blob: 14f0c9723a3330e0807899cbec7646d28f428e99 [file] [log] [blame]
Austin Schuh36244a12019-09-21 17:52:38 -07001// Copyright 2018 The Abseil Authors.
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// https://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15// This library provides Symbolize() function that symbolizes program
16// counters to their corresponding symbol names on linux platforms.
17// This library has a minimal implementation of an ELF symbol table
18// reader (i.e. it doesn't depend on libelf, etc.).
19//
20// The algorithm used in Symbolize() is as follows.
21//
22// 1. Go through a list of maps in /proc/self/maps and find the map
23// containing the program counter.
24//
25// 2. Open the mapped file and find a regular symbol table inside.
26// Iterate over symbols in the symbol table and look for the symbol
27// containing the program counter. If such a symbol is found,
28// obtain the symbol name, and demangle the symbol if possible.
29// If the symbol isn't found in the regular symbol table (binary is
30// stripped), try the same thing with a dynamic symbol table.
31//
32// Note that Symbolize() is originally implemented to be used in
33// signal handlers, hence it doesn't use malloc() and other unsafe
34// operations. It should be both thread-safe and async-signal-safe.
35//
36// Implementation note:
37//
38// We don't use heaps but only use stacks. We want to reduce the
39// stack consumption so that the symbolizer can run on small stacks.
40//
41// Here are some numbers collected with GCC 4.1.0 on x86:
42// - sizeof(Elf32_Sym) = 16
43// - sizeof(Elf32_Shdr) = 40
44// - sizeof(Elf64_Sym) = 24
45// - sizeof(Elf64_Shdr) = 64
46//
47// This implementation is intended to be async-signal-safe but uses some
48// functions which are not guaranteed to be so, such as memchr() and
49// memmove(). We assume they are async-signal-safe.
50
51#include <dlfcn.h>
52#include <elf.h>
53#include <fcntl.h>
54#include <link.h> // For ElfW() macro.
55#include <sys/stat.h>
56#include <sys/types.h>
57#include <unistd.h>
58
59#include <algorithm>
60#include <atomic>
61#include <cerrno>
62#include <cinttypes>
63#include <climits>
64#include <cstdint>
65#include <cstdio>
66#include <cstdlib>
67#include <cstring>
68
69#include "absl/base/casts.h"
70#include "absl/base/dynamic_annotations.h"
71#include "absl/base/internal/low_level_alloc.h"
72#include "absl/base/internal/raw_logging.h"
73#include "absl/base/internal/spinlock.h"
74#include "absl/base/port.h"
75#include "absl/debugging/internal/demangle.h"
76#include "absl/debugging/internal/vdso_support.h"
77
78namespace absl {
79
80// Value of argv[0]. Used by MaybeInitializeObjFile().
81static char *argv0_value = nullptr;
82
83void InitializeSymbolizer(const char *argv0) {
84 if (argv0_value != nullptr) {
85 free(argv0_value);
86 argv0_value = nullptr;
87 }
88 if (argv0 != nullptr && argv0[0] != '\0') {
89 argv0_value = strdup(argv0);
90 }
91}
92
93namespace debugging_internal {
94namespace {
95
96// Re-runs fn until it doesn't cause EINTR.
97#define NO_INTR(fn) \
98 do { \
99 } while ((fn) < 0 && errno == EINTR)
100
101// On Linux, ELF_ST_* are defined in <linux/elf.h>. To make this portable
102// we define our own ELF_ST_BIND and ELF_ST_TYPE if not available.
103#ifndef ELF_ST_BIND
104#define ELF_ST_BIND(info) (((unsigned char)(info)) >> 4)
105#endif
106
107#ifndef ELF_ST_TYPE
108#define ELF_ST_TYPE(info) (((unsigned char)(info)) & 0xF)
109#endif
110
111// Some platforms use a special .opd section to store function pointers.
112const char kOpdSectionName[] = ".opd";
113
114#if (defined(__powerpc__) && !(_CALL_ELF > 1)) || defined(__ia64)
115// Use opd section for function descriptors on these platforms, the function
116// address is the first word of the descriptor.
117enum { kPlatformUsesOPDSections = 1 };
118#else // not PPC or IA64
119enum { kPlatformUsesOPDSections = 0 };
120#endif
121
122// This works for PowerPC & IA64 only. A function descriptor consist of two
123// pointers and the first one is the function's entry.
124const size_t kFunctionDescriptorSize = sizeof(void *) * 2;
125
126const int kMaxDecorators = 10; // Seems like a reasonable upper limit.
127
128struct InstalledSymbolDecorator {
129 SymbolDecorator fn;
130 void *arg;
131 int ticket;
132};
133
134int g_num_decorators;
135InstalledSymbolDecorator g_decorators[kMaxDecorators];
136
137struct FileMappingHint {
138 const void *start;
139 const void *end;
140 uint64_t offset;
141 const char *filename;
142};
143
144// Protects g_decorators.
145// We are using SpinLock and not a Mutex here, because we may be called
146// from inside Mutex::Lock itself, and it prohibits recursive calls.
147// This happens in e.g. base/stacktrace_syscall_unittest.
148// Moreover, we are using only TryLock(), if the decorator list
149// is being modified (is busy), we skip all decorators, and possibly
150// loose some info. Sorry, that's the best we could do.
151base_internal::SpinLock g_decorators_mu(base_internal::kLinkerInitialized);
152
153const int kMaxFileMappingHints = 8;
154int g_num_file_mapping_hints;
155FileMappingHint g_file_mapping_hints[kMaxFileMappingHints];
156// Protects g_file_mapping_hints.
157base_internal::SpinLock g_file_mapping_mu(base_internal::kLinkerInitialized);
158
159// Async-signal-safe function to zero a buffer.
160// memset() is not guaranteed to be async-signal-safe.
161static void SafeMemZero(void* p, size_t size) {
162 unsigned char *c = static_cast<unsigned char *>(p);
163 while (size--) {
164 *c++ = 0;
165 }
166}
167
168struct ObjFile {
169 ObjFile()
170 : filename(nullptr),
171 start_addr(nullptr),
172 end_addr(nullptr),
173 offset(0),
174 fd(-1),
175 elf_type(-1) {
176 SafeMemZero(&elf_header, sizeof(elf_header));
177 }
178
179 char *filename;
180 const void *start_addr;
181 const void *end_addr;
182 uint64_t offset;
183
184 // The following fields are initialized on the first access to the
185 // object file.
186 int fd;
187 int elf_type;
188 ElfW(Ehdr) elf_header;
189};
190
191// Build 4-way associative cache for symbols. Within each cache line, symbols
192// are replaced in LRU order.
193enum {
194 ASSOCIATIVITY = 4,
195};
196struct SymbolCacheLine {
197 const void *pc[ASSOCIATIVITY];
198 char *name[ASSOCIATIVITY];
199
200 // age[i] is incremented when a line is accessed. it's reset to zero if the
201 // i'th entry is read.
202 uint32_t age[ASSOCIATIVITY];
203};
204
205// ---------------------------------------------------------------
206// An async-signal-safe arena for LowLevelAlloc
207static std::atomic<base_internal::LowLevelAlloc::Arena *> g_sig_safe_arena;
208
209static base_internal::LowLevelAlloc::Arena *SigSafeArena() {
210 return g_sig_safe_arena.load(std::memory_order_acquire);
211}
212
213static void InitSigSafeArena() {
214 if (SigSafeArena() == nullptr) {
215 base_internal::LowLevelAlloc::Arena *new_arena =
216 base_internal::LowLevelAlloc::NewArena(
217 base_internal::LowLevelAlloc::kAsyncSignalSafe);
218 base_internal::LowLevelAlloc::Arena *old_value = nullptr;
219 if (!g_sig_safe_arena.compare_exchange_strong(old_value, new_arena,
220 std::memory_order_release,
221 std::memory_order_relaxed)) {
222 // We lost a race to allocate an arena; deallocate.
223 base_internal::LowLevelAlloc::DeleteArena(new_arena);
224 }
225 }
226}
227
228// ---------------------------------------------------------------
229// An AddrMap is a vector of ObjFile, using SigSafeArena() for allocation.
230
231class AddrMap {
232 public:
233 AddrMap() : size_(0), allocated_(0), obj_(nullptr) {}
234 ~AddrMap() { base_internal::LowLevelAlloc::Free(obj_); }
235 int Size() const { return size_; }
236 ObjFile *At(int i) { return &obj_[i]; }
237 ObjFile *Add();
238 void Clear();
239
240 private:
241 int size_; // count of valid elements (<= allocated_)
242 int allocated_; // count of allocated elements
243 ObjFile *obj_; // array of allocated_ elements
244 AddrMap(const AddrMap &) = delete;
245 AddrMap &operator=(const AddrMap &) = delete;
246};
247
248void AddrMap::Clear() {
249 for (int i = 0; i != size_; i++) {
250 At(i)->~ObjFile();
251 }
252 size_ = 0;
253}
254
255ObjFile *AddrMap::Add() {
256 if (size_ == allocated_) {
257 int new_allocated = allocated_ * 2 + 50;
258 ObjFile *new_obj_ =
259 static_cast<ObjFile *>(base_internal::LowLevelAlloc::AllocWithArena(
260 new_allocated * sizeof(*new_obj_), SigSafeArena()));
261 if (obj_) {
262 memcpy(new_obj_, obj_, allocated_ * sizeof(*new_obj_));
263 base_internal::LowLevelAlloc::Free(obj_);
264 }
265 obj_ = new_obj_;
266 allocated_ = new_allocated;
267 }
268 return new (&obj_[size_++]) ObjFile;
269}
270
271// ---------------------------------------------------------------
272
273enum FindSymbolResult { SYMBOL_NOT_FOUND = 1, SYMBOL_TRUNCATED, SYMBOL_FOUND };
274
275class Symbolizer {
276 public:
277 Symbolizer();
278 ~Symbolizer();
279 const char *GetSymbol(const void *const pc);
280
281 private:
282 char *CopyString(const char *s) {
283 int len = strlen(s);
284 char *dst = static_cast<char *>(
285 base_internal::LowLevelAlloc::AllocWithArena(len + 1, SigSafeArena()));
286 ABSL_RAW_CHECK(dst != nullptr, "out of memory");
287 memcpy(dst, s, len + 1);
288 return dst;
289 }
290 ObjFile *FindObjFile(const void *const start,
291 size_t size) ABSL_ATTRIBUTE_NOINLINE;
292 static bool RegisterObjFile(const char *filename,
293 const void *const start_addr,
294 const void *const end_addr, uint64_t offset,
295 void *arg);
296 SymbolCacheLine *GetCacheLine(const void *const pc);
297 const char *FindSymbolInCache(const void *const pc);
298 const char *InsertSymbolInCache(const void *const pc, const char *name);
299 void AgeSymbols(SymbolCacheLine *line);
300 void ClearAddrMap();
301 FindSymbolResult GetSymbolFromObjectFile(const ObjFile &obj,
302 const void *const pc,
303 const ptrdiff_t relocation,
304 char *out, int out_size,
305 char *tmp_buf, int tmp_buf_size);
306
307 enum {
308 SYMBOL_BUF_SIZE = 3072,
309 TMP_BUF_SIZE = 1024,
310 SYMBOL_CACHE_LINES = 128,
311 };
312
313 AddrMap addr_map_;
314
315 bool ok_;
316 bool addr_map_read_;
317
318 char symbol_buf_[SYMBOL_BUF_SIZE];
319
320 // tmp_buf_ will be used to store arrays of ElfW(Shdr) and ElfW(Sym)
321 // so we ensure that tmp_buf_ is properly aligned to store either.
322 alignas(16) char tmp_buf_[TMP_BUF_SIZE];
323 static_assert(alignof(ElfW(Shdr)) <= 16,
324 "alignment of tmp buf too small for Shdr");
325 static_assert(alignof(ElfW(Sym)) <= 16,
326 "alignment of tmp buf too small for Sym");
327
328 SymbolCacheLine symbol_cache_[SYMBOL_CACHE_LINES];
329};
330
331static std::atomic<Symbolizer *> g_cached_symbolizer;
332
333} // namespace
334
335static int SymbolizerSize() {
336#if defined(__wasm__) || defined(__asmjs__)
337 int pagesize = getpagesize();
338#else
339 int pagesize = sysconf(_SC_PAGESIZE);
340#endif
341 return ((sizeof(Symbolizer) - 1) / pagesize + 1) * pagesize;
342}
343
344// Return (and set null) g_cached_symbolized_state if it is not null.
345// Otherwise return a new symbolizer.
346static Symbolizer *AllocateSymbolizer() {
347 InitSigSafeArena();
348 Symbolizer *symbolizer =
349 g_cached_symbolizer.exchange(nullptr, std::memory_order_acquire);
350 if (symbolizer != nullptr) {
351 return symbolizer;
352 }
353 return new (base_internal::LowLevelAlloc::AllocWithArena(
354 SymbolizerSize(), SigSafeArena())) Symbolizer();
355}
356
357// Set g_cached_symbolize_state to s if it is null, otherwise
358// delete s.
359static void FreeSymbolizer(Symbolizer *s) {
360 Symbolizer *old_cached_symbolizer = nullptr;
361 if (!g_cached_symbolizer.compare_exchange_strong(old_cached_symbolizer, s,
362 std::memory_order_release,
363 std::memory_order_relaxed)) {
364 s->~Symbolizer();
365 base_internal::LowLevelAlloc::Free(s);
366 }
367}
368
369Symbolizer::Symbolizer() : ok_(true), addr_map_read_(false) {
370 for (SymbolCacheLine &symbol_cache_line : symbol_cache_) {
371 for (size_t j = 0; j < ABSL_ARRAYSIZE(symbol_cache_line.name); ++j) {
372 symbol_cache_line.pc[j] = nullptr;
373 symbol_cache_line.name[j] = nullptr;
374 symbol_cache_line.age[j] = 0;
375 }
376 }
377}
378
379Symbolizer::~Symbolizer() {
380 for (SymbolCacheLine &symbol_cache_line : symbol_cache_) {
381 for (char *s : symbol_cache_line.name) {
382 base_internal::LowLevelAlloc::Free(s);
383 }
384 }
385 ClearAddrMap();
386}
387
388// We don't use assert() since it's not guaranteed to be
389// async-signal-safe. Instead we define a minimal assertion
390// macro. So far, we don't need pretty printing for __FILE__, etc.
391#define SAFE_ASSERT(expr) ((expr) ? static_cast<void>(0) : abort())
392
393// Read up to "count" bytes from file descriptor "fd" into the buffer
394// starting at "buf" while handling short reads and EINTR. On
395// success, return the number of bytes read. Otherwise, return -1.
396static ssize_t ReadPersistent(int fd, void *buf, size_t count) {
397 SAFE_ASSERT(fd >= 0);
398 SAFE_ASSERT(count <= SSIZE_MAX);
399 char *buf0 = reinterpret_cast<char *>(buf);
400 size_t num_bytes = 0;
401 while (num_bytes < count) {
402 ssize_t len;
403 NO_INTR(len = read(fd, buf0 + num_bytes, count - num_bytes));
404 if (len < 0) { // There was an error other than EINTR.
405 ABSL_RAW_LOG(WARNING, "read failed: errno=%d", errno);
406 return -1;
407 }
408 if (len == 0) { // Reached EOF.
409 break;
410 }
411 num_bytes += len;
412 }
413 SAFE_ASSERT(num_bytes <= count);
414 return static_cast<ssize_t>(num_bytes);
415}
416
417// Read up to "count" bytes from "offset" in the file pointed by file
418// descriptor "fd" into the buffer starting at "buf". On success,
419// return the number of bytes read. Otherwise, return -1.
420static ssize_t ReadFromOffset(const int fd, void *buf, const size_t count,
421 const off_t offset) {
422 off_t off = lseek(fd, offset, SEEK_SET);
423 if (off == (off_t)-1) {
424 ABSL_RAW_LOG(WARNING, "lseek(%d, %ju, SEEK_SET) failed: errno=%d", fd,
425 static_cast<uintmax_t>(offset), errno);
426 return -1;
427 }
428 return ReadPersistent(fd, buf, count);
429}
430
431// Try reading exactly "count" bytes from "offset" bytes in a file
432// pointed by "fd" into the buffer starting at "buf" while handling
433// short reads and EINTR. On success, return true. Otherwise, return
434// false.
435static bool ReadFromOffsetExact(const int fd, void *buf, const size_t count,
436 const off_t offset) {
437 ssize_t len = ReadFromOffset(fd, buf, count, offset);
438 return len >= 0 && static_cast<size_t>(len) == count;
439}
440
441// Returns elf_header.e_type if the file pointed by fd is an ELF binary.
442static int FileGetElfType(const int fd) {
443 ElfW(Ehdr) elf_header;
444 if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
445 return -1;
446 }
447 if (memcmp(elf_header.e_ident, ELFMAG, SELFMAG) != 0) {
448 return -1;
449 }
450 return elf_header.e_type;
451}
452
453// Read the section headers in the given ELF binary, and if a section
454// of the specified type is found, set the output to this section header
455// and return true. Otherwise, return false.
456// To keep stack consumption low, we would like this function to not get
457// inlined.
458static ABSL_ATTRIBUTE_NOINLINE bool GetSectionHeaderByType(
459 const int fd, ElfW(Half) sh_num, const off_t sh_offset, ElfW(Word) type,
460 ElfW(Shdr) * out, char *tmp_buf, int tmp_buf_size) {
461 ElfW(Shdr) *buf = reinterpret_cast<ElfW(Shdr) *>(tmp_buf);
462 const int buf_entries = tmp_buf_size / sizeof(buf[0]);
463 const int buf_bytes = buf_entries * sizeof(buf[0]);
464
465 for (int i = 0; i < sh_num;) {
466 const ssize_t num_bytes_left = (sh_num - i) * sizeof(buf[0]);
467 const ssize_t num_bytes_to_read =
468 (buf_bytes > num_bytes_left) ? num_bytes_left : buf_bytes;
469 const off_t offset = sh_offset + i * sizeof(buf[0]);
470 const ssize_t len = ReadFromOffset(fd, buf, num_bytes_to_read, offset);
471 if (len % sizeof(buf[0]) != 0) {
472 ABSL_RAW_LOG(
473 WARNING,
474 "Reading %zd bytes from offset %ju returned %zd which is not a "
475 "multiple of %zu.",
476 num_bytes_to_read, static_cast<uintmax_t>(offset), len,
477 sizeof(buf[0]));
478 return false;
479 }
480 const ssize_t num_headers_in_buf = len / sizeof(buf[0]);
481 SAFE_ASSERT(num_headers_in_buf <= buf_entries);
482 for (int j = 0; j < num_headers_in_buf; ++j) {
483 if (buf[j].sh_type == type) {
484 *out = buf[j];
485 return true;
486 }
487 }
488 i += num_headers_in_buf;
489 }
490 return false;
491}
492
493// There is no particular reason to limit section name to 63 characters,
494// but there has (as yet) been no need for anything longer either.
495const int kMaxSectionNameLen = 64;
496
497bool ForEachSection(int fd,
498 const std::function<bool(const std::string &name,
499 const ElfW(Shdr) &)> &callback) {
500 ElfW(Ehdr) elf_header;
501 if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
502 return false;
503 }
504
505 ElfW(Shdr) shstrtab;
506 off_t shstrtab_offset =
507 (elf_header.e_shoff + elf_header.e_shentsize * elf_header.e_shstrndx);
508 if (!ReadFromOffsetExact(fd, &shstrtab, sizeof(shstrtab), shstrtab_offset)) {
509 return false;
510 }
511
512 for (int i = 0; i < elf_header.e_shnum; ++i) {
513 ElfW(Shdr) out;
514 off_t section_header_offset =
515 (elf_header.e_shoff + elf_header.e_shentsize * i);
516 if (!ReadFromOffsetExact(fd, &out, sizeof(out), section_header_offset)) {
517 return false;
518 }
519 off_t name_offset = shstrtab.sh_offset + out.sh_name;
520 char header_name[kMaxSectionNameLen + 1];
521 ssize_t n_read =
522 ReadFromOffset(fd, &header_name, kMaxSectionNameLen, name_offset);
523 if (n_read == -1) {
524 return false;
525 } else if (n_read > kMaxSectionNameLen) {
526 // Long read?
527 return false;
528 }
529 header_name[n_read] = '\0';
530
531 std::string name(header_name);
532 if (!callback(name, out)) {
533 break;
534 }
535 }
536 return true;
537}
538
539// name_len should include terminating '\0'.
540bool GetSectionHeaderByName(int fd, const char *name, size_t name_len,
541 ElfW(Shdr) * out) {
542 char header_name[kMaxSectionNameLen];
543 if (sizeof(header_name) < name_len) {
544 ABSL_RAW_LOG(WARNING,
545 "Section name '%s' is too long (%zu); "
546 "section will not be found (even if present).",
547 name, name_len);
548 // No point in even trying.
549 return false;
550 }
551
552 ElfW(Ehdr) elf_header;
553 if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
554 return false;
555 }
556
557 ElfW(Shdr) shstrtab;
558 off_t shstrtab_offset =
559 (elf_header.e_shoff + elf_header.e_shentsize * elf_header.e_shstrndx);
560 if (!ReadFromOffsetExact(fd, &shstrtab, sizeof(shstrtab), shstrtab_offset)) {
561 return false;
562 }
563
564 for (int i = 0; i < elf_header.e_shnum; ++i) {
565 off_t section_header_offset =
566 (elf_header.e_shoff + elf_header.e_shentsize * i);
567 if (!ReadFromOffsetExact(fd, out, sizeof(*out), section_header_offset)) {
568 return false;
569 }
570 off_t name_offset = shstrtab.sh_offset + out->sh_name;
571 ssize_t n_read = ReadFromOffset(fd, &header_name, name_len, name_offset);
572 if (n_read < 0) {
573 return false;
574 } else if (static_cast<size_t>(n_read) != name_len) {
575 // Short read -- name could be at end of file.
576 continue;
577 }
578 if (memcmp(header_name, name, name_len) == 0) {
579 return true;
580 }
581 }
582 return false;
583}
584
585// Compare symbols at in the same address.
586// Return true if we should pick symbol1.
587static bool ShouldPickFirstSymbol(const ElfW(Sym) & symbol1,
588 const ElfW(Sym) & symbol2) {
589 // If one of the symbols is weak and the other is not, pick the one
590 // this is not a weak symbol.
591 char bind1 = ELF_ST_BIND(symbol1.st_info);
592 char bind2 = ELF_ST_BIND(symbol1.st_info);
593 if (bind1 == STB_WEAK && bind2 != STB_WEAK) return false;
594 if (bind2 == STB_WEAK && bind1 != STB_WEAK) return true;
595
596 // If one of the symbols has zero size and the other is not, pick the
597 // one that has non-zero size.
598 if (symbol1.st_size != 0 && symbol2.st_size == 0) {
599 return true;
600 }
601 if (symbol1.st_size == 0 && symbol2.st_size != 0) {
602 return false;
603 }
604
605 // If one of the symbols has no type and the other is not, pick the
606 // one that has a type.
607 char type1 = ELF_ST_TYPE(symbol1.st_info);
608 char type2 = ELF_ST_TYPE(symbol1.st_info);
609 if (type1 != STT_NOTYPE && type2 == STT_NOTYPE) {
610 return true;
611 }
612 if (type1 == STT_NOTYPE && type2 != STT_NOTYPE) {
613 return false;
614 }
615
616 // Pick the first one, if we still cannot decide.
617 return true;
618}
619
620// Return true if an address is inside a section.
621static bool InSection(const void *address, const ElfW(Shdr) * section) {
622 const char *start = reinterpret_cast<const char *>(section->sh_addr);
623 size_t size = static_cast<size_t>(section->sh_size);
624 return start <= address && address < (start + size);
625}
626
627static const char *ComputeOffset(const char *base, ptrdiff_t offset) {
628 // Note: cast to uintptr_t to avoid undefined behavior when base evaluates to
629 // zero and offset is non-zero.
630 return reinterpret_cast<const char *>(
631 reinterpret_cast<uintptr_t>(base) + offset);
632}
633
634// Read a symbol table and look for the symbol containing the
635// pc. Iterate over symbols in a symbol table and look for the symbol
636// containing "pc". If the symbol is found, and its name fits in
637// out_size, the name is written into out and SYMBOL_FOUND is returned.
638// If the name does not fit, truncated name is written into out,
639// and SYMBOL_TRUNCATED is returned. Out is NUL-terminated.
640// If the symbol is not found, SYMBOL_NOT_FOUND is returned;
641// To keep stack consumption low, we would like this function to not get
642// inlined.
643static ABSL_ATTRIBUTE_NOINLINE FindSymbolResult FindSymbol(
644 const void *const pc, const int fd, char *out, int out_size,
645 ptrdiff_t relocation, const ElfW(Shdr) * strtab, const ElfW(Shdr) * symtab,
646 const ElfW(Shdr) * opd, char *tmp_buf, int tmp_buf_size) {
647 if (symtab == nullptr) {
648 return SYMBOL_NOT_FOUND;
649 }
650
651 // Read multiple symbols at once to save read() calls.
652 ElfW(Sym) *buf = reinterpret_cast<ElfW(Sym) *>(tmp_buf);
653 const int buf_entries = tmp_buf_size / sizeof(buf[0]);
654
655 const int num_symbols = symtab->sh_size / symtab->sh_entsize;
656
657 // On platforms using an .opd section (PowerPC & IA64), a function symbol
658 // has the address of a function descriptor, which contains the real
659 // starting address. However, we do not always want to use the real
660 // starting address because we sometimes want to symbolize a function
661 // pointer into the .opd section, e.g. FindSymbol(&foo,...).
662 const bool pc_in_opd =
663 kPlatformUsesOPDSections && opd != nullptr && InSection(pc, opd);
664 const bool deref_function_descriptor_pointer =
665 kPlatformUsesOPDSections && opd != nullptr && !pc_in_opd;
666
667 ElfW(Sym) best_match;
668 SafeMemZero(&best_match, sizeof(best_match));
669 bool found_match = false;
670 for (int i = 0; i < num_symbols;) {
671 off_t offset = symtab->sh_offset + i * symtab->sh_entsize;
672 const int num_remaining_symbols = num_symbols - i;
673 const int entries_in_chunk = std::min(num_remaining_symbols, buf_entries);
674 const int bytes_in_chunk = entries_in_chunk * sizeof(buf[0]);
675 const ssize_t len = ReadFromOffset(fd, buf, bytes_in_chunk, offset);
676 SAFE_ASSERT(len % sizeof(buf[0]) == 0);
677 const ssize_t num_symbols_in_buf = len / sizeof(buf[0]);
678 SAFE_ASSERT(num_symbols_in_buf <= entries_in_chunk);
679 for (int j = 0; j < num_symbols_in_buf; ++j) {
680 const ElfW(Sym) &symbol = buf[j];
681
682 // For a DSO, a symbol address is relocated by the loading address.
683 // We keep the original address for opd redirection below.
684 const char *const original_start_address =
685 reinterpret_cast<const char *>(symbol.st_value);
686 const char *start_address =
687 ComputeOffset(original_start_address, relocation);
688
689 if (deref_function_descriptor_pointer &&
690 InSection(original_start_address, opd)) {
691 // The opd section is mapped into memory. Just dereference
692 // start_address to get the first double word, which points to the
693 // function entry.
694 start_address = *reinterpret_cast<const char *const *>(start_address);
695 }
696
697 // If pc is inside the .opd section, it points to a function descriptor.
698 const size_t size = pc_in_opd ? kFunctionDescriptorSize : symbol.st_size;
699 const void *const end_address = ComputeOffset(start_address, size);
700 if (symbol.st_value != 0 && // Skip null value symbols.
701 symbol.st_shndx != 0 && // Skip undefined symbols.
702#ifdef STT_TLS
703 ELF_ST_TYPE(symbol.st_info) != STT_TLS && // Skip thread-local data.
704#endif // STT_TLS
705 ((start_address <= pc && pc < end_address) ||
706 (start_address == pc && pc == end_address))) {
707 if (!found_match || ShouldPickFirstSymbol(symbol, best_match)) {
708 found_match = true;
709 best_match = symbol;
710 }
711 }
712 }
713 i += num_symbols_in_buf;
714 }
715
716 if (found_match) {
717 const size_t off = strtab->sh_offset + best_match.st_name;
718 const ssize_t n_read = ReadFromOffset(fd, out, out_size, off);
719 if (n_read <= 0) {
720 // This should never happen.
721 ABSL_RAW_LOG(WARNING,
722 "Unable to read from fd %d at offset %zu: n_read = %zd", fd,
723 off, n_read);
724 return SYMBOL_NOT_FOUND;
725 }
726 ABSL_RAW_CHECK(n_read <= out_size, "ReadFromOffset read too much data.");
727
728 // strtab->sh_offset points into .strtab-like section that contains
729 // NUL-terminated strings: '\0foo\0barbaz\0...".
730 //
731 // sh_offset+st_name points to the start of symbol name, but we don't know
732 // how long the symbol is, so we try to read as much as we have space for,
733 // and usually over-read (i.e. there is a NUL somewhere before n_read).
734 if (memchr(out, '\0', n_read) == nullptr) {
735 // Either out_size was too small (n_read == out_size and no NUL), or
736 // we tried to read past the EOF (n_read < out_size) and .strtab is
737 // corrupt (missing terminating NUL; should never happen for valid ELF).
738 out[n_read - 1] = '\0';
739 return SYMBOL_TRUNCATED;
740 }
741 return SYMBOL_FOUND;
742 }
743
744 return SYMBOL_NOT_FOUND;
745}
746
747// Get the symbol name of "pc" from the file pointed by "fd". Process
748// both regular and dynamic symbol tables if necessary.
749// See FindSymbol() comment for description of return value.
750FindSymbolResult Symbolizer::GetSymbolFromObjectFile(
751 const ObjFile &obj, const void *const pc, const ptrdiff_t relocation,
752 char *out, int out_size, char *tmp_buf, int tmp_buf_size) {
753 ElfW(Shdr) symtab;
754 ElfW(Shdr) strtab;
755 ElfW(Shdr) opd;
756 ElfW(Shdr) *opd_ptr = nullptr;
757
758 // On platforms using an .opd sections for function descriptor, read
759 // the section header. The .opd section is in data segment and should be
760 // loaded but we check that it is mapped just to be extra careful.
761 if (kPlatformUsesOPDSections) {
762 if (GetSectionHeaderByName(obj.fd, kOpdSectionName,
763 sizeof(kOpdSectionName) - 1, &opd) &&
764 FindObjFile(reinterpret_cast<const char *>(opd.sh_addr) + relocation,
765 opd.sh_size) != nullptr) {
766 opd_ptr = &opd;
767 } else {
768 return SYMBOL_NOT_FOUND;
769 }
770 }
771
772 // Consult a regular symbol table, then fall back to the dynamic symbol table.
773 for (const auto symbol_table_type : {SHT_SYMTAB, SHT_DYNSYM}) {
774 if (!GetSectionHeaderByType(obj.fd, obj.elf_header.e_shnum,
775 obj.elf_header.e_shoff, symbol_table_type,
776 &symtab, tmp_buf, tmp_buf_size)) {
777 continue;
778 }
779 if (!ReadFromOffsetExact(
780 obj.fd, &strtab, sizeof(strtab),
781 obj.elf_header.e_shoff + symtab.sh_link * sizeof(symtab))) {
782 continue;
783 }
784 const FindSymbolResult rc =
785 FindSymbol(pc, obj.fd, out, out_size, relocation, &strtab, &symtab,
786 opd_ptr, tmp_buf, tmp_buf_size);
787 if (rc != SYMBOL_NOT_FOUND) {
788 return rc;
789 }
790 }
791
792 return SYMBOL_NOT_FOUND;
793}
794
795namespace {
796// Thin wrapper around a file descriptor so that the file descriptor
797// gets closed for sure.
798class FileDescriptor {
799 public:
800 explicit FileDescriptor(int fd) : fd_(fd) {}
801 FileDescriptor(const FileDescriptor &) = delete;
802 FileDescriptor &operator=(const FileDescriptor &) = delete;
803
804 ~FileDescriptor() {
805 if (fd_ >= 0) {
806 NO_INTR(close(fd_));
807 }
808 }
809
810 int get() const { return fd_; }
811
812 private:
813 const int fd_;
814};
815
816// Helper class for reading lines from file.
817//
818// Note: we don't use ProcMapsIterator since the object is big (it has
819// a 5k array member) and uses async-unsafe functions such as sscanf()
820// and snprintf().
821class LineReader {
822 public:
823 explicit LineReader(int fd, char *buf, int buf_len)
824 : fd_(fd),
825 buf_len_(buf_len),
826 buf_(buf),
827 bol_(buf),
828 eol_(buf),
829 eod_(buf) {}
830
831 LineReader(const LineReader &) = delete;
832 LineReader &operator=(const LineReader &) = delete;
833
834 // Read '\n'-terminated line from file. On success, modify "bol"
835 // and "eol", then return true. Otherwise, return false.
836 //
837 // Note: if the last line doesn't end with '\n', the line will be
838 // dropped. It's an intentional behavior to make the code simple.
839 bool ReadLine(const char **bol, const char **eol) {
840 if (BufferIsEmpty()) { // First time.
841 const ssize_t num_bytes = ReadPersistent(fd_, buf_, buf_len_);
842 if (num_bytes <= 0) { // EOF or error.
843 return false;
844 }
845 eod_ = buf_ + num_bytes;
846 bol_ = buf_;
847 } else {
848 bol_ = eol_ + 1; // Advance to the next line in the buffer.
849 SAFE_ASSERT(bol_ <= eod_); // "bol_" can point to "eod_".
850 if (!HasCompleteLine()) {
851 const int incomplete_line_length = eod_ - bol_;
852 // Move the trailing incomplete line to the beginning.
853 memmove(buf_, bol_, incomplete_line_length);
854 // Read text from file and append it.
855 char *const append_pos = buf_ + incomplete_line_length;
856 const int capacity_left = buf_len_ - incomplete_line_length;
857 const ssize_t num_bytes =
858 ReadPersistent(fd_, append_pos, capacity_left);
859 if (num_bytes <= 0) { // EOF or error.
860 return false;
861 }
862 eod_ = append_pos + num_bytes;
863 bol_ = buf_;
864 }
865 }
866 eol_ = FindLineFeed();
867 if (eol_ == nullptr) { // '\n' not found. Malformed line.
868 return false;
869 }
870 *eol_ = '\0'; // Replace '\n' with '\0'.
871
872 *bol = bol_;
873 *eol = eol_;
874 return true;
875 }
876
877 private:
878 char *FindLineFeed() const {
879 return reinterpret_cast<char *>(memchr(bol_, '\n', eod_ - bol_));
880 }
881
882 bool BufferIsEmpty() const { return buf_ == eod_; }
883
884 bool HasCompleteLine() const {
885 return !BufferIsEmpty() && FindLineFeed() != nullptr;
886 }
887
888 const int fd_;
889 const int buf_len_;
890 char *const buf_;
891 char *bol_;
892 char *eol_;
893 const char *eod_; // End of data in "buf_".
894};
895} // namespace
896
897// Place the hex number read from "start" into "*hex". The pointer to
898// the first non-hex character or "end" is returned.
899static const char *GetHex(const char *start, const char *end,
900 uint64_t *const value) {
901 uint64_t hex = 0;
902 const char *p;
903 for (p = start; p < end; ++p) {
904 int ch = *p;
905 if ((ch >= '0' && ch <= '9') || (ch >= 'A' && ch <= 'F') ||
906 (ch >= 'a' && ch <= 'f')) {
907 hex = (hex << 4) | (ch < 'A' ? ch - '0' : (ch & 0xF) + 9);
908 } else { // Encountered the first non-hex character.
909 break;
910 }
911 }
912 SAFE_ASSERT(p <= end);
913 *value = hex;
914 return p;
915}
916
917static const char *GetHex(const char *start, const char *end,
918 const void **const addr) {
919 uint64_t hex = 0;
920 const char *p = GetHex(start, end, &hex);
921 *addr = reinterpret_cast<void *>(hex);
922 return p;
923}
924
925// Normally we are only interested in "r?x" maps.
926// On the PowerPC, function pointers point to descriptors in the .opd
927// section. The descriptors themselves are not executable code, so
928// we need to relax the check below to "r??".
929static bool ShouldUseMapping(const char *const flags) {
930 return flags[0] == 'r' && (kPlatformUsesOPDSections || flags[2] == 'x');
931}
932
933// Read /proc/self/maps and run "callback" for each mmapped file found. If
934// "callback" returns false, stop scanning and return true. Else continue
935// scanning /proc/self/maps. Return true if no parse error is found.
936static ABSL_ATTRIBUTE_NOINLINE bool ReadAddrMap(
937 bool (*callback)(const char *filename, const void *const start_addr,
938 const void *const end_addr, uint64_t offset, void *arg),
939 void *arg, void *tmp_buf, int tmp_buf_size) {
940 // Use /proc/self/task/<pid>/maps instead of /proc/self/maps. The latter
941 // requires kernel to stop all threads, and is significantly slower when there
942 // are 1000s of threads.
943 char maps_path[80];
944 snprintf(maps_path, sizeof(maps_path), "/proc/self/task/%d/maps", getpid());
945
946 int maps_fd;
947 NO_INTR(maps_fd = open(maps_path, O_RDONLY));
948 FileDescriptor wrapped_maps_fd(maps_fd);
949 if (wrapped_maps_fd.get() < 0) {
950 ABSL_RAW_LOG(WARNING, "%s: errno=%d", maps_path, errno);
951 return false;
952 }
953
954 // Iterate over maps and look for the map containing the pc. Then
955 // look into the symbol tables inside.
956 LineReader reader(wrapped_maps_fd.get(), static_cast<char *>(tmp_buf),
957 tmp_buf_size);
958 while (true) {
959 const char *cursor;
960 const char *eol;
961 if (!reader.ReadLine(&cursor, &eol)) { // EOF or malformed line.
962 break;
963 }
964
965 const char *line = cursor;
966 const void *start_address;
967 // Start parsing line in /proc/self/maps. Here is an example:
968 //
969 // 08048000-0804c000 r-xp 00000000 08:01 2142121 /bin/cat
970 //
971 // We want start address (08048000), end address (0804c000), flags
972 // (r-xp) and file name (/bin/cat).
973
974 // Read start address.
975 cursor = GetHex(cursor, eol, &start_address);
976 if (cursor == eol || *cursor != '-') {
977 ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps line: %s", line);
978 return false;
979 }
980 ++cursor; // Skip '-'.
981
982 // Read end address.
983 const void *end_address;
984 cursor = GetHex(cursor, eol, &end_address);
985 if (cursor == eol || *cursor != ' ') {
986 ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps line: %s", line);
987 return false;
988 }
989 ++cursor; // Skip ' '.
990
991 // Read flags. Skip flags until we encounter a space or eol.
992 const char *const flags_start = cursor;
993 while (cursor < eol && *cursor != ' ') {
994 ++cursor;
995 }
996 // We expect at least four letters for flags (ex. "r-xp").
997 if (cursor == eol || cursor < flags_start + 4) {
998 ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps: %s", line);
999 return false;
1000 }
1001
1002 // Check flags.
1003 if (!ShouldUseMapping(flags_start)) {
1004 continue; // We skip this map.
1005 }
1006 ++cursor; // Skip ' '.
1007
1008 // Read file offset.
1009 uint64_t offset;
1010 cursor = GetHex(cursor, eol, &offset);
1011 ++cursor; // Skip ' '.
1012
1013 // Skip to file name. "cursor" now points to dev. We need to skip at least
1014 // two spaces for dev and inode.
1015 int num_spaces = 0;
1016 while (cursor < eol) {
1017 if (*cursor == ' ') {
1018 ++num_spaces;
1019 } else if (num_spaces >= 2) {
1020 // The first non-space character after skipping two spaces
1021 // is the beginning of the file name.
1022 break;
1023 }
1024 ++cursor;
1025 }
1026
1027 // Check whether this entry corresponds to our hint table for the true
1028 // filename.
1029 bool hinted =
1030 GetFileMappingHint(&start_address, &end_address, &offset, &cursor);
1031 if (!hinted && (cursor == eol || cursor[0] == '[')) {
1032 // not an object file, typically [vdso] or [vsyscall]
1033 continue;
1034 }
1035 if (!callback(cursor, start_address, end_address, offset, arg)) break;
1036 }
1037 return true;
1038}
1039
1040// Find the objfile mapped in address region containing [addr, addr + len).
1041ObjFile *Symbolizer::FindObjFile(const void *const addr, size_t len) {
1042 for (int i = 0; i < 2; ++i) {
1043 if (!ok_) return nullptr;
1044
1045 // Read /proc/self/maps if necessary
1046 if (!addr_map_read_) {
1047 addr_map_read_ = true;
1048 if (!ReadAddrMap(RegisterObjFile, this, tmp_buf_, TMP_BUF_SIZE)) {
1049 ok_ = false;
1050 return nullptr;
1051 }
1052 }
1053
1054 int lo = 0;
1055 int hi = addr_map_.Size();
1056 while (lo < hi) {
1057 int mid = (lo + hi) / 2;
1058 if (addr < addr_map_.At(mid)->end_addr) {
1059 hi = mid;
1060 } else {
1061 lo = mid + 1;
1062 }
1063 }
1064 if (lo != addr_map_.Size()) {
1065 ObjFile *obj = addr_map_.At(lo);
1066 SAFE_ASSERT(obj->end_addr > addr);
1067 if (addr >= obj->start_addr &&
1068 reinterpret_cast<const char *>(addr) + len <= obj->end_addr)
1069 return obj;
1070 }
1071
1072 // The address mapping may have changed since it was last read. Retry.
1073 ClearAddrMap();
1074 }
1075 return nullptr;
1076}
1077
1078void Symbolizer::ClearAddrMap() {
1079 for (int i = 0; i != addr_map_.Size(); i++) {
1080 ObjFile *o = addr_map_.At(i);
1081 base_internal::LowLevelAlloc::Free(o->filename);
1082 if (o->fd >= 0) {
1083 NO_INTR(close(o->fd));
1084 }
1085 }
1086 addr_map_.Clear();
1087 addr_map_read_ = false;
1088}
1089
1090// Callback for ReadAddrMap to register objfiles in an in-memory table.
1091bool Symbolizer::RegisterObjFile(const char *filename,
1092 const void *const start_addr,
1093 const void *const end_addr, uint64_t offset,
1094 void *arg) {
1095 Symbolizer *impl = static_cast<Symbolizer *>(arg);
1096
1097 // Files are supposed to be added in the increasing address order. Make
1098 // sure that's the case.
1099 int addr_map_size = impl->addr_map_.Size();
1100 if (addr_map_size != 0) {
1101 ObjFile *old = impl->addr_map_.At(addr_map_size - 1);
1102 if (old->end_addr > end_addr) {
1103 ABSL_RAW_LOG(ERROR,
1104 "Unsorted addr map entry: 0x%" PRIxPTR ": %s <-> 0x%" PRIxPTR
1105 ": %s",
1106 reinterpret_cast<uintptr_t>(end_addr), filename,
1107 reinterpret_cast<uintptr_t>(old->end_addr), old->filename);
1108 return true;
1109 } else if (old->end_addr == end_addr) {
1110 // The same entry appears twice. This sometimes happens for [vdso].
1111 if (old->start_addr != start_addr ||
1112 strcmp(old->filename, filename) != 0) {
1113 ABSL_RAW_LOG(ERROR,
1114 "Duplicate addr 0x%" PRIxPTR ": %s <-> 0x%" PRIxPTR ": %s",
1115 reinterpret_cast<uintptr_t>(end_addr), filename,
1116 reinterpret_cast<uintptr_t>(old->end_addr), old->filename);
1117 }
1118 return true;
1119 }
1120 }
1121 ObjFile *obj = impl->addr_map_.Add();
1122 obj->filename = impl->CopyString(filename);
1123 obj->start_addr = start_addr;
1124 obj->end_addr = end_addr;
1125 obj->offset = offset;
1126 obj->elf_type = -1; // filled on demand
1127 obj->fd = -1; // opened on demand
1128 return true;
1129}
1130
1131// This function wraps the Demangle function to provide an interface
1132// where the input symbol is demangled in-place.
1133// To keep stack consumption low, we would like this function to not
1134// get inlined.
1135static ABSL_ATTRIBUTE_NOINLINE void DemangleInplace(char *out, int out_size,
1136 char *tmp_buf,
1137 int tmp_buf_size) {
1138 if (Demangle(out, tmp_buf, tmp_buf_size)) {
1139 // Demangling succeeded. Copy to out if the space allows.
1140 int len = strlen(tmp_buf);
1141 if (len + 1 <= out_size) { // +1 for '\0'.
1142 SAFE_ASSERT(len < tmp_buf_size);
1143 memmove(out, tmp_buf, len + 1);
1144 }
1145 }
1146}
1147
1148SymbolCacheLine *Symbolizer::GetCacheLine(const void *const pc) {
1149 uintptr_t pc0 = reinterpret_cast<uintptr_t>(pc);
1150 pc0 >>= 3; // drop the low 3 bits
1151
1152 // Shuffle bits.
1153 pc0 ^= (pc0 >> 6) ^ (pc0 >> 12) ^ (pc0 >> 18);
1154 return &symbol_cache_[pc0 % SYMBOL_CACHE_LINES];
1155}
1156
1157void Symbolizer::AgeSymbols(SymbolCacheLine *line) {
1158 for (uint32_t &age : line->age) {
1159 ++age;
1160 }
1161}
1162
1163const char *Symbolizer::FindSymbolInCache(const void *const pc) {
1164 if (pc == nullptr) return nullptr;
1165
1166 SymbolCacheLine *line = GetCacheLine(pc);
1167 for (size_t i = 0; i < ABSL_ARRAYSIZE(line->pc); ++i) {
1168 if (line->pc[i] == pc) {
1169 AgeSymbols(line);
1170 line->age[i] = 0;
1171 return line->name[i];
1172 }
1173 }
1174 return nullptr;
1175}
1176
1177const char *Symbolizer::InsertSymbolInCache(const void *const pc,
1178 const char *name) {
1179 SAFE_ASSERT(pc != nullptr);
1180
1181 SymbolCacheLine *line = GetCacheLine(pc);
1182 uint32_t max_age = 0;
1183 int oldest_index = -1;
1184 for (size_t i = 0; i < ABSL_ARRAYSIZE(line->pc); ++i) {
1185 if (line->pc[i] == nullptr) {
1186 AgeSymbols(line);
1187 line->pc[i] = pc;
1188 line->name[i] = CopyString(name);
1189 line->age[i] = 0;
1190 return line->name[i];
1191 }
1192 if (line->age[i] >= max_age) {
1193 max_age = line->age[i];
1194 oldest_index = i;
1195 }
1196 }
1197
1198 AgeSymbols(line);
1199 ABSL_RAW_CHECK(oldest_index >= 0, "Corrupt cache");
1200 base_internal::LowLevelAlloc::Free(line->name[oldest_index]);
1201 line->pc[oldest_index] = pc;
1202 line->name[oldest_index] = CopyString(name);
1203 line->age[oldest_index] = 0;
1204 return line->name[oldest_index];
1205}
1206
1207static void MaybeOpenFdFromSelfExe(ObjFile *obj) {
1208 if (memcmp(obj->start_addr, ELFMAG, SELFMAG) != 0) {
1209 return;
1210 }
1211 int fd = open("/proc/self/exe", O_RDONLY);
1212 if (fd == -1) {
1213 return;
1214 }
1215 // Verify that contents of /proc/self/exe matches in-memory image of
1216 // the binary. This can fail if the "deleted" binary is in fact not
1217 // the main executable, or for binaries that have the first PT_LOAD
1218 // segment smaller than 4K. We do it in four steps so that the
1219 // buffer is smaller and we don't consume too much stack space.
1220 const char *mem = reinterpret_cast<const char *>(obj->start_addr);
1221 for (int i = 0; i < 4; ++i) {
1222 char buf[1024];
1223 ssize_t n = read(fd, buf, sizeof(buf));
1224 if (n != sizeof(buf) || memcmp(buf, mem, sizeof(buf)) != 0) {
1225 close(fd);
1226 return;
1227 }
1228 mem += sizeof(buf);
1229 }
1230 obj->fd = fd;
1231}
1232
1233static bool MaybeInitializeObjFile(ObjFile *obj) {
1234 if (obj->fd < 0) {
1235 obj->fd = open(obj->filename, O_RDONLY);
1236
1237 if (obj->fd < 0) {
1238 // Getting /proc/self/exe here means that we were hinted.
1239 if (strcmp(obj->filename, "/proc/self/exe") == 0) {
1240 // /proc/self/exe may be inaccessible (due to setuid, etc.), so try
1241 // accessing the binary via argv0.
1242 if (argv0_value != nullptr) {
1243 obj->fd = open(argv0_value, O_RDONLY);
1244 }
1245 } else {
1246 MaybeOpenFdFromSelfExe(obj);
1247 }
1248 }
1249
1250 if (obj->fd < 0) {
1251 ABSL_RAW_LOG(WARNING, "%s: open failed: errno=%d", obj->filename, errno);
1252 return false;
1253 }
1254 obj->elf_type = FileGetElfType(obj->fd);
1255 if (obj->elf_type < 0) {
1256 ABSL_RAW_LOG(WARNING, "%s: wrong elf type: %d", obj->filename,
1257 obj->elf_type);
1258 return false;
1259 }
1260
1261 if (!ReadFromOffsetExact(obj->fd, &obj->elf_header, sizeof(obj->elf_header),
1262 0)) {
1263 ABSL_RAW_LOG(WARNING, "%s: failed to read elf header", obj->filename);
1264 return false;
1265 }
1266 }
1267 return true;
1268}
1269
1270// The implementation of our symbolization routine. If it
1271// successfully finds the symbol containing "pc" and obtains the
1272// symbol name, returns pointer to that symbol. Otherwise, returns nullptr.
1273// If any symbol decorators have been installed via InstallSymbolDecorator(),
1274// they are called here as well.
1275// To keep stack consumption low, we would like this function to not
1276// get inlined.
1277const char *Symbolizer::GetSymbol(const void *const pc) {
1278 const char *entry = FindSymbolInCache(pc);
1279 if (entry != nullptr) {
1280 return entry;
1281 }
1282 symbol_buf_[0] = '\0';
1283
1284 ObjFile *const obj = FindObjFile(pc, 1);
1285 ptrdiff_t relocation = 0;
1286 int fd = -1;
1287 if (obj != nullptr) {
1288 if (MaybeInitializeObjFile(obj)) {
1289 if (obj->elf_type == ET_DYN &&
1290 reinterpret_cast<uint64_t>(obj->start_addr) >= obj->offset) {
1291 // This object was relocated.
1292 //
1293 // For obj->offset > 0, adjust the relocation since a mapping at offset
1294 // X in the file will have a start address of [true relocation]+X.
1295 relocation = reinterpret_cast<ptrdiff_t>(obj->start_addr) - obj->offset;
1296 }
1297
1298 fd = obj->fd;
1299 }
1300 if (GetSymbolFromObjectFile(*obj, pc, relocation, symbol_buf_,
1301 sizeof(symbol_buf_), tmp_buf_,
1302 sizeof(tmp_buf_)) == SYMBOL_FOUND) {
1303 // Only try to demangle the symbol name if it fit into symbol_buf_.
1304 DemangleInplace(symbol_buf_, sizeof(symbol_buf_), tmp_buf_,
1305 sizeof(tmp_buf_));
1306 }
1307 } else {
1308#if ABSL_HAVE_VDSO_SUPPORT
1309 VDSOSupport vdso;
1310 if (vdso.IsPresent()) {
1311 VDSOSupport::SymbolInfo symbol_info;
1312 if (vdso.LookupSymbolByAddress(pc, &symbol_info)) {
1313 // All VDSO symbols are known to be short.
1314 size_t len = strlen(symbol_info.name);
1315 ABSL_RAW_CHECK(len + 1 < sizeof(symbol_buf_),
1316 "VDSO symbol unexpectedly long");
1317 memcpy(symbol_buf_, symbol_info.name, len + 1);
1318 }
1319 }
1320#endif
1321 }
1322
1323 if (g_decorators_mu.TryLock()) {
1324 if (g_num_decorators > 0) {
1325 SymbolDecoratorArgs decorator_args = {
1326 pc, relocation, fd, symbol_buf_, sizeof(symbol_buf_),
1327 tmp_buf_, sizeof(tmp_buf_), nullptr};
1328 for (int i = 0; i < g_num_decorators; ++i) {
1329 decorator_args.arg = g_decorators[i].arg;
1330 g_decorators[i].fn(&decorator_args);
1331 }
1332 }
1333 g_decorators_mu.Unlock();
1334 }
1335 if (symbol_buf_[0] == '\0') {
1336 return nullptr;
1337 }
1338 symbol_buf_[sizeof(symbol_buf_) - 1] = '\0'; // Paranoia.
1339 return InsertSymbolInCache(pc, symbol_buf_);
1340}
1341
1342bool RemoveAllSymbolDecorators(void) {
1343 if (!g_decorators_mu.TryLock()) {
1344 // Someone else is using decorators. Get out.
1345 return false;
1346 }
1347 g_num_decorators = 0;
1348 g_decorators_mu.Unlock();
1349 return true;
1350}
1351
1352bool RemoveSymbolDecorator(int ticket) {
1353 if (!g_decorators_mu.TryLock()) {
1354 // Someone else is using decorators. Get out.
1355 return false;
1356 }
1357 for (int i = 0; i < g_num_decorators; ++i) {
1358 if (g_decorators[i].ticket == ticket) {
1359 while (i < g_num_decorators - 1) {
1360 g_decorators[i] = g_decorators[i + 1];
1361 ++i;
1362 }
1363 g_num_decorators = i;
1364 break;
1365 }
1366 }
1367 g_decorators_mu.Unlock();
1368 return true; // Decorator is known to be removed.
1369}
1370
1371int InstallSymbolDecorator(SymbolDecorator decorator, void *arg) {
1372 static int ticket = 0;
1373
1374 if (!g_decorators_mu.TryLock()) {
1375 // Someone else is using decorators. Get out.
1376 return false;
1377 }
1378 int ret = ticket;
1379 if (g_num_decorators >= kMaxDecorators) {
1380 ret = -1;
1381 } else {
1382 g_decorators[g_num_decorators] = {decorator, arg, ticket++};
1383 ++g_num_decorators;
1384 }
1385 g_decorators_mu.Unlock();
1386 return ret;
1387}
1388
1389bool RegisterFileMappingHint(const void *start, const void *end, uint64_t offset,
1390 const char *filename) {
1391 SAFE_ASSERT(start <= end);
1392 SAFE_ASSERT(filename != nullptr);
1393
1394 InitSigSafeArena();
1395
1396 if (!g_file_mapping_mu.TryLock()) {
1397 return false;
1398 }
1399
1400 bool ret = true;
1401 if (g_num_file_mapping_hints >= kMaxFileMappingHints) {
1402 ret = false;
1403 } else {
1404 // TODO(ckennelly): Move this into a std::string copy routine.
1405 int len = strlen(filename);
1406 char *dst = static_cast<char *>(
1407 base_internal::LowLevelAlloc::AllocWithArena(len + 1, SigSafeArena()));
1408 ABSL_RAW_CHECK(dst != nullptr, "out of memory");
1409 memcpy(dst, filename, len + 1);
1410
1411 auto &hint = g_file_mapping_hints[g_num_file_mapping_hints++];
1412 hint.start = start;
1413 hint.end = end;
1414 hint.offset = offset;
1415 hint.filename = dst;
1416 }
1417
1418 g_file_mapping_mu.Unlock();
1419 return ret;
1420}
1421
1422bool GetFileMappingHint(const void **start, const void **end, uint64_t *offset,
1423 const char **filename) {
1424 if (!g_file_mapping_mu.TryLock()) {
1425 return false;
1426 }
1427 bool found = false;
1428 for (int i = 0; i < g_num_file_mapping_hints; i++) {
1429 if (g_file_mapping_hints[i].start <= *start &&
1430 *end <= g_file_mapping_hints[i].end) {
1431 // We assume that the start_address for the mapping is the base
1432 // address of the ELF section, but when [start_address,end_address) is
1433 // not strictly equal to [hint.start, hint.end), that assumption is
1434 // invalid.
1435 //
1436 // This uses the hint's start address (even though hint.start is not
1437 // necessarily equal to start_address) to ensure the correct
1438 // relocation is computed later.
1439 *start = g_file_mapping_hints[i].start;
1440 *end = g_file_mapping_hints[i].end;
1441 *offset = g_file_mapping_hints[i].offset;
1442 *filename = g_file_mapping_hints[i].filename;
1443 found = true;
1444 break;
1445 }
1446 }
1447 g_file_mapping_mu.Unlock();
1448 return found;
1449}
1450
1451} // namespace debugging_internal
1452
1453bool Symbolize(const void *pc, char *out, int out_size) {
1454 // Symbolization is very slow under tsan.
1455 ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN();
1456 SAFE_ASSERT(out_size >= 0);
1457 debugging_internal::Symbolizer *s = debugging_internal::AllocateSymbolizer();
1458 const char *name = s->GetSymbol(pc);
1459 bool ok = false;
1460 if (name != nullptr && out_size > 0) {
1461 strncpy(out, name, out_size);
1462 ok = true;
1463 if (out[out_size - 1] != '\0') {
1464 // strncpy() does not '\0' terminate when it truncates. Do so, with
1465 // trailing ellipsis.
1466 static constexpr char kEllipsis[] = "...";
1467 int ellipsis_size =
1468 std::min(implicit_cast<int>(strlen(kEllipsis)), out_size - 1);
1469 memcpy(out + out_size - ellipsis_size - 1, kEllipsis, ellipsis_size);
1470 out[out_size - 1] = '\0';
1471 }
1472 }
1473 debugging_internal::FreeSymbolizer(s);
1474 ANNOTATE_IGNORE_READS_AND_WRITES_END();
1475 return ok;
1476}
1477
1478} // namespace absl