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