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__) || defined(__loongarch_lp64)) && \
70     !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 // Runtime detected VMA size.
90 uptr vmaSize;
91 
92 enum {
93   MemTotal,
94   MemShadow,
95   MemMeta,
96   MemFile,
97   MemMmap,
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
116     mem[MemOther] += rss;
117 }
118 
119 void WriteMemoryProfile(char *buf, uptr buf_size, u64 uptime_ns) {
120   uptr mem[MemCount];
121   internal_memset(mem, 0, sizeof(mem));
122   GetMemoryProfile(FillProfileCallback, mem);
123   auto meta = ctx->metamap.GetMemoryStats();
124   StackDepotStats stacks = StackDepotGetStats();
125   uptr nthread, nlive;
126   ctx->thread_registry.GetNumberOfThreads(&nthread, &nlive);
127   uptr trace_mem;
128   {
129     Lock l(&ctx->slot_mtx);
130     trace_mem = ctx->trace_part_total_allocated * sizeof(TracePart);
131   }
132   uptr internal_stats[AllocatorStatCount];
133   internal_allocator()->GetStats(internal_stats);
134   // All these are allocated from the common mmap region.
135   mem[MemMmap] -= meta.mem_block + meta.sync_obj + trace_mem +
136                   stacks.allocated + internal_stats[AllocatorStatMapped];
137   if (s64(mem[MemMmap]) < 0)
138     mem[MemMmap] = 0;
139   internal_snprintf(
140       buf, buf_size,
141       "==%zu== %llus [%zu]: RSS %zd MB: shadow:%zd meta:%zd file:%zd"
142       " mmap:%zd heap:%zd other:%zd intalloc:%zd memblocks:%zd syncobj:%zu"
143       " trace:%zu stacks=%zd threads=%zu/%zu\n",
144       internal_getpid(), uptime_ns / (1000 * 1000 * 1000), ctx->global_epoch,
145       mem[MemTotal] >> 20, mem[MemShadow] >> 20, mem[MemMeta] >> 20,
146       mem[MemFile] >> 20, mem[MemMmap] >> 20, mem[MemHeap] >> 20,
147       mem[MemOther] >> 20, internal_stats[AllocatorStatMapped] >> 20,
148       meta.mem_block >> 20, meta.sync_obj >> 20, trace_mem >> 20,
149       stacks.allocated >> 20, nlive, nthread);
150 }
151 
152 #if !SANITIZER_GO
153 // Mark shadow for .rodata sections with the special Shadow::kRodata marker.
154 // Accesses to .rodata can't race, so this saves time, memory and trace space.
155 static void MapRodata() {
156   // First create temp file.
157   const char *tmpdir = GetEnv("TMPDIR");
158   if (tmpdir == 0)
159     tmpdir = GetEnv("TEST_TMPDIR");
160 #ifdef P_tmpdir
161   if (tmpdir == 0)
162     tmpdir = P_tmpdir;
163 #endif
164   if (tmpdir == 0)
165     return;
166   char name[256];
167   internal_snprintf(name, sizeof(name), "%s/tsan.rodata.%d",
168                     tmpdir, (int)internal_getpid());
169   uptr openrv = internal_open(name, O_RDWR | O_CREAT | O_EXCL, 0600);
170   if (internal_iserror(openrv))
171     return;
172   internal_unlink(name);  // Unlink it now, so that we can reuse the buffer.
173   fd_t fd = openrv;
174   // Fill the file with Shadow::kRodata.
175   const uptr kMarkerSize = 512 * 1024 / sizeof(RawShadow);
176   InternalMmapVector<RawShadow> marker(kMarkerSize);
177   // volatile to prevent insertion of memset
178   for (volatile RawShadow *p = marker.data(); p < marker.data() + kMarkerSize;
179        p++)
180     *p = Shadow::kRodata;
181   internal_write(fd, marker.data(), marker.size() * sizeof(RawShadow));
182   // Map the file into memory.
183   uptr page = internal_mmap(0, GetPageSizeCached(), PROT_READ | PROT_WRITE,
184                             MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
185   if (internal_iserror(page)) {
186     internal_close(fd);
187     return;
188   }
189   // Map the file into shadow of .rodata sections.
190   MemoryMappingLayout proc_maps(/*cache_enabled*/true);
191   // Reusing the buffer 'name'.
192   MemoryMappedSegment segment(name, ARRAY_SIZE(name));
193   while (proc_maps.Next(&segment)) {
194     if (segment.filename[0] != 0 && segment.filename[0] != '[' &&
195         segment.IsReadable() && segment.IsExecutable() &&
196         !segment.IsWritable() && IsAppMem(segment.start)) {
197       // Assume it's .rodata
198       char *shadow_start = (char *)MemToShadow(segment.start);
199       char *shadow_end = (char *)MemToShadow(segment.end);
200       for (char *p = shadow_start; p < shadow_end;
201            p += marker.size() * sizeof(RawShadow)) {
202         internal_mmap(
203             p, Min<uptr>(marker.size() * sizeof(RawShadow), shadow_end - p),
204             PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
205       }
206     }
207   }
208   internal_close(fd);
209 }
210 
211 void InitializeShadowMemoryPlatform() {
212   MapRodata();
213 }
214 
215 #endif  // #if !SANITIZER_GO
216 
217 void InitializePlatformEarly() {
218   vmaSize =
219     (MostSignificantSetBitIndex(GET_CURRENT_FRAME()) + 1);
220 #if defined(__aarch64__)
221 # if !SANITIZER_GO
222   if (vmaSize != 39 && vmaSize != 42 && vmaSize != 48) {
223     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
224     Printf("FATAL: Found %zd - Supported 39, 42 and 48\n", vmaSize);
225     Die();
226   }
227 #else
228   if (vmaSize != 48) {
229     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
230     Printf("FATAL: Found %zd - Supported 48\n", vmaSize);
231     Die();
232   }
233 #endif
234 #elif SANITIZER_LOONGARCH64
235 # if !SANITIZER_GO
236   if (vmaSize != 47) {
237     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
238     Printf("FATAL: Found %zd - Supported 47\n", vmaSize);
239     Die();
240   }
241 # endif
242 #elif defined(__powerpc64__)
243 # if !SANITIZER_GO
244   if (vmaSize != 44 && vmaSize != 46 && vmaSize != 47) {
245     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
246     Printf("FATAL: Found %zd - Supported 44, 46, and 47\n", vmaSize);
247     Die();
248   }
249 # else
250   if (vmaSize != 46 && vmaSize != 47) {
251     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
252     Printf("FATAL: Found %zd - Supported 46, and 47\n", vmaSize);
253     Die();
254   }
255 # endif
256 #elif defined(__mips64)
257 # if !SANITIZER_GO
258   if (vmaSize != 40) {
259     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
260     Printf("FATAL: Found %zd - Supported 40\n", vmaSize);
261     Die();
262   }
263 # else
264   if (vmaSize != 47) {
265     Printf("FATAL: ThreadSanitizer: unsupported VMA range\n");
266     Printf("FATAL: Found %zd - Supported 47\n", vmaSize);
267     Die();
268   }
269 # endif
270 #endif
271 }
272 
273 void InitializePlatform() {
274   DisableCoreDumperIfNecessary();
275 
276   // Go maps shadow memory lazily and works fine with limited address space.
277   // Unlimited stack is not a problem as well, because the executable
278   // is not compiled with -pie.
279 #if !SANITIZER_GO
280   {
281     bool reexec = false;
282     // TSan doesn't play well with unlimited stack size (as stack
283     // overlaps with shadow memory). If we detect unlimited stack size,
284     // we re-exec the program with limited stack size as a best effort.
285     if (StackSizeIsUnlimited()) {
286       const uptr kMaxStackSize = 32 * 1024 * 1024;
287       VReport(1, "Program is run with unlimited stack size, which wouldn't "
288                  "work with ThreadSanitizer.\n"
289                  "Re-execing with stack size limited to %zd bytes.\n",
290               kMaxStackSize);
291       SetStackSizeLimitInBytes(kMaxStackSize);
292       reexec = true;
293     }
294 
295     if (!AddressSpaceIsUnlimited()) {
296       Report("WARNING: Program is run with limited virtual address space,"
297              " which wouldn't work with ThreadSanitizer.\n");
298       Report("Re-execing with unlimited virtual address space.\n");
299       SetAddressSpaceUnlimited();
300       reexec = true;
301     }
302 #if SANITIZER_ANDROID && (defined(__aarch64__) || defined(__x86_64__))
303     // After patch "arm64: mm: support ARCH_MMAP_RND_BITS." is introduced in
304     // linux kernel, the random gap between stack and mapped area is increased
305     // from 128M to 36G on 39-bit aarch64. As it is almost impossible to cover
306     // this big range, we should disable randomized virtual space on aarch64.
307     // ASLR personality check.
308     int old_personality = personality(0xffffffff);
309     if (old_personality != -1 && (old_personality & ADDR_NO_RANDOMIZE) == 0) {
310       VReport(1, "WARNING: Program is run with randomized virtual address "
311               "space, which wouldn't work with ThreadSanitizer.\n"
312               "Re-execing with fixed virtual address space.\n");
313       CHECK_NE(personality(old_personality | ADDR_NO_RANDOMIZE), -1);
314       reexec = true;
315     }
316 
317 #endif
318 #if SANITIZER_LINUX && (defined(__aarch64__) || defined(__loongarch_lp64))
319     // Initialize the xor key used in {sig}{set,long}jump.
320     InitializeLongjmpXorKey();
321 #endif
322     if (reexec)
323       ReExec();
324   }
325 
326   CheckAndProtect();
327   InitTlsSize();
328 #endif  // !SANITIZER_GO
329 }
330 
331 #if !SANITIZER_GO
332 // Extract file descriptors passed to glibc internal __res_iclose function.
333 // This is required to properly "close" the fds, because we do not see internal
334 // closes within glibc. The code is a pure hack.
335 int ExtractResolvFDs(void *state, int *fds, int nfd) {
336 #if SANITIZER_LINUX && !SANITIZER_ANDROID
337   int cnt = 0;
338   struct __res_state *statp = (struct __res_state*)state;
339   for (int i = 0; i < MAXNS && cnt < nfd; i++) {
340     if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1)
341       fds[cnt++] = statp->_u._ext.nssocks[i];
342   }
343   return cnt;
344 #else
345   return 0;
346 #endif
347 }
348 
349 // Extract file descriptors passed via UNIX domain sockets.
350 // This is required to properly handle "open" of these fds.
351 // see 'man recvmsg' and 'man 3 cmsg'.
352 int ExtractRecvmsgFDs(void *msgp, int *fds, int nfd) {
353   int res = 0;
354   msghdr *msg = (msghdr*)msgp;
355   struct cmsghdr *cmsg = CMSG_FIRSTHDR(msg);
356   for (; cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
357     if (cmsg->cmsg_level != SOL_SOCKET || cmsg->cmsg_type != SCM_RIGHTS)
358       continue;
359     int n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(fds[0]);
360     for (int i = 0; i < n; i++) {
361       fds[res++] = ((int*)CMSG_DATA(cmsg))[i];
362       if (res == nfd)
363         return res;
364     }
365   }
366   return res;
367 }
368 
369 // Reverse operation of libc stack pointer mangling
370 static uptr UnmangleLongJmpSp(uptr mangled_sp) {
371 #if defined(__x86_64__)
372 # if SANITIZER_LINUX
373   // Reverse of:
374   //   xor  %fs:0x30, %rsi
375   //   rol  $0x11, %rsi
376   uptr sp;
377   asm("ror  $0x11,     %0 \n"
378       "xor  %%fs:0x30, %0 \n"
379       : "=r" (sp)
380       : "0" (mangled_sp));
381   return sp;
382 # else
383   return mangled_sp;
384 # endif
385 #elif defined(__aarch64__)
386 # if SANITIZER_LINUX
387   return mangled_sp ^ longjmp_xor_key;
388 # else
389   return mangled_sp;
390 # endif
391 #elif defined(__loongarch_lp64)
392   return mangled_sp ^ longjmp_xor_key;
393 #elif defined(__powerpc64__)
394   // Reverse of:
395   //   ld   r4, -28696(r13)
396   //   xor  r4, r3, r4
397   uptr xor_key;
398   asm("ld  %0, -28696(%%r13)" : "=r" (xor_key));
399   return mangled_sp ^ xor_key;
400 #elif defined(__mips__)
401   return mangled_sp;
402 #elif defined(__s390x__)
403   // tcbhead_t.stack_guard
404   uptr xor_key = ((uptr *)__builtin_thread_pointer())[5];
405   return mangled_sp ^ xor_key;
406 #else
407   #error "Unknown platform"
408 #endif
409 }
410 
411 #if SANITIZER_NETBSD
412 # ifdef __x86_64__
413 #  define LONG_JMP_SP_ENV_SLOT 6
414 # else
415 #  error unsupported
416 # endif
417 #elif defined(__powerpc__)
418 # define LONG_JMP_SP_ENV_SLOT 0
419 #elif SANITIZER_FREEBSD
420 # ifdef __aarch64__
421 #  define LONG_JMP_SP_ENV_SLOT 1
422 # else
423 #  define LONG_JMP_SP_ENV_SLOT 2
424 # endif
425 #elif SANITIZER_LINUX
426 # ifdef __aarch64__
427 #  define LONG_JMP_SP_ENV_SLOT 13
428 # elif defined(__loongarch__)
429 #  define LONG_JMP_SP_ENV_SLOT 1
430 # elif defined(__mips64)
431 #  define LONG_JMP_SP_ENV_SLOT 1
432 # elif defined(__s390x__)
433 #  define LONG_JMP_SP_ENV_SLOT 9
434 # else
435 #  define LONG_JMP_SP_ENV_SLOT 6
436 # endif
437 #endif
438 
439 uptr ExtractLongJmpSp(uptr *env) {
440   uptr mangled_sp = env[LONG_JMP_SP_ENV_SLOT];
441   return UnmangleLongJmpSp(mangled_sp);
442 }
443 
444 #if INIT_LONGJMP_XOR_KEY
445 // GLIBC mangles the function pointers in jmp_buf (used in {set,long}*jmp
446 // functions) by XORing them with a random key.  For AArch64 it is a global
447 // variable rather than a TCB one (as for x86_64/powerpc).  We obtain the key by
448 // issuing a setjmp and XORing the SP pointer values to derive the key.
449 static void InitializeLongjmpXorKey() {
450   // 1. Call REAL(setjmp), which stores the mangled SP in env.
451   jmp_buf env;
452   REAL(_setjmp)(env);
453 
454   // 2. Retrieve vanilla/mangled SP.
455   uptr sp;
456 #ifdef __loongarch__
457   asm("move  %0, $sp" : "=r" (sp));
458 #else
459   asm("mov  %0, sp" : "=r" (sp));
460 #endif
461   uptr mangled_sp = ((uptr *)&env)[LONG_JMP_SP_ENV_SLOT];
462 
463   // 3. xor SPs to obtain key.
464   longjmp_xor_key = mangled_sp ^ sp;
465 }
466 #endif
467 
468 extern "C" void __tsan_tls_initialization() {}
469 
470 void ImitateTlsWrite(ThreadState *thr, uptr tls_addr, uptr tls_size) {
471   // Check that the thr object is in tls;
472   const uptr thr_beg = (uptr)thr;
473   const uptr thr_end = (uptr)thr + sizeof(*thr);
474   CHECK_GE(thr_beg, tls_addr);
475   CHECK_LE(thr_beg, tls_addr + tls_size);
476   CHECK_GE(thr_end, tls_addr);
477   CHECK_LE(thr_end, tls_addr + tls_size);
478   // Since the thr object is huge, skip it.
479   const uptr pc = StackTrace::GetNextInstructionPc(
480       reinterpret_cast<uptr>(__tsan_tls_initialization));
481   MemoryRangeImitateWrite(thr, pc, tls_addr, thr_beg - tls_addr);
482   MemoryRangeImitateWrite(thr, pc, thr_end, tls_addr + tls_size - thr_end);
483 }
484 
485 // Note: this function runs with async signals enabled,
486 // so it must not touch any tsan state.
487 int call_pthread_cancel_with_cleanup(int (*fn)(void *arg),
488                                      void (*cleanup)(void *arg), void *arg) {
489   // pthread_cleanup_push/pop are hardcore macros mess.
490   // We can't intercept nor call them w/o including pthread.h.
491   int res;
492   pthread_cleanup_push(cleanup, arg);
493   res = fn(arg);
494   pthread_cleanup_pop(0);
495   return res;
496 }
497 #endif  // !SANITIZER_GO
498 
499 #if !SANITIZER_GO
500 void ReplaceSystemMalloc() { }
501 #endif
502 
503 #if !SANITIZER_GO
504 #if SANITIZER_ANDROID
505 // On Android, one thread can call intercepted functions after
506 // DestroyThreadState(), so add a fake thread state for "dead" threads.
507 static ThreadState *dead_thread_state = nullptr;
508 
509 ThreadState *cur_thread() {
510   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
511   if (thr == nullptr) {
512     __sanitizer_sigset_t emptyset;
513     internal_sigfillset(&emptyset);
514     __sanitizer_sigset_t oldset;
515     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
516     thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
517     if (thr == nullptr) {
518       thr = reinterpret_cast<ThreadState*>(MmapOrDie(sizeof(ThreadState),
519                                                      "ThreadState"));
520       *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
521       if (dead_thread_state == nullptr) {
522         dead_thread_state = reinterpret_cast<ThreadState*>(
523             MmapOrDie(sizeof(ThreadState), "ThreadState"));
524         dead_thread_state->fast_state.SetIgnoreBit();
525         dead_thread_state->ignore_interceptors = 1;
526         dead_thread_state->is_dead = true;
527         *const_cast<u32*>(&dead_thread_state->tid) = -1;
528         CHECK_EQ(0, internal_mprotect(dead_thread_state, sizeof(ThreadState),
529                                       PROT_READ));
530       }
531     }
532     CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
533   }
534   return thr;
535 }
536 
537 void set_cur_thread(ThreadState *thr) {
538   *get_android_tls_ptr() = reinterpret_cast<uptr>(thr);
539 }
540 
541 void cur_thread_finalize() {
542   __sanitizer_sigset_t emptyset;
543   internal_sigfillset(&emptyset);
544   __sanitizer_sigset_t oldset;
545   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &emptyset, &oldset));
546   ThreadState* thr = reinterpret_cast<ThreadState*>(*get_android_tls_ptr());
547   if (thr != dead_thread_state) {
548     *get_android_tls_ptr() = reinterpret_cast<uptr>(dead_thread_state);
549     UnmapOrDie(thr, sizeof(ThreadState));
550   }
551   CHECK_EQ(0, internal_sigprocmask(SIG_SETMASK, &oldset, nullptr));
552 }
553 #endif  // SANITIZER_ANDROID
554 #endif  // if !SANITIZER_GO
555 
556 }  // namespace __tsan
557 
558 #endif  // SANITIZER_LINUX || SANITIZER_FREEBSD || SANITIZER_NETBSD
559