1 //===-- tsan_rtl.cc -------------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a part of ThreadSanitizer (TSan), a race detector.
12 // Main file (entry points) for the TSan run-time.
13 //===----------------------------------------------------------------------===//
15 #include "sanitizer_common/sanitizer_atomic.h"
16 #include "sanitizer_common/sanitizer_common.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"
24 #include "tsan_mman.h"
25 #include "tsan_suppressions.h"
26 #include "tsan_symbolize.h"
27 #include "ubsan/ubsan_init.h"
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
35 #include <emmintrin.h>
39 volatile int __tsan_resumed = 0;
41 extern "C" void __tsan_resume() {
47 #if !SANITIZER_GO && !SANITIZER_MAC
48 __attribute__((tls_model("initial-exec")))
49 THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
51 static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
54 // Can be overriden by a front-end.
55 #ifdef TSAN_EXTERNAL_HOOKS
56 bool OnFinalize(bool failed);
59 SANITIZER_WEAK_CXX_DEFAULT_IMPL
60 bool OnFinalize(bool failed) {
63 SANITIZER_WEAK_CXX_DEFAULT_IMPL
64 void OnInitialize() {}
67 static char thread_registry_placeholder[sizeof(ThreadRegistry)];
69 static ThreadContextBase *CreateThreadContext(u32 tid) {
70 // Map thread trace when context is created.
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);
91 static const u32 kThreadQuarantineSize = 16;
93 static const u32 kThreadQuarantineSize = 64;
98 , report_mtx(MutexTypeReport, StatMtxReport)
101 , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
102 CreateThreadContext, kMaxTid, kThreadQuarantineSize, kMaxTidReuse))
103 , racy_mtx(MutexTypeRacy, StatMtxRacy)
104 , racy_stacks(MBlockRacyStacks)
105 , racy_addresses(MBlockRacyAddresses)
106 , fired_suppressions_mtx(MutexTypeFired, StatMtxFired)
107 , fired_suppressions(8) {
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)
122 , jmp_bufs(MBlockJmpBuf)
125 , unique_id(unique_id)
131 , last_sleep_clock(tid)
137 static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
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()));
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
151 cur_thread()->ignore_interceptors++;
152 const u64 kMs2Ns = 1000 * 1000;
154 fd_t mprof_fd = kInvalidFd;
155 if (flags()->profile_memory && flags()->profile_memory[0]) {
156 if (internal_strcmp(flags()->profile_memory, "stdout") == 0) {
158 } else if (internal_strcmp(flags()->profile_memory, "stderr") == 0) {
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",
173 u64 last_flush = NanoTime();
176 atomic_load(&ctx->stop_background_thread, memory_order_relaxed) == 0;
179 u64 now = NanoTime();
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");
186 last_flush = NanoTime();
189 // GetRSS can be expensive on huge programs, so don't do it every 100ms.
190 if (flags()->memory_limit_mb > 0) {
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");
200 VPrintf(1, "ThreadSanitizer: memory flushed RSS=%llu\n", (u64)rss>>20);
205 // Write memory profile if requested.
206 if (mprof_fd != kInvalidFd)
207 MemoryProfiler(ctx, mprof_fd, i);
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);
217 atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
223 static void StartBackgroundThread() {
224 ctx->background_thread = internal_start_thread(&BackgroundThread, 0);
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;
236 void DontNeedShadowFor(uptr addr, uptr size) {
237 ReleaseMemoryPagesToOS(MemToShadow(addr), MemToShadow(addr + size));
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");
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);
257 // First call maps data+bss.
259 MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow");
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)
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;
272 VPrintf(2, "mapped meta shadow for (%p-%p) at (%p-%p)\n",
273 addr, addr+size, meta_begin, meta_end);
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);
283 Printf("FATAL: ThreadSanitizer can not mmap thread trace (%p/%p->%p)\n",
289 static void CheckShadowMapping() {
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).
296 VPrintf(3, "checking shadow region %p-%p\n", beg, end);
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)
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);
307 CHECK(IsShadowMem(s));
308 CHECK_EQ(p, ShadowToMem(s));
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);
325 void Initialize(ThreadState *thr) {
326 // Thread safe because done before all threads exist.
327 static bool is_initialized = false;
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);
337 ctx = new(ctx_placeholder) Context;
338 const char *options = GetEnv(SANITIZER_GO ? "GORACE" : "TSAN_OPTIONS");
340 InitializeFlags(&ctx->flags, options);
341 AvoidCVE_2016_2143();
342 InitializePlatformEarly();
344 // Re-exec ourselves if we need to set additional env or command line args.
347 InitializeAllocator();
348 ReplaceSystemMalloc();
350 if (common_flags()->detect_deadlocks)
351 ctx->dd = DDetector::Create(flags());
352 Processor *proc = ProcCreate();
354 InitializeInterceptors();
355 CheckShadowMapping();
356 InitializePlatform();
358 InitializeDynamicAnnotations();
360 InitializeShadowMemory();
361 InitializeAllocatorLate();
363 // Setup correct file descriptor for error reports.
364 __sanitizer_set_report_path(common_flags()->log_path);
365 InitializeSuppressions();
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
373 StartBackgroundThread();
374 SetSandboxingCallback(StopBackgroundThread);
378 VPrintf(1, "***** Running under ThreadSanitizer v2 (pid %d) *****\n",
379 (int)internal_getpid());
381 // Initialize thread 0.
382 int tid = ThreadCreate(thr, 0, 0, true);
384 ThreadStart(thr, tid, GetTid(), /*workerthread*/ false);
385 #if TSAN_CONTAINS_UBSAN
386 __ubsan::InitAsPlugin();
388 ctx->initialized = true;
391 Symbolizer::LateInitialize();
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) {}
404 int Finalize(ThreadState *thr) {
407 if (common_flags()->print_module_map == 1) PrintModuleMap();
409 if (flags()->atexit_sleep_ms > 0 && ThreadCount(thr) > 1)
410 SleepForMillis(flags()->atexit_sleep_ms);
412 // Wait for pending reports.
413 ctx->report_mtx.Lock();
414 CommonSanitizerReportMutex.Lock();
415 CommonSanitizerReportMutex.Unlock();
416 ctx->report_mtx.Unlock();
419 if (Verbosity()) AllocatorPrintStats();
424 if (ctx->nreported) {
427 Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
429 Printf("Found %d data race(s)\n", ctx->nreported);
433 if (ctx->nmissed_expected) {
435 Printf("ThreadSanitizer: missed %d expected races\n",
436 ctx->nmissed_expected);
439 if (common_flags()->print_suppressions)
440 PrintMatchedSuppressions();
442 if (flags()->print_benign)
443 PrintMatchedBenignRaces();
446 failed = OnFinalize(failed);
448 #if TSAN_COLLECT_STATS
449 StatAggregate(ctx->stat, thr->stat);
450 StatOutput(ctx->stat);
453 return failed ? common_flags()->exitcode : 0;
457 void ForkBefore(ThreadState *thr, uptr pc) {
458 ctx->thread_registry->Lock();
459 ctx->report_mtx.Lock();
462 void ForkParentAfter(ThreadState *thr, uptr pc) {
463 ctx->report_mtx.Unlock();
464 ctx->thread_registry->Unlock();
467 void ForkChildAfter(ThreadState *thr, uptr pc) {
468 ctx->report_mtx.Unlock();
469 ctx->thread_registry->Unlock();
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);
476 StartBackgroundThread();
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);
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;
504 u32 CurrentStackId(ThreadState *thr, uptr pc) {
505 if (!thr->is_inited) // May happen during bootstrap.
509 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
511 if (thr->shadow_stack_pos == thr->shadow_stack_end)
512 GrowShadowStack(thr);
514 thr->shadow_stack_pos[0] = pc;
515 thr->shadow_stack_pos++;
517 u32 id = StackDepotPut(
518 StackTrace(thr->shadow_stack, thr->shadow_stack_pos - thr->shadow_stack));
520 thr->shadow_stack_pos--;
524 void TraceSwitch(ThreadState *thr) {
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;
536 Trace *ThreadTrace(int tid) {
537 return (Trace*)GetThreadTraceHeader(tid);
540 uptr TraceTopPC(ThreadState *thr) {
541 Event *events = (Event*)GetThreadTrace(thr->tid);
542 uptr pc = events[thr->fast_state.GetTracePos()];
547 return (uptr)(1ull << (kTracePartSizeBits + flags()->history_size + 1));
551 return TraceSize() / kTracePartSize;
555 extern "C" void __tsan_trace_switch() {
556 TraceSwitch(cur_thread());
559 extern "C" void __tsan_report_race() {
560 ReportRace(cur_thread());
565 Shadow LoadShadow(u64 *p) {
566 u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
571 void StoreShadow(u64 *sp, u64 s) {
572 atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
576 void StoreIfNotYetStored(u64 *sp, u64 *s) {
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;
588 HACKY_CALL(__tsan_report_race);
594 static inline bool HappensBefore(Shadow old, ThreadState *thr) {
595 return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
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));
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();
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).
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.
628 for (int idx = 0; idx < 4; idx++) {
629 #include "tsan_update_shadow_word_inl.h"
633 #include "tsan_update_shadow_word_inl.h"
635 #include "tsan_update_shadow_word_inl.h"
637 #include "tsan_update_shadow_word_inl.h"
639 #include "tsan_update_shadow_word_inl.h"
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))
646 // choose a random candidate slot and replace it
647 StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
648 StatInc(thr, StatShadowReplace);
651 HandleRace(thr, shadow_mem, cur, old);
655 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
656 int size, bool kAccessIsWrite, bool kIsAtomic) {
659 int kAccessSizeLog = kSizeLog1;
660 if (size >= 8 && (addr & ~7) == ((addr + 7) & ~7)) {
662 kAccessSizeLog = kSizeLog8;
663 } else if (size >= 4 && (addr & ~7) == ((addr + 3) & ~7)) {
665 kAccessSizeLog = kSizeLog4;
666 } else if (size >= 2 && (addr & ~7) == ((addr + 1) & ~7)) {
668 kAccessSizeLog = kSizeLog2;
670 MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
677 bool ContainsSameAccessSlow(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
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())
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))
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);
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);
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);
743 // mask[15] = res[127]
744 const int mask = _mm_movemask_epi8(res);
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));
759 return ContainsSameAccessSlow(s, a, sync_epoch, is_write);
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]);
774 if (!IsAppMem(addr)) {
775 Printf("Access to non app mem %zx\n", addr);
776 DCHECK(IsAppMem(addr));
778 if (!IsShadowMem((uptr)shadow_mem)) {
779 Printf("Bad shadow addr %p (%zx)\n", shadow_mem, addr);
780 DCHECK(IsShadowMem((uptr)shadow_mem));
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);
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);
803 Shadow cur(fast_state);
804 cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
805 cur.SetWrite(kAccessIsWrite);
806 cur.SetAtomic(kIsAtomic);
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);
817 if (kCollectHistory) {
818 fast_state.IncrementEpoch();
819 thr->fast_state = fast_state;
820 TraceAddEvent(thr, fast_state, EventTypeMop, pc);
821 cur.IncrementEpoch();
824 MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
828 // Called by MemoryAccessRange in tsan_rtl_thread.cc
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);
842 MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
846 static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
853 uptr offset = addr % kShadowCell;
855 offset = kShadowCell - offset;
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))
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;) {
877 for (uptr j = 1; j < kShadowCnt; j++)
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;
886 // Set at least first kPageSize/2 to page boundary.
887 while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
889 for (uptr j = 1; j < kShadowCnt; j++)
892 // Reset middle part.
894 p = RoundDown(end, kPageSize);
895 UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
896 MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1);
900 for (uptr j = 1; j < kShadowCnt; j++)
906 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
907 MemoryRangeSet(thr, pc, addr, size, 0);
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.
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);
924 Shadow s(thr->fast_state);
928 s.SetAddr0AndSizeLog(0, 3);
929 MemoryRangeSet(thr, pc, addr, size, s.raw());
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);
937 Shadow s(thr->fast_state);
940 s.SetAddr0AndSizeLog(0, 3);
941 MemoryRangeSet(thr, pc, addr, size, s.raw());
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);
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);
957 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
959 if (thr->shadow_stack_pos == thr->shadow_stack_end)
960 GrowShadowStack(thr);
962 thr->shadow_stack_pos[0] = pc;
963 thr->shadow_stack_pos++;
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);
975 DCHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
977 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
979 thr->shadow_stack_pos--;
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();
988 if (save_stack && !ctx->after_multithreaded_fork)
989 thr->mop_ignore_set.Add(CurrentStackId(thr, pc));
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();
1000 thr->mop_ignore_set.Reset();
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;
1013 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc, bool save_stack) {
1014 DPrintf("#%d: ThreadIgnoreSyncBegin\n", thr->tid);
1016 CHECK_GT(thr->ignore_sync, 0);
1018 if (save_stack && !ctx->after_multithreaded_fork)
1019 thr->sync_ignore_set.Add(CurrentStackId(thr, pc));
1023 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc) {
1024 DPrintf("#%d: ThreadIgnoreSyncEnd\n", thr->tid);
1025 CHECK_GT(thr->ignore_sync, 0);
1028 if (thr->ignore_sync == 0)
1029 thr->sync_ignore_set.Reset();
1033 bool MD5Hash::operator==(const MD5Hash &other) const {
1034 return hash[0] == other.hash[0] && hash[1] == other.hash[1];
1038 void build_consistency_debug() {}
1040 void build_consistency_release() {}
1043 #if TSAN_COLLECT_STATS
1044 void build_consistency_stats() {}
1046 void build_consistency_nostats() {}
1049 } // namespace __tsan
1052 // Must be included in this file to make sure everything is inlined.
1053 #include "tsan_interface_inl.h"