plt.c revision 6ca19a3c08bc75fb451df30f4c6e6c1a128731fe
1#include <gelf.h> 2#include <sys/ptrace.h> 3#include <errno.h> 4#include <error.h> 5#include <inttypes.h> 6#include <assert.h> 7#include <string.h> 8 9#include "proc.h" 10#include "common.h" 11#include "library.h" 12#include "breakpoint.h" 13#include "linux-gnu/trace.h" 14 15/* There are two PLT types on 32-bit PPC: old-style, BSS PLT, and 16 * new-style "secure" PLT. We can tell one from the other by the 17 * flags on the .plt section. If it's +X (executable), it's BSS PLT, 18 * otherwise it's secure. 19 * 20 * BSS PLT works the same way as most architectures: the .plt section 21 * contains trampolines and we put breakpoints to those. If not 22 * prelinked, .plt contains zeroes, and dynamic linker fills in the 23 * initial set of trampolines, which means that we need to delay 24 * enabling breakpoints until after binary entry point is hit. 25 * Additionally, after first call, dynamic linker updates .plt with 26 * branch to resolved address. That means that on first hit, we must 27 * do something similar to the PPC64 gambit described below. 28 * 29 * With secure PLT, the .plt section doesn't contain instructions but 30 * addresses. The real PLT table is stored in .text. Addresses of 31 * those PLT entries can be computed, and apart from the fact that 32 * they are in .text, they are ordinary PLT entries. 33 * 34 * 64-bit PPC is more involved. Program linker creates for each 35 * library call a _stub_ symbol named xxxxxxxx.plt_call.<callee> 36 * (where xxxxxxxx is a hexadecimal number). That stub does the call 37 * dispatch: it loads an address of a function to call from the 38 * section .plt, and branches. PLT entries themselves are essentially 39 * a curried call to the resolver. When the symbol is resolved, the 40 * resolver updates the value stored in .plt, and the next time 41 * around, the stub calls the library function directly. So we make 42 * at most one trip (none if the binary is prelinked) through each PLT 43 * entry, and correspondingly that is useless as a breakpoint site. 44 * 45 * Note the three confusing terms: stubs (that play the role of PLT 46 * entries), PLT entries, .plt section. 47 * 48 * We first check symbol tables and see if we happen to have stub 49 * symbols available. If yes we just put breakpoints to those, and 50 * treat them as usual breakpoints. The only tricky part is realizing 51 * that there can be more than one breakpoint per symbol. 52 * 53 * The case that we don't have the stub symbols available is harder. 54 * The following scheme uses two kinds of PLT breakpoints: unresolved 55 * and resolved (to some address). When the process starts (or when 56 * we attach), we distribute unresolved PLT breakpoints to the PLT 57 * entries (not stubs). Then we look in .plt, and for each entry 58 * whose value is different than the corresponding PLT entry address, 59 * we assume it was already resolved, and convert the breakpoint to 60 * resolved. We also rewrite the resolved value in .plt back to the 61 * PLT address. 62 * 63 * When a PLT entry hits a resolved breakpoint (which happens because 64 * we rewrite .plt with the original unresolved addresses), we move 65 * the instruction pointer to the corresponding address and continue 66 * the process as if nothing happened. 67 * 68 * When unresolved PLT entry is called for the first time, we need to 69 * catch the new value that the resolver will write to a .plt slot. 70 * We also need to prevent another thread from racing through and 71 * taking the branch without ltrace noticing. So when unresolved PLT 72 * entry hits, we have to stop all threads. We then single-step 73 * through the resolver, until the .plt slot changes. When it does, 74 * we treat it the same way as above: convert the PLT breakpoint to 75 * resolved, and rewrite the .plt value back to PLT address. We then 76 * start all threads again. 77 * 78 * As an optimization, we remember the address where the address was 79 * resolved, and put a breakpoint there. The next time around (when 80 * the next PLT entry is to be resolved), instead of single-stepping 81 * through half the dynamic linker, we just let the thread run and hit 82 * this breakpoint. When it hits, we know the PLT entry was resolved. 83 * 84 * XXX TODO If we have hardware watch point, we might put a read watch 85 * on .plt slot, and discover the offenders this way. I don't know 86 * the details, but I assume at most a handful (like, one or two, if 87 * available at all) addresses may be watched at a time, and thus this 88 * would be used as an amendment of the above rather than full-on 89 * solution to PLT tracing on PPC. 90 */ 91 92#define PPC_PLT_STUB_SIZE 16 93#define PPC64_PLT_STUB_SIZE 8 //xxx 94 95static inline int 96host_powerpc64() 97{ 98#ifdef __powerpc64__ 99 return 1; 100#else 101 return 0; 102#endif 103} 104 105int 106read_target_4(struct Process *proc, target_address_t addr, uint32_t *lp) 107{ 108 unsigned long l = ptrace(PTRACE_PEEKTEXT, proc->pid, addr, 0); 109 if (l == -1UL && errno) 110 return -1; 111#ifdef __powerpc64__ 112 l >>= 32; 113#endif 114 *lp = l; 115 return 0; 116} 117 118static int 119read_target_8(struct Process *proc, target_address_t addr, uint64_t *lp) 120{ 121 unsigned long l = ptrace(PTRACE_PEEKTEXT, proc->pid, addr, 0); 122 if (l == -1UL && errno) 123 return -1; 124 if (host_powerpc64()) { 125 *lp = l; 126 } else { 127 unsigned long l2 = ptrace(PTRACE_PEEKTEXT, proc->pid, 128 addr + 4, 0); 129 if (l2 == -1UL && errno) 130 return -1; 131 *lp = ((uint64_t)l << 32) | l2; 132 } 133 return 0; 134} 135 136int 137read_target_long(struct Process *proc, target_address_t addr, uint64_t *lp) 138{ 139 if (proc->e_machine == EM_PPC) { 140 uint32_t w; 141 int ret = read_target_4(proc, addr, &w); 142 if (ret >= 0) 143 *lp = (uint64_t)w; 144 return ret; 145 } else { 146 return read_target_8(proc, addr, lp); 147 } 148} 149 150static enum callback_status 151reenable_breakpoint(struct Process *proc, struct breakpoint *bp, void *data) 152{ 153 /* We don't need to re-enable non-PLT breakpoints and 154 * breakpoints that are not PPC32 BSS unprelinked. */ 155 if (bp->libsym == NULL 156 || bp->libsym->plt_type == LS_TOPLT_NONE 157 || bp->libsym->lib->arch.bss_plt_prelinked != 0) 158 return CBS_CONT; 159 160 debug(DEBUG_PROCESS, "pid=%d reenable_breakpoint %s", 161 proc->pid, breakpoint_name(bp)); 162 163 assert(proc->e_machine == EM_PPC); 164 uint64_t l; 165 if (read_target_8(proc, bp->addr, &l) < 0) { 166 error(0, errno, "couldn't read PLT value for %s(%p)", 167 breakpoint_name(bp), bp->addr); 168 return CBS_CONT; 169 } 170 171 /* XXX double cast */ 172 bp->libsym->arch.plt_slot_addr = (GElf_Addr)(uintptr_t)bp->addr; 173 174 /* If necessary, re-enable the breakpoint if it was 175 * overwritten by the dynamic linker. */ 176 union { 177 uint32_t insn; 178 char buf[4]; 179 } u = { .buf = BREAKPOINT_VALUE }; 180 if (l >> 32 == u.insn) 181 debug(DEBUG_PROCESS, "pid=%d, breakpoint still present" 182 " at %p, avoiding reenable", proc->pid, bp->addr); 183 else 184 enable_breakpoint(proc, bp); 185 186 bp->libsym->arch.resolved_value = l; 187 188 return CBS_CONT; 189} 190 191void 192arch_dynlink_done(struct Process *proc) 193{ 194 /* On PPC32, .plt of objects that use BSS PLT are overwritten 195 * by the dynamic linker (unless that object was prelinked). 196 * We need to re-enable breakpoints in those objects. */ 197 proc_each_breakpoint(proc, NULL, reenable_breakpoint, NULL); 198} 199 200GElf_Addr 201arch_plt_sym_val(struct ltelf *lte, size_t ndx, GElf_Rela *rela) 202{ 203 if (lte->ehdr.e_machine == EM_PPC && lte->arch.secure_plt) { 204 assert(lte->arch.plt_stub_vma != 0); 205 return lte->arch.plt_stub_vma + PPC_PLT_STUB_SIZE * ndx; 206 207 } else if (lte->ehdr.e_machine == EM_PPC) { 208 return rela->r_offset; 209 210 } else { 211 /* If we get here, we don't have stub symbols. In 212 * that case we put brakpoints to PLT entries the same 213 * as the PPC32 secure PLT case does. */ 214 assert(lte->arch.plt_stub_vma != 0); 215 return lte->arch.plt_stub_vma + PPC64_PLT_STUB_SIZE * ndx; 216 } 217} 218 219/* This entry point is called when ltelf is not available 220 * anymore--during runtime. At that point we don't have to concern 221 * ourselves with bias, as the values in OPD have been resolved 222 * already. */ 223int 224arch_translate_address_dyn(struct Process *proc, 225 target_address_t addr, target_address_t *ret) 226{ 227 if (proc->e_machine == EM_PPC64) { 228 uint64_t value; 229 if (read_target_8(proc, addr, &value) < 0) { 230 error(0, errno, "dynamic .opd translation of %p", addr); 231 return -1; 232 } 233 *ret = (target_address_t)value; 234 return 0; 235 } 236 237 *ret = addr; 238 return 0; 239} 240 241int 242arch_translate_address(struct ltelf *lte, 243 target_address_t addr, target_address_t *ret) 244{ 245 if (lte->ehdr.e_machine == EM_PPC64) { 246 GElf_Xword offset = (GElf_Addr)addr - lte->arch.opd_base; 247 uint64_t value; 248 if (elf_read_u64(lte->arch.opd_data, offset, &value) < 0) { 249 error(0, 0, "static .opd translation of %p: %s", addr, 250 elf_errmsg(-1)); 251 return -1; 252 } 253 *ret = (target_address_t)(value + lte->bias); 254 return 0; 255 } 256 257 *ret = addr; 258 return 0; 259} 260 261static int 262load_opd_data(struct ltelf *lte, struct library *lib) 263{ 264 Elf_Scn *sec; 265 GElf_Shdr shdr; 266 if (elf_get_section_named(lte, ".opd", &sec, &shdr) < 0) { 267 fail: 268 fprintf(stderr, "couldn't find .opd data\n"); 269 return -1; 270 } 271 272 lte->arch.opd_data = elf_rawdata(sec, NULL); 273 if (lte->arch.opd_data == NULL) 274 goto fail; 275 276 lte->arch.opd_base = shdr.sh_addr + lte->bias; 277 lte->arch.opd_size = shdr.sh_size; 278 279 return 0; 280} 281 282void * 283sym2addr(struct Process *proc, struct library_symbol *sym) 284{ 285 return sym->enter_addr; 286} 287 288static GElf_Addr 289get_glink_vma(struct ltelf *lte, GElf_Addr ppcgot, Elf_Data *plt_data) 290{ 291 Elf_Scn *ppcgot_sec = NULL; 292 GElf_Shdr ppcgot_shdr; 293 if (ppcgot != 0 294 && elf_get_section_covering(lte, ppcgot, 295 &ppcgot_sec, &ppcgot_shdr) < 0) 296 error(0, 0, "DT_PPC_GOT=%#"PRIx64", but no such section found", 297 ppcgot); 298 299 if (ppcgot_sec != NULL) { 300 Elf_Data *data = elf_loaddata(ppcgot_sec, &ppcgot_shdr); 301 if (data == NULL || data->d_size < 8 ) { 302 error(0, 0, "couldn't read GOT data"); 303 } else { 304 // where PPCGOT begins in .got 305 size_t offset = ppcgot - ppcgot_shdr.sh_addr; 306 assert(offset % 4 == 0); 307 uint32_t glink_vma; 308 if (elf_read_u32(data, offset + 4, &glink_vma) < 0) { 309 error(0, 0, "couldn't read glink VMA address" 310 " at %zd@GOT", offset); 311 return 0; 312 } 313 if (glink_vma != 0) { 314 debug(1, "PPC GOT glink_vma address: %#" PRIx32, 315 glink_vma); 316 return (GElf_Addr)glink_vma; 317 } 318 } 319 } 320 321 if (plt_data != NULL) { 322 uint32_t glink_vma; 323 if (elf_read_u32(plt_data, 0, &glink_vma) < 0) { 324 error(0, 0, "couldn't read glink VMA address"); 325 return 0; 326 } 327 debug(1, ".plt glink_vma address: %#" PRIx32, glink_vma); 328 return (GElf_Addr)glink_vma; 329 } 330 331 return 0; 332} 333 334static int 335load_dynamic_entry(struct ltelf *lte, int tag, GElf_Addr *valuep) 336{ 337 Elf_Scn *scn; 338 GElf_Shdr shdr; 339 if (elf_get_section_type(lte, SHT_DYNAMIC, &scn, &shdr) < 0 340 || scn == NULL) { 341 fail: 342 error(0, 0, "Couldn't get SHT_DYNAMIC: %s", 343 elf_errmsg(-1)); 344 return -1; 345 } 346 347 Elf_Data *data = elf_loaddata(scn, &shdr); 348 if (data == NULL) 349 goto fail; 350 351 size_t j; 352 for (j = 0; j < shdr.sh_size / shdr.sh_entsize; ++j) { 353 GElf_Dyn dyn; 354 if (gelf_getdyn(data, j, &dyn) == NULL) 355 goto fail; 356 357 if(dyn.d_tag == tag) { 358 *valuep = dyn.d_un.d_ptr; 359 return 0; 360 } 361 } 362 363 return -1; 364} 365 366static int 367load_ppcgot(struct ltelf *lte, GElf_Addr *ppcgotp) 368{ 369 return load_dynamic_entry(lte, DT_PPC_GOT, ppcgotp); 370} 371 372static int 373load_ppc64_glink(struct ltelf *lte, GElf_Addr *glinkp) 374{ 375 return load_dynamic_entry(lte, DT_PPC64_GLINK, glinkp); 376} 377 378static int 379nonzero_data(Elf_Data *data) 380{ 381 /* We are not supposed to get here if there's no PLT. */ 382 assert(data != NULL); 383 384 unsigned char *buf = data->d_buf; 385 if (buf == NULL) 386 return 0; 387 388 size_t i; 389 for (i = 0; i < data->d_size; ++i) 390 if (buf[i] != 0) 391 return 1; 392 return 0; 393} 394 395int 396arch_elf_init(struct ltelf *lte, struct library *lib) 397{ 398 if (lte->ehdr.e_machine == EM_PPC64 399 && load_opd_data(lte, lib) < 0) 400 return -1; 401 402 lte->arch.secure_plt = !(lte->plt_flags & SHF_EXECINSTR); 403 404 /* For PPC32 BSS, it is important whether the binary was 405 * prelinked. If .plt section is NODATA, or if it contains 406 * zeroes, then this library is not prelinked, and we need to 407 * delay breakpoints. */ 408 if (lte->ehdr.e_machine == EM_PPC && !lte->arch.secure_plt) 409 lib->arch.bss_plt_prelinked = nonzero_data(lte->plt_data); 410 else 411 /* For cases where it's irrelevant, initialize the 412 * value to something conspicuous. */ 413 lib->arch.bss_plt_prelinked = -1; 414 415 if (lte->ehdr.e_machine == EM_PPC && lte->arch.secure_plt) { 416 GElf_Addr ppcgot; 417 if (load_ppcgot(lte, &ppcgot) < 0) { 418 error(0, 0, "couldn't find DT_PPC_GOT"); 419 return -1; 420 } 421 GElf_Addr glink_vma = get_glink_vma(lte, ppcgot, lte->plt_data); 422 423 assert (lte->relplt_size % 12 == 0); 424 size_t count = lte->relplt_size / 12; // size of RELA entry 425 lte->arch.plt_stub_vma = glink_vma 426 - (GElf_Addr)count * PPC_PLT_STUB_SIZE; 427 debug(1, "stub_vma is %#" PRIx64, lte->arch.plt_stub_vma); 428 429 } else if (lte->ehdr.e_machine == EM_PPC64) { 430 GElf_Addr glink_vma; 431 if (load_ppc64_glink(lte, &glink_vma) < 0) { 432 error(0, 0, "couldn't find DT_PPC64_GLINK"); 433 return -1; 434 } 435 436 /* The first glink stub starts at offset 32. */ 437 lte->arch.plt_stub_vma = glink_vma + 32; 438 } 439 440 /* On PPC64, look for stub symbols in symbol table. These are 441 * called: xxxxxxxx.plt_call.callee_name@version+addend. */ 442 if (lte->ehdr.e_machine == EM_PPC64 443 && lte->symtab != NULL && lte->strtab != NULL) { 444 445 /* N.B. We can't simply skip the symbols that we fail 446 * to read or malloc. There may be more than one stub 447 * per symbol name, and if we failed in one but 448 * succeeded in another, the PLT enabling code would 449 * have no way to tell that something is missing. We 450 * could work around that, of course, but it doesn't 451 * seem worth the trouble. So if anything fails, we 452 * just pretend that we don't have stub symbols at 453 * all, as if the binary is stripped. */ 454 455 size_t i; 456 for (i = 0; i < lte->symtab_count; ++i) { 457 GElf_Sym sym; 458 if (gelf_getsym(lte->symtab, i, &sym) == NULL) { 459 struct library_symbol *sym, *next; 460 fail: 461 for (sym = lte->arch.stubs; sym != NULL; ) { 462 next = sym->next; 463 library_symbol_destroy(sym); 464 free(sym); 465 sym = next; 466 } 467 lte->arch.stubs = NULL; 468 break; 469 } 470 471 const char *name = lte->strtab + sym.st_name; 472 473#define STUBN ".plt_call." 474 if ((name = strstr(name, STUBN)) == NULL) 475 continue; 476 name += sizeof(STUBN) - 1; 477#undef STUBN 478 479 size_t len; 480 const char *ver = strchr(name, '@'); 481 if (ver != NULL) { 482 len = ver - name; 483 484 } else { 485 /* If there is "+" at all, check that 486 * the symbol name ends in "+0". */ 487 const char *add = strrchr(name, '+'); 488 if (add != NULL) { 489 assert(strcmp(add, "+0") == 0); 490 len = add - name; 491 } else { 492 len = strlen(name); 493 } 494 } 495 496 char *sym_name = strndup(name, len); 497 struct library_symbol *libsym = malloc(sizeof(*libsym)); 498 if (sym_name == NULL || libsym == NULL) { 499 fail2: 500 free(sym_name); 501 free(libsym); 502 goto fail; 503 } 504 505 /* XXX The double cast should be removed when 506 * target_address_t becomes integral type. */ 507 target_address_t addr = (target_address_t) 508 (uintptr_t)sym.st_value + lte->bias; 509 if (library_symbol_init(libsym, addr, sym_name, 1, 510 LS_TOPLT_EXEC) < 0) 511 goto fail2; 512 libsym->arch.type = PPC64_PLT_STUB; 513 libsym->next = lte->arch.stubs; 514 lte->arch.stubs = libsym; 515 } 516 } 517 518 return 0; 519} 520 521static int 522read_plt_slot_value(struct Process *proc, GElf_Addr addr, GElf_Addr *valp) 523{ 524 /* On PPC64, we read from .plt, which contains 8 byte 525 * addresses. On PPC32 we read from .plt, which contains 4 526 * byte instructions, but the PLT is two instructions, and 527 * either can change. */ 528 uint64_t l; 529 /* XXX double cast. */ 530 if (read_target_8(proc, (target_address_t)(uintptr_t)addr, &l) < 0) { 531 error(0, errno, "ptrace .plt slot value @%#" PRIx64, addr); 532 return -1; 533 } 534 535 *valp = (GElf_Addr)l; 536 return 0; 537} 538 539static int 540unresolve_plt_slot(struct Process *proc, GElf_Addr addr, GElf_Addr value) 541{ 542 /* We only modify plt_entry[0], which holds the resolved 543 * address of the routine. We keep the TOC and environment 544 * pointers intact. Hence the only adjustment that we need to 545 * do is to IP. */ 546 if (ptrace(PTRACE_POKETEXT, proc->pid, addr, value) < 0) { 547 error(0, errno, "unresolve .plt slot"); 548 return -1; 549 } 550 return 0; 551} 552 553static void 554mark_as_resolved(struct library_symbol *libsym, GElf_Addr value) 555{ 556 libsym->arch.type = PPC_PLT_RESOLVED; 557 libsym->arch.resolved_value = value; 558} 559 560enum plt_status 561arch_elf_add_plt_entry(struct Process *proc, struct ltelf *lte, 562 const char *a_name, GElf_Rela *rela, size_t ndx, 563 struct library_symbol **ret) 564{ 565 if (lte->ehdr.e_machine == EM_PPC) 566 return plt_default; 567 568 /* PPC64. If we have stubs, we return a chain of breakpoint 569 * sites, one for each stub that corresponds to this PLT 570 * entry. */ 571 struct library_symbol *chain = NULL; 572 struct library_symbol **symp; 573 for (symp = <e->arch.stubs; *symp != NULL; ) { 574 struct library_symbol *sym = *symp; 575 if (strcmp(sym->name, a_name) != 0) { 576 symp = &(*symp)->next; 577 continue; 578 } 579 580 /* Re-chain the symbol from stubs to CHAIN. */ 581 *symp = sym->next; 582 sym->next = chain; 583 chain = sym; 584 } 585 586 if (chain != NULL) { 587 *ret = chain; 588 return plt_ok; 589 } 590 591 /* We don't have stub symbols. Find corresponding .plt slot, 592 * and check whether it contains the corresponding PLT address 593 * (or 0 if the dynamic linker hasn't run yet). N.B. we don't 594 * want read this from ELF file, but from process image. That 595 * makes a difference if we are attaching to a running 596 * process. */ 597 598 GElf_Addr plt_entry_addr = arch_plt_sym_val(lte, ndx, rela); 599 GElf_Addr plt_slot_addr = rela->r_offset; 600 assert(plt_slot_addr >= lte->plt_addr 601 || plt_slot_addr < lte->plt_addr + lte->plt_size); 602 603 GElf_Addr plt_slot_value; 604 if (read_plt_slot_value(proc, plt_slot_addr, &plt_slot_value) < 0) 605 return plt_fail; 606 607 char *name = strdup(a_name); 608 struct library_symbol *libsym = malloc(sizeof(*libsym)); 609 if (name == NULL || libsym == NULL) { 610 error(0, errno, "allocation for .plt slot"); 611 fail: 612 free(name); 613 free(libsym); 614 return plt_fail; 615 } 616 617 /* XXX The double cast should be removed when 618 * target_address_t becomes integral type. */ 619 if (library_symbol_init(libsym, 620 (target_address_t)(uintptr_t)plt_entry_addr, 621 name, 1, LS_TOPLT_EXEC) < 0) 622 goto fail; 623 libsym->arch.plt_slot_addr = plt_slot_addr; 624 625 if (plt_slot_value == plt_entry_addr || plt_slot_value == 0) { 626 libsym->arch.type = PPC_PLT_UNRESOLVED; 627 libsym->arch.resolved_value = plt_entry_addr; 628 629 } else { 630 /* Unresolve the .plt slot. If the binary was 631 * prelinked, this makes the code invalid, because in 632 * case of prelinked binary, the dynamic linker 633 * doesn't update .plt[0] and .plt[1] with addresses 634 * of the resover. But we don't care, we will never 635 * need to enter the resolver. That just means that 636 * we have to un-un-resolve this back before we 637 * detach. */ 638 639 if (unresolve_plt_slot(proc, plt_slot_addr, plt_entry_addr) < 0) { 640 library_symbol_destroy(libsym); 641 goto fail; 642 } 643 mark_as_resolved(libsym, plt_slot_value); 644 } 645 646 *ret = libsym; 647 return plt_ok; 648} 649 650void 651arch_elf_destroy(struct ltelf *lte) 652{ 653 struct library_symbol *sym; 654 for (sym = lte->arch.stubs; sym != NULL; ) { 655 struct library_symbol *next = sym->next; 656 library_symbol_destroy(sym); 657 free(sym); 658 sym = next; 659 } 660} 661 662static void 663dl_plt_update_bp_on_hit(struct breakpoint *bp, struct Process *proc) 664{ 665 debug(DEBUG_PROCESS, "pid=%d dl_plt_update_bp_on_hit %s(%p)", 666 proc->pid, breakpoint_name(bp), bp->addr); 667 struct process_stopping_handler *self = proc->arch.handler; 668 assert(self != NULL); 669 670 struct library_symbol *libsym = self->breakpoint_being_enabled->libsym; 671 GElf_Addr value; 672 if (read_plt_slot_value(proc, libsym->arch.plt_slot_addr, &value) < 0) 673 return; 674 675 /* On PPC64, we rewrite the slot value. */ 676 if (proc->e_machine == EM_PPC64) 677 unresolve_plt_slot(proc, libsym->arch.plt_slot_addr, 678 libsym->arch.resolved_value); 679 /* We mark the breakpoint as resolved on both arches. */ 680 mark_as_resolved(libsym, value); 681 682 /* cb_on_all_stopped looks if HANDLER is set to NULL as a way 683 * to check that this was run. It's an error if it 684 * wasn't. */ 685 proc->arch.handler = NULL; 686 687 breakpoint_turn_off(bp, proc); 688} 689 690static void 691cb_on_all_stopped(struct process_stopping_handler *self) 692{ 693 /* Put that in for dl_plt_update_bp_on_hit to see. */ 694 assert(self->task_enabling_breakpoint->arch.handler == NULL); 695 self->task_enabling_breakpoint->arch.handler = self; 696 697 linux_ptrace_disable_and_continue(self); 698} 699 700static enum callback_status 701cb_keep_stepping_p(struct process_stopping_handler *self) 702{ 703 struct Process *proc = self->task_enabling_breakpoint; 704 struct library_symbol *libsym = self->breakpoint_being_enabled->libsym; 705 706 GElf_Addr value; 707 if (read_plt_slot_value(proc, libsym->arch.plt_slot_addr, &value) < 0) 708 return CBS_FAIL; 709 710 /* In UNRESOLVED state, the RESOLVED_VALUE in fact contains 711 * the PLT entry value. */ 712 if (value == libsym->arch.resolved_value) 713 return CBS_CONT; 714 715 debug(DEBUG_PROCESS, "pid=%d PLT got resolved to value %#"PRIx64, 716 proc->pid, value); 717 718 /* The .plt slot got resolved! We can migrate the breakpoint 719 * to RESOLVED and stop single-stepping. */ 720 if (proc->e_machine == EM_PPC64 721 && unresolve_plt_slot(proc, libsym->arch.plt_slot_addr, 722 libsym->arch.resolved_value) < 0) 723 return CBS_FAIL; 724 725 /* Resolving on PPC64 consists of overwriting a doubleword in 726 * .plt. That doubleword is than read back by a stub, and 727 * jumped on. Hopefully we can assume that double word update 728 * is done on a single place only, as it contains a final 729 * address. We still need to look around for any sync 730 * instruction, but essentially it is safe to optimize away 731 * the single stepping next time and install a post-update 732 * breakpoint. 733 * 734 * The situation on PPC32 BSS is more complicated. The 735 * dynamic linker here updates potentially several 736 * instructions (XXX currently we assume two) and the rules 737 * are more complicated. Sometimes it's enough to adjust just 738 * one of the addresses--the logic for generating optimal 739 * dispatch depends on relative addresses of the .plt entry 740 * and the jump destination. We can't assume that the some 741 * instruction block does the update every time. So on PPC32, 742 * we turn the optimization off and just step through it each 743 * time. */ 744 if (proc->e_machine == EM_PPC) 745 goto done; 746 747 /* Install breakpoint to the address where the change takes 748 * place. If we fail, then that just means that we'll have to 749 * singlestep the next time around as well. */ 750 struct Process *leader = proc->leader; 751 if (leader == NULL || leader->arch.dl_plt_update_bp != NULL) 752 goto done; 753 754 /* We need to install to the next instruction. ADDR points to 755 * a store instruction, so moving the breakpoint one 756 * instruction forward is safe. */ 757 target_address_t addr = get_instruction_pointer(proc) + 4; 758 leader->arch.dl_plt_update_bp = insert_breakpoint(proc, addr, NULL); 759 if (leader->arch.dl_plt_update_bp == NULL) 760 goto done; 761 762 static struct bp_callbacks dl_plt_update_cbs = { 763 .on_hit = dl_plt_update_bp_on_hit, 764 }; 765 leader->arch.dl_plt_update_bp->cbs = &dl_plt_update_cbs; 766 767 /* Turn it off for now. We will turn it on again when we hit 768 * the PLT entry that needs this. */ 769 breakpoint_turn_off(leader->arch.dl_plt_update_bp, proc); 770 771done: 772 mark_as_resolved(libsym, value); 773 774 return CBS_STOP; 775} 776 777static void 778jump_to_entry_point(struct Process *proc, struct breakpoint *bp) 779{ 780 /* XXX The double cast should be removed when 781 * target_address_t becomes integral type. */ 782 target_address_t rv = (target_address_t) 783 (uintptr_t)bp->libsym->arch.resolved_value; 784 set_instruction_pointer(proc, rv); 785} 786 787static void 788ppc_plt_bp_continue(struct breakpoint *bp, struct Process *proc) 789{ 790 switch (bp->libsym->arch.type) { 791 struct Process *leader; 792 void (*on_all_stopped)(struct process_stopping_handler *); 793 enum callback_status (*keep_stepping_p) 794 (struct process_stopping_handler *); 795 796 case PPC_DEFAULT: 797 assert(proc->e_machine == EM_PPC); 798 assert(bp->libsym != NULL); 799 assert(bp->libsym->lib->arch.bss_plt_prelinked == 0); 800 /* fall-through */ 801 802 case PPC_PLT_UNRESOLVED: 803 on_all_stopped = NULL; 804 keep_stepping_p = NULL; 805 leader = proc->leader; 806 807 if (leader != NULL && leader->arch.dl_plt_update_bp != NULL 808 && breakpoint_turn_on(leader->arch.dl_plt_update_bp, 809 proc) >= 0) 810 on_all_stopped = cb_on_all_stopped; 811 else 812 keep_stepping_p = cb_keep_stepping_p; 813 814 if (process_install_stopping_handler 815 (proc, bp, on_all_stopped, keep_stepping_p, NULL) < 0) { 816 error(0, 0, "ppc_plt_bp_continue: couldn't install" 817 " event handler"); 818 continue_after_breakpoint(proc, bp); 819 } 820 return; 821 822 case PPC_PLT_RESOLVED: 823 if (proc->e_machine == EM_PPC) { 824 continue_after_breakpoint(proc, bp); 825 return; 826 } 827 828 jump_to_entry_point(proc, bp); 829 continue_process(proc->pid); 830 return; 831 832 case PPC64_PLT_STUB: 833 /* These should never hit here. */ 834 break; 835 } 836 837 assert(bp->libsym->arch.type != bp->libsym->arch.type); 838 abort(); 839} 840 841/* When a process is in a PLT stub, it may have already read the data 842 * in .plt that we changed. If we detach now, it will jump to PLT 843 * entry and continue to the dynamic linker, where it will SIGSEGV, 844 * because zeroth .plt slot is not filled in prelinked binaries, and 845 * the dynamic linker needs that data. Moreover, the process may 846 * actually have hit the breakpoint already. This functions tries to 847 * detect both cases and do any fix-ups necessary to mend this 848 * situation. */ 849static enum callback_status 850detach_task_cb(struct Process *task, void *data) 851{ 852 struct breakpoint *bp = data; 853 854 if (get_instruction_pointer(task) == bp->addr) { 855 debug(DEBUG_PROCESS, "%d at %p, which is PLT slot", 856 task->pid, bp->addr); 857 jump_to_entry_point(task, bp); 858 return CBS_CONT; 859 } 860 861 /* XXX There's still a window of several instructions where we 862 * might catch the task inside a stub such that it has already 863 * read destination address from .plt, but hasn't jumped yet, 864 * thus avoiding the breakpoint. */ 865 866 return CBS_CONT; 867} 868 869static void 870ppc_plt_bp_retract(struct breakpoint *bp, struct Process *proc) 871{ 872 /* On PPC64, we rewrite .plt with PLT entry addresses. This 873 * needs to be undone. Unfortunately, the program may have 874 * made decisions based on that value */ 875 if (proc->e_machine == EM_PPC64 876 && bp->libsym != NULL 877 && bp->libsym->arch.type == PPC_PLT_RESOLVED) { 878 each_task(proc->leader, NULL, detach_task_cb, bp); 879 unresolve_plt_slot(proc, bp->libsym->arch.plt_slot_addr, 880 bp->libsym->arch.resolved_value); 881 } 882} 883 884void 885arch_library_init(struct library *lib) 886{ 887} 888 889void 890arch_library_destroy(struct library *lib) 891{ 892} 893 894void 895arch_library_clone(struct library *retp, struct library *lib) 896{ 897} 898 899int 900arch_library_symbol_init(struct library_symbol *libsym) 901{ 902 /* We set type explicitly in the code above, where we have the 903 * necessary context. This is for calls from ltrace-elf.c and 904 * such. */ 905 libsym->arch.type = PPC_DEFAULT; 906 return 0; 907} 908 909void 910arch_library_symbol_destroy(struct library_symbol *libsym) 911{ 912} 913 914int 915arch_library_symbol_clone(struct library_symbol *retp, 916 struct library_symbol *libsym) 917{ 918 retp->arch = libsym->arch; 919 return 0; 920} 921 922/* For some symbol types, we need to set up custom callbacks. XXX we 923 * don't need PROC here, we can store the data in BP if it is of 924 * interest to us. */ 925int 926arch_breakpoint_init(struct Process *proc, struct breakpoint *bp) 927{ 928 /* Artificial and entry-point breakpoints are plain. */ 929 if (bp->libsym == NULL || bp->libsym->plt_type != LS_TOPLT_EXEC) 930 return 0; 931 932 /* On PPC, secure PLT and prelinked BSS PLT are plain. */ 933 if (proc->e_machine == EM_PPC 934 && bp->libsym->lib->arch.bss_plt_prelinked != 0) 935 return 0; 936 937 /* On PPC64, stub PLT breakpoints are plain. */ 938 if (proc->e_machine == EM_PPC64 939 && bp->libsym->arch.type == PPC64_PLT_STUB) 940 return 0; 941 942 static struct bp_callbacks cbs = { 943 .on_continue = ppc_plt_bp_continue, 944 .on_retract = ppc_plt_bp_retract, 945 }; 946 breakpoint_set_callbacks(bp, &cbs); 947 return 0; 948} 949 950void 951arch_breakpoint_destroy(struct breakpoint *bp) 952{ 953} 954 955int 956arch_breakpoint_clone(struct breakpoint *retp, struct breakpoint *sbp) 957{ 958 retp->arch = sbp->arch; 959 return 0; 960} 961 962int 963arch_process_init(struct Process *proc) 964{ 965 proc->arch.dl_plt_update_bp = NULL; 966 proc->arch.handler = NULL; 967 return 0; 968} 969 970void 971arch_process_destroy(struct Process *proc) 972{ 973} 974 975int 976arch_process_clone(struct Process *retp, struct Process *proc) 977{ 978 retp->arch = proc->arch; 979 return 0; 980} 981 982int 983arch_process_exec(struct Process *proc) 984{ 985 return arch_process_init(proc); 986} 987