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 #ifdef TSAN_RUNTIME_VMA
89 // Runtime detected VMA size.
90 uptr vmaSize;
91 #endif
92 
93 enum {
94   MemTotal  = 0,
95   MemShadow = 1,
96   MemMeta   = 2,
97   MemFile   = 3,
98   MemMmap   = 4,
99   MemTrace  = 5,
100   MemHeap   = 6,
101   MemOther  = 7,
102   MemCount  = 8,
103 };
104 
FillProfileCallback(uptr p,uptr rss,bool file,uptr * mem,uptr stats_size)105 void FillProfileCallback(uptr p, uptr rss, bool file,
106                          uptr *mem, uptr stats_size) {
107   mem[MemTotal] += rss;
108   if (p >= ShadowBeg() && p < ShadowEnd())
109     mem[MemShadow] += rss;
110   else if (p >= MetaShadowBeg() && p < MetaShadowEnd())
111     mem[MemMeta] += rss;
112 #if !SANITIZER_GO
113   else if (p >= HeapMemBeg() && p < HeapMemEnd())
114     mem[MemHeap] += rss;
115   else if (p >= LoAppMemBeg() && p < LoAppMemEnd())
116     mem[file ? MemFile : MemMmap] += rss;
117   else if (p >= HiAppMemBeg() && p < HiAppMemEnd())
118     mem[file ? MemFile : MemMmap] += rss;
119 #else
120   else if (p >= AppMemBeg() && p < AppMemEnd())
121     mem[file ? MemFile : MemMmap] += rss;
122 #endif
123   else if (p >= TraceMemBeg() && p < TraceMemEnd())
124     mem[MemTrace] += rss;
125   else
126     mem[MemOther] += rss;
127 }
128 
WriteMemoryProfile(char * buf,uptr buf_size,uptr nthread,uptr nlive)129 void WriteMemoryProfile(char *buf, uptr buf_size, uptr nthread, uptr nlive) {
130   uptr mem[MemCount];
131   internal_memset(mem, 0, sizeof(mem[0]) * MemCount);
132   __sanitizer::GetMemoryProfile(FillProfileCallback, mem, 7);
133   StackDepotStats *stacks = StackDepotGetStats();
134   internal_snprintf(buf, buf_size,
135       "RSS %zd MB: shadow:%zd meta:%zd file:%zd mmap:%zd"
136       " trace:%zd heap:%zd other:%zd stacks=%zd[%zd] nthr=%zd/%zd\n",
137       mem[MemTotal] >> 20, mem[MemShadow] >> 20, mem[MemMeta] >> 20,
138       mem[MemFile] >> 20, mem[MemMmap] >> 20, mem[MemTrace] >> 20,
139       mem[MemHeap] >> 20, mem[MemOther] >> 20,
140       stacks->allocated >> 20, stacks->n_uniq_ids,
141       nlive, nthread);
142 }
143 
144 #if SANITIZER_LINUX
FlushShadowMemoryCallback(const SuspendedThreadsList & suspended_threads_list,void * argument)145 void FlushShadowMemoryCallback(
146     const SuspendedThreadsList &suspended_threads_list,
147     void *argument) {
148   ReleaseMemoryPagesToOS(ShadowBeg(), ShadowEnd());
149 }
150 #endif
151 
FlushShadowMemory()152 void FlushShadowMemory() {
153 #if SANITIZER_LINUX
154   StopTheWorld(FlushShadowMemoryCallback, 0);
155 #endif
156 }
157 
158 #if !SANITIZER_GO
159 // Mark shadow for .rodata sections with the special kShadowRodata marker.
160 // Accesses to .rodata can't race, so this saves time, memory and trace space.
MapRodata()161 static void MapRodata() {
162   // First create temp file.
163   const char *tmpdir = GetEnv("TMPDIR");
164   if (tmpdir == 0)
165     tmpdir = GetEnv("TEST_TMPDIR");
166 #ifdef P_tmpdir
167   if (tmpdir == 0)
168     tmpdir = P_tmpdir;
169 #endif
170   if (tmpdir == 0)
171     return;
172   char name[256];
173   internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d",
174                     tmpdir, (int)internal_getpid());
175   uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600);
176   if (internal_iserror(openrv))
177     return;
178   internal_unlink(name);  // Unlink it now, so that we can reuse the buffer.
179   fd_t fd = openrv;
180   // Fill the file with kShadowRodata.
181   const uptr kMarkerSize = 512 * 1024 / sizeof(u64);
182   InternalMmapVector<u64> marker(kMarkerSize);
183   // volatile to prevent insertion of memset
184   for (volatile u64 *p = marker.data(); p < marker.data() + kMarkerSize; p++)
185     *p = kShadowRodata;
186   internal_write(fd, marker.data(), marker.size() * sizeof(u64));
187   // Map the file into memory.
188   uptr page = internal_mmap(0, GetPageSizeCached(), PROT_READ | PROT_WRITE,
189                             MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
190   if (internal_iserror(page)) {
191     internal_close(fd);
192     return;
193   }
194   // Map the file into shadow of .rodata sections.
195   MemoryMappingLayout proc_maps(/*cache_enabled*/true);
196   // Reusing the buffer 'name'.
197   MemoryMappedSegment segment(name, ARRAY_SIZE(name));
198   while (proc_maps.Next(&segment)) {
199     if (segment.filename[0] != 0 && segment.filename[0] != '[' &&
200         segment.IsReadable() && segment.IsExecutable() &&
201         !segment.IsWritable() && IsAppMem(segment.start)) {
202       // Assume it's .rodata
203       char *shadow_start = (char *)MemToShadow(segment.start);
204       char *shadow_end = (char *)MemToShadow(segment.end);
205       for (char *p = shadow_start; p < shadow_end;
206            p += marker.size() * sizeof(u64)) {
207         internal_mmap(p, Min<uptr>(marker.size() * sizeof(u64), shadow_end - p),
208                       PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
209       }
210     }
211   }
212   internal_close(fd);
213 }
214 
InitializeShadowMemoryPlatform()215 void InitializeShadowMemoryPlatform() {
216   MapRodata();
217 }
218 
219 #endif  // #if !SANITIZER_GO
220 
InitializePlatformEarly()221 void InitializePlatformEarly() {
222 #ifdef TSAN_RUNTIME_VMA
223   vmaSize =
224     (MostSignificantSetBitIndex(GET_CURRENT_FRAME()) + 1);
225 #if defined(__aarch64__)
226 # if !SANITIZER_GO
227   if (vmaSize != 39 && vmaSize != 42 && vmaSize != 48) {
228     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
229     Printf("FATAL: Found %zd - Supported 39, 42 and 48\n", vmaSize);
230     Die();
231   }
232 #else
233   if (vmaSize != 48) {
234     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
235     Printf("FATAL: Found %zd - Supported 48\n", vmaSize);
236     Die();
237   }
238 #endif
239 #elif defined(__powerpc64__)
240 # if !SANITIZER_GO
241   if (vmaSize != 44 && vmaSize != 46 && vmaSize != 47) {
242     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
243     Printf("FATAL: Found %zd - Supported 44, 46, and 47\n", vmaSize);
244     Die();
245   }
246 # else
247   if (vmaSize != 46 && vmaSize != 47) {
248     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
249     Printf("FATAL: Found %zd - Supported 46, and 47\n", vmaSize);
250     Die();
251   }
252 # endif
253 #elif defined(__mips64)
254 # if !SANITIZER_GO
255   if (vmaSize != 40) {
256     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
257     Printf("FATAL: Found %zd - Supported 40\n", vmaSize);
258     Die();
259   }
260 # else
261   if (vmaSize != 47) {
262     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
263     Printf("FATAL: Found %zd - Supported 47\n", vmaSize);
264     Die();
265   }
266 # endif
267 #endif
268 #endif
269 }
270 
InitializePlatform()271 void InitializePlatform() {
272   DisableCoreDumperIfNecessary();
273 
274   // Go maps shadow memory lazily and works fine with limited address space.
275   // Unlimited stack is not a problem as well, because the executable
276   // is not compiled with -pie.
277 #if !SANITIZER_GO
278   {
279     bool reexec = false;
280     // TSan doesn't play well with unlimited stack size (as stack
281     // overlaps with shadow memory). If we detect unlimited stack size,
282     // we re-exec the program with limited stack size as a best effort.
283     if (StackSizeIsUnlimited()) {
284       const uptr kMaxStackSize = 32 * 1024 * 1024;
285       VReport(1, "Program is run with unlimited stack size, which wouldn't "
286                  "work with ThreadSanitizer.\n"
287                  "Re-execing with stack size limited to %zd bytes.\n",
288               kMaxStackSize);
289       SetStackSizeLimitInBytes(kMaxStackSize);
290       reexec = true;
291     }
292 
293     if (!AddressSpaceIsUnlimited()) {
294       Report("WARNING: Program is run with limited virtual address space,"
295              " which wouldn't work with ThreadSanitizer.\n");
296       Report("Re-execing with unlimited virtual address space.\n");
297       SetAddressSpaceUnlimited();
298       reexec = true;
299     }
300 #if SANITIZER_LINUX && defined(__aarch64__)
301     // After patch "arm64: mm: support ARCH_MMAP_RND_BITS." is introduced in
302     // linux kernel, the random gap between stack and mapped area is increased
303     // from 128M to 36G on 39-bit aarch64. As it is almost impossible to cover
304     // this big range, we should disable randomized virtual space on aarch64.
305     int old_personality = personality(0xffffffff);
306     if (old_personality != -1 && (old_personality & ADDR_NO_RANDOMIZE) == 0) {
307       VReport(1, "WARNING: Program is run with randomized virtual address "
308               "space, which wouldn't work with ThreadSanitizer.\n"
309               "Re-execing with fixed virtual address space.\n");
310       CHECK_NE(personality(old_personality | ADDR_NO_RANDOMIZE), -1);
311       reexec = true;
312     }
313     // Initialize the xor key used in {sig}{set,long}jump.
314     InitializeLongjmpXorKey();
315 #endif
316     if (reexec)
317       ReExec();
318   }
319 
320   CheckAndProtect();
321   InitTlsSize();
322 #endif  // !SANITIZER_GO
323 }
324 
325 #if !SANITIZER_GO
326 // Extract file descriptors passed to glibc internal __res_iclose function.
327 // This is required to properly "close" the fds, because we do not see internal
328 // closes within glibc. The code is a pure hack.
ExtractResolvFDs(void * state,int * fds,int nfd)329 int ExtractResolvFDs(void *state, int *fds, int nfd) {
330 #if SANITIZER_LINUX && !SANITIZER_ANDROID
331   int cnt = 0;
332   struct __res_state *statp = (struct __res_state*)state;
333   for (int i = 0; i < MAXNS && cnt < nfd; i++) {
334     if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1)
335       fds[cnt++] = statp->_u._ext.nssocks[i];
336   }
337   return cnt;
338 #else
339   return 0;
340 #endif
341 }
342 
343 // Extract file descriptors passed via UNIX domain sockets.
344 // This is requried to properly handle "open" of these fds.
345 // see 'man recvmsg' and 'man 3 cmsg'.
ExtractRecvmsgFDs(void * msgp,int * fds,int nfd)346 int ExtractRecvmsgFDs(void *msgp, int *fds, int nfd) {
347   int res = 0;
348   msghdr *msg = (msghdr*)msgp;
349   struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg);
350   for (; cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
351     if (cmsg->cmsg_level != SOL_SOCKET || cmsg->cmsg_type != SCM_RIGHTS)
352       continue;
353     int n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(fds[0]);
354     for (int i = 0; i < n; i++) {
355       fds[res++] = ((int*)CMSG_DATA(cmsg))[i];
356       if (res == nfd)
357         return res;
358     }
359   }
360   return res;
361 }
362 
363 // Reverse operation of libc stack pointer mangling
UnmangleLongJmpSp(uptr mangled_sp)364 static uptr UnmangleLongJmpSp(uptr mangled_sp) {
365 #if defined(__x86_64__)
366 # if SANITIZER_LINUX
367   // Reverse of:
368   //   xor  %fs:0x30, %rsi
369   //   rol  $0x11, %rsi
370   uptr sp;
371   asm("ror  $0x11,     %0 \n"
372       "xor  %%fs:0x30, %0 \n"
373       : "=r" (sp)
374       : "0" (mangled_sp));
375   return sp;
376 # else
377   return mangled_sp;
378 # endif
379 #elif defined(__aarch64__)
380 # if SANITIZER_LINUX
381   return mangled_sp ^ longjmp_xor_key;
382 # else
383   return mangled_sp;
384 # endif
385 #elif defined(__powerpc64__)
386   // Reverse of:
387   //   ld   r4, -28696(r13)
388   //   xor  r4, r3, r4
389   uptr xor_key;
390   asm("ld  %0, -28696(%%r13)" : "=r" (xor_key));
391   return mangled_sp ^ xor_key;
392 #elif defined(__mips__)
393   return mangled_sp;
394 #elif defined(__s390x__)
395   // tcbhead_t.stack_guard
396   uptr xor_key = ((uptr *)__builtin_thread_pointer())[5];
397   return mangled_sp ^ xor_key;
398 #else
399   #error "Unknown platform"
400 #endif
401 }
402 
403 #if SANITIZER_NETBSD
404 # ifdef __x86_64__
405 #  define LONG_JMP_SP_ENV_SLOT 6
406 # else
407 #  error unsupported
408 # endif
409 #elif defined(__powerpc__)
410 # define LONG_JMP_SP_ENV_SLOT 0
411 #elif SANITIZER_FREEBSD
412 # define LONG_JMP_SP_ENV_SLOT 2
413 #elif SANITIZER_LINUX
414 # ifdef __aarch64__
415 #  define LONG_JMP_SP_ENV_SLOT 13
416 # elif defined(__mips64)
417 #  define LONG_JMP_SP_ENV_SLOT 1
418 # elif defined(__s390x__)
419 #  define LONG_JMP_SP_ENV_SLOT 9
420 # else
421 #  define LONG_JMP_SP_ENV_SLOT 6
422 # endif
423 #endif
424 
ExtractLongJmpSp(uptr * env)425 uptr ExtractLongJmpSp(uptr *env) {
426   uptr mangled_sp = env[LONG_JMP_SP_ENV_SLOT];
427   return UnmangleLongJmpSp(mangled_sp);
428 }
429 
430 #if INIT_LONGJMP_XOR_KEY
431 // GLIBC mangles the function pointers in jmp_buf (used in {set,long}*jmp
432 // functions) by XORing them with a random key.  For AArch64 it is a global
433 // variable rather than a TCB one (as for x86_64/powerpc).  We obtain the key by
434 // issuing a setjmp and XORing the SP pointer values to derive the key.
InitializeLongjmpXorKey()435 static void InitializeLongjmpXorKey() {
436   // 1. Call REAL(setjmp), which stores the mangled SP in env.
437   jmp_buf env;
438   REAL(_setjmp)(env);
439 
440   // 2. Retrieve vanilla/mangled SP.
441   uptr sp;
442   asm("mov  %0, sp" : "=r" (sp));
443   uptr mangled_sp = ((uptr *)&env)[LONG_JMP_SP_ENV_SLOT];
444 
445   // 3. xor SPs to obtain key.
446   longjmp_xor_key = mangled_sp ^ sp;
447 }
448 #endif
449 
ImitateTlsWrite(ThreadState * thr,uptr tls_addr,uptr tls_size)450 void ImitateTlsWrite(ThreadState *thr, uptr tls_addr, uptr tls_size) {
451   // Check that the thr object is in tls;
452   const uptr thr_beg = (uptr)thr;
453   const uptr thr_end = (uptr)thr + sizeof(*thr);
454   CHECK_GE(thr_beg, tls_addr);
455   CHECK_LE(thr_beg, tls_addr + tls_size);
456   CHECK_GE(thr_end, tls_addr);
457   CHECK_LE(thr_end, tls_addr + tls_size);
458   // Since the thr object is huge, skip it.
459   MemoryRangeImitateWrite(thr, /*pc=*/2, tls_addr, thr_beg - tls_addr);
460   MemoryRangeImitateWrite(thr, /*pc=*/2, thr_end,
461                           tls_addr + tls_size - thr_end);
462 }
463 
464 // Note: this function runs with async signals enabled,
465 // so it must not touch any tsan state.
call_pthread_cancel_with_cleanup(int (* fn)(void * arg),void (* cleanup)(void * arg),void * arg)466 int call_pthread_cancel_with_cleanup(int (*fn)(void *arg),
467                                      void (*cleanup)(void *arg), void *arg) {
468   // pthread_cleanup_push/pop are hardcore macros mess.
469   // We can't intercept nor call them w/o including pthread.h.
470   int res;
471   pthread_cleanup_push(cleanup, arg);
472   res = fn(arg);
473   pthread_cleanup_pop(0);
474   return res;
475 }
476 #endif  // !SANITIZER_GO
477 
478 #if !SANITIZER_GO
ReplaceSystemMalloc()479 void ReplaceSystemMalloc() { }
480 #endif
481 
482 #if !SANITIZER_GO
483 #if SANITIZER_ANDROID
484 // On Android, one thread can call intercepted functions after
485 // DestroyThreadState(), so add a fake thread state for "dead" threads.
486 static ThreadState *dead_thread_state = nullptr;
487 
cur_thread()488 ThreadState *cur_thread() {
489   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
490   if (thr == nullptr) {
491     __sanitizer_sigset_t emptyset;
492     internal_sigfillset(&emptyset);
493     __sanitizer_sigset_t oldset;
494     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
495     thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
496     if (thr == nullptr) {
497       thr = reinterpret_cast<ThreadState*>(MmapOrDie(sizeof(ThreadState),
498                                                      "ThreadState"));
499       *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
500       if (dead_thread_state == nullptr) {
501         dead_thread_state = reinterpret_cast<ThreadState*>(
502             MmapOrDie(sizeof(ThreadState), "ThreadState"));
503         dead_thread_state->fast_state.SetIgnoreBit();
504         dead_thread_state->ignore_interceptors = 1;
505         dead_thread_state->is_dead = true;
506         *const_cast<u32*>(&dead_thread_state->tid) = -1;
507         CHECK_EQ(0, internal_mprotect(dead_thread_state, sizeof(ThreadState),
508                                       PROT_READ));
509       }
510     }
511     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
512   }
513   return thr;
514 }
515 
set_cur_thread(ThreadState * thr)516 void set_cur_thread(ThreadState *thr) {
517   *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
518 }
519 
cur_thread_finalize()520 void cur_thread_finalize() {
521   __sanitizer_sigset_t emptyset;
522   internal_sigfillset(&emptyset);
523   __sanitizer_sigset_t oldset;
524   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
525   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
526   if (thr != dead_thread_state) {
527     *get_android_tls_ptr() = reinterpret_cast<uptr>(dead_thread_state);
528     UnmapOrDie(thr, sizeof(ThreadState));
529   }
530   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
531 }
532 #endif  // SANITIZER_ANDROID
533 #endif  // if !SANITIZER_GO
534 
535 }  // namespace __tsan
536 
537 #endif  // SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD
538