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 FreeBSD-specific code.
12 //===----------------------------------------------------------------------===//
13 
14 
15 #include "sanitizer_common/sanitizer_platform.h"
16 #if SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD
17 
18 #include "sanitizer_common/sanitizer_common.h"
19 #include "sanitizer_common/sanitizer_libc.h"
20 #include "sanitizer_common/sanitizer_linux.h"
21 #include "sanitizer_common/sanitizer_platform_limits_netbsd.h"
22 #include "sanitizer_common/sanitizer_platform_limits_posix.h"
23 #include "sanitizer_common/sanitizer_posix.h"
24 #include "sanitizer_common/sanitizer_procmaps.h"
25 #include "sanitizer_common/sanitizer_stoptheworld.h"
26 #include "sanitizer_common/sanitizer_stackdepot.h"
27 #include "tsan_platform.h"
28 #include "tsan_rtl.h"
29 #include "tsan_flags.h"
30 
31 #include <fcntl.h>
32 #include <pthread.h>
33 #include <signal.h>
34 #include <stdio.h>
35 #include <stdlib.h>
36 #include <string.h>
37 #include <stdarg.h>
38 #include <sys/mman.h>
39 #if SANITIZER_LINUX
40 #include <sys/personality.h>
41 #include <setjmp.h>
42 #endif
43 #include <sys/syscall.h>
44 #include <sys/socket.h>
45 #include <sys/time.h>
46 #include <sys/types.h>
47 #include <sys/resource.h>
48 #include <sys/stat.h>
49 #include <unistd.h>
50 #include <sched.h>
51 #include <dlfcn.h>
52 #if SANITIZER_LINUX
53 #define __need_res_state
54 #include <resolv.h>
55 #endif
56 
57 #ifdef sa_handler
58 # undef sa_handler
59 #endif
60 
61 #ifdef sa_sigaction
62 # undef sa_sigaction
63 #endif
64 
65 #if SANITIZER_FREEBSD
66 extern "C" void *__libc_stack_end;
67 void *__libc_stack_end = 0;
68 #endif
69 
70 #if SANITIZER_LINUX && defined(__aarch64__) && !SANITIZER_GO
71 # define INIT_LONGJMP_XOR_KEY 1
72 #else
73 # define INIT_LONGJMP_XOR_KEY 0
74 #endif
75 
76 #if INIT_LONGJMP_XOR_KEY
77 #include "interception/interception.h"
78 // Must be declared outside of other namespaces.
79 DECLARE_REAL(int, _setjmp, void *env)
80 #endif
81 
82 namespace __tsan {
83 
84 #if INIT_LONGJMP_XOR_KEY
85 static void InitializeLongjmpXorKey();
86 static uptr longjmp_xor_key;
87 #endif
88 
89 #ifdef TSAN_RUNTIME_VMA
90 // Runtime detected VMA size.
91 uptr vmaSize;
92 #endif
93 
94 enum {
95   MemTotal  = 0,
96   MemShadow = 1,
97   MemMeta   = 2,
98   MemFile   = 3,
99   MemMmap   = 4,
100   MemTrace  = 5,
101   MemHeap   = 6,
102   MemOther  = 7,
103   MemCount  = 8,
104 };
105 
106 void FillProfileCallback(uptr p, uptr rss, bool file,
107                          uptr *mem, uptr stats_size) {
108   mem[MemTotal] += rss;
109   if (p >= ShadowBeg() && p < ShadowEnd())
110     mem[MemShadow] += rss;
111   else if (p >= MetaShadowBeg() && p < MetaShadowEnd())
112     mem[MemMeta] += rss;
113 #if !SANITIZER_GO
114   else if (p >= HeapMemBeg() && p < HeapMemEnd())
115     mem[MemHeap] += rss;
116   else if (p >= LoAppMemBeg() && p < LoAppMemEnd())
117     mem[file ? MemFile : MemMmap] += rss;
118   else if (p >= HiAppMemBeg() && p < HiAppMemEnd())
119     mem[file ? MemFile : MemMmap] += rss;
120 #else
121   else if (p >= AppMemBeg() && p < AppMemEnd())
122     mem[file ? MemFile : MemMmap] += rss;
123 #endif
124   else if (p >= TraceMemBeg() && p < TraceMemEnd())
125     mem[MemTrace] += rss;
126   else
127     mem[MemOther] += rss;
128 }
129 
130 void WriteMemoryProfile(char *buf, uptr buf_size, uptr nthread, uptr nlive) {
131   uptr mem[MemCount];
132   internal_memset(mem, 0, sizeof(mem[0]) * MemCount);
133   __sanitizer::GetMemoryProfile(FillProfileCallback, mem, 7);
134   StackDepotStats *stacks = StackDepotGetStats();
135   internal_snprintf(buf, buf_size,
136       "RSS %zd MB: shadow:%zd meta:%zd file:%zd mmap:%zd"
137       " trace:%zd heap:%zd other:%zd stacks=%zd[%zd] nthr=%zd/%zd\n",
138       mem[MemTotal] >> 20, mem[MemShadow] >> 20, mem[MemMeta] >> 20,
139       mem[MemFile] >> 20, mem[MemMmap] >> 20, mem[MemTrace] >> 20,
140       mem[MemHeap] >> 20, mem[MemOther] >> 20,
141       stacks->allocated >> 20, stacks->n_uniq_ids,
142       nlive, nthread);
143 }
144 
145 #if SANITIZER_LINUX
146 void FlushShadowMemoryCallback(
147     const SuspendedThreadsList &suspended_threads_list,
148     void *argument) {
149   ReleaseMemoryPagesToOS(ShadowBeg(), ShadowEnd());
150 }
151 #endif
152 
153 void FlushShadowMemory() {
154 #if SANITIZER_LINUX
155   StopTheWorld(FlushShadowMemoryCallback, 0);
156 #endif
157 }
158 
159 #if !SANITIZER_GO
160 // Mark shadow for .rodata sections with the special kShadowRodata marker.
161 // Accesses to .rodata can't race, so this saves time, memory and trace space.
162 static void MapRodata() {
163   // First create temp file.
164   const char *tmpdir = GetEnv("TMPDIR");
165   if (tmpdir == 0)
166     tmpdir = GetEnv("TEST_TMPDIR");
167 #ifdef P_tmpdir
168   if (tmpdir == 0)
169     tmpdir = P_tmpdir;
170 #endif
171   if (tmpdir == 0)
172     return;
173   char name[256];
174   internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d",
175                     tmpdir, (int)internal_getpid());
176   uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600);
177   if (internal_iserror(openrv))
178     return;
179   internal_unlink(name);  // Unlink it now, so that we can reuse the buffer.
180   fd_t fd = openrv;
181   // Fill the file with kShadowRodata.
182   const uptr kMarkerSize = 512 * 1024 / sizeof(u64);
183   InternalMmapVector<u64> marker(kMarkerSize);
184   // volatile to prevent insertion of memset
185   for (volatile u64 *p = marker.data(); p < marker.data() + kMarkerSize; p++)
186     *p = kShadowRodata;
187   internal_write(fd, marker.data(), marker.size() * sizeof(u64));
188   // Map the file into memory.
189   uptr page = internal_mmap(0, GetPageSizeCached(), PROT_READ | PROT_WRITE,
190                             MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
191   if (internal_iserror(page)) {
192     internal_close(fd);
193     return;
194   }
195   // Map the file into shadow of .rodata sections.
196   MemoryMappingLayout proc_maps(/*cache_enabled*/true);
197   // Reusing the buffer 'name'.
198   MemoryMappedSegment segment(name, ARRAY_SIZE(name));
199   while (proc_maps.Next(&segment)) {
200     if (segment.filename[0] != 0 && segment.filename[0] != '[' &&
201         segment.IsReadable() && segment.IsExecutable() &&
202         !segment.IsWritable() && IsAppMem(segment.start)) {
203       // Assume it's .rodata
204       char *shadow_start = (char *)MemToShadow(segment.start);
205       char *shadow_end = (char *)MemToShadow(segment.end);
206       for (char *p = shadow_start; p < shadow_end;
207            p += marker.size() * sizeof(u64)) {
208         internal_mmap(p, Min<uptr>(marker.size() * sizeof(u64), shadow_end - p),
209                       PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
210       }
211     }
212   }
213   internal_close(fd);
214 }
215 
216 void InitializeShadowMemoryPlatform() {
217   MapRodata();
218 }
219 
220 #endif  // #if !SANITIZER_GO
221 
222 void InitializePlatformEarly() {
223 #ifdef TSAN_RUNTIME_VMA
224   vmaSize =
225     (MostSignificantSetBitIndex(GET_CURRENT_FRAME()) + 1);
226 #if defined(__aarch64__)
227 # if !SANITIZER_GO
228   if (vmaSize != 39 && vmaSize != 42 && vmaSize != 48) {
229     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
230     Printf("FATAL: Found %zd - Supported 39, 42 and 48\n", vmaSize);
231     Die();
232   }
233 #else
234   if (vmaSize != 48) {
235     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
236     Printf("FATAL: Found %zd - Supported 48\n", vmaSize);
237     Die();
238   }
239 #endif
240 #elif defined(__powerpc64__)
241 # if !SANITIZER_GO
242   if (vmaSize != 44 && vmaSize != 46 && vmaSize != 47) {
243     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
244     Printf("FATAL: Found %zd - Supported 44, 46, and 47\n", vmaSize);
245     Die();
246   }
247 # else
248   if (vmaSize != 46 && vmaSize != 47) {
249     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
250     Printf("FATAL: Found %zd - Supported 46, and 47\n", vmaSize);
251     Die();
252   }
253 # endif
254 #endif
255 #endif
256 }
257 
258 void InitializePlatform() {
259   DisableCoreDumperIfNecessary();
260 
261   // Go maps shadow memory lazily and works fine with limited address space.
262   // Unlimited stack is not a problem as well, because the executable
263   // is not compiled with -pie.
264 #if !SANITIZER_GO
265   {
266     bool reexec = false;
267     // TSan doesn't play well with unlimited stack size (as stack
268     // overlaps with shadow memory). If we detect unlimited stack size,
269     // we re-exec the program with limited stack size as a best effort.
270     if (StackSizeIsUnlimited()) {
271       const uptr kMaxStackSize = 32 * 1024 * 1024;
272       VReport(1, "Program is run with unlimited stack size, which wouldn't "
273                  "work with ThreadSanitizer.\n"
274                  "Re-execing with stack size limited to %zd bytes.\n",
275               kMaxStackSize);
276       SetStackSizeLimitInBytes(kMaxStackSize);
277       reexec = true;
278     }
279 
280     if (!AddressSpaceIsUnlimited()) {
281       Report("WARNING: Program is run with limited virtual address space,"
282              " which wouldn't work with ThreadSanitizer.\n");
283       Report("Re-execing with unlimited virtual address space.\n");
284       SetAddressSpaceUnlimited();
285       reexec = true;
286     }
287 #if SANITIZER_LINUX && defined(__aarch64__)
288     // After patch "arm64: mm: support ARCH_MMAP_RND_BITS." is introduced in
289     // linux kernel, the random gap between stack and mapped area is increased
290     // from 128M to 36G on 39-bit aarch64. As it is almost impossible to cover
291     // this big range, we should disable randomized virtual space on aarch64.
292     int old_personality = personality(0xffffffff);
293     if (old_personality != -1 && (old_personality & ADDR_NO_RANDOMIZE) == 0) {
294       VReport(1, "WARNING: Program is run with randomized virtual address "
295               "space, which wouldn't work with ThreadSanitizer.\n"
296               "Re-execing with fixed virtual address space.\n");
297       CHECK_NE(personality(old_personality | ADDR_NO_RANDOMIZE), -1);
298       reexec = true;
299     }
300     // Initialize the xor key used in {sig}{set,long}jump.
301     InitializeLongjmpXorKey();
302 #endif
303     if (reexec)
304       ReExec();
305   }
306 
307   CheckAndProtect();
308   InitTlsSize();
309 #endif  // !SANITIZER_GO
310 }
311 
312 #if !SANITIZER_GO
313 // Extract file descriptors passed to glibc internal __res_iclose function.
314 // This is required to properly "close" the fds, because we do not see internal
315 // closes within glibc. The code is a pure hack.
316 int ExtractResolvFDs(void *state, int *fds, int nfd) {
317 #if SANITIZER_LINUX && !SANITIZER_ANDROID
318   int cnt = 0;
319   struct __res_state *statp = (struct __res_state*)state;
320   for (int i = 0; i < MAXNS && cnt < nfd; i++) {
321     if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1)
322       fds[cnt++] = statp->_u._ext.nssocks[i];
323   }
324   return cnt;
325 #else
326   return 0;
327 #endif
328 }
329 
330 // Extract file descriptors passed via UNIX domain sockets.
331 // This is requried to properly handle "open" of these fds.
332 // see 'man recvmsg' and 'man 3 cmsg'.
333 int ExtractRecvmsgFDs(void *msgp, int *fds, int nfd) {
334   int res = 0;
335   msghdr *msg = (msghdr*)msgp;
336   struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg);
337   for (; cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
338     if (cmsg->cmsg_level != SOL_SOCKET || cmsg->cmsg_type != SCM_RIGHTS)
339       continue;
340     int n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(fds[0]);
341     for (int i = 0; i < n; i++) {
342       fds[res++] = ((int*)CMSG_DATA(cmsg))[i];
343       if (res == nfd)
344         return res;
345     }
346   }
347   return res;
348 }
349 
350 // Reverse operation of libc stack pointer mangling
351 static uptr UnmangleLongJmpSp(uptr mangled_sp) {
352 #if defined(__x86_64__)
353 # if SANITIZER_LINUX
354   // Reverse of:
355   //   xor  %fs:0x30, %rsi
356   //   rol  $0x11, %rsi
357   uptr sp;
358   asm("ror  $0x11,     %0 \n"
359       "xor  %%fs:0x30, %0 \n"
360       : "=r" (sp)
361       : "0" (mangled_sp));
362   return sp;
363 # else
364   return mangled_sp;
365 # endif
366 #elif defined(__aarch64__)
367 # if SANITIZER_LINUX
368   return mangled_sp ^ longjmp_xor_key;
369 # else
370   return mangled_sp;
371 # endif
372 #elif defined(__powerpc64__)
373   // Reverse of:
374   //   ld   r4, -28696(r13)
375   //   xor  r4, r3, r4
376   uptr xor_key;
377   asm("ld  %0, -28696(%%r13)" : "=r" (xor_key));
378   return mangled_sp ^ xor_key;
379 #elif defined(__mips__)
380   return mangled_sp;
381 #else
382   #error "Unknown platform"
383 #endif
384 }
385 
386 #ifdef __powerpc__
387 # define LONG_JMP_SP_ENV_SLOT 0
388 #elif SANITIZER_FREEBSD
389 # define LONG_JMP_SP_ENV_SLOT 2
390 #elif SANITIZER_NETBSD
391 # define LONG_JMP_SP_ENV_SLOT 6
392 #elif SANITIZER_LINUX
393 # ifdef __aarch64__
394 #  define LONG_JMP_SP_ENV_SLOT 13
395 # elif defined(__mips64)
396 #  define LONG_JMP_SP_ENV_SLOT 1
397 # else
398 #  define LONG_JMP_SP_ENV_SLOT 6
399 # endif
400 #endif
401 
402 uptr ExtractLongJmpSp(uptr *env) {
403   uptr mangled_sp = env[LONG_JMP_SP_ENV_SLOT];
404   return UnmangleLongJmpSp(mangled_sp);
405 }
406 
407 #if INIT_LONGJMP_XOR_KEY
408 // GLIBC mangles the function pointers in jmp_buf (used in {set,long}*jmp
409 // functions) by XORing them with a random key.  For AArch64 it is a global
410 // variable rather than a TCB one (as for x86_64/powerpc).  We obtain the key by
411 // issuing a setjmp and XORing the SP pointer values to derive the key.
412 static void InitializeLongjmpXorKey() {
413   // 1. Call REAL(setjmp), which stores the mangled SP in env.
414   jmp_buf env;
415   REAL(_setjmp)(env);
416 
417   // 2. Retrieve vanilla/mangled SP.
418   uptr sp;
419   asm("mov  %0, sp" : "=r" (sp));
420   uptr mangled_sp = ((uptr *)&env)[LONG_JMP_SP_ENV_SLOT];
421 
422   // 3. xor SPs to obtain key.
423   longjmp_xor_key = mangled_sp ^ sp;
424 }
425 #endif
426 
427 void ImitateTlsWrite(ThreadState *thr, uptr tls_addr, uptr tls_size) {
428   // Check that the thr object is in tls;
429   const uptr thr_beg = (uptr)thr;
430   const uptr thr_end = (uptr)thr + sizeof(*thr);
431   CHECK_GE(thr_beg, tls_addr);
432   CHECK_LE(thr_beg, tls_addr + tls_size);
433   CHECK_GE(thr_end, tls_addr);
434   CHECK_LE(thr_end, tls_addr + tls_size);
435   // Since the thr object is huge, skip it.
436   MemoryRangeImitateWrite(thr, /*pc=*/2, tls_addr, thr_beg - tls_addr);
437   MemoryRangeImitateWrite(thr, /*pc=*/2, thr_end,
438                           tls_addr + tls_size - thr_end);
439 }
440 
441 // Note: this function runs with async signals enabled,
442 // so it must not touch any tsan state.
443 int call_pthread_cancel_with_cleanup(int(*fn)(void *c, void *m,
444     void *abstime), void *c, void *m, void *abstime,
445     void(*cleanup)(void *arg), void *arg) {
446   // pthread_cleanup_push/pop are hardcore macros mess.
447   // We can't intercept nor call them w/o including pthread.h.
448   int res;
449   pthread_cleanup_push(cleanup, arg);
450   res = fn(c, m, abstime);
451   pthread_cleanup_pop(0);
452   return res;
453 }
454 #endif  // !SANITIZER_GO
455 
456 #if !SANITIZER_GO
457 void ReplaceSystemMalloc() { }
458 #endif
459 
460 #if !SANITIZER_GO
461 #if SANITIZER_ANDROID
462 // On Android, one thread can call intercepted functions after
463 // DestroyThreadState(), so add a fake thread state for "dead" threads.
464 static ThreadState *dead_thread_state = nullptr;
465 
466 ThreadState *cur_thread() {
467   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
468   if (thr == nullptr) {
469     __sanitizer_sigset_t emptyset;
470     internal_sigfillset(&emptyset);
471     __sanitizer_sigset_t oldset;
472     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
473     thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
474     if (thr == nullptr) {
475       thr = reinterpret_cast<ThreadState*>(MmapOrDie(sizeof(ThreadState),
476                                                      "ThreadState"));
477       *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
478       if (dead_thread_state == nullptr) {
479         dead_thread_state = reinterpret_cast<ThreadState*>(
480             MmapOrDie(sizeof(ThreadState), "ThreadState"));
481         dead_thread_state->fast_state.SetIgnoreBit();
482         dead_thread_state->ignore_interceptors = 1;
483         dead_thread_state->is_dead = true;
484         *const_cast<int*>(&dead_thread_state->tid) = -1;
485         CHECK_EQ(0, internal_mprotect(dead_thread_state, sizeof(ThreadState),
486                                       PROT_READ));
487       }
488     }
489     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
490   }
491   return thr;
492 }
493 
494 void set_cur_thread(ThreadState *thr) {
495   *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
496 }
497 
498 void cur_thread_finalize() {
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   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
504   if (thr != dead_thread_state) {
505     *get_android_tls_ptr() = reinterpret_cast<uptr>(dead_thread_state);
506     UnmapOrDie(thr, sizeof(ThreadState));
507   }
508   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
509 }
510 #endif  // SANITIZER_ANDROID
511 #endif  // if !SANITIZER_GO
512 
513 }  // namespace __tsan
514 
515 #endif  // SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD
516