libsanitizer/tsan/tsan_platform_linux.cc

   1 //===-- tsan_platform_linux.cc --------------------------------------------===//
   2 //
   3 // This file is distributed under the University of Illinois Open Source
   4 // License. See LICENSE.TXT for details.
   5 //
   6 //===----------------------------------------------------------------------===//
   7 //
   8 // This file is a part of ThreadSanitizer (TSan), a race detector.
   9 //
  10 // Linux-specific code.
  11 //===----------------------------------------------------------------------===//
  12
  13
  14 #include "sanitizer_common/sanitizer_platform.h"
  15 #if SANITIZER_LINUX || SANITIZER_FREEBSD
  16
  17 #include "sanitizer_common/sanitizer_common.h"
  18 #include "sanitizer_common/sanitizer_libc.h"
  19 #include "sanitizer_common/sanitizer_procmaps.h"
  20 #include "sanitizer_common/sanitizer_stoptheworld.h"
  21 #include "tsan_platform.h"
  22 #include "tsan_rtl.h"
  23 #include "tsan_flags.h"
  24
  25 #include <fcntl.h>
  26 #include <pthread.h>
  27 #include <signal.h>
  28 #include <stdio.h>
  29 #include <stdlib.h>
  30 #include <string.h>
  31 #include <stdarg.h>
  32 #include <sys/mman.h>
  33 #include <sys/syscall.h>
  34 #include <sys/socket.h>
  35 #include <sys/time.h>
  36 #include <sys/types.h>
  37 #include <sys/resource.h>
  38 #include <sys/stat.h>
  39 #include <unistd.h>
  40 #include <errno.h>
  41 #include <sched.h>
  42 #include <dlfcn.h>
  43 #if SANITIZER_LINUX
  44 #define __need_res_state
  45 #include <resolv.h>
  46 #endif
  47
  48 #ifdef sa_handler
  49 # undef sa_handler
  50 #endif
  51
  52 #ifdef sa_sigaction
  53 # undef sa_sigaction
  54 #endif
  55
  56 #if SANITIZER_FREEBSD
  57 extern "C" void *__libc_stack_end;
  58 void *__libc_stack_end = 0;
  59 #endif
  60
  61 namespace __tsan {
  62
  63 const uptr kPageSize = 4096;
  64
  65 enum {
  66   MemTotal  = 0,
  67   MemShadow = 1,
  68   MemMeta   = 2,
  69   MemFile   = 3,
  70   MemMmap   = 4,
  71   MemTrace  = 5,
  72   MemHeap   = 6,
  73   MemOther  = 7,
  74   MemCount  = 8,
  75 };
  76
  77 void FillProfileCallback(uptr start, uptr rss, bool file,
  78                          uptr *mem, uptr stats_size) {
  79   mem[MemTotal] += rss;
  80   start >>= 40;
  81   if (start < 0x10)
  82     mem[MemShadow] += rss;
  83   else if (start >= 0x20 && start < 0x30)
  84     mem[file ? MemFile : MemMmap] += rss;
  85   else if (start >= 0x30 && start < 0x40)
  86     mem[MemMeta] += rss;
  87   else if (start >= 0x7e)
  88     mem[file ? MemFile : MemMmap] += rss;
  89   else if (start >= 0x60 && start < 0x62)
  90     mem[MemTrace] += rss;
  91   else if (start >= 0x7d && start < 0x7e)
  92     mem[MemHeap] += rss;
  93   else
  94     mem[MemOther] += rss;
  95 }
  96
  97 void WriteMemoryProfile(char *buf, uptr buf_size, uptr nthread, uptr nlive) {
  98   uptr mem[MemCount] = {};
  99   __sanitizer::GetMemoryProfile(FillProfileCallback, mem, 7);
 100   internal_snprintf(buf, buf_size,
 101       "RSS %zd MB: shadow:%zd meta:%zd file:%zd mmap:%zd"
 102       " trace:%zd heap:%zd other:%zd nthr=%zd/%zd\n",
 103       mem[MemTotal] >> 20, mem[MemShadow] >> 20, mem[MemMeta] >> 20,
 104       mem[MemFile] >> 20, mem[MemMmap] >> 20, mem[MemTrace] >> 20,
 105       mem[MemHeap] >> 20, mem[MemOther] >> 20,
 106       nlive, nthread);
 107 }
 108
 109 uptr GetRSS() {
 110   uptr fd = OpenFile("/proc/self/statm", false);
 111   if ((sptr)fd < 0)
 112     return 0;
 113   char buf[64];
 114   uptr len = internal_read(fd, buf, sizeof(buf) - 1);
 115   internal_close(fd);
 116   if ((sptr)len <= 0)
 117     return 0;
 118   buf[len] = 0;
 119   // The format of the file is:
 120   // 1084 89 69 11 0 79 0
 121   // We need the second number which is RSS in 4K units.
 122   char *pos = buf;
 123   // Skip the first number.
 124   while (*pos >= '0' && *pos <= '9')
 125     pos++;
 126   // Skip whitespaces.
 127   while (!(*pos >= '0' && *pos <= '9') && *pos != 0)
 128     pos++;
 129   // Read the number.
 130   uptr rss = 0;
 131   while (*pos >= '0' && *pos <= '9')
 132     rss = rss * 10 + *pos++ - '0';
 133   return rss * 4096;
 134 }
 135
 136 #if SANITIZER_LINUX
 137 void FlushShadowMemoryCallback(
 138     const SuspendedThreadsList &suspended_threads_list,
 139     void *argument) {
 140   FlushUnneededShadowMemory(kLinuxShadowBeg, kLinuxShadowEnd - kLinuxShadowBeg);
 141 }
 142 #endif
 143
 144 void FlushShadowMemory() {
 145 #if SANITIZER_LINUX
 146   StopTheWorld(FlushShadowMemoryCallback, 0);
 147 #endif
 148 }
 149
 150 #ifndef TSAN_GO
 151 static void ProtectRange(uptr beg, uptr end) {
 152   CHECK_LE(beg, end);
 153   if (beg == end)
 154     return;
 155   if (beg != (uptr)Mprotect(beg, end - beg)) {
 156     Printf("FATAL: ThreadSanitizer can not protect [%zx,%zx]\n", beg, end);
 157     Printf("FATAL: Make sure you are not using unlimited stack\n");
 158     Die();
 159   }
 160 }
 161
 162 // Mark shadow for .rodata sections with the special kShadowRodata marker.
 163 // Accesses to .rodata can't race, so this saves time, memory and trace space.
 164 static void MapRodata() {
 165   // First create temp file.
 166   const char *tmpdir = GetEnv("TMPDIR");
 167   if (tmpdir == 0)
 168     tmpdir = GetEnv("TEST_TMPDIR");
 169 #ifdef P_tmpdir
 170   if (tmpdir == 0)
 171     tmpdir = P_tmpdir;
 172 #endif
 173   if (tmpdir == 0)
 174     return;
 175   char name[256];
 176   internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d",
 177                     tmpdir, (int)internal_getpid());
 178   uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600);
 179   if (internal_iserror(openrv))
 180     return;
 181   internal_unlink(name);  // Unlink it now, so that we can reuse the buffer.
 182   fd_t fd = openrv;
 183   // Fill the file with kShadowRodata.
 184   const uptr kMarkerSize = 512 * 1024 / sizeof(u64);
 185   InternalScopedBuffer<u64> marker(kMarkerSize);
 186   // volatile to prevent insertion of memset
 187   for (volatile u64 *p = marker.data(); p < marker.data() + kMarkerSize; p++)
 188     *p = kShadowRodata;
 189   internal_write(fd, marker.data(), marker.size());
 190   // Map the file into memory.
 191   uptr page = internal_mmap(0, kPageSize, PROT_READ | PROT_WRITE,
 192                             MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
 193   if (internal_iserror(page)) {
 194     internal_close(fd);
 195     return;
 196   }
 197   // Map the file into shadow of .rodata sections.
 198   MemoryMappingLayout proc_maps(/*cache_enabled*/true);
 199   uptr start, end, offset, prot;
 200   // Reusing the buffer 'name'.
 201   while (proc_maps.Next(&start, &end, &offset, name, ARRAY_SIZE(name), &prot)) {
 202     if (name[0] != 0 && name[0] != '['
 203         && (prot & MemoryMappingLayout::kProtectionRead)
 204         && (prot & MemoryMappingLayout::kProtectionExecute)
 205         && !(prot & MemoryMappingLayout::kProtectionWrite)
 206         && IsAppMem(start)) {
 207       // Assume it's .rodata
 208       char *shadow_start = (char*)MemToShadow(start);
 209       char *shadow_end = (char*)MemToShadow(end);
 210       for (char *p = shadow_start; p < shadow_end; p += marker.size()) {
 211         internal_mmap(p, Min<uptr>(marker.size(), shadow_end - p),
 212                       PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
 213       }
 214     }
 215   }
 216   internal_close(fd);
 217 }
 218
 219 void InitializeShadowMemory() {
 220   // Map memory shadow.
 221   uptr shadow = (uptr)MmapFixedNoReserve(kLinuxShadowBeg,
 222     kLinuxShadowEnd - kLinuxShadowBeg);
 223   if (shadow != kLinuxShadowBeg) {
 224     Printf("FATAL: ThreadSanitizer can not mmap the shadow memory\n");
 225     Printf("FATAL: Make sure to compile with -fPIE and "
 226                "to link with -pie (%p, %p).\n", shadow, kLinuxShadowBeg);
 227     Die();
 228   }
 229   // This memory range is used for thread stacks and large user mmaps.
 230   // Frequently a thread uses only a small part of stack and similarly
 231   // a program uses a small part of large mmap. On some programs
 232   // we see 20% memory usage reduction without huge pages for this range.
 233 #ifdef MADV_NOHUGEPAGE
 234   madvise((void*)MemToShadow(0x7f0000000000ULL),
 235       0x10000000000ULL * kShadowMultiplier, MADV_NOHUGEPAGE);
 236 #endif
 237   DPrintf("memory shadow: %zx-%zx (%zuGB)\n",
 238       kLinuxShadowBeg, kLinuxShadowEnd,
 239       (kLinuxShadowEnd - kLinuxShadowBeg) >> 30);
 240
 241   // Map meta shadow.
 242   if (MemToMeta(kLinuxAppMemBeg) < (u32*)kMetaShadow) {
 243     Printf("ThreadSanitizer: bad meta shadow (%p -> %p < %p)\n",
 244         kLinuxAppMemBeg, MemToMeta(kLinuxAppMemBeg), kMetaShadow);
 245     Die();
 246   }
 247   if (MemToMeta(kLinuxAppMemEnd) >= (u32*)(kMetaShadow + kMetaSize)) {
 248     Printf("ThreadSanitizer: bad meta shadow (%p -> %p >= %p)\n",
 249         kLinuxAppMemEnd, MemToMeta(kLinuxAppMemEnd), kMetaShadow + kMetaSize);
 250     Die();
 251   }
 252   uptr meta = (uptr)MmapFixedNoReserve(kMetaShadow, kMetaSize);
 253   if (meta != kMetaShadow) {
 254     Printf("FATAL: ThreadSanitizer can not mmap the shadow memory\n");
 255     Printf("FATAL: Make sure to compile with -fPIE and "
 256                "to link with -pie (%p, %p).\n", meta, kMetaShadow);
 257     Die();
 258   }
 259   DPrintf("meta shadow: %zx-%zx (%zuGB)\n",
 260       kMetaShadow, kMetaShadow + kMetaSize, kMetaSize >> 30);
 261
 262   // Protect gaps.
 263   const uptr kClosedLowBeg  = 0x200000;
 264   const uptr kClosedLowEnd  = kLinuxShadowBeg - 1;
 265   const uptr kClosedMidBeg = kLinuxShadowEnd + 1;
 266   const uptr kClosedMidEnd = min(min(kLinuxAppMemBeg, kTraceMemBegin),
 267       kMetaShadow);
 268
 269   ProtectRange(kClosedLowBeg, kClosedLowEnd);
 270   ProtectRange(kClosedMidBeg, kClosedMidEnd);
 271   VPrintf(2, "kClosedLow   %zx-%zx (%zuGB)\n",
 272       kClosedLowBeg, kClosedLowEnd, (kClosedLowEnd - kClosedLowBeg) >> 30);
 273   VPrintf(2, "kClosedMid   %zx-%zx (%zuGB)\n",
 274       kClosedMidBeg, kClosedMidEnd, (kClosedMidEnd - kClosedMidBeg) >> 30);
 275   VPrintf(2, "app mem: %zx-%zx (%zuGB)\n",
 276       kLinuxAppMemBeg, kLinuxAppMemEnd,
 277       (kLinuxAppMemEnd - kLinuxAppMemBeg) >> 30);
 278   VPrintf(2, "stack: %zx\n", (uptr)&shadow);
 279
 280   MapRodata();
 281 }
 282 #endif
 283
 284 static uptr g_data_start;
 285 static uptr g_data_end;
 286
 287 #ifndef TSAN_GO
 288 static void CheckPIE() {
 289   // Ensure that the binary is indeed compiled with -pie.
 290   MemoryMappingLayout proc_maps(true);
 291   uptr start, end;
 292   if (proc_maps.Next(&start, &end,
 293                      /*offset*/0, /*filename*/0, /*filename_size*/0,
 294                      /*protection*/0)) {
 295     if ((u64)start < kLinuxAppMemBeg) {
 296       Printf("FATAL: ThreadSanitizer can not mmap the shadow memory ("
 297              "something is mapped at 0x%zx < 0x%zx)\n",
 298              start, kLinuxAppMemBeg);
 299       Printf("FATAL: Make sure to compile with -fPIE"
 300              " and to link with -pie.\n");
 301       Die();
 302     }
 303   }
 304 }
 305
 306 static void InitDataSeg() {
 307   MemoryMappingLayout proc_maps(true);
 308   uptr start, end, offset;
 309   char name[128];
 310   bool prev_is_data = false;
 311   while (proc_maps.Next(&start, &end, &offset, name, ARRAY_SIZE(name),
 312                         /*protection*/ 0)) {
 313     DPrintf("%p-%p %p %s\n", start, end, offset, name);
 314     bool is_data = offset != 0 && name[0] != 0;
 315     // BSS may get merged with [heap] in /proc/self/maps. This is not very
 316     // reliable.
 317     bool is_bss = offset == 0 &&
 318       (name[0] == 0 || internal_strcmp(name, "[heap]") == 0) && prev_is_data;
 319     if (g_data_start == 0 && is_data)
 320       g_data_start = start;
 321     if (is_bss)
 322       g_data_end = end;
 323     prev_is_data = is_data;
 324   }
 325   DPrintf("guessed data_start=%p data_end=%p\n",  g_data_start, g_data_end);
 326   CHECK_LT(g_data_start, g_data_end);
 327   CHECK_GE((uptr)&g_data_start, g_data_start);
 328   CHECK_LT((uptr)&g_data_start, g_data_end);
 329 }
 330
 331 #endif  // #ifndef TSAN_GO
 332
 333 void InitializePlatform() {
 334   DisableCoreDumperIfNecessary();
 335
 336   // Go maps shadow memory lazily and works fine with limited address space.
 337   // Unlimited stack is not a problem as well, because the executable
 338   // is not compiled with -pie.
 339   if (kCppMode) {
 340     bool reexec = false;
 341     // TSan doesn't play well with unlimited stack size (as stack
 342     // overlaps with shadow memory). If we detect unlimited stack size,
 343     // we re-exec the program with limited stack size as a best effort.
 344     if (StackSizeIsUnlimited()) {
 345       const uptr kMaxStackSize = 32 * 1024 * 1024;
 346       VReport(1, "Program is run with unlimited stack size, which wouldn't "
 347                  "work with ThreadSanitizer.\n"
 348                  "Re-execing with stack size limited to %zd bytes.\n",
 349               kMaxStackSize);
 350       SetStackSizeLimitInBytes(kMaxStackSize);
 351       reexec = true;
 352     }
 353
 354     if (!AddressSpaceIsUnlimited()) {
 355       Report("WARNING: Program is run with limited virtual address space,"
 356              " which wouldn't work with ThreadSanitizer.\n");
 357       Report("Re-execing with unlimited virtual address space.\n");
 358       SetAddressSpaceUnlimited();
 359       reexec = true;
 360     }
 361     if (reexec)
 362       ReExec();
 363   }
 364
 365 #ifndef TSAN_GO
 366   CheckPIE();
 367   InitTlsSize();
 368   InitDataSeg();
 369 #endif
 370 }
 371
 372 bool IsGlobalVar(uptr addr) {
 373   return g_data_start && addr >= g_data_start && addr < g_data_end;
 374 }
 375
 376 #ifndef TSAN_GO
 377 // Extract file descriptors passed to glibc internal __res_iclose function.
 378 // This is required to properly "close" the fds, because we do not see internal
 379 // closes within glibc. The code is a pure hack.
 380 int ExtractResolvFDs(void *state, int *fds, int nfd) {
 381 #if SANITIZER_LINUX
 382   int cnt = 0;
 383   __res_state *statp = (__res_state*)state;
 384   for (int i = 0; i < MAXNS && cnt < nfd; i++) {
 385     if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1)
 386       fds[cnt++] = statp->_u._ext.nssocks[i];
 387   }
 388   return cnt;
 389 #else
 390   return 0;
 391 #endif
 392 }
 393
 394 // Extract file descriptors passed via UNIX domain sockets.
 395 // This is requried to properly handle "open" of these fds.
 396 // see 'man recvmsg' and 'man 3 cmsg'.
 397 int ExtractRecvmsgFDs(void *msgp, int *fds, int nfd) {
 398   int res = 0;
 399   msghdr *msg = (msghdr*)msgp;
 400   struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg);
 401   for (; cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 402     if (cmsg->cmsg_level != SOL_SOCKET || cmsg->cmsg_type != SCM_RIGHTS)
 403       continue;
 404     int n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(fds[0]);
 405     for (int i = 0; i < n; i++) {
 406       fds[res++] = ((int*)CMSG_DATA(cmsg))[i];
 407       if (res == nfd)
 408         return res;
 409     }
 410   }
 411   return res;
 412 }
 413
 414 int call_pthread_cancel_with_cleanup(int(*fn)(void *c, void *m,
 415     void *abstime), void *c, void *m, void *abstime,
 416     void(*cleanup)(void *arg), void *arg) {
 417   // pthread_cleanup_push/pop are hardcore macros mess.
 418   // We can't intercept nor call them w/o including pthread.h.
 419   int res;
 420   pthread_cleanup_push(cleanup, arg);
 421   res = fn(c, m, abstime);
 422   pthread_cleanup_pop(0);
 423   return res;
 424 }
 425 #endif
 426
 427 }  // namespace __tsan
 428
 429 #endif  // SANITIZER_LINUX