tsan_rtl.cpp revision bf6963673a64b343707bcf31f709319df3bb7502
1//===-- tsan_rtl.cpp ------------------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file is a part of ThreadSanitizer (TSan), a race detector.
10//
11// Main file (entry points) for the TSan run-time.
12//===----------------------------------------------------------------------===//
13
14#include "sanitizer_common/sanitizer_atomic.h"
15#include "sanitizer_common/sanitizer_common.h"
16#include "sanitizer_common/sanitizer_file.h"
17#include "sanitizer_common/sanitizer_libc.h"
18#include "sanitizer_common/sanitizer_stackdepot.h"
19#include "sanitizer_common/sanitizer_placement_new.h"
20#include "sanitizer_common/sanitizer_symbolizer.h"
21#include "tsan_defs.h"
22#include "tsan_platform.h"
23#include "tsan_rtl.h"
24#include "tsan_mman.h"
25#include "tsan_suppressions.h"
26#include "tsan_symbolize.h"
27#include "ubsan/ubsan_init.h"
28
29#ifdef __SSE3__
30// <emmintrin.h> transitively includes <stdlib.h>,
31// and it's prohibited to include std headers into tsan runtime.
32// So we do this dirty trick.
33#define _MM_MALLOC_H_INCLUDED
34#define __MM_MALLOC_H
35#include <emmintrin.h>
36typedef __m128i m128;
37#endif
38
39volatile int __tsan_resumed = 0;
40
41extern "C" void __tsan_resume() {
42  __tsan_resumed = 1;
43}
44
45namespace __tsan {
46
47#if !SANITIZER_GO && !SANITIZER_MAC
48__attribute__((tls_model("initial-exec")))
49THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
50#endif
51static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
52Context *ctx;
53
54// Can be overriden by a front-end.
55#ifdef TSAN_EXTERNAL_HOOKS
56bool OnFinalize(bool failed);
57void OnInitialize();
58#else
59SANITIZER_WEAK_CXX_DEFAULT_IMPL
60bool OnFinalize(bool failed) {
61  return failed;
62}
63SANITIZER_WEAK_CXX_DEFAULT_IMPL
64void OnInitialize() {}
65#endif
66
67static char thread_registry_placeholder[sizeof(ThreadRegistry)];
68
69static ThreadContextBase *CreateThreadContext(u32 tid) {
70  // Map thread trace when context is created.
71  char name[50];
72  internal_snprintf(name, sizeof(name), "trace %u", tid);
73  MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event), name);
74  const uptr hdr = GetThreadTraceHeader(tid);
75  internal_snprintf(name, sizeof(name), "trace header %u", tid);
76  MapThreadTrace(hdr, sizeof(Trace), name);
77  new((void*)hdr) Trace();
78  // We are going to use only a small part of the trace with the default
79  // value of history_size. However, the constructor writes to the whole trace.
80  // Unmap the unused part.
81  uptr hdr_end = hdr + sizeof(Trace);
82  hdr_end -= sizeof(TraceHeader) * (kTraceParts - TraceParts());
83  hdr_end = RoundUp(hdr_end, GetPageSizeCached());
84  if (hdr_end < hdr + sizeof(Trace))
85    UnmapOrDie((void*)hdr_end, hdr + sizeof(Trace) - hdr_end);
86  void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
87  return new(mem) ThreadContext(tid);
88}
89
90#if !SANITIZER_GO
91static const u32 kThreadQuarantineSize = 16;
92#else
93static const u32 kThreadQuarantineSize = 64;
94#endif
95
96Context::Context()
97  : initialized()
98  , report_mtx(MutexTypeReport, StatMtxReport)
99  , nreported()
100  , nmissed_expected()
101  , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
102      CreateThreadContext, kMaxTid, kThreadQuarantineSize, kMaxTidReuse))
103  , racy_mtx(MutexTypeRacy, StatMtxRacy)
104  , racy_stacks()
105  , racy_addresses()
106  , fired_suppressions_mtx(MutexTypeFired, StatMtxFired)
107  , clock_alloc("clock allocator") {
108  fired_suppressions.reserve(8);
109}
110
111// The objects are allocated in TLS, so one may rely on zero-initialization.
112ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
113                         unsigned reuse_count,
114                         uptr stk_addr, uptr stk_size,
115                         uptr tls_addr, uptr tls_size)
116  : fast_state(tid, epoch)
117  // Do not touch these, rely on zero initialization,
118  // they may be accessed before the ctor.
119  // , ignore_reads_and_writes()
120  // , ignore_interceptors()
121  , clock(tid, reuse_count)
122#if !SANITIZER_GO
123  , jmp_bufs()
124#endif
125  , tid(tid)
126  , unique_id(unique_id)
127  , stk_addr(stk_addr)
128  , stk_size(stk_size)
129  , tls_addr(tls_addr)
130  , tls_size(tls_size)
131#if !SANITIZER_GO
132  , last_sleep_clock(tid)
133#endif
134{
135}
136
137#if !SANITIZER_GO
138static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
139  uptr n_threads;
140  uptr n_running_threads;
141  ctx->thread_registry->GetNumberOfThreads(&n_threads, &n_running_threads);
142  InternalMmapVector<char> buf(4096);
143  WriteMemoryProfile(buf.data(), buf.size(), n_threads, n_running_threads);
144  WriteToFile(fd, buf.data(), internal_strlen(buf.data()));
145}
146
147static void BackgroundThread(void *arg) {
148  // This is a non-initialized non-user thread, nothing to see here.
149  // We don't use ScopedIgnoreInterceptors, because we want ignores to be
150  // enabled even when the thread function exits (e.g. during pthread thread
151  // shutdown code).
152  cur_thread_init();
153  cur_thread()->ignore_interceptors++;
154  const u64 kMs2Ns = 1000 * 1000;
155
156  fd_t mprof_fd = kInvalidFd;
157  if (flags()->profile_memory && flags()->profile_memory[0]) {
158    if (internal_strcmp(flags()->profile_memory, "stdout") == 0) {
159      mprof_fd = 1;
160    } else if (internal_strcmp(flags()->profile_memory, "stderr") == 0) {
161      mprof_fd = 2;
162    } else {
163      InternalScopedString filename(kMaxPathLength);
164      filename.append("%s.%d", flags()->profile_memory, (int)internal_getpid());
165      fd_t fd = OpenFile(filename.data(), WrOnly);
166      if (fd == kInvalidFd) {
167        Printf("ThreadSanitizer: failed to open memory profile file '%s'\n",
168            &filename[0]);
169      } else {
170        mprof_fd = fd;
171      }
172    }
173  }
174
175  u64 last_flush = NanoTime();
176  uptr last_rss = 0;
177  for (int i = 0;
178      atomic_load(&ctx->stop_background_thread, memory_order_relaxed) == 0;
179      i++) {
180    SleepForMillis(100);
181    u64 now = NanoTime();
182
183    // Flush memory if requested.
184    if (flags()->flush_memory_ms > 0) {
185      if (last_flush + flags()->flush_memory_ms * kMs2Ns < now) {
186        VPrintf(1, "ThreadSanitizer: periodic memory flush\n");
187        FlushShadowMemory();
188        last_flush = NanoTime();
189      }
190    }
191    // GetRSS can be expensive on huge programs, so don't do it every 100ms.
192    if (flags()->memory_limit_mb > 0) {
193      uptr rss = GetRSS();
194      uptr limit = uptr(flags()->memory_limit_mb) << 20;
195      VPrintf(1, "ThreadSanitizer: memory flush check"
196                 " RSS=%llu LAST=%llu LIMIT=%llu\n",
197              (u64)rss >> 20, (u64)last_rss >> 20, (u64)limit >> 20);
198      if (2 * rss > limit + last_rss) {
199        VPrintf(1, "ThreadSanitizer: flushing memory due to RSS\n");
200        FlushShadowMemory();
201        rss = GetRSS();
202        VPrintf(1, "ThreadSanitizer: memory flushed RSS=%llu\n", (u64)rss>>20);
203      }
204      last_rss = rss;
205    }
206
207    // Write memory profile if requested.
208    if (mprof_fd != kInvalidFd)
209      MemoryProfiler(ctx, mprof_fd, i);
210
211    // Flush symbolizer cache if requested.
212    if (flags()->flush_symbolizer_ms > 0) {
213      u64 last = atomic_load(&ctx->last_symbolize_time_ns,
214                             memory_order_relaxed);
215      if (last != 0 && last + flags()->flush_symbolizer_ms * kMs2Ns < now) {
216        Lock l(&ctx->report_mtx);
217        ScopedErrorReportLock l2;
218        SymbolizeFlush();
219        atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
220      }
221    }
222  }
223}
224
225static void StartBackgroundThread() {
226  ctx->background_thread = internal_start_thread(&BackgroundThread, 0);
227}
228
229#ifndef __mips__
230static void StopBackgroundThread() {
231  atomic_store(&ctx->stop_background_thread, 1, memory_order_relaxed);
232  internal_join_thread(ctx->background_thread);
233  ctx->background_thread = 0;
234}
235#endif
236#endif
237
238void DontNeedShadowFor(uptr addr, uptr size) {
239  ReleaseMemoryPagesToOS(MemToShadow(addr), MemToShadow(addr + size));
240}
241
242#if !SANITIZER_GO
243void UnmapShadow(ThreadState *thr, uptr addr, uptr size) {
244  if (size == 0) return;
245  DontNeedShadowFor(addr, size);
246  ScopedGlobalProcessor sgp;
247  ctx->metamap.ResetRange(thr->proc(), addr, size);
248}
249#endif
250
251void MapShadow(uptr addr, uptr size) {
252  // Global data is not 64K aligned, but there are no adjacent mappings,
253  // so we can get away with unaligned mapping.
254  // CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
255  const uptr kPageSize = GetPageSizeCached();
256  uptr shadow_begin = RoundDownTo((uptr)MemToShadow(addr), kPageSize);
257  uptr shadow_end = RoundUpTo((uptr)MemToShadow(addr + size), kPageSize);
258  if (!MmapFixedNoReserve(shadow_begin, shadow_end - shadow_begin, "shadow"))
259    Die();
260
261  // Meta shadow is 2:1, so tread carefully.
262  static bool data_mapped = false;
263  static uptr mapped_meta_end = 0;
264  uptr meta_begin = (uptr)MemToMeta(addr);
265  uptr meta_end = (uptr)MemToMeta(addr + size);
266  meta_begin = RoundDownTo(meta_begin, 64 << 10);
267  meta_end = RoundUpTo(meta_end, 64 << 10);
268  if (!data_mapped) {
269    // First call maps data+bss.
270    data_mapped = true;
271    if (!MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow"))
272      Die();
273  } else {
274    // Mapping continous heap.
275    // Windows wants 64K alignment.
276    meta_begin = RoundDownTo(meta_begin, 64 << 10);
277    meta_end = RoundUpTo(meta_end, 64 << 10);
278    if (meta_end <= mapped_meta_end)
279      return;
280    if (meta_begin < mapped_meta_end)
281      meta_begin = mapped_meta_end;
282    if (!MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow"))
283      Die();
284    mapped_meta_end = meta_end;
285  }
286  VPrintf(2, "mapped meta shadow for (%p-%p) at (%p-%p)\n",
287      addr, addr+size, meta_begin, meta_end);
288}
289
290void MapThreadTrace(uptr addr, uptr size, const char *name) {
291  DPrintf("#0: Mapping trace at %p-%p(0x%zx)\n", addr, addr + size, size);
292  CHECK_GE(addr, TraceMemBeg());
293  CHECK_LE(addr + size, TraceMemEnd());
294  CHECK_EQ(addr, addr & ~((64 << 10) - 1));  // windows wants 64K alignment
295  if (!MmapFixedNoReserve(addr, size, name)) {
296    Printf("FATAL: ThreadSanitizer can not mmap thread trace (%p/%p)\n",
297        addr, size);
298    Die();
299  }
300}
301
302static void CheckShadowMapping() {
303  uptr beg, end;
304  for (int i = 0; GetUserRegion(i, &beg, &end); i++) {
305    // Skip cases for empty regions (heap definition for architectures that
306    // do not use 64-bit allocator).
307    if (beg == end)
308      continue;
309    VPrintf(3, "checking shadow region %p-%p\n", beg, end);
310    uptr prev = 0;
311    for (uptr p0 = beg; p0 <= end; p0 += (end - beg) / 4) {
312      for (int x = -(int)kShadowCell; x <= (int)kShadowCell; x += kShadowCell) {
313        const uptr p = RoundDown(p0 + x, kShadowCell);
314        if (p < beg || p >= end)
315          continue;
316        const uptr s = MemToShadow(p);
317        const uptr m = (uptr)MemToMeta(p);
318        VPrintf(3, "  checking pointer %p: shadow=%p meta=%p\n", p, s, m);
319        CHECK(IsAppMem(p));
320        CHECK(IsShadowMem(s));
321        CHECK_EQ(p, ShadowToMem(s));
322        CHECK(IsMetaMem(m));
323        if (prev) {
324          // Ensure that shadow and meta mappings are linear within a single
325          // user range. Lots of code that processes memory ranges assumes it.
326          const uptr prev_s = MemToShadow(prev);
327          const uptr prev_m = (uptr)MemToMeta(prev);
328          CHECK_EQ(s - prev_s, (p - prev) * kShadowMultiplier);
329          CHECK_EQ((m - prev_m) / kMetaShadowSize,
330                   (p - prev) / kMetaShadowCell);
331        }
332        prev = p;
333      }
334    }
335  }
336}
337
338#if !SANITIZER_GO
339static void OnStackUnwind(const SignalContext &sig, const void *,
340                          BufferedStackTrace *stack) {
341  stack->Unwind(StackTrace::GetNextInstructionPc(sig.pc), sig.bp, sig.context,
342                common_flags()->fast_unwind_on_fatal);
343}
344
345static void TsanOnDeadlySignal(int signo, void *siginfo, void *context) {
346  HandleDeadlySignal(siginfo, context, GetTid(), &OnStackUnwind, nullptr);
347}
348#endif
349
350void Initialize(ThreadState *thr) {
351  // Thread safe because done before all threads exist.
352  static bool is_initialized = false;
353  if (is_initialized)
354    return;
355  is_initialized = true;
356  // We are not ready to handle interceptors yet.
357  ScopedIgnoreInterceptors ignore;
358  SanitizerToolName = "ThreadSanitizer";
359  // Install tool-specific callbacks in sanitizer_common.
360  SetCheckFailedCallback(TsanCheckFailed);
361
362  ctx = new(ctx_placeholder) Context;
363  const char *env_name = SANITIZER_GO ? "GORACE" : "TSAN_OPTIONS";
364  const char *options = GetEnv(env_name);
365  CacheBinaryName();
366  CheckASLR();
367  InitializeFlags(&ctx->flags, options, env_name);
368  AvoidCVE_2016_2143();
369  __sanitizer::InitializePlatformEarly();
370  __tsan::InitializePlatformEarly();
371
372#if !SANITIZER_GO
373  // Re-exec ourselves if we need to set additional env or command line args.
374  MaybeReexec();
375
376  InitializeAllocator();
377  ReplaceSystemMalloc();
378#endif
379  if (common_flags()->detect_deadlocks)
380    ctx->dd = DDetector::Create(flags());
381  Processor *proc = ProcCreate();
382  ProcWire(proc, thr);
383  InitializeInterceptors();
384  CheckShadowMapping();
385  InitializePlatform();
386  InitializeMutex();
387  InitializeDynamicAnnotations();
388#if !SANITIZER_GO
389  InitializeShadowMemory();
390  InitializeAllocatorLate();
391  InstallDeadlySignalHandlers(TsanOnDeadlySignal);
392#endif
393  // Setup correct file descriptor for error reports.
394  __sanitizer_set_report_path(common_flags()->log_path);
395  InitializeSuppressions();
396#if !SANITIZER_GO
397  InitializeLibIgnore();
398  Symbolizer::GetOrInit()->AddHooks(EnterSymbolizer, ExitSymbolizer);
399#endif
400
401  VPrintf(1, "***** Running under ThreadSanitizer v2 (pid %d) *****\n",
402          (int)internal_getpid());
403
404  // Initialize thread 0.
405  int tid = ThreadCreate(thr, 0, 0, true);
406  CHECK_EQ(tid, 0);
407  ThreadStart(thr, tid, GetTid(), ThreadType::Regular);
408#if TSAN_CONTAINS_UBSAN
409  __ubsan::InitAsPlugin();
410#endif
411  ctx->initialized = true;
412
413#if !SANITIZER_GO
414  Symbolizer::LateInitialize();
415#endif
416
417  if (flags()->stop_on_start) {
418    Printf("ThreadSanitizer is suspended at startup (pid %d)."
419           " Call __tsan_resume().\n",
420           (int)internal_getpid());
421    while (__tsan_resumed == 0) {}
422  }
423
424  OnInitialize();
425}
426
427void MaybeSpawnBackgroundThread() {
428  // On MIPS, TSan initialization is run before
429  // __pthread_initialize_minimal_internal() is finished, so we can not spawn
430  // new threads.
431#if !SANITIZER_GO && !defined(__mips__)
432  static atomic_uint32_t bg_thread = {};
433  if (atomic_load(&bg_thread, memory_order_relaxed) == 0 &&
434      atomic_exchange(&bg_thread, 1, memory_order_relaxed) == 0) {
435    StartBackgroundThread();
436    SetSandboxingCallback(StopBackgroundThread);
437  }
438#endif
439}
440
441
442int Finalize(ThreadState *thr) {
443  bool failed = false;
444
445  if (common_flags()->print_module_map == 1) PrintModuleMap();
446
447  if (flags()->atexit_sleep_ms > 0 && ThreadCount(thr) > 1)
448    SleepForMillis(flags()->atexit_sleep_ms);
449
450  // Wait for pending reports.
451  ctx->report_mtx.Lock();
452  { ScopedErrorReportLock l; }
453  ctx->report_mtx.Unlock();
454
455#if !SANITIZER_GO
456  if (Verbosity()) AllocatorPrintStats();
457#endif
458
459  ThreadFinalize(thr);
460
461  if (ctx->nreported) {
462    failed = true;
463#if !SANITIZER_GO
464    Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
465#else
466    Printf("Found %d data race(s)\n", ctx->nreported);
467#endif
468  }
469
470  if (ctx->nmissed_expected) {
471    failed = true;
472    Printf("ThreadSanitizer: missed %d expected races\n",
473        ctx->nmissed_expected);
474  }
475
476  if (common_flags()->print_suppressions)
477    PrintMatchedSuppressions();
478#if !SANITIZER_GO
479  if (flags()->print_benign)
480    PrintMatchedBenignRaces();
481#endif
482
483  failed = OnFinalize(failed);
484
485#if TSAN_COLLECT_STATS
486  StatAggregate(ctx->stat, thr->stat);
487  StatOutput(ctx->stat);
488#endif
489
490  return failed ? common_flags()->exitcode : 0;
491}
492
493#if !SANITIZER_GO
494void ForkBefore(ThreadState *thr, uptr pc) {
495  ctx->thread_registry->Lock();
496  ctx->report_mtx.Lock();
497}
498
499void ForkParentAfter(ThreadState *thr, uptr pc) {
500  ctx->report_mtx.Unlock();
501  ctx->thread_registry->Unlock();
502}
503
504void ForkChildAfter(ThreadState *thr, uptr pc) {
505  ctx->report_mtx.Unlock();
506  ctx->thread_registry->Unlock();
507
508  uptr nthread = 0;
509  ctx->thread_registry->GetNumberOfThreads(0, 0, &nthread /* alive threads */);
510  VPrintf(1, "ThreadSanitizer: forked new process with pid %d,"
511      " parent had %d threads\n", (int)internal_getpid(), (int)nthread);
512  if (nthread == 1) {
513    StartBackgroundThread();
514  } else {
515    // We've just forked a multi-threaded process. We cannot reasonably function
516    // after that (some mutexes may be locked before fork). So just enable
517    // ignores for everything in the hope that we will exec soon.
518    ctx->after_multithreaded_fork = true;
519    thr->ignore_interceptors++;
520    ThreadIgnoreBegin(thr, pc);
521    ThreadIgnoreSyncBegin(thr, pc);
522  }
523}
524#endif
525
526#if SANITIZER_GO
527NOINLINE
528void GrowShadowStack(ThreadState *thr) {
529  const int sz = thr->shadow_stack_end - thr->shadow_stack;
530  const int newsz = 2 * sz;
531  uptr *newstack = (uptr*)internal_alloc(MBlockShadowStack,
532      newsz * sizeof(uptr));
533  internal_memcpy(newstack, thr->shadow_stack, sz * sizeof(uptr));
534  internal_free(thr->shadow_stack);
535  thr->shadow_stack = newstack;
536  thr->shadow_stack_pos = newstack + sz;
537  thr->shadow_stack_end = newstack + newsz;
538}
539#endif
540
541u32 CurrentStackId(ThreadState *thr, uptr pc) {
542  if (!thr->is_inited)  // May happen during bootstrap.
543    return 0;
544  if (pc != 0) {
545#if !SANITIZER_GO
546    DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
547#else
548    if (thr->shadow_stack_pos == thr->shadow_stack_end)
549      GrowShadowStack(thr);
550#endif
551    thr->shadow_stack_pos[0] = pc;
552    thr->shadow_stack_pos++;
553  }
554  u32 id = StackDepotPut(
555      StackTrace(thr->shadow_stack, thr->shadow_stack_pos - thr->shadow_stack));
556  if (pc != 0)
557    thr->shadow_stack_pos--;
558  return id;
559}
560
561void TraceSwitch(ThreadState *thr) {
562#if !SANITIZER_GO
563  if (ctx->after_multithreaded_fork)
564    return;
565#endif
566  thr->nomalloc++;
567  Trace *thr_trace = ThreadTrace(thr->tid);
568  Lock l(&thr_trace->mtx);
569  unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % TraceParts();
570  TraceHeader *hdr = &thr_trace->headers[trace];
571  hdr->epoch0 = thr->fast_state.epoch();
572  ObtainCurrentStack(thr, 0, &hdr->stack0);
573  hdr->mset0 = thr->mset;
574  thr->nomalloc--;
575}
576
577Trace *ThreadTrace(int tid) {
578  return (Trace*)GetThreadTraceHeader(tid);
579}
580
581uptr TraceTopPC(ThreadState *thr) {
582  Event *events = (Event*)GetThreadTrace(thr->tid);
583  uptr pc = events[thr->fast_state.GetTracePos()];
584  return pc;
585}
586
587uptr TraceSize() {
588  return (uptr)(1ull << (kTracePartSizeBits + flags()->history_size + 1));
589}
590
591uptr TraceParts() {
592  return TraceSize() / kTracePartSize;
593}
594
595#if !SANITIZER_GO
596extern "C" void __tsan_trace_switch() {
597  TraceSwitch(cur_thread());
598}
599
600extern "C" void __tsan_report_race() {
601  ReportRace(cur_thread());
602}
603#endif
604
605ALWAYS_INLINE
606Shadow LoadShadow(u64 *p) {
607  u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
608  return Shadow(raw);
609}
610
611ALWAYS_INLINE
612void StoreShadow(u64 *sp, u64 s) {
613  atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
614}
615
616ALWAYS_INLINE
617void StoreIfNotYetStored(u64 *sp, u64 *s) {
618  StoreShadow(sp, *s);
619  *s = 0;
620}
621
622ALWAYS_INLINE
623void HandleRace(ThreadState *thr, u64 *shadow_mem,
624                              Shadow cur, Shadow old) {
625  thr->racy_state[0] = cur.raw();
626  thr->racy_state[1] = old.raw();
627  thr->racy_shadow_addr = shadow_mem;
628#if !SANITIZER_GO
629  HACKY_CALL(__tsan_report_race);
630#else
631  ReportRace(thr);
632#endif
633}
634
635static inline bool HappensBefore(Shadow old, ThreadState *thr) {
636  return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
637}
638
639ALWAYS_INLINE
640void MemoryAccessImpl1(ThreadState *thr, uptr addr,
641    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
642    u64 *shadow_mem, Shadow cur) {
643  StatInc(thr, StatMop);
644  StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
645  StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
646
647  // This potentially can live in an MMX/SSE scratch register.
648  // The required intrinsics are:
649  // __m128i _mm_move_epi64(__m128i*);
650  // _mm_storel_epi64(u64*, __m128i);
651  u64 store_word = cur.raw();
652  bool stored = false;
653
654  // scan all the shadow values and dispatch to 4 categories:
655  // same, replace, candidate and race (see comments below).
656  // we consider only 3 cases regarding access sizes:
657  // equal, intersect and not intersect. initially I considered
658  // larger and smaller as well, it allowed to replace some
659  // 'candidates' with 'same' or 'replace', but I think
660  // it's just not worth it (performance- and complexity-wise).
661
662  Shadow old(0);
663
664  // It release mode we manually unroll the loop,
665  // because empirically gcc generates better code this way.
666  // However, we can't afford unrolling in debug mode, because the function
667  // consumes almost 4K of stack. Gtest gives only 4K of stack to death test
668  // threads, which is not enough for the unrolled loop.
669#if SANITIZER_DEBUG
670  for (int idx = 0; idx < 4; idx++) {
671#include "tsan_update_shadow_word_inl.h"
672  }
673#else
674  int idx = 0;
675#include "tsan_update_shadow_word_inl.h"
676  idx = 1;
677  if (stored) {
678#include "tsan_update_shadow_word_inl.h"
679  } else {
680#include "tsan_update_shadow_word_inl.h"
681  }
682  idx = 2;
683  if (stored) {
684#include "tsan_update_shadow_word_inl.h"
685  } else {
686#include "tsan_update_shadow_word_inl.h"
687  }
688  idx = 3;
689  if (stored) {
690#include "tsan_update_shadow_word_inl.h"
691  } else {
692#include "tsan_update_shadow_word_inl.h"
693  }
694#endif
695
696  // we did not find any races and had already stored
697  // the current access info, so we are done
698  if (LIKELY(stored))
699    return;
700  // choose a random candidate slot and replace it
701  StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
702  StatInc(thr, StatShadowReplace);
703  return;
704 RACE:
705  HandleRace(thr, shadow_mem, cur, old);
706  return;
707}
708
709void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
710    int size, bool kAccessIsWrite, bool kIsAtomic) {
711  while (size) {
712    int size1 = 1;
713    int kAccessSizeLog = kSizeLog1;
714    if (size >= 8 && (addr & ~7) == ((addr + 7) & ~7)) {
715      size1 = 8;
716      kAccessSizeLog = kSizeLog8;
717    } else if (size >= 4 && (addr & ~7) == ((addr + 3) & ~7)) {
718      size1 = 4;
719      kAccessSizeLog = kSizeLog4;
720    } else if (size >= 2 && (addr & ~7) == ((addr + 1) & ~7)) {
721      size1 = 2;
722      kAccessSizeLog = kSizeLog2;
723    }
724    MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
725    addr += size1;
726    size -= size1;
727  }
728}
729
730ALWAYS_INLINE
731bool ContainsSameAccessSlow(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
732  Shadow cur(a);
733  for (uptr i = 0; i < kShadowCnt; i++) {
734    Shadow old(LoadShadow(&s[i]));
735    if (Shadow::Addr0AndSizeAreEqual(cur, old) &&
736        old.TidWithIgnore() == cur.TidWithIgnore() &&
737        old.epoch() > sync_epoch &&
738        old.IsAtomic() == cur.IsAtomic() &&
739        old.IsRead() <= cur.IsRead())
740      return true;
741  }
742  return false;
743}
744
745#if defined(__SSE3__)
746#define SHUF(v0, v1, i0, i1, i2, i3) _mm_castps_si128(_mm_shuffle_ps( \
747    _mm_castsi128_ps(v0), _mm_castsi128_ps(v1), \
748    (i0)*1 + (i1)*4 + (i2)*16 + (i3)*64))
749ALWAYS_INLINE
750bool ContainsSameAccessFast(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
751  // This is an optimized version of ContainsSameAccessSlow.
752  // load current access into access[0:63]
753  const m128 access     = _mm_cvtsi64_si128(a);
754  // duplicate high part of access in addr0:
755  // addr0[0:31]        = access[32:63]
756  // addr0[32:63]       = access[32:63]
757  // addr0[64:95]       = access[32:63]
758  // addr0[96:127]      = access[32:63]
759  const m128 addr0      = SHUF(access, access, 1, 1, 1, 1);
760  // load 4 shadow slots
761  const m128 shadow0    = _mm_load_si128((__m128i*)s);
762  const m128 shadow1    = _mm_load_si128((__m128i*)s + 1);
763  // load high parts of 4 shadow slots into addr_vect:
764  // addr_vect[0:31]    = shadow0[32:63]
765  // addr_vect[32:63]   = shadow0[96:127]
766  // addr_vect[64:95]   = shadow1[32:63]
767  // addr_vect[96:127]  = shadow1[96:127]
768  m128 addr_vect        = SHUF(shadow0, shadow1, 1, 3, 1, 3);
769  if (!is_write) {
770    // set IsRead bit in addr_vect
771    const m128 rw_mask1 = _mm_cvtsi64_si128(1<<15);
772    const m128 rw_mask  = SHUF(rw_mask1, rw_mask1, 0, 0, 0, 0);
773    addr_vect           = _mm_or_si128(addr_vect, rw_mask);
774  }
775  // addr0 == addr_vect?
776  const m128 addr_res   = _mm_cmpeq_epi32(addr0, addr_vect);
777  // epoch1[0:63]       = sync_epoch
778  const m128 epoch1     = _mm_cvtsi64_si128(sync_epoch);
779  // epoch[0:31]        = sync_epoch[0:31]
780  // epoch[32:63]       = sync_epoch[0:31]
781  // epoch[64:95]       = sync_epoch[0:31]
782  // epoch[96:127]      = sync_epoch[0:31]
783  const m128 epoch      = SHUF(epoch1, epoch1, 0, 0, 0, 0);
784  // load low parts of shadow cell epochs into epoch_vect:
785  // epoch_vect[0:31]   = shadow0[0:31]
786  // epoch_vect[32:63]  = shadow0[64:95]
787  // epoch_vect[64:95]  = shadow1[0:31]
788  // epoch_vect[96:127] = shadow1[64:95]
789  const m128 epoch_vect = SHUF(shadow0, shadow1, 0, 2, 0, 2);
790  // epoch_vect >= sync_epoch?
791  const m128 epoch_res  = _mm_cmpgt_epi32(epoch_vect, epoch);
792  // addr_res & epoch_res
793  const m128 res        = _mm_and_si128(addr_res, epoch_res);
794  // mask[0] = res[7]
795  // mask[1] = res[15]
796  // ...
797  // mask[15] = res[127]
798  const int mask        = _mm_movemask_epi8(res);
799  return mask != 0;
800}
801#endif
802
803ALWAYS_INLINE
804bool ContainsSameAccess(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
805#if defined(__SSE3__)
806  bool res = ContainsSameAccessFast(s, a, sync_epoch, is_write);
807  // NOTE: this check can fail if the shadow is concurrently mutated
808  // by other threads. But it still can be useful if you modify
809  // ContainsSameAccessFast and want to ensure that it's not completely broken.
810  // DCHECK_EQ(res, ContainsSameAccessSlow(s, a, sync_epoch, is_write));
811  return res;
812#else
813  return ContainsSameAccessSlow(s, a, sync_epoch, is_write);
814#endif
815}
816
817ALWAYS_INLINE USED
818void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
819    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic) {
820  u64 *shadow_mem = (u64*)MemToShadow(addr);
821  DPrintf2("#%d: MemoryAccess: @%p %p size=%d"
822      " is_write=%d shadow_mem=%p {%zx, %zx, %zx, %zx}\n",
823      (int)thr->fast_state.tid(), (void*)pc, (void*)addr,
824      (int)(1 << kAccessSizeLog), kAccessIsWrite, shadow_mem,
825      (uptr)shadow_mem[0], (uptr)shadow_mem[1],
826      (uptr)shadow_mem[2], (uptr)shadow_mem[3]);
827#if SANITIZER_DEBUG
828  if (!IsAppMem(addr)) {
829    Printf("Access to non app mem %zx\n", addr);
830    DCHECK(IsAppMem(addr));
831  }
832  if (!IsShadowMem((uptr)shadow_mem)) {
833    Printf("Bad shadow addr %p (%zx)\n", shadow_mem, addr);
834    DCHECK(IsShadowMem((uptr)shadow_mem));
835  }
836#endif
837
838  if (!SANITIZER_GO && !kAccessIsWrite && *shadow_mem == kShadowRodata) {
839    // Access to .rodata section, no races here.
840    // Measurements show that it can be 10-20% of all memory accesses.
841    StatInc(thr, StatMop);
842    StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
843    StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
844    StatInc(thr, StatMopRodata);
845    return;
846  }
847
848  FastState fast_state = thr->fast_state;
849  if (UNLIKELY(fast_state.GetIgnoreBit())) {
850    StatInc(thr, StatMop);
851    StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
852    StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
853    StatInc(thr, StatMopIgnored);
854    return;
855  }
856
857  Shadow cur(fast_state);
858  cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
859  cur.SetWrite(kAccessIsWrite);
860  cur.SetAtomic(kIsAtomic);
861
862  if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
863      thr->fast_synch_epoch, kAccessIsWrite))) {
864    StatInc(thr, StatMop);
865    StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
866    StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
867    StatInc(thr, StatMopSame);
868    return;
869  }
870
871  if (kCollectHistory) {
872    fast_state.IncrementEpoch();
873    thr->fast_state = fast_state;
874    TraceAddEvent(thr, fast_state, EventTypeMop, pc);
875    cur.IncrementEpoch();
876  }
877
878  MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
879      shadow_mem, cur);
880}
881
882// Called by MemoryAccessRange in tsan_rtl_thread.cpp
883ALWAYS_INLINE USED
884void MemoryAccessImpl(ThreadState *thr, uptr addr,
885    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
886    u64 *shadow_mem, Shadow cur) {
887  if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
888      thr->fast_synch_epoch, kAccessIsWrite))) {
889    StatInc(thr, StatMop);
890    StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
891    StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
892    StatInc(thr, StatMopSame);
893    return;
894  }
895
896  MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
897      shadow_mem, cur);
898}
899
900static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
901                           u64 val) {
902  (void)thr;
903  (void)pc;
904  if (size == 0)
905    return;
906  // FIXME: fix me.
907  uptr offset = addr % kShadowCell;
908  if (offset) {
909    offset = kShadowCell - offset;
910    if (size <= offset)
911      return;
912    addr += offset;
913    size -= offset;
914  }
915  DCHECK_EQ(addr % 8, 0);
916  // If a user passes some insane arguments (memset(0)),
917  // let it just crash as usual.
918  if (!IsAppMem(addr) || !IsAppMem(addr + size - 1))
919    return;
920  // Don't want to touch lots of shadow memory.
921  // If a program maps 10MB stack, there is no need reset the whole range.
922  size = (size + (kShadowCell - 1)) & ~(kShadowCell - 1);
923  // UnmapOrDie/MmapFixedNoReserve does not work on Windows.
924  if (SANITIZER_WINDOWS || size < common_flags()->clear_shadow_mmap_threshold) {
925    u64 *p = (u64*)MemToShadow(addr);
926    CHECK(IsShadowMem((uptr)p));
927    CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
928    // FIXME: may overwrite a part outside the region
929    for (uptr i = 0; i < size / kShadowCell * kShadowCnt;) {
930      p[i++] = val;
931      for (uptr j = 1; j < kShadowCnt; j++)
932        p[i++] = 0;
933    }
934  } else {
935    // The region is big, reset only beginning and end.
936    const uptr kPageSize = GetPageSizeCached();
937    u64 *begin = (u64*)MemToShadow(addr);
938    u64 *end = begin + size / kShadowCell * kShadowCnt;
939    u64 *p = begin;
940    // Set at least first kPageSize/2 to page boundary.
941    while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
942      *p++ = val;
943      for (uptr j = 1; j < kShadowCnt; j++)
944        *p++ = 0;
945    }
946    // Reset middle part.
947    u64 *p1 = p;
948    p = RoundDown(end, kPageSize);
949    UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
950    if (!MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1))
951      Die();
952    // Set the ending.
953    while (p < end) {
954      *p++ = val;
955      for (uptr j = 1; j < kShadowCnt; j++)
956        *p++ = 0;
957    }
958  }
959}
960
961void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
962  MemoryRangeSet(thr, pc, addr, size, 0);
963}
964
965void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
966  // Processing more than 1k (4k of shadow) is expensive,
967  // can cause excessive memory consumption (user does not necessary touch
968  // the whole range) and most likely unnecessary.
969  if (size > 1024)
970    size = 1024;
971  CHECK_EQ(thr->is_freeing, false);
972  thr->is_freeing = true;
973  MemoryAccessRange(thr, pc, addr, size, true);
974  thr->is_freeing = false;
975  if (kCollectHistory) {
976    thr->fast_state.IncrementEpoch();
977    TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
978  }
979  Shadow s(thr->fast_state);
980  s.ClearIgnoreBit();
981  s.MarkAsFreed();
982  s.SetWrite(true);
983  s.SetAddr0AndSizeLog(0, 3);
984  MemoryRangeSet(thr, pc, addr, size, s.raw());
985}
986
987void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
988  if (kCollectHistory) {
989    thr->fast_state.IncrementEpoch();
990    TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
991  }
992  Shadow s(thr->fast_state);
993  s.ClearIgnoreBit();
994  s.SetWrite(true);
995  s.SetAddr0AndSizeLog(0, 3);
996  MemoryRangeSet(thr, pc, addr, size, s.raw());
997}
998
999void MemoryRangeImitateWriteOrResetRange(ThreadState *thr, uptr pc, uptr addr,
1000                                         uptr size) {
1001  if (thr->ignore_reads_and_writes == 0)
1002    MemoryRangeImitateWrite(thr, pc, addr, size);
1003  else
1004    MemoryResetRange(thr, pc, addr, size);
1005}
1006
1007ALWAYS_INLINE USED
1008void FuncEntry(ThreadState *thr, uptr pc) {
1009  StatInc(thr, StatFuncEnter);
1010  DPrintf2("#%d: FuncEntry %p\n", (int)thr->fast_state.tid(), (void*)pc);
1011  if (kCollectHistory) {
1012    thr->fast_state.IncrementEpoch();
1013    TraceAddEvent(thr, thr->fast_state, EventTypeFuncEnter, pc);
1014  }
1015
1016  // Shadow stack maintenance can be replaced with
1017  // stack unwinding during trace switch (which presumably must be faster).
1018  DCHECK_GE(thr->shadow_stack_pos, thr->shadow_stack);
1019#if !SANITIZER_GO
1020  DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
1021#else
1022  if (thr->shadow_stack_pos == thr->shadow_stack_end)
1023    GrowShadowStack(thr);
1024#endif
1025  thr->shadow_stack_pos[0] = pc;
1026  thr->shadow_stack_pos++;
1027}
1028
1029ALWAYS_INLINE USED
1030void FuncExit(ThreadState *thr) {
1031  StatInc(thr, StatFuncExit);
1032  DPrintf2("#%d: FuncExit\n", (int)thr->fast_state.tid());
1033  if (kCollectHistory) {
1034    thr->fast_state.IncrementEpoch();
1035    TraceAddEvent(thr, thr->fast_state, EventTypeFuncExit, 0);
1036  }
1037
1038  DCHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
1039#if !SANITIZER_GO
1040  DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
1041#endif
1042  thr->shadow_stack_pos--;
1043}
1044
1045void ThreadIgnoreBegin(ThreadState *thr, uptr pc, bool save_stack) {
1046  DPrintf("#%d: ThreadIgnoreBegin\n", thr->tid);
1047  thr->ignore_reads_and_writes++;
1048  CHECK_GT(thr->ignore_reads_and_writes, 0);
1049  thr->fast_state.SetIgnoreBit();
1050#if !SANITIZER_GO
1051  if (save_stack && !ctx->after_multithreaded_fork)
1052    thr->mop_ignore_set.Add(CurrentStackId(thr, pc));
1053#endif
1054}
1055
1056void ThreadIgnoreEnd(ThreadState *thr, uptr pc) {
1057  DPrintf("#%d: ThreadIgnoreEnd\n", thr->tid);
1058  CHECK_GT(thr->ignore_reads_and_writes, 0);
1059  thr->ignore_reads_and_writes--;
1060  if (thr->ignore_reads_and_writes == 0) {
1061    thr->fast_state.ClearIgnoreBit();
1062#if !SANITIZER_GO
1063    thr->mop_ignore_set.Reset();
1064#endif
1065  }
1066}
1067
1068#if !SANITIZER_GO
1069extern "C" SANITIZER_INTERFACE_ATTRIBUTE
1070uptr __tsan_testonly_shadow_stack_current_size() {
1071  ThreadState *thr = cur_thread();
1072  return thr->shadow_stack_pos - thr->shadow_stack;
1073}
1074#endif
1075
1076void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack) {
1077  DPrintf("#%d: ThreadIgnoreSyncBegin\n", thr->tid);
1078  thr->ignore_sync++;
1079  CHECK_GT(thr->ignore_sync, 0);
1080#if !SANITIZER_GO
1081  if (save_stack && !ctx->after_multithreaded_fork)
1082    thr->sync_ignore_set.Add(CurrentStackId(thr, pc));
1083#endif
1084}
1085
1086void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc) {
1087  DPrintf("#%d: ThreadIgnoreSyncEnd\n", thr->tid);
1088  CHECK_GT(thr->ignore_sync, 0);
1089  thr->ignore_sync--;
1090#if !SANITIZER_GO
1091  if (thr->ignore_sync == 0)
1092    thr->sync_ignore_set.Reset();
1093#endif
1094}
1095
1096bool MD5Hash::operator==(const MD5Hash &other) const {
1097  return hash[0] == other.hash[0] && hash[1] == other.hash[1];
1098}
1099
1100#if SANITIZER_DEBUG
1101void build_consistency_debug() {}
1102#else
1103void build_consistency_release() {}
1104#endif
1105
1106#if TSAN_COLLECT_STATS
1107void build_consistency_stats() {}
1108#else
1109void build_consistency_nostats() {}
1110#endif
1111
1112}  // namespace __tsan
1113
1114#if !SANITIZER_GO
1115// Must be included in this file to make sure everything is inlined.
1116#include "tsan_interface_inl.h"
1117#endif
1118