sys/kern/link_elf_obj.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 1998-2000 Doug Rabson
   5  * Copyright (c) 2004 Peter Wemm
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  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  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/cdefs.h>
  31 __FBSDID("$FreeBSD$");
  32
  33 #include "opt_ddb.h"
  34
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/fcntl.h>
  38 #include <sys/kernel.h>
  39 #include <sys/lock.h>
  40 #include <sys/malloc.h>
  41 #include <sys/linker.h>
  42 #include <sys/mutex.h>
  43 #include <sys/mount.h>
  44 #include <sys/namei.h>
  45 #include <sys/proc.h>
  46 #include <sys/rwlock.h>
  47 #include <sys/vnode.h>
  48
  49 #include <machine/elf.h>
  50
  51 #include <net/vnet.h>
  52
  53 #include <security/mac/mac_framework.h>
  54
  55 #include <vm/vm.h>
  56 #include <vm/vm_param.h>
  57 #include <vm/pmap.h>
  58 #include <vm/vm_extern.h>
  59 #include <vm/vm_kern.h>
  60 #include <vm/vm_map.h>
  61 #include <vm/vm_object.h>
  62 #include <vm/vm_page.h>
  63 #include <vm/vm_pager.h>
  64
  65 #include <sys/link_elf.h>
  66
  67 #ifdef DDB_CTF
  68 #include <contrib/zlib/zlib.h>
  69 #endif
  70
  71 #include "linker_if.h"
  72
  73 typedef struct {
  74         void            *addr;
  75         Elf_Off         size;
  76         int             flags;  /* Section flags. */
  77         int             sec;    /* Original section number. */
  78         char            *name;
  79 } Elf_progent;
  80
  81 typedef struct {
  82         Elf_Rel         *rel;
  83         int             nrel;
  84         int             sec;
  85 } Elf_relent;
  86
  87 typedef struct {
  88         Elf_Rela        *rela;
  89         int             nrela;
  90         int             sec;
  91 } Elf_relaent;
  92
  93 typedef struct elf_file {
  94         struct linker_file lf;          /* Common fields */
  95
  96         int             preloaded;
  97         caddr_t         address;        /* Relocation address */
  98         vm_object_t     object;         /* VM object to hold file pages */
  99         Elf_Shdr        *e_shdr;
 100
 101         Elf_progent     *progtab;
 102         u_int           nprogtab;
 103
 104         Elf_relaent     *relatab;
 105         u_int           nrelatab;
 106
 107         Elf_relent      *reltab;
 108         int             nreltab;
 109
 110         Elf_Sym         *ddbsymtab;     /* The symbol table we are using */
 111         long            ddbsymcnt;      /* Number of symbols */
 112         caddr_t         ddbstrtab;      /* String table */
 113         long            ddbstrcnt;      /* number of bytes in string table */
 114
 115         caddr_t         shstrtab;       /* Section name string table */
 116         long            shstrcnt;       /* number of bytes in string table */
 117
 118         caddr_t         ctftab;         /* CTF table */
 119         long            ctfcnt;         /* number of bytes in CTF table */
 120         caddr_t         ctfoff;         /* CTF offset table */
 121         caddr_t         typoff;         /* Type offset table */
 122         long            typlen;         /* Number of type entries. */
 123
 124 } *elf_file_t;
 125
 126 #include <kern/kern_ctf.c>
 127
 128 static int      link_elf_link_preload(linker_class_t cls,
 129                     const char *, linker_file_t *);
 130 static int      link_elf_link_preload_finish(linker_file_t);
 131 static int      link_elf_load_file(linker_class_t, const char *, linker_file_t *);
 132 static int      link_elf_lookup_symbol(linker_file_t, const char *,
 133                     c_linker_sym_t *);
 134 static int      link_elf_symbol_values(linker_file_t, c_linker_sym_t,
 135                     linker_symval_t *);
 136 static int      link_elf_search_symbol(linker_file_t, caddr_t value,
 137                     c_linker_sym_t *sym, long *diffp);
 138
 139 static void     link_elf_unload_file(linker_file_t);
 140 static int      link_elf_lookup_set(linker_file_t, const char *,
 141                     void ***, void ***, int *);
 142 static int      link_elf_each_function_name(linker_file_t,
 143                     int (*)(const char *, void *), void *);
 144 static int      link_elf_each_function_nameval(linker_file_t,
 145                                 linker_function_nameval_callback_t,
 146                                 void *);
 147 static int      link_elf_reloc_local(linker_file_t, bool);
 148 static long     link_elf_symtab_get(linker_file_t, const Elf_Sym **);
 149 static long     link_elf_strtab_get(linker_file_t, caddr_t *);
 150
 151 static int      elf_obj_lookup(linker_file_t lf, Elf_Size symidx, int deps,
 152                     Elf_Addr *);
 153
 154 static kobj_method_t link_elf_methods[] = {
 155         KOBJMETHOD(linker_lookup_symbol,        link_elf_lookup_symbol),
 156         KOBJMETHOD(linker_symbol_values,        link_elf_symbol_values),
 157         KOBJMETHOD(linker_search_symbol,        link_elf_search_symbol),
 158         KOBJMETHOD(linker_unload,               link_elf_unload_file),
 159         KOBJMETHOD(linker_load_file,            link_elf_load_file),
 160         KOBJMETHOD(linker_link_preload,         link_elf_link_preload),
 161         KOBJMETHOD(linker_link_preload_finish,  link_elf_link_preload_finish),
 162         KOBJMETHOD(linker_lookup_set,           link_elf_lookup_set),
 163         KOBJMETHOD(linker_each_function_name,   link_elf_each_function_name),
 164         KOBJMETHOD(linker_each_function_nameval, link_elf_each_function_nameval),
 165         KOBJMETHOD(linker_ctf_get,              link_elf_ctf_get),
 166         KOBJMETHOD(linker_symtab_get,           link_elf_symtab_get),
 167         KOBJMETHOD(linker_strtab_get,           link_elf_strtab_get),
 168         KOBJMETHOD_END
 169 };
 170
 171 static struct linker_class link_elf_class = {
 172 #if ELF_TARG_CLASS == ELFCLASS32
 173         "elf32_obj",
 174 #else
 175         "elf64_obj",
 176 #endif
 177         link_elf_methods, sizeof(struct elf_file)
 178 };
 179
 180 static int      relocate_file(elf_file_t ef);
 181 static void     elf_obj_cleanup_globals_cache(elf_file_t);
 182
 183 static void
 184 link_elf_error(const char *filename, const char *s)
 185 {
 186         if (filename == NULL)
 187                 printf("kldload: %s\n", s);
 188         else
 189                 printf("kldload: %s: %s\n", filename, s);
 190 }
 191
 192 static void
 193 link_elf_init(void *arg)
 194 {
 195
 196         linker_add_class(&link_elf_class);
 197 }
 198 SYSINIT(link_elf_obj, SI_SUB_KLD, SI_ORDER_SECOND, link_elf_init, NULL);
 199
 200 static void
 201 link_elf_protect_range(elf_file_t ef, vm_offset_t start, vm_offset_t end,
 202     vm_prot_t prot)
 203 {
 204         int error __unused;
 205
 206         KASSERT(start <= end && start >= (vm_offset_t)ef->address &&
 207             end <= round_page((vm_offset_t)ef->address + ef->lf.size),
 208             ("link_elf_protect_range: invalid range %#jx-%#jx",
 209             (uintmax_t)start, (uintmax_t)end));
 210
 211         if (start == end)
 212                 return;
 213         if (ef->preloaded) {
 214 #ifdef __amd64__
 215                 error = pmap_change_prot(start, end - start, prot);
 216                 KASSERT(error == 0,
 217                     ("link_elf_protect_range: pmap_change_prot() returned %d",
 218                     error));
 219 #endif
 220                 return;
 221         }
 222         error = vm_map_protect(kernel_map, start, end, prot, 0,
 223             VM_MAP_PROTECT_SET_PROT);
 224         KASSERT(error == KERN_SUCCESS,
 225             ("link_elf_protect_range: vm_map_protect() returned %d", error));
 226 }
 227
 228 /*
 229  * Restrict permissions on linker file memory based on section flags.
 230  * Sections need not be page-aligned, so overlap within a page is possible.
 231  */
 232 static void
 233 link_elf_protect(elf_file_t ef)
 234 {
 235         vm_offset_t end, segend, segstart, start;
 236         vm_prot_t gapprot, prot, segprot;
 237         int i;
 238
 239         /*
 240          * If the file was preloaded, the last page may contain other preloaded
 241          * data which may need to be writeable.  ELF files are always
 242          * page-aligned, but other preloaded data, such as entropy or CPU
 243          * microcode may be loaded with a smaller alignment.
 244          */
 245         gapprot = ef->preloaded ? VM_PROT_RW : VM_PROT_READ;
 246
 247         start = end = (vm_offset_t)ef->address;
 248         prot = VM_PROT_READ;
 249         for (i = 0; i < ef->nprogtab; i++) {
 250                 /*
 251                  * VNET and DPCPU sections have their memory allocated by their
 252                  * respective subsystems.
 253                  */
 254                 if (ef->progtab[i].name != NULL && (
 255 #ifdef VIMAGE
 256                     strcmp(ef->progtab[i].name, VNET_SETNAME) == 0 ||
 257 #endif
 258                     strcmp(ef->progtab[i].name, DPCPU_SETNAME) == 0))
 259                         continue;
 260
 261                 segstart = trunc_page((vm_offset_t)ef->progtab[i].addr);
 262                 segend = round_page((vm_offset_t)ef->progtab[i].addr +
 263                     ef->progtab[i].size);
 264                 segprot = VM_PROT_READ;
 265                 if ((ef->progtab[i].flags & SHF_WRITE) != 0)
 266                         segprot |= VM_PROT_WRITE;
 267                 if ((ef->progtab[i].flags & SHF_EXECINSTR) != 0)
 268                         segprot |= VM_PROT_EXECUTE;
 269
 270                 if (end <= segstart) {
 271                         /*
 272                          * Case 1: there is no overlap between the previous
 273                          * segment and this one.  Apply protections to the
 274                          * previous segment, and protect the gap between the
 275                          * previous and current segments, if any.
 276                          */
 277                         link_elf_protect_range(ef, start, end, prot);
 278                         link_elf_protect_range(ef, end, segstart, gapprot);
 279
 280                         start = segstart;
 281                         end = segend;
 282                         prot = segprot;
 283                 } else if (start < segstart && end == segend) {
 284                         /*
 285                          * Case 2: the current segment is a subrange of the
 286                          * previous segment.  Apply protections to the
 287                          * non-overlapping portion of the previous segment.
 288                          */
 289                         link_elf_protect_range(ef, start, segstart, prot);
 290
 291                         start = segstart;
 292                         prot |= segprot;
 293                 } else if (end < segend) {
 294                         /*
 295                          * Case 3: there is partial overlap between the previous
 296                          * and current segments.  Apply protections to the
 297                          * non-overlapping portion of the previous segment, and
 298                          * then the overlap, which must use the union of the two
 299                          * segments' protections.
 300                          */
 301                         link_elf_protect_range(ef, start, segstart, prot);
 302                         link_elf_protect_range(ef, segstart, end,
 303                             prot | segprot);
 304                         start = end;
 305                         end = segend;
 306                         prot = segprot;
 307                 } else {
 308                         /*
 309                          * Case 4: the two segments reside in the same page.
 310                          */
 311                         prot |= segprot;
 312                 }
 313         }
 314
 315         /*
 316          * Fix up the last unprotected segment and trailing data.
 317          */
 318         link_elf_protect_range(ef, start, end, prot);
 319         link_elf_protect_range(ef, end,
 320             round_page((vm_offset_t)ef->address + ef->lf.size), gapprot);
 321 }
 322
 323 static int
 324 link_elf_link_preload(linker_class_t cls, const char *filename,
 325     linker_file_t *result)
 326 {
 327         Elf_Ehdr *hdr;
 328         Elf_Shdr *shdr;
 329         Elf_Sym *es;
 330         void *modptr, *baseptr, *sizeptr;
 331         char *type;
 332         elf_file_t ef;
 333         linker_file_t lf;
 334         Elf_Addr off;
 335         int error, i, j, pb, ra, rl, shstrindex, symstrindex, symtabindex;
 336
 337         /* Look to see if we have the file preloaded */
 338         modptr = preload_search_by_name(filename);
 339         if (modptr == NULL)
 340                 return ENOENT;
 341
 342         type = (char *)preload_search_info(modptr, MODINFO_TYPE);
 343         baseptr = preload_search_info(modptr, MODINFO_ADDR);
 344         sizeptr = preload_search_info(modptr, MODINFO_SIZE);
 345         hdr = (Elf_Ehdr *)preload_search_info(modptr, MODINFO_METADATA |
 346             MODINFOMD_ELFHDR);
 347         shdr = (Elf_Shdr *)preload_search_info(modptr, MODINFO_METADATA |
 348             MODINFOMD_SHDR);
 349         if (type == NULL || (strcmp(type, "elf" __XSTRING(__ELF_WORD_SIZE)
 350             " obj module") != 0 &&
 351             strcmp(type, "elf obj module") != 0)) {
 352                 return (EFTYPE);
 353         }
 354         if (baseptr == NULL || sizeptr == NULL || hdr == NULL ||
 355             shdr == NULL)
 356                 return (EINVAL);
 357
 358         lf = linker_make_file(filename, &link_elf_class);
 359         if (lf == NULL)
 360                 return (ENOMEM);
 361
 362         ef = (elf_file_t)lf;
 363         ef->preloaded = 1;
 364         ef->address = *(caddr_t *)baseptr;
 365         lf->address = *(caddr_t *)baseptr;
 366         lf->size = *(size_t *)sizeptr;
 367
 368         if (hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
 369             hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
 370             hdr->e_ident[EI_VERSION] != EV_CURRENT ||
 371             hdr->e_version != EV_CURRENT ||
 372             hdr->e_type != ET_REL ||
 373             hdr->e_machine != ELF_TARG_MACH) {
 374                 error = EFTYPE;
 375                 goto out;
 376         }
 377         ef->e_shdr = shdr;
 378
 379         /* Scan the section header for information and table sizing. */
 380         symtabindex = -1;
 381         symstrindex = -1;
 382         for (i = 0; i < hdr->e_shnum; i++) {
 383                 switch (shdr[i].sh_type) {
 384                 case SHT_PROGBITS:
 385                 case SHT_NOBITS:
 386 #ifdef __amd64__
 387                 case SHT_X86_64_UNWIND:
 388 #endif
 389                         /* Ignore sections not loaded by the loader. */
 390                         if (shdr[i].sh_addr == 0)
 391                                 break;
 392                         ef->nprogtab++;
 393                         break;
 394                 case SHT_SYMTAB:
 395                         symtabindex = i;
 396                         symstrindex = shdr[i].sh_link;
 397                         break;
 398                 case SHT_REL:
 399                         /*
 400                          * Ignore relocation tables for sections not
 401                          * loaded by the loader.
 402                          */
 403                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 404                                 break;
 405                         ef->nreltab++;
 406                         break;
 407                 case SHT_RELA:
 408                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 409                                 break;
 410                         ef->nrelatab++;
 411                         break;
 412                 }
 413         }
 414
 415         shstrindex = hdr->e_shstrndx;
 416         if (ef->nprogtab == 0 || symstrindex < 0 ||
 417             symstrindex >= hdr->e_shnum ||
 418             shdr[symstrindex].sh_type != SHT_STRTAB || shstrindex == 0 ||
 419             shstrindex >= hdr->e_shnum ||
 420             shdr[shstrindex].sh_type != SHT_STRTAB) {
 421                 printf("%s: bad/missing section headers\n", filename);
 422                 error = ENOEXEC;
 423                 goto out;
 424         }
 425
 426         /* Allocate space for tracking the load chunks */
 427         if (ef->nprogtab != 0)
 428                 ef->progtab = malloc(ef->nprogtab * sizeof(*ef->progtab),
 429                     M_LINKER, M_WAITOK | M_ZERO);
 430         if (ef->nreltab != 0)
 431                 ef->reltab = malloc(ef->nreltab * sizeof(*ef->reltab),
 432                     M_LINKER, M_WAITOK | M_ZERO);
 433         if (ef->nrelatab != 0)
 434                 ef->relatab = malloc(ef->nrelatab * sizeof(*ef->relatab),
 435                     M_LINKER, M_WAITOK | M_ZERO);
 436         if ((ef->nprogtab != 0 && ef->progtab == NULL) ||
 437             (ef->nreltab != 0 && ef->reltab == NULL) ||
 438             (ef->nrelatab != 0 && ef->relatab == NULL)) {
 439                 error = ENOMEM;
 440                 goto out;
 441         }
 442
 443         /* XXX, relocate the sh_addr fields saved by the loader. */
 444         off = 0;
 445         for (i = 0; i < hdr->e_shnum; i++) {
 446                 if (shdr[i].sh_addr != 0 && (off == 0 || shdr[i].sh_addr < off))
 447                         off = shdr[i].sh_addr;
 448         }
 449         for (i = 0; i < hdr->e_shnum; i++) {
 450                 if (shdr[i].sh_addr != 0)
 451                         shdr[i].sh_addr = shdr[i].sh_addr - off +
 452                             (Elf_Addr)ef->address;
 453         }
 454
 455         ef->ddbsymcnt = shdr[symtabindex].sh_size / sizeof(Elf_Sym);
 456         ef->ddbsymtab = (Elf_Sym *)shdr[symtabindex].sh_addr;
 457         ef->ddbstrcnt = shdr[symstrindex].sh_size;
 458         ef->ddbstrtab = (char *)shdr[symstrindex].sh_addr;
 459         ef->shstrcnt = shdr[shstrindex].sh_size;
 460         ef->shstrtab = (char *)shdr[shstrindex].sh_addr;
 461
 462         /* Now fill out progtab and the relocation tables. */
 463         pb = 0;
 464         rl = 0;
 465         ra = 0;
 466         for (i = 0; i < hdr->e_shnum; i++) {
 467                 switch (shdr[i].sh_type) {
 468                 case SHT_PROGBITS:
 469                 case SHT_NOBITS:
 470 #ifdef __amd64__
 471                 case SHT_X86_64_UNWIND:
 472 #endif
 473                         if (shdr[i].sh_addr == 0)
 474                                 break;
 475                         ef->progtab[pb].addr = (void *)shdr[i].sh_addr;
 476                         if (shdr[i].sh_type == SHT_PROGBITS)
 477                                 ef->progtab[pb].name = "<<PROGBITS>>";
 478 #ifdef __amd64__
 479                         else if (shdr[i].sh_type == SHT_X86_64_UNWIND)
 480                                 ef->progtab[pb].name = "<<UNWIND>>";
 481 #endif
 482                         else
 483                                 ef->progtab[pb].name = "<<NOBITS>>";
 484                         ef->progtab[pb].size = shdr[i].sh_size;
 485                         ef->progtab[pb].flags = shdr[i].sh_flags;
 486                         ef->progtab[pb].sec = i;
 487                         if (ef->shstrtab && shdr[i].sh_name != 0)
 488                                 ef->progtab[pb].name =
 489                                     ef->shstrtab + shdr[i].sh_name;
 490                         if (ef->progtab[pb].name != NULL &&
 491                             !strcmp(ef->progtab[pb].name, DPCPU_SETNAME)) {
 492                                 void *dpcpu;
 493
 494                                 dpcpu = dpcpu_alloc(shdr[i].sh_size);
 495                                 if (dpcpu == NULL) {
 496                                         printf("%s: pcpu module space is out "
 497                                             "of space; cannot allocate %#jx "
 498                                             "for %s\n", __func__,
 499                                             (uintmax_t)shdr[i].sh_size,
 500                                             filename);
 501                                         error = ENOSPC;
 502                                         goto out;
 503                                 }
 504                                 memcpy(dpcpu, ef->progtab[pb].addr,
 505                                     ef->progtab[pb].size);
 506                                 dpcpu_copy(dpcpu, shdr[i].sh_size);
 507                                 ef->progtab[pb].addr = dpcpu;
 508 #ifdef VIMAGE
 509                         } else if (ef->progtab[pb].name != NULL &&
 510                             !strcmp(ef->progtab[pb].name, VNET_SETNAME)) {
 511                                 void *vnet_data;
 512
 513                                 vnet_data = vnet_data_alloc(shdr[i].sh_size);
 514                                 if (vnet_data == NULL) {
 515                                         printf("%s: vnet module space is out "
 516                                             "of space; cannot allocate %#jx "
 517                                             "for %s\n", __func__,
 518                                             (uintmax_t)shdr[i].sh_size,
 519                                             filename);
 520                                         error = ENOSPC;
 521                                         goto out;
 522                                 }
 523                                 memcpy(vnet_data, ef->progtab[pb].addr,
 524                                     ef->progtab[pb].size);
 525                                 vnet_data_copy(vnet_data, shdr[i].sh_size);
 526                                 ef->progtab[pb].addr = vnet_data;
 527 #endif
 528                         } else if (ef->progtab[pb].name != NULL &&
 529                             !strcmp(ef->progtab[pb].name, ".ctors")) {
 530                                 lf->ctors_addr = ef->progtab[pb].addr;
 531                                 lf->ctors_size = shdr[i].sh_size;
 532                         }
 533
 534                         /* Update all symbol values with the offset. */
 535                         for (j = 0; j < ef->ddbsymcnt; j++) {
 536                                 es = &ef->ddbsymtab[j];
 537                                 if (es->st_shndx != i)
 538                                         continue;
 539                                 es->st_value += (Elf_Addr)ef->progtab[pb].addr;
 540                         }
 541                         pb++;
 542                         break;
 543                 case SHT_REL:
 544                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 545                                 break;
 546                         ef->reltab[rl].rel = (Elf_Rel *)shdr[i].sh_addr;
 547                         ef->reltab[rl].nrel = shdr[i].sh_size / sizeof(Elf_Rel);
 548                         ef->reltab[rl].sec = shdr[i].sh_info;
 549                         rl++;
 550                         break;
 551                 case SHT_RELA:
 552                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 553                                 break;
 554                         ef->relatab[ra].rela = (Elf_Rela *)shdr[i].sh_addr;
 555                         ef->relatab[ra].nrela =
 556                             shdr[i].sh_size / sizeof(Elf_Rela);
 557                         ef->relatab[ra].sec = shdr[i].sh_info;
 558                         ra++;
 559                         break;
 560                 }
 561         }
 562         if (pb != ef->nprogtab) {
 563                 printf("%s: lost progbits\n", filename);
 564                 error = ENOEXEC;
 565                 goto out;
 566         }
 567         if (rl != ef->nreltab) {
 568                 printf("%s: lost reltab\n", filename);
 569                 error = ENOEXEC;
 570                 goto out;
 571         }
 572         if (ra != ef->nrelatab) {
 573                 printf("%s: lost relatab\n", filename);
 574                 error = ENOEXEC;
 575                 goto out;
 576         }
 577
 578         /*
 579          * The file needs to be writeable and executable while applying
 580          * relocations.  Mapping protections are applied once relocation
 581          * processing is complete.
 582          */
 583         link_elf_protect_range(ef, (vm_offset_t)ef->address,
 584             round_page((vm_offset_t)ef->address + ef->lf.size), VM_PROT_ALL);
 585
 586         /* Local intra-module relocations */
 587         error = link_elf_reloc_local(lf, false);
 588         if (error != 0)
 589                 goto out;
 590         *result = lf;
 591         return (0);
 592
 593 out:
 594         /* preload not done this way */
 595         linker_file_unload(lf, LINKER_UNLOAD_FORCE);
 596         return (error);
 597 }
 598
 599 static void
 600 link_elf_invoke_ctors(caddr_t addr, size_t size)
 601 {
 602         void (**ctor)(void);
 603         size_t i, cnt;
 604
 605         if (addr == NULL || size == 0)
 606                 return;
 607         cnt = size / sizeof(*ctor);
 608         ctor = (void *)addr;
 609         for (i = 0; i < cnt; i++) {
 610                 if (ctor[i] != NULL)
 611                         (*ctor[i])();
 612         }
 613 }
 614
 615 static int
 616 link_elf_link_preload_finish(linker_file_t lf)
 617 {
 618         elf_file_t ef;
 619         int error;
 620
 621         ef = (elf_file_t)lf;
 622         error = relocate_file(ef);
 623         if (error)
 624                 return (error);
 625
 626         /* Notify MD code that a module is being loaded. */
 627         error = elf_cpu_load_file(lf);
 628         if (error)
 629                 return (error);
 630
 631 #if defined(__i386__) || defined(__amd64__)
 632         /* Now ifuncs. */
 633         error = link_elf_reloc_local(lf, true);
 634         if (error != 0)
 635                 return (error);
 636 #endif
 637
 638         /* Apply protections now that relocation processing is complete. */
 639         link_elf_protect(ef);
 640
 641         link_elf_invoke_ctors(lf->ctors_addr, lf->ctors_size);
 642         return (0);
 643 }
 644
 645 static int
 646 link_elf_load_file(linker_class_t cls, const char *filename,
 647     linker_file_t *result)
 648 {
 649         struct nameidata *nd;
 650         struct thread *td = curthread;  /* XXX */
 651         Elf_Ehdr *hdr;
 652         Elf_Shdr *shdr;
 653         Elf_Sym *es;
 654         int nbytes, i, j;
 655         vm_offset_t mapbase;
 656         size_t mapsize;
 657         int error = 0;
 658         ssize_t resid;
 659         int flags;
 660         elf_file_t ef;
 661         linker_file_t lf;
 662         int symtabindex;
 663         int symstrindex;
 664         int shstrindex;
 665         int nsym;
 666         int pb, rl, ra;
 667         int alignmask;
 668
 669         shdr = NULL;
 670         lf = NULL;
 671         mapsize = 0;
 672         hdr = NULL;
 673
 674         nd = malloc(sizeof(struct nameidata), M_TEMP, M_WAITOK);
 675         NDINIT(nd, LOOKUP, FOLLOW, UIO_SYSSPACE, filename, td);
 676         flags = FREAD;
 677         error = vn_open(nd, &flags, 0, NULL);
 678         if (error) {
 679                 free(nd, M_TEMP);
 680                 return error;
 681         }
 682         NDFREE(nd, NDF_ONLY_PNBUF);
 683         if (nd->ni_vp->v_type != VREG) {
 684                 error = ENOEXEC;
 685                 goto out;
 686         }
 687 #ifdef MAC
 688         error = mac_kld_check_load(td->td_ucred, nd->ni_vp);
 689         if (error) {
 690                 goto out;
 691         }
 692 #endif
 693
 694         /* Read the elf header from the file. */
 695         hdr = malloc(sizeof(*hdr), M_LINKER, M_WAITOK);
 696         error = vn_rdwr(UIO_READ, nd->ni_vp, (void *)hdr, sizeof(*hdr), 0,
 697             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 698             &resid, td);
 699         if (error)
 700                 goto out;
 701         if (resid != 0){
 702                 error = ENOEXEC;
 703                 goto out;
 704         }
 705
 706         if (!IS_ELF(*hdr)) {
 707                 error = ENOEXEC;
 708                 goto out;
 709         }
 710
 711         if (hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS
 712             || hdr->e_ident[EI_DATA] != ELF_TARG_DATA) {
 713                 link_elf_error(filename, "Unsupported file layout");
 714                 error = ENOEXEC;
 715                 goto out;
 716         }
 717         if (hdr->e_ident[EI_VERSION] != EV_CURRENT
 718             || hdr->e_version != EV_CURRENT) {
 719                 link_elf_error(filename, "Unsupported file version");
 720                 error = ENOEXEC;
 721                 goto out;
 722         }
 723         if (hdr->e_type != ET_REL) {
 724                 error = ENOSYS;
 725                 goto out;
 726         }
 727         if (hdr->e_machine != ELF_TARG_MACH) {
 728                 link_elf_error(filename, "Unsupported machine");
 729                 error = ENOEXEC;
 730                 goto out;
 731         }
 732
 733         lf = linker_make_file(filename, &link_elf_class);
 734         if (!lf) {
 735                 error = ENOMEM;
 736                 goto out;
 737         }
 738         ef = (elf_file_t) lf;
 739         ef->nprogtab = 0;
 740         ef->e_shdr = 0;
 741         ef->nreltab = 0;
 742         ef->nrelatab = 0;
 743
 744         /* Allocate and read in the section header */
 745         nbytes = hdr->e_shnum * hdr->e_shentsize;
 746         if (nbytes == 0 || hdr->e_shoff == 0 ||
 747             hdr->e_shentsize != sizeof(Elf_Shdr)) {
 748                 error = ENOEXEC;
 749                 goto out;
 750         }
 751         shdr = malloc(nbytes, M_LINKER, M_WAITOK);
 752         ef->e_shdr = shdr;
 753         error = vn_rdwr(UIO_READ, nd->ni_vp, (caddr_t)shdr, nbytes,
 754             hdr->e_shoff, UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred,
 755             NOCRED, &resid, td);
 756         if (error)
 757                 goto out;
 758         if (resid) {
 759                 error = ENOEXEC;
 760                 goto out;
 761         }
 762
 763         /* Scan the section header for information and table sizing. */
 764         nsym = 0;
 765         symtabindex = -1;
 766         symstrindex = -1;
 767         for (i = 0; i < hdr->e_shnum; i++) {
 768                 if (shdr[i].sh_size == 0)
 769                         continue;
 770                 switch (shdr[i].sh_type) {
 771                 case SHT_PROGBITS:
 772                 case SHT_NOBITS:
 773 #ifdef __amd64__
 774                 case SHT_X86_64_UNWIND:
 775 #endif
 776                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 777                                 break;
 778                         ef->nprogtab++;
 779                         break;
 780                 case SHT_SYMTAB:
 781                         nsym++;
 782                         symtabindex = i;
 783                         symstrindex = shdr[i].sh_link;
 784                         break;
 785                 case SHT_REL:
 786                         /*
 787                          * Ignore relocation tables for unallocated
 788                          * sections.
 789                          */
 790                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
 791                                 break;
 792                         ef->nreltab++;
 793                         break;
 794                 case SHT_RELA:
 795                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
 796                                 break;
 797                         ef->nrelatab++;
 798                         break;
 799                 case SHT_STRTAB:
 800                         break;
 801                 }
 802         }
 803         if (ef->nprogtab == 0) {
 804                 link_elf_error(filename, "file has no contents");
 805                 error = ENOEXEC;
 806                 goto out;
 807         }
 808         if (nsym != 1) {
 809                 /* Only allow one symbol table for now */
 810                 link_elf_error(filename,
 811                     "file must have exactly one symbol table");
 812                 error = ENOEXEC;
 813                 goto out;
 814         }
 815         if (symstrindex < 0 || symstrindex > hdr->e_shnum ||
 816             shdr[symstrindex].sh_type != SHT_STRTAB) {
 817                 link_elf_error(filename, "file has invalid symbol strings");
 818                 error = ENOEXEC;
 819                 goto out;
 820         }
 821
 822         /* Allocate space for tracking the load chunks */
 823         if (ef->nprogtab != 0)
 824                 ef->progtab = malloc(ef->nprogtab * sizeof(*ef->progtab),
 825                     M_LINKER, M_WAITOK | M_ZERO);
 826         if (ef->nreltab != 0)
 827                 ef->reltab = malloc(ef->nreltab * sizeof(*ef->reltab),
 828                     M_LINKER, M_WAITOK | M_ZERO);
 829         if (ef->nrelatab != 0)
 830                 ef->relatab = malloc(ef->nrelatab * sizeof(*ef->relatab),
 831                     M_LINKER, M_WAITOK | M_ZERO);
 832
 833         if (symtabindex == -1) {
 834                 link_elf_error(filename, "lost symbol table index");
 835                 error = ENOEXEC;
 836                 goto out;
 837         }
 838         /* Allocate space for and load the symbol table */
 839         ef->ddbsymcnt = shdr[symtabindex].sh_size / sizeof(Elf_Sym);
 840         ef->ddbsymtab = malloc(shdr[symtabindex].sh_size, M_LINKER, M_WAITOK);
 841         error = vn_rdwr(UIO_READ, nd->ni_vp, (void *)ef->ddbsymtab,
 842             shdr[symtabindex].sh_size, shdr[symtabindex].sh_offset,
 843             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 844             &resid, td);
 845         if (error)
 846                 goto out;
 847         if (resid != 0){
 848                 error = EINVAL;
 849                 goto out;
 850         }
 851
 852         /* Allocate space for and load the symbol strings */
 853         ef->ddbstrcnt = shdr[symstrindex].sh_size;
 854         ef->ddbstrtab = malloc(shdr[symstrindex].sh_size, M_LINKER, M_WAITOK);
 855         error = vn_rdwr(UIO_READ, nd->ni_vp, ef->ddbstrtab,
 856             shdr[symstrindex].sh_size, shdr[symstrindex].sh_offset,
 857             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 858             &resid, td);
 859         if (error)
 860                 goto out;
 861         if (resid != 0){
 862                 error = EINVAL;
 863                 goto out;
 864         }
 865
 866         /* Do we have a string table for the section names?  */
 867         shstrindex = -1;
 868         if (hdr->e_shstrndx != 0 &&
 869             shdr[hdr->e_shstrndx].sh_type == SHT_STRTAB) {
 870                 shstrindex = hdr->e_shstrndx;
 871                 ef->shstrcnt = shdr[shstrindex].sh_size;
 872                 ef->shstrtab = malloc(shdr[shstrindex].sh_size, M_LINKER,
 873                     M_WAITOK);
 874                 error = vn_rdwr(UIO_READ, nd->ni_vp, ef->shstrtab,
 875                     shdr[shstrindex].sh_size, shdr[shstrindex].sh_offset,
 876                     UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 877                     &resid, td);
 878                 if (error)
 879                         goto out;
 880                 if (resid != 0){
 881                         error = EINVAL;
 882                         goto out;
 883                 }
 884         }
 885
 886         /* Size up code/data(progbits) and bss(nobits). */
 887         alignmask = 0;
 888         for (i = 0; i < hdr->e_shnum; i++) {
 889                 if (shdr[i].sh_size == 0)
 890                         continue;
 891                 switch (shdr[i].sh_type) {
 892                 case SHT_PROGBITS:
 893                 case SHT_NOBITS:
 894 #ifdef __amd64__
 895                 case SHT_X86_64_UNWIND:
 896 #endif
 897                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 898                                 break;
 899                         alignmask = shdr[i].sh_addralign - 1;
 900                         mapsize += alignmask;
 901                         mapsize &= ~alignmask;
 902                         mapsize += shdr[i].sh_size;
 903                         break;
 904                 }
 905         }
 906
 907         /*
 908          * We know how much space we need for the text/data/bss/etc.
 909          * This stuff needs to be in a single chunk so that profiling etc
 910          * can get the bounds and gdb can associate offsets with modules
 911          */
 912         ef->object = vm_pager_allocate(OBJT_PHYS, NULL, round_page(mapsize),
 913             VM_PROT_ALL, 0, thread0.td_ucred);
 914         if (ef->object == NULL) {
 915                 error = ENOMEM;
 916                 goto out;
 917         }
 918 #if VM_NRESERVLEVEL > 0
 919         vm_object_color(ef->object, 0);
 920 #endif
 921
 922         /*
 923          * In order to satisfy amd64's architectural requirements on the
 924          * location of code and data in the kernel's address space, request a
 925          * mapping that is above the kernel.
 926          *
 927          * Protections will be restricted once relocations are applied.
 928          */
 929 #ifdef __amd64__
 930         mapbase = KERNBASE;
 931 #else
 932         mapbase = VM_MIN_KERNEL_ADDRESS;
 933 #endif
 934         error = vm_map_find(kernel_map, ef->object, 0, &mapbase,
 935             round_page(mapsize), 0, VMFS_OPTIMAL_SPACE, VM_PROT_ALL,
 936             VM_PROT_ALL, 0);
 937         if (error != KERN_SUCCESS) {
 938                 vm_object_deallocate(ef->object);
 939                 ef->object = NULL;
 940                 error = ENOMEM;
 941                 goto out;
 942         }
 943
 944         /* Wire the pages */
 945         error = vm_map_wire(kernel_map, mapbase,
 946             mapbase + round_page(mapsize),
 947             VM_MAP_WIRE_SYSTEM|VM_MAP_WIRE_NOHOLES);
 948         if (error != KERN_SUCCESS) {
 949                 error = ENOMEM;
 950                 goto out;
 951         }
 952
 953         /* Inform the kld system about the situation */
 954         lf->address = ef->address = (caddr_t)mapbase;
 955         lf->size = mapsize;
 956
 957         /*
 958          * Now load code/data(progbits), zero bss(nobits), allocate space for
 959          * and load relocs
 960          */
 961         pb = 0;
 962         rl = 0;
 963         ra = 0;
 964         alignmask = 0;
 965         for (i = 0; i < hdr->e_shnum; i++) {
 966                 if (shdr[i].sh_size == 0)
 967                         continue;
 968                 switch (shdr[i].sh_type) {
 969                 case SHT_PROGBITS:
 970                 case SHT_NOBITS:
 971 #ifdef __amd64__
 972                 case SHT_X86_64_UNWIND:
 973 #endif
 974                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 975                                 break;
 976                         alignmask = shdr[i].sh_addralign - 1;
 977                         mapbase += alignmask;
 978                         mapbase &= ~alignmask;
 979                         if (ef->shstrtab != NULL && shdr[i].sh_name != 0) {
 980                                 ef->progtab[pb].name =
 981                                     ef->shstrtab + shdr[i].sh_name;
 982                                 if (!strcmp(ef->progtab[pb].name, ".ctors")) {
 983                                         lf->ctors_addr = (caddr_t)mapbase;
 984                                         lf->ctors_size = shdr[i].sh_size;
 985                                 }
 986                         } else if (shdr[i].sh_type == SHT_PROGBITS)
 987                                 ef->progtab[pb].name = "<<PROGBITS>>";
 988 #ifdef __amd64__
 989                         else if (shdr[i].sh_type == SHT_X86_64_UNWIND)
 990                                 ef->progtab[pb].name = "<<UNWIND>>";
 991 #endif
 992                         else
 993                                 ef->progtab[pb].name = "<<NOBITS>>";
 994                         if (ef->progtab[pb].name != NULL &&
 995                             !strcmp(ef->progtab[pb].name, DPCPU_SETNAME)) {
 996                                 ef->progtab[pb].addr =
 997                                     dpcpu_alloc(shdr[i].sh_size);
 998                                 if (ef->progtab[pb].addr == NULL) {
 999                                         printf("%s: pcpu module space is out "
1000                                             "of space; cannot allocate %#jx "
1001                                             "for %s\n", __func__,
1002                                             (uintmax_t)shdr[i].sh_size,
1003                                             filename);
1004                                 }
1005                         }
1006 #ifdef VIMAGE
1007                         else if (ef->progtab[pb].name != NULL &&
1008                             !strcmp(ef->progtab[pb].name, VNET_SETNAME)) {
1009                                 ef->progtab[pb].addr =
1010                                     vnet_data_alloc(shdr[i].sh_size);
1011                                 if (ef->progtab[pb].addr == NULL) {
1012                                         printf("%s: vnet module space is out "
1013                                             "of space; cannot allocate %#jx "
1014                                             "for %s\n", __func__,
1015                                             (uintmax_t)shdr[i].sh_size,
1016                                             filename);
1017                                 }
1018                         }
1019 #endif
1020                         else
1021                                 ef->progtab[pb].addr =
1022                                     (void *)(uintptr_t)mapbase;
1023                         if (ef->progtab[pb].addr == NULL) {
1024                                 error = ENOSPC;
1025                                 goto out;
1026                         }
1027                         ef->progtab[pb].size = shdr[i].sh_size;
1028                         ef->progtab[pb].flags = shdr[i].sh_flags;
1029                         ef->progtab[pb].sec = i;
1030                         if (shdr[i].sh_type == SHT_PROGBITS
1031 #ifdef __amd64__
1032                             || shdr[i].sh_type == SHT_X86_64_UNWIND
1033 #endif
1034                             ) {
1035                                 error = vn_rdwr(UIO_READ, nd->ni_vp,
1036                                     ef->progtab[pb].addr,
1037                                     shdr[i].sh_size, shdr[i].sh_offset,
1038                                     UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred,
1039                                     NOCRED, &resid, td);
1040                                 if (error)
1041                                         goto out;
1042                                 if (resid != 0){
1043                                         error = EINVAL;
1044                                         goto out;
1045                                 }
1046                                 /* Initialize the per-cpu or vnet area. */
1047                                 if (ef->progtab[pb].addr != (void *)mapbase &&
1048                                     !strcmp(ef->progtab[pb].name, DPCPU_SETNAME))
1049                                         dpcpu_copy(ef->progtab[pb].addr,
1050                                             shdr[i].sh_size);
1051 #ifdef VIMAGE
1052                                 else if (ef->progtab[pb].addr !=
1053                                     (void *)mapbase &&
1054                                     !strcmp(ef->progtab[pb].name, VNET_SETNAME))
1055                                         vnet_data_copy(ef->progtab[pb].addr,
1056                                             shdr[i].sh_size);
1057 #endif
1058                         } else
1059                                 bzero(ef->progtab[pb].addr, shdr[i].sh_size);
1060
1061                         /* Update all symbol values with the offset. */
1062                         for (j = 0; j < ef->ddbsymcnt; j++) {
1063                                 es = &ef->ddbsymtab[j];
1064                                 if (es->st_shndx != i)
1065                                         continue;
1066                                 es->st_value += (Elf_Addr)ef->progtab[pb].addr;
1067                         }
1068                         mapbase += shdr[i].sh_size;
1069                         pb++;
1070                         break;
1071                 case SHT_REL:
1072                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
1073                                 break;
1074                         ef->reltab[rl].rel = malloc(shdr[i].sh_size, M_LINKER,
1075                             M_WAITOK);
1076                         ef->reltab[rl].nrel = shdr[i].sh_size / sizeof(Elf_Rel);
1077                         ef->reltab[rl].sec = shdr[i].sh_info;
1078                         error = vn_rdwr(UIO_READ, nd->ni_vp,
1079                             (void *)ef->reltab[rl].rel,
1080                             shdr[i].sh_size, shdr[i].sh_offset,
1081                             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
1082                             &resid, td);
1083                         if (error)
1084                                 goto out;
1085                         if (resid != 0){
1086                                 error = EINVAL;
1087                                 goto out;
1088                         }
1089                         rl++;
1090                         break;
1091                 case SHT_RELA:
1092                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
1093                                 break;
1094                         ef->relatab[ra].rela = malloc(shdr[i].sh_size, M_LINKER,
1095                             M_WAITOK);
1096                         ef->relatab[ra].nrela =
1097                             shdr[i].sh_size / sizeof(Elf_Rela);
1098                         ef->relatab[ra].sec = shdr[i].sh_info;
1099                         error = vn_rdwr(UIO_READ, nd->ni_vp,
1100                             (void *)ef->relatab[ra].rela,
1101                             shdr[i].sh_size, shdr[i].sh_offset,
1102                             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
1103                             &resid, td);
1104                         if (error)
1105                                 goto out;
1106                         if (resid != 0){
1107                                 error = EINVAL;
1108                                 goto out;
1109                         }
1110                         ra++;
1111                         break;
1112                 }
1113         }
1114         if (pb != ef->nprogtab) {
1115                 link_elf_error(filename, "lost progbits");
1116                 error = ENOEXEC;
1117                 goto out;
1118         }
1119         if (rl != ef->nreltab) {
1120                 link_elf_error(filename, "lost reltab");
1121                 error = ENOEXEC;
1122                 goto out;
1123         }
1124         if (ra != ef->nrelatab) {
1125                 link_elf_error(filename, "lost relatab");
1126                 error = ENOEXEC;
1127                 goto out;
1128         }
1129         if (mapbase != (vm_offset_t)ef->address + mapsize) {
1130                 printf(
1131                     "%s: mapbase 0x%lx != address %p + mapsize 0x%lx (0x%lx)\n",
1132                     filename != NULL ? filename : "<none>",
1133                     (u_long)mapbase, ef->address, (u_long)mapsize,
1134                     (u_long)(vm_offset_t)ef->address + mapsize);
1135                 error = ENOMEM;
1136                 goto out;
1137         }
1138
1139         /* Local intra-module relocations */
1140         error = link_elf_reloc_local(lf, false);
1141         if (error != 0)
1142                 goto out;
1143
1144         /* Pull in dependencies */
1145         VOP_UNLOCK(nd->ni_vp);
1146         error = linker_load_dependencies(lf);
1147         vn_lock(nd->ni_vp, LK_EXCLUSIVE | LK_RETRY);
1148         if (error)
1149                 goto out;
1150
1151         /* External relocations */
1152         error = relocate_file(ef);
1153         if (error)
1154                 goto out;
1155
1156         /* Notify MD code that a module is being loaded. */
1157         error = elf_cpu_load_file(lf);
1158         if (error)
1159                 goto out;
1160
1161 #if defined(__i386__) || defined(__amd64__)
1162         /* Now ifuncs. */
1163         error = link_elf_reloc_local(lf, true);
1164         if (error != 0)
1165                 goto out;
1166 #endif
1167
1168         link_elf_protect(ef);
1169         link_elf_invoke_ctors(lf->ctors_addr, lf->ctors_size);
1170         *result = lf;
1171
1172 out:
1173         VOP_UNLOCK(nd->ni_vp);
1174         vn_close(nd->ni_vp, FREAD, td->td_ucred, td);
1175         free(nd, M_TEMP);
1176         if (error && lf)
1177                 linker_file_unload(lf, LINKER_UNLOAD_FORCE);
1178         free(hdr, M_LINKER);
1179
1180         return error;
1181 }
1182
1183 static void
1184 link_elf_unload_file(linker_file_t file)
1185 {
1186         elf_file_t ef = (elf_file_t) file;
1187         u_int i;
1188
1189         /* Notify MD code that a module is being unloaded. */
1190         elf_cpu_unload_file(file);
1191
1192         if (ef->progtab) {
1193                 for (i = 0; i < ef->nprogtab; i++) {
1194                         if (ef->progtab[i].size == 0)
1195                                 continue;
1196                         if (ef->progtab[i].name == NULL)
1197                                 continue;
1198                         if (!strcmp(ef->progtab[i].name, DPCPU_SETNAME))
1199                                 dpcpu_free(ef->progtab[i].addr,
1200                                     ef->progtab[i].size);
1201 #ifdef VIMAGE
1202                         else if (!strcmp(ef->progtab[i].name, VNET_SETNAME))
1203                                 vnet_data_free(ef->progtab[i].addr,
1204                                     ef->progtab[i].size);
1205 #endif
1206                 }
1207         }
1208         if (ef->preloaded) {
1209                 free(ef->reltab, M_LINKER);
1210                 free(ef->relatab, M_LINKER);
1211                 free(ef->progtab, M_LINKER);
1212                 free(ef->ctftab, M_LINKER);
1213                 free(ef->ctfoff, M_LINKER);
1214                 free(ef->typoff, M_LINKER);
1215                 if (file->pathname != NULL)
1216                         preload_delete_name(file->pathname);
1217                 return;
1218         }
1219
1220         for (i = 0; i < ef->nreltab; i++)
1221                 free(ef->reltab[i].rel, M_LINKER);
1222         for (i = 0; i < ef->nrelatab; i++)
1223                 free(ef->relatab[i].rela, M_LINKER);
1224         free(ef->reltab, M_LINKER);
1225         free(ef->relatab, M_LINKER);
1226         free(ef->progtab, M_LINKER);
1227
1228         if (ef->object != NULL)
1229                 vm_map_remove(kernel_map, (vm_offset_t)ef->address,
1230                     (vm_offset_t)ef->address + ptoa(ef->object->size));
1231         free(ef->e_shdr, M_LINKER);
1232         free(ef->ddbsymtab, M_LINKER);
1233         free(ef->ddbstrtab, M_LINKER);
1234         free(ef->shstrtab, M_LINKER);
1235         free(ef->ctftab, M_LINKER);
1236         free(ef->ctfoff, M_LINKER);
1237         free(ef->typoff, M_LINKER);
1238 }
1239
1240 static const char *
1241 symbol_name(elf_file_t ef, Elf_Size r_info)
1242 {
1243         const Elf_Sym *ref;
1244
1245         if (ELF_R_SYM(r_info)) {
1246                 ref = ef->ddbsymtab + ELF_R_SYM(r_info);
1247                 return ef->ddbstrtab + ref->st_name;
1248         } else
1249                 return NULL;
1250 }
1251
1252 static Elf_Addr
1253 findbase(elf_file_t ef, int sec)
1254 {
1255         int i;
1256         Elf_Addr base = 0;
1257
1258         for (i = 0; i < ef->nprogtab; i++) {
1259                 if (sec == ef->progtab[i].sec) {
1260                         base = (Elf_Addr)ef->progtab[i].addr;
1261                         break;
1262                 }
1263         }
1264         return base;
1265 }
1266
1267 static int
1268 relocate_file(elf_file_t ef)
1269 {
1270         const Elf_Rel *rellim;
1271         const Elf_Rel *rel;
1272         const Elf_Rela *relalim;
1273         const Elf_Rela *rela;
1274         const char *symname;
1275         const Elf_Sym *sym;
1276         int i;
1277         Elf_Size symidx;
1278         Elf_Addr base;
1279
1280         /* Perform relocations without addend if there are any: */
1281         for (i = 0; i < ef->nreltab; i++) {
1282                 rel = ef->reltab[i].rel;
1283                 if (rel == NULL) {
1284                         link_elf_error(ef->lf.filename, "lost a reltab!");
1285                         return (ENOEXEC);
1286                 }
1287                 rellim = rel + ef->reltab[i].nrel;
1288                 base = findbase(ef, ef->reltab[i].sec);
1289                 if (base == 0) {
1290                         link_elf_error(ef->lf.filename, "lost base for reltab");
1291                         return (ENOEXEC);
1292                 }
1293                 for ( ; rel < rellim; rel++) {
1294                         symidx = ELF_R_SYM(rel->r_info);
1295                         if (symidx >= ef->ddbsymcnt)
1296                                 continue;
1297                         sym = ef->ddbsymtab + symidx;
1298                         /* Local relocs are already done */
1299                         if (ELF_ST_BIND(sym->st_info) == STB_LOCAL)
1300                                 continue;
1301                         if (elf_reloc(&ef->lf, base, rel, ELF_RELOC_REL,
1302                             elf_obj_lookup)) {
1303                                 symname = symbol_name(ef, rel->r_info);
1304                                 printf("link_elf_obj: symbol %s undefined\n",
1305                                     symname);
1306                                 return (ENOENT);
1307                         }
1308                 }
1309         }
1310
1311         /* Perform relocations with addend if there are any: */
1312         for (i = 0; i < ef->nrelatab; i++) {
1313                 rela = ef->relatab[i].rela;
1314                 if (rela == NULL) {
1315                         link_elf_error(ef->lf.filename, "lost a relatab!");
1316                         return (ENOEXEC);
1317                 }
1318                 relalim = rela + ef->relatab[i].nrela;
1319                 base = findbase(ef, ef->relatab[i].sec);
1320                 if (base == 0) {
1321                         link_elf_error(ef->lf.filename,
1322                             "lost base for relatab");
1323                         return (ENOEXEC);
1324                 }
1325                 for ( ; rela < relalim; rela++) {
1326                         symidx = ELF_R_SYM(rela->r_info);
1327                         if (symidx >= ef->ddbsymcnt)
1328                                 continue;
1329                         sym = ef->ddbsymtab + symidx;
1330                         /* Local relocs are already done */
1331                         if (ELF_ST_BIND(sym->st_info) == STB_LOCAL)
1332                                 continue;
1333                         if (elf_reloc(&ef->lf, base, rela, ELF_RELOC_RELA,
1334                             elf_obj_lookup)) {
1335                                 symname = symbol_name(ef, rela->r_info);
1336                                 printf("link_elf_obj: symbol %s undefined\n",
1337                                     symname);
1338                                 return (ENOENT);
1339                         }
1340                 }
1341         }
1342
1343         /*
1344          * Only clean SHN_FBSD_CACHED for successful return.  If we
1345          * modified symbol table for the object but found an
1346          * unresolved symbol, there is no reason to roll back.
1347          */
1348         elf_obj_cleanup_globals_cache(ef);
1349
1350         return (0);
1351 }
1352
1353 static int
1354 link_elf_lookup_symbol(linker_file_t lf, const char *name, c_linker_sym_t *sym)
1355 {
1356         elf_file_t ef = (elf_file_t) lf;
1357         const Elf_Sym *symp;
1358         const char *strp;
1359         int i;
1360
1361         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1362                 strp = ef->ddbstrtab + symp->st_name;
1363                 if (symp->st_shndx != SHN_UNDEF && strcmp(name, strp) == 0) {
1364                         *sym = (c_linker_sym_t) symp;
1365                         return 0;
1366                 }
1367         }
1368         return ENOENT;
1369 }
1370
1371 static int
1372 link_elf_symbol_values(linker_file_t lf, c_linker_sym_t sym,
1373     linker_symval_t *symval)
1374 {
1375         elf_file_t ef;
1376         const Elf_Sym *es;
1377         caddr_t val;
1378
1379         ef = (elf_file_t) lf;
1380         es = (const Elf_Sym*) sym;
1381         val = (caddr_t)es->st_value;
1382         if (es >= ef->ddbsymtab && es < (ef->ddbsymtab + ef->ddbsymcnt)) {
1383                 symval->name = ef->ddbstrtab + es->st_name;
1384                 val = (caddr_t)es->st_value;
1385                 if (ELF_ST_TYPE(es->st_info) == STT_GNU_IFUNC)
1386                         val = ((caddr_t (*)(void))val)();
1387                 symval->value = val;
1388                 symval->size = es->st_size;
1389                 return 0;
1390         }
1391         return ENOENT;
1392 }
1393
1394 static int
1395 link_elf_search_symbol(linker_file_t lf, caddr_t value,
1396     c_linker_sym_t *sym, long *diffp)
1397 {
1398         elf_file_t ef = (elf_file_t) lf;
1399         u_long off = (uintptr_t) (void *) value;
1400         u_long diff = off;
1401         u_long st_value;
1402         const Elf_Sym *es;
1403         const Elf_Sym *best = NULL;
1404         int i;
1405
1406         for (i = 0, es = ef->ddbsymtab; i < ef->ddbsymcnt; i++, es++) {
1407                 if (es->st_name == 0)
1408                         continue;
1409                 st_value = es->st_value;
1410                 if (off >= st_value) {
1411                         if (off - st_value < diff) {
1412                                 diff = off - st_value;
1413                                 best = es;
1414                                 if (diff == 0)
1415                                         break;
1416                         } else if (off - st_value == diff) {
1417                                 best = es;
1418                         }
1419                 }
1420         }
1421         if (best == NULL)
1422                 *diffp = off;
1423         else
1424                 *diffp = diff;
1425         *sym = (c_linker_sym_t) best;
1426
1427         return 0;
1428 }
1429
1430 /*
1431  * Look up a linker set on an ELF system.
1432  */
1433 static int
1434 link_elf_lookup_set(linker_file_t lf, const char *name,
1435     void ***startp, void ***stopp, int *countp)
1436 {
1437         elf_file_t ef = (elf_file_t)lf;
1438         void **start, **stop;
1439         int i, count;
1440
1441         /* Relative to section number */
1442         for (i = 0; i < ef->nprogtab; i++) {
1443                 if ((strncmp(ef->progtab[i].name, "set_", 4) == 0) &&
1444                     strcmp(ef->progtab[i].name + 4, name) == 0) {
1445                         start  = (void **)ef->progtab[i].addr;
1446                         stop = (void **)((char *)ef->progtab[i].addr +
1447                             ef->progtab[i].size);
1448                         count = stop - start;
1449                         if (startp)
1450                                 *startp = start;
1451                         if (stopp)
1452                                 *stopp = stop;
1453                         if (countp)
1454                                 *countp = count;
1455                         return (0);
1456                 }
1457         }
1458         return (ESRCH);
1459 }
1460
1461 static int
1462 link_elf_each_function_name(linker_file_t file,
1463     int (*callback)(const char *, void *), void *opaque)
1464 {
1465         elf_file_t ef = (elf_file_t)file;
1466         const Elf_Sym *symp;
1467         int i, error;
1468
1469         /* Exhaustive search */
1470         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1471                 if (symp->st_value != 0 &&
1472                     (ELF_ST_TYPE(symp->st_info) == STT_FUNC ||
1473                     ELF_ST_TYPE(symp->st_info) == STT_GNU_IFUNC)) {
1474                         error = callback(ef->ddbstrtab + symp->st_name, opaque);
1475                         if (error)
1476                                 return (error);
1477                 }
1478         }
1479         return (0);
1480 }
1481
1482 static int
1483 link_elf_each_function_nameval(linker_file_t file,
1484     linker_function_nameval_callback_t callback, void *opaque)
1485 {
1486         linker_symval_t symval;
1487         elf_file_t ef = (elf_file_t)file;
1488         const Elf_Sym* symp;
1489         int i, error;
1490
1491         /* Exhaustive search */
1492         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1493                 if (symp->st_value != 0 &&
1494                     (ELF_ST_TYPE(symp->st_info) == STT_FUNC ||
1495                     ELF_ST_TYPE(symp->st_info) == STT_GNU_IFUNC)) {
1496                         error = link_elf_symbol_values(file,
1497                             (c_linker_sym_t)symp, &symval);
1498                         if (error)
1499                                 return (error);
1500                         error = callback(file, i, &symval, opaque);
1501                         if (error)
1502                                 return (error);
1503                 }
1504         }
1505         return (0);
1506 }
1507
1508 static void
1509 elf_obj_cleanup_globals_cache(elf_file_t ef)
1510 {
1511         Elf_Sym *sym;
1512         Elf_Size i;
1513
1514         for (i = 0; i < ef->ddbsymcnt; i++) {
1515                 sym = ef->ddbsymtab + i;
1516                 if (sym->st_shndx == SHN_FBSD_CACHED) {
1517                         sym->st_shndx = SHN_UNDEF;
1518                         sym->st_value = 0;
1519                 }
1520         }
1521 }
1522
1523 /*
1524  * Symbol lookup function that can be used when the symbol index is known (ie
1525  * in relocations). It uses the symbol index instead of doing a fully fledged
1526  * hash table based lookup when such is valid. For example for local symbols.
1527  * This is not only more efficient, it's also more correct. It's not always
1528  * the case that the symbol can be found through the hash table.
1529  */
1530 static int
1531 elf_obj_lookup(linker_file_t lf, Elf_Size symidx, int deps, Elf_Addr *res)
1532 {
1533         elf_file_t ef = (elf_file_t)lf;
1534         Elf_Sym *sym;
1535         const char *symbol;
1536         Elf_Addr res1;
1537
1538         /* Don't even try to lookup the symbol if the index is bogus. */
1539         if (symidx >= ef->ddbsymcnt) {
1540                 *res = 0;
1541                 return (EINVAL);
1542         }
1543
1544         sym = ef->ddbsymtab + symidx;
1545
1546         /* Quick answer if there is a definition included. */
1547         if (sym->st_shndx != SHN_UNDEF) {
1548                 res1 = (Elf_Addr)sym->st_value;
1549                 if (ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC)
1550                         res1 = ((Elf_Addr (*)(void))res1)();
1551                 *res = res1;
1552                 return (0);
1553         }
1554
1555         /* If we get here, then it is undefined and needs a lookup. */
1556         switch (ELF_ST_BIND(sym->st_info)) {
1557         case STB_LOCAL:
1558                 /* Local, but undefined? huh? */
1559                 *res = 0;
1560                 return (EINVAL);
1561
1562         case STB_GLOBAL:
1563         case STB_WEAK:
1564                 /* Relative to Data or Function name */
1565                 symbol = ef->ddbstrtab + sym->st_name;
1566
1567                 /* Force a lookup failure if the symbol name is bogus. */
1568                 if (*symbol == 0) {
1569                         *res = 0;
1570                         return (EINVAL);
1571                 }
1572                 res1 = (Elf_Addr)linker_file_lookup_symbol(lf, symbol, deps);
1573
1574                 /*
1575                  * Cache global lookups during module relocation. The failure
1576                  * case is particularly expensive for callers, who must scan
1577                  * through the entire globals table doing strcmp(). Cache to
1578                  * avoid doing such work repeatedly.
1579                  *
1580                  * After relocation is complete, undefined globals will be
1581                  * restored to SHN_UNDEF in elf_obj_cleanup_globals_cache(),
1582                  * above.
1583                  */
1584                 if (res1 != 0) {
1585                         sym->st_shndx = SHN_FBSD_CACHED;
1586                         sym->st_value = res1;
1587                         *res = res1;
1588                         return (0);
1589                 } else if (ELF_ST_BIND(sym->st_info) == STB_WEAK) {
1590                         sym->st_value = 0;
1591                         *res = 0;
1592                         return (0);
1593                 }
1594                 return (EINVAL);
1595
1596         default:
1597                 return (EINVAL);
1598         }
1599 }
1600
1601 static void
1602 link_elf_fix_link_set(elf_file_t ef)
1603 {
1604         static const char startn[] = "__start_";
1605         static const char stopn[] = "__stop_";
1606         Elf_Sym *sym;
1607         const char *sym_name, *linkset_name;
1608         Elf_Addr startp, stopp;
1609         Elf_Size symidx;
1610         int start, i;
1611
1612         startp = stopp = 0;
1613         for (symidx = 1 /* zero entry is special */;
1614                 symidx < ef->ddbsymcnt; symidx++) {
1615                 sym = ef->ddbsymtab + symidx;
1616                 if (sym->st_shndx != SHN_UNDEF)
1617                         continue;
1618
1619                 sym_name = ef->ddbstrtab + sym->st_name;
1620                 if (strncmp(sym_name, startn, sizeof(startn) - 1) == 0) {
1621                         start = 1;
1622                         linkset_name = sym_name + sizeof(startn) - 1;
1623                 }
1624                 else if (strncmp(sym_name, stopn, sizeof(stopn) - 1) == 0) {
1625                         start = 0;
1626                         linkset_name = sym_name + sizeof(stopn) - 1;
1627                 }
1628                 else
1629                         continue;
1630
1631                 for (i = 0; i < ef->nprogtab; i++) {
1632                         if (strcmp(ef->progtab[i].name, linkset_name) == 0) {
1633                                 startp = (Elf_Addr)ef->progtab[i].addr;
1634                                 stopp = (Elf_Addr)(startp + ef->progtab[i].size);
1635                                 break;
1636                         }
1637                 }
1638                 if (i == ef->nprogtab)
1639                         continue;
1640
1641                 sym->st_value = start ? startp : stopp;
1642                 sym->st_shndx = i;
1643         }
1644 }
1645
1646 static int
1647 link_elf_reloc_local(linker_file_t lf, bool ifuncs)
1648 {
1649         elf_file_t ef = (elf_file_t)lf;
1650         const Elf_Rel *rellim;
1651         const Elf_Rel *rel;
1652         const Elf_Rela *relalim;
1653         const Elf_Rela *rela;
1654         const Elf_Sym *sym;
1655         Elf_Addr base;
1656         int i;
1657         Elf_Size symidx;
1658
1659         link_elf_fix_link_set(ef);
1660
1661         /* Perform relocations without addend if there are any: */
1662         for (i = 0; i < ef->nreltab; i++) {
1663                 rel = ef->reltab[i].rel;
1664                 if (rel == NULL) {
1665                         link_elf_error(ef->lf.filename, "lost a reltab");
1666                         return (ENOEXEC);
1667                 }
1668                 rellim = rel + ef->reltab[i].nrel;
1669                 base = findbase(ef, ef->reltab[i].sec);
1670                 if (base == 0) {
1671                         link_elf_error(ef->lf.filename, "lost base for reltab");
1672                         return (ENOEXEC);
1673                 }
1674                 for ( ; rel < rellim; rel++) {
1675                         symidx = ELF_R_SYM(rel->r_info);
1676                         if (symidx >= ef->ddbsymcnt)
1677                                 continue;
1678                         sym = ef->ddbsymtab + symidx;
1679                         /* Only do local relocs */
1680                         if (ELF_ST_BIND(sym->st_info) != STB_LOCAL)
1681                                 continue;
1682                         if ((ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC ||
1683                             elf_is_ifunc_reloc(rel->r_info)) != ifuncs)
1684                                 continue;
1685                         if (elf_reloc_local(lf, base, rel, ELF_RELOC_REL,
1686                             elf_obj_lookup) != 0)
1687                                 return (ENOEXEC);
1688                 }
1689         }
1690
1691         /* Perform relocations with addend if there are any: */
1692         for (i = 0; i < ef->nrelatab; i++) {
1693                 rela = ef->relatab[i].rela;
1694                 if (rela == NULL) {
1695                         link_elf_error(ef->lf.filename, "lost a relatab!");
1696                         return (ENOEXEC);
1697                 }
1698                 relalim = rela + ef->relatab[i].nrela;
1699                 base = findbase(ef, ef->relatab[i].sec);
1700                 if (base == 0) {
1701                         link_elf_error(ef->lf.filename, "lost base for reltab");
1702                         return (ENOEXEC);
1703                 }
1704                 for ( ; rela < relalim; rela++) {
1705                         symidx = ELF_R_SYM(rela->r_info);
1706                         if (symidx >= ef->ddbsymcnt)
1707                                 continue;
1708                         sym = ef->ddbsymtab + symidx;
1709                         /* Only do local relocs */
1710                         if (ELF_ST_BIND(sym->st_info) != STB_LOCAL)
1711                                 continue;
1712                         if ((ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC ||
1713                             elf_is_ifunc_reloc(rela->r_info)) != ifuncs)
1714                                 continue;
1715                         if (elf_reloc_local(lf, base, rela, ELF_RELOC_RELA,
1716                             elf_obj_lookup) != 0)
1717                                 return (ENOEXEC);
1718                 }
1719         }
1720         return (0);
1721 }
1722
1723 static long
1724 link_elf_symtab_get(linker_file_t lf, const Elf_Sym **symtab)
1725 {
1726     elf_file_t ef = (elf_file_t)lf;
1727
1728     *symtab = ef->ddbsymtab;
1729
1730     if (*symtab == NULL)
1731         return (0);
1732
1733     return (ef->ddbsymcnt);
1734 }
1735
1736 static long
1737 link_elf_strtab_get(linker_file_t lf, caddr_t *strtab)
1738 {
1739     elf_file_t ef = (elf_file_t)lf;
1740
1741     *strtab = ef->ddbstrtab;
1742
1743     if (*strtab == NULL)
1744         return (0);
1745
1746     return (ef->ddbstrcnt);
1747 }