] |= (u64)stk >> (64 - 44 - 1); 67 DCHECK_EQ(Size(), siz); 68 DCHECK_EQ(Tid(), tid); 69 DCHECK_EQ(StackId(), stk); 70 } 71 72 u32 Tid() const { 73 return GetLsb(raw[1] >> ((1 + 44 + 31) % 64), kTidBits); 74 } 75 76 uptr Size() const { 77 return raw[1] >> ((1 + 31 + 44 + kTidBits) % 64); 78 } 79 80 u32 StackId() const { 81 return (raw[0] >> (1 + 44)) | GetLsb(raw[1] << (64 - 44 - 1), 31); 82 } 83 84 SyncVar *ListHead() const { 85 return (SyncVar*)(GetLsb(raw[0] >> 1, 44) << 3); 86 } 87 88 void ListPush(SyncVar *v) { 89 SyncVar *lst = ListHead(); 90 v->next = lst; 91 u64 x = (u64)v ^ (u64)lst; 92 x = (x >> 3) << 1; 93 raw[0] ^= x; 94 DCHECK_EQ(ListHead(), v); 95 } 96 97 SyncVar *ListPop() { 98 SyncVar *lst = ListHead(); 99 SyncVar *nxt = lst->next; 100 lst->next = 0; 101 u64 x = (u64)lst ^ (u64)nxt; 102 x = (x >> 3) << 1; 103 raw[0] ^= x; 104 DCHECK_EQ(ListHead(), nxt); 105 return lst; 106 } 107 108 void ListReset() { 109 SyncVar *lst = ListHead(); 110 u64 x = (u64)lst; 111 x = (x >> 3) << 1; 112 raw[0] ^= x; 113 DCHECK_EQ(ListHead(), 0); 114 } 115 116 void Lock(); 117 void Unlock(); 118 typedef GenericScopedLock<MBlock> ScopedLock; 119 }; 120 121 #ifndef TSAN_GO 122 #if defined(TSAN_COMPAT_SHADOW) && TSAN_COMPAT_SHADOW 123 const uptr kAllocatorSpace = 0x7d0000000000ULL; 124 #else 125 const uptr kAllocatorSpace = 0x7d0000000000ULL; 126 #endif 127 const uptr kAllocatorSize = 0x10000000000ULL; // 1T. 128 129 struct MapUnmapCallback; 130 typedef SizeClassAllocator64<kAllocatorSpace, kAllocatorSize, sizeof(MBlock), 131 DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator; 132 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache; 133 typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator; 134 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache, 135 SecondaryAllocator> Allocator; 136 Allocator *allocator(); 137 #endif 138 139 void TsanCheckFailed(const char *file, int line, const char *cond, 140 u64 v1, u64 v2); 141 142 const u64 kShadowRodata = (u64)-1; // .rodata shadow marker 143 144 // FastState (from most significant bit): 145 // ignore : 1 146 // tid : kTidBits 147 // epoch : kClkBits 148 // unused : - 149 // history_size : 3 150 class FastState { 151 public: 152 FastState(u64 tid, u64 epoch) { 153 x_ = tid << kTidShift; 154 x_ |= epoch << kClkShift; 155 DCHECK_EQ(tid, this->tid()); 156 DCHECK_EQ(epoch, this->epoch()); 157 DCHECK_EQ(GetIgnoreBit(), false); 158 } 159 160 explicit FastState(u64 x) 161 : x_(x) { 162 } 163 164 u64 raw() const { 165 return x_; 166 } 167 168 u64 tid() const { 169 u64 res = (x_ & ~kIgnoreBit) >> kTidShift; 170 return res; 171 } 172 173 u64 TidWithIgnore() const { 174 u64 res = x_ >> kTidShift; 175 return res; 176 } 177 178 u64 epoch() const { 179 u64 res = (x_ << (kTidBits + 1)) >> (64 - kClkBits); 180 return res; 181 } 182 183 void IncrementEpoch() { 184 u64 old_epoch = epoch(); 185 x_ += 1 << kClkShift; 186 DCHECK_EQ(old_epoch + 1, epoch()); 187 (void)old_epoch; 188 } 189 190 void SetIgnoreBit() { x_ |= kIgnoreBit; } 191 void ClearIgnoreBit() { x_ &= ~kIgnoreBit; } 192 bool GetIgnoreBit() const { return (s64)x_ < 0; } 193 194 void SetHistorySize(int hs) { 195 CHECK_GE(hs, 0); 196 CHECK_LE(hs, 7); 197 x_ = (x_ & ~7) | hs; 198 } 199 200 int GetHistorySize() const { 201 return (int)(x_ & 7); 202 } 203 204 void ClearHistorySize() { 205 x_ &= ~7; 206 } 207 208 u64 GetTracePos() const { 209 const int hs = GetHistorySize(); 210 // When hs == 0, the trace consists of 2 parts. 211 const u64 mask = (1ull << (kTracePartSizeBits + hs + 1)) - 1; 212 return epoch() & mask; 213 } 214 215 private: 216 friend class Shadow; 217 static const int kTidShift = 64 - kTidBits - 1; 218 static const int kClkShift = kTidShift - kClkBits; 219 static const u64 kIgnoreBit = 1ull << 63; 220 static const u64 kFreedBit = 1ull << 63; 221 u64 x_; 222 }; 223 224 // Shadow (from most significant bit): 225 // freed : 1 226 // tid : kTidBits 227 // epoch : kClkBits 228 // is_atomic : 1 229 // is_read : 1 230 // size_log : 2 231 // addr0 : 3 232 class Shadow : public FastState { 233 public: 234 explicit Shadow(u64 x) 235 : FastState(x) { 236 } 237 238 explicit Shadow(const FastState &s) 239 : FastState(s.x_) { 240 ClearHistorySize(); 241 } 242 243 void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) { 244 DCHECK_EQ(x_ & 31, 0); 245 DCHECK_LE(addr0, 7); 246 DCHECK_LE(kAccessSizeLog, 3); 247 x_ |= (kAccessSizeLog << 3) | addr0; 248 DCHECK_EQ(kAccessSizeLog, size_log()); 249 DCHECK_EQ(addr0, this->addr0()); 250 } 251 252 void SetWrite(unsigned kAccessIsWrite) { 253 DCHECK_EQ(x_ & kReadBit, 0); 254 if (!kAccessIsWrite) 255 x_ |= kReadBit; 256 DCHECK_EQ(kAccessIsWrite, IsWrite()); 257 } 258 259 void SetAtomic(bool kIsAtomic) { 260 DCHECK(!IsAtomic()); 261 if (kIsAtomic) 262 x_ |= kAtomicBit; 263 DCHECK_EQ(IsAtomic(), kIsAtomic); 264 } 265 266 bool IsAtomic() const { 267 return x_ & kAtomicBit; 268 } 269 270 bool IsZero() const { 271 return x_ == 0; 272 } 273 274 static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) { 275 u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift; 276 DCHECK_EQ(shifted_xor == 0, s1.TidWithIgnore() == s2.TidWithIgnore()); 277 return shifted_xor == 0; 278 } 279 280 static inline bool Addr0AndSizeAreEqual(const Shadow s1, const Shadow s2) { 281 u64 masked_xor = (s1.x_ ^ s2.x_) & 31; 282 return masked_xor == 0; 283 } 284 285 static inline bool TwoRangesIntersect(Shadow s1, Shadow s2, 286 unsigned kS2AccessSize) { 287 bool res = false; 288 u64 diff = s1.addr0() - s2.addr0(); 289 if ((s64)diff < 0) { // s1.addr0 < s2.addr0 // NOLINT 290 // if (s1.addr0() + size1) > s2.addr0()) return true; 291 if (s1.size() > -diff) res = true; 292 } else { 293 // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true; 294 if (kS2AccessSize > diff) res = true; 295 } 296 DCHECK_EQ(res, TwoRangesIntersectSLOW(s1, s2)); 297 DCHECK_EQ(res, TwoRangesIntersectSLOW(s2, s1)); 298 return res; 299 } 300 301 // The idea behind the offset is as follows. 302 // Consider that we have 8 bool's contained within a single 8-byte block 303 // (mapped to a single shadow "cell"). Now consider that we write to the bools 304 // from a single thread (which we consider the common case). 305 // W/o offsetting each access will have to scan 4 shadow values at average 306 // to find the corresponding shadow value for the bool. 307 // With offsetting we start scanning shadow with the offset so that 308 // each access hits necessary shadow straight off (at least in an expected 309 // optimistic case). 310 // This logic works seamlessly for any layout of user data. For example, 311 // if user data is {int, short, char, char}, then accesses to the int are 312 // offsetted to 0, short - 4, 1st char - 6, 2nd char - 7. Hopefully, accesses 313 // from a single thread won't need to scan all 8 shadow values. 314 unsigned ComputeSearchOffset() { 315 return x_ & 7; 316 } 317 u64 addr0() const { return x_ & 7; } 318 u64 size() const { return 1ull << size_log(); } 319 bool IsWrite() const { return !IsRead(); } 320 bool IsRead() const { return x_ & kReadBit; } 321 322 // The idea behind the freed bit is as follows. 323 // When the memory is freed (or otherwise unaccessible) we write to the shadow 324 // values with tid/epoch related to the free and the freed bit set. 325 // During memory accesses processing the freed bit is considered 326 // as msb of tid. So any access races with shadow with freed bit set 327 // (it is as if write from a thread with which we never synchronized before). 328 // This allows us to detect accesses to freed memory w/o additional 329 // overheads in memory access processing and at the same time restore 330 // tid/epoch of free. 331 void MarkAsFreed() { 332 x_ |= kFreedBit; 333 } 334 335 bool IsFreed() const { 336 return x_ & kFreedBit; 337 } 338 339 bool GetFreedAndReset() { 340 bool res = x_ & kFreedBit; 341 x_ &= ~kFreedBit; 342 return res; 343 } 344 345 bool IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const { 346 // analyzes 5-th bit (is_read) and 6-th bit (is_atomic) 347 bool v = x_ & u64(((kIsWrite ^ 1) << kReadShift) 348 | (kIsAtomic << kAtomicShift)); 349 DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic)); 350 return v; 351 } 352 353 bool IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const { 354 bool v = ((x_ >> kReadShift) & 3) 355 <= u64((kIsWrite ^ 1) | (kIsAtomic << 1)); 356 DCHECK_EQ(v, (IsAtomic() < kIsAtomic) || 357 (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite)); 358 return v; 359 } 360 361 bool IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const { 362 bool v = ((x_ >> kReadShift) & 3) 363 >= u64((kIsWrite ^ 1) | (kIsAtomic << 1)); 364 DCHECK_EQ(v, (IsAtomic() > kIsAtomic) || 365 (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite)); 366 return v; 367 } 368 369 private: 370 static const u64 kReadShift = 5; 371 static const u64 kReadBit = 1ull << kReadShift; 372 static const u64 kAtomicShift = 6; 373 static const u64 kAtomicBit = 1ull << kAtomicShift; 374 375 u64 size_log() const { return (x_ >> 3) & 3; } 376 377 static bool TwoRangesIntersectSLOW(const Shadow s1, const Shadow s2) { 378 if (s1.addr0() == s2.addr0()) return true; 379 if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0()) 380 return true; 381 if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0()) 382 return true; 383 return false; 384 } 385 }; 386 387 struct SignalContext; 388 389 struct JmpBuf { 390 uptr sp; 391 uptr mangled_sp; 392 uptr *shadow_stack_pos; 393 }; 394 395 // This struct is stored in TLS. 396 struct ThreadState { 397 FastState fast_state; 398 // Synch epoch represents the threads's epoch before the last synchronization 399 // action. It allows to reduce number of shadow state updates. 400 // For example, fast_synch_epoch=100, last write to addr X was at epoch=150, 401 // if we are processing write to X from the same thread at epoch=200, 402 // we do nothing, because both writes happen in the same 'synch epoch'. 403 // That is, if another memory access does not race with the former write, 404 // it does not race with the latter as well. 405 // QUESTION: can we can squeeze this into ThreadState::Fast? 406 // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are 407 // taken by epoch between synchs. 408 // This way we can save one load from tls. 409 u64 fast_synch_epoch; 410 // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read. 411 // We do not distinguish beteween ignoring reads and writes 412 // for better performance. 413 int ignore_reads_and_writes; 414 uptr *shadow_stack_pos; 415 u64 *racy_shadow_addr; 416 u64 racy_state[2]; 417 #ifndef TSAN_GO 418 // C/C++ uses embed shadow stack of fixed size. 419 uptr shadow_stack[kShadowStackSize]; 420 #else 421 // Go uses satellite shadow stack with dynamic size. 422 uptr *shadow_stack; 423 uptr *shadow_stack_end; 424 #endif 425 MutexSet mset; 426 ThreadClock clock; 427 #ifndef TSAN_GO 428 AllocatorCache alloc_cache; 429 InternalAllocatorCache internal_alloc_cache; 430 Vector<JmpBuf> jmp_bufs; 431 #endif 432 u64 stat[StatCnt]; 433 const int tid; 434 const int unique_id; 435 int in_rtl; 436 bool in_symbolizer; 437 bool is_alive; 438 bool is_freeing; 439 bool is_vptr_access; 440 const uptr stk_addr; 441 const uptr stk_size; 442 const uptr tls_addr; 443 const uptr tls_size; 444 445 DeadlockDetector deadlock_detector; 446 447 bool in_signal_handler; 448 SignalContext *signal_ctx; 449 450 #ifndef TSAN_GO 451 u32 last_sleep_stack_id; 452 ThreadClock last_sleep_clock; 453 #endif 454 455 // Set in regions of runtime that must be signal-safe and fork-safe. 456 // If set, malloc must not be called. 457 int nomalloc; 458 459 explicit ThreadState(Context *ctx, int tid, int unique_id, u64 epoch, 460 uptr stk_addr, uptr stk_size, 461 uptr tls_addr, uptr tls_size); 462 }; 463 464 Context *CTX(); 465 466 #ifndef TSAN_GO 467 extern THREADLOCAL char cur_thread_placeholder[]; 468 INLINE ThreadState *cur_thread() { 469 return reinterpret_cast<ThreadState *>(&cur_thread_placeholder); 470 } 471 #endif 472 473 class ThreadContext : public ThreadContextBase { 474 public: 475 explicit ThreadContext(int tid); 476 ~ThreadContext(); 477 ThreadState *thr; 478 #ifdef TSAN_GO 479 StackTrace creation_stack; 480 #else 481 u32 creation_stack_id; 482 #endif 483 SyncClock sync; 484 // Epoch at which the thread had started. 485 // If we see an event from the thread stamped by an older epoch, 486 // the event is from a dead thread that shared tid with this thread. 487 u64 epoch0; 488 u64 epoch1; 489 490 // Override superclass callbacks. 491 void OnDead(); 492 void OnJoined(void *arg); 493 void OnFinished(); 494 void OnStarted(void *arg); 495 void OnCreated(void *arg); 496 void OnReset(); 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; 518 Suppression *supp; 519 }; 520 521 struct Context { 522 Context(); 523 524 bool initialized; 525 526 SyncTab synctab; 527 528 Mutex report_mtx; 529 int nreported; 530 int nmissed_expected; 531 atomic_uint64_t last_symbolize_time_ns; 532 533 ThreadRegistry *thread_registry; 534 535 Vector<RacyStacks> racy_stacks; 536 Vector<RacyAddress> racy_addresses; 537 // Number of fired suppressions may be large enough. 538 InternalMmapVector<FiredSuppression> fired_suppressions; 539 540 Flags flags; 541 542 u64 stat[StatCnt]; 543 u64 int_alloc_cnt[MBlockTypeCount]; 544 u64 int_alloc_siz[MBlockTypeCount]; 545 }; 546 547 class ScopedInRtl { 548 public: 549 ScopedInRtl(); 550 ~ScopedInRtl(); 551 private: 552 ThreadState*thr_; 553 int in_rtl_; 554 int errno_; 555 }; 556 557 class ScopedReport { 558 public: 559 explicit ScopedReport(ReportType typ); 560 ~ScopedReport(); 561 562 void AddStack(const StackTrace *stack); 563 void AddMemoryAccess(uptr addr, Shadow s, const StackTrace *stack, 564 const MutexSet *mset); 565 void AddThread(const ThreadContext *tctx); 566 void AddMutex(const SyncVar *s); 567 void AddLocation(uptr addr, uptr size); 568 void AddSleep(u32 stack_id); 569 void SetCount(int count); 570 571 const ReportDesc *GetReport() const; 572 573 private: 574 Context *ctx_; 575 ReportDesc *rep_; 576 577 void AddMutex(u64 id); 578 579 ScopedReport(const ScopedReport&); 580 void operator = (const ScopedReport&); 581 }; 582 583 void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset); 584 585 void StatAggregate(u64 *dst, u64 *src); 586 void StatOutput(u64 *stat); 587 void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) { 588 if (kCollectStats) 589 thr->stat[typ] += n; 590 } 591 void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) { 592 if (kCollectStats) 593 thr->stat[typ] = n; 594 } 595 596 void MapShadow(uptr addr, uptr size); 597 void MapThreadTrace(uptr addr, uptr size); 598 void DontNeedShadowFor(uptr addr, uptr size); 599 void InitializeShadowMemory(); 600 void InitializeInterceptors(); 601 void InitializeDynamicAnnotations(); 602 603 void ReportRace(ThreadState *thr); 604 bool OutputReport(Context *ctx, 605 const ScopedReport &srep, 606 const ReportStack *suppress_stack1 = 0, 607 const ReportStack *suppress_stack2 = 0, 608 const ReportLocation *suppress_loc = 0); 609 bool IsFiredSuppression(Context *ctx, 610 const ScopedReport &srep, 611 const StackTrace &trace); 612 bool IsExpectedReport(uptr addr, uptr size); 613 void PrintMatchedBenignRaces(); 614 bool FrameIsInternal(const ReportStack *frame); 615 ReportStack *SkipTsanInternalFrames(ReportStack *ent); 616 617 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1 618 # define DPrintf Printf 619 #else 620 # define DPrintf(...) 621 #endif 622 623 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2 624 # define DPrintf2 Printf 625 #else 626 # define DPrintf2(...) 627 #endif 628 629 u32 CurrentStackId(ThreadState *thr, uptr pc); 630 void PrintCurrentStack(ThreadState *thr, uptr pc); 631 void PrintCurrentStackSlow(); // uses libunwind 632 633 void Initialize(ThreadState *thr); 634 int Finalize(ThreadState *thr); 635 636 SyncVar* GetJavaSync(ThreadState *thr, uptr pc, uptr addr, 637 bool write_lock, bool create); 638 SyncVar* GetAndRemoveJavaSync(ThreadState *thr, uptr pc, uptr addr); 639 640 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr, 641 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic); 642 void MemoryAccessImpl(ThreadState *thr, uptr addr, 643 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic, 644 u64 *shadow_mem, Shadow cur); 645 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr, 646 uptr size, bool is_write); 647 void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr, 648 uptr size, uptr step, bool is_write); 649 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr, 650 int size, bool kAccessIsWrite, bool kIsAtomic); 651 652 const int kSizeLog1 = 0; 653 const int kSizeLog2 = 1; 654 const int kSizeLog4 = 2; 655 const int kSizeLog8 = 3; 656 657 void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc, 658 uptr addr, int kAccessSizeLog) { 659 MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false); 660 } 661 662 void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc, 663 uptr addr, int kAccessSizeLog) { 664 MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false); 665 } 666 667 void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc, 668 uptr addr, int kAccessSizeLog) { 669 MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true); 670 } 671 672 void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc, 673 uptr addr, int kAccessSizeLog) { 674 MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true); 675 } 676 677 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size); 678 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size); 679 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size); 680 void IgnoreCtl(ThreadState *thr, bool write, bool begin); 681 682 void FuncEntry(ThreadState *thr, uptr pc); 683 void FuncExit(ThreadState *thr); 684 685 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached); 686 void ThreadStart(ThreadState *thr, int tid, uptr os_id); 687 void ThreadFinish(ThreadState *thr); 688 int ThreadTid(ThreadState *thr, uptr pc, uptr uid); 689 void ThreadJoin(ThreadState *thr, uptr pc, int tid); 690 void ThreadDetach(ThreadState *thr, uptr pc, int tid); 691 void ThreadFinalize(ThreadState *thr); 692 void ThreadSetName(ThreadState *thr, const char *name); 693 int ThreadCount(ThreadState *thr); 694 void ProcessPendingSignals(ThreadState *thr); 695 696 void MutexCreate(ThreadState *thr, uptr pc, uptr addr, 697 bool rw, bool recursive, bool linker_init); 698 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr); 699 void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1); 700 int MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false); 701 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr); 702 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr); 703 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr); 704 705 void Acquire(ThreadState *thr, uptr pc, uptr addr); 706 void AcquireGlobal(ThreadState *thr, uptr pc); 707 void Release(ThreadState *thr, uptr pc, uptr addr); 708 void ReleaseStore(ThreadState *thr, uptr pc, uptr addr); 709 void AfterSleep(ThreadState *thr, uptr pc); 710 711 // The hacky call uses custom calling convention and an assembly thunk. 712 // It is considerably faster that a normal call for the caller 713 // if it is not executed (it is intended for slow paths from hot functions). 714 // The trick is that the call preserves all registers and the compiler 715 // does not treat it as a call. 716 // If it does not work for you, use normal call. 717 #if TSAN_DEBUG == 0 718 // The caller may not create the stack frame for itself at all, 719 // so we create a reserve stack frame for it (1024b must be enough). 720 #define HACKY_CALL(f) \ 721 __asm__ __volatile__("sub $1024, %%rsp;" \ 722 "/*.cfi_adjust_cfa_offset 1024;*/" \ 723 ".hidden " #f "_thunk;" \ 724 "call " #f "_thunk;" \ 725 "add $1024, %%rsp;" \ 726 "/*.cfi_adjust_cfa_offset -1024;*/" \ 727 ::: "memory", "cc"); 728 #else 729 #define HACKY_CALL(f) f() 730 #endif 731 732 void TraceSwitch(ThreadState *thr); 733 uptr TraceTopPC(ThreadState *thr); 734 uptr TraceSize(); 735 uptr TraceParts(); 736 Trace *ThreadTrace(int tid); 737 738 extern "C" void __tsan_trace_switch(); 739 void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs, 740 EventType typ, u64 addr) { 741 DCHECK_GE((int)typ, 0); 742 DCHECK_LE((int)typ, 7); 743 DCHECK_EQ(GetLsb(addr, 61), addr); 744 StatInc(thr, StatEvents); 745 u64 pos = fs.GetTracePos(); 746 if (UNLIKELY((pos % kTracePartSize) == 0)) { 747 #ifndef TSAN_GO 748 HACKY_CALL(__tsan_trace_switch); 749 #else 750 TraceSwitch(thr); 751 #endif 752 } 753 Event *trace = (Event*)GetThreadTrace(fs.tid()); 754 Event *evp = &trace[pos]; 755 Event ev = (u64)addr | ((u64)typ << 61); 756 *evp = ev; 757 } 758 759 } // namespace __tsan 760 761 #endif // TSAN_RTL_H 762
