libsanitizer/tsan/tsan_rtl.cc

   1 //===-- tsan_rtl.cc -------------------------------------------------------===//
   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 file (entry points) for the TSan run-time.
  11 //===----------------------------------------------------------------------===//
  12
  13 #include "sanitizer_common/sanitizer_atomic.h"
  14 #include "sanitizer_common/sanitizer_common.h"
  15 #include "sanitizer_common/sanitizer_libc.h"
  16 #include "sanitizer_common/sanitizer_stackdepot.h"
  17 #include "sanitizer_common/sanitizer_placement_new.h"
  18 #include "sanitizer_common/sanitizer_symbolizer.h"
  19 #include "tsan_defs.h"
  20 #include "tsan_platform.h"
  21 #include "tsan_rtl.h"
  22 #include "tsan_mman.h"
  23 #include "tsan_suppressions.h"
  24 #include "tsan_symbolize.h"
  25
  26 #ifdef __SSE3__
  27 // <emmintrin.h> transitively includes <stdlib.h>,
  28 // and it's prohibited to include std headers into tsan runtime.
  29 // So we do this dirty trick.
  30 #define _MM_MALLOC_H_INCLUDED
  31 #define __MM_MALLOC_H
  32 #include <emmintrin.h>
  33 typedef __m128i m128;
  34 #endif
  35
  36 volatile int __tsan_resumed = 0;
  37
  38 extern "C" void __tsan_resume() {
  39   __tsan_resumed = 1;
  40 }
  41
  42 namespace __tsan {
  43
  44 #ifndef TSAN_GO
  45 THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
  46 #endif
  47 static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
  48 Context *ctx;
  49
  50 // Can be overriden by a front-end.
  51 #ifdef TSAN_EXTERNAL_HOOKS
  52 bool OnFinalize(bool failed);
  53 void OnInitialize();
  54 #else
  55 SANITIZER_INTERFACE_ATTRIBUTE
  56 bool WEAK OnFinalize(bool failed) {
  57   return failed;
  58 }
  59 SANITIZER_INTERFACE_ATTRIBUTE
  60 void WEAK OnInitialize() {}
  61 #endif
  62
  63 static char thread_registry_placeholder[sizeof(ThreadRegistry)];
  64
  65 static ThreadContextBase *CreateThreadContext(u32 tid) {
  66   // Map thread trace when context is created.
  67   MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event));
  68   MapThreadTrace(GetThreadTraceHeader(tid), sizeof(Trace));
  69   new(ThreadTrace(tid)) Trace();
  70   void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
  71   return new(mem) ThreadContext(tid);
  72 }
  73
  74 #ifndef TSAN_GO
  75 static const u32 kThreadQuarantineSize = 16;
  76 #else
  77 static const u32 kThreadQuarantineSize = 64;
  78 #endif
  79
  80 Context::Context()
  81   : initialized()
  82   , report_mtx(MutexTypeReport, StatMtxReport)
  83   , nreported()
  84   , nmissed_expected()
  85   , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
  86       CreateThreadContext, kMaxTid, kThreadQuarantineSize, kMaxTidReuse))
  87   , racy_stacks(MBlockRacyStacks)
  88   , racy_addresses(MBlockRacyAddresses)
  89   , fired_suppressions(8) {
  90 }
  91
  92 // The objects are allocated in TLS, so one may rely on zero-initialization.
  93 ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
  94                          unsigned reuse_count,
  95                          uptr stk_addr, uptr stk_size,
  96                          uptr tls_addr, uptr tls_size)
  97   : fast_state(tid, epoch)
  98   // Do not touch these, rely on zero initialization,
  99   // they may be accessed before the ctor.
 100   // , ignore_reads_and_writes()
 101   // , ignore_interceptors()
 102   , clock(tid, reuse_count)
 103 #ifndef TSAN_GO
 104   , jmp_bufs(MBlockJmpBuf)
 105 #endif
 106   , tid(tid)
 107   , unique_id(unique_id)
 108   , stk_addr(stk_addr)
 109   , stk_size(stk_size)
 110   , tls_addr(tls_addr)
 111   , tls_size(tls_size)
 112 #ifndef TSAN_GO
 113   , last_sleep_clock(tid)
 114 #endif
 115 {
 116 }
 117
 118 static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
 119   uptr n_threads;
 120   uptr n_running_threads;
 121   ctx->thread_registry->GetNumberOfThreads(&n_threads, &n_running_threads);
 122   InternalScopedBuffer<char> buf(4096);
 123   WriteMemoryProfile(buf.data(), buf.size(), n_threads, n_running_threads);
 124   internal_write(fd, buf.data(), internal_strlen(buf.data()));
 125 }
 126
 127 static void BackgroundThread(void *arg) {
 128 #ifndef TSAN_GO
 129   // This is a non-initialized non-user thread, nothing to see here.
 130   // We don't use ScopedIgnoreInterceptors, because we want ignores to be
 131   // enabled even when the thread function exits (e.g. during pthread thread
 132   // shutdown code).
 133   cur_thread()->ignore_interceptors++;
 134 #endif
 135   const u64 kMs2Ns = 1000 * 1000;
 136
 137   fd_t mprof_fd = kInvalidFd;
 138   if (flags()->profile_memory && flags()->profile_memory[0]) {
 139     if (internal_strcmp(flags()->profile_memory, "stdout") == 0) {
 140       mprof_fd = 1;
 141     } else if (internal_strcmp(flags()->profile_memory, "stderr") == 0) {
 142       mprof_fd = 2;
 143     } else {
 144       InternalScopedBuffer<char> filename(4096);
 145       internal_snprintf(filename.data(), filename.size(), "%s.%d",
 146           flags()->profile_memory, (int)internal_getpid());
 147       uptr openrv = OpenFile(filename.data(), true);
 148       if (internal_iserror(openrv)) {
 149         Printf("ThreadSanitizer: failed to open memory profile file '%s'\n",
 150             &filename[0]);
 151       } else {
 152         mprof_fd = openrv;
 153       }
 154     }
 155   }
 156
 157   u64 last_flush = NanoTime();
 158   uptr last_rss = 0;
 159   for (int i = 0;
 160       atomic_load(&ctx->stop_background_thread, memory_order_relaxed) == 0;
 161       i++) {
 162     SleepForMillis(100);
 163     u64 now = NanoTime();
 164
 165     // Flush memory if requested.
 166     if (flags()->flush_memory_ms > 0) {
 167       if (last_flush + flags()->flush_memory_ms * kMs2Ns < now) {
 168         VPrintf(1, "ThreadSanitizer: periodic memory flush\n");
 169         FlushShadowMemory();
 170         last_flush = NanoTime();
 171       }
 172     }
 173     // GetRSS can be expensive on huge programs, so don't do it every 100ms.
 174     if (flags()->memory_limit_mb > 0) {
 175       uptr rss = GetRSS();
 176       uptr limit = uptr(flags()->memory_limit_mb) << 20;
 177       VPrintf(1, "ThreadSanitizer: memory flush check"
 178                  " RSS=%llu LAST=%llu LIMIT=%llu\n",
 179               (u64)rss >> 20, (u64)last_rss >> 20, (u64)limit >> 20);
 180       if (2 * rss > limit + last_rss) {
 181         VPrintf(1, "ThreadSanitizer: flushing memory due to RSS\n");
 182         FlushShadowMemory();
 183         rss = GetRSS();
 184         VPrintf(1, "ThreadSanitizer: memory flushed RSS=%llu\n", (u64)rss>>20);
 185       }
 186       last_rss = rss;
 187     }
 188
 189     // Write memory profile if requested.
 190     if (mprof_fd != kInvalidFd)
 191       MemoryProfiler(ctx, mprof_fd, i);
 192
 193 #ifndef TSAN_GO
 194     // Flush symbolizer cache if requested.
 195     if (flags()->flush_symbolizer_ms > 0) {
 196       u64 last = atomic_load(&ctx->last_symbolize_time_ns,
 197                              memory_order_relaxed);
 198       if (last != 0 && last + flags()->flush_symbolizer_ms * kMs2Ns < now) {
 199         Lock l(&ctx->report_mtx);
 200         SpinMutexLock l2(&CommonSanitizerReportMutex);
 201         SymbolizeFlush();
 202         atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
 203       }
 204     }
 205 #endif
 206   }
 207 }
 208
 209 static void StartBackgroundThread() {
 210   ctx->background_thread = internal_start_thread(&BackgroundThread, 0);
 211 }
 212
 213 #ifndef TSAN_GO
 214 static void StopBackgroundThread() {
 215   atomic_store(&ctx->stop_background_thread, 1, memory_order_relaxed);
 216   internal_join_thread(ctx->background_thread);
 217   ctx->background_thread = 0;
 218 }
 219 #endif
 220
 221 void DontNeedShadowFor(uptr addr, uptr size) {
 222   uptr shadow_beg = MemToShadow(addr);
 223   uptr shadow_end = MemToShadow(addr + size);
 224   FlushUnneededShadowMemory(shadow_beg, shadow_end - shadow_beg);
 225 }
 226
 227 void MapShadow(uptr addr, uptr size) {
 228   // Global data is not 64K aligned, but there are no adjacent mappings,
 229   // so we can get away with unaligned mapping.
 230   // CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
 231   MmapFixedNoReserve(MemToShadow(addr), size * kShadowMultiplier);
 232
 233   // Meta shadow is 2:1, so tread carefully.
 234   static bool data_mapped = false;
 235   static uptr mapped_meta_end = 0;
 236   uptr meta_begin = (uptr)MemToMeta(addr);
 237   uptr meta_end = (uptr)MemToMeta(addr + size);
 238   meta_begin = RoundDownTo(meta_begin, 64 << 10);
 239   meta_end = RoundUpTo(meta_end, 64 << 10);
 240   if (!data_mapped) {
 241     // First call maps data+bss.
 242     data_mapped = true;
 243     MmapFixedNoReserve(meta_begin, meta_end - meta_begin);
 244   } else {
 245     // Mapping continous heap.
 246     // Windows wants 64K alignment.
 247     meta_begin = RoundDownTo(meta_begin, 64 << 10);
 248     meta_end = RoundUpTo(meta_end, 64 << 10);
 249     if (meta_end <= mapped_meta_end)
 250       return;
 251     if (meta_begin < mapped_meta_end)
 252       meta_begin = mapped_meta_end;
 253     MmapFixedNoReserve(meta_begin, meta_end - meta_begin);
 254     mapped_meta_end = meta_end;
 255   }
 256   VPrintf(2, "mapped meta shadow for (%p-%p) at (%p-%p)\n",
 257       addr, addr+size, meta_begin, meta_end);
 258 }
 259
 260 void MapThreadTrace(uptr addr, uptr size) {
 261   DPrintf("#0: Mapping trace at %p-%p(0x%zx)\n", addr, addr + size, size);
 262   CHECK_GE(addr, kTraceMemBeg);
 263   CHECK_LE(addr + size, kTraceMemEnd);
 264   CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
 265   uptr addr1 = (uptr)MmapFixedNoReserve(addr, size);
 266   if (addr1 != addr) {
 267     Printf("FATAL: ThreadSanitizer can not mmap thread trace (%p/%p->%p)\n",
 268         addr, size, addr1);
 269     Die();
 270   }
 271 }
 272
 273 static void CheckShadowMapping() {
 274   for (uptr i = 0; i < ARRAY_SIZE(UserRegions); i += 2) {
 275     const uptr beg = UserRegions[i];
 276     const uptr end = UserRegions[i + 1];
 277     VPrintf(3, "checking shadow region %p-%p\n", beg, end);
 278     for (uptr p0 = beg; p0 <= end; p0 += (end - beg) / 4) {
 279       for (int x = -1; x <= 1; x++) {
 280         const uptr p = p0 + x;
 281         if (p < beg || p >= end)
 282           continue;
 283         const uptr s = MemToShadow(p);
 284         VPrintf(3, "  checking pointer %p -> %p\n", p, s);
 285         CHECK(IsAppMem(p));
 286         CHECK(IsShadowMem(s));
 287         CHECK_EQ(p & ~(kShadowCell - 1), ShadowToMem(s));
 288         const uptr m = (uptr)MemToMeta(p);
 289         CHECK(IsMetaMem(m));
 290       }
 291     }
 292   }
 293 }
 294
 295 void Initialize(ThreadState *thr) {
 296   // Thread safe because done before all threads exist.
 297   static bool is_initialized = false;
 298   if (is_initialized)
 299     return;
 300   is_initialized = true;
 301   // We are not ready to handle interceptors yet.
 302   ScopedIgnoreInterceptors ignore;
 303   SanitizerToolName = "ThreadSanitizer";
 304   // Install tool-specific callbacks in sanitizer_common.
 305   SetCheckFailedCallback(TsanCheckFailed);
 306
 307   ctx = new(ctx_placeholder) Context;
 308   const char *options = GetEnv(kTsanOptionsEnv);
 309   InitializeFlags(&ctx->flags, options);
 310 #ifndef TSAN_GO
 311   InitializeAllocator();
 312 #endif
 313   InitializeInterceptors();
 314   CheckShadowMapping();
 315   InitializePlatform();
 316   InitializeMutex();
 317   InitializeDynamicAnnotations();
 318 #ifndef TSAN_GO
 319   InitializeShadowMemory();
 320 #endif
 321   // Setup correct file descriptor for error reports.
 322   __sanitizer_set_report_path(common_flags()->log_path);
 323   InitializeSuppressions();
 324 #ifndef TSAN_GO
 325   InitializeLibIgnore();
 326   Symbolizer::GetOrInit()->AddHooks(EnterSymbolizer, ExitSymbolizer);
 327 #endif
 328   StartBackgroundThread();
 329 #ifndef TSAN_GO
 330   SetSandboxingCallback(StopBackgroundThread);
 331 #endif
 332   if (common_flags()->detect_deadlocks)
 333     ctx->dd = DDetector::Create(flags());
 334
 335   VPrintf(1, "***** Running under ThreadSanitizer v2 (pid %d) *****\n",
 336           (int)internal_getpid());
 337
 338   // Initialize thread 0.
 339   int tid = ThreadCreate(thr, 0, 0, true);
 340   CHECK_EQ(tid, 0);
 341   ThreadStart(thr, tid, internal_getpid());
 342   ctx->initialized = true;
 343
 344   if (flags()->stop_on_start) {
 345     Printf("ThreadSanitizer is suspended at startup (pid %d)."
 346            " Call __tsan_resume().\n",
 347            (int)internal_getpid());
 348     while (__tsan_resumed == 0) {}
 349   }
 350
 351   OnInitialize();
 352 }
 353
 354 int Finalize(ThreadState *thr) {
 355   bool failed = false;
 356
 357   if (flags()->atexit_sleep_ms > 0 && ThreadCount(thr) > 1)
 358     SleepForMillis(flags()->atexit_sleep_ms);
 359
 360   // Wait for pending reports.
 361   ctx->report_mtx.Lock();
 362   CommonSanitizerReportMutex.Lock();
 363   CommonSanitizerReportMutex.Unlock();
 364   ctx->report_mtx.Unlock();
 365
 366 #ifndef TSAN_GO
 367   if (common_flags()->verbosity)
 368     AllocatorPrintStats();
 369 #endif
 370
 371   ThreadFinalize(thr);
 372
 373   if (ctx->nreported) {
 374     failed = true;
 375 #ifndef TSAN_GO
 376     Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
 377 #else
 378     Printf("Found %d data race(s)\n", ctx->nreported);
 379 #endif
 380   }
 381
 382   if (ctx->nmissed_expected) {
 383     failed = true;
 384     Printf("ThreadSanitizer: missed %d expected races\n",
 385         ctx->nmissed_expected);
 386   }
 387
 388   if (common_flags()->print_suppressions)
 389     PrintMatchedSuppressions();
 390 #ifndef TSAN_GO
 391   if (flags()->print_benign)
 392     PrintMatchedBenignRaces();
 393 #endif
 394
 395   failed = OnFinalize(failed);
 396
 397   StatAggregate(ctx->stat, thr->stat);
 398   StatOutput(ctx->stat);
 399   return failed ? flags()->exitcode : 0;
 400 }
 401
 402 #ifndef TSAN_GO
 403 void ForkBefore(ThreadState *thr, uptr pc) {
 404   ctx->thread_registry->Lock();
 405   ctx->report_mtx.Lock();
 406 }
 407
 408 void ForkParentAfter(ThreadState *thr, uptr pc) {
 409   ctx->report_mtx.Unlock();
 410   ctx->thread_registry->Unlock();
 411 }
 412
 413 void ForkChildAfter(ThreadState *thr, uptr pc) {
 414   ctx->report_mtx.Unlock();
 415   ctx->thread_registry->Unlock();
 416
 417   uptr nthread = 0;
 418   ctx->thread_registry->GetNumberOfThreads(0, 0, &nthread /* alive threads */);
 419   VPrintf(1, "ThreadSanitizer: forked new process with pid %d,"
 420       " parent had %d threads\n", (int)internal_getpid(), (int)nthread);
 421   if (nthread == 1) {
 422     internal_start_thread(&BackgroundThread, 0);
 423   } else {
 424     // We've just forked a multi-threaded process. We cannot reasonably function
 425     // after that (some mutexes may be locked before fork). So just enable
 426     // ignores for everything in the hope that we will exec soon.
 427     ctx->after_multithreaded_fork = true;
 428     thr->ignore_interceptors++;
 429     ThreadIgnoreBegin(thr, pc);
 430     ThreadIgnoreSyncBegin(thr, pc);
 431   }
 432 }
 433 #endif
 434
 435 #ifdef TSAN_GO
 436 NOINLINE
 437 void GrowShadowStack(ThreadState *thr) {
 438   const int sz = thr->shadow_stack_end - thr->shadow_stack;
 439   const int newsz = 2 * sz;
 440   uptr *newstack = (uptr*)internal_alloc(MBlockShadowStack,
 441       newsz * sizeof(uptr));
 442   internal_memcpy(newstack, thr->shadow_stack, sz * sizeof(uptr));
 443   internal_free(thr->shadow_stack);
 444   thr->shadow_stack = newstack;
 445   thr->shadow_stack_pos = newstack + sz;
 446   thr->shadow_stack_end = newstack + newsz;
 447 }
 448 #endif
 449
 450 u32 CurrentStackId(ThreadState *thr, uptr pc) {
 451   if (thr->shadow_stack_pos == 0)  // May happen during bootstrap.
 452     return 0;
 453   if (pc != 0) {
 454 #ifndef TSAN_GO
 455     DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 456 #else
 457     if (thr->shadow_stack_pos == thr->shadow_stack_end)
 458       GrowShadowStack(thr);
 459 #endif
 460     thr->shadow_stack_pos[0] = pc;
 461     thr->shadow_stack_pos++;
 462   }
 463   u32 id = StackDepotPut(
 464       StackTrace(thr->shadow_stack, thr->shadow_stack_pos - thr->shadow_stack));
 465   if (pc != 0)
 466     thr->shadow_stack_pos--;
 467   return id;
 468 }
 469
 470 void TraceSwitch(ThreadState *thr) {
 471   thr->nomalloc++;
 472   Trace *thr_trace = ThreadTrace(thr->tid);
 473   Lock l(&thr_trace->mtx);
 474   unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % TraceParts();
 475   TraceHeader *hdr = &thr_trace->headers[trace];
 476   hdr->epoch0 = thr->fast_state.epoch();
 477   ObtainCurrentStack(thr, 0, &hdr->stack0);
 478   hdr->mset0 = thr->mset;
 479   thr->nomalloc--;
 480 }
 481
 482 Trace *ThreadTrace(int tid) {
 483   return (Trace*)GetThreadTraceHeader(tid);
 484 }
 485
 486 uptr TraceTopPC(ThreadState *thr) {
 487   Event *events = (Event*)GetThreadTrace(thr->tid);
 488   uptr pc = events[thr->fast_state.GetTracePos()];
 489   return pc;
 490 }
 491
 492 uptr TraceSize() {
 493   return (uptr)(1ull << (kTracePartSizeBits + flags()->history_size + 1));
 494 }
 495
 496 uptr TraceParts() {
 497   return TraceSize() / kTracePartSize;
 498 }
 499
 500 #ifndef TSAN_GO
 501 extern "C" void __tsan_trace_switch() {
 502   TraceSwitch(cur_thread());
 503 }
 504
 505 extern "C" void __tsan_report_race() {
 506   ReportRace(cur_thread());
 507 }
 508 #endif
 509
 510 ALWAYS_INLINE
 511 Shadow LoadShadow(u64 *p) {
 512   u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
 513   return Shadow(raw);
 514 }
 515
 516 ALWAYS_INLINE
 517 void StoreShadow(u64 *sp, u64 s) {
 518   atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
 519 }
 520
 521 ALWAYS_INLINE
 522 void StoreIfNotYetStored(u64 *sp, u64 *s) {
 523   StoreShadow(sp, *s);
 524   *s = 0;
 525 }
 526
 527 ALWAYS_INLINE
 528 void HandleRace(ThreadState *thr, u64 *shadow_mem,
 529                               Shadow cur, Shadow old) {
 530   thr->racy_state[0] = cur.raw();
 531   thr->racy_state[1] = old.raw();
 532   thr->racy_shadow_addr = shadow_mem;
 533 #ifndef TSAN_GO
 534   HACKY_CALL(__tsan_report_race);
 535 #else
 536   ReportRace(thr);
 537 #endif
 538 }
 539
 540 static inline bool HappensBefore(Shadow old, ThreadState *thr) {
 541   return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
 542 }
 543
 544 ALWAYS_INLINE
 545 void MemoryAccessImpl1(ThreadState *thr, uptr addr,
 546     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 547     u64 *shadow_mem, Shadow cur) {
 548   StatInc(thr, StatMop);
 549   StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 550   StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 551
 552   // This potentially can live in an MMX/SSE scratch register.
 553   // The required intrinsics are:
 554   // __m128i _mm_move_epi64(__m128i*);
 555   // _mm_storel_epi64(u64*, __m128i);
 556   u64 store_word = cur.raw();
 557
 558   // scan all the shadow values and dispatch to 4 categories:
 559   // same, replace, candidate and race (see comments below).
 560   // we consider only 3 cases regarding access sizes:
 561   // equal, intersect and not intersect. initially I considered
 562   // larger and smaller as well, it allowed to replace some
 563   // 'candidates' with 'same' or 'replace', but I think
 564   // it's just not worth it (performance- and complexity-wise).
 565
 566   Shadow old(0);
 567   if (kShadowCnt == 1) {
 568     int idx = 0;
 569 #include "tsan_update_shadow_word_inl.h"
 570   } else if (kShadowCnt == 2) {
 571     int idx = 0;
 572 #include "tsan_update_shadow_word_inl.h"
 573     idx = 1;
 574 #include "tsan_update_shadow_word_inl.h"
 575   } else if (kShadowCnt == 4) {
 576     int idx = 0;
 577 #include "tsan_update_shadow_word_inl.h"
 578     idx = 1;
 579 #include "tsan_update_shadow_word_inl.h"
 580     idx = 2;
 581 #include "tsan_update_shadow_word_inl.h"
 582     idx = 3;
 583 #include "tsan_update_shadow_word_inl.h"
 584   } else if (kShadowCnt == 8) {
 585     int idx = 0;
 586 #include "tsan_update_shadow_word_inl.h"
 587     idx = 1;
 588 #include "tsan_update_shadow_word_inl.h"
 589     idx = 2;
 590 #include "tsan_update_shadow_word_inl.h"
 591     idx = 3;
 592 #include "tsan_update_shadow_word_inl.h"
 593     idx = 4;
 594 #include "tsan_update_shadow_word_inl.h"
 595     idx = 5;
 596 #include "tsan_update_shadow_word_inl.h"
 597     idx = 6;
 598 #include "tsan_update_shadow_word_inl.h"
 599     idx = 7;
 600 #include "tsan_update_shadow_word_inl.h"
 601   } else {
 602     CHECK(false);
 603   }
 604
 605   // we did not find any races and had already stored
 606   // the current access info, so we are done
 607   if (LIKELY(store_word == 0))
 608     return;
 609   // choose a random candidate slot and replace it
 610   StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
 611   StatInc(thr, StatShadowReplace);
 612   return;
 613  RACE:
 614   HandleRace(thr, shadow_mem, cur, old);
 615   return;
 616 }
 617
 618 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 619     int size, bool kAccessIsWrite, bool kIsAtomic) {
 620   while (size) {
 621     int size1 = 1;
 622     int kAccessSizeLog = kSizeLog1;
 623     if (size >= 8 && (addr & ~7) == ((addr + 7) & ~7)) {
 624       size1 = 8;
 625       kAccessSizeLog = kSizeLog8;
 626     } else if (size >= 4 && (addr & ~7) == ((addr + 3) & ~7)) {
 627       size1 = 4;
 628       kAccessSizeLog = kSizeLog4;
 629     } else if (size >= 2 && (addr & ~7) == ((addr + 1) & ~7)) {
 630       size1 = 2;
 631       kAccessSizeLog = kSizeLog2;
 632     }
 633     MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
 634     addr += size1;
 635     size -= size1;
 636   }
 637 }
 638
 639 ALWAYS_INLINE
 640 bool ContainsSameAccessSlow(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 641   Shadow cur(a);
 642   for (uptr i = 0; i < kShadowCnt; i++) {
 643     Shadow old(LoadShadow(&s[i]));
 644     if (Shadow::Addr0AndSizeAreEqual(cur, old) &&
 645         old.TidWithIgnore() == cur.TidWithIgnore() &&
 646         old.epoch() > sync_epoch &&
 647         old.IsAtomic() == cur.IsAtomic() &&
 648         old.IsRead() <= cur.IsRead())
 649       return true;
 650   }
 651   return false;
 652 }
 653
 654 #if defined(__SSE3__) && TSAN_SHADOW_COUNT == 4
 655 #define SHUF(v0, v1, i0, i1, i2, i3) _mm_castps_si128(_mm_shuffle_ps( \
 656     _mm_castsi128_ps(v0), _mm_castsi128_ps(v1), \
 657     (i0)*1 + (i1)*4 + (i2)*16 + (i3)*64))
 658 ALWAYS_INLINE
 659 bool ContainsSameAccessFast(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 660   // This is an optimized version of ContainsSameAccessSlow.
 661   // load current access into access[0:63]
 662   const m128 access     = _mm_cvtsi64_si128(a);
 663   // duplicate high part of access in addr0:
 664   // addr0[0:31]        = access[32:63]
 665   // addr0[32:63]       = access[32:63]
 666   // addr0[64:95]       = access[32:63]
 667   // addr0[96:127]      = access[32:63]
 668   const m128 addr0      = SHUF(access, access, 1, 1, 1, 1);
 669   // load 4 shadow slots
 670   const m128 shadow0    = _mm_load_si128((__m128i*)s);
 671   const m128 shadow1    = _mm_load_si128((__m128i*)s + 1);
 672   // load high parts of 4 shadow slots into addr_vect:
 673   // addr_vect[0:31]    = shadow0[32:63]
 674   // addr_vect[32:63]   = shadow0[96:127]
 675   // addr_vect[64:95]   = shadow1[32:63]
 676   // addr_vect[96:127]  = shadow1[96:127]
 677   m128 addr_vect        = SHUF(shadow0, shadow1, 1, 3, 1, 3);
 678   if (!is_write) {
 679     // set IsRead bit in addr_vect
 680     const m128 rw_mask1 = _mm_cvtsi64_si128(1<<15);
 681     const m128 rw_mask  = SHUF(rw_mask1, rw_mask1, 0, 0, 0, 0);
 682     addr_vect           = _mm_or_si128(addr_vect, rw_mask);
 683   }
 684   // addr0 == addr_vect?
 685   const m128 addr_res   = _mm_cmpeq_epi32(addr0, addr_vect);
 686   // epoch1[0:63]       = sync_epoch
 687   const m128 epoch1     = _mm_cvtsi64_si128(sync_epoch);
 688   // epoch[0:31]        = sync_epoch[0:31]
 689   // epoch[32:63]       = sync_epoch[0:31]
 690   // epoch[64:95]       = sync_epoch[0:31]
 691   // epoch[96:127]      = sync_epoch[0:31]
 692   const m128 epoch      = SHUF(epoch1, epoch1, 0, 0, 0, 0);
 693   // load low parts of shadow cell epochs into epoch_vect:
 694   // epoch_vect[0:31]   = shadow0[0:31]
 695   // epoch_vect[32:63]  = shadow0[64:95]
 696   // epoch_vect[64:95]  = shadow1[0:31]
 697   // epoch_vect[96:127] = shadow1[64:95]
 698   const m128 epoch_vect = SHUF(shadow0, shadow1, 0, 2, 0, 2);
 699   // epoch_vect >= sync_epoch?
 700   const m128 epoch_res  = _mm_cmpgt_epi32(epoch_vect, epoch);
 701   // addr_res & epoch_res
 702   const m128 res        = _mm_and_si128(addr_res, epoch_res);
 703   // mask[0] = res[7]
 704   // mask[1] = res[15]
 705   // ...
 706   // mask[15] = res[127]
 707   const int mask        = _mm_movemask_epi8(res);
 708   return mask != 0;
 709 }
 710 #endif
 711
 712 ALWAYS_INLINE
 713 bool ContainsSameAccess(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 714 #if defined(__SSE3__) && TSAN_SHADOW_COUNT == 4
 715   bool res = ContainsSameAccessFast(s, a, sync_epoch, is_write);
 716   // NOTE: this check can fail if the shadow is concurrently mutated
 717   // by other threads.
 718   DCHECK_EQ(res, ContainsSameAccessSlow(s, a, sync_epoch, is_write));
 719   return res;
 720 #else
 721   return ContainsSameAccessSlow(s, a, sync_epoch, is_write);
 722 #endif
 723 }
 724
 725 ALWAYS_INLINE USED
 726 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 727     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic) {
 728   u64 *shadow_mem = (u64*)MemToShadow(addr);
 729   DPrintf2("#%d: MemoryAccess: @%p %p size=%d"
 730       " is_write=%d shadow_mem=%p {%zx, %zx, %zx, %zx}\n",
 731       (int)thr->fast_state.tid(), (void*)pc, (void*)addr,
 732       (int)(1 << kAccessSizeLog), kAccessIsWrite, shadow_mem,
 733       (uptr)shadow_mem[0], (uptr)shadow_mem[1],
 734       (uptr)shadow_mem[2], (uptr)shadow_mem[3]);
 735 #if TSAN_DEBUG
 736   if (!IsAppMem(addr)) {
 737     Printf("Access to non app mem %zx\n", addr);
 738     DCHECK(IsAppMem(addr));
 739   }
 740   if (!IsShadowMem((uptr)shadow_mem)) {
 741     Printf("Bad shadow addr %p (%zx)\n", shadow_mem, addr);
 742     DCHECK(IsShadowMem((uptr)shadow_mem));
 743   }
 744 #endif
 745
 746   if (kCppMode && *shadow_mem == kShadowRodata) {
 747     // Access to .rodata section, no races here.
 748     // Measurements show that it can be 10-20% of all memory accesses.
 749     StatInc(thr, StatMop);
 750     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 751     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 752     StatInc(thr, StatMopRodata);
 753     return;
 754   }
 755
 756   FastState fast_state = thr->fast_state;
 757   if (fast_state.GetIgnoreBit()) {
 758     StatInc(thr, StatMop);
 759     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 760     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 761     StatInc(thr, StatMopIgnored);
 762     return;
 763   }
 764
 765   Shadow cur(fast_state);
 766   cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
 767   cur.SetWrite(kAccessIsWrite);
 768   cur.SetAtomic(kIsAtomic);
 769
 770   if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
 771       thr->fast_synch_epoch, kAccessIsWrite))) {
 772     StatInc(thr, StatMop);
 773     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 774     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 775     StatInc(thr, StatMopSame);
 776     return;
 777   }
 778
 779   if (kCollectHistory) {
 780     fast_state.IncrementEpoch();
 781     thr->fast_state = fast_state;
 782     TraceAddEvent(thr, fast_state, EventTypeMop, pc);
 783     cur.IncrementEpoch();
 784   }
 785
 786   MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
 787       shadow_mem, cur);
 788 }
 789
 790 // Called by MemoryAccessRange in tsan_rtl_thread.cc
 791 ALWAYS_INLINE USED
 792 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 793     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 794     u64 *shadow_mem, Shadow cur) {
 795   if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
 796       thr->fast_synch_epoch, kAccessIsWrite))) {
 797     StatInc(thr, StatMop);
 798     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 799     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 800     StatInc(thr, StatMopSame);
 801     return;
 802   }
 803
 804   MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
 805       shadow_mem, cur);
 806 }
 807
 808 static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
 809                            u64 val) {
 810   (void)thr;
 811   (void)pc;
 812   if (size == 0)
 813     return;
 814   // FIXME: fix me.
 815   uptr offset = addr % kShadowCell;
 816   if (offset) {
 817     offset = kShadowCell - offset;
 818     if (size <= offset)
 819       return;
 820     addr += offset;
 821     size -= offset;
 822   }
 823   DCHECK_EQ(addr % 8, 0);
 824   // If a user passes some insane arguments (memset(0)),
 825   // let it just crash as usual.
 826   if (!IsAppMem(addr) || !IsAppMem(addr + size - 1))
 827     return;
 828   // Don't want to touch lots of shadow memory.
 829   // If a program maps 10MB stack, there is no need reset the whole range.
 830   size = (size + (kShadowCell - 1)) & ~(kShadowCell - 1);
 831   // UnmapOrDie/MmapFixedNoReserve does not work on Windows,
 832   // so we do it only for C/C++.
 833   if (kGoMode || size < common_flags()->clear_shadow_mmap_threshold) {
 834     u64 *p = (u64*)MemToShadow(addr);
 835     CHECK(IsShadowMem((uptr)p));
 836     CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
 837     // FIXME: may overwrite a part outside the region
 838     for (uptr i = 0; i < size / kShadowCell * kShadowCnt;) {
 839       p[i++] = val;
 840       for (uptr j = 1; j < kShadowCnt; j++)
 841         p[i++] = 0;
 842     }
 843   } else {
 844     // The region is big, reset only beginning and end.
 845     const uptr kPageSize = 4096;
 846     u64 *begin = (u64*)MemToShadow(addr);
 847     u64 *end = begin + size / kShadowCell * kShadowCnt;
 848     u64 *p = begin;
 849     // Set at least first kPageSize/2 to page boundary.
 850     while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
 851       *p++ = val;
 852       for (uptr j = 1; j < kShadowCnt; j++)
 853         *p++ = 0;
 854     }
 855     // Reset middle part.
 856     u64 *p1 = p;
 857     p = RoundDown(end, kPageSize);
 858     UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
 859     MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1);
 860     // Set the ending.
 861     while (p < end) {
 862       *p++ = val;
 863       for (uptr j = 1; j < kShadowCnt; j++)
 864         *p++ = 0;
 865     }
 866   }
 867 }
 868
 869 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 870   MemoryRangeSet(thr, pc, addr, size, 0);
 871 }
 872
 873 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 874   // Processing more than 1k (4k of shadow) is expensive,
 875   // can cause excessive memory consumption (user does not necessary touch
 876   // the whole range) and most likely unnecessary.
 877   if (size > 1024)
 878     size = 1024;
 879   CHECK_EQ(thr->is_freeing, false);
 880   thr->is_freeing = true;
 881   MemoryAccessRange(thr, pc, addr, size, true);
 882   thr->is_freeing = false;
 883   if (kCollectHistory) {
 884     thr->fast_state.IncrementEpoch();
 885     TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
 886   }
 887   Shadow s(thr->fast_state);
 888   s.ClearIgnoreBit();
 889   s.MarkAsFreed();
 890   s.SetWrite(true);
 891   s.SetAddr0AndSizeLog(0, 3);
 892   MemoryRangeSet(thr, pc, addr, size, s.raw());
 893 }
 894
 895 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 896   if (kCollectHistory) {
 897     thr->fast_state.IncrementEpoch();
 898     TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
 899   }
 900   Shadow s(thr->fast_state);
 901   s.ClearIgnoreBit();
 902   s.SetWrite(true);
 903   s.SetAddr0AndSizeLog(0, 3);
 904   MemoryRangeSet(thr, pc, addr, size, s.raw());
 905 }
 906
 907 ALWAYS_INLINE USED
 908 void FuncEntry(ThreadState *thr, uptr pc) {
 909   StatInc(thr, StatFuncEnter);
 910   DPrintf2("#%d: FuncEntry %p\n", (int)thr->fast_state.tid(), (void*)pc);
 911   if (kCollectHistory) {
 912     thr->fast_state.IncrementEpoch();
 913     TraceAddEvent(thr, thr->fast_state, EventTypeFuncEnter, pc);
 914   }
 915
 916   // Shadow stack maintenance can be replaced with
 917   // stack unwinding during trace switch (which presumably must be faster).
 918   DCHECK_GE(thr->shadow_stack_pos, thr->shadow_stack);
 919 #ifndef TSAN_GO
 920   DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 921 #else
 922   if (thr->shadow_stack_pos == thr->shadow_stack_end)
 923     GrowShadowStack(thr);
 924 #endif
 925   thr->shadow_stack_pos[0] = pc;
 926   thr->shadow_stack_pos++;
 927 }
 928
 929 ALWAYS_INLINE USED
 930 void FuncExit(ThreadState *thr) {
 931   StatInc(thr, StatFuncExit);
 932   DPrintf2("#%d: FuncExit\n", (int)thr->fast_state.tid());
 933   if (kCollectHistory) {
 934     thr->fast_state.IncrementEpoch();
 935     TraceAddEvent(thr, thr->fast_state, EventTypeFuncExit, 0);
 936   }
 937
 938   DCHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
 939 #ifndef TSAN_GO
 940   DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 941 #endif
 942   thr->shadow_stack_pos--;
 943 }
 944
 945 void ThreadIgnoreBegin(ThreadState *thr, uptr pc) {
 946   DPrintf("#%d: ThreadIgnoreBegin\n", thr->tid);
 947   thr->ignore_reads_and_writes++;
 948   CHECK_GT(thr->ignore_reads_and_writes, 0);
 949   thr->fast_state.SetIgnoreBit();
 950 #ifndef TSAN_GO
 951   if (!ctx->after_multithreaded_fork)
 952     thr->mop_ignore_set.Add(CurrentStackId(thr, pc));
 953 #endif
 954 }
 955
 956 void ThreadIgnoreEnd(ThreadState *thr, uptr pc) {
 957   DPrintf("#%d: ThreadIgnoreEnd\n", thr->tid);
 958   thr->ignore_reads_and_writes--;
 959   CHECK_GE(thr->ignore_reads_and_writes, 0);
 960   if (thr->ignore_reads_and_writes == 0) {
 961     thr->fast_state.ClearIgnoreBit();
 962 #ifndef TSAN_GO
 963     thr->mop_ignore_set.Reset();
 964 #endif
 965   }
 966 }
 967
 968 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc) {
 969   DPrintf("#%d: ThreadIgnoreSyncBegin\n", thr->tid);
 970   thr->ignore_sync++;
 971   CHECK_GT(thr->ignore_sync, 0);
 972 #ifndef TSAN_GO
 973   if (!ctx->after_multithreaded_fork)
 974     thr->sync_ignore_set.Add(CurrentStackId(thr, pc));
 975 #endif
 976 }
 977
 978 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc) {
 979   DPrintf("#%d: ThreadIgnoreSyncEnd\n", thr->tid);
 980   thr->ignore_sync--;
 981   CHECK_GE(thr->ignore_sync, 0);
 982 #ifndef TSAN_GO
 983   if (thr->ignore_sync == 0)
 984     thr->sync_ignore_set.Reset();
 985 #endif
 986 }
 987
 988 bool MD5Hash::operator==(const MD5Hash &other) const {
 989   return hash[0] == other.hash[0] && hash[1] == other.hash[1];
 990 }
 991
 992 #if TSAN_DEBUG
 993 void build_consistency_debug() {}
 994 #else
 995 void build_consistency_release() {}
 996 #endif
 997
 998 #if TSAN_COLLECT_STATS
 999 void build_consistency_stats() {}
1000 #else
1001 void build_consistency_nostats() {}
1002 #endif
1003
1004 #if TSAN_SHADOW_COUNT == 1
1005 void build_consistency_shadow1() {}
1006 #elif TSAN_SHADOW_COUNT == 2
1007 void build_consistency_shadow2() {}
1008 #elif TSAN_SHADOW_COUNT == 4
1009 void build_consistency_shadow4() {}
1010 #else
1011 void build_consistency_shadow8() {}
1012 #endif
1013
1014 }  // namespace __tsan
1015
1016 #ifndef TSAN_GO
1017 // Must be included in this file to make sure everything is inlined.
1018 #include "tsan_interface_inl.h"
1019 #endif