ltrace-elf.c revision 97d13666cd84589135ba593fa43a800d098026d0
1/* 2 * This file is part of ltrace. 3 * Copyright (C) 2006,2010,2011,2012,2013 Petr Machata, Red Hat Inc. 4 * Copyright (C) 2010 Zachary T Welch, CodeSourcery 5 * Copyright (C) 2010 Joe Damato 6 * Copyright (C) 1997,1998,2001,2004,2007,2008,2009 Juan Cespedes 7 * Copyright (C) 2006 Olaf Hering, SUSE Linux GmbH 8 * Copyright (C) 2006 Eric Vaitl, Cisco Systems, Inc. 9 * Copyright (C) 2006 Paul Gilliam, IBM Corporation 10 * Copyright (C) 2006 Ian Wienand 11 * 12 * This program is free software; you can redistribute it and/or 13 * modify it under the terms of the GNU General Public License as 14 * published by the Free Software Foundation; either version 2 of the 15 * License, or (at your option) any later version. 16 * 17 * This program is distributed in the hope that it will be useful, but 18 * WITHOUT ANY WARRANTY; without even the implied warranty of 19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 20 * General Public License for more details. 21 * 22 * You should have received a copy of the GNU General Public License 23 * along with this program; if not, write to the Free Software 24 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 25 * 02110-1301 USA 26 */ 27 28#include "config.h" 29 30#include <assert.h> 31#ifdef __linux__ 32#include <endian.h> 33#endif 34#include <errno.h> 35#include <fcntl.h> 36#include <gelf.h> 37#include <inttypes.h> 38#include <search.h> 39#include <stdbool.h> 40#include <stdint.h> 41#include <stdio.h> 42#include <stdlib.h> 43#include <string.h> 44#include <strings.h> 45#include <unistd.h> 46 47#include "backend.h" 48#include "filter.h" 49#include "library.h" 50#include "ltrace-elf.h" 51#include "proc.h" 52#include "debug.h" 53#include "options.h" 54 55#ifndef ARCH_HAVE_LTELF_DATA 56int 57arch_elf_init(struct ltelf *lte, struct library *lib) 58{ 59 return 0; 60} 61 62void 63arch_elf_destroy(struct ltelf *lte) 64{ 65} 66#endif 67 68#ifndef OS_HAVE_ADD_PLT_ENTRY 69enum plt_status 70os_elf_add_plt_entry(struct process *proc, struct ltelf *lte, 71 const char *a_name, GElf_Rela *rela, size_t ndx, 72 struct library_symbol **ret) 73{ 74 return PLT_DEFAULT; 75} 76#endif 77 78#ifndef ARCH_HAVE_ADD_PLT_ENTRY 79enum plt_status 80arch_elf_add_plt_entry(struct process *proc, struct ltelf *lte, 81 const char *a_name, GElf_Rela *rela, size_t ndx, 82 struct library_symbol **ret) 83{ 84 return PLT_DEFAULT; 85} 86#endif 87 88#ifndef OS_HAVE_ADD_FUNC_ENTRY 89enum plt_status 90os_elf_add_func_entry(struct process *proc, struct ltelf *lte, 91 const GElf_Sym *sym, 92 arch_addr_t addr, const char *name, 93 struct library_symbol **ret) 94{ 95 if (GELF_ST_TYPE(sym->st_info) != STT_FUNC) { 96 *ret = NULL; 97 return PLT_OK; 98 } else { 99 return PLT_DEFAULT; 100 } 101} 102#endif 103 104#ifndef ARCH_HAVE_ADD_FUNC_ENTRY 105enum plt_status 106arch_elf_add_func_entry(struct process *proc, struct ltelf *lte, 107 const GElf_Sym *sym, 108 arch_addr_t addr, const char *name, 109 struct library_symbol **ret) 110{ 111 return PLT_DEFAULT; 112} 113#endif 114 115Elf_Data * 116elf_loaddata(Elf_Scn *scn, GElf_Shdr *shdr) 117{ 118 Elf_Data *data = elf_getdata(scn, NULL); 119 if (data == NULL || elf_getdata(scn, data) != NULL 120 || data->d_off || data->d_size != shdr->sh_size) 121 return NULL; 122 return data; 123} 124 125static int 126elf_get_section_if(struct ltelf *lte, Elf_Scn **tgt_sec, GElf_Shdr *tgt_shdr, 127 int (*predicate)(Elf_Scn *, GElf_Shdr *, void *data), 128 void *data) 129{ 130 int i; 131 for (i = 1; i < lte->ehdr.e_shnum; ++i) { 132 Elf_Scn *scn; 133 GElf_Shdr shdr; 134 135 scn = elf_getscn(lte->elf, i); 136 if (scn == NULL || gelf_getshdr(scn, &shdr) == NULL) { 137 debug(1, "Couldn't read section or header."); 138 return -1; 139 } 140 if (predicate(scn, &shdr, data)) { 141 *tgt_sec = scn; 142 *tgt_shdr = shdr; 143 return 0; 144 } 145 } 146 147 *tgt_sec = NULL; 148 return 0; 149} 150 151static int 152inside_p(Elf_Scn *scn, GElf_Shdr *shdr, void *data) 153{ 154 GElf_Addr addr = *(GElf_Addr *)data; 155 return addr >= shdr->sh_addr 156 && addr < shdr->sh_addr + shdr->sh_size; 157} 158 159int 160elf_get_section_covering(struct ltelf *lte, GElf_Addr addr, 161 Elf_Scn **tgt_sec, GElf_Shdr *tgt_shdr) 162{ 163 return elf_get_section_if(lte, tgt_sec, tgt_shdr, 164 &inside_p, &addr); 165} 166 167static int 168type_p(Elf_Scn *scn, GElf_Shdr *shdr, void *data) 169{ 170 GElf_Word type = *(GElf_Word *)data; 171 return shdr->sh_type == type; 172} 173 174int 175elf_get_section_type(struct ltelf *lte, GElf_Word type, 176 Elf_Scn **tgt_sec, GElf_Shdr *tgt_shdr) 177{ 178 return elf_get_section_if(lte, tgt_sec, tgt_shdr, 179 &type_p, &type); 180} 181 182struct section_named_data { 183 struct ltelf *lte; 184 const char *name; 185}; 186 187static int 188name_p(Elf_Scn *scn, GElf_Shdr *shdr, void *d) 189{ 190 struct section_named_data *data = d; 191 const char *name = elf_strptr(data->lte->elf, 192 data->lte->ehdr.e_shstrndx, 193 shdr->sh_name); 194 return strcmp(name, data->name) == 0; 195} 196 197int 198elf_get_section_named(struct ltelf *lte, const char *name, 199 Elf_Scn **tgt_sec, GElf_Shdr *tgt_shdr) 200{ 201 struct section_named_data data = { 202 .lte = lte, 203 .name = name, 204 }; 205 return elf_get_section_if(lte, tgt_sec, tgt_shdr, 206 &name_p, &data); 207} 208 209static struct elf_each_symbol_t 210each_symbol_in(Elf_Data *symtab, const char *strtab, size_t count, 211 unsigned i, 212 enum callback_status (*cb)(GElf_Sym *symbol, 213 const char *name, void *data), 214 void *data) 215{ 216 for (; i < count; ++i) { 217 GElf_Sym sym; 218 if (gelf_getsym(symtab, i, &sym) == NULL) 219 return (struct elf_each_symbol_t){ i, -2 }; 220 221 switch (cb(&sym, strtab + sym.st_name, data)) { 222 case CBS_FAIL: 223 return (struct elf_each_symbol_t){ i, -1 }; 224 case CBS_STOP: 225 return (struct elf_each_symbol_t){ i + 1, 0 }; 226 case CBS_CONT: 227 break; 228 } 229 } 230 231 return (struct elf_each_symbol_t){ 0, 0 }; 232} 233 234/* N.B.: gelf_getsym takes integer argument. Since negative values 235 * are invalid as indices, we can use the extra bit to encode which 236 * symbol table we are looking into. ltrace currently doesn't handle 237 * more than two symbol tables anyway, nor does it handle the xindex 238 * stuff. */ 239struct elf_each_symbol_t 240elf_each_symbol(struct ltelf *lte, unsigned start_after, 241 enum callback_status (*cb)(GElf_Sym *symbol, 242 const char *name, void *data), 243 void *data) 244{ 245 unsigned index = start_after == 0 ? 0 : start_after >> 1; 246 247 /* Go through static symbol table first. */ 248 if ((start_after & 0x1) == 0) { 249 struct elf_each_symbol_t st 250 = each_symbol_in(lte->symtab, lte->strtab, 251 lte->symtab_count, index, cb, data); 252 253 /* If the iteration stopped prematurely, bail out. */ 254 if (st.restart != 0) 255 return ((struct elf_each_symbol_t) 256 { st.restart << 1, st.status }); 257 } 258 259 struct elf_each_symbol_t st 260 = each_symbol_in(lte->dynsym, lte->dynstr, lte->dynsym_count, 261 index, cb, data); 262 if (st.restart != 0) 263 return ((struct elf_each_symbol_t) 264 { st.restart << 1 | 0x1, st.status }); 265 266 return (struct elf_each_symbol_t){ 0, 0 }; 267} 268 269int 270elf_can_read_next(Elf_Data *data, GElf_Xword offset, GElf_Xword size) 271{ 272 assert(data != NULL); 273 if (data->d_size < size || offset > data->d_size - size) { 274 debug(1, "Not enough data to read %"PRId64"-byte value" 275 " at offset %"PRId64".", size, offset); 276 return 0; 277 } 278 return 1; 279} 280 281#define DEF_READER(NAME, SIZE) \ 282 int \ 283 NAME(Elf_Data *data, GElf_Xword offset, uint##SIZE##_t *retp) \ 284 { \ 285 if (!elf_can_read_next(data, offset, SIZE / 8)) \ 286 return -1; \ 287 \ 288 if (data->d_buf == NULL) /* NODATA section */ { \ 289 *retp = 0; \ 290 return 0; \ 291 } \ 292 \ 293 union { \ 294 uint##SIZE##_t dst; \ 295 char buf[0]; \ 296 } u; \ 297 memcpy(u.buf, data->d_buf + offset, sizeof(u.dst)); \ 298 *retp = u.dst; \ 299 return 0; \ 300 } 301 302DEF_READER(elf_read_u8, 8) 303DEF_READER(elf_read_u16, 16) 304DEF_READER(elf_read_u32, 32) 305DEF_READER(elf_read_u64, 64) 306 307#undef DEF_READER 308 309#define DEF_READER(NAME, SIZE) \ 310 int \ 311 NAME(Elf_Data *data, GElf_Xword *offset, uint##SIZE##_t *retp) \ 312 { \ 313 int rc = elf_read_u##SIZE(data, *offset, retp); \ 314 if (rc < 0) \ 315 return rc; \ 316 *offset += SIZE / 8; \ 317 return 0; \ 318 } 319 320DEF_READER(elf_read_next_u8, 8) 321DEF_READER(elf_read_next_u16, 16) 322DEF_READER(elf_read_next_u32, 32) 323DEF_READER(elf_read_next_u64, 64) 324 325#undef DEF_READER 326 327int 328elf_read_next_uleb128(Elf_Data *data, GElf_Xword *offset, uint64_t *retp) 329{ 330 uint64_t result = 0; 331 int shift = 0; 332 int size = 8 * sizeof result; 333 334 while (1) { 335 uint8_t byte; 336 if (elf_read_next_u8(data, offset, &byte) < 0) 337 return -1; 338 339 uint8_t payload = byte & 0x7f; 340 result |= (uint64_t)payload << shift; 341 shift += 7; 342 if (shift > size && byte != 0x1) 343 return -1; 344 if ((byte & 0x80) == 0) 345 break; 346 } 347 348 if (retp != NULL) 349 *retp = result; 350 return 0; 351} 352 353int 354elf_read_uleb128(Elf_Data *data, GElf_Xword offset, uint64_t *retp) 355{ 356 return elf_read_next_uleb128(data, &offset, retp); 357} 358 359int 360ltelf_init(struct ltelf *lte, const char *filename) 361{ 362 memset(lte, 0, sizeof *lte); 363 lte->fd = open(filename, O_RDONLY); 364 if (lte->fd == -1) 365 return 1; 366 367 elf_version(EV_CURRENT); 368 369#ifdef HAVE_ELF_C_READ_MMAP 370 lte->elf = elf_begin(lte->fd, ELF_C_READ_MMAP, NULL); 371#else 372 lte->elf = elf_begin(lte->fd, ELF_C_READ, NULL); 373#endif 374 375 if (lte->elf == NULL || elf_kind(lte->elf) != ELF_K_ELF) { 376 fprintf(stderr, "\"%s\" is not an ELF file\n", filename); 377 exit(EXIT_FAILURE); 378 } 379 380 if (gelf_getehdr(lte->elf, <e->ehdr) == NULL) { 381 fprintf(stderr, "can't read ELF header of \"%s\": %s\n", 382 filename, elf_errmsg(-1)); 383 exit(EXIT_FAILURE); 384 } 385 386 if (lte->ehdr.e_type != ET_EXEC && lte->ehdr.e_type != ET_DYN) { 387 fprintf(stderr, "\"%s\" is neither an ELF executable" 388 " nor a shared library\n", filename); 389 exit(EXIT_FAILURE); 390 } 391 392 if (1 393#ifdef LT_ELF_MACHINE 394 && (lte->ehdr.e_ident[EI_CLASS] != LT_ELFCLASS 395 || lte->ehdr.e_machine != LT_ELF_MACHINE) 396#endif 397#ifdef LT_ELF_MACHINE2 398 && (lte->ehdr.e_ident[EI_CLASS] != LT_ELFCLASS2 399 || lte->ehdr.e_machine != LT_ELF_MACHINE2) 400#endif 401#ifdef LT_ELF_MACHINE3 402 && (lte->ehdr.e_ident[EI_CLASS] != LT_ELFCLASS3 403 || lte->ehdr.e_machine != LT_ELF_MACHINE3) 404#endif 405 ) { 406 fprintf(stderr, 407 "\"%s\" is ELF from incompatible architecture\n", 408 filename); 409 exit(EXIT_FAILURE); 410 } 411 412 VECT_INIT(<e->plt_relocs, GElf_Rela); 413 414 return 0; 415} 416 417void 418ltelf_destroy(struct ltelf *lte) 419{ 420 debug(DEBUG_FUNCTION, "close_elf()"); 421 elf_end(lte->elf); 422 close(lte->fd); 423 VECT_DESTROY(<e->plt_relocs, GElf_Rela, NULL, NULL); 424} 425 426static void 427read_symbol_table(struct ltelf *lte, const char *filename, 428 Elf_Scn *scn, GElf_Shdr *shdr, const char *name, 429 Elf_Data **datap, size_t *countp, const char **strsp) 430{ 431 *datap = elf_getdata(scn, NULL); 432 *countp = shdr->sh_size / shdr->sh_entsize; 433 if ((*datap == NULL || elf_getdata(scn, *datap) != NULL) 434 && options.static_filter != NULL) { 435 fprintf(stderr, "Couldn't get data of section" 436 " %s from \"%s\": %s\n", 437 name, filename, elf_errmsg(-1)); 438 exit(EXIT_FAILURE); 439 } 440 441 scn = elf_getscn(lte->elf, shdr->sh_link); 442 GElf_Shdr shdr2; 443 if (scn == NULL || gelf_getshdr(scn, &shdr2) == NULL) { 444 fprintf(stderr, "Couldn't get header of section" 445 " #%d from \"%s\": %s\n", 446 shdr->sh_link, filename, elf_errmsg(-1)); 447 exit(EXIT_FAILURE); 448 } 449 450 Elf_Data *data = elf_getdata(scn, NULL); 451 if (data == NULL || elf_getdata(scn, data) != NULL 452 || shdr2.sh_size != data->d_size || data->d_off) { 453 fprintf(stderr, "Couldn't get data of section" 454 " #%d from \"%s\": %s\n", 455 shdr2.sh_link, filename, elf_errmsg(-1)); 456 exit(EXIT_FAILURE); 457 } 458 459 *strsp = data->d_buf; 460} 461 462static int 463rel_to_rela(struct ltelf *lte, const GElf_Rel *rel, GElf_Rela *rela) 464{ 465 rela->r_offset = rel->r_offset; 466 rela->r_info = rel->r_info; 467 468 Elf_Scn *sec; 469 GElf_Shdr shdr; 470 if (elf_get_section_covering(lte, rel->r_offset, &sec, &shdr) < 0 471 || sec == NULL) 472 return -1; 473 474 Elf_Data *data = elf_loaddata(sec, &shdr); 475 if (data == NULL) 476 return -1; 477 478 GElf_Xword offset = rel->r_offset - shdr.sh_addr - data->d_off; 479 uint64_t value; 480 if (lte->ehdr.e_ident[EI_CLASS] == ELFCLASS32) { 481 uint32_t tmp; 482 if (elf_read_u32(data, offset, &tmp) < 0) 483 return -1; 484 value = tmp; 485 } else if (elf_read_u64(data, offset, &value) < 0) { 486 return -1; 487 } 488 489 rela->r_addend = value; 490 return 0; 491} 492 493int 494elf_read_relocs(struct ltelf *lte, Elf_Scn *scn, GElf_Shdr *shdr, 495 struct vect *rela_vec) 496{ 497 if (vect_reserve_additional(rela_vec, lte->ehdr.e_shnum) < 0) 498 return -1; 499 500 Elf_Data *relplt = elf_loaddata(scn, shdr); 501 if (relplt == NULL) { 502 fprintf(stderr, "Couldn't load .rel*.plt data.\n"); 503 return -1; 504 } 505 506 if ((shdr->sh_size % shdr->sh_entsize) != 0) { 507 fprintf(stderr, ".rel*.plt size (%" PRIx64 "d) not a multiple " 508 "of its sh_entsize (%" PRIx64 "d).\n", 509 shdr->sh_size, shdr->sh_entsize); 510 return -1; 511 } 512 513 GElf_Xword relplt_count = shdr->sh_size / shdr->sh_entsize; 514 GElf_Xword i; 515 for (i = 0; i < relplt_count; ++i) { 516 GElf_Rela rela; 517 if (relplt->d_type == ELF_T_REL) { 518 GElf_Rel rel; 519 if (gelf_getrel(relplt, i, &rel) == NULL 520 || rel_to_rela(lte, &rel, &rela) < 0) 521 return -1; 522 523 } else if (gelf_getrela(relplt, i, &rela) == NULL) { 524 return -1; 525 } 526 527 if (VECT_PUSHBACK(rela_vec, &rela) < 0) 528 return -1; 529 } 530 531 return 0; 532} 533 534static int 535ltelf_read_elf(struct ltelf *lte, const char *filename) 536{ 537 int i; 538 GElf_Addr relplt_addr = 0; 539 GElf_Addr soname_offset = 0; 540 GElf_Xword relplt_size = 0; 541 542 debug(DEBUG_FUNCTION, "ltelf_read_elf(filename=%s)", filename); 543 debug(1, "Reading ELF from %s...", filename); 544 545 for (i = 1; i < lte->ehdr.e_shnum; ++i) { 546 Elf_Scn *scn; 547 GElf_Shdr shdr; 548 const char *name; 549 550 scn = elf_getscn(lte->elf, i); 551 if (scn == NULL || gelf_getshdr(scn, &shdr) == NULL) { 552 fprintf(stderr, "Couldn't get section #%d from" 553 " \"%s\": %s\n", i, filename, elf_errmsg(-1)); 554 exit(EXIT_FAILURE); 555 } 556 557 name = elf_strptr(lte->elf, lte->ehdr.e_shstrndx, shdr.sh_name); 558 if (name == NULL) { 559 fprintf(stderr, "Couldn't get name of section #%d from" 560 " \"%s\": %s\n", i, filename, elf_errmsg(-1)); 561 exit(EXIT_FAILURE); 562 } 563 564 if (shdr.sh_type == SHT_SYMTAB) { 565 read_symbol_table(lte, filename, 566 scn, &shdr, name, <e->symtab, 567 <e->symtab_count, <e->strtab); 568 569 } else if (shdr.sh_type == SHT_DYNSYM) { 570 read_symbol_table(lte, filename, 571 scn, &shdr, name, <e->dynsym, 572 <e->dynsym_count, <e->dynstr); 573 574 } else if (shdr.sh_type == SHT_DYNAMIC) { 575 Elf_Data *data; 576 size_t j; 577 578 lte->dyn_addr = shdr.sh_addr + lte->bias; 579 lte->dyn_sz = shdr.sh_size; 580 581 data = elf_getdata(scn, NULL); 582 if (data == NULL || elf_getdata(scn, data) != NULL) { 583 fprintf(stderr, "Couldn't get .dynamic data" 584 " from \"%s\": %s\n", 585 filename, strerror(errno)); 586 exit(EXIT_FAILURE); 587 } 588 589 for (j = 0; j < shdr.sh_size / shdr.sh_entsize; ++j) { 590 GElf_Dyn dyn; 591 592 if (gelf_getdyn(data, j, &dyn) == NULL) { 593 fprintf(stderr, "Couldn't get .dynamic" 594 " data from \"%s\": %s\n", 595 filename, strerror(errno)); 596 exit(EXIT_FAILURE); 597 } 598 if (dyn.d_tag == DT_JMPREL) 599 relplt_addr = dyn.d_un.d_ptr; 600 else if (dyn.d_tag == DT_PLTRELSZ) 601 relplt_size = dyn.d_un.d_val; 602 else if (dyn.d_tag == DT_SONAME) 603 soname_offset = dyn.d_un.d_val; 604 } 605 } else if (shdr.sh_type == SHT_PROGBITS 606 || shdr.sh_type == SHT_NOBITS) { 607 if (strcmp(name, ".plt") == 0) { 608 lte->plt_addr = shdr.sh_addr; 609 lte->plt_size = shdr.sh_size; 610 lte->plt_data = elf_loaddata(scn, &shdr); 611 if (lte->plt_data == NULL) 612 fprintf(stderr, 613 "Can't load .plt data\n"); 614 lte->plt_flags = shdr.sh_flags; 615 } 616#ifdef ARCH_SUPPORTS_OPD 617 else if (strcmp(name, ".opd") == 0) { 618 lte->opd_addr = (GElf_Addr *) (long) shdr.sh_addr; 619 lte->opd_size = shdr.sh_size; 620 lte->opd = elf_rawdata(scn, NULL); 621 } 622#endif 623 } 624 } 625 626 if (lte->dynsym == NULL || lte->dynstr == NULL) { 627 fprintf(stderr, "Couldn't find .dynsym or .dynstr in \"%s\"\n", 628 filename); 629 exit(EXIT_FAILURE); 630 } 631 632 if (!relplt_addr || !lte->plt_addr) { 633 debug(1, "%s has no PLT relocations", filename); 634 } else if (relplt_size == 0) { 635 debug(1, "%s has unknown PLT size", filename); 636 } else { 637 for (i = 1; i < lte->ehdr.e_shnum; ++i) { 638 Elf_Scn *scn; 639 GElf_Shdr shdr; 640 641 scn = elf_getscn(lte->elf, i); 642 if (scn == NULL || gelf_getshdr(scn, &shdr) == NULL) { 643 fprintf(stderr, "Couldn't get section header" 644 " from \"%s\": %s\n", 645 filename, elf_errmsg(-1)); 646 exit(EXIT_FAILURE); 647 } 648 if (shdr.sh_addr == relplt_addr 649 && shdr.sh_size == relplt_size) { 650 if (elf_read_relocs(lte, scn, &shdr, 651 <e->plt_relocs) < 0) { 652 fprintf(stderr, "Couldn't get .rel*.plt" 653 " data from \"%s\": %s\n", 654 filename, elf_errmsg(-1)); 655 exit(EXIT_FAILURE); 656 } 657 break; 658 } 659 } 660 661 if (i == lte->ehdr.e_shnum) { 662 fprintf(stderr, 663 "Couldn't find .rel*.plt section in \"%s\"\n", 664 filename); 665 exit(EXIT_FAILURE); 666 } 667 } 668 debug(1, "%s %zd PLT relocations", filename, 669 vect_size(<e->plt_relocs)); 670 671 if (soname_offset != 0) 672 lte->soname = lte->dynstr + soname_offset; 673 674 return 0; 675} 676 677#ifndef ARCH_HAVE_GET_SYMINFO 678int 679arch_get_sym_info(struct ltelf *lte, const char *filename, 680 size_t sym_index, GElf_Rela *rela, GElf_Sym *sym) 681{ 682 return gelf_getsym(lte->dynsym, 683 ELF64_R_SYM(rela->r_info), sym) != NULL ? 0 : -1; 684} 685#endif 686 687int 688default_elf_add_plt_entry(struct process *proc, struct ltelf *lte, 689 const char *a_name, GElf_Rela *rela, size_t ndx, 690 struct library_symbol **ret) 691{ 692 char *name = strdup(a_name); 693 if (name == NULL) { 694 fail_message: 695 fprintf(stderr, "Couldn't create symbol for PLT entry: %s\n", 696 strerror(errno)); 697 fail: 698 free(name); 699 return -1; 700 } 701 702 GElf_Addr addr = arch_plt_sym_val(lte, ndx, rela); 703 704 struct library_symbol *libsym = malloc(sizeof(*libsym)); 705 if (libsym == NULL) 706 goto fail_message; 707 708 /* XXX The double cast should be removed when 709 * arch_addr_t becomes integral type. */ 710 arch_addr_t taddr = (arch_addr_t) 711 (uintptr_t)(addr + lte->bias); 712 713 if (library_symbol_init(libsym, taddr, name, 1, LS_TOPLT_EXEC) < 0) { 714 free(libsym); 715 goto fail; 716 } 717 718 libsym->next = *ret; 719 *ret = libsym; 720 return 0; 721} 722 723int 724elf_add_plt_entry(struct process *proc, struct ltelf *lte, 725 const char *name, GElf_Rela *rela, size_t idx, 726 struct library_symbol **ret) 727{ 728 enum plt_status plts 729 = arch_elf_add_plt_entry(proc, lte, name, rela, idx, ret); 730 731 if (plts == PLT_DEFAULT) 732 plts = os_elf_add_plt_entry(proc, lte, name, rela, idx, ret); 733 734 switch (plts) { 735 case PLT_DEFAULT: 736 return default_elf_add_plt_entry(proc, lte, name, 737 rela, idx, ret); 738 case PLT_FAIL: 739 return -1; 740 case PLT_OK: 741 return 0; 742 } 743 744 assert(! "Invalid return from X_elf_add_plt_entry!"); 745 abort(); 746} 747 748static void 749mark_chain_latent(struct library_symbol *libsym) 750{ 751 for (; libsym != NULL; libsym = libsym->next) { 752 debug(DEBUG_FUNCTION, "marking %s latent", libsym->name); 753 libsym->latent = 1; 754 } 755} 756 757static void 758filter_symbol_chain(struct filter *filter, 759 struct library_symbol **libsymp, struct library *lib) 760{ 761 assert(libsymp != NULL); 762 struct library_symbol **ptr = libsymp; 763 while (*ptr != NULL) { 764 if (filter_matches_symbol(filter, (*ptr)->name, lib)) { 765 ptr = &(*ptr)->next; 766 } else { 767 struct library_symbol *sym = *ptr; 768 *ptr = (*ptr)->next; 769 library_symbol_destroy(sym); 770 free(sym); 771 } 772 } 773} 774 775static int 776populate_plt(struct process *proc, const char *filename, 777 struct ltelf *lte, struct library *lib) 778{ 779 const bool latent_plts = options.export_filter != NULL; 780 const size_t count = vect_size(<e->plt_relocs); 781 782 size_t i; 783 for (i = 0; i < count; ++i) { 784 GElf_Rela *rela = VECT_ELEMENT(<e->plt_relocs, GElf_Rela, i); 785 GElf_Sym sym; 786 787 switch (arch_get_sym_info(lte, filename, i, rela, &sym)) { 788 default: 789 fprintf(stderr, 790 "Couldn't get relocation for symbol #%zd" 791 " from \"%s\": %s\n", 792 i, filename, elf_errmsg(-1)); 793 /* Fall through. */ 794 case 1: 795 continue; /* Skip this entry. */ 796 case 0: 797 break; 798 } 799 800 char const *name = lte->dynstr + sym.st_name; 801 int matched = filter_matches_symbol(options.plt_filter, 802 name, lib); 803 804 struct library_symbol *libsym = NULL; 805 if (elf_add_plt_entry(proc, lte, name, rela, i, &libsym) < 0) 806 return -1; 807 808 /* If we didn't match the PLT entry, filter the chain 809 * to only include the matching symbols (but include 810 * all if we are adding latent symbols) to allow 811 * backends to override the PLT symbol's name. */ 812 813 if (! matched && ! latent_plts) 814 filter_symbol_chain(options.plt_filter, &libsym, lib); 815 816 if (libsym != NULL) { 817 /* If we are adding those symbols just for 818 * tracing exports, mark them all latent. */ 819 if (! matched && latent_plts) 820 mark_chain_latent(libsym); 821 library_add_symbol(lib, libsym); 822 } 823 } 824 return 0; 825} 826 827static void 828delete_symbol_chain(struct library_symbol *libsym) 829{ 830 while (libsym != NULL) { 831 struct library_symbol *tmp = libsym->next; 832 library_symbol_destroy(libsym); 833 free(libsym); 834 libsym = tmp; 835 } 836} 837 838/* When -x rules result in request to trace several aliases, we only 839 * want to add such symbol once. The only way that those symbols 840 * differ in is their name, e.g. in glibc you have __GI___libc_free, 841 * __cfree, __free, __libc_free, cfree and free all defined on the 842 * same address. So instead we keep this unique symbol struct for 843 * each address, and replace name in libsym with a shorter variant if 844 * we find it. */ 845struct unique_symbol { 846 arch_addr_t addr; 847 struct library_symbol *libsym; 848}; 849 850static int 851unique_symbol_cmp(const void *key, const void *val) 852{ 853 const struct unique_symbol *sym_key = key; 854 const struct unique_symbol *sym_val = val; 855 return sym_key->addr != sym_val->addr; 856} 857 858static enum callback_status 859symbol_with_address(struct library_symbol *sym, void *addrptr) 860{ 861 return sym->enter_addr == *(arch_addr_t *)addrptr 862 ? CBS_STOP : CBS_CONT; 863} 864 865static int 866populate_this_symtab(struct process *proc, const char *filename, 867 struct ltelf *lte, struct library *lib, 868 Elf_Data *symtab, const char *strtab, size_t count, 869 struct library_exported_name **names) 870{ 871 /* If a valid NAMES is passed, we pass in *NAMES a list of 872 * symbol names that this library exports. */ 873 if (names != NULL) 874 *names = NULL; 875 876 /* Using sorted array would be arguably better, but this 877 * should be well enough for the number of symbols that we 878 * typically deal with. */ 879 size_t num_symbols = 0; 880 struct unique_symbol *symbols = malloc(sizeof(*symbols) * count); 881 if (symbols == NULL) { 882 fprintf(stderr, "couldn't insert symbols for -x: %s\n", 883 strerror(errno)); 884 return -1; 885 } 886 887 GElf_Word secflags[lte->ehdr.e_shnum]; 888 size_t i; 889 for (i = 1; i < lte->ehdr.e_shnum; ++i) { 890 Elf_Scn *scn = elf_getscn(lte->elf, i); 891 GElf_Shdr shdr; 892 if (scn == NULL || gelf_getshdr(scn, &shdr) == NULL) 893 secflags[i] = 0; 894 else 895 secflags[i] = shdr.sh_flags; 896 } 897 898 for (i = 0; i < count; ++i) { 899 GElf_Sym sym; 900 if (gelf_getsym(symtab, i, &sym) == NULL) { 901 fprintf(stderr, 902 "couldn't get symbol #%zd from %s: %s\n", 903 i, filename, elf_errmsg(-1)); 904 continue; 905 } 906 907 if (sym.st_value == 0 || sym.st_shndx == STN_UNDEF 908 /* Also ignore any special values besides direct 909 * section references. */ 910 || sym.st_shndx >= lte->ehdr.e_shnum) 911 continue; 912 913 /* Find symbol name and snip version. */ 914 const char *orig_name = strtab + sym.st_name; 915 const char *version = strchr(orig_name, '@'); 916 size_t len = version != NULL ? (assert(version > orig_name), 917 (size_t)(version - orig_name)) 918 : strlen(orig_name); 919 char name[len + 1]; 920 memcpy(name, orig_name, len); 921 name[len] = 0; 922 923 /* If we are interested in exports, store this name. */ 924 if (names != NULL) { 925 struct library_exported_name *export 926 = malloc(sizeof *export); 927 char *name_copy = strdup(name); 928 929 if (name_copy == NULL || export == NULL) { 930 free(name_copy); 931 free(export); 932 fprintf(stderr, "Couldn't store symbol %s. " 933 "Tracing may be incomplete.\n", name); 934 } else { 935 export->name = name_copy; 936 export->own_name = 1; 937 export->next = *names; 938 *names = export; 939 } 940 } 941 942 /* If the symbol is not matched, skip it. We already 943 * stored it to export list above. */ 944 if (!filter_matches_symbol(options.static_filter, name, lib)) 945 continue; 946 947 arch_addr_t addr = (arch_addr_t) 948 (uintptr_t)(sym.st_value + lte->bias); 949 arch_addr_t naddr; 950 951 /* On arches that support OPD, the value of typical 952 * function symbol will be a pointer to .opd, but some 953 * will point directly to .text. We don't want to 954 * translate those. */ 955 if (secflags[sym.st_shndx] & SHF_EXECINSTR) { 956 naddr = addr; 957 } else if (arch_translate_address(lte, addr, &naddr) < 0) { 958 fprintf(stderr, 959 "couldn't translate address of %s@%s: %s\n", 960 name, lib->soname, strerror(errno)); 961 continue; 962 } 963 964 char *full_name = strdup(name); 965 if (full_name == NULL) { 966 fprintf(stderr, "couldn't copy name of %s@%s: %s\n", 967 name, lib->soname, strerror(errno)); 968 continue; 969 } 970 971 struct library_symbol *libsym = NULL; 972 enum plt_status plts 973 = arch_elf_add_func_entry(proc, lte, &sym, 974 naddr, full_name, &libsym); 975 if (plts == PLT_DEFAULT) 976 plts = os_elf_add_func_entry(proc, lte, &sym, 977 naddr, full_name, &libsym); 978 979 switch (plts) { 980 case PLT_DEFAULT:; 981 /* Put the default symbol to the chain. */ 982 struct library_symbol *tmp = malloc(sizeof *tmp); 983 if (tmp == NULL 984 || library_symbol_init(tmp, naddr, full_name, 1, 985 LS_TOPLT_NONE) < 0) { 986 free(tmp); 987 988 /* Either add the whole bunch, or none 989 * of it. Note that for PLT_FAIL we 990 * don't do this--it's the callee's 991 * job to clean up after itself before 992 * it bails out. */ 993 delete_symbol_chain(libsym); 994 libsym = NULL; 995 996 case PLT_FAIL: 997 fprintf(stderr, "Couldn't add symbol %s@%s " 998 "for tracing.\n", name, lib->soname); 999 1000 break; 1001 } 1002 1003 full_name = NULL; 1004 tmp->next = libsym; 1005 libsym = tmp; 1006 break; 1007 1008 case PLT_OK: 1009 break; 1010 } 1011 1012 free(full_name); 1013 1014 struct library_symbol *tmp; 1015 for (tmp = libsym; tmp != NULL; ) { 1016 /* Look whether we already have a symbol for 1017 * this address. If not, add this one. If 1018 * yes, look if we should pick the new symbol 1019 * name. */ 1020 1021 struct unique_symbol key = { tmp->enter_addr, NULL }; 1022 struct unique_symbol *unique 1023 = lsearch(&key, symbols, &num_symbols, 1024 sizeof *symbols, &unique_symbol_cmp); 1025 1026 if (unique->libsym == NULL) { 1027 unique->libsym = tmp; 1028 unique->addr = tmp->enter_addr; 1029 tmp = tmp->next; 1030 unique->libsym->next = NULL; 1031 } else { 1032 if (strlen(tmp->name) 1033 < strlen(unique->libsym->name)) { 1034 library_symbol_set_name 1035 (unique->libsym, tmp->name, 1); 1036 tmp->name = NULL; 1037 } 1038 struct library_symbol *next = tmp->next; 1039 library_symbol_destroy(tmp); 1040 free(tmp); 1041 tmp = next; 1042 } 1043 } 1044 } 1045 1046 /* Now we do the union of this set of unique symbols with 1047 * what's already in the library. */ 1048 for (i = 0; i < num_symbols; ++i) { 1049 struct library_symbol *this_sym = symbols[i].libsym; 1050 assert(this_sym != NULL); 1051 struct library_symbol *other 1052 = library_each_symbol(lib, NULL, symbol_with_address, 1053 &this_sym->enter_addr); 1054 if (other != NULL) { 1055 library_symbol_destroy(this_sym); 1056 free(this_sym); 1057 symbols[i].libsym = NULL; 1058 } 1059 } 1060 1061 for (i = 0; i < num_symbols; ++i) 1062 if (symbols[i].libsym != NULL) 1063 library_add_symbol(lib, symbols[i].libsym); 1064 1065 free(symbols); 1066 return 0; 1067} 1068 1069static int 1070populate_symtab(struct process *proc, const char *filename, 1071 struct ltelf *lte, struct library *lib, 1072 int symtabs, int exports) 1073{ 1074 int status; 1075 if (symtabs && lte->symtab != NULL && lte->strtab != NULL 1076 && (status = populate_this_symtab(proc, filename, lte, lib, 1077 lte->symtab, lte->strtab, 1078 lte->symtab_count, NULL)) < 0) 1079 return status; 1080 1081 /* Check whether we want to trace symbols implemented by this 1082 * library (-l). */ 1083 struct library_exported_name **names = NULL; 1084 if (exports) { 1085 debug(DEBUG_FUNCTION, "-l matches %s", lib->soname); 1086 names = &lib->exported_names; 1087 } 1088 1089 return populate_this_symtab(proc, filename, lte, lib, 1090 lte->dynsym, lte->dynstr, 1091 lte->dynsym_count, names); 1092} 1093 1094static int 1095read_module(struct library *lib, struct process *proc, 1096 const char *filename, GElf_Addr bias, int main) 1097{ 1098 struct ltelf lte; 1099 if (ltelf_init(<e, filename) < 0) 1100 return -1; 1101 1102 /* XXX When we abstract ABI into a module, this should instead 1103 * become something like 1104 * 1105 * proc->abi = arch_get_abi(lte.ehdr); 1106 * 1107 * The code in ltelf_init needs to be replaced by this logic. 1108 * Be warned that libltrace.c calls ltelf_init as well to 1109 * determine whether ABI is supported. This is to get 1110 * reasonable error messages when trying to run 64-bit binary 1111 * with 32-bit ltrace. It is desirable to preserve this. */ 1112 proc->e_machine = lte.ehdr.e_machine; 1113 proc->e_class = lte.ehdr.e_ident[EI_CLASS]; 1114 get_arch_dep(proc); 1115 1116 /* Find out the base address. For PIE main binaries we look 1117 * into auxv, otherwise we scan phdrs. */ 1118 if (main && lte.ehdr.e_type == ET_DYN) { 1119 arch_addr_t entry; 1120 if (process_get_entry(proc, &entry, NULL) < 0) { 1121 fprintf(stderr, "Couldn't find entry of PIE %s\n", 1122 filename); 1123 fail: 1124 ltelf_destroy(<e); 1125 return -1; 1126 } 1127 /* XXX The double cast should be removed when 1128 * arch_addr_t becomes integral type. */ 1129 lte.entry_addr = (GElf_Addr)(uintptr_t)entry; 1130 lte.bias = (GElf_Addr)(uintptr_t)entry - lte.ehdr.e_entry; 1131 1132 } else { 1133 GElf_Phdr phdr; 1134 size_t i; 1135 for (i = 0; gelf_getphdr (lte.elf, i, &phdr) != NULL; ++i) { 1136 if (phdr.p_type == PT_LOAD) { 1137 lte.base_addr = phdr.p_vaddr + bias; 1138 break; 1139 } 1140 } 1141 1142 lte.bias = bias; 1143 lte.entry_addr = lte.ehdr.e_entry + lte.bias; 1144 1145 if (lte.base_addr == 0) { 1146 fprintf(stderr, 1147 "Couldn't determine base address of %s\n", 1148 filename); 1149 goto fail; 1150 } 1151 } 1152 1153 if (ltelf_read_elf(<e, filename) < 0) 1154 goto fail; 1155 1156 if (arch_elf_init(<e, lib) < 0) { 1157 fprintf(stderr, "Backend initialization failed.\n"); 1158 goto fail; 1159 } 1160 1161 if (lib == NULL) 1162 goto fail; 1163 1164 /* Note that we set soname and pathname as soon as they are 1165 * allocated, so in case of further errors, this get released 1166 * when LIB is released, which should happen in the caller 1167 * when we return error. */ 1168 1169 if (lib->pathname == NULL) { 1170 char *pathname = strdup(filename); 1171 if (pathname == NULL) 1172 goto fail; 1173 library_set_pathname(lib, pathname, 1); 1174 } 1175 1176 if (lte.soname != NULL) { 1177 char *soname = strdup(lte.soname); 1178 if (soname == NULL) 1179 goto fail; 1180 library_set_soname(lib, soname, 1); 1181 } else { 1182 const char *soname = rindex(lib->pathname, '/'); 1183 if (soname != NULL) 1184 soname += 1; 1185 else 1186 soname = lib->pathname; 1187 library_set_soname(lib, soname, 0); 1188 } 1189 1190 /* XXX The double cast should be removed when 1191 * arch_addr_t becomes integral type. */ 1192 arch_addr_t entry = (arch_addr_t)(uintptr_t)lte.entry_addr; 1193 if (arch_translate_address(<e, entry, &entry) < 0) 1194 goto fail; 1195 1196 /* XXX The double cast should be removed when 1197 * arch_addr_t becomes integral type. */ 1198 lib->base = (arch_addr_t)(uintptr_t)lte.base_addr; 1199 lib->entry = entry; 1200 /* XXX The double cast should be removed when 1201 * arch_addr_t becomes integral type. */ 1202 lib->dyn_addr = (arch_addr_t)(uintptr_t)lte.dyn_addr; 1203 1204 /* There are two reasons that we need to inspect symbol tables 1205 * or populate PLT entries. Either the user requested 1206 * corresponding tracing features (respectively -x and -e), or 1207 * they requested tracing exported symbols (-l). 1208 * 1209 * In the latter case we need to keep even those PLT slots 1210 * that are not requested by -e (but we keep them latent). We 1211 * also need to inspect .dynsym to find what exports this 1212 * library provide, to turn on existing latent PLT 1213 * entries. */ 1214 1215 int plts = filter_matches_library(options.plt_filter, lib); 1216 if ((plts || options.export_filter != NULL) 1217 && populate_plt(proc, filename, <e, lib) < 0) 1218 goto fail; 1219 1220 int exports = filter_matches_library(options.export_filter, lib); 1221 int symtabs = filter_matches_library(options.static_filter, lib); 1222 if ((symtabs || exports) 1223 && populate_symtab(proc, filename, <e, lib, 1224 symtabs, exports) < 0) 1225 goto fail; 1226 1227 arch_elf_destroy(<e); 1228 ltelf_destroy(<e); 1229 return 0; 1230} 1231 1232int 1233ltelf_read_library(struct library *lib, struct process *proc, 1234 const char *filename, GElf_Addr bias) 1235{ 1236 return read_module(lib, proc, filename, bias, 0); 1237} 1238 1239 1240struct library * 1241ltelf_read_main_binary(struct process *proc, const char *path) 1242{ 1243 struct library *lib = malloc(sizeof(*lib)); 1244 if (lib == NULL || library_init(lib, LT_LIBTYPE_MAIN) < 0) { 1245 free(lib); 1246 return NULL; 1247 } 1248 library_set_pathname(lib, path, 0); 1249 1250 /* There is a race between running the process and reading its 1251 * binary for internal consumption. So open the binary from 1252 * the /proc filesystem. XXX Note that there is similar race 1253 * for libraries, but there we don't have a nice answer like 1254 * that. Presumably we could read the DSOs from the process 1255 * memory image, but that's not currently done. */ 1256 char *fname = pid2name(proc->pid); 1257 if (fname == NULL 1258 || read_module(lib, proc, fname, 0, 1) < 0) { 1259 library_destroy(lib); 1260 free(lib); 1261 lib = NULL; 1262 } 1263 1264 free(fname); 1265 return lib; 1266} 1267