1 //===-- tsan_platform_linux.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 // Linux- and BSD-specific code. 12 //===----------------------------------------------------------------------===// 13 14 #include "sanitizer_common/sanitizer_platform.h" 15 #if SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD 16 17 #include "sanitizer_common/sanitizer_common.h" 18 #include "sanitizer_common/sanitizer_libc.h" 19 #include "sanitizer_common/sanitizer_linux.h" 20 #include "sanitizer_common/sanitizer_platform_limits_netbsd.h" 21 #include "sanitizer_common/sanitizer_platform_limits_posix.h" 22 #include "sanitizer_common/sanitizer_posix.h" 23 #include "sanitizer_common/sanitizer_procmaps.h" 24 #include "sanitizer_common/sanitizer_stackdepot.h" 25 #include "sanitizer_common/sanitizer_stoptheworld.h" 26 #include "tsan_flags.h" 27 #include "tsan_platform.h" 28 #include "tsan_rtl.h" 29 30 #include <fcntl.h> 31 #include <pthread.h> 32 #include <signal.h> 33 #include <stdio.h> 34 #include <stdlib.h> 35 #include <string.h> 36 #include <stdarg.h> 37 #include <sys/mman.h> 38 #if SANITIZER_LINUX 39 #include <sys/personality.h> 40 #include <setjmp.h> 41 #endif 42 #include <sys/syscall.h> 43 #include <sys/socket.h> 44 #include <sys/time.h> 45 #include <sys/types.h> 46 #include <sys/resource.h> 47 #include <sys/stat.h> 48 #include <unistd.h> 49 #include <sched.h> 50 #include <dlfcn.h> 51 #if SANITIZER_LINUX 52 #define __need_res_state 53 #include <resolv.h> 54 #endif 55 56 #ifdef sa_handler 57 # undef sa_handler 58 #endif 59 60 #ifdef sa_sigaction 61 # undef sa_sigaction 62 #endif 63 64 #if SANITIZER_FREEBSD 65 extern "C" void *__libc_stack_end; 66 void *__libc_stack_end = 0; 67 #endif 68 69 #if SANITIZER_LINUX && defined(__aarch64__) && !SANITIZER_GO 70 # define INIT_LONGJMP_XOR_KEY 1 71 #else 72 # define INIT_LONGJMP_XOR_KEY 0 73 #endif 74 75 #if INIT_LONGJMP_XOR_KEY 76 #include "interception/interception.h" 77 // Must be declared outside of other namespaces. 78 DECLARE_REAL(int, _setjmp, void *env) 79 #endif 80 81 namespace __tsan { 82 83 #if INIT_LONGJMP_XOR_KEY 84 static void InitializeLongjmpXorKey(); 85 static uptr longjmp_xor_key; 86 #endif 87 88 // Runtime detected VMA size. 89 uptr vmaSize; 90 91 enum { 92 MemTotal, 93 MemShadow, 94 MemMeta, 95 MemFile, 96 MemMmap, 97 MemTrace, 98 MemHeap, 99 MemOther, 100 MemCount, 101 }; 102 103 void FillProfileCallback(uptr p, uptr rss, bool file, uptr *mem) { 104 mem[MemTotal] += rss; 105 if (p >= ShadowBeg() && p < ShadowEnd()) 106 mem[MemShadow] += rss; 107 else if (p >= MetaShadowBeg() && p < MetaShadowEnd()) 108 mem[MemMeta] += rss; 109 else if ((p >= LoAppMemBeg() && p < LoAppMemEnd()) || 110 (p >= MidAppMemBeg() && p < MidAppMemEnd()) || 111 (p >= HiAppMemBeg() && p < HiAppMemEnd())) 112 mem[file ? MemFile : MemMmap] += rss; 113 else if (p >= HeapMemBeg() && p < HeapMemEnd()) 114 mem[MemHeap] += rss; 115 else if (p >= TraceMemBeg() && p < TraceMemEnd()) 116 mem[MemTrace] += rss; 117 else 118 mem[MemOther] += rss; 119 } 120 121 void WriteMemoryProfile(char *buf, uptr buf_size, u64 uptime_ns) { 122 uptr mem[MemCount]; 123 internal_memset(mem, 0, sizeof(mem)); 124 GetMemoryProfile(FillProfileCallback, mem); 125 auto meta = ctx->metamap.GetMemoryStats(); 126 StackDepotStats stacks = StackDepotGetStats(); 127 uptr nthread, nlive; 128 ctx->thread_registry.GetNumberOfThreads(&nthread, &nlive); 129 uptr internal_stats[AllocatorStatCount]; 130 internal_allocator()->GetStats(internal_stats); 131 // All these are allocated from the common mmap region. 132 mem[MemMmap] -= meta.mem_block + meta.sync_obj + stacks.allocated + 133 internal_stats[AllocatorStatMapped]; 134 if (s64(mem[MemMmap]) < 0) 135 mem[MemMmap] = 0; 136 internal_snprintf( 137 buf, buf_size, 138 "%llus: RSS %zd MB: shadow:%zd meta:%zd file:%zd mmap:%zd" 139 " trace:%zd heap:%zd other:%zd intalloc:%zd memblocks:%zd syncobj:%zu" 140 " stacks=%zd[%zd] nthr=%zd/%zd\n", 141 uptime_ns / (1000 * 1000 * 1000), mem[MemTotal] >> 20, 142 mem[MemShadow] >> 20, mem[MemMeta] >> 20, mem[MemFile] >> 20, 143 mem[MemMmap] >> 20, mem[MemTrace] >> 20, mem[MemHeap] >> 20, 144 mem[MemOther] >> 20, internal_stats[AllocatorStatMapped] >> 20, 145 meta.mem_block >> 20, meta.sync_obj >> 20, stacks.allocated >> 20, 146 stacks.n_uniq_ids, nlive, nthread); 147 } 148 149 # if SANITIZER_LINUX 150 void FlushShadowMemoryCallback( 151 const SuspendedThreadsList &suspended_threads_list, 152 void *argument) { 153 ReleaseMemoryPagesToOS(ShadowBeg(), ShadowEnd()); 154 } 155 #endif 156 157 void FlushShadowMemory() { 158 #if SANITIZER_LINUX 159 StopTheWorld(FlushShadowMemoryCallback, 0); 160 #endif 161 } 162 163 #if !SANITIZER_GO 164 // Mark shadow for .rodata sections with the special kShadowRodata marker. 165 // Accesses to .rodata can't race, so this saves time, memory and trace space. 166 static void MapRodata() { 167 // First create temp file. 168 const char *tmpdir = GetEnv("TMPDIR"); 169 if (tmpdir == 0) 170 tmpdir = GetEnv("TEST_TMPDIR"); 171 #ifdef P_tmpdir 172 if (tmpdir == 0) 173 tmpdir = P_tmpdir; 174 #endif 175 if (tmpdir == 0) 176 return; 177 char name[256]; 178 internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d", 179 tmpdir, (int)internal_getpid()); 180 uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600); 181 if (internal_iserror(openrv)) 182 return; 183 internal_unlink(name); // Unlink it now, so that we can reuse the buffer. 184 fd_t fd = openrv; 185 // Fill the file with kShadowRodata. 186 const uptr kMarkerSize = 512 * 1024 / sizeof(RawShadow); 187 InternalMmapVector<RawShadow> marker(kMarkerSize); 188 // volatile to prevent insertion of memset 189 for (volatile RawShadow *p = marker.data(); p < marker.data() + kMarkerSize; 190 p++) 191 *p = kShadowRodata; 192 internal_write(fd, marker.data(), marker.size() * sizeof(RawShadow)); 193 // Map the file into memory. 194 uptr page = internal_mmap(0, GetPageSizeCached(), PROT_READ | PROT_WRITE, 195 MAP_PRIVATE | MAP_ANONYMOUS, fd, 0); 196 if (internal_iserror(page)) { 197 internal_close(fd); 198 return; 199 } 200 // Map the file into shadow of .rodata sections. 201 MemoryMappingLayout proc_maps(/*cache_enabled*/true); 202 // Reusing the buffer 'name'. 203 MemoryMappedSegment segment(name, ARRAY_SIZE(name)); 204 while (proc_maps.Next(&segment)) { 205 if (segment.filename[0] != 0 && segment.filename[0] != '[' && 206 segment.IsReadable() && segment.IsExecutable() && 207 !segment.IsWritable() && IsAppMem(segment.start)) { 208 // Assume it's .rodata 209 char *shadow_start = (char *)MemToShadow(segment.start); 210 char *shadow_end = (char *)MemToShadow(segment.end); 211 for (char *p = shadow_start; p < shadow_end; 212 p += marker.size() * sizeof(RawShadow)) { 213 internal_mmap( 214 p, Min<uptr>(marker.size() * sizeof(RawShadow), shadow_end - p), 215 PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0); 216 } 217 } 218 } 219 internal_close(fd); 220 } 221 222 void InitializeShadowMemoryPlatform() { 223 MapRodata(); 224 } 225 226 #endif // #if !SANITIZER_GO 227 228 void InitializePlatformEarly() { 229 vmaSize = 230 (MostSignificantSetBitIndex(GET_CURRENT_FRAME()) + 1); 231 #if defined(__aarch64__) 232 # if !SANITIZER_GO 233 if (vmaSize != 39 && vmaSize != 42 && vmaSize != 48) { 234 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 235 Printf("FATAL: Found %zd - Supported 39, 42 and 48\n", vmaSize); 236 Die(); 237 } 238 #else 239 if (vmaSize != 48) { 240 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 241 Printf("FATAL: Found %zd - Supported 48\n", vmaSize); 242 Die(); 243 } 244 #endif 245 #elif defined(__powerpc64__) 246 # if !SANITIZER_GO 247 if (vmaSize != 44 && vmaSize != 46 && vmaSize != 47) { 248 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 249 Printf("FATAL: Found %zd - Supported 44, 46, and 47\n", vmaSize); 250 Die(); 251 } 252 # else 253 if (vmaSize != 46 && vmaSize != 47) { 254 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 255 Printf("FATAL: Found %zd - Supported 46, and 47\n", vmaSize); 256 Die(); 257 } 258 # endif 259 #elif defined(__mips64) 260 # if !SANITIZER_GO 261 if (vmaSize != 40) { 262 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 263 Printf("FATAL: Found %zd - Supported 40\n", vmaSize); 264 Die(); 265 } 266 # else 267 if (vmaSize != 47) { 268 Printf("FATAL: ThreadSanitizer: unsupported VMA range\n"); 269 Printf("FATAL: Found %zd - Supported 47\n", vmaSize); 270 Die(); 271 } 272 # endif 273 #endif 274 } 275 276 void InitializePlatform() { 277 DisableCoreDumperIfNecessary(); 278 279 // Go maps shadow memory lazily and works fine with limited address space. 280 // Unlimited stack is not a problem as well, because the executable 281 // is not compiled with -pie. 282 #if !SANITIZER_GO 283 { 284 bool reexec = false; 285 // TSan doesn't play well with unlimited stack size (as stack 286 // overlaps with shadow memory). If we detect unlimited stack size, 287 // we re-exec the program with limited stack size as a best effort. 288 if (StackSizeIsUnlimited()) { 289 const uptr kMaxStackSize = 32 * 1024 * 1024; 290 VReport(1, "Program is run with unlimited stack size, which wouldn't " 291 "work with ThreadSanitizer.\n" 292 "Re-execing with stack size limited to %zd bytes.\n", 293 kMaxStackSize); 294 SetStackSizeLimitInBytes(kMaxStackSize); 295 reexec = true; 296 } 297 298 if (!AddressSpaceIsUnlimited()) { 299 Report("WARNING: Program is run with limited virtual address space," 300 " which wouldn't work with ThreadSanitizer.\n"); 301 Report("Re-execing with unlimited virtual address space.\n"); 302 SetAddressSpaceUnlimited(); 303 reexec = true; 304 } 305 #if SANITIZER_LINUX && defined(__aarch64__) 306 // After patch "arm64: mm: support ARCH_MMAP_RND_BITS." is introduced in 307 // linux kernel, the random gap between stack and mapped area is increased 308 // from 128M to 36G on 39-bit aarch64. As it is almost impossible to cover 309 // this big range, we should disable randomized virtual space on aarch64. 310 int old_personality = personality(0xffffffff); 311 if (old_personality != -1 && (old_personality & ADDR_NO_RANDOMIZE) == 0) { 312 VReport(1, "WARNING: Program is run with randomized virtual address " 313 "space, which wouldn't work with ThreadSanitizer.\n" 314 "Re-execing with fixed virtual address space.\n"); 315 CHECK_NE(personality(old_personality | ADDR_NO_RANDOMIZE), -1); 316 reexec = true; 317 } 318 // Initialize the xor key used in {sig}{set,long}jump. 319 InitializeLongjmpXorKey(); 320 #endif 321 if (reexec) 322 ReExec(); 323 } 324 325 CheckAndProtect(); 326 InitTlsSize(); 327 #endif // !SANITIZER_GO 328 } 329 330 #if !SANITIZER_GO 331 // Extract file descriptors passed to glibc internal __res_iclose function. 332 // This is required to properly "close" the fds, because we do not see internal 333 // closes within glibc. The code is a pure hack. 334 int ExtractResolvFDs(void *state, int *fds, int nfd) { 335 #if SANITIZER_LINUX && !SANITIZER_ANDROID 336 int cnt = 0; 337 struct __res_state *statp = (struct __res_state*)state; 338 for (int i = 0; i < MAXNS && cnt < nfd; i++) { 339 if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1) 340 fds[cnt++] = statp->_u._ext.nssocks[i]; 341 } 342 return cnt; 343 #else 344 return 0; 345 #endif 346 } 347 348 // Extract file descriptors passed via UNIX domain sockets. 349 // This is required to properly handle "open" of these fds. 350 // see 'man recvmsg' and 'man 3 cmsg'. 351 int ExtractRecvmsgFDs(void *msgp, int *fds, int nfd) { 352 int res = 0; 353 msghdr *msg = (msghdr*)msgp; 354 struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg); 355 for (; cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { 356 if (cmsg->cmsg_level != SOL_SOCKET || cmsg->cmsg_type != SCM_RIGHTS) 357 continue; 358 int n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(fds[0]); 359 for (int i = 0; i < n; i++) { 360 fds[res++] = ((int*)CMSG_DATA(cmsg))[i]; 361 if (res == nfd) 362 return res; 363 } 364 } 365 return res; 366 } 367 368 // Reverse operation of libc stack pointer mangling 369 static uptr UnmangleLongJmpSp(uptr mangled_sp) { 370 #if defined(__x86_64__) 371 # if SANITIZER_LINUX 372 // Reverse of: 373 // xor %fs:0x30, %rsi 374 // rol $0x11, %rsi 375 uptr sp; 376 asm("ror $0x11, %0 \n" 377 "xor %%fs:0x30, %0 \n" 378 : "=r" (sp) 379 : "0" (mangled_sp)); 380 return sp; 381 # else 382 return mangled_sp; 383 # endif 384 #elif defined(__aarch64__) 385 # if SANITIZER_LINUX 386 return mangled_sp ^ longjmp_xor_key; 387 # else 388 return mangled_sp; 389 # endif 390 #elif defined(__powerpc64__) 391 // Reverse of: 392 // ld r4, -28696(r13) 393 // xor r4, r3, r4 394 uptr xor_key; 395 asm("ld %0, -28696(%%r13)" : "=r" (xor_key)); 396 return mangled_sp ^ xor_key; 397 #elif defined(__mips__) 398 return mangled_sp; 399 #elif defined(__s390x__) 400 // tcbhead_t.stack_guard 401 uptr xor_key = ((uptr *)__builtin_thread_pointer())[5]; 402 return mangled_sp ^ xor_key; 403 #else 404 #error "Unknown platform" 405 #endif 406 } 407 408 #if SANITIZER_NETBSD 409 # ifdef __x86_64__ 410 # define LONG_JMP_SP_ENV_SLOT 6 411 # else 412 # error unsupported 413 # endif 414 #elif defined(__powerpc__) 415 # define LONG_JMP_SP_ENV_SLOT 0 416 #elif SANITIZER_FREEBSD 417 # define LONG_JMP_SP_ENV_SLOT 2 418 #elif SANITIZER_LINUX 419 # ifdef __aarch64__ 420 # define LONG_JMP_SP_ENV_SLOT 13 421 # elif defined(__mips64) 422 # define LONG_JMP_SP_ENV_SLOT 1 423 # elif defined(__s390x__) 424 # define LONG_JMP_SP_ENV_SLOT 9 425 # else 426 # define LONG_JMP_SP_ENV_SLOT 6 427 # endif 428 #endif 429 430 uptr ExtractLongJmpSp(uptr *env) { 431 uptr mangled_sp = env[LONG_JMP_SP_ENV_SLOT]; 432 return UnmangleLongJmpSp(mangled_sp); 433 } 434 435 #if INIT_LONGJMP_XOR_KEY 436 // GLIBC mangles the function pointers in jmp_buf (used in {set,long}*jmp 437 // functions) by XORing them with a random key. For AArch64 it is a global 438 // variable rather than a TCB one (as for x86_64/powerpc). We obtain the key by 439 // issuing a setjmp and XORing the SP pointer values to derive the key. 440 static void InitializeLongjmpXorKey() { 441 // 1. Call REAL(setjmp), which stores the mangled SP in env. 442 jmp_buf env; 443 REAL(_setjmp)(env); 444 445 // 2. Retrieve vanilla/mangled SP. 446 uptr sp; 447 asm("mov %0, sp" : "=r" (sp)); 448 uptr mangled_sp = ((uptr *)&env)[LONG_JMP_SP_ENV_SLOT]; 449 450 // 3. xor SPs to obtain key. 451 longjmp_xor_key = mangled_sp ^ sp; 452 } 453 #endif 454 455 extern "C" void __tsan_tls_initialization() {} 456 457 void ImitateTlsWrite(ThreadState *thr, uptr tls_addr, uptr tls_size) { 458 // Check that the thr object is in tls; 459 const uptr thr_beg = (uptr)thr; 460 const uptr thr_end = (uptr)thr + sizeof(*thr); 461 CHECK_GE(thr_beg, tls_addr); 462 CHECK_LE(thr_beg, tls_addr + tls_size); 463 CHECK_GE(thr_end, tls_addr); 464 CHECK_LE(thr_end, tls_addr + tls_size); 465 // Since the thr object is huge, skip it. 466 const uptr pc = StackTrace::GetNextInstructionPc( 467 reinterpret_cast<uptr>(__tsan_tls_initialization)); 468 MemoryRangeImitateWrite(thr, pc, tls_addr, thr_beg - tls_addr); 469 MemoryRangeImitateWrite(thr, pc, thr_end, tls_addr + tls_size - thr_end); 470 } 471 472 // Note: this function runs with async signals enabled, 473 // so it must not touch any tsan state. 474 int call_pthread_cancel_with_cleanup(int (*fn)(void *arg), 475 void (*cleanup)(void *arg), void *arg) { 476 // pthread_cleanup_push/pop are hardcore macros mess. 477 // We can't intercept nor call them w/o including pthread.h. 478 int res; 479 pthread_cleanup_push(cleanup, arg); 480 res = fn(arg); 481 pthread_cleanup_pop(0); 482 return res; 483 } 484 #endif // !SANITIZER_GO 485 486 #if !SANITIZER_GO 487 void ReplaceSystemMalloc() { } 488 #endif 489 490 #if !SANITIZER_GO 491 #if SANITIZER_ANDROID 492 // On Android, one thread can call intercepted functions after 493 // DestroyThreadState(), so add a fake thread state for "dead" threads. 494 static ThreadState *dead_thread_state = nullptr; 495 496 ThreadState *cur_thread() { 497 ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr()); 498 if (thr == nullptr) { 499 __sanitizer_sigset_t emptyset; 500 internal_sigfillset(&emptyset); 501 __sanitizer_sigset_t oldset; 502 CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset)); 503 thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr()); 504 if (thr == nullptr) { 505 thr = reinterpret_cast<ThreadState*>(MmapOrDie(sizeof(ThreadState), 506 "ThreadState")); 507 *get_android_tls_ptr() = reinterpret_cast<uptr>(thr); 508 if (dead_thread_state == nullptr) { 509 dead_thread_state = reinterpret_cast<ThreadState*>( 510 MmapOrDie(sizeof(ThreadState), "ThreadState")); 511 dead_thread_state->fast_state.SetIgnoreBit(); 512 dead_thread_state->ignore_interceptors = 1; 513 dead_thread_state->is_dead = true; 514 *const_cast<u32*>(&dead_thread_state->tid) = -1; 515 CHECK_EQ(0, internal_mprotect(dead_thread_state, sizeof(ThreadState), 516 PROT_READ)); 517 } 518 } 519 CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr)); 520 } 521 return thr; 522 } 523 524 void set_cur_thread(ThreadState *thr) { 525 *get_android_tls_ptr() = reinterpret_cast<uptr>(thr); 526 } 527 528 void cur_thread_finalize() { 529 __sanitizer_sigset_t emptyset; 530 internal_sigfillset(&emptyset); 531 __sanitizer_sigset_t oldset; 532 CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset)); 533 ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr()); 534 if (thr != dead_thread_state) { 535 *get_android_tls_ptr() = reinterpret_cast<uptr>(dead_thread_state); 536 UnmapOrDie(thr, sizeof(ThreadState)); 537 } 538 CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr)); 539 } 540 #endif // SANITIZER_ANDROID 541 #endif // if !SANITIZER_GO 542 543 } // namespace __tsan 544 545 #endif // SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD 546