1 /*- 2 * Copyright (c) 2000 David O'Brien 3 * Copyright (c) 1995-1996 Søren Schmidt 4 * Copyright (c) 1996 Peter Wemm 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer 12 * in this position and unchanged. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. The name of the author may not be used to endorse or promote products 17 * derived from this software without specific prior written permission 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 * 30 * $FreeBSD: src/sys/kern/imgact_elf.c,v 1.73.2.13 2002/12/28 19:49:41 dillon Exp $ 31 */ 32 33 #include <sys/param.h> 34 #include <sys/exec.h> 35 #include <sys/fcntl.h> 36 #include <sys/file.h> 37 #include <sys/imgact.h> 38 #include <sys/imgact_elf.h> 39 #include <sys/kernel.h> 40 #include <sys/malloc.h> 41 #include <sys/mman.h> 42 #include <sys/systm.h> 43 #include <sys/proc.h> 44 #include <sys/nlookup.h> 45 #include <sys/pioctl.h> 46 #include <sys/procfs.h> 47 #include <sys/resourcevar.h> 48 #include <sys/signalvar.h> 49 #include <sys/stat.h> 50 #include <sys/syscall.h> 51 #include <sys/sysctl.h> 52 #include <sys/sysent.h> 53 #include <sys/vnode.h> 54 #include <sys/eventhandler.h> 55 56 #include <cpu/lwbuf.h> 57 58 #include <vm/vm.h> 59 #include <vm/vm_kern.h> 60 #include <vm/vm_param.h> 61 #include <vm/pmap.h> 62 #include <sys/lock.h> 63 #include <vm/vm_map.h> 64 #include <vm/vm_object.h> 65 #include <vm/vm_extern.h> 66 67 #include <machine/elf.h> 68 #include <machine/md_var.h> 69 #include <sys/mount.h> 70 #include <sys/ckpt.h> 71 72 #define OLD_EI_BRAND 8 73 #define truncps(va,ps) ((va) & ~(ps - 1)) 74 #define aligned(a,t) (truncps((u_long)(a), sizeof(t)) == (u_long)(a)) 75 76 static int __elfN(check_header)(const Elf_Ehdr *hdr); 77 static Elf_Brandinfo *__elfN(get_brandinfo)(struct image_params *imgp, 78 const char *interp, int32_t *osrel); 79 static int __elfN(load_file)(struct proc *p, const char *file, u_long *addr, 80 u_long *entry); 81 static int __elfN(load_section)(struct proc *p, 82 struct vmspace *vmspace, struct vnode *vp, 83 vm_offset_t offset, caddr_t vmaddr, size_t memsz, size_t filsz, 84 vm_prot_t prot); 85 static int __CONCAT(exec_, __elfN(imgact))(struct image_params *imgp); 86 static boolean_t __elfN(bsd_trans_osrel)(const Elf_Note *note, 87 int32_t *osrel); 88 static boolean_t __elfN(check_note)(struct image_params *imgp, 89 Elf_Brandnote *checknote, int32_t *osrel); 90 static vm_prot_t __elfN(trans_prot)(Elf_Word); 91 static Elf_Word __elfN(untrans_prot)(vm_prot_t); 92 static boolean_t check_PT_NOTE(struct image_params *imgp, 93 Elf_Brandnote *checknote, int32_t *osrel, const Elf_Phdr * pnote); 94 static boolean_t extract_interpreter(struct image_params *imgp, 95 const Elf_Phdr *pinterpreter, char *data); 96 static u_long pie_base_hint(struct proc *p); 97 98 static int elf_legacy_coredump = 0; 99 static int __elfN(fallback_brand) = -1; 100 static int elf_pie_base_mmap = 0; 101 #if defined(__x86_64__) 102 SYSCTL_NODE(_kern, OID_AUTO, elf64, CTLFLAG_RW, 0, ""); 103 SYSCTL_INT(_debug, OID_AUTO, elf64_legacy_coredump, CTLFLAG_RW, 104 &elf_legacy_coredump, 0, "legacy coredump mode"); 105 SYSCTL_INT(_kern_elf64, OID_AUTO, fallback_brand, CTLFLAG_RW, 106 &elf64_fallback_brand, 0, "ELF64 brand of last resort"); 107 TUNABLE_INT("kern.elf64.fallback_brand", &elf64_fallback_brand); 108 SYSCTL_INT(_kern_elf64, OID_AUTO, pie_base_mmap, CTLFLAG_RW, 109 &elf_pie_base_mmap, 0, 110 "choose a base address for PIE as if it is mapped with mmap()"); 111 TUNABLE_INT("kern.elf64.pie_base_mmap", &elf_pie_base_mmap); 112 #else /* i386 assumed */ 113 SYSCTL_NODE(_kern, OID_AUTO, elf32, CTLFLAG_RW, 0, ""); 114 SYSCTL_INT(_debug, OID_AUTO, elf32_legacy_coredump, CTLFLAG_RW, 115 &elf_legacy_coredump, 0, "legacy coredump mode"); 116 SYSCTL_INT(_kern_elf32, OID_AUTO, fallback_brand, CTLFLAG_RW, 117 &elf32_fallback_brand, 0, "ELF32 brand of last resort"); 118 TUNABLE_INT("kern.elf32.fallback_brand", &elf32_fallback_brand); 119 SYSCTL_INT(_kern_elf32, OID_AUTO, pie_base_mmap, CTLFLAG_RW, 120 &elf_pie_base_mmap, 0, 121 "choose a base address for PIE as if it is mapped with mmap()"); 122 TUNABLE_INT("kern.elf32.pie_base_mmap", &elf_pie_base_mmap); 123 #endif 124 125 static Elf_Brandinfo *elf_brand_list[MAX_BRANDS]; 126 127 static const char DRAGONFLY_ABI_VENDOR[] = "DragonFly"; 128 static const char FREEBSD_ABI_VENDOR[] = "FreeBSD"; 129 130 Elf_Brandnote __elfN(dragonfly_brandnote) = { 131 .hdr.n_namesz = sizeof(DRAGONFLY_ABI_VENDOR), 132 .hdr.n_descsz = sizeof(int32_t), 133 .hdr.n_type = 1, 134 .vendor = DRAGONFLY_ABI_VENDOR, 135 .flags = BN_TRANSLATE_OSREL, 136 .trans_osrel = __elfN(bsd_trans_osrel), 137 }; 138 139 Elf_Brandnote __elfN(freebsd_brandnote) = { 140 .hdr.n_namesz = sizeof(FREEBSD_ABI_VENDOR), 141 .hdr.n_descsz = sizeof(int32_t), 142 .hdr.n_type = 1, 143 .vendor = FREEBSD_ABI_VENDOR, 144 .flags = BN_TRANSLATE_OSREL, 145 .trans_osrel = __elfN(bsd_trans_osrel), 146 }; 147 148 int 149 __elfN(insert_brand_entry)(Elf_Brandinfo *entry) 150 { 151 int i; 152 153 for (i = 0; i < MAX_BRANDS; i++) { 154 if (elf_brand_list[i] == NULL) { 155 elf_brand_list[i] = entry; 156 break; 157 } 158 } 159 if (i == MAX_BRANDS) { 160 uprintf("WARNING: %s: could not insert brandinfo entry: %p\n", 161 __func__, entry); 162 return (-1); 163 } 164 return (0); 165 } 166 167 int 168 __elfN(remove_brand_entry)(Elf_Brandinfo *entry) 169 { 170 int i; 171 172 for (i = 0; i < MAX_BRANDS; i++) { 173 if (elf_brand_list[i] == entry) { 174 elf_brand_list[i] = NULL; 175 break; 176 } 177 } 178 if (i == MAX_BRANDS) 179 return (-1); 180 return (0); 181 } 182 183 /* 184 * Check if an elf brand is being used anywhere in the system. 185 * 186 * Used by the linux emulation module unloader. This isn't safe from 187 * races. 188 */ 189 struct elf_brand_inuse_info { 190 int rval; 191 Elf_Brandinfo *entry; 192 }; 193 194 static int elf_brand_inuse_callback(struct proc *p, void *data); 195 196 int 197 __elfN(brand_inuse)(Elf_Brandinfo *entry) 198 { 199 struct elf_brand_inuse_info info; 200 201 info.rval = FALSE; 202 info.entry = entry; 203 allproc_scan(elf_brand_inuse_callback, &info, 0); 204 return (info.rval); 205 } 206 207 static 208 int 209 elf_brand_inuse_callback(struct proc *p, void *data) 210 { 211 struct elf_brand_inuse_info *info = data; 212 213 if (p->p_sysent == info->entry->sysvec) { 214 info->rval = TRUE; 215 return (-1); 216 } 217 return (0); 218 } 219 220 static int 221 __elfN(check_header)(const Elf_Ehdr *hdr) 222 { 223 Elf_Brandinfo *bi; 224 int i; 225 226 if (!IS_ELF(*hdr) || 227 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS || 228 hdr->e_ident[EI_DATA] != ELF_TARG_DATA || 229 hdr->e_ident[EI_VERSION] != EV_CURRENT || 230 hdr->e_phentsize != sizeof(Elf_Phdr) || 231 hdr->e_ehsize != sizeof(Elf_Ehdr) || 232 hdr->e_version != ELF_TARG_VER) 233 return (ENOEXEC); 234 235 /* 236 * Make sure we have at least one brand for this machine. 237 */ 238 239 for (i = 0; i < MAX_BRANDS; i++) { 240 bi = elf_brand_list[i]; 241 if (bi != NULL && bi->machine == hdr->e_machine) 242 break; 243 } 244 if (i == MAX_BRANDS) 245 return (ENOEXEC); 246 247 return (0); 248 } 249 250 static int 251 __elfN(load_section)(struct proc *p, struct vmspace *vmspace, struct vnode *vp, 252 vm_offset_t offset, caddr_t vmaddr, size_t memsz, 253 size_t filsz, vm_prot_t prot) 254 { 255 size_t map_len; 256 vm_offset_t map_addr; 257 int error, rv, cow; 258 int count; 259 int shared; 260 size_t copy_len; 261 vm_object_t object; 262 vm_offset_t file_addr; 263 264 object = vp->v_object; 265 error = 0; 266 267 /* 268 * In most cases we will be able to use a shared lock on the 269 * object we are inserting into the map. The lock will be 270 * upgraded in situations where new VM pages must be allocated. 271 */ 272 vm_object_hold_shared(object); 273 shared = 1; 274 275 /* 276 * It's necessary to fail if the filsz + offset taken from the 277 * header is greater than the actual file pager object's size. 278 * If we were to allow this, then the vm_map_find() below would 279 * walk right off the end of the file object and into the ether. 280 * 281 * While I'm here, might as well check for something else that 282 * is invalid: filsz cannot be greater than memsz. 283 */ 284 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) { 285 uprintf("elf_load_section: truncated ELF file\n"); 286 vm_object_drop(object); 287 return (ENOEXEC); 288 } 289 290 map_addr = trunc_page((vm_offset_t)vmaddr); 291 file_addr = trunc_page(offset); 292 293 /* 294 * We have two choices. We can either clear the data in the last page 295 * of an oversized mapping, or we can start the anon mapping a page 296 * early and copy the initialized data into that first page. We 297 * choose the second.. 298 */ 299 if (memsz > filsz) 300 map_len = trunc_page(offset+filsz) - file_addr; 301 else 302 map_len = round_page(offset+filsz) - file_addr; 303 304 if (map_len != 0) { 305 vm_object_reference_locked(object); 306 307 /* cow flags: don't dump readonly sections in core */ 308 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT; 309 if ((prot & VM_PROT_WRITE) == 0) 310 cow |= MAP_DISABLE_COREDUMP; 311 if (shared == 0) 312 cow |= MAP_PREFAULT_RELOCK; 313 314 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 315 vm_map_lock(&vmspace->vm_map); 316 rv = vm_map_insert(&vmspace->vm_map, &count, 317 object, NULL, 318 file_addr, /* file offset */ 319 map_addr, /* virtual start */ 320 map_addr + map_len,/* virtual end */ 321 VM_MAPTYPE_NORMAL, 322 VM_SUBSYS_IMGACT, 323 prot, VM_PROT_ALL, cow); 324 vm_map_unlock(&vmspace->vm_map); 325 vm_map_entry_release(count); 326 327 /* 328 * NOTE: Object must have a hold ref when calling 329 * vm_object_deallocate(). 330 */ 331 if (rv != KERN_SUCCESS) { 332 vm_object_drop(object); 333 vm_object_deallocate(object); 334 return (EINVAL); 335 } 336 337 /* we can stop now if we've covered it all */ 338 if (memsz == filsz) { 339 vm_object_drop(object); 340 return (0); 341 } 342 } 343 344 /* 345 * We have to get the remaining bit of the file into the first part 346 * of the oversized map segment. This is normally because the .data 347 * segment in the file is extended to provide bss. It's a neat idea 348 * to try and save a page, but it's a pain in the behind to implement. 349 */ 350 copy_len = (offset + filsz) - trunc_page(offset + filsz); 351 map_addr = trunc_page((vm_offset_t)vmaddr + filsz); 352 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr; 353 354 /* This had damn well better be true! */ 355 if (map_len != 0) { 356 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 357 vm_map_lock(&vmspace->vm_map); 358 rv = vm_map_insert(&vmspace->vm_map, &count, 359 NULL, NULL, 360 0, 361 map_addr, 362 map_addr + map_len, 363 VM_MAPTYPE_NORMAL, 364 VM_SUBSYS_IMGACT, 365 VM_PROT_ALL, VM_PROT_ALL, 0); 366 vm_map_unlock(&vmspace->vm_map); 367 vm_map_entry_release(count); 368 if (rv != KERN_SUCCESS) { 369 vm_object_drop(object); 370 return (EINVAL); 371 } 372 } 373 374 if (copy_len != 0) { 375 struct lwbuf *lwb; 376 struct lwbuf lwb_cache; 377 vm_page_t m; 378 379 m = vm_fault_object_page(object, trunc_page(offset + filsz), 380 VM_PROT_READ, 0, &shared, &error); 381 vm_object_drop(object); 382 if (m) { 383 lwb = lwbuf_alloc(m, &lwb_cache); 384 error = copyout((caddr_t)lwbuf_kva(lwb), 385 (caddr_t)map_addr, copy_len); 386 lwbuf_free(lwb); 387 vm_page_unhold(m); 388 } 389 } else { 390 vm_object_drop(object); 391 } 392 393 /* 394 * set it to the specified protection 395 */ 396 if (error == 0) { 397 vm_map_protect(&vmspace->vm_map, 398 map_addr, map_addr + map_len, 399 prot, FALSE); 400 } 401 return (error); 402 } 403 404 /* 405 * Load the file "file" into memory. It may be either a shared object 406 * or an executable. 407 * 408 * The "addr" reference parameter is in/out. On entry, it specifies 409 * the address where a shared object should be loaded. If the file is 410 * an executable, this value is ignored. On exit, "addr" specifies 411 * where the file was actually loaded. 412 * 413 * The "entry" reference parameter is out only. On exit, it specifies 414 * the entry point for the loaded file. 415 */ 416 static int 417 __elfN(load_file)(struct proc *p, const char *file, u_long *addr, u_long *entry) 418 { 419 struct { 420 struct nlookupdata nd; 421 struct vattr attr; 422 struct image_params image_params; 423 } *tempdata; 424 const Elf_Ehdr *hdr = NULL; 425 const Elf_Phdr *phdr = NULL; 426 struct nlookupdata *nd; 427 struct vmspace *vmspace = p->p_vmspace; 428 struct vattr *attr; 429 struct image_params *imgp; 430 struct mount *topmnt; 431 vm_prot_t prot; 432 u_long rbase; 433 u_long base_addr = 0; 434 int error, i, numsegs; 435 436 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_WAITOK); 437 nd = &tempdata->nd; 438 attr = &tempdata->attr; 439 imgp = &tempdata->image_params; 440 441 /* 442 * Initialize part of the common data 443 */ 444 imgp->proc = p; 445 imgp->attr = attr; 446 imgp->firstpage = NULL; 447 imgp->image_header = NULL; 448 imgp->vp = NULL; 449 450 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW); 451 if (error == 0) 452 error = nlookup(nd); 453 if (error == 0) 454 error = cache_vget(&nd->nl_nch, nd->nl_cred, 455 LK_SHARED, &imgp->vp); 456 topmnt = nd->nl_nch.mount; 457 nlookup_done(nd); 458 if (error) 459 goto fail; 460 461 /* 462 * Check permissions, modes, uid, etc on the file, and "open" it. 463 */ 464 error = exec_check_permissions(imgp, topmnt); 465 if (error) { 466 vn_unlock(imgp->vp); 467 goto fail; 468 } 469 470 error = exec_map_first_page(imgp); 471 /* 472 * Also make certain that the interpreter stays the same, so set 473 * its VTEXT flag, too. 474 */ 475 if (error == 0) 476 vsetflags(imgp->vp, VTEXT); 477 vn_unlock(imgp->vp); 478 if (error) 479 goto fail; 480 481 hdr = (const Elf_Ehdr *)imgp->image_header; 482 if ((error = __elfN(check_header)(hdr)) != 0) 483 goto fail; 484 if (hdr->e_type == ET_DYN) 485 rbase = *addr; 486 else if (hdr->e_type == ET_EXEC) 487 rbase = 0; 488 else { 489 error = ENOEXEC; 490 goto fail; 491 } 492 493 /* Only support headers that fit within first page for now */ 494 /* (multiplication of two Elf_Half fields will not overflow) */ 495 if ((hdr->e_phoff > PAGE_SIZE) || 496 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) { 497 error = ENOEXEC; 498 goto fail; 499 } 500 501 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff); 502 if (!aligned(phdr, Elf_Addr)) { 503 error = ENOEXEC; 504 goto fail; 505 } 506 507 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) { 508 if (phdr[i].p_type == PT_LOAD && phdr[i].p_memsz != 0) { 509 /* Loadable segment */ 510 prot = __elfN(trans_prot)(phdr[i].p_flags); 511 error = __elfN(load_section)( 512 p, vmspace, imgp->vp, 513 phdr[i].p_offset, 514 (caddr_t)phdr[i].p_vaddr + 515 rbase, 516 phdr[i].p_memsz, 517 phdr[i].p_filesz, prot); 518 if (error != 0) 519 goto fail; 520 /* 521 * Establish the base address if this is the 522 * first segment. 523 */ 524 if (numsegs == 0) 525 base_addr = trunc_page(phdr[i].p_vaddr + rbase); 526 numsegs++; 527 } 528 } 529 *addr = base_addr; 530 *entry = (unsigned long)hdr->e_entry + rbase; 531 532 fail: 533 if (imgp->firstpage) 534 exec_unmap_first_page(imgp); 535 if (imgp->vp) { 536 vrele(imgp->vp); 537 imgp->vp = NULL; 538 } 539 kfree(tempdata, M_TEMP); 540 541 return (error); 542 } 543 544 static Elf_Brandinfo * 545 __elfN(get_brandinfo)(struct image_params *imgp, const char *interp, 546 int32_t *osrel) 547 { 548 const Elf_Ehdr *hdr = (const Elf_Ehdr *)imgp->image_header; 549 Elf_Brandinfo *bi; 550 boolean_t ret; 551 int i; 552 553 /* We support four types of branding -- (1) the ELF EI_OSABI field 554 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string 555 * branding within the ELF header, (3) path of the `interp_path' field, 556 * and (4) the ".note.ABI-tag" ELF section. 557 */ 558 559 /* Look for an ".note.ABI-tag" ELF section */ 560 for (i = 0; i < MAX_BRANDS; i++) { 561 bi = elf_brand_list[i]; 562 563 if (bi == NULL) 564 continue; 565 if (hdr->e_machine == bi->machine && (bi->flags & 566 (BI_BRAND_NOTE|BI_BRAND_NOTE_MANDATORY)) != 0) { 567 ret = __elfN(check_note)(imgp, bi->brand_note, osrel); 568 if (ret) 569 return (bi); 570 } 571 } 572 573 /* If the executable has a brand, search for it in the brand list. */ 574 for (i = 0; i < MAX_BRANDS; i++) { 575 bi = elf_brand_list[i]; 576 577 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY) 578 continue; 579 if (hdr->e_machine == bi->machine && 580 (hdr->e_ident[EI_OSABI] == bi->brand || 581 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND], 582 bi->compat_3_brand, strlen(bi->compat_3_brand)) == 0)) 583 return (bi); 584 } 585 586 /* Lacking a known brand, search for a recognized interpreter. */ 587 if (interp != NULL) { 588 for (i = 0; i < MAX_BRANDS; i++) { 589 bi = elf_brand_list[i]; 590 591 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY) 592 continue; 593 if (hdr->e_machine == bi->machine && 594 strcmp(interp, bi->interp_path) == 0) 595 return (bi); 596 } 597 } 598 599 /* Lacking a recognized interpreter, try the default brand */ 600 for (i = 0; i < MAX_BRANDS; i++) { 601 bi = elf_brand_list[i]; 602 603 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY) 604 continue; 605 if (hdr->e_machine == bi->machine && 606 __elfN(fallback_brand) == bi->brand) 607 return (bi); 608 } 609 return (NULL); 610 } 611 612 static int 613 __CONCAT(exec_,__elfN(imgact))(struct image_params *imgp) 614 { 615 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header; 616 const Elf_Phdr *phdr; 617 Elf_Auxargs *elf_auxargs; 618 struct vmspace *vmspace; 619 vm_prot_t prot; 620 u_long text_size = 0, data_size = 0, total_size = 0; 621 u_long text_addr = 0, data_addr = 0; 622 u_long seg_size, seg_addr; 623 u_long addr, baddr, et_dyn_addr = 0, entry = 0, proghdr = 0; 624 int32_t osrel = 0; 625 int error = 0, i, n; 626 boolean_t failure; 627 char *interp = NULL; 628 const char *newinterp = NULL; 629 Elf_Brandinfo *brand_info; 630 char *path; 631 632 /* 633 * Do we have a valid ELF header ? 634 * 635 * Only allow ET_EXEC & ET_DYN here, reject ET_DYN later if a particular 636 * brand doesn't support it. Both DragonFly platforms do by default. 637 */ 638 if (__elfN(check_header)(hdr) != 0 || 639 (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN)) 640 return (-1); 641 642 /* 643 * From here on down, we return an errno, not -1, as we've 644 * detected an ELF file. 645 */ 646 647 if ((hdr->e_phoff > PAGE_SIZE) || 648 (hdr->e_phoff + hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE) { 649 /* Only support headers in first page for now */ 650 return (ENOEXEC); 651 } 652 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff); 653 if (!aligned(phdr, Elf_Addr)) 654 return (ENOEXEC); 655 n = 0; 656 baddr = 0; 657 for (i = 0; i < hdr->e_phnum; i++) { 658 if (phdr[i].p_type == PT_LOAD) { 659 if (n == 0) 660 baddr = phdr[i].p_vaddr; 661 n++; 662 continue; 663 } 664 if (phdr[i].p_type == PT_INTERP) { 665 /* 666 * If interp is already defined there are more than 667 * one PT_INTERP program headers present. Take only 668 * the first one and ignore the rest. 669 */ 670 if (interp != NULL) 671 continue; 672 673 if (phdr[i].p_filesz == 0 || 674 phdr[i].p_filesz > PAGE_SIZE || 675 phdr[i].p_filesz > MAXPATHLEN) 676 return (ENOEXEC); 677 678 interp = kmalloc(phdr[i].p_filesz, M_TEMP, M_WAITOK); 679 failure = extract_interpreter(imgp, &phdr[i], interp); 680 if (failure) { 681 kfree(interp, M_TEMP); 682 return (ENOEXEC); 683 } 684 continue; 685 } 686 } 687 688 brand_info = __elfN(get_brandinfo)(imgp, interp, &osrel); 689 if (brand_info == NULL) { 690 uprintf("ELF binary type \"%u\" not known.\n", 691 hdr->e_ident[EI_OSABI]); 692 if (interp != NULL) 693 kfree(interp, M_TEMP); 694 return (ENOEXEC); 695 } 696 if (hdr->e_type == ET_DYN) { 697 if ((brand_info->flags & BI_CAN_EXEC_DYN) == 0) { 698 if (interp != NULL) 699 kfree(interp, M_TEMP); 700 return (ENOEXEC); 701 } 702 /* 703 * If p_vaddr field of PT_LOAD program header is zero and type of an 704 * executale is ET_DYN, then it must be a position independent 705 * executable (PIE). In this case the system needs to pick a base 706 * address for us. Set et_dyn_addr to non-zero and choose the actual 707 * address when we are ready. 708 */ 709 if (baddr == 0) 710 et_dyn_addr = 1; 711 } 712 713 if (interp != NULL && brand_info->interp_newpath != NULL) 714 newinterp = brand_info->interp_newpath; 715 716 exec_new_vmspace(imgp, NULL); 717 718 /* 719 * Yeah, I'm paranoid. There is every reason in the world to get 720 * VTEXT now since from here on out, there are places we can have 721 * a context switch. Better safe than sorry; I really don't want 722 * the file to change while it's being loaded. 723 */ 724 vsetflags(imgp->vp, VTEXT); 725 726 vmspace = imgp->proc->p_vmspace; 727 /* Choose the base address for dynamic executables if we need to. */ 728 if (et_dyn_addr) 729 et_dyn_addr = pie_base_hint(imgp->proc); 730 731 for (i = 0; i < hdr->e_phnum; i++) { 732 switch (phdr[i].p_type) { 733 case PT_LOAD: /* Loadable segment */ 734 if (phdr[i].p_memsz == 0) 735 break; 736 prot = __elfN(trans_prot)(phdr[i].p_flags); 737 738 if ((error = __elfN(load_section)( 739 imgp->proc, 740 vmspace, 741 imgp->vp, 742 phdr[i].p_offset, 743 (caddr_t)phdr[i].p_vaddr + et_dyn_addr, 744 phdr[i].p_memsz, 745 phdr[i].p_filesz, 746 prot)) != 0) { 747 if (interp != NULL) 748 kfree (interp, M_TEMP); 749 return (error); 750 } 751 752 /* 753 * If this segment contains the program headers, 754 * remember their virtual address for the AT_PHDR 755 * aux entry. Static binaries don't usually include 756 * a PT_PHDR entry. 757 */ 758 if (phdr[i].p_offset == 0 && 759 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize 760 <= phdr[i].p_filesz) 761 proghdr = phdr[i].p_vaddr + hdr->e_phoff + 762 et_dyn_addr; 763 764 seg_addr = trunc_page(phdr[i].p_vaddr + et_dyn_addr); 765 seg_size = round_page(phdr[i].p_memsz + 766 phdr[i].p_vaddr + et_dyn_addr - seg_addr); 767 768 /* 769 * Is this .text or .data? We can't use 770 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the 771 * alpha terribly and possibly does other bad 772 * things so we stick to the old way of figuring 773 * it out: If the segment contains the program 774 * entry point, it's a text segment, otherwise it 775 * is a data segment. 776 * 777 * Note that obreak() assumes that data_addr + 778 * data_size == end of data load area, and the ELF 779 * file format expects segments to be sorted by 780 * address. If multiple data segments exist, the 781 * last one will be used. 782 */ 783 if (hdr->e_entry >= phdr[i].p_vaddr && 784 hdr->e_entry < (phdr[i].p_vaddr + 785 phdr[i].p_memsz)) { 786 text_size = seg_size; 787 text_addr = seg_addr; 788 entry = (u_long)hdr->e_entry + et_dyn_addr; 789 } else { 790 data_size = seg_size; 791 data_addr = seg_addr; 792 } 793 total_size += seg_size; 794 795 /* 796 * Check limits. It should be safe to check the 797 * limits after loading the segment since we do 798 * not actually fault in all the segment's pages. 799 */ 800 if (data_size > 801 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur || 802 text_size > maxtsiz || 803 total_size > 804 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 805 if (interp != NULL) 806 kfree(interp, M_TEMP); 807 error = ENOMEM; 808 return (error); 809 } 810 break; 811 case PT_PHDR: /* Program header table info */ 812 proghdr = phdr[i].p_vaddr + et_dyn_addr; 813 break; 814 default: 815 break; 816 } 817 } 818 819 vmspace->vm_tsize = text_size >> PAGE_SHIFT; 820 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr; 821 vmspace->vm_dsize = data_size >> PAGE_SHIFT; 822 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr; 823 824 addr = ELF_RTLD_ADDR(vmspace); 825 826 imgp->entry_addr = entry; 827 828 imgp->proc->p_sysent = brand_info->sysvec; 829 830 if (interp != NULL) { 831 int have_interp = FALSE; 832 if (brand_info->emul_path != NULL && 833 brand_info->emul_path[0] != '\0') { 834 path = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 835 ksnprintf(path, MAXPATHLEN, "%s%s", 836 brand_info->emul_path, interp); 837 error = __elfN(load_file)(imgp->proc, path, &addr, 838 &imgp->entry_addr); 839 kfree(path, M_TEMP); 840 if (error == 0) 841 have_interp = TRUE; 842 } 843 if (!have_interp && newinterp != NULL) { 844 error = __elfN(load_file)(imgp->proc, newinterp, 845 &addr, &imgp->entry_addr); 846 if (error == 0) 847 have_interp = TRUE; 848 } 849 if (!have_interp) { 850 error = __elfN(load_file)(imgp->proc, interp, &addr, 851 &imgp->entry_addr); 852 } 853 if (error != 0) { 854 uprintf("ELF interpreter %s not found\n", interp); 855 kfree(interp, M_TEMP); 856 return (error); 857 } 858 kfree(interp, M_TEMP); 859 } else 860 addr = et_dyn_addr; 861 862 /* 863 * Construct auxargs table (used by the fixup routine) 864 */ 865 elf_auxargs = kmalloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK); 866 elf_auxargs->execfd = -1; 867 elf_auxargs->phdr = proghdr; 868 elf_auxargs->phent = hdr->e_phentsize; 869 elf_auxargs->phnum = hdr->e_phnum; 870 elf_auxargs->pagesz = PAGE_SIZE; 871 elf_auxargs->base = addr; 872 elf_auxargs->flags = 0; 873 elf_auxargs->entry = entry; 874 875 imgp->auxargs = elf_auxargs; 876 imgp->interpreted = 0; 877 imgp->proc->p_osrel = osrel; 878 879 return (error); 880 } 881 882 int 883 __elfN(dragonfly_fixup)(register_t **stack_base, struct image_params *imgp) 884 { 885 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs; 886 Elf_Addr *base; 887 Elf_Addr *pos; 888 889 base = (Elf_Addr *)*stack_base; 890 pos = base + (imgp->args->argc + imgp->args->envc + 2); 891 892 if (args->execfd != -1) 893 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd); 894 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr); 895 AUXARGS_ENTRY(pos, AT_PHENT, args->phent); 896 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum); 897 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz); 898 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags); 899 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry); 900 AUXARGS_ENTRY(pos, AT_BASE, args->base); 901 if (imgp->execpathp != 0) 902 AUXARGS_ENTRY(pos, AT_EXECPATH, imgp->execpathp); 903 AUXARGS_ENTRY(pos, AT_OSRELDATE, osreldate); 904 AUXARGS_ENTRY(pos, AT_NULL, 0); 905 906 kfree(imgp->auxargs, M_TEMP); 907 imgp->auxargs = NULL; 908 909 base--; 910 suword64(base, (long)imgp->args->argc); 911 *stack_base = (register_t *)base; 912 return (0); 913 } 914 915 /* 916 * Code for generating ELF core dumps. 917 */ 918 919 typedef int (*segment_callback)(vm_map_entry_t, void *); 920 921 /* Closure for cb_put_phdr(). */ 922 struct phdr_closure { 923 Elf_Phdr *phdr; /* Program header to fill in (incremented) */ 924 Elf_Phdr *phdr_max; /* Pointer bound for error check */ 925 Elf_Off offset; /* Offset of segment in core file */ 926 }; 927 928 /* Closure for cb_size_segment(). */ 929 struct sseg_closure { 930 int count; /* Count of writable segments. */ 931 size_t vsize; /* Total size of all writable segments. */ 932 }; 933 934 /* Closure for cb_put_fp(). */ 935 struct fp_closure { 936 struct vn_hdr *vnh; 937 struct vn_hdr *vnh_max; 938 int count; 939 struct stat *sb; 940 }; 941 942 typedef struct elf_buf { 943 char *buf; 944 size_t off; 945 size_t off_max; 946 } *elf_buf_t; 947 948 static void *target_reserve(elf_buf_t target, size_t bytes, int *error); 949 950 static int cb_put_phdr (vm_map_entry_t, void *); 951 static int cb_size_segment (vm_map_entry_t, void *); 952 static int cb_fpcount_segment(vm_map_entry_t, void *); 953 static int cb_put_fp(vm_map_entry_t, void *); 954 955 956 static int each_segment (struct proc *, segment_callback, void *, int); 957 static int __elfN(corehdr)(struct lwp *, int, struct file *, struct ucred *, 958 int, elf_buf_t); 959 enum putmode { WRITE, DRYRUN }; 960 static int __elfN(puthdr)(struct lwp *, elf_buf_t, int sig, enum putmode, 961 int, struct file *); 962 static int elf_putallnotes(struct lwp *, elf_buf_t, int, enum putmode); 963 static int __elfN(putnote)(elf_buf_t, const char *, int, const void *, size_t); 964 965 static int elf_putsigs(struct lwp *, elf_buf_t); 966 static int elf_puttextvp(struct proc *, elf_buf_t); 967 static int elf_putfiles(struct proc *, elf_buf_t, struct file *); 968 969 int 970 __elfN(coredump)(struct lwp *lp, int sig, struct vnode *vp, off_t limit) 971 { 972 struct file *fp; 973 int error; 974 975 if ((error = falloc(NULL, &fp, NULL)) != 0) 976 return (error); 977 fsetcred(fp, lp->lwp_proc->p_ucred); 978 979 /* 980 * XXX fixme. 981 */ 982 fp->f_type = DTYPE_VNODE; 983 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW; 984 fp->f_ops = &vnode_fileops; 985 fp->f_data = vp; 986 987 error = generic_elf_coredump(lp, sig, fp, limit); 988 989 fp->f_type = 0; 990 fp->f_flag = 0; 991 fp->f_ops = &badfileops; 992 fp->f_data = NULL; 993 fdrop(fp); 994 return (error); 995 } 996 997 int 998 generic_elf_coredump(struct lwp *lp, int sig, struct file *fp, off_t limit) 999 { 1000 struct proc *p = lp->lwp_proc; 1001 struct ucred *cred = p->p_ucred; 1002 int error = 0; 1003 struct sseg_closure seginfo; 1004 struct elf_buf target; 1005 1006 if (!fp) 1007 kprintf("can't dump core - null fp\n"); 1008 1009 /* 1010 * Size the program segments 1011 */ 1012 seginfo.count = 0; 1013 seginfo.vsize = 0; 1014 each_segment(p, cb_size_segment, &seginfo, 1); 1015 1016 /* 1017 * Calculate the size of the core file header area by making 1018 * a dry run of generating it. Nothing is written, but the 1019 * size is calculated. 1020 */ 1021 bzero(&target, sizeof(target)); 1022 __elfN(puthdr)(lp, &target, sig, DRYRUN, seginfo.count, fp); 1023 1024 if (target.off + seginfo.vsize >= limit) 1025 return (EFAULT); 1026 1027 /* 1028 * Allocate memory for building the header, fill it up, 1029 * and write it out. 1030 */ 1031 target.off_max = target.off; 1032 target.off = 0; 1033 target.buf = kmalloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO); 1034 1035 error = __elfN(corehdr)(lp, sig, fp, cred, seginfo.count, &target); 1036 1037 /* Write the contents of all of the writable segments. */ 1038 if (error == 0) { 1039 Elf_Phdr *php; 1040 int i; 1041 ssize_t nbytes; 1042 1043 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1; 1044 for (i = 0; i < seginfo.count; i++) { 1045 error = fp_write(fp, (caddr_t)php->p_vaddr, 1046 php->p_filesz, &nbytes, UIO_USERSPACE); 1047 if (error != 0) 1048 break; 1049 php++; 1050 } 1051 } 1052 kfree(target.buf, M_TEMP); 1053 1054 return (error); 1055 } 1056 1057 /* 1058 * A callback for each_segment() to write out the segment's 1059 * program header entry. 1060 */ 1061 static int 1062 cb_put_phdr(vm_map_entry_t entry, void *closure) 1063 { 1064 struct phdr_closure *phc = closure; 1065 Elf_Phdr *phdr = phc->phdr; 1066 1067 if (phc->phdr == phc->phdr_max) 1068 return (EINVAL); 1069 1070 phc->offset = round_page(phc->offset); 1071 1072 phdr->p_type = PT_LOAD; 1073 phdr->p_offset = phc->offset; 1074 phdr->p_vaddr = entry->start; 1075 phdr->p_paddr = 0; 1076 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start; 1077 phdr->p_align = PAGE_SIZE; 1078 phdr->p_flags = __elfN(untrans_prot)(entry->protection); 1079 1080 phc->offset += phdr->p_filesz; 1081 ++phc->phdr; 1082 return (0); 1083 } 1084 1085 /* 1086 * A callback for each_writable_segment() to gather information about 1087 * the number of segments and their total size. 1088 */ 1089 static int 1090 cb_size_segment(vm_map_entry_t entry, void *closure) 1091 { 1092 struct sseg_closure *ssc = closure; 1093 1094 ++ssc->count; 1095 ssc->vsize += entry->end - entry->start; 1096 return (0); 1097 } 1098 1099 /* 1100 * A callback for each_segment() to gather information about 1101 * the number of text segments. 1102 */ 1103 static int 1104 cb_fpcount_segment(vm_map_entry_t entry, void *closure) 1105 { 1106 int *count = closure; 1107 struct vnode *vp; 1108 1109 if (entry->object.vm_object->type == OBJT_VNODE) { 1110 vp = (struct vnode *)entry->object.vm_object->handle; 1111 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp) 1112 return (0); 1113 ++*count; 1114 } 1115 return (0); 1116 } 1117 1118 static int 1119 cb_put_fp(vm_map_entry_t entry, void *closure) 1120 { 1121 struct fp_closure *fpc = closure; 1122 struct vn_hdr *vnh = fpc->vnh; 1123 Elf_Phdr *phdr = &vnh->vnh_phdr; 1124 struct vnode *vp; 1125 int error; 1126 1127 /* 1128 * If an entry represents a vnode then write out a file handle. 1129 * 1130 * If we are checkpointing a checkpoint-restored program we do 1131 * NOT record the filehandle for the old checkpoint vnode (which 1132 * is mapped all over the place). Instead we rely on the fact 1133 * that a checkpoint-restored program does not mmap() the checkpt 1134 * vnode NOCORE, so its contents will be written out to the 1135 * new checkpoint file. This is necessary because the 'old' 1136 * checkpoint file is typically destroyed when a new one is created 1137 * and thus cannot be used to restore the new checkpoint. 1138 * 1139 * Theoretically we could create a chain of checkpoint files and 1140 * operate the checkpointing operation kinda like an incremental 1141 * checkpoint, but a checkpoint restore would then likely wind up 1142 * referencing many prior checkpoint files and that is a bit over 1143 * the top for the purpose of the checkpoint API. 1144 */ 1145 if (entry->object.vm_object->type == OBJT_VNODE) { 1146 vp = (struct vnode *)entry->object.vm_object->handle; 1147 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp) 1148 return (0); 1149 if (vnh == fpc->vnh_max) 1150 return (EINVAL); 1151 1152 if (vp->v_mount) 1153 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid; 1154 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid); 1155 if (error) { 1156 char *freepath, *fullpath; 1157 1158 /* 1159 * This is actually a relatively common occurance, 1160 * so don't spew on the console by default. 1161 */ 1162 if (vn_fullpath(curproc, vp, &fullpath, &freepath, 0)) { 1163 if (bootverbose) 1164 kprintf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp); 1165 } else { 1166 if (bootverbose) 1167 kprintf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath); 1168 kfree(freepath, M_TEMP); 1169 } 1170 error = 0; 1171 } 1172 1173 phdr->p_type = PT_LOAD; 1174 phdr->p_offset = 0; /* not written to core */ 1175 phdr->p_vaddr = entry->start; 1176 phdr->p_paddr = 0; 1177 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start; 1178 phdr->p_align = PAGE_SIZE; 1179 phdr->p_flags = 0; 1180 if (entry->protection & VM_PROT_READ) 1181 phdr->p_flags |= PF_R; 1182 if (entry->protection & VM_PROT_WRITE) 1183 phdr->p_flags |= PF_W; 1184 if (entry->protection & VM_PROT_EXECUTE) 1185 phdr->p_flags |= PF_X; 1186 ++fpc->vnh; 1187 ++fpc->count; 1188 } 1189 return (0); 1190 } 1191 1192 /* 1193 * For each writable segment in the process's memory map, call the given 1194 * function with a pointer to the map entry and some arbitrary 1195 * caller-supplied data. 1196 */ 1197 static int 1198 each_segment(struct proc *p, segment_callback func, void *closure, int writable) 1199 { 1200 int error = 0; 1201 vm_map_t map = &p->p_vmspace->vm_map; 1202 vm_map_entry_t entry; 1203 1204 for (entry = map->header.next; error == 0 && entry != &map->header; 1205 entry = entry->next) { 1206 vm_object_t obj; 1207 vm_object_t lobj; 1208 vm_object_t tobj; 1209 1210 /* 1211 * Don't dump inaccessible mappings, deal with legacy 1212 * coredump mode. 1213 * 1214 * Note that read-only segments related to the elf binary 1215 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer 1216 * need to arbitrarily ignore such segments. 1217 */ 1218 if (elf_legacy_coredump) { 1219 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW) 1220 continue; 1221 } else { 1222 if (writable && (entry->protection & VM_PROT_ALL) == 0) 1223 continue; 1224 } 1225 1226 /* 1227 * Dont include memory segment in the coredump if 1228 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in 1229 * madvise(2). 1230 * 1231 * Currently we only dump normal VM object maps. We do 1232 * not dump submaps or virtual page tables. 1233 */ 1234 if (writable && (entry->eflags & MAP_ENTRY_NOCOREDUMP)) 1235 continue; 1236 if (entry->maptype != VM_MAPTYPE_NORMAL) 1237 continue; 1238 if ((obj = entry->object.vm_object) == NULL) 1239 continue; 1240 1241 /* 1242 * Find the bottom-most object, leaving the base object 1243 * and the bottom-most object held (but only one hold 1244 * if they happen to be the same). 1245 */ 1246 vm_object_hold_shared(obj); 1247 1248 lobj = obj; 1249 while (lobj && (tobj = lobj->backing_object) != NULL) { 1250 KKASSERT(tobj != obj); 1251 vm_object_hold_shared(tobj); 1252 if (tobj == lobj->backing_object) { 1253 if (lobj != obj) { 1254 vm_object_lock_swap(); 1255 vm_object_drop(lobj); 1256 } 1257 lobj = tobj; 1258 } else { 1259 vm_object_drop(tobj); 1260 } 1261 } 1262 1263 /* 1264 * The callback only applies to default, swap, or vnode 1265 * objects. Other types of objects such as memory-mapped 1266 * devices are ignored. 1267 */ 1268 if (lobj->type == OBJT_DEFAULT || lobj->type == OBJT_SWAP || 1269 lobj->type == OBJT_VNODE) { 1270 error = (*func)(entry, closure); 1271 } 1272 if (lobj != obj) 1273 vm_object_drop(lobj); 1274 vm_object_drop(obj); 1275 } 1276 return (error); 1277 } 1278 1279 static 1280 void * 1281 target_reserve(elf_buf_t target, size_t bytes, int *error) 1282 { 1283 void *res = NULL; 1284 1285 if (target->buf) { 1286 if (target->off + bytes > target->off_max) 1287 *error = EINVAL; 1288 else 1289 res = target->buf + target->off; 1290 } 1291 target->off += bytes; 1292 return (res); 1293 } 1294 1295 /* 1296 * Write the core file header to the file, including padding up to 1297 * the page boundary. 1298 */ 1299 static int 1300 __elfN(corehdr)(struct lwp *lp, int sig, struct file *fp, struct ucred *cred, 1301 int numsegs, elf_buf_t target) 1302 { 1303 int error; 1304 ssize_t nbytes; 1305 1306 /* 1307 * Fill in the header. The fp is passed so we can detect and flag 1308 * a checkpoint file pointer within the core file itself, because 1309 * it may not be restored from the same file handle. 1310 */ 1311 error = __elfN(puthdr)(lp, target, sig, WRITE, numsegs, fp); 1312 1313 /* Write it to the core file. */ 1314 if (error == 0) { 1315 error = fp_write(fp, target->buf, target->off, &nbytes, 1316 UIO_SYSSPACE); 1317 } 1318 return (error); 1319 } 1320 1321 static int 1322 __elfN(puthdr)(struct lwp *lp, elf_buf_t target, int sig, enum putmode mode, 1323 int numsegs, struct file *fp) 1324 { 1325 struct proc *p = lp->lwp_proc; 1326 int error = 0; 1327 size_t phoff; 1328 size_t noteoff; 1329 size_t notesz; 1330 Elf_Ehdr *ehdr; 1331 Elf_Phdr *phdr; 1332 1333 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error); 1334 1335 phoff = target->off; 1336 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error); 1337 1338 noteoff = target->off; 1339 if (error == 0) 1340 elf_putallnotes(lp, target, sig, mode); 1341 notesz = target->off - noteoff; 1342 1343 /* 1344 * put extra cruft for dumping process state here 1345 * - we really want it be before all the program 1346 * mappings 1347 * - we just need to update the offset accordingly 1348 * and GDB will be none the wiser. 1349 */ 1350 if (error == 0) 1351 error = elf_puttextvp(p, target); 1352 if (error == 0) 1353 error = elf_putsigs(lp, target); 1354 if (error == 0) 1355 error = elf_putfiles(p, target, fp); 1356 1357 /* 1358 * Align up to a page boundary for the program segments. The 1359 * actual data will be written to the outptu file, not to elf_buf_t, 1360 * so we do not have to do any further bounds checking. 1361 */ 1362 target->off = round_page(target->off); 1363 if (error == 0 && ehdr != NULL) { 1364 /* 1365 * Fill in the ELF header. 1366 */ 1367 ehdr->e_ident[EI_MAG0] = ELFMAG0; 1368 ehdr->e_ident[EI_MAG1] = ELFMAG1; 1369 ehdr->e_ident[EI_MAG2] = ELFMAG2; 1370 ehdr->e_ident[EI_MAG3] = ELFMAG3; 1371 ehdr->e_ident[EI_CLASS] = ELF_CLASS; 1372 ehdr->e_ident[EI_DATA] = ELF_DATA; 1373 ehdr->e_ident[EI_VERSION] = EV_CURRENT; 1374 ehdr->e_ident[EI_OSABI] = ELFOSABI_NONE; 1375 ehdr->e_ident[EI_ABIVERSION] = 0; 1376 ehdr->e_ident[EI_PAD] = 0; 1377 ehdr->e_type = ET_CORE; 1378 ehdr->e_machine = ELF_ARCH; 1379 ehdr->e_version = EV_CURRENT; 1380 ehdr->e_entry = 0; 1381 ehdr->e_phoff = phoff; 1382 ehdr->e_flags = 0; 1383 ehdr->e_ehsize = sizeof(Elf_Ehdr); 1384 ehdr->e_phentsize = sizeof(Elf_Phdr); 1385 ehdr->e_phnum = numsegs + 1; 1386 ehdr->e_shentsize = sizeof(Elf_Shdr); 1387 ehdr->e_shnum = 0; 1388 ehdr->e_shstrndx = SHN_UNDEF; 1389 } 1390 if (error == 0 && phdr != NULL) { 1391 /* 1392 * Fill in the program header entries. 1393 */ 1394 struct phdr_closure phc; 1395 1396 /* The note segement. */ 1397 phdr->p_type = PT_NOTE; 1398 phdr->p_offset = noteoff; 1399 phdr->p_vaddr = 0; 1400 phdr->p_paddr = 0; 1401 phdr->p_filesz = notesz; 1402 phdr->p_memsz = 0; 1403 phdr->p_flags = 0; 1404 phdr->p_align = 0; 1405 ++phdr; 1406 1407 /* All the writable segments from the program. */ 1408 phc.phdr = phdr; 1409 phc.phdr_max = phdr + numsegs; 1410 phc.offset = target->off; 1411 each_segment(p, cb_put_phdr, &phc, 1); 1412 } 1413 return (error); 1414 } 1415 1416 /* 1417 * Append core dump notes to target ELF buffer or simply update target size 1418 * if dryrun selected. 1419 */ 1420 static int 1421 elf_putallnotes(struct lwp *corelp, elf_buf_t target, int sig, 1422 enum putmode mode) 1423 { 1424 struct proc *p = corelp->lwp_proc; 1425 int error; 1426 struct { 1427 prstatus_t status; 1428 prfpregset_t fpregs; 1429 prpsinfo_t psinfo; 1430 } *tmpdata; 1431 prstatus_t *status; 1432 prfpregset_t *fpregs; 1433 prpsinfo_t *psinfo; 1434 struct lwp *lp; 1435 1436 /* 1437 * Allocate temporary storage for notes on heap to avoid stack overflow. 1438 */ 1439 if (mode != DRYRUN) { 1440 tmpdata = kmalloc(sizeof(*tmpdata), M_TEMP, M_ZERO | M_WAITOK); 1441 status = &tmpdata->status; 1442 fpregs = &tmpdata->fpregs; 1443 psinfo = &tmpdata->psinfo; 1444 } else { 1445 tmpdata = NULL; 1446 status = NULL; 1447 fpregs = NULL; 1448 psinfo = NULL; 1449 } 1450 1451 /* 1452 * Append LWP-agnostic note. 1453 */ 1454 if (mode != DRYRUN) { 1455 psinfo->pr_version = PRPSINFO_VERSION; 1456 psinfo->pr_psinfosz = sizeof(prpsinfo_t); 1457 strlcpy(psinfo->pr_fname, p->p_comm, 1458 sizeof(psinfo->pr_fname)); 1459 /* 1460 * XXX - We don't fill in the command line arguments 1461 * properly yet. 1462 */ 1463 strlcpy(psinfo->pr_psargs, p->p_comm, 1464 sizeof(psinfo->pr_psargs)); 1465 } 1466 error = 1467 __elfN(putnote)(target, "CORE", NT_PRPSINFO, psinfo, sizeof *psinfo); 1468 if (error) 1469 goto exit; 1470 1471 /* 1472 * Append first note for LWP that triggered core so that it is 1473 * the selected one when the debugger starts. 1474 */ 1475 if (mode != DRYRUN) { 1476 status->pr_version = PRSTATUS_VERSION; 1477 status->pr_statussz = sizeof(prstatus_t); 1478 status->pr_gregsetsz = sizeof(gregset_t); 1479 status->pr_fpregsetsz = sizeof(fpregset_t); 1480 status->pr_osreldate = osreldate; 1481 status->pr_cursig = sig; 1482 status->pr_pid = corelp->lwp_tid; 1483 fill_regs(corelp, &status->pr_reg); 1484 fill_fpregs(corelp, fpregs); 1485 } 1486 error = 1487 __elfN(putnote)(target, "CORE", NT_PRSTATUS, status, sizeof *status); 1488 if (error) 1489 goto exit; 1490 error = 1491 __elfN(putnote)(target, "CORE", NT_FPREGSET, fpregs, sizeof *fpregs); 1492 if (error) 1493 goto exit; 1494 1495 /* 1496 * Then append notes for other LWPs. 1497 */ 1498 FOREACH_LWP_IN_PROC(lp, p) { 1499 if (lp == corelp) 1500 continue; 1501 /* skip lwps being created */ 1502 if (lp->lwp_thread == NULL) 1503 continue; 1504 if (mode != DRYRUN) { 1505 status->pr_pid = lp->lwp_tid; 1506 fill_regs(lp, &status->pr_reg); 1507 fill_fpregs(lp, fpregs); 1508 } 1509 error = __elfN(putnote)(target, "CORE", NT_PRSTATUS, 1510 status, sizeof *status); 1511 if (error) 1512 goto exit; 1513 error = __elfN(putnote)(target, "CORE", NT_FPREGSET, 1514 fpregs, sizeof *fpregs); 1515 if (error) 1516 goto exit; 1517 } 1518 1519 exit: 1520 if (tmpdata != NULL) 1521 kfree(tmpdata, M_TEMP); 1522 return (error); 1523 } 1524 1525 /* 1526 * Generate a note sub-structure. 1527 * 1528 * NOTE: 4-byte alignment. 1529 */ 1530 static int 1531 __elfN(putnote)(elf_buf_t target, const char *name, int type, 1532 const void *desc, size_t descsz) 1533 { 1534 int error = 0; 1535 char *dst; 1536 Elf_Note note; 1537 1538 note.n_namesz = strlen(name) + 1; 1539 note.n_descsz = descsz; 1540 note.n_type = type; 1541 dst = target_reserve(target, sizeof(note), &error); 1542 if (dst != NULL) 1543 bcopy(¬e, dst, sizeof note); 1544 dst = target_reserve(target, note.n_namesz, &error); 1545 if (dst != NULL) 1546 bcopy(name, dst, note.n_namesz); 1547 target->off = roundup2(target->off, sizeof(Elf_Word)); 1548 dst = target_reserve(target, note.n_descsz, &error); 1549 if (dst != NULL) 1550 bcopy(desc, dst, note.n_descsz); 1551 target->off = roundup2(target->off, sizeof(Elf_Word)); 1552 return (error); 1553 } 1554 1555 1556 static int 1557 elf_putsigs(struct lwp *lp, elf_buf_t target) 1558 { 1559 /* XXX lwp handle more than one lwp */ 1560 struct proc *p = lp->lwp_proc; 1561 int error = 0; 1562 struct ckpt_siginfo *csi; 1563 1564 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error); 1565 if (csi) { 1566 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo); 1567 bcopy(p->p_sigacts, &csi->csi_sigacts, sizeof(*p->p_sigacts)); 1568 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval)); 1569 bcopy(&lp->lwp_sigmask, &csi->csi_sigmask, 1570 sizeof(sigset_t)); 1571 csi->csi_sigparent = p->p_sigparent; 1572 } 1573 return (error); 1574 } 1575 1576 static int 1577 elf_putfiles(struct proc *p, elf_buf_t target, struct file *ckfp) 1578 { 1579 thread_t td = curthread; 1580 int error = 0; 1581 int i; 1582 struct ckpt_filehdr *cfh = NULL; 1583 struct ckpt_fileinfo *cfi; 1584 struct file *fp; 1585 struct vnode *vp; 1586 1587 /* 1588 * the duplicated loop is gross, but it was the only way 1589 * to eliminate uninitialized variable warnings 1590 */ 1591 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error); 1592 if (cfh) { 1593 cfh->cfh_nfiles = 0; 1594 } 1595 1596 /* 1597 * ignore STDIN/STDERR/STDOUT. 1598 */ 1599 KKASSERT(td->td_proc == p); 1600 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) { 1601 fp = holdfp(td, i, -1); 1602 if (fp == NULL) 1603 continue; 1604 /* 1605 * XXX Only checkpoint vnodes for now. 1606 */ 1607 if (fp->f_type != DTYPE_VNODE) { 1608 fdrop(fp); 1609 continue; 1610 } 1611 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo), 1612 &error); 1613 if (cfi == NULL) { 1614 fdrop(fp); 1615 continue; 1616 } 1617 cfi->cfi_index = -1; 1618 cfi->cfi_type = fp->f_type; 1619 cfi->cfi_flags = fp->f_flag; 1620 cfi->cfi_offset = fp->f_offset; 1621 cfi->cfi_ckflags = 0; 1622 1623 if (fp == ckfp) 1624 cfi->cfi_ckflags |= CKFIF_ISCKPTFD; 1625 /* f_count and f_msgcount should not be saved/restored */ 1626 /* XXX save cred info */ 1627 1628 switch(fp->f_type) { 1629 case DTYPE_VNODE: 1630 vp = (struct vnode *)fp->f_data; 1631 /* 1632 * it looks like a bug in ptrace is marking 1633 * a non-vnode as a vnode - until we find the 1634 * root cause this will at least prevent 1635 * further panics from truss 1636 */ 1637 if (vp == NULL || vp->v_mount == NULL) 1638 break; 1639 cfh->cfh_nfiles++; 1640 cfi->cfi_index = i; 1641 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid; 1642 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid); 1643 break; 1644 default: 1645 break; 1646 } 1647 fdrop(fp); 1648 } 1649 return (error); 1650 } 1651 1652 static int 1653 elf_puttextvp(struct proc *p, elf_buf_t target) 1654 { 1655 int error = 0; 1656 int *vn_count; 1657 struct fp_closure fpc; 1658 struct ckpt_vminfo *vminfo; 1659 1660 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error); 1661 if (vminfo != NULL) { 1662 vminfo->cvm_dsize = p->p_vmspace->vm_dsize; 1663 vminfo->cvm_tsize = p->p_vmspace->vm_tsize; 1664 vminfo->cvm_daddr = p->p_vmspace->vm_daddr; 1665 vminfo->cvm_taddr = p->p_vmspace->vm_taddr; 1666 } 1667 1668 fpc.count = 0; 1669 vn_count = target_reserve(target, sizeof(int), &error); 1670 if (target->buf != NULL) { 1671 fpc.vnh = (struct vn_hdr *)(target->buf + target->off); 1672 fpc.vnh_max = fpc.vnh + 1673 (target->off_max - target->off) / sizeof(struct vn_hdr); 1674 error = each_segment(p, cb_put_fp, &fpc, 0); 1675 if (vn_count) 1676 *vn_count = fpc.count; 1677 } else { 1678 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0); 1679 } 1680 target->off += fpc.count * sizeof(struct vn_hdr); 1681 return (error); 1682 } 1683 1684 /* 1685 * Try to find the appropriate ABI-note section for checknote, 1686 * The entire image is searched if necessary, not only the first page. 1687 */ 1688 static boolean_t 1689 __elfN(check_note)(struct image_params *imgp, Elf_Brandnote *checknote, 1690 int32_t *osrel) 1691 { 1692 boolean_t valid_note_found; 1693 const Elf_Phdr *phdr, *pnote; 1694 const Elf_Ehdr *hdr; 1695 int i; 1696 1697 valid_note_found = FALSE; 1698 hdr = (const Elf_Ehdr *)imgp->image_header; 1699 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff); 1700 1701 for (i = 0; i < hdr->e_phnum; i++) { 1702 if (phdr[i].p_type == PT_NOTE) { 1703 pnote = &phdr[i]; 1704 valid_note_found = check_PT_NOTE (imgp, checknote, 1705 osrel, pnote); 1706 if (valid_note_found) 1707 break; 1708 } 1709 } 1710 return valid_note_found; 1711 } 1712 1713 /* 1714 * Be careful not to create new overflow conditions when checking 1715 * for overflow. 1716 */ 1717 static boolean_t 1718 note_overflow(const Elf_Note *note, size_t maxsize) 1719 { 1720 if (sizeof(*note) > maxsize) 1721 return TRUE; 1722 if (note->n_namesz > maxsize - sizeof(*note)) 1723 return TRUE; 1724 return FALSE; 1725 } 1726 1727 static boolean_t 1728 hdr_overflow(__ElfN(Off) off_beg, __ElfN(Size) size) 1729 { 1730 __ElfN(Off) off_end; 1731 1732 off_end = off_beg + size; 1733 if (off_end < off_beg) 1734 return TRUE; 1735 return FALSE; 1736 } 1737 1738 static boolean_t 1739 check_PT_NOTE(struct image_params *imgp, Elf_Brandnote *checknote, 1740 int32_t *osrel, const Elf_Phdr * pnote) 1741 { 1742 boolean_t limited_to_first_page; 1743 boolean_t found = FALSE; 1744 const Elf_Note *note, *note0, *note_end; 1745 const char *note_name; 1746 __ElfN(Off) noteloc, firstloc; 1747 __ElfN(Size) notesz, firstlen, endbyte; 1748 struct lwbuf *lwb; 1749 struct lwbuf lwb_cache; 1750 const char *page; 1751 char *data = NULL; 1752 int n; 1753 1754 if (hdr_overflow(pnote->p_offset, pnote->p_filesz)) 1755 return (FALSE); 1756 notesz = pnote->p_filesz; 1757 noteloc = pnote->p_offset; 1758 endbyte = noteloc + notesz; 1759 limited_to_first_page = noteloc < PAGE_SIZE && endbyte < PAGE_SIZE; 1760 1761 if (limited_to_first_page) { 1762 note = (const Elf_Note *)(imgp->image_header + noteloc); 1763 note_end = (const Elf_Note *)(imgp->image_header + endbyte); 1764 note0 = note; 1765 } else { 1766 firstloc = noteloc & PAGE_MASK; 1767 firstlen = PAGE_SIZE - firstloc; 1768 if (notesz < sizeof(Elf_Note) || notesz > PAGE_SIZE) 1769 return (FALSE); 1770 1771 lwb = &lwb_cache; 1772 if (exec_map_page(imgp, noteloc >> PAGE_SHIFT, &lwb, &page)) 1773 return (FALSE); 1774 if (firstlen < notesz) { /* crosses page boundary */ 1775 data = kmalloc(notesz, M_TEMP, M_WAITOK); 1776 bcopy(page + firstloc, data, firstlen); 1777 1778 exec_unmap_page(lwb); 1779 lwb = &lwb_cache; 1780 if (exec_map_page(imgp, (noteloc >> PAGE_SHIFT) + 1, 1781 &lwb, &page)) { 1782 kfree(data, M_TEMP); 1783 return (FALSE); 1784 } 1785 bcopy(page, data + firstlen, notesz - firstlen); 1786 note = note0 = (const Elf_Note *)(data); 1787 note_end = (const Elf_Note *)(data + notesz); 1788 } else { 1789 note = note0 = (const Elf_Note *)(page + firstloc); 1790 note_end = (const Elf_Note *)(page + firstloc + 1791 firstlen); 1792 } 1793 } 1794 1795 for (n = 0; n < 100 && note >= note0 && note < note_end; n++) { 1796 if (!aligned(note, Elf32_Addr)) 1797 break; 1798 if (note_overflow(note, (const char *)note_end - 1799 (const char *)note)) { 1800 break; 1801 } 1802 note_name = (const char *)(note + 1); 1803 1804 if (note->n_namesz == checknote->hdr.n_namesz 1805 && note->n_descsz == checknote->hdr.n_descsz 1806 && note->n_type == checknote->hdr.n_type 1807 && (strncmp(checknote->vendor, note_name, 1808 checknote->hdr.n_namesz) == 0)) { 1809 /* Fetch osreldata from ABI.note-tag */ 1810 if ((checknote->flags & BN_TRANSLATE_OSREL) != 0 && 1811 checknote->trans_osrel != NULL) 1812 checknote->trans_osrel(note, osrel); 1813 found = TRUE; 1814 break; 1815 } 1816 note = (const Elf_Note *)((const char *)(note + 1) + 1817 roundup2(note->n_namesz, sizeof(Elf32_Addr)) + 1818 roundup2(note->n_descsz, sizeof(Elf32_Addr))); 1819 } 1820 1821 if (!limited_to_first_page) { 1822 if (data != NULL) 1823 kfree(data, M_TEMP); 1824 exec_unmap_page(lwb); 1825 } 1826 return (found); 1827 } 1828 1829 /* 1830 * The interpreter program header may be located beyond the first page, so 1831 * regardless of its location, a copy of the interpreter path is created so 1832 * that it may be safely referenced by the calling function in all case. The 1833 * memory is allocated by calling function, and the copying is done here. 1834 */ 1835 static boolean_t 1836 extract_interpreter(struct image_params *imgp, const Elf_Phdr *pinterpreter, 1837 char *data) 1838 { 1839 boolean_t limited_to_first_page; 1840 const boolean_t result_success = FALSE; 1841 const boolean_t result_failure = TRUE; 1842 __ElfN(Off) pathloc, firstloc; 1843 __ElfN(Size) pathsz, firstlen, endbyte; 1844 struct lwbuf *lwb; 1845 struct lwbuf lwb_cache; 1846 const char *page; 1847 1848 if (hdr_overflow(pinterpreter->p_offset, pinterpreter->p_filesz)) 1849 return (result_failure); 1850 pathsz = pinterpreter->p_filesz; 1851 pathloc = pinterpreter->p_offset; 1852 endbyte = pathloc + pathsz; 1853 1854 limited_to_first_page = pathloc < PAGE_SIZE && endbyte < PAGE_SIZE; 1855 if (limited_to_first_page) { 1856 bcopy(imgp->image_header + pathloc, data, pathsz); 1857 return (result_success); 1858 } 1859 1860 firstloc = pathloc & PAGE_MASK; 1861 firstlen = PAGE_SIZE - firstloc; 1862 1863 lwb = &lwb_cache; 1864 if (exec_map_page(imgp, pathloc >> PAGE_SHIFT, &lwb, &page)) 1865 return (result_failure); 1866 1867 if (firstlen < pathsz) { /* crosses page boundary */ 1868 bcopy(page + firstloc, data, firstlen); 1869 1870 exec_unmap_page(lwb); 1871 lwb = &lwb_cache; 1872 if (exec_map_page(imgp, (pathloc >> PAGE_SHIFT) + 1, &lwb, 1873 &page)) 1874 return (result_failure); 1875 bcopy(page, data + firstlen, pathsz - firstlen); 1876 } else 1877 bcopy(page + firstloc, data, pathsz); 1878 1879 exec_unmap_page(lwb); 1880 return (result_success); 1881 } 1882 1883 static boolean_t 1884 __elfN(bsd_trans_osrel)(const Elf_Note *note, int32_t *osrel) 1885 { 1886 uintptr_t p; 1887 1888 p = (uintptr_t)(note + 1); 1889 p += roundup2(note->n_namesz, sizeof(Elf32_Addr)); 1890 *osrel = *(const int32_t *)(p); 1891 1892 return (TRUE); 1893 } 1894 1895 /* 1896 * Tell kern_execve.c about it, with a little help from the linker. 1897 */ 1898 #if defined(__x86_64__) 1899 static struct execsw elf_execsw = {exec_elf64_imgact, "ELF64"}; 1900 EXEC_SET_ORDERED(elf64, elf_execsw, SI_ORDER_FIRST); 1901 #else /* i386 assumed */ 1902 static struct execsw elf_execsw = {exec_elf32_imgact, "ELF32"}; 1903 EXEC_SET_ORDERED(elf32, elf_execsw, SI_ORDER_FIRST); 1904 #endif 1905 1906 static vm_prot_t 1907 __elfN(trans_prot)(Elf_Word flags) 1908 { 1909 vm_prot_t prot; 1910 1911 prot = 0; 1912 if (flags & PF_X) 1913 prot |= VM_PROT_EXECUTE; 1914 if (flags & PF_W) 1915 prot |= VM_PROT_WRITE; 1916 if (flags & PF_R) 1917 prot |= VM_PROT_READ; 1918 return (prot); 1919 } 1920 1921 static Elf_Word 1922 __elfN(untrans_prot)(vm_prot_t prot) 1923 { 1924 Elf_Word flags; 1925 1926 flags = 0; 1927 if (prot & VM_PROT_EXECUTE) 1928 flags |= PF_X; 1929 if (prot & VM_PROT_READ) 1930 flags |= PF_R; 1931 if (prot & VM_PROT_WRITE) 1932 flags |= PF_W; 1933 return (flags); 1934 } 1935 1936 static u_long 1937 pie_base_hint(struct proc *p) 1938 { 1939 u_long base; 1940 1941 if (elf_pie_base_mmap) 1942 base = vm_map_hint(p, 0, VM_PROT_READ | VM_PROT_EXECUTE); 1943 else 1944 base = ET_DYN_LOAD_ADDR; 1945 return base; 1946 } 1947