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_file.h"
  16 #include "sanitizer_common/sanitizer_libc.h"
  17 #include "sanitizer_common/sanitizer_stackdepot.h"
  18 #include "sanitizer_common/sanitizer_placement_new.h"
  19 #include "sanitizer_common/sanitizer_symbolizer.h"
  20 #include "tsan_defs.h"
  21 #include "tsan_platform.h"
  22 #include "tsan_rtl.h"
  23 #include "tsan_mman.h"
  24 #include "tsan_suppressions.h"
  25 #include "tsan_symbolize.h"
  26 #include "ubsan/ubsan_init.h"
  27
  28 #ifdef __SSE3__
  29 // <emmintrin.h> transitively includes <stdlib.h>,
  30 // and it's prohibited to include std headers into tsan runtime.
  31 // So we do this dirty trick.
  32 #define _MM_MALLOC_H_INCLUDED
  33 #define __MM_MALLOC_H
  34 #include <emmintrin.h>
  35 typedef __m128i m128;
  36 #endif
  37
  38 volatile int __tsan_resumed = 0;
  39
  40 extern "C" void __tsan_resume() {
  41   __tsan_resumed = 1;
  42 }
  43
  44 namespace __tsan {
  45
  46 #if !SANITIZER_GO && !SANITIZER_MAC
  47 __attribute__((tls_model("initial-exec")))
  48 THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
  49 #endif
  50 static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
  51 Context *ctx;
  52
  53 // Can be overriden by a front-end.
  54 #ifdef TSAN_EXTERNAL_HOOKS
  55 bool OnFinalize(bool failed);
  56 void OnInitialize();
  57 #else
  58 SANITIZER_WEAK_CXX_DEFAULT_IMPL
  59 bool OnFinalize(bool failed) {
  60   return failed;
  61 }
  62 SANITIZER_WEAK_CXX_DEFAULT_IMPL
  63 void OnInitialize() {}
  64 #endif
  65
  66 static char thread_registry_placeholder[sizeof(ThreadRegistry)];
  67
  68 static ThreadContextBase *CreateThreadContext(u32 tid) {
  69   // Map thread trace when context is created.
  70   char name[50];
  71   internal_snprintf(name, sizeof(name), "trace %u", tid);
  72   MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event), name);
  73   const uptr hdr = GetThreadTraceHeader(tid);
  74   internal_snprintf(name, sizeof(name), "trace header %u", tid);
  75   MapThreadTrace(hdr, sizeof(Trace), name);
  76   new((void*)hdr) Trace();
  77   // We are going to use only a small part of the trace with the default
  78   // value of history_size. However, the constructor writes to the whole trace.
  79   // Unmap the unused part.
  80   uptr hdr_end = hdr + sizeof(Trace);
  81   hdr_end -= sizeof(TraceHeader) * (kTraceParts - TraceParts());
  82   hdr_end = RoundUp(hdr_end, GetPageSizeCached());
  83   if (hdr_end < hdr + sizeof(Trace))
  84     UnmapOrDie((void*)hdr_end, hdr + sizeof(Trace) - hdr_end);
  85   void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
  86   return new(mem) ThreadContext(tid);
  87 }
  88
  89 #if !SANITIZER_GO
  90 static const u32 kThreadQuarantineSize = 16;
  91 #else
  92 static const u32 kThreadQuarantineSize = 64;
  93 #endif
  94
  95 Context::Context()
  96   : initialized()
  97   , report_mtx(MutexTypeReport, StatMtxReport)
  98   , nreported()
  99   , nmissed_expected()
 100   , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
 101       CreateThreadContext, kMaxTid, kThreadQuarantineSize, kMaxTidReuse))
 102   , racy_mtx(MutexTypeRacy, StatMtxRacy)
 103   , racy_stacks(MBlockRacyStacks)
 104   , racy_addresses(MBlockRacyAddresses)
 105   , fired_suppressions_mtx(MutexTypeFired, StatMtxFired)
 106   , fired_suppressions(8)
 107   , clock_alloc("clock allocator") {
 108 }
 109
 110 // The objects are allocated in TLS, so one may rely on zero-initialization.
 111 ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
 112                          unsigned reuse_count,
 113                          uptr stk_addr, uptr stk_size,
 114                          uptr tls_addr, uptr tls_size)
 115   : fast_state(tid, epoch)
 116   // Do not touch these, rely on zero initialization,
 117   // they may be accessed before the ctor.
 118   // , ignore_reads_and_writes()
 119   // , ignore_interceptors()
 120   , clock(tid, reuse_count)
 121 #if !SANITIZER_GO
 122   , jmp_bufs(MBlockJmpBuf)
 123 #endif
 124   , tid(tid)
 125   , unique_id(unique_id)
 126   , stk_addr(stk_addr)
 127   , stk_size(stk_size)
 128   , tls_addr(tls_addr)
 129   , tls_size(tls_size)
 130 #if !SANITIZER_GO
 131   , last_sleep_clock(tid)
 132 #endif
 133 {
 134 }
 135
 136 #if !SANITIZER_GO
 137 static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
 138   uptr n_threads;
 139   uptr n_running_threads;
 140   ctx->thread_registry->GetNumberOfThreads(&n_threads, &n_running_threads);
 141   InternalScopedBuffer<char> buf(4096);
 142   WriteMemoryProfile(buf.data(), buf.size(), n_threads, n_running_threads);
 143   WriteToFile(fd, buf.data(), internal_strlen(buf.data()));
 144 }
 145
 146 static void BackgroundThread(void *arg) {
 147   // This is a non-initialized non-user thread, nothing to see here.
 148   // We don't use ScopedIgnoreInterceptors, because we want ignores to be
 149   // enabled even when the thread function exits (e.g. during pthread thread
 150   // shutdown code).
 151   cur_thread()->ignore_interceptors++;
 152   const u64 kMs2Ns = 1000 * 1000;
 153
 154   fd_t mprof_fd = kInvalidFd;
 155   if (flags()->profile_memory && flags()->profile_memory[0]) {
 156     if (internal_strcmp(flags()->profile_memory, "stdout") == 0) {
 157       mprof_fd = 1;
 158     } else if (internal_strcmp(flags()->profile_memory, "stderr") == 0) {
 159       mprof_fd = 2;
 160     } else {
 161       InternalScopedString filename(kMaxPathLength);
 162       filename.append("%s.%d", flags()->profile_memory, (int)internal_getpid());
 163       fd_t fd = OpenFile(filename.data(), WrOnly);
 164       if (fd == kInvalidFd) {
 165         Printf("ThreadSanitizer: failed to open memory profile file '%s'\n",
 166             &filename[0]);
 167       } else {
 168         mprof_fd = fd;
 169       }
 170     }
 171   }
 172
 173   u64 last_flush = NanoTime();
 174   uptr last_rss = 0;
 175   for (int i = 0;
 176       atomic_load(&ctx->stop_background_thread, memory_order_relaxed) == 0;
 177       i++) {
 178     SleepForMillis(100);
 179     u64 now = NanoTime();
 180
 181     // Flush memory if requested.
 182     if (flags()->flush_memory_ms > 0) {
 183       if (last_flush + flags()->flush_memory_ms * kMs2Ns < now) {
 184         VPrintf(1, "ThreadSanitizer: periodic memory flush\n");
 185         FlushShadowMemory();
 186         last_flush = NanoTime();
 187       }
 188     }
 189     // GetRSS can be expensive on huge programs, so don't do it every 100ms.
 190     if (flags()->memory_limit_mb > 0) {
 191       uptr rss = GetRSS();
 192       uptr limit = uptr(flags()->memory_limit_mb) << 20;
 193       VPrintf(1, "ThreadSanitizer: memory flush check"
 194                  " RSS=%llu LAST=%llu LIMIT=%llu\n",
 195               (u64)rss >> 20, (u64)last_rss >> 20, (u64)limit >> 20);
 196       if (2 * rss > limit + last_rss) {
 197         VPrintf(1, "ThreadSanitizer: flushing memory due to RSS\n");
 198         FlushShadowMemory();
 199         rss = GetRSS();
 200         VPrintf(1, "ThreadSanitizer: memory flushed RSS=%llu\n", (u64)rss>>20);
 201       }
 202       last_rss = rss;
 203     }
 204
 205     // Write memory profile if requested.
 206     if (mprof_fd != kInvalidFd)
 207       MemoryProfiler(ctx, mprof_fd, i);
 208
 209     // Flush symbolizer cache if requested.
 210     if (flags()->flush_symbolizer_ms > 0) {
 211       u64 last = atomic_load(&ctx->last_symbolize_time_ns,
 212                              memory_order_relaxed);
 213       if (last != 0 && last + flags()->flush_symbolizer_ms * kMs2Ns < now) {
 214         Lock l(&ctx->report_mtx);
 215         SpinMutexLock l2(&CommonSanitizerReportMutex);
 216         SymbolizeFlush();
 217         atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
 218       }
 219     }
 220   }
 221 }
 222
 223 static void StartBackgroundThread() {
 224   ctx->background_thread = internal_start_thread(&BackgroundThread, 0);
 225 }
 226
 227 #ifndef __mips__
 228 static void StopBackgroundThread() {
 229   atomic_store(&ctx->stop_background_thread, 1, memory_order_relaxed);
 230   internal_join_thread(ctx->background_thread);
 231   ctx->background_thread = 0;
 232 }
 233 #endif
 234 #endif
 235
 236 void DontNeedShadowFor(uptr addr, uptr size) {
 237   ReleaseMemoryPagesToOS(MemToShadow(addr), MemToShadow(addr + size));
 238 }
 239
 240 void MapShadow(uptr addr, uptr size) {
 241   // Global data is not 64K aligned, but there are no adjacent mappings,
 242   // so we can get away with unaligned mapping.
 243   // CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
 244   const uptr kPageSize = GetPageSizeCached();
 245   uptr shadow_begin = RoundDownTo((uptr)MemToShadow(addr), kPageSize);
 246   uptr shadow_end = RoundUpTo((uptr)MemToShadow(addr + size), kPageSize);
 247   MmapFixedNoReserve(shadow_begin, shadow_end - shadow_begin, "shadow");
 248
 249   // Meta shadow is 2:1, so tread carefully.
 250   static bool data_mapped = false;
 251   static uptr mapped_meta_end = 0;
 252   uptr meta_begin = (uptr)MemToMeta(addr);
 253   uptr meta_end = (uptr)MemToMeta(addr + size);
 254   meta_begin = RoundDownTo(meta_begin, 64 << 10);
 255   meta_end = RoundUpTo(meta_end, 64 << 10);
 256   if (!data_mapped) {
 257     // First call maps data+bss.
 258     data_mapped = true;
 259     MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow");
 260   } else {
 261     // Mapping continous heap.
 262     // Windows wants 64K alignment.
 263     meta_begin = RoundDownTo(meta_begin, 64 << 10);
 264     meta_end = RoundUpTo(meta_end, 64 << 10);
 265     if (meta_end <= mapped_meta_end)
 266       return;
 267     if (meta_begin < mapped_meta_end)
 268       meta_begin = mapped_meta_end;
 269     MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow");
 270     mapped_meta_end = meta_end;
 271   }
 272   VPrintf(2, "mapped meta shadow for (%p-%p) at (%p-%p)\n",
 273       addr, addr+size, meta_begin, meta_end);
 274 }
 275
 276 void MapThreadTrace(uptr addr, uptr size, const char *name) {
 277   DPrintf("#0: Mapping trace at %p-%p(0x%zx)\n", addr, addr + size, size);
 278   CHECK_GE(addr, TraceMemBeg());
 279   CHECK_LE(addr + size, TraceMemEnd());
 280   CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
 281   uptr addr1 = (uptr)MmapFixedNoReserve(addr, size, name);
 282   if (addr1 != addr) {
 283     Printf("FATAL: ThreadSanitizer can not mmap thread trace (%p/%p->%p)\n",
 284         addr, size, addr1);
 285     Die();
 286   }
 287 }
 288
 289 static void CheckShadowMapping() {
 290   uptr beg, end;
 291   for (int i = 0; GetUserRegion(i, &beg, &end); i++) {
 292     // Skip cases for empty regions (heap definition for architectures that
 293     // do not use 64-bit allocator).
 294     if (beg == end)
 295       continue;
 296     VPrintf(3, "checking shadow region %p-%p\n", beg, end);
 297     uptr prev = 0;
 298     for (uptr p0 = beg; p0 <= end; p0 += (end - beg) / 4) {
 299       for (int x = -(int)kShadowCell; x <= (int)kShadowCell; x += kShadowCell) {
 300         const uptr p = RoundDown(p0 + x, kShadowCell);
 301         if (p < beg || p >= end)
 302           continue;
 303         const uptr s = MemToShadow(p);
 304         const uptr m = (uptr)MemToMeta(p);
 305         VPrintf(3, "  checking pointer %p: shadow=%p meta=%p\n", p, s, m);
 306         CHECK(IsAppMem(p));
 307         CHECK(IsShadowMem(s));
 308         CHECK_EQ(p, ShadowToMem(s));
 309         CHECK(IsMetaMem(m));
 310         if (prev) {
 311           // Ensure that shadow and meta mappings are linear within a single
 312           // user range. Lots of code that processes memory ranges assumes it.
 313           const uptr prev_s = MemToShadow(prev);
 314           const uptr prev_m = (uptr)MemToMeta(prev);
 315           CHECK_EQ(s - prev_s, (p - prev) * kShadowMultiplier);
 316           CHECK_EQ((m - prev_m) / kMetaShadowSize,
 317                    (p - prev) / kMetaShadowCell);
 318         }
 319         prev = p;
 320       }
 321     }
 322   }
 323 }
 324
 325 void Initialize(ThreadState *thr) {
 326   // Thread safe because done before all threads exist.
 327   static bool is_initialized = false;
 328   if (is_initialized)
 329     return;
 330   is_initialized = true;
 331   // We are not ready to handle interceptors yet.
 332   ScopedIgnoreInterceptors ignore;
 333   SanitizerToolName = "ThreadSanitizer";
 334   // Install tool-specific callbacks in sanitizer_common.
 335   SetCheckFailedCallback(TsanCheckFailed);
 336
 337   ctx = new(ctx_placeholder) Context;
 338   const char *options = GetEnv(SANITIZER_GO ? "GORACE" : "TSAN_OPTIONS");
 339   CacheBinaryName();
 340   InitializeFlags(&ctx->flags, options);
 341   AvoidCVE_2016_2143();
 342   InitializePlatformEarly();
 343 #if !SANITIZER_GO
 344   // Re-exec ourselves if we need to set additional env or command line args.
 345   MaybeReexec();
 346
 347   InitializeAllocator();
 348   ReplaceSystemMalloc();
 349 #endif
 350   if (common_flags()->detect_deadlocks)
 351     ctx->dd = DDetector::Create(flags());
 352   Processor *proc = ProcCreate();
 353   ProcWire(proc, thr);
 354   InitializeInterceptors();
 355   CheckShadowMapping();
 356   InitializePlatform();
 357   InitializeMutex();
 358   InitializeDynamicAnnotations();
 359 #if !SANITIZER_GO
 360   InitializeShadowMemory();
 361   InitializeAllocatorLate();
 362 #endif
 363   // Setup correct file descriptor for error reports.
 364   __sanitizer_set_report_path(common_flags()->log_path);
 365   InitializeSuppressions();
 366 #if !SANITIZER_GO
 367   InitializeLibIgnore();
 368   Symbolizer::GetOrInit()->AddHooks(EnterSymbolizer, ExitSymbolizer);
 369   // On MIPS, TSan initialization is run before
 370   // __pthread_initialize_minimal_internal() is finished, so we can not spawn
 371   // new threads.
 372 #ifndef __mips__
 373   StartBackgroundThread();
 374   SetSandboxingCallback(StopBackgroundThread);
 375 #endif
 376 #endif
 377
 378   VPrintf(1, "***** Running under ThreadSanitizer v2 (pid %d) *****\n",
 379           (int)internal_getpid());
 380
 381   // Initialize thread 0.
 382   int tid = ThreadCreate(thr, 0, 0, true);
 383   CHECK_EQ(tid, 0);
 384   ThreadStart(thr, tid, GetTid(), /*workerthread*/ false);
 385 #if TSAN_CONTAINS_UBSAN
 386   __ubsan::InitAsPlugin();
 387 #endif
 388   ctx->initialized = true;
 389
 390 #if !SANITIZER_GO
 391   Symbolizer::LateInitialize();
 392 #endif
 393
 394   if (flags()->stop_on_start) {
 395     Printf("ThreadSanitizer is suspended at startup (pid %d)."
 396            " Call __tsan_resume().\n",
 397            (int)internal_getpid());
 398     while (__tsan_resumed == 0) {}
 399   }
 400
 401   OnInitialize();
 402 }
 403
 404 int Finalize(ThreadState *thr) {
 405   bool failed = false;
 406
 407   if (common_flags()->print_module_map == 1) PrintModuleMap();
 408
 409   if (flags()->atexit_sleep_ms > 0 && ThreadCount(thr) > 1)
 410     SleepForMillis(flags()->atexit_sleep_ms);
 411
 412   // Wait for pending reports.
 413   ctx->report_mtx.Lock();
 414   CommonSanitizerReportMutex.Lock();
 415   CommonSanitizerReportMutex.Unlock();
 416   ctx->report_mtx.Unlock();
 417
 418 #if !SANITIZER_GO
 419   if (Verbosity()) AllocatorPrintStats();
 420 #endif
 421
 422   ThreadFinalize(thr);
 423
 424   if (ctx->nreported) {
 425     failed = true;
 426 #if !SANITIZER_GO
 427     Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
 428 #else
 429     Printf("Found %d data race(s)\n", ctx->nreported);
 430 #endif
 431   }
 432
 433   if (ctx->nmissed_expected) {
 434     failed = true;
 435     Printf("ThreadSanitizer: missed %d expected races\n",
 436         ctx->nmissed_expected);
 437   }
 438
 439   if (common_flags()->print_suppressions)
 440     PrintMatchedSuppressions();
 441 #if !SANITIZER_GO
 442   if (flags()->print_benign)
 443     PrintMatchedBenignRaces();
 444 #endif
 445
 446   failed = OnFinalize(failed);
 447
 448 #if TSAN_COLLECT_STATS
 449   StatAggregate(ctx->stat, thr->stat);
 450   StatOutput(ctx->stat);
 451 #endif
 452
 453   return failed ? common_flags()->exitcode : 0;
 454 }
 455
 456 #if !SANITIZER_GO
 457 void ForkBefore(ThreadState *thr, uptr pc) {
 458   ctx->thread_registry->Lock();
 459   ctx->report_mtx.Lock();
 460 }
 461
 462 void ForkParentAfter(ThreadState *thr, uptr pc) {
 463   ctx->report_mtx.Unlock();
 464   ctx->thread_registry->Unlock();
 465 }
 466
 467 void ForkChildAfter(ThreadState *thr, uptr pc) {
 468   ctx->report_mtx.Unlock();
 469   ctx->thread_registry->Unlock();
 470
 471   uptr nthread = 0;
 472   ctx->thread_registry->GetNumberOfThreads(0, 0, &nthread /* alive threads */);
 473   VPrintf(1, "ThreadSanitizer: forked new process with pid %d,"
 474       " parent had %d threads\n", (int)internal_getpid(), (int)nthread);
 475   if (nthread == 1) {
 476     StartBackgroundThread();
 477   } else {
 478     // We've just forked a multi-threaded process. We cannot reasonably function
 479     // after that (some mutexes may be locked before fork). So just enable
 480     // ignores for everything in the hope that we will exec soon.
 481     ctx->after_multithreaded_fork = true;
 482     thr->ignore_interceptors++;
 483     ThreadIgnoreBegin(thr, pc);
 484     ThreadIgnoreSyncBegin(thr, pc);
 485   }
 486 }
 487 #endif
 488
 489 #if SANITIZER_GO
 490 NOINLINE
 491 void GrowShadowStack(ThreadState *thr) {
 492   const int sz = thr->shadow_stack_end - thr->shadow_stack;
 493   const int newsz = 2 * sz;
 494   uptr *newstack = (uptr*)internal_alloc(MBlockShadowStack,
 495       newsz * sizeof(uptr));
 496   internal_memcpy(newstack, thr->shadow_stack, sz * sizeof(uptr));
 497   internal_free(thr->shadow_stack);
 498   thr->shadow_stack = newstack;
 499   thr->shadow_stack_pos = newstack + sz;
 500   thr->shadow_stack_end = newstack + newsz;
 501 }
 502 #endif
 503
 504 u32 CurrentStackId(ThreadState *thr, uptr pc) {
 505   if (!thr->is_inited)  // May happen during bootstrap.
 506     return 0;
 507   if (pc != 0) {
 508 #if !SANITIZER_GO
 509     DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 510 #else
 511     if (thr->shadow_stack_pos == thr->shadow_stack_end)
 512       GrowShadowStack(thr);
 513 #endif
 514     thr->shadow_stack_pos[0] = pc;
 515     thr->shadow_stack_pos++;
 516   }
 517   u32 id = StackDepotPut(
 518       StackTrace(thr->shadow_stack, thr->shadow_stack_pos - thr->shadow_stack));
 519   if (pc != 0)
 520     thr->shadow_stack_pos--;
 521   return id;
 522 }
 523
 524 void TraceSwitch(ThreadState *thr) {
 525   thr->nomalloc++;
 526   Trace *thr_trace = ThreadTrace(thr->tid);
 527   Lock l(&thr_trace->mtx);
 528   unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % TraceParts();
 529   TraceHeader *hdr = &thr_trace->headers[trace];
 530   hdr->epoch0 = thr->fast_state.epoch();
 531   ObtainCurrentStack(thr, 0, &hdr->stack0);
 532   hdr->mset0 = thr->mset;
 533   thr->nomalloc--;
 534 }
 535
 536 Trace *ThreadTrace(int tid) {
 537   return (Trace*)GetThreadTraceHeader(tid);
 538 }
 539
 540 uptr TraceTopPC(ThreadState *thr) {
 541   Event *events = (Event*)GetThreadTrace(thr->tid);
 542   uptr pc = events[thr->fast_state.GetTracePos()];
 543   return pc;
 544 }
 545
 546 uptr TraceSize() {
 547   return (uptr)(1ull << (kTracePartSizeBits + flags()->history_size + 1));
 548 }
 549
 550 uptr TraceParts() {
 551   return TraceSize() / kTracePartSize;
 552 }
 553
 554 #if !SANITIZER_GO
 555 extern "C" void __tsan_trace_switch() {
 556   TraceSwitch(cur_thread());
 557 }
 558
 559 extern "C" void __tsan_report_race() {
 560   ReportRace(cur_thread());
 561 }
 562 #endif
 563
 564 ALWAYS_INLINE
 565 Shadow LoadShadow(u64 *p) {
 566   u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
 567   return Shadow(raw);
 568 }
 569
 570 ALWAYS_INLINE
 571 void StoreShadow(u64 *sp, u64 s) {
 572   atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
 573 }
 574
 575 ALWAYS_INLINE
 576 void StoreIfNotYetStored(u64 *sp, u64 *s) {
 577   StoreShadow(sp, *s);
 578   *s = 0;
 579 }
 580
 581 ALWAYS_INLINE
 582 void HandleRace(ThreadState *thr, u64 *shadow_mem,
 583                               Shadow cur, Shadow old) {
 584   thr->racy_state[0] = cur.raw();
 585   thr->racy_state[1] = old.raw();
 586   thr->racy_shadow_addr = shadow_mem;
 587 #if !SANITIZER_GO
 588   HACKY_CALL(__tsan_report_race);
 589 #else
 590   ReportRace(thr);
 591 #endif
 592 }
 593
 594 static inline bool HappensBefore(Shadow old, ThreadState *thr) {
 595   return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
 596 }
 597
 598 ALWAYS_INLINE
 599 void MemoryAccessImpl1(ThreadState *thr, uptr addr,
 600     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 601     u64 *shadow_mem, Shadow cur) {
 602   StatInc(thr, StatMop);
 603   StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 604   StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 605
 606   // This potentially can live in an MMX/SSE scratch register.
 607   // The required intrinsics are:
 608   // __m128i _mm_move_epi64(__m128i*);
 609   // _mm_storel_epi64(u64*, __m128i);
 610   u64 store_word = cur.raw();
 611
 612   // scan all the shadow values and dispatch to 4 categories:
 613   // same, replace, candidate and race (see comments below).
 614   // we consider only 3 cases regarding access sizes:
 615   // equal, intersect and not intersect. initially I considered
 616   // larger and smaller as well, it allowed to replace some
 617   // 'candidates' with 'same' or 'replace', but I think
 618   // it's just not worth it (performance- and complexity-wise).
 619
 620   Shadow old(0);
 621
 622   // It release mode we manually unroll the loop,
 623   // because empirically gcc generates better code this way.
 624   // However, we can't afford unrolling in debug mode, because the function
 625   // consumes almost 4K of stack. Gtest gives only 4K of stack to death test
 626   // threads, which is not enough for the unrolled loop.
 627 #if SANITIZER_DEBUG
 628   for (int idx = 0; idx < 4; idx++) {
 629 #include "tsan_update_shadow_word_inl.h"
 630   }
 631 #else
 632   int idx = 0;
 633 #include "tsan_update_shadow_word_inl.h"
 634   idx = 1;
 635 #include "tsan_update_shadow_word_inl.h"
 636   idx = 2;
 637 #include "tsan_update_shadow_word_inl.h"
 638   idx = 3;
 639 #include "tsan_update_shadow_word_inl.h"
 640 #endif
 641
 642   // we did not find any races and had already stored
 643   // the current access info, so we are done
 644   if (LIKELY(store_word == 0))
 645     return;
 646   // choose a random candidate slot and replace it
 647   StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
 648   StatInc(thr, StatShadowReplace);
 649   return;
 650  RACE:
 651   HandleRace(thr, shadow_mem, cur, old);
 652   return;
 653 }
 654
 655 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 656     int size, bool kAccessIsWrite, bool kIsAtomic) {
 657   while (size) {
 658     int size1 = 1;
 659     int kAccessSizeLog = kSizeLog1;
 660     if (size >= 8 && (addr & ~7) == ((addr + 7) & ~7)) {
 661       size1 = 8;
 662       kAccessSizeLog = kSizeLog8;
 663     } else if (size >= 4 && (addr & ~7) == ((addr + 3) & ~7)) {
 664       size1 = 4;
 665       kAccessSizeLog = kSizeLog4;
 666     } else if (size >= 2 && (addr & ~7) == ((addr + 1) & ~7)) {
 667       size1 = 2;
 668       kAccessSizeLog = kSizeLog2;
 669     }
 670     MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
 671     addr += size1;
 672     size -= size1;
 673   }
 674 }
 675
 676 ALWAYS_INLINE
 677 bool ContainsSameAccessSlow(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 678   Shadow cur(a);
 679   for (uptr i = 0; i < kShadowCnt; i++) {
 680     Shadow old(LoadShadow(&s[i]));
 681     if (Shadow::Addr0AndSizeAreEqual(cur, old) &&
 682         old.TidWithIgnore() == cur.TidWithIgnore() &&
 683         old.epoch() > sync_epoch &&
 684         old.IsAtomic() == cur.IsAtomic() &&
 685         old.IsRead() <= cur.IsRead())
 686       return true;
 687   }
 688   return false;
 689 }
 690
 691 #if defined(__SSE3__)
 692 #define SHUF(v0, v1, i0, i1, i2, i3) _mm_castps_si128(_mm_shuffle_ps( \
 693     _mm_castsi128_ps(v0), _mm_castsi128_ps(v1), \
 694     (i0)*1 + (i1)*4 + (i2)*16 + (i3)*64))
 695 ALWAYS_INLINE
 696 bool ContainsSameAccessFast(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 697   // This is an optimized version of ContainsSameAccessSlow.
 698   // load current access into access[0:63]
 699   const m128 access     = _mm_cvtsi64_si128(a);
 700   // duplicate high part of access in addr0:
 701   // addr0[0:31]        = access[32:63]
 702   // addr0[32:63]       = access[32:63]
 703   // addr0[64:95]       = access[32:63]
 704   // addr0[96:127]      = access[32:63]
 705   const m128 addr0      = SHUF(access, access, 1, 1, 1, 1);
 706   // load 4 shadow slots
 707   const m128 shadow0    = _mm_load_si128((__m128i*)s);
 708   const m128 shadow1    = _mm_load_si128((__m128i*)s + 1);
 709   // load high parts of 4 shadow slots into addr_vect:
 710   // addr_vect[0:31]    = shadow0[32:63]
 711   // addr_vect[32:63]   = shadow0[96:127]
 712   // addr_vect[64:95]   = shadow1[32:63]
 713   // addr_vect[96:127]  = shadow1[96:127]
 714   m128 addr_vect        = SHUF(shadow0, shadow1, 1, 3, 1, 3);
 715   if (!is_write) {
 716     // set IsRead bit in addr_vect
 717     const m128 rw_mask1 = _mm_cvtsi64_si128(1<<15);
 718     const m128 rw_mask  = SHUF(rw_mask1, rw_mask1, 0, 0, 0, 0);
 719     addr_vect           = _mm_or_si128(addr_vect, rw_mask);
 720   }
 721   // addr0 == addr_vect?
 722   const m128 addr_res   = _mm_cmpeq_epi32(addr0, addr_vect);
 723   // epoch1[0:63]       = sync_epoch
 724   const m128 epoch1     = _mm_cvtsi64_si128(sync_epoch);
 725   // epoch[0:31]        = sync_epoch[0:31]
 726   // epoch[32:63]       = sync_epoch[0:31]
 727   // epoch[64:95]       = sync_epoch[0:31]
 728   // epoch[96:127]      = sync_epoch[0:31]
 729   const m128 epoch      = SHUF(epoch1, epoch1, 0, 0, 0, 0);
 730   // load low parts of shadow cell epochs into epoch_vect:
 731   // epoch_vect[0:31]   = shadow0[0:31]
 732   // epoch_vect[32:63]  = shadow0[64:95]
 733   // epoch_vect[64:95]  = shadow1[0:31]
 734   // epoch_vect[96:127] = shadow1[64:95]
 735   const m128 epoch_vect = SHUF(shadow0, shadow1, 0, 2, 0, 2);
 736   // epoch_vect >= sync_epoch?
 737   const m128 epoch_res  = _mm_cmpgt_epi32(epoch_vect, epoch);
 738   // addr_res & epoch_res
 739   const m128 res        = _mm_and_si128(addr_res, epoch_res);
 740   // mask[0] = res[7]
 741   // mask[1] = res[15]
 742   // ...
 743   // mask[15] = res[127]
 744   const int mask        = _mm_movemask_epi8(res);
 745   return mask != 0;
 746 }
 747 #endif
 748
 749 ALWAYS_INLINE
 750 bool ContainsSameAccess(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
 751 #if defined(__SSE3__)
 752   bool res = ContainsSameAccessFast(s, a, sync_epoch, is_write);
 753   // NOTE: this check can fail if the shadow is concurrently mutated
 754   // by other threads. But it still can be useful if you modify
 755   // ContainsSameAccessFast and want to ensure that it's not completely broken.
 756   // DCHECK_EQ(res, ContainsSameAccessSlow(s, a, sync_epoch, is_write));
 757   return res;
 758 #else
 759   return ContainsSameAccessSlow(s, a, sync_epoch, is_write);
 760 #endif
 761 }
 762
 763 ALWAYS_INLINE USED
 764 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
 765     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic) {
 766   u64 *shadow_mem = (u64*)MemToShadow(addr);
 767   DPrintf2("#%d: MemoryAccess: @%p %p size=%d"
 768       " is_write=%d shadow_mem=%p {%zx, %zx, %zx, %zx}\n",
 769       (int)thr->fast_state.tid(), (void*)pc, (void*)addr,
 770       (int)(1 << kAccessSizeLog), kAccessIsWrite, shadow_mem,
 771       (uptr)shadow_mem[0], (uptr)shadow_mem[1],
 772       (uptr)shadow_mem[2], (uptr)shadow_mem[3]);
 773 #if SANITIZER_DEBUG
 774   if (!IsAppMem(addr)) {
 775     Printf("Access to non app mem %zx\n", addr);
 776     DCHECK(IsAppMem(addr));
 777   }
 778   if (!IsShadowMem((uptr)shadow_mem)) {
 779     Printf("Bad shadow addr %p (%zx)\n", shadow_mem, addr);
 780     DCHECK(IsShadowMem((uptr)shadow_mem));
 781   }
 782 #endif
 783
 784   if (!SANITIZER_GO && *shadow_mem == kShadowRodata) {
 785     // Access to .rodata section, no races here.
 786     // Measurements show that it can be 10-20% of all memory accesses.
 787     StatInc(thr, StatMop);
 788     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 789     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 790     StatInc(thr, StatMopRodata);
 791     return;
 792   }
 793
 794   FastState fast_state = thr->fast_state;
 795   if (fast_state.GetIgnoreBit()) {
 796     StatInc(thr, StatMop);
 797     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 798     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 799     StatInc(thr, StatMopIgnored);
 800     return;
 801   }
 802
 803   Shadow cur(fast_state);
 804   cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
 805   cur.SetWrite(kAccessIsWrite);
 806   cur.SetAtomic(kIsAtomic);
 807
 808   if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
 809       thr->fast_synch_epoch, kAccessIsWrite))) {
 810     StatInc(thr, StatMop);
 811     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 812     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 813     StatInc(thr, StatMopSame);
 814     return;
 815   }
 816
 817   if (kCollectHistory) {
 818     fast_state.IncrementEpoch();
 819     thr->fast_state = fast_state;
 820     TraceAddEvent(thr, fast_state, EventTypeMop, pc);
 821     cur.IncrementEpoch();
 822   }
 823
 824   MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
 825       shadow_mem, cur);
 826 }
 827
 828 // Called by MemoryAccessRange in tsan_rtl_thread.cc
 829 ALWAYS_INLINE USED
 830 void MemoryAccessImpl(ThreadState *thr, uptr addr,
 831     int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
 832     u64 *shadow_mem, Shadow cur) {
 833   if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
 834       thr->fast_synch_epoch, kAccessIsWrite))) {
 835     StatInc(thr, StatMop);
 836     StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
 837     StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
 838     StatInc(thr, StatMopSame);
 839     return;
 840   }
 841
 842   MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
 843       shadow_mem, cur);
 844 }
 845
 846 static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
 847                            u64 val) {
 848   (void)thr;
 849   (void)pc;
 850   if (size == 0)
 851     return;
 852   // FIXME: fix me.
 853   uptr offset = addr % kShadowCell;
 854   if (offset) {
 855     offset = kShadowCell - offset;
 856     if (size <= offset)
 857       return;
 858     addr += offset;
 859     size -= offset;
 860   }
 861   DCHECK_EQ(addr % 8, 0);
 862   // If a user passes some insane arguments (memset(0)),
 863   // let it just crash as usual.
 864   if (!IsAppMem(addr) || !IsAppMem(addr + size - 1))
 865     return;
 866   // Don't want to touch lots of shadow memory.
 867   // If a program maps 10MB stack, there is no need reset the whole range.
 868   size = (size + (kShadowCell - 1)) & ~(kShadowCell - 1);
 869   // UnmapOrDie/MmapFixedNoReserve does not work on Windows.
 870   if (SANITIZER_WINDOWS || size < common_flags()->clear_shadow_mmap_threshold) {
 871     u64 *p = (u64*)MemToShadow(addr);
 872     CHECK(IsShadowMem((uptr)p));
 873     CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
 874     // FIXME: may overwrite a part outside the region
 875     for (uptr i = 0; i < size / kShadowCell * kShadowCnt;) {
 876       p[i++] = val;
 877       for (uptr j = 1; j < kShadowCnt; j++)
 878         p[i++] = 0;
 879     }
 880   } else {
 881     // The region is big, reset only beginning and end.
 882     const uptr kPageSize = GetPageSizeCached();
 883     u64 *begin = (u64*)MemToShadow(addr);
 884     u64 *end = begin + size / kShadowCell * kShadowCnt;
 885     u64 *p = begin;
 886     // Set at least first kPageSize/2 to page boundary.
 887     while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
 888       *p++ = val;
 889       for (uptr j = 1; j < kShadowCnt; j++)
 890         *p++ = 0;
 891     }
 892     // Reset middle part.
 893     u64 *p1 = p;
 894     p = RoundDown(end, kPageSize);
 895     UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
 896     MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1);
 897     // Set the ending.
 898     while (p < end) {
 899       *p++ = val;
 900       for (uptr j = 1; j < kShadowCnt; j++)
 901         *p++ = 0;
 902     }
 903   }
 904 }
 905
 906 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 907   MemoryRangeSet(thr, pc, addr, size, 0);
 908 }
 909
 910 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 911   // Processing more than 1k (4k of shadow) is expensive,
 912   // can cause excessive memory consumption (user does not necessary touch
 913   // the whole range) and most likely unnecessary.
 914   if (size > 1024)
 915     size = 1024;
 916   CHECK_EQ(thr->is_freeing, false);
 917   thr->is_freeing = true;
 918   MemoryAccessRange(thr, pc, addr, size, true);
 919   thr->is_freeing = false;
 920   if (kCollectHistory) {
 921     thr->fast_state.IncrementEpoch();
 922     TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
 923   }
 924   Shadow s(thr->fast_state);
 925   s.ClearIgnoreBit();
 926   s.MarkAsFreed();
 927   s.SetWrite(true);
 928   s.SetAddr0AndSizeLog(0, 3);
 929   MemoryRangeSet(thr, pc, addr, size, s.raw());
 930 }
 931
 932 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 933   if (kCollectHistory) {
 934     thr->fast_state.IncrementEpoch();
 935     TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
 936   }
 937   Shadow s(thr->fast_state);
 938   s.ClearIgnoreBit();
 939   s.SetWrite(true);
 940   s.SetAddr0AndSizeLog(0, 3);
 941   MemoryRangeSet(thr, pc, addr, size, s.raw());
 942 }
 943
 944 ALWAYS_INLINE USED
 945 void FuncEntry(ThreadState *thr, uptr pc) {
 946   StatInc(thr, StatFuncEnter);
 947   DPrintf2("#%d: FuncEntry %p\n", (int)thr->fast_state.tid(), (void*)pc);
 948   if (kCollectHistory) {
 949     thr->fast_state.IncrementEpoch();
 950     TraceAddEvent(thr, thr->fast_state, EventTypeFuncEnter, pc);
 951   }
 952
 953   // Shadow stack maintenance can be replaced with
 954   // stack unwinding during trace switch (which presumably must be faster).
 955   DCHECK_GE(thr->shadow_stack_pos, thr->shadow_stack);
 956 #if !SANITIZER_GO
 957   DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 958 #else
 959   if (thr->shadow_stack_pos == thr->shadow_stack_end)
 960     GrowShadowStack(thr);
 961 #endif
 962   thr->shadow_stack_pos[0] = pc;
 963   thr->shadow_stack_pos++;
 964 }
 965
 966 ALWAYS_INLINE USED
 967 void FuncExit(ThreadState *thr) {
 968   StatInc(thr, StatFuncExit);
 969   DPrintf2("#%d: FuncExit\n", (int)thr->fast_state.tid());
 970   if (kCollectHistory) {
 971     thr->fast_state.IncrementEpoch();
 972     TraceAddEvent(thr, thr->fast_state, EventTypeFuncExit, 0);
 973   }
 974
 975   DCHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
 976 #if !SANITIZER_GO
 977   DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
 978 #endif
 979   thr->shadow_stack_pos--;
 980 }
 981
 982 void ThreadIgnoreBegin(ThreadState *thr, uptr pc, bool save_stack) {
 983   DPrintf("#%d: ThreadIgnoreBegin\n", thr->tid);
 984   thr->ignore_reads_and_writes++;
 985   CHECK_GT(thr->ignore_reads_and_writes, 0);
 986   thr->fast_state.SetIgnoreBit();
 987 #if !SANITIZER_GO
 988   if (save_stack && !ctx->after_multithreaded_fork)
 989     thr->mop_ignore_set.Add(CurrentStackId(thr, pc));
 990 #endif
 991 }
 992
 993 void ThreadIgnoreEnd(ThreadState *thr, uptr pc) {
 994   DPrintf("#%d: ThreadIgnoreEnd\n", thr->tid);
 995   CHECK_GT(thr->ignore_reads_and_writes, 0);
 996   thr->ignore_reads_and_writes--;
 997   if (thr->ignore_reads_and_writes == 0) {
 998     thr->fast_state.ClearIgnoreBit();
 999 #if !SANITIZER_GO
1000     thr->mop_ignore_set.Reset();
1001 #endif
1002   }
1003 }
1004
1005 #if !SANITIZER_GO
1006 extern "C" SANITIZER_INTERFACE_ATTRIBUTE
1007 uptr __tsan_testonly_shadow_stack_current_size() {
1008   ThreadState *thr = cur_thread();
1009   return thr->shadow_stack_pos - thr->shadow_stack;
1010 }
1011 #endif
1012
1013 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack) {
1014   DPrintf("#%d: ThreadIgnoreSyncBegin\n", thr->tid);
1015   thr->ignore_sync++;
1016   CHECK_GT(thr->ignore_sync, 0);
1017 #if !SANITIZER_GO
1018   if (save_stack && !ctx->after_multithreaded_fork)
1019     thr->sync_ignore_set.Add(CurrentStackId(thr, pc));
1020 #endif
1021 }
1022
1023 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc) {
1024   DPrintf("#%d: ThreadIgnoreSyncEnd\n", thr->tid);
1025   CHECK_GT(thr->ignore_sync, 0);
1026   thr->ignore_sync--;
1027 #if !SANITIZER_GO
1028   if (thr->ignore_sync == 0)
1029     thr->sync_ignore_set.Reset();
1030 #endif
1031 }
1032
1033 bool MD5Hash::operator==(const MD5Hash &other) const {
1034   return hash[0] == other.hash[0] && hash[1] == other.hash[1];
1035 }
1036
1037 #if SANITIZER_DEBUG
1038 void build_consistency_debug() {}
1039 #else
1040 void build_consistency_release() {}
1041 #endif
1042
1043 #if TSAN_COLLECT_STATS
1044 void build_consistency_stats() {}
1045 #else
1046 void build_consistency_nostats() {}
1047 #endif
1048
1049 }  // namespace __tsan
1050
1051 #if !SANITIZER_GO
1052 // Must be included in this file to make sure everything is inlined.
1053 #include "tsan_interface_inl.h"
1054 #endif