1 //===-- tsan_rtl.cc -------------------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file is a part of ThreadSanitizer (TSan), a race detector.
11 //
12 // Main file (entry points) for the TSan run-time.
13 //===----------------------------------------------------------------------===//
14
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"
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>
36 typedef __m128i m128;
37 #endif
38
39 volatile int __tsan_resumed = 0;
40
__tsan_resume()41 extern "C" void __tsan_resume() {
42 __tsan_resumed = 1;
43 }
44
45 namespace __tsan {
46
47 #if !defined(SANITIZER_GO) && !SANITIZER_MAC
48 THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
49 #endif
50 static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
51 Context *ctx;
52
53 // Can be overriden by a front-end.
54 #ifdef TSAN_EXTERNAL_HOOKS
55 bool OnFinalize(bool failed);
56 void OnInitialize();
57 #else
58 SANITIZER_WEAK_CXX_DEFAULT_IMPL
OnFinalize(bool failed)59 bool OnFinalize(bool failed) {
60 return failed;
61 }
62 SANITIZER_WEAK_CXX_DEFAULT_IMPL
OnInitialize()63 void OnInitialize() {}
64 #endif
65
66 static char thread_registry_placeholder[sizeof(ThreadRegistry)];
67
CreateThreadContext(u32 tid)68 static ThreadContextBase *CreateThreadContext(u32 tid) {
69 // Map thread trace when context is created.
70 char name[50];
71 internal_snprintf(name, sizeof(name), "trace %u", tid);
72 MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event), name);
73 const uptr hdr = GetThreadTraceHeader(tid);
74 internal_snprintf(name, sizeof(name), "trace header %u", tid);
75 MapThreadTrace(hdr, sizeof(Trace), name);
76 new((void*)hdr) Trace();
77 // We are going to use only a small part of the trace with the default
78 // value of history_size. However, the constructor writes to the whole trace.
79 // Unmap the unused part.
80 uptr hdr_end = hdr + sizeof(Trace);
81 hdr_end -= sizeof(TraceHeader) * (kTraceParts - TraceParts());
82 hdr_end = RoundUp(hdr_end, GetPageSizeCached());
83 if (hdr_end < hdr + sizeof(Trace))
84 UnmapOrDie((void*)hdr_end, hdr + sizeof(Trace) - hdr_end);
85 void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
86 return new(mem) ThreadContext(tid);
87 }
88
89 #ifndef SANITIZER_GO
90 static const u32 kThreadQuarantineSize = 16;
91 #else
92 static const u32 kThreadQuarantineSize = 64;
93 #endif
94
Context()95 Context::Context()
96 : initialized()
97 , report_mtx(MutexTypeReport, StatMtxReport)
98 , nreported()
99 , nmissed_expected()
100 , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
101 CreateThreadContext, kMaxTid, kThreadQuarantineSize, kMaxTidReuse))
102 , racy_mtx(MutexTypeRacy, StatMtxRacy)
103 , racy_stacks(MBlockRacyStacks)
104 , racy_addresses(MBlockRacyAddresses)
105 , fired_suppressions_mtx(MutexTypeFired, StatMtxFired)
106 , fired_suppressions(8) {
107 }
108
109 // The objects are allocated in TLS, so one may rely on zero-initialization.
ThreadState(Context * ctx,int tid,int unique_id,u64 epoch,unsigned reuse_count,uptr stk_addr,uptr stk_size,uptr tls_addr,uptr tls_size)110 ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
111 unsigned reuse_count,
112 uptr stk_addr, uptr stk_size,
113 uptr tls_addr, uptr tls_size)
114 : fast_state(tid, epoch)
115 // Do not touch these, rely on zero initialization,
116 // they may be accessed before the ctor.
117 // , ignore_reads_and_writes()
118 // , ignore_interceptors()
119 , clock(tid, reuse_count)
120 #ifndef SANITIZER_GO
121 , jmp_bufs(MBlockJmpBuf)
122 #endif
123 , tid(tid)
124 , unique_id(unique_id)
125 , stk_addr(stk_addr)
126 , stk_size(stk_size)
127 , tls_addr(tls_addr)
128 , tls_size(tls_size)
129 #ifndef SANITIZER_GO
130 , last_sleep_clock(tid)
131 #endif
132 {
133 }
134
135 #ifndef SANITIZER_GO
MemoryProfiler(Context * ctx,fd_t fd,int i)136 static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
137 uptr n_threads;
138 uptr n_running_threads;
139 ctx->thread_registry->GetNumberOfThreads(&n_threads, &n_running_threads);
140 InternalScopedBuffer<char> buf(4096);
141 WriteMemoryProfile(buf.data(), buf.size(), n_threads, n_running_threads);
142 WriteToFile(fd, buf.data(), internal_strlen(buf.data()));
143 }
144
BackgroundThread(void * arg)145 static void BackgroundThread(void *arg) {
146 // This is a non-initialized non-user thread, nothing to see here.
147 // We don't use ScopedIgnoreInterceptors, because we want ignores to be
148 // enabled even when the thread function exits (e.g. during pthread thread
149 // shutdown code).
150 cur_thread()->ignore_interceptors++;
151 const u64 kMs2Ns = 1000 * 1000;
152
153 fd_t mprof_fd = kInvalidFd;
154 if (flags()->profile_memory && flags()->profile_memory[0]) {
155 if (internal_strcmp(flags()->profile_memory, "stdout") == 0) {
156 mprof_fd = 1;
157 } else if (internal_strcmp(flags()->profile_memory, "stderr") == 0) {
158 mprof_fd = 2;
159 } else {
160 InternalScopedString filename(kMaxPathLength);
161 filename.append("%s.%d", flags()->profile_memory, (int)internal_getpid());
162 fd_t fd = OpenFile(filename.data(), WrOnly);
163 if (fd == kInvalidFd) {
164 Printf("ThreadSanitizer: failed to open memory profile file '%s'\n",
165 &filename[0]);
166 } else {
167 mprof_fd = fd;
168 }
169 }
170 }
171
172 u64 last_flush = NanoTime();
173 uptr last_rss = 0;
174 for (int i = 0;
175 atomic_load(&ctx->stop_background_thread, memory_order_relaxed) == 0;
176 i++) {
177 SleepForMillis(100);
178 u64 now = NanoTime();
179
180 // Flush memory if requested.
181 if (flags()->flush_memory_ms > 0) {
182 if (last_flush + flags()->flush_memory_ms * kMs2Ns < now) {
183 VPrintf(1, "ThreadSanitizer: periodic memory flush\n");
184 FlushShadowMemory();
185 last_flush = NanoTime();
186 }
187 }
188 // GetRSS can be expensive on huge programs, so don't do it every 100ms.
189 if (flags()->memory_limit_mb > 0) {
190 uptr rss = GetRSS();
191 uptr limit = uptr(flags()->memory_limit_mb) << 20;
192 VPrintf(1, "ThreadSanitizer: memory flush check"
193 " RSS=%llu LAST=%llu LIMIT=%llu\n",
194 (u64)rss >> 20, (u64)last_rss >> 20, (u64)limit >> 20);
195 if (2 * rss > limit + last_rss) {
196 VPrintf(1, "ThreadSanitizer: flushing memory due to RSS\n");
197 FlushShadowMemory();
198 rss = GetRSS();
199 VPrintf(1, "ThreadSanitizer: memory flushed RSS=%llu\n", (u64)rss>>20);
200 }
201 last_rss = rss;
202 }
203
204 // Write memory profile if requested.
205 if (mprof_fd != kInvalidFd)
206 MemoryProfiler(ctx, mprof_fd, i);
207
208 // Flush symbolizer cache if requested.
209 if (flags()->flush_symbolizer_ms > 0) {
210 u64 last = atomic_load(&ctx->last_symbolize_time_ns,
211 memory_order_relaxed);
212 if (last != 0 && last + flags()->flush_symbolizer_ms * kMs2Ns < now) {
213 Lock l(&ctx->report_mtx);
214 SpinMutexLock l2(&CommonSanitizerReportMutex);
215 SymbolizeFlush();
216 atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
217 }
218 }
219 }
220 }
221
StartBackgroundThread()222 static void StartBackgroundThread() {
223 ctx->background_thread = internal_start_thread(&BackgroundThread, 0);
224 }
225
226 #ifndef __mips__
StopBackgroundThread()227 static void StopBackgroundThread() {
228 atomic_store(&ctx->stop_background_thread, 1, memory_order_relaxed);
229 internal_join_thread(ctx->background_thread);
230 ctx->background_thread = 0;
231 }
232 #endif
233 #endif
234
DontNeedShadowFor(uptr addr,uptr size)235 void DontNeedShadowFor(uptr addr, uptr size) {
236 uptr shadow_beg = MemToShadow(addr);
237 uptr shadow_end = MemToShadow(addr + size);
238 FlushUnneededShadowMemory(shadow_beg, shadow_end - shadow_beg);
239 }
240
MapShadow(uptr addr,uptr size)241 void MapShadow(uptr addr, uptr size) {
242 // Global data is not 64K aligned, but there are no adjacent mappings,
243 // so we can get away with unaligned mapping.
244 // CHECK_EQ(addr, addr & ~((64 << 10) - 1)); // windows wants 64K alignment
245 MmapFixedNoReserve(MemToShadow(addr), size * kShadowMultiplier, "shadow");
246
247 // Meta shadow is 2:1, so tread carefully.
248 static bool data_mapped = false;
249 static uptr mapped_meta_end = 0;
250 uptr meta_begin = (uptr)MemToMeta(addr);
251 uptr meta_end = (uptr)MemToMeta(addr + size);
252 meta_begin = RoundDownTo(meta_begin, 64 << 10);
253 meta_end = RoundUpTo(meta_end, 64 << 10);
254 if (!data_mapped) {
255 // First call maps data+bss.
256 data_mapped = true;
257 MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow");
258 } else {
259 // Mapping continous heap.
260 // Windows wants 64K alignment.
261 meta_begin = RoundDownTo(meta_begin, 64 << 10);
262 meta_end = RoundUpTo(meta_end, 64 << 10);
263 if (meta_end <= mapped_meta_end)
264 return;
265 if (meta_begin < mapped_meta_end)
266 meta_begin = mapped_meta_end;
267 MmapFixedNoReserve(meta_begin, meta_end - meta_begin, "meta shadow");
268 mapped_meta_end = meta_end;
269 }
270 VPrintf(2, "mapped meta shadow for (%p-%p) at (%p-%p)\n",
271 addr, addr+size, meta_begin, meta_end);
272 }
273
MapThreadTrace(uptr addr,uptr size,const char * name)274 void MapThreadTrace(uptr addr, uptr size, const char *name) {
275 DPrintf("#0: Mapping trace at %p-%p(0x%zx)\n", addr, addr + size, size);
276 CHECK_GE(addr, TraceMemBeg());
277 CHECK_LE(addr + size, TraceMemEnd());
278 CHECK_EQ(addr, addr & ~((64 << 10) - 1)); // windows wants 64K alignment
279 uptr addr1 = (uptr)MmapFixedNoReserve(addr, size, name);
280 if (addr1 != addr) {
281 Printf("FATAL: ThreadSanitizer can not mmap thread trace (%p/%p->%p)\n",
282 addr, size, addr1);
283 Die();
284 }
285 }
286
CheckShadowMapping()287 static void CheckShadowMapping() {
288 uptr beg, end;
289 for (int i = 0; GetUserRegion(i, &beg, &end); i++) {
290 VPrintf(3, "checking shadow region %p-%p\n", beg, end);
291 for (uptr p0 = beg; p0 <= end; p0 += (end - beg) / 4) {
292 for (int x = -1; x <= 1; x++) {
293 const uptr p = p0 + x;
294 if (p < beg || p >= end)
295 continue;
296 const uptr s = MemToShadow(p);
297 const uptr m = (uptr)MemToMeta(p);
298 VPrintf(3, " checking pointer %p: shadow=%p meta=%p\n", p, s, m);
299 CHECK(IsAppMem(p));
300 CHECK(IsShadowMem(s));
301 CHECK_EQ(p & ~(kShadowCell - 1), ShadowToMem(s));
302 CHECK(IsMetaMem(m));
303 }
304 }
305 }
306 }
307
Initialize(ThreadState * thr)308 void Initialize(ThreadState *thr) {
309 // Thread safe because done before all threads exist.
310 static bool is_initialized = false;
311 if (is_initialized)
312 return;
313 is_initialized = true;
314 // We are not ready to handle interceptors yet.
315 ScopedIgnoreInterceptors ignore;
316 SanitizerToolName = "ThreadSanitizer";
317 // Install tool-specific callbacks in sanitizer_common.
318 SetCheckFailedCallback(TsanCheckFailed);
319
320 ctx = new(ctx_placeholder) Context;
321 const char *options = GetEnv(kTsanOptionsEnv);
322 CacheBinaryName();
323 InitializeFlags(&ctx->flags, options);
324 AvoidCVE_2016_2143();
325 InitializePlatformEarly();
326 #ifndef SANITIZER_GO
327 // Re-exec ourselves if we need to set additional env or command line args.
328 MaybeReexec();
329
330 InitializeAllocator();
331 ReplaceSystemMalloc();
332 #endif
333 if (common_flags()->detect_deadlocks)
334 ctx->dd = DDetector::Create(flags());
335 Processor *proc = ProcCreate();
336 ProcWire(proc, thr);
337 InitializeInterceptors();
338 CheckShadowMapping();
339 InitializePlatform();
340 InitializeMutex();
341 InitializeDynamicAnnotations();
342 #ifndef SANITIZER_GO
343 InitializeShadowMemory();
344 InitializeAllocatorLate();
345 #endif
346 // Setup correct file descriptor for error reports.
347 __sanitizer_set_report_path(common_flags()->log_path);
348 InitializeSuppressions();
349 #ifndef SANITIZER_GO
350 InitializeLibIgnore();
351 Symbolizer::GetOrInit()->AddHooks(EnterSymbolizer, ExitSymbolizer);
352 // On MIPS, TSan initialization is run before
353 // __pthread_initialize_minimal_internal() is finished, so we can not spawn
354 // new threads.
355 #ifndef __mips__
356 StartBackgroundThread();
357 SetSandboxingCallback(StopBackgroundThread);
358 #endif
359 #endif
360
361 VPrintf(1, "***** Running under ThreadSanitizer v2 (pid %d) *****\n",
362 (int)internal_getpid());
363
364 // Initialize thread 0.
365 int tid = ThreadCreate(thr, 0, 0, true);
366 CHECK_EQ(tid, 0);
367 ThreadStart(thr, tid, internal_getpid());
368 #if TSAN_CONTAINS_UBSAN
369 __ubsan::InitAsPlugin();
370 #endif
371 ctx->initialized = true;
372
373 #ifndef SANITIZER_GO
374 Symbolizer::LateInitialize();
375 #endif
376
377 if (flags()->stop_on_start) {
378 Printf("ThreadSanitizer is suspended at startup (pid %d)."
379 " Call __tsan_resume().\n",
380 (int)internal_getpid());
381 while (__tsan_resumed == 0) {}
382 }
383
384 OnInitialize();
385 }
386
Finalize(ThreadState * thr)387 int Finalize(ThreadState *thr) {
388 bool failed = false;
389
390 if (flags()->atexit_sleep_ms > 0 && ThreadCount(thr) > 1)
391 SleepForMillis(flags()->atexit_sleep_ms);
392
393 // Wait for pending reports.
394 ctx->report_mtx.Lock();
395 CommonSanitizerReportMutex.Lock();
396 CommonSanitizerReportMutex.Unlock();
397 ctx->report_mtx.Unlock();
398
399 #ifndef SANITIZER_GO
400 if (Verbosity()) AllocatorPrintStats();
401 #endif
402
403 ThreadFinalize(thr);
404
405 if (ctx->nreported) {
406 failed = true;
407 #ifndef SANITIZER_GO
408 Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
409 #else
410 Printf("Found %d data race(s)\n", ctx->nreported);
411 #endif
412 }
413
414 if (ctx->nmissed_expected) {
415 failed = true;
416 Printf("ThreadSanitizer: missed %d expected races\n",
417 ctx->nmissed_expected);
418 }
419
420 if (common_flags()->print_suppressions)
421 PrintMatchedSuppressions();
422 #ifndef SANITIZER_GO
423 if (flags()->print_benign)
424 PrintMatchedBenignRaces();
425 #endif
426
427 failed = OnFinalize(failed);
428
429 #if TSAN_COLLECT_STATS
430 StatAggregate(ctx->stat, thr->stat);
431 StatOutput(ctx->stat);
432 #endif
433
434 return failed ? common_flags()->exitcode : 0;
435 }
436
437 #ifndef SANITIZER_GO
ForkBefore(ThreadState * thr,uptr pc)438 void ForkBefore(ThreadState *thr, uptr pc) {
439 ctx->thread_registry->Lock();
440 ctx->report_mtx.Lock();
441 }
442
ForkParentAfter(ThreadState * thr,uptr pc)443 void ForkParentAfter(ThreadState *thr, uptr pc) {
444 ctx->report_mtx.Unlock();
445 ctx->thread_registry->Unlock();
446 }
447
ForkChildAfter(ThreadState * thr,uptr pc)448 void ForkChildAfter(ThreadState *thr, uptr pc) {
449 ctx->report_mtx.Unlock();
450 ctx->thread_registry->Unlock();
451
452 uptr nthread = 0;
453 ctx->thread_registry->GetNumberOfThreads(0, 0, &nthread /* alive threads */);
454 VPrintf(1, "ThreadSanitizer: forked new process with pid %d,"
455 " parent had %d threads\n", (int)internal_getpid(), (int)nthread);
456 if (nthread == 1) {
457 StartBackgroundThread();
458 } else {
459 // We've just forked a multi-threaded process. We cannot reasonably function
460 // after that (some mutexes may be locked before fork). So just enable
461 // ignores for everything in the hope that we will exec soon.
462 ctx->after_multithreaded_fork = true;
463 thr->ignore_interceptors++;
464 ThreadIgnoreBegin(thr, pc);
465 ThreadIgnoreSyncBegin(thr, pc);
466 }
467 }
468 #endif
469
470 #ifdef SANITIZER_GO
471 NOINLINE
GrowShadowStack(ThreadState * thr)472 void GrowShadowStack(ThreadState *thr) {
473 const int sz = thr->shadow_stack_end - thr->shadow_stack;
474 const int newsz = 2 * sz;
475 uptr *newstack = (uptr*)internal_alloc(MBlockShadowStack,
476 newsz * sizeof(uptr));
477 internal_memcpy(newstack, thr->shadow_stack, sz * sizeof(uptr));
478 internal_free(thr->shadow_stack);
479 thr->shadow_stack = newstack;
480 thr->shadow_stack_pos = newstack + sz;
481 thr->shadow_stack_end = newstack + newsz;
482 }
483 #endif
484
CurrentStackId(ThreadState * thr,uptr pc)485 u32 CurrentStackId(ThreadState *thr, uptr pc) {
486 if (!thr->is_inited) // May happen during bootstrap.
487 return 0;
488 if (pc != 0) {
489 #ifndef SANITIZER_GO
490 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
491 #else
492 if (thr->shadow_stack_pos == thr->shadow_stack_end)
493 GrowShadowStack(thr);
494 #endif
495 thr->shadow_stack_pos[0] = pc;
496 thr->shadow_stack_pos++;
497 }
498 u32 id = StackDepotPut(
499 StackTrace(thr->shadow_stack, thr->shadow_stack_pos - thr->shadow_stack));
500 if (pc != 0)
501 thr->shadow_stack_pos--;
502 return id;
503 }
504
TraceSwitch(ThreadState * thr)505 void TraceSwitch(ThreadState *thr) {
506 thr->nomalloc++;
507 Trace *thr_trace = ThreadTrace(thr->tid);
508 Lock l(&thr_trace->mtx);
509 unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % TraceParts();
510 TraceHeader *hdr = &thr_trace->headers[trace];
511 hdr->epoch0 = thr->fast_state.epoch();
512 ObtainCurrentStack(thr, 0, &hdr->stack0);
513 hdr->mset0 = thr->mset;
514 thr->nomalloc--;
515 }
516
ThreadTrace(int tid)517 Trace *ThreadTrace(int tid) {
518 return (Trace*)GetThreadTraceHeader(tid);
519 }
520
TraceTopPC(ThreadState * thr)521 uptr TraceTopPC(ThreadState *thr) {
522 Event *events = (Event*)GetThreadTrace(thr->tid);
523 uptr pc = events[thr->fast_state.GetTracePos()];
524 return pc;
525 }
526
TraceSize()527 uptr TraceSize() {
528 return (uptr)(1ull << (kTracePartSizeBits + flags()->history_size + 1));
529 }
530
TraceParts()531 uptr TraceParts() {
532 return TraceSize() / kTracePartSize;
533 }
534
535 #ifndef SANITIZER_GO
__tsan_trace_switch()536 extern "C" void __tsan_trace_switch() {
537 TraceSwitch(cur_thread());
538 }
539
__tsan_report_race()540 extern "C" void __tsan_report_race() {
541 ReportRace(cur_thread());
542 }
543 #endif
544
545 ALWAYS_INLINE
LoadShadow(u64 * p)546 Shadow LoadShadow(u64 *p) {
547 u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
548 return Shadow(raw);
549 }
550
551 ALWAYS_INLINE
StoreShadow(u64 * sp,u64 s)552 void StoreShadow(u64 *sp, u64 s) {
553 atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
554 }
555
556 ALWAYS_INLINE
StoreIfNotYetStored(u64 * sp,u64 * s)557 void StoreIfNotYetStored(u64 *sp, u64 *s) {
558 StoreShadow(sp, *s);
559 *s = 0;
560 }
561
562 ALWAYS_INLINE
HandleRace(ThreadState * thr,u64 * shadow_mem,Shadow cur,Shadow old)563 void HandleRace(ThreadState *thr, u64 *shadow_mem,
564 Shadow cur, Shadow old) {
565 thr->racy_state[0] = cur.raw();
566 thr->racy_state[1] = old.raw();
567 thr->racy_shadow_addr = shadow_mem;
568 #ifndef SANITIZER_GO
569 HACKY_CALL(__tsan_report_race);
570 #else
571 ReportRace(thr);
572 #endif
573 }
574
HappensBefore(Shadow old,ThreadState * thr)575 static inline bool HappensBefore(Shadow old, ThreadState *thr) {
576 return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
577 }
578
579 ALWAYS_INLINE
MemoryAccessImpl1(ThreadState * thr,uptr addr,int kAccessSizeLog,bool kAccessIsWrite,bool kIsAtomic,u64 * shadow_mem,Shadow cur)580 void MemoryAccessImpl1(ThreadState *thr, uptr addr,
581 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
582 u64 *shadow_mem, Shadow cur) {
583 StatInc(thr, StatMop);
584 StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
585 StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
586
587 // This potentially can live in an MMX/SSE scratch register.
588 // The required intrinsics are:
589 // __m128i _mm_move_epi64(__m128i*);
590 // _mm_storel_epi64(u64*, __m128i);
591 u64 store_word = cur.raw();
592
593 // scan all the shadow values and dispatch to 4 categories:
594 // same, replace, candidate and race (see comments below).
595 // we consider only 3 cases regarding access sizes:
596 // equal, intersect and not intersect. initially I considered
597 // larger and smaller as well, it allowed to replace some
598 // 'candidates' with 'same' or 'replace', but I think
599 // it's just not worth it (performance- and complexity-wise).
600
601 Shadow old(0);
602
603 // It release mode we manually unroll the loop,
604 // because empirically gcc generates better code this way.
605 // However, we can't afford unrolling in debug mode, because the function
606 // consumes almost 4K of stack. Gtest gives only 4K of stack to death test
607 // threads, which is not enough for the unrolled loop.
608 #if SANITIZER_DEBUG
609 for (int idx = 0; idx < 4; idx++) {
610 #include "tsan_update_shadow_word_inl.h"
611 }
612 #else
613 int idx = 0;
614 #include "tsan_update_shadow_word_inl.h"
615 idx = 1;
616 #include "tsan_update_shadow_word_inl.h"
617 idx = 2;
618 #include "tsan_update_shadow_word_inl.h"
619 idx = 3;
620 #include "tsan_update_shadow_word_inl.h"
621 #endif
622
623 // we did not find any races and had already stored
624 // the current access info, so we are done
625 if (LIKELY(store_word == 0))
626 return;
627 // choose a random candidate slot and replace it
628 StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
629 StatInc(thr, StatShadowReplace);
630 return;
631 RACE:
632 HandleRace(thr, shadow_mem, cur, old);
633 return;
634 }
635
UnalignedMemoryAccess(ThreadState * thr,uptr pc,uptr addr,int size,bool kAccessIsWrite,bool kIsAtomic)636 void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
637 int size, bool kAccessIsWrite, bool kIsAtomic) {
638 while (size) {
639 int size1 = 1;
640 int kAccessSizeLog = kSizeLog1;
641 if (size >= 8 && (addr & ~7) == ((addr + 7) & ~7)) {
642 size1 = 8;
643 kAccessSizeLog = kSizeLog8;
644 } else if (size >= 4 && (addr & ~7) == ((addr + 3) & ~7)) {
645 size1 = 4;
646 kAccessSizeLog = kSizeLog4;
647 } else if (size >= 2 && (addr & ~7) == ((addr + 1) & ~7)) {
648 size1 = 2;
649 kAccessSizeLog = kSizeLog2;
650 }
651 MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
652 addr += size1;
653 size -= size1;
654 }
655 }
656
657 ALWAYS_INLINE
ContainsSameAccessSlow(u64 * s,u64 a,u64 sync_epoch,bool is_write)658 bool ContainsSameAccessSlow(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
659 Shadow cur(a);
660 for (uptr i = 0; i < kShadowCnt; i++) {
661 Shadow old(LoadShadow(&s[i]));
662 if (Shadow::Addr0AndSizeAreEqual(cur, old) &&
663 old.TidWithIgnore() == cur.TidWithIgnore() &&
664 old.epoch() > sync_epoch &&
665 old.IsAtomic() == cur.IsAtomic() &&
666 old.IsRead() <= cur.IsRead())
667 return true;
668 }
669 return false;
670 }
671
672 #if defined(__SSE3__)
673 #define SHUF(v0, v1, i0, i1, i2, i3) _mm_castps_si128(_mm_shuffle_ps( \
674 _mm_castsi128_ps(v0), _mm_castsi128_ps(v1), \
675 (i0)*1 + (i1)*4 + (i2)*16 + (i3)*64))
676 ALWAYS_INLINE
ContainsSameAccessFast(u64 * s,u64 a,u64 sync_epoch,bool is_write)677 bool ContainsSameAccessFast(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
678 // This is an optimized version of ContainsSameAccessSlow.
679 // load current access into access[0:63]
680 const m128 access = _mm_cvtsi64_si128(a);
681 // duplicate high part of access in addr0:
682 // addr0[0:31] = access[32:63]
683 // addr0[32:63] = access[32:63]
684 // addr0[64:95] = access[32:63]
685 // addr0[96:127] = access[32:63]
686 const m128 addr0 = SHUF(access, access, 1, 1, 1, 1);
687 // load 4 shadow slots
688 const m128 shadow0 = _mm_load_si128((__m128i*)s);
689 const m128 shadow1 = _mm_load_si128((__m128i*)s + 1);
690 // load high parts of 4 shadow slots into addr_vect:
691 // addr_vect[0:31] = shadow0[32:63]
692 // addr_vect[32:63] = shadow0[96:127]
693 // addr_vect[64:95] = shadow1[32:63]
694 // addr_vect[96:127] = shadow1[96:127]
695 m128 addr_vect = SHUF(shadow0, shadow1, 1, 3, 1, 3);
696 if (!is_write) {
697 // set IsRead bit in addr_vect
698 const m128 rw_mask1 = _mm_cvtsi64_si128(1<<15);
699 const m128 rw_mask = SHUF(rw_mask1, rw_mask1, 0, 0, 0, 0);
700 addr_vect = _mm_or_si128(addr_vect, rw_mask);
701 }
702 // addr0 == addr_vect?
703 const m128 addr_res = _mm_cmpeq_epi32(addr0, addr_vect);
704 // epoch1[0:63] = sync_epoch
705 const m128 epoch1 = _mm_cvtsi64_si128(sync_epoch);
706 // epoch[0:31] = sync_epoch[0:31]
707 // epoch[32:63] = sync_epoch[0:31]
708 // epoch[64:95] = sync_epoch[0:31]
709 // epoch[96:127] = sync_epoch[0:31]
710 const m128 epoch = SHUF(epoch1, epoch1, 0, 0, 0, 0);
711 // load low parts of shadow cell epochs into epoch_vect:
712 // epoch_vect[0:31] = shadow0[0:31]
713 // epoch_vect[32:63] = shadow0[64:95]
714 // epoch_vect[64:95] = shadow1[0:31]
715 // epoch_vect[96:127] = shadow1[64:95]
716 const m128 epoch_vect = SHUF(shadow0, shadow1, 0, 2, 0, 2);
717 // epoch_vect >= sync_epoch?
718 const m128 epoch_res = _mm_cmpgt_epi32(epoch_vect, epoch);
719 // addr_res & epoch_res
720 const m128 res = _mm_and_si128(addr_res, epoch_res);
721 // mask[0] = res[7]
722 // mask[1] = res[15]
723 // ...
724 // mask[15] = res[127]
725 const int mask = _mm_movemask_epi8(res);
726 return mask != 0;
727 }
728 #endif
729
730 ALWAYS_INLINE
ContainsSameAccess(u64 * s,u64 a,u64 sync_epoch,bool is_write)731 bool ContainsSameAccess(u64 *s, u64 a, u64 sync_epoch, bool is_write) {
732 #if defined(__SSE3__)
733 bool res = ContainsSameAccessFast(s, a, sync_epoch, is_write);
734 // NOTE: this check can fail if the shadow is concurrently mutated
735 // by other threads. But it still can be useful if you modify
736 // ContainsSameAccessFast and want to ensure that it's not completely broken.
737 // DCHECK_EQ(res, ContainsSameAccessSlow(s, a, sync_epoch, is_write));
738 return res;
739 #else
740 return ContainsSameAccessSlow(s, a, sync_epoch, is_write);
741 #endif
742 }
743
744 ALWAYS_INLINE USED
MemoryAccess(ThreadState * thr,uptr pc,uptr addr,int kAccessSizeLog,bool kAccessIsWrite,bool kIsAtomic)745 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
746 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic) {
747 u64 *shadow_mem = (u64*)MemToShadow(addr);
748 DPrintf2("#%d: MemoryAccess: @%p %p size=%d"
749 " is_write=%d shadow_mem=%p {%zx, %zx, %zx, %zx}\n",
750 (int)thr->fast_state.tid(), (void*)pc, (void*)addr,
751 (int)(1 << kAccessSizeLog), kAccessIsWrite, shadow_mem,
752 (uptr)shadow_mem[0], (uptr)shadow_mem[1],
753 (uptr)shadow_mem[2], (uptr)shadow_mem[3]);
754 #if SANITIZER_DEBUG
755 if (!IsAppMem(addr)) {
756 Printf("Access to non app mem %zx\n", addr);
757 DCHECK(IsAppMem(addr));
758 }
759 if (!IsShadowMem((uptr)shadow_mem)) {
760 Printf("Bad shadow addr %p (%zx)\n", shadow_mem, addr);
761 DCHECK(IsShadowMem((uptr)shadow_mem));
762 }
763 #endif
764
765 if (kCppMode && *shadow_mem == kShadowRodata) {
766 // Access to .rodata section, no races here.
767 // Measurements show that it can be 10-20% of all memory accesses.
768 StatInc(thr, StatMop);
769 StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
770 StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
771 StatInc(thr, StatMopRodata);
772 return;
773 }
774
775 FastState fast_state = thr->fast_state;
776 if (fast_state.GetIgnoreBit()) {
777 StatInc(thr, StatMop);
778 StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
779 StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
780 StatInc(thr, StatMopIgnored);
781 return;
782 }
783
784 Shadow cur(fast_state);
785 cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
786 cur.SetWrite(kAccessIsWrite);
787 cur.SetAtomic(kIsAtomic);
788
789 if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
790 thr->fast_synch_epoch, kAccessIsWrite))) {
791 StatInc(thr, StatMop);
792 StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
793 StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
794 StatInc(thr, StatMopSame);
795 return;
796 }
797
798 if (kCollectHistory) {
799 fast_state.IncrementEpoch();
800 thr->fast_state = fast_state;
801 TraceAddEvent(thr, fast_state, EventTypeMop, pc);
802 cur.IncrementEpoch();
803 }
804
805 MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
806 shadow_mem, cur);
807 }
808
809 // Called by MemoryAccessRange in tsan_rtl_thread.cc
810 ALWAYS_INLINE USED
MemoryAccessImpl(ThreadState * thr,uptr addr,int kAccessSizeLog,bool kAccessIsWrite,bool kIsAtomic,u64 * shadow_mem,Shadow cur)811 void MemoryAccessImpl(ThreadState *thr, uptr addr,
812 int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
813 u64 *shadow_mem, Shadow cur) {
814 if (LIKELY(ContainsSameAccess(shadow_mem, cur.raw(),
815 thr->fast_synch_epoch, kAccessIsWrite))) {
816 StatInc(thr, StatMop);
817 StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
818 StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
819 StatInc(thr, StatMopSame);
820 return;
821 }
822
823 MemoryAccessImpl1(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
824 shadow_mem, cur);
825 }
826
MemoryRangeSet(ThreadState * thr,uptr pc,uptr addr,uptr size,u64 val)827 static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
828 u64 val) {
829 (void)thr;
830 (void)pc;
831 if (size == 0)
832 return;
833 // FIXME: fix me.
834 uptr offset = addr % kShadowCell;
835 if (offset) {
836 offset = kShadowCell - offset;
837 if (size <= offset)
838 return;
839 addr += offset;
840 size -= offset;
841 }
842 DCHECK_EQ(addr % 8, 0);
843 // If a user passes some insane arguments (memset(0)),
844 // let it just crash as usual.
845 if (!IsAppMem(addr) || !IsAppMem(addr + size - 1))
846 return;
847 // Don't want to touch lots of shadow memory.
848 // If a program maps 10MB stack, there is no need reset the whole range.
849 size = (size + (kShadowCell - 1)) & ~(kShadowCell - 1);
850 // UnmapOrDie/MmapFixedNoReserve does not work on Windows,
851 // so we do it only for C/C++.
852 if (kGoMode || size < common_flags()->clear_shadow_mmap_threshold) {
853 u64 *p = (u64*)MemToShadow(addr);
854 CHECK(IsShadowMem((uptr)p));
855 CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
856 // FIXME: may overwrite a part outside the region
857 for (uptr i = 0; i < size / kShadowCell * kShadowCnt;) {
858 p[i++] = val;
859 for (uptr j = 1; j < kShadowCnt; j++)
860 p[i++] = 0;
861 }
862 } else {
863 // The region is big, reset only beginning and end.
864 const uptr kPageSize = GetPageSizeCached();
865 u64 *begin = (u64*)MemToShadow(addr);
866 u64 *end = begin + size / kShadowCell * kShadowCnt;
867 u64 *p = begin;
868 // Set at least first kPageSize/2 to page boundary.
869 while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
870 *p++ = val;
871 for (uptr j = 1; j < kShadowCnt; j++)
872 *p++ = 0;
873 }
874 // Reset middle part.
875 u64 *p1 = p;
876 p = RoundDown(end, kPageSize);
877 UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
878 MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1);
879 // Set the ending.
880 while (p < end) {
881 *p++ = val;
882 for (uptr j = 1; j < kShadowCnt; j++)
883 *p++ = 0;
884 }
885 }
886 }
887
MemoryResetRange(ThreadState * thr,uptr pc,uptr addr,uptr size)888 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
889 MemoryRangeSet(thr, pc, addr, size, 0);
890 }
891
MemoryRangeFreed(ThreadState * thr,uptr pc,uptr addr,uptr size)892 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
893 // Processing more than 1k (4k of shadow) is expensive,
894 // can cause excessive memory consumption (user does not necessary touch
895 // the whole range) and most likely unnecessary.
896 if (size > 1024)
897 size = 1024;
898 CHECK_EQ(thr->is_freeing, false);
899 thr->is_freeing = true;
900 MemoryAccessRange(thr, pc, addr, size, true);
901 thr->is_freeing = false;
902 if (kCollectHistory) {
903 thr->fast_state.IncrementEpoch();
904 TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
905 }
906 Shadow s(thr->fast_state);
907 s.ClearIgnoreBit();
908 s.MarkAsFreed();
909 s.SetWrite(true);
910 s.SetAddr0AndSizeLog(0, 3);
911 MemoryRangeSet(thr, pc, addr, size, s.raw());
912 }
913
MemoryRangeImitateWrite(ThreadState * thr,uptr pc,uptr addr,uptr size)914 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
915 if (kCollectHistory) {
916 thr->fast_state.IncrementEpoch();
917 TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
918 }
919 Shadow s(thr->fast_state);
920 s.ClearIgnoreBit();
921 s.SetWrite(true);
922 s.SetAddr0AndSizeLog(0, 3);
923 MemoryRangeSet(thr, pc, addr, size, s.raw());
924 }
925
926 ALWAYS_INLINE USED
FuncEntry(ThreadState * thr,uptr pc)927 void FuncEntry(ThreadState *thr, uptr pc) {
928 StatInc(thr, StatFuncEnter);
929 DPrintf2("#%d: FuncEntry %p\n", (int)thr->fast_state.tid(), (void*)pc);
930 if (kCollectHistory) {
931 thr->fast_state.IncrementEpoch();
932 TraceAddEvent(thr, thr->fast_state, EventTypeFuncEnter, pc);
933 }
934
935 // Shadow stack maintenance can be replaced with
936 // stack unwinding during trace switch (which presumably must be faster).
937 DCHECK_GE(thr->shadow_stack_pos, thr->shadow_stack);
938 #ifndef SANITIZER_GO
939 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
940 #else
941 if (thr->shadow_stack_pos == thr->shadow_stack_end)
942 GrowShadowStack(thr);
943 #endif
944 thr->shadow_stack_pos[0] = pc;
945 thr->shadow_stack_pos++;
946 }
947
948 ALWAYS_INLINE USED
FuncExit(ThreadState * thr)949 void FuncExit(ThreadState *thr) {
950 StatInc(thr, StatFuncExit);
951 DPrintf2("#%d: FuncExit\n", (int)thr->fast_state.tid());
952 if (kCollectHistory) {
953 thr->fast_state.IncrementEpoch();
954 TraceAddEvent(thr, thr->fast_state, EventTypeFuncExit, 0);
955 }
956
957 DCHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
958 #ifndef SANITIZER_GO
959 DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
960 #endif
961 thr->shadow_stack_pos--;
962 }
963
ThreadIgnoreBegin(ThreadState * thr,uptr pc)964 void ThreadIgnoreBegin(ThreadState *thr, uptr pc) {
965 DPrintf("#%d: ThreadIgnoreBegin\n", thr->tid);
966 thr->ignore_reads_and_writes++;
967 CHECK_GT(thr->ignore_reads_and_writes, 0);
968 thr->fast_state.SetIgnoreBit();
969 #ifndef SANITIZER_GO
970 if (!ctx->after_multithreaded_fork)
971 thr->mop_ignore_set.Add(CurrentStackId(thr, pc));
972 #endif
973 }
974
ThreadIgnoreEnd(ThreadState * thr,uptr pc)975 void ThreadIgnoreEnd(ThreadState *thr, uptr pc) {
976 DPrintf("#%d: ThreadIgnoreEnd\n", thr->tid);
977 thr->ignore_reads_and_writes--;
978 CHECK_GE(thr->ignore_reads_and_writes, 0);
979 if (thr->ignore_reads_and_writes == 0) {
980 thr->fast_state.ClearIgnoreBit();
981 #ifndef SANITIZER_GO
982 thr->mop_ignore_set.Reset();
983 #endif
984 }
985 }
986
ThreadIgnoreSyncBegin(ThreadState * thr,uptr pc)987 void ThreadIgnoreSyncBegin(ThreadState *thr, uptr pc) {
988 DPrintf("#%d: ThreadIgnoreSyncBegin\n", thr->tid);
989 thr->ignore_sync++;
990 CHECK_GT(thr->ignore_sync, 0);
991 #ifndef SANITIZER_GO
992 if (!ctx->after_multithreaded_fork)
993 thr->sync_ignore_set.Add(CurrentStackId(thr, pc));
994 #endif
995 }
996
ThreadIgnoreSyncEnd(ThreadState * thr,uptr pc)997 void ThreadIgnoreSyncEnd(ThreadState *thr, uptr pc) {
998 DPrintf("#%d: ThreadIgnoreSyncEnd\n", thr->tid);
999 thr->ignore_sync--;
1000 CHECK_GE(thr->ignore_sync, 0);
1001 #ifndef SANITIZER_GO
1002 if (thr->ignore_sync == 0)
1003 thr->sync_ignore_set.Reset();
1004 #endif
1005 }
1006
operator ==(const MD5Hash & other) const1007 bool MD5Hash::operator==(const MD5Hash &other) const {
1008 return hash[0] == other.hash[0] && hash[1] == other.hash[1];
1009 }
1010
1011 #if SANITIZER_DEBUG
build_consistency_debug()1012 void build_consistency_debug() {}
1013 #else
build_consistency_release()1014 void build_consistency_release() {}
1015 #endif
1016
1017 #if TSAN_COLLECT_STATS
build_consistency_stats()1018 void build_consistency_stats() {}
1019 #else
build_consistency_nostats()1020 void build_consistency_nostats() {}
1021 #endif
1022
1023 } // namespace __tsan
1024
1025 #ifndef SANITIZER_GO
1026 // Must be included in this file to make sure everything is inlined.
1027 #include "tsan_interface_inl.h"
1028 #endif
1029