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   MemHeap,
98   MemOther,
99   MemCount,
100 };
101 
102 void FillProfileCallback(uptr p, uptr rss, bool file, uptr *mem) {
103   mem[MemTotal] += rss;
104   if (p >= ShadowBeg() && p < ShadowEnd())
105     mem[MemShadow] += rss;
106   else if (p >= MetaShadowBeg() && p < MetaShadowEnd())
107     mem[MemMeta] += rss;
108   else if ((p >= LoAppMemBeg() && p < LoAppMemEnd()) ||
109            (p >= MidAppMemBeg() && p < MidAppMemEnd()) ||
110            (p >= HiAppMemBeg() && p < HiAppMemEnd()))
111     mem[file ? MemFile : MemMmap] += rss;
112   else if (p >= HeapMemBeg() && p < HeapMemEnd())
113     mem[MemHeap] += rss;
114   else
115     mem[MemOther] += rss;
116 }
117 
118 void WriteMemoryProfile(char *buf, uptr buf_size, u64 uptime_ns) {
119   uptr mem[MemCount];
120   internal_memset(mem, 0, sizeof(mem));
121   GetMemoryProfile(FillProfileCallback, mem);
122   auto meta = ctx->metamap.GetMemoryStats();
123   StackDepotStats stacks = StackDepotGetStats();
124   uptr nthread, nlive;
125   ctx->thread_registry.GetNumberOfThreads(&nthread, &nlive);
126   uptr trace_mem;
127   {
128     Lock l(&ctx->slot_mtx);
129     trace_mem = ctx->trace_part_total_allocated * sizeof(TracePart);
130   }
131   uptr internal_stats[AllocatorStatCount];
132   internal_allocator()->GetStats(internal_stats);
133   // All these are allocated from the common mmap region.
134   mem[MemMmap] -= meta.mem_block + meta.sync_obj + trace_mem +
135                   stacks.allocated + internal_stats[AllocatorStatMapped];
136   if (s64(mem[MemMmap]) < 0)
137     mem[MemMmap] = 0;
138   internal_snprintf(
139       buf, buf_size,
140       "==%zu== %llus [%zu]: RSS %zd MB: shadow:%zd meta:%zd file:%zd"
141       " mmap:%zd heap:%zd other:%zd intalloc:%zd memblocks:%zd syncobj:%zu"
142       " trace:%zu stacks=%zd threads=%zu/%zu\n",
143       internal_getpid(), uptime_ns / (1000 * 1000 * 1000), ctx->global_epoch,
144       mem[MemTotal] >> 20, mem[MemShadow] >> 20, mem[MemMeta] >> 20,
145       mem[MemFile] >> 20, mem[MemMmap] >> 20, mem[MemHeap] >> 20,
146       mem[MemOther] >> 20, internal_stats[AllocatorStatMapped] >> 20,
147       meta.mem_block >> 20, meta.sync_obj >> 20, trace_mem >> 20,
148       stacks.allocated >> 20, nlive, nthread);
149 }
150 
151 #if !SANITIZER_GO
152 // Mark shadow for .rodata sections with the special Shadow::kRodata marker.
153 // Accesses to .rodata can't race, so this saves time, memory and trace space.
154 static void MapRodata() {
155   // First create temp file.
156   const char *tmpdir = GetEnv("TMPDIR");
157   if (tmpdir == 0)
158     tmpdir = GetEnv("TEST_TMPDIR");
159 #ifdef P_tmpdir
160   if (tmpdir == 0)
161     tmpdir = P_tmpdir;
162 #endif
163   if (tmpdir == 0)
164     return;
165   char name[256];
166   internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d",
167                     tmpdir, (int)internal_getpid());
168   uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600);
169   if (internal_iserror(openrv))
170     return;
171   internal_unlink(name);  // Unlink it now, so that we can reuse the buffer.
172   fd_t fd = openrv;
173   // Fill the file with Shadow::kRodata.
174   const uptr kMarkerSize = 512 * 1024 / sizeof(RawShadow);
175   InternalMmapVector<RawShadow> marker(kMarkerSize);
176   // volatile to prevent insertion of memset
177   for (volatile RawShadow *p = marker.data(); p < marker.data() + kMarkerSize;
178        p++)
179     *p = Shadow::kRodata;
180   internal_write(fd, marker.data(), marker.size() * sizeof(RawShadow));
181   // Map the file into memory.
182   uptr page = internal_mmap(0, GetPageSizeCached(), PROT_READ | PROT_WRITE,
183                             MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
184   if (internal_iserror(page)) {
185     internal_close(fd);
186     return;
187   }
188   // Map the file into shadow of .rodata sections.
189   MemoryMappingLayout proc_maps(/*cache_enabled*/true);
190   // Reusing the buffer 'name'.
191   MemoryMappedSegment segment(name, ARRAY_SIZE(name));
192   while (proc_maps.Next(&segment)) {
193     if (segment.filename[0] != 0 && segment.filename[0] != '[' &&
194         segment.IsReadable() && segment.IsExecutable() &&
195         !segment.IsWritable() && IsAppMem(segment.start)) {
196       // Assume it's .rodata
197       char *shadow_start = (char *)MemToShadow(segment.start);
198       char *shadow_end = (char *)MemToShadow(segment.end);
199       for (char *p = shadow_start; p < shadow_end;
200            p += marker.size() * sizeof(RawShadow)) {
201         internal_mmap(
202             p, Min<uptr>(marker.size() * sizeof(RawShadow), shadow_end - p),
203             PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
204       }
205     }
206   }
207   internal_close(fd);
208 }
209 
210 void InitializeShadowMemoryPlatform() {
211   MapRodata();
212 }
213 
214 #endif  // #if !SANITIZER_GO
215 
216 void InitializePlatformEarly() {
217   vmaSize =
218     (MostSignificantSetBitIndex(GET_CURRENT_FRAME()) + 1);
219 #if defined(__aarch64__)
220 # if !SANITIZER_GO
221   if (vmaSize != 39 && vmaSize != 42 && vmaSize != 48) {
222     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
223     Printf("FATAL: Found %zd - Supported 39, 42 and 48\n", vmaSize);
224     Die();
225   }
226 #else
227   if (vmaSize != 48) {
228     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
229     Printf("FATAL: Found %zd - Supported 48\n", vmaSize);
230     Die();
231   }
232 #endif
233 #elif SANITIZER_LOONGARCH64
234 # if !SANITIZER_GO
235   if (vmaSize != 47) {
236     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
237     Printf("FATAL: Found %zd - Supported 47\n", vmaSize);
238     Die();
239   }
240 # endif
241 #elif defined(__powerpc64__)
242 # if !SANITIZER_GO
243   if (vmaSize != 44 && vmaSize != 46 && vmaSize != 47) {
244     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
245     Printf("FATAL: Found %zd - Supported 44, 46, and 47\n", vmaSize);
246     Die();
247   }
248 # else
249   if (vmaSize != 46 && vmaSize != 47) {
250     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
251     Printf("FATAL: Found %zd - Supported 46, and 47\n", vmaSize);
252     Die();
253   }
254 # endif
255 #elif defined(__mips64)
256 # if !SANITIZER_GO
257   if (vmaSize != 40) {
258     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
259     Printf("FATAL: Found %zd - Supported 40\n", vmaSize);
260     Die();
261   }
262 # else
263   if (vmaSize != 47) {
264     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
265     Printf("FATAL: Found %zd - Supported 47\n", vmaSize);
266     Die();
267   }
268 # endif
269 #endif
270 }
271 
272 void InitializePlatform() {
273   DisableCoreDumperIfNecessary();
274 
275   // Go maps shadow memory lazily and works fine with limited address space.
276   // Unlimited stack is not a problem as well, because the executable
277   // is not compiled with -pie.
278 #if !SANITIZER_GO
279   {
280     bool reexec = false;
281     // TSan doesn't play well with unlimited stack size (as stack
282     // overlaps with shadow memory). If we detect unlimited stack size,
283     // we re-exec the program with limited stack size as a best effort.
284     if (StackSizeIsUnlimited()) {
285       const uptr kMaxStackSize = 32 * 1024 * 1024;
286       VReport(1, "Program is run with unlimited stack size, which wouldn't "
287                  "work with ThreadSanitizer.\n"
288                  "Re-execing with stack size limited to %zd bytes.\n",
289               kMaxStackSize);
290       SetStackSizeLimitInBytes(kMaxStackSize);
291       reexec = true;
292     }
293 
294     if (!AddressSpaceIsUnlimited()) {
295       Report("WARNING: Program is run with limited virtual address space,"
296              " which wouldn't work with ThreadSanitizer.\n");
297       Report("Re-execing with unlimited virtual address space.\n");
298       SetAddressSpaceUnlimited();
299       reexec = true;
300     }
301 #if SANITIZER_ANDROID && (defined(__aarch64__) || defined(__x86_64__))
302     // After patch "arm64: mm: support ARCH_MMAP_RND_BITS." is introduced in
303     // linux kernel, the random gap between stack and mapped area is increased
304     // from 128M to 36G on 39-bit aarch64. As it is almost impossible to cover
305     // this big range, we should disable randomized virtual space on aarch64.
306     // ASLR personality check.
307     int old_personality = personality(0xffffffff);
308     if (old_personality != -1 && (old_personality & ADDR_NO_RANDOMIZE) == 0) {
309       VReport(1, "WARNING: Program is run with randomized virtual address "
310               "space, which wouldn't work with ThreadSanitizer.\n"
311               "Re-execing with fixed virtual address space.\n");
312       CHECK_NE(personality(old_personality | ADDR_NO_RANDOMIZE), -1);
313       reexec = true;
314     }
315 
316 #endif
317 #if SANITIZER_LINUX && defined(__aarch64__)
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(__loongarch__)
391   return mangled_sp;
392 #elif defined(__powerpc64__)
393   // Reverse of:
394   //   ld   r4, -28696(r13)
395   //   xor  r4, r3, r4
396   uptr xor_key;
397   asm("ld  %0, -28696(%%r13)" : "=r" (xor_key));
398   return mangled_sp ^ xor_key;
399 #elif defined(__mips__)
400   return mangled_sp;
401 #elif defined(__s390x__)
402   // tcbhead_t.stack_guard
403   uptr xor_key = ((uptr *)__builtin_thread_pointer())[5];
404   return mangled_sp ^ xor_key;
405 #else
406   #error "Unknown platform"
407 #endif
408 }
409 
410 #if SANITIZER_NETBSD
411 # ifdef __x86_64__
412 #  define LONG_JMP_SP_ENV_SLOT 6
413 # else
414 #  error unsupported
415 # endif
416 #elif defined(__powerpc__)
417 # define LONG_JMP_SP_ENV_SLOT 0
418 #elif SANITIZER_FREEBSD
419 # ifdef __aarch64__
420 #  define LONG_JMP_SP_ENV_SLOT 1
421 # else
422 #  define LONG_JMP_SP_ENV_SLOT 2
423 # endif
424 #elif SANITIZER_LINUX
425 # ifdef __aarch64__
426 #  define LONG_JMP_SP_ENV_SLOT 13
427 # elif defined(__loongarch__)
428 #  define LONG_JMP_SP_ENV_SLOT 1
429 # elif defined(__mips64)
430 #  define LONG_JMP_SP_ENV_SLOT 1
431 # elif defined(__s390x__)
432 #  define LONG_JMP_SP_ENV_SLOT 9
433 # else
434 #  define LONG_JMP_SP_ENV_SLOT 6
435 # endif
436 #endif
437 
438 uptr ExtractLongJmpSp(uptr *env) {
439   uptr mangled_sp = env[LONG_JMP_SP_ENV_SLOT];
440   return UnmangleLongJmpSp(mangled_sp);
441 }
442 
443 #if INIT_LONGJMP_XOR_KEY
444 // GLIBC mangles the function pointers in jmp_buf (used in {set,long}*jmp
445 // functions) by XORing them with a random key.  For AArch64 it is a global
446 // variable rather than a TCB one (as for x86_64/powerpc).  We obtain the key by
447 // issuing a setjmp and XORing the SP pointer values to derive the key.
448 static void InitializeLongjmpXorKey() {
449   // 1. Call REAL(setjmp), which stores the mangled SP in env.
450   jmp_buf env;
451   REAL(_setjmp)(env);
452 
453   // 2. Retrieve vanilla/mangled SP.
454   uptr sp;
455   asm("mov  %0, sp" : "=r" (sp));
456   uptr mangled_sp = ((uptr *)&env)[LONG_JMP_SP_ENV_SLOT];
457 
458   // 3. xor SPs to obtain key.
459   longjmp_xor_key = mangled_sp ^ sp;
460 }
461 #endif
462 
463 extern "C" void __tsan_tls_initialization() {}
464 
465 void ImitateTlsWrite(ThreadState *thr, uptr tls_addr, uptr tls_size) {
466   // Check that the thr object is in tls;
467   const uptr thr_beg = (uptr)thr;
468   const uptr thr_end = (uptr)thr + sizeof(*thr);
469   CHECK_GE(thr_beg, tls_addr);
470   CHECK_LE(thr_beg, tls_addr + tls_size);
471   CHECK_GE(thr_end, tls_addr);
472   CHECK_LE(thr_end, tls_addr + tls_size);
473   // Since the thr object is huge, skip it.
474   const uptr pc = StackTrace::GetNextInstructionPc(
475       reinterpret_cast<uptr>(__tsan_tls_initialization));
476   MemoryRangeImitateWrite(thr, pc, tls_addr, thr_beg - tls_addr);
477   MemoryRangeImitateWrite(thr, pc, thr_end, tls_addr + tls_size - thr_end);
478 }
479 
480 // Note: this function runs with async signals enabled,
481 // so it must not touch any tsan state.
482 int call_pthread_cancel_with_cleanup(int (*fn)(void *arg),
483                                      void (*cleanup)(void *arg), void *arg) {
484   // pthread_cleanup_push/pop are hardcore macros mess.
485   // We can't intercept nor call them w/o including pthread.h.
486   int res;
487   pthread_cleanup_push(cleanup, arg);
488   res = fn(arg);
489   pthread_cleanup_pop(0);
490   return res;
491 }
492 #endif  // !SANITIZER_GO
493 
494 #if !SANITIZER_GO
495 void ReplaceSystemMalloc() { }
496 #endif
497 
498 #if !SANITIZER_GO
499 #if SANITIZER_ANDROID
500 // On Android, one thread can call intercepted functions after
501 // DestroyThreadState(), so add a fake thread state for "dead" threads.
502 static ThreadState *dead_thread_state = nullptr;
503 
504 ThreadState *cur_thread() {
505   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
506   if (thr == nullptr) {
507     __sanitizer_sigset_t emptyset;
508     internal_sigfillset(&emptyset);
509     __sanitizer_sigset_t oldset;
510     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
511     thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
512     if (thr == nullptr) {
513       thr = reinterpret_cast<ThreadState*>(MmapOrDie(sizeof(ThreadState),
514                                                      "ThreadState"));
515       *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
516       if (dead_thread_state == nullptr) {
517         dead_thread_state = reinterpret_cast<ThreadState*>(
518             MmapOrDie(sizeof(ThreadState), "ThreadState"));
519         dead_thread_state->fast_state.SetIgnoreBit();
520         dead_thread_state->ignore_interceptors = 1;
521         dead_thread_state->is_dead = true;
522         *const_cast<u32*>(&dead_thread_state->tid) = -1;
523         CHECK_EQ(0, internal_mprotect(dead_thread_state, sizeof(ThreadState),
524                                       PROT_READ));
525       }
526     }
527     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
528   }
529   return thr;
530 }
531 
532 void set_cur_thread(ThreadState *thr) {
533   *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
534 }
535 
536 void cur_thread_finalize() {
537   __sanitizer_sigset_t emptyset;
538   internal_sigfillset(&emptyset);
539   __sanitizer_sigset_t oldset;
540   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
541   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
542   if (thr != dead_thread_state) {
543     *get_android_tls_ptr() = reinterpret_cast<uptr>(dead_thread_state);
544     UnmapOrDie(thr, sizeof(ThreadState));
545   }
546   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
547 }
548 #endif  // SANITIZER_ANDROID
549 #endif  // if !SANITIZER_GO
550 
551 }  // namespace __tsan
552 
553 #endif  // SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD
554