plt.c revision 02a41a5a996ffb774078cafe6311150720d2fd5b
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 219int 220arch_translate_address(struct Process *proc, 221 target_address_t addr, target_address_t *ret) 222{ 223 if (proc->e_machine == EM_PPC64) { 224 assert(host_powerpc64()); 225 long l = ptrace(PTRACE_PEEKTEXT, proc->pid, addr, 0); 226 if (l == -1 && errno) { 227 error(0, errno, ".opd translation of %p", addr); 228 return -1; 229 } 230 *ret = (target_address_t)l; 231 return 0; 232 } 233 234 *ret = addr; 235 return 0; 236} 237 238void * 239sym2addr(struct Process *proc, struct library_symbol *sym) 240{ 241 return sym->enter_addr; 242} 243 244static GElf_Addr 245get_glink_vma(struct ltelf *lte, GElf_Addr ppcgot, Elf_Data *plt_data) 246{ 247 Elf_Scn *ppcgot_sec = NULL; 248 GElf_Shdr ppcgot_shdr; 249 if (ppcgot != 0 250 && elf_get_section_covering(lte, ppcgot, 251 &ppcgot_sec, &ppcgot_shdr) < 0) 252 error(0, 0, "DT_PPC_GOT=%#"PRIx64", but no such section found", 253 ppcgot); 254 255 if (ppcgot_sec != NULL) { 256 Elf_Data *data = elf_loaddata(ppcgot_sec, &ppcgot_shdr); 257 if (data == NULL || data->d_size < 8 ) { 258 error(0, 0, "couldn't read GOT data"); 259 } else { 260 // where PPCGOT begins in .got 261 size_t offset = ppcgot - ppcgot_shdr.sh_addr; 262 assert(offset % 4 == 0); 263 uint32_t glink_vma; 264 if (elf_read_u32(data, offset + 4, &glink_vma) < 0) { 265 error(0, 0, "couldn't read glink VMA address" 266 " at %zd@GOT", offset); 267 return 0; 268 } 269 if (glink_vma != 0) { 270 debug(1, "PPC GOT glink_vma address: %#" PRIx32, 271 glink_vma); 272 return (GElf_Addr)glink_vma; 273 } 274 } 275 } 276 277 if (plt_data != NULL) { 278 uint32_t glink_vma; 279 if (elf_read_u32(plt_data, 0, &glink_vma) < 0) { 280 error(0, 0, "couldn't read glink VMA address"); 281 return 0; 282 } 283 debug(1, ".plt glink_vma address: %#" PRIx32, glink_vma); 284 return (GElf_Addr)glink_vma; 285 } 286 287 return 0; 288} 289 290static int 291load_dynamic_entry(struct ltelf *lte, int tag, GElf_Addr *valuep) 292{ 293 Elf_Scn *scn; 294 GElf_Shdr shdr; 295 if (elf_get_section_type(lte, SHT_DYNAMIC, &scn, &shdr) < 0 296 || scn == NULL) { 297 fail: 298 error(0, 0, "Couldn't get SHT_DYNAMIC: %s", 299 elf_errmsg(-1)); 300 return -1; 301 } 302 303 Elf_Data *data = elf_loaddata(scn, &shdr); 304 if (data == NULL) 305 goto fail; 306 307 size_t j; 308 for (j = 0; j < shdr.sh_size / shdr.sh_entsize; ++j) { 309 GElf_Dyn dyn; 310 if (gelf_getdyn(data, j, &dyn) == NULL) 311 goto fail; 312 313 if(dyn.d_tag == tag) { 314 *valuep = dyn.d_un.d_ptr; 315 return 0; 316 } 317 } 318 319 return -1; 320} 321 322static int 323load_ppcgot(struct ltelf *lte, GElf_Addr *ppcgotp) 324{ 325 return load_dynamic_entry(lte, DT_PPC_GOT, ppcgotp); 326} 327 328static int 329load_ppc64_glink(struct ltelf *lte, GElf_Addr *glinkp) 330{ 331 return load_dynamic_entry(lte, DT_PPC64_GLINK, glinkp); 332} 333 334static int 335nonzero_data(Elf_Data *data) 336{ 337 /* We are not supposed to get here if there's no PLT. */ 338 assert(data != NULL); 339 340 unsigned char *buf = data->d_buf; 341 if (buf == NULL) 342 return 0; 343 344 size_t i; 345 for (i = 0; i < data->d_size; ++i) 346 if (buf[i] != 0) 347 return 1; 348 return 0; 349} 350 351int 352arch_elf_init(struct ltelf *lte, struct library *lib) 353{ 354 lte->arch.secure_plt = !(lte->plt_flags & SHF_EXECINSTR); 355 356 /* For PPC32 BSS, it is important whether the binary was 357 * prelinked. If .plt section is NODATA, or if it contains 358 * zeroes, then this library is not prelinked, and we need to 359 * delay breakpoints. */ 360 if (lte->ehdr.e_machine == EM_PPC && !lte->arch.secure_plt) 361 lib->arch.bss_plt_prelinked = nonzero_data(lte->plt_data); 362 else 363 /* For cases where it's irrelevant, initialize the 364 * value to something conspicuous. */ 365 lib->arch.bss_plt_prelinked = -1; 366 367 if (lte->ehdr.e_machine == EM_PPC && lte->arch.secure_plt) { 368 GElf_Addr ppcgot; 369 if (load_ppcgot(lte, &ppcgot) < 0) { 370 error(0, 0, "couldn't find DT_PPC_GOT"); 371 return -1; 372 } 373 GElf_Addr glink_vma = get_glink_vma(lte, ppcgot, lte->plt_data); 374 375 assert (lte->relplt_size % 12 == 0); 376 size_t count = lte->relplt_size / 12; // size of RELA entry 377 lte->arch.plt_stub_vma = glink_vma 378 - (GElf_Addr)count * PPC_PLT_STUB_SIZE; 379 debug(1, "stub_vma is %#" PRIx64, lte->arch.plt_stub_vma); 380 381 } else if (lte->ehdr.e_machine == EM_PPC64) { 382 GElf_Addr glink_vma; 383 if (load_ppc64_glink(lte, &glink_vma) < 0) { 384 error(0, 0, "couldn't find DT_PPC64_GLINK"); 385 return -1; 386 } 387 388 /* The first glink stub starts at offset 32. */ 389 lte->arch.plt_stub_vma = glink_vma + 32; 390 } 391 392 /* On PPC64, look for stub symbols in symbol table. These are 393 * called: xxxxxxxx.plt_call.callee_name@version+addend. */ 394 if (lte->ehdr.e_machine == EM_PPC64 395 && lte->symtab != NULL && lte->strtab != NULL) { 396 397 /* N.B. We can't simply skip the symbols that we fail 398 * to read or malloc. There may be more than one stub 399 * per symbol name, and if we failed in one but 400 * succeeded in another, the PLT enabling code would 401 * have no way to tell that something is missing. We 402 * could work around that, of course, but it doesn't 403 * seem worth the trouble. So if anything fails, we 404 * just pretend that we don't have stub symbols at 405 * all, as if the binary is stripped. */ 406 407 size_t i; 408 for (i = 0; i < lte->symtab_count; ++i) { 409 GElf_Sym sym; 410 if (gelf_getsym(lte->symtab, i, &sym) == NULL) { 411 struct library_symbol *sym, *next; 412 fail: 413 for (sym = lte->arch.stubs; sym != NULL; ) { 414 next = sym->next; 415 library_symbol_destroy(sym); 416 free(sym); 417 sym = next; 418 } 419 lte->arch.stubs = NULL; 420 break; 421 } 422 423 const char *name = lte->strtab + sym.st_name; 424 425#define STUBN ".plt_call." 426 if ((name = strstr(name, STUBN)) == NULL) 427 continue; 428 name += sizeof(STUBN) - 1; 429#undef STUBN 430 431 size_t len; 432 const char *ver = strchr(name, '@'); 433 if (ver != NULL) { 434 len = ver - name; 435 436 } else { 437 /* If there is "+" at all, check that 438 * the symbol name ends in "+0". */ 439 const char *add = strrchr(name, '+'); 440 if (add != NULL) { 441 assert(strcmp(add, "+0") == 0); 442 len = add - name; 443 } else { 444 len = strlen(name); 445 } 446 } 447 448 char *sym_name = strndup(name, len); 449 struct library_symbol *libsym = malloc(sizeof(*libsym)); 450 if (sym_name == NULL || libsym == NULL) { 451 fail2: 452 free(sym_name); 453 free(libsym); 454 goto fail; 455 } 456 457 /* XXX The double cast should be removed when 458 * target_address_t becomes integral type. */ 459 target_address_t addr = (target_address_t) 460 (uintptr_t)sym.st_value + lte->bias; 461 if (library_symbol_init(libsym, addr, sym_name, 1, 462 LS_TOPLT_EXEC) < 0) 463 goto fail2; 464 libsym->arch.type = PPC64_PLT_STUB; 465 libsym->next = lte->arch.stubs; 466 lte->arch.stubs = libsym; 467 } 468 } 469 470 return 0; 471} 472 473static int 474read_plt_slot_value(struct Process *proc, GElf_Addr addr, GElf_Addr *valp) 475{ 476 /* On PPC64, we read from .plt, which contains 8 byte 477 * addresses. On PPC32 we read from .plt, which contains 4 478 * byte instructions, but the PLT is two instructions, and 479 * either can change. */ 480 uint64_t l; 481 /* XXX double cast. */ 482 if (read_target_8(proc, (target_address_t)(uintptr_t)addr, &l) < 0) { 483 error(0, errno, "ptrace .plt slot value @%#" PRIx64, addr); 484 return -1; 485 } 486 487 *valp = (GElf_Addr)l; 488 return 0; 489} 490 491static int 492unresolve_plt_slot(struct Process *proc, GElf_Addr addr, GElf_Addr value) 493{ 494 /* We only modify plt_entry[0], which holds the resolved 495 * address of the routine. We keep the TOC and environment 496 * pointers intact. Hence the only adjustment that we need to 497 * do is to IP. */ 498 if (ptrace(PTRACE_POKETEXT, proc->pid, addr, value) < 0) { 499 error(0, errno, "unresolve .plt slot"); 500 return -1; 501 } 502 return 0; 503} 504 505static void 506mark_as_resolved(struct library_symbol *libsym, GElf_Addr value) 507{ 508 libsym->arch.type = PPC_PLT_RESOLVED; 509 libsym->arch.resolved_value = value; 510} 511 512enum plt_status 513arch_elf_add_plt_entry(struct Process *proc, struct ltelf *lte, 514 const char *a_name, GElf_Rela *rela, size_t ndx, 515 struct library_symbol **ret) 516{ 517 if (lte->ehdr.e_machine == EM_PPC) 518 return plt_default; 519 520 /* PPC64. If we have stubs, we return a chain of breakpoint 521 * sites, one for each stub that corresponds to this PLT 522 * entry. */ 523 struct library_symbol *chain = NULL; 524 struct library_symbol **symp; 525 for (symp = <e->arch.stubs; *symp != NULL; ) { 526 struct library_symbol *sym = *symp; 527 if (strcmp(sym->name, a_name) != 0) { 528 symp = &(*symp)->next; 529 continue; 530 } 531 532 /* Re-chain the symbol from stubs to CHAIN. */ 533 *symp = sym->next; 534 sym->next = chain; 535 chain = sym; 536 } 537 538 if (chain != NULL) { 539 *ret = chain; 540 return plt_ok; 541 } 542 543 /* We don't have stub symbols. Find corresponding .plt slot, 544 * and check whether it contains the corresponding PLT address 545 * (or 0 if the dynamic linker hasn't run yet). N.B. we don't 546 * want read this from ELF file, but from process image. That 547 * makes a difference if we are attaching to a running 548 * process. */ 549 550 GElf_Addr plt_entry_addr = arch_plt_sym_val(lte, ndx, rela); 551 GElf_Addr plt_slot_addr = rela->r_offset; 552 assert(plt_slot_addr >= lte->plt_addr 553 || plt_slot_addr < lte->plt_addr + lte->plt_size); 554 555 GElf_Addr plt_slot_value; 556 if (read_plt_slot_value(proc, plt_slot_addr, &plt_slot_value) < 0) 557 return plt_fail; 558 559 char *name = strdup(a_name); 560 struct library_symbol *libsym = malloc(sizeof(*libsym)); 561 if (name == NULL || libsym == NULL) { 562 error(0, errno, "allocation for .plt slot"); 563 fail: 564 free(name); 565 free(libsym); 566 return plt_fail; 567 } 568 569 /* XXX The double cast should be removed when 570 * target_address_t becomes integral type. */ 571 if (library_symbol_init(libsym, 572 (target_address_t)(uintptr_t)plt_entry_addr, 573 name, 1, LS_TOPLT_EXEC) < 0) 574 goto fail; 575 libsym->arch.plt_slot_addr = plt_slot_addr; 576 577 if (plt_slot_value == plt_entry_addr || plt_slot_value == 0) { 578 libsym->arch.type = PPC_PLT_UNRESOLVED; 579 libsym->arch.resolved_value = plt_entry_addr; 580 581 } else { 582 /* Unresolve the .plt slot. If the binary was 583 * prelinked, this makes the code invalid, because in 584 * case of prelinked binary, the dynamic linker 585 * doesn't update .plt[0] and .plt[1] with addresses 586 * of the resover. But we don't care, we will never 587 * need to enter the resolver. That just means that 588 * we have to un-un-resolve this back before we 589 * detach. */ 590 591 if (unresolve_plt_slot(proc, plt_slot_addr, plt_entry_addr) < 0) { 592 library_symbol_destroy(libsym); 593 goto fail; 594 } 595 mark_as_resolved(libsym, plt_slot_value); 596 } 597 598 *ret = libsym; 599 return plt_ok; 600} 601 602void 603arch_elf_destroy(struct ltelf *lte) 604{ 605 struct library_symbol *sym; 606 for (sym = lte->arch.stubs; sym != NULL; ) { 607 struct library_symbol *next = sym->next; 608 library_symbol_destroy(sym); 609 free(sym); 610 sym = next; 611 } 612} 613 614static void 615dl_plt_update_bp_on_hit(struct breakpoint *bp, struct Process *proc) 616{ 617 debug(DEBUG_PROCESS, "pid=%d dl_plt_update_bp_on_hit %s(%p)", 618 proc->pid, breakpoint_name(bp), bp->addr); 619 struct process_stopping_handler *self = proc->arch.handler; 620 assert(self != NULL); 621 622 struct library_symbol *libsym = self->breakpoint_being_enabled->libsym; 623 GElf_Addr value; 624 if (read_plt_slot_value(proc, libsym->arch.plt_slot_addr, &value) < 0) 625 return; 626 627 /* On PPC64, we rewrite the slot value. */ 628 if (proc->e_machine == EM_PPC64) 629 unresolve_plt_slot(proc, libsym->arch.plt_slot_addr, 630 libsym->arch.resolved_value); 631 /* We mark the breakpoint as resolved on both arches. */ 632 mark_as_resolved(libsym, value); 633 634 /* cb_on_all_stopped looks if HANDLER is set to NULL as a way 635 * to check that this was run. It's an error if it 636 * wasn't. */ 637 proc->arch.handler = NULL; 638 639 breakpoint_turn_off(bp, proc); 640} 641 642static void 643cb_on_all_stopped(struct process_stopping_handler *self) 644{ 645 /* Put that in for dl_plt_update_bp_on_hit to see. */ 646 assert(self->task_enabling_breakpoint->arch.handler == NULL); 647 self->task_enabling_breakpoint->arch.handler = self; 648 649 linux_ptrace_disable_and_continue(self); 650} 651 652static enum callback_status 653cb_keep_stepping_p(struct process_stopping_handler *self) 654{ 655 struct Process *proc = self->task_enabling_breakpoint; 656 struct library_symbol *libsym = self->breakpoint_being_enabled->libsym; 657 658 GElf_Addr value; 659 if (read_plt_slot_value(proc, libsym->arch.plt_slot_addr, &value) < 0) 660 return CBS_FAIL; 661 662 /* In UNRESOLVED state, the RESOLVED_VALUE in fact contains 663 * the PLT entry value. */ 664 if (value == libsym->arch.resolved_value) 665 return CBS_CONT; 666 667 debug(DEBUG_PROCESS, "pid=%d PLT got resolved to value %#"PRIx64, 668 proc->pid, value); 669 670 /* The .plt slot got resolved! We can migrate the breakpoint 671 * to RESOLVED and stop single-stepping. */ 672 if (proc->e_machine == EM_PPC64 673 && unresolve_plt_slot(proc, libsym->arch.plt_slot_addr, 674 libsym->arch.resolved_value) < 0) 675 return CBS_FAIL; 676 677 /* Resolving on PPC64 consists of overwriting a doubleword in 678 * .plt. That doubleword is than read back by a stub, and 679 * jumped on. Hopefully we can assume that double word update 680 * is done on a single place only, as it contains a final 681 * address. We still need to look around for any sync 682 * instruction, but essentially it is safe to optimize away 683 * the single stepping next time and install a post-update 684 * breakpoint. 685 * 686 * The situation on PPC32 BSS is more complicated. The 687 * dynamic linker here updates potentially several 688 * instructions (XXX currently we assume two) and the rules 689 * are more complicated. Sometimes it's enough to adjust just 690 * one of the addresses--the logic for generating optimal 691 * dispatch depends on relative addresses of the .plt entry 692 * and the jump destination. We can't assume that the some 693 * instruction block does the update every time. So on PPC32, 694 * we turn the optimization off and just step through it each 695 * time. */ 696 if (proc->e_machine == EM_PPC) 697 goto done; 698 699 /* Install breakpoint to the address where the change takes 700 * place. If we fail, then that just means that we'll have to 701 * singlestep the next time around as well. */ 702 struct Process *leader = proc->leader; 703 if (leader == NULL || leader->arch.dl_plt_update_bp != NULL) 704 goto done; 705 706 /* We need to install to the next instruction. ADDR points to 707 * a store instruction, so moving the breakpoint one 708 * instruction forward is safe. */ 709 target_address_t addr = get_instruction_pointer(proc) + 4; 710 leader->arch.dl_plt_update_bp = insert_breakpoint(proc, addr, NULL); 711 if (leader->arch.dl_plt_update_bp == NULL) 712 goto done; 713 714 static struct bp_callbacks dl_plt_update_cbs = { 715 .on_hit = dl_plt_update_bp_on_hit, 716 }; 717 leader->arch.dl_plt_update_bp->cbs = &dl_plt_update_cbs; 718 719 /* Turn it off for now. We will turn it on again when we hit 720 * the PLT entry that needs this. */ 721 breakpoint_turn_off(leader->arch.dl_plt_update_bp, proc); 722 723done: 724 mark_as_resolved(libsym, value); 725 726 return CBS_STOP; 727} 728 729static void 730ppc_plt_bp_continue(struct breakpoint *bp, struct Process *proc) 731{ 732 switch (bp->libsym->arch.type) { 733 target_address_t rv; 734 struct Process *leader; 735 void (*on_all_stopped)(struct process_stopping_handler *); 736 enum callback_status (*keep_stepping_p) 737 (struct process_stopping_handler *); 738 739 case PPC_DEFAULT: 740 assert(proc->e_machine == EM_PPC); 741 assert(bp->libsym != NULL); 742 assert(bp->libsym->lib->arch.bss_plt_prelinked == 0); 743 /* fall-through */ 744 745 case PPC_PLT_UNRESOLVED: 746 on_all_stopped = NULL; 747 keep_stepping_p = NULL; 748 leader = proc->leader; 749 750 if (leader != NULL && leader->arch.dl_plt_update_bp != NULL 751 && breakpoint_turn_on(leader->arch.dl_plt_update_bp, 752 proc) >= 0) 753 on_all_stopped = cb_on_all_stopped; 754 else 755 keep_stepping_p = cb_keep_stepping_p; 756 757 if (process_install_stopping_handler 758 (proc, bp, on_all_stopped, keep_stepping_p, NULL) < 0) { 759 error(0, 0, "ppc_plt_bp_continue: couldn't install" 760 " event handler"); 761 continue_after_breakpoint(proc, bp); 762 } 763 return; 764 765 case PPC_PLT_RESOLVED: 766 if (proc->e_machine == EM_PPC) { 767 continue_after_breakpoint(proc, bp); 768 return; 769 } 770 771 /* XXX The double cast should be removed when 772 * target_address_t becomes integral type. */ 773 rv = (target_address_t) 774 (uintptr_t)bp->libsym->arch.resolved_value; 775 set_instruction_pointer(proc, rv); 776 continue_process(proc->pid); 777 return; 778 779 case PPC64_PLT_STUB: 780 /* These should never hit here. */ 781 break; 782 } 783 784 assert(bp->libsym->arch.type != bp->libsym->arch.type); 785 abort(); 786} 787 788void 789arch_library_init(struct library *lib) 790{ 791} 792 793void 794arch_library_destroy(struct library *lib) 795{ 796} 797 798void 799arch_library_clone(struct library *retp, struct library *lib) 800{ 801} 802 803int 804arch_library_symbol_init(struct library_symbol *libsym) 805{ 806 /* We set type explicitly in the code above, where we have the 807 * necessary context. This is for calls from ltrace-elf.c and 808 * such. */ 809 libsym->arch.type = PPC_DEFAULT; 810 return 0; 811} 812 813void 814arch_library_symbol_destroy(struct library_symbol *libsym) 815{ 816} 817 818int 819arch_library_symbol_clone(struct library_symbol *retp, 820 struct library_symbol *libsym) 821{ 822 retp->arch = libsym->arch; 823 return 0; 824} 825 826/* For some symbol types, we need to set up custom callbacks. XXX we 827 * don't need PROC here, we can store the data in BP if it is of 828 * interest to us. */ 829int 830arch_breakpoint_init(struct Process *proc, struct breakpoint *bp) 831{ 832 /* Artificial and entry-point breakpoints are plain. */ 833 if (bp->libsym == NULL || bp->libsym->plt_type != LS_TOPLT_EXEC) 834 return 0; 835 836 /* On PPC, secure PLT and prelinked BSS PLT are plain. */ 837 if (proc->e_machine == EM_PPC 838 && bp->libsym->lib->arch.bss_plt_prelinked != 0) 839 return 0; 840 841 /* On PPC64, stub PLT breakpoints are plain. */ 842 if (proc->e_machine == EM_PPC64 843 && bp->libsym->arch.type == PPC64_PLT_STUB) 844 return 0; 845 846 static struct bp_callbacks cbs = { 847 .on_continue = ppc_plt_bp_continue, 848 }; 849 breakpoint_set_callbacks(bp, &cbs); 850 return 0; 851} 852 853void 854arch_breakpoint_destroy(struct breakpoint *bp) 855{ 856} 857 858int 859arch_breakpoint_clone(struct breakpoint *retp, struct breakpoint *sbp) 860{ 861 retp->arch = sbp->arch; 862 return 0; 863} 864 865int 866arch_process_init(struct Process *proc) 867{ 868 proc->arch.dl_plt_update_bp = NULL; 869 proc->arch.handler = NULL; 870 return 0; 871} 872 873void 874arch_process_destroy(struct Process *proc) 875{ 876} 877 878int 879arch_process_clone(struct Process *retp, struct Process *proc) 880{ 881 retp->arch = proc->arch; 882 return 0; 883} 884 885int 886arch_process_exec(struct Process *proc) 887{ 888 return arch_process_init(proc); 889} 890