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                 case SHT_INIT_ARRAY:
 390                 case SHT_FINI_ARRAY:
 391                         /* Ignore sections not loaded by the loader. */
 392                         if (shdr[i].sh_addr == 0)
 393                                 break;
 394                         ef->nprogtab++;
 395                         break;
 396                 case SHT_SYMTAB:
 397                         symtabindex = i;
 398                         symstrindex = shdr[i].sh_link;
 399                         break;
 400                 case SHT_REL:
 401                         /*
 402                          * Ignore relocation tables for sections not
 403                          * loaded by the loader.
 404                          */
 405                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 406                                 break;
 407                         ef->nreltab++;
 408                         break;
 409                 case SHT_RELA:
 410                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 411                                 break;
 412                         ef->nrelatab++;
 413                         break;
 414                 }
 415         }
 416
 417         shstrindex = hdr->e_shstrndx;
 418         if (ef->nprogtab == 0 || symstrindex < 0 ||
 419             symstrindex >= hdr->e_shnum ||
 420             shdr[symstrindex].sh_type != SHT_STRTAB || shstrindex == 0 ||
 421             shstrindex >= hdr->e_shnum ||
 422             shdr[shstrindex].sh_type != SHT_STRTAB) {
 423                 printf("%s: bad/missing section headers\n", filename);
 424                 error = ENOEXEC;
 425                 goto out;
 426         }
 427
 428         /* Allocate space for tracking the load chunks */
 429         if (ef->nprogtab != 0)
 430                 ef->progtab = malloc(ef->nprogtab * sizeof(*ef->progtab),
 431                     M_LINKER, M_WAITOK | M_ZERO);
 432         if (ef->nreltab != 0)
 433                 ef->reltab = malloc(ef->nreltab * sizeof(*ef->reltab),
 434                     M_LINKER, M_WAITOK | M_ZERO);
 435         if (ef->nrelatab != 0)
 436                 ef->relatab = malloc(ef->nrelatab * sizeof(*ef->relatab),
 437                     M_LINKER, M_WAITOK | M_ZERO);
 438         if ((ef->nprogtab != 0 && ef->progtab == NULL) ||
 439             (ef->nreltab != 0 && ef->reltab == NULL) ||
 440             (ef->nrelatab != 0 && ef->relatab == NULL)) {
 441                 error = ENOMEM;
 442                 goto out;
 443         }
 444
 445         /* XXX, relocate the sh_addr fields saved by the loader. */
 446         off = 0;
 447         for (i = 0; i < hdr->e_shnum; i++) {
 448                 if (shdr[i].sh_addr != 0 && (off == 0 || shdr[i].sh_addr < off))
 449                         off = shdr[i].sh_addr;
 450         }
 451         for (i = 0; i < hdr->e_shnum; i++) {
 452                 if (shdr[i].sh_addr != 0)
 453                         shdr[i].sh_addr = shdr[i].sh_addr - off +
 454                             (Elf_Addr)ef->address;
 455         }
 456
 457         ef->ddbsymcnt = shdr[symtabindex].sh_size / sizeof(Elf_Sym);
 458         ef->ddbsymtab = (Elf_Sym *)shdr[symtabindex].sh_addr;
 459         ef->ddbstrcnt = shdr[symstrindex].sh_size;
 460         ef->ddbstrtab = (char *)shdr[symstrindex].sh_addr;
 461         ef->shstrcnt = shdr[shstrindex].sh_size;
 462         ef->shstrtab = (char *)shdr[shstrindex].sh_addr;
 463
 464         /* Now fill out progtab and the relocation tables. */
 465         pb = 0;
 466         rl = 0;
 467         ra = 0;
 468         for (i = 0; i < hdr->e_shnum; i++) {
 469                 switch (shdr[i].sh_type) {
 470                 case SHT_PROGBITS:
 471                 case SHT_NOBITS:
 472 #ifdef __amd64__
 473                 case SHT_X86_64_UNWIND:
 474 #endif
 475                 case SHT_FINI_ARRAY:
 476                         if (shdr[i].sh_addr == 0)
 477                                 break;
 478                         ef->progtab[pb].addr = (void *)shdr[i].sh_addr;
 479                         if (shdr[i].sh_type == SHT_PROGBITS)
 480                                 ef->progtab[pb].name = "<<PROGBITS>>";
 481 #ifdef __amd64__
 482                         else if (shdr[i].sh_type == SHT_X86_64_UNWIND)
 483                                 ef->progtab[pb].name = "<<UNWIND>>";
 484 #endif
 485                         else if (shdr[i].sh_type == SHT_INIT_ARRAY)
 486                                 ef->progtab[pb].name = "<<INIT_ARRAY>>";
 487                         else if (shdr[i].sh_type == SHT_FINI_ARRAY)
 488                                 ef->progtab[pb].name = "<<FINI_ARRAY>>";
 489                         else
 490                                 ef->progtab[pb].name = "<<NOBITS>>";
 491                         ef->progtab[pb].size = shdr[i].sh_size;
 492                         ef->progtab[pb].flags = shdr[i].sh_flags;
 493                         ef->progtab[pb].sec = i;
 494                         if (ef->shstrtab && shdr[i].sh_name != 0)
 495                                 ef->progtab[pb].name =
 496                                     ef->shstrtab + shdr[i].sh_name;
 497                         if (ef->progtab[pb].name != NULL &&
 498                             !strcmp(ef->progtab[pb].name, DPCPU_SETNAME)) {
 499                                 void *dpcpu;
 500
 501                                 dpcpu = dpcpu_alloc(shdr[i].sh_size);
 502                                 if (dpcpu == NULL) {
 503                                         printf("%s: pcpu module space is out "
 504                                             "of space; cannot allocate %#jx "
 505                                             "for %s\n", __func__,
 506                                             (uintmax_t)shdr[i].sh_size,
 507                                             filename);
 508                                         error = ENOSPC;
 509                                         goto out;
 510                                 }
 511                                 memcpy(dpcpu, ef->progtab[pb].addr,
 512                                     ef->progtab[pb].size);
 513                                 dpcpu_copy(dpcpu, shdr[i].sh_size);
 514                                 ef->progtab[pb].addr = dpcpu;
 515 #ifdef VIMAGE
 516                         } else if (ef->progtab[pb].name != NULL &&
 517                             !strcmp(ef->progtab[pb].name, VNET_SETNAME)) {
 518                                 void *vnet_data;
 519
 520                                 vnet_data = vnet_data_alloc(shdr[i].sh_size);
 521                                 if (vnet_data == NULL) {
 522                                         printf("%s: vnet module space is out "
 523                                             "of space; cannot allocate %#jx "
 524                                             "for %s\n", __func__,
 525                                             (uintmax_t)shdr[i].sh_size,
 526                                             filename);
 527                                         error = ENOSPC;
 528                                         goto out;
 529                                 }
 530                                 memcpy(vnet_data, ef->progtab[pb].addr,
 531                                     ef->progtab[pb].size);
 532                                 vnet_data_copy(vnet_data, shdr[i].sh_size);
 533                                 ef->progtab[pb].addr = vnet_data;
 534 #endif
 535                         } else if ((ef->progtab[pb].name != NULL &&
 536                             strcmp(ef->progtab[pb].name, ".ctors") == 0) ||
 537                             shdr[i].sh_type == SHT_INIT_ARRAY) {
 538                                 if (lf->ctors_addr != 0) {
 539                                         printf(
 540                                     "%s: multiple ctor sections in %s\n",
 541                                             __func__, filename);
 542                                 } else {
 543                                         lf->ctors_addr = ef->progtab[pb].addr;
 544                                         lf->ctors_size = shdr[i].sh_size;
 545                                 }
 546                         }
 547
 548                         /* Update all symbol values with the offset. */
 549                         for (j = 0; j < ef->ddbsymcnt; j++) {
 550                                 es = &ef->ddbsymtab[j];
 551                                 if (es->st_shndx != i)
 552                                         continue;
 553                                 es->st_value += (Elf_Addr)ef->progtab[pb].addr;
 554                         }
 555                         pb++;
 556                         break;
 557                 case SHT_REL:
 558                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 559                                 break;
 560                         ef->reltab[rl].rel = (Elf_Rel *)shdr[i].sh_addr;
 561                         ef->reltab[rl].nrel = shdr[i].sh_size / sizeof(Elf_Rel);
 562                         ef->reltab[rl].sec = shdr[i].sh_info;
 563                         rl++;
 564                         break;
 565                 case SHT_RELA:
 566                         if (shdr[shdr[i].sh_info].sh_addr == 0)
 567                                 break;
 568                         ef->relatab[ra].rela = (Elf_Rela *)shdr[i].sh_addr;
 569                         ef->relatab[ra].nrela =
 570                             shdr[i].sh_size / sizeof(Elf_Rela);
 571                         ef->relatab[ra].sec = shdr[i].sh_info;
 572                         ra++;
 573                         break;
 574                 }
 575         }
 576         if (pb != ef->nprogtab) {
 577                 printf("%s: lost progbits\n", filename);
 578                 error = ENOEXEC;
 579                 goto out;
 580         }
 581         if (rl != ef->nreltab) {
 582                 printf("%s: lost reltab\n", filename);
 583                 error = ENOEXEC;
 584                 goto out;
 585         }
 586         if (ra != ef->nrelatab) {
 587                 printf("%s: lost relatab\n", filename);
 588                 error = ENOEXEC;
 589                 goto out;
 590         }
 591
 592         /*
 593          * The file needs to be writeable and executable while applying
 594          * relocations.  Mapping protections are applied once relocation
 595          * processing is complete.
 596          */
 597         link_elf_protect_range(ef, (vm_offset_t)ef->address,
 598             round_page((vm_offset_t)ef->address + ef->lf.size), VM_PROT_ALL);
 599
 600         /* Local intra-module relocations */
 601         error = link_elf_reloc_local(lf, false);
 602         if (error != 0)
 603                 goto out;
 604         *result = lf;
 605         return (0);
 606
 607 out:
 608         /* preload not done this way */
 609         linker_file_unload(lf, LINKER_UNLOAD_FORCE);
 610         return (error);
 611 }
 612
 613 static void
 614 link_elf_invoke_ctors(caddr_t addr, size_t size)
 615 {
 616         void (**ctor)(void);
 617         size_t i, cnt;
 618
 619         if (addr == NULL || size == 0)
 620                 return;
 621         cnt = size / sizeof(*ctor);
 622         ctor = (void *)addr;
 623         for (i = 0; i < cnt; i++) {
 624                 if (ctor[i] != NULL)
 625                         (*ctor[i])();
 626         }
 627 }
 628
 629 static int
 630 link_elf_link_preload_finish(linker_file_t lf)
 631 {
 632         elf_file_t ef;
 633         int error;
 634
 635         ef = (elf_file_t)lf;
 636         error = relocate_file(ef);
 637         if (error)
 638                 return (error);
 639
 640         /* Notify MD code that a module is being loaded. */
 641         error = elf_cpu_load_file(lf);
 642         if (error)
 643                 return (error);
 644
 645 #if defined(__i386__) || defined(__amd64__)
 646         /* Now ifuncs. */
 647         error = link_elf_reloc_local(lf, true);
 648         if (error != 0)
 649                 return (error);
 650 #endif
 651
 652         /* Apply protections now that relocation processing is complete. */
 653         link_elf_protect(ef);
 654
 655         link_elf_invoke_ctors(lf->ctors_addr, lf->ctors_size);
 656         return (0);
 657 }
 658
 659 static int
 660 link_elf_load_file(linker_class_t cls, const char *filename,
 661     linker_file_t *result)
 662 {
 663         struct nameidata *nd;
 664         struct thread *td = curthread;  /* XXX */
 665         Elf_Ehdr *hdr;
 666         Elf_Shdr *shdr;
 667         Elf_Sym *es;
 668         int nbytes, i, j;
 669         vm_offset_t mapbase;
 670         size_t mapsize;
 671         int error = 0;
 672         ssize_t resid;
 673         int flags;
 674         elf_file_t ef;
 675         linker_file_t lf;
 676         int symtabindex;
 677         int symstrindex;
 678         int shstrindex;
 679         int nsym;
 680         int pb, rl, ra;
 681         int alignmask;
 682
 683         shdr = NULL;
 684         lf = NULL;
 685         mapsize = 0;
 686         hdr = NULL;
 687
 688         nd = malloc(sizeof(struct nameidata), M_TEMP, M_WAITOK);
 689         NDINIT(nd, LOOKUP, FOLLOW, UIO_SYSSPACE, filename, td);
 690         flags = FREAD;
 691         error = vn_open(nd, &flags, 0, NULL);
 692         if (error) {
 693                 free(nd, M_TEMP);
 694                 return error;
 695         }
 696         NDFREE(nd, NDF_ONLY_PNBUF);
 697         if (nd->ni_vp->v_type != VREG) {
 698                 error = ENOEXEC;
 699                 goto out;
 700         }
 701 #ifdef MAC
 702         error = mac_kld_check_load(td->td_ucred, nd->ni_vp);
 703         if (error) {
 704                 goto out;
 705         }
 706 #endif
 707
 708         /* Read the elf header from the file. */
 709         hdr = malloc(sizeof(*hdr), M_LINKER, M_WAITOK);
 710         error = vn_rdwr(UIO_READ, nd->ni_vp, (void *)hdr, sizeof(*hdr), 0,
 711             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 712             &resid, td);
 713         if (error)
 714                 goto out;
 715         if (resid != 0){
 716                 error = ENOEXEC;
 717                 goto out;
 718         }
 719
 720         if (!IS_ELF(*hdr)) {
 721                 error = ENOEXEC;
 722                 goto out;
 723         }
 724
 725         if (hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS
 726             || hdr->e_ident[EI_DATA] != ELF_TARG_DATA) {
 727                 link_elf_error(filename, "Unsupported file layout");
 728                 error = ENOEXEC;
 729                 goto out;
 730         }
 731         if (hdr->e_ident[EI_VERSION] != EV_CURRENT
 732             || hdr->e_version != EV_CURRENT) {
 733                 link_elf_error(filename, "Unsupported file version");
 734                 error = ENOEXEC;
 735                 goto out;
 736         }
 737         if (hdr->e_type != ET_REL) {
 738                 error = ENOSYS;
 739                 goto out;
 740         }
 741         if (hdr->e_machine != ELF_TARG_MACH) {
 742                 link_elf_error(filename, "Unsupported machine");
 743                 error = ENOEXEC;
 744                 goto out;
 745         }
 746
 747         lf = linker_make_file(filename, &link_elf_class);
 748         if (!lf) {
 749                 error = ENOMEM;
 750                 goto out;
 751         }
 752         ef = (elf_file_t) lf;
 753         ef->nprogtab = 0;
 754         ef->e_shdr = 0;
 755         ef->nreltab = 0;
 756         ef->nrelatab = 0;
 757
 758         /* Allocate and read in the section header */
 759         nbytes = hdr->e_shnum * hdr->e_shentsize;
 760         if (nbytes == 0 || hdr->e_shoff == 0 ||
 761             hdr->e_shentsize != sizeof(Elf_Shdr)) {
 762                 error = ENOEXEC;
 763                 goto out;
 764         }
 765         shdr = malloc(nbytes, M_LINKER, M_WAITOK);
 766         ef->e_shdr = shdr;
 767         error = vn_rdwr(UIO_READ, nd->ni_vp, (caddr_t)shdr, nbytes,
 768             hdr->e_shoff, UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred,
 769             NOCRED, &resid, td);
 770         if (error)
 771                 goto out;
 772         if (resid) {
 773                 error = ENOEXEC;
 774                 goto out;
 775         }
 776
 777         /* Scan the section header for information and table sizing. */
 778         nsym = 0;
 779         symtabindex = -1;
 780         symstrindex = -1;
 781         for (i = 0; i < hdr->e_shnum; i++) {
 782                 if (shdr[i].sh_size == 0)
 783                         continue;
 784                 switch (shdr[i].sh_type) {
 785                 case SHT_PROGBITS:
 786                 case SHT_NOBITS:
 787 #ifdef __amd64__
 788                 case SHT_X86_64_UNWIND:
 789 #endif
 790                 case SHT_INIT_ARRAY:
 791                 case SHT_FINI_ARRAY:
 792                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 793                                 break;
 794                         ef->nprogtab++;
 795                         break;
 796                 case SHT_SYMTAB:
 797                         nsym++;
 798                         symtabindex = i;
 799                         symstrindex = shdr[i].sh_link;
 800                         break;
 801                 case SHT_REL:
 802                         /*
 803                          * Ignore relocation tables for unallocated
 804                          * sections.
 805                          */
 806                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
 807                                 break;
 808                         ef->nreltab++;
 809                         break;
 810                 case SHT_RELA:
 811                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
 812                                 break;
 813                         ef->nrelatab++;
 814                         break;
 815                 case SHT_STRTAB:
 816                         break;
 817                 }
 818         }
 819         if (ef->nprogtab == 0) {
 820                 link_elf_error(filename, "file has no contents");
 821                 error = ENOEXEC;
 822                 goto out;
 823         }
 824         if (nsym != 1) {
 825                 /* Only allow one symbol table for now */
 826                 link_elf_error(filename,
 827                     "file must have exactly one symbol table");
 828                 error = ENOEXEC;
 829                 goto out;
 830         }
 831         if (symstrindex < 0 || symstrindex > hdr->e_shnum ||
 832             shdr[symstrindex].sh_type != SHT_STRTAB) {
 833                 link_elf_error(filename, "file has invalid symbol strings");
 834                 error = ENOEXEC;
 835                 goto out;
 836         }
 837
 838         /* Allocate space for tracking the load chunks */
 839         if (ef->nprogtab != 0)
 840                 ef->progtab = malloc(ef->nprogtab * sizeof(*ef->progtab),
 841                     M_LINKER, M_WAITOK | M_ZERO);
 842         if (ef->nreltab != 0)
 843                 ef->reltab = malloc(ef->nreltab * sizeof(*ef->reltab),
 844                     M_LINKER, M_WAITOK | M_ZERO);
 845         if (ef->nrelatab != 0)
 846                 ef->relatab = malloc(ef->nrelatab * sizeof(*ef->relatab),
 847                     M_LINKER, M_WAITOK | M_ZERO);
 848
 849         if (symtabindex == -1) {
 850                 link_elf_error(filename, "lost symbol table index");
 851                 error = ENOEXEC;
 852                 goto out;
 853         }
 854         /* Allocate space for and load the symbol table */
 855         ef->ddbsymcnt = shdr[symtabindex].sh_size / sizeof(Elf_Sym);
 856         ef->ddbsymtab = malloc(shdr[symtabindex].sh_size, M_LINKER, M_WAITOK);
 857         error = vn_rdwr(UIO_READ, nd->ni_vp, (void *)ef->ddbsymtab,
 858             shdr[symtabindex].sh_size, shdr[symtabindex].sh_offset,
 859             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 860             &resid, td);
 861         if (error)
 862                 goto out;
 863         if (resid != 0){
 864                 error = EINVAL;
 865                 goto out;
 866         }
 867
 868         /* Allocate space for and load the symbol strings */
 869         ef->ddbstrcnt = shdr[symstrindex].sh_size;
 870         ef->ddbstrtab = malloc(shdr[symstrindex].sh_size, M_LINKER, M_WAITOK);
 871         error = vn_rdwr(UIO_READ, nd->ni_vp, ef->ddbstrtab,
 872             shdr[symstrindex].sh_size, shdr[symstrindex].sh_offset,
 873             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 874             &resid, td);
 875         if (error)
 876                 goto out;
 877         if (resid != 0){
 878                 error = EINVAL;
 879                 goto out;
 880         }
 881
 882         /* Do we have a string table for the section names?  */
 883         shstrindex = -1;
 884         if (hdr->e_shstrndx != 0 &&
 885             shdr[hdr->e_shstrndx].sh_type == SHT_STRTAB) {
 886                 shstrindex = hdr->e_shstrndx;
 887                 ef->shstrcnt = shdr[shstrindex].sh_size;
 888                 ef->shstrtab = malloc(shdr[shstrindex].sh_size, M_LINKER,
 889                     M_WAITOK);
 890                 error = vn_rdwr(UIO_READ, nd->ni_vp, ef->shstrtab,
 891                     shdr[shstrindex].sh_size, shdr[shstrindex].sh_offset,
 892                     UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
 893                     &resid, td);
 894                 if (error)
 895                         goto out;
 896                 if (resid != 0){
 897                         error = EINVAL;
 898                         goto out;
 899                 }
 900         }
 901
 902         /* Size up code/data(progbits) and bss(nobits). */
 903         alignmask = 0;
 904         for (i = 0; i < hdr->e_shnum; i++) {
 905                 if (shdr[i].sh_size == 0)
 906                         continue;
 907                 switch (shdr[i].sh_type) {
 908                 case SHT_PROGBITS:
 909                 case SHT_NOBITS:
 910 #ifdef __amd64__
 911                 case SHT_X86_64_UNWIND:
 912 #endif
 913                 case SHT_INIT_ARRAY:
 914                 case SHT_FINI_ARRAY:
 915                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 916                                 break;
 917                         alignmask = shdr[i].sh_addralign - 1;
 918                         mapsize += alignmask;
 919                         mapsize &= ~alignmask;
 920                         mapsize += shdr[i].sh_size;
 921                         break;
 922                 }
 923         }
 924
 925         /*
 926          * We know how much space we need for the text/data/bss/etc.
 927          * This stuff needs to be in a single chunk so that profiling etc
 928          * can get the bounds and gdb can associate offsets with modules
 929          */
 930         ef->object = vm_pager_allocate(OBJT_PHYS, NULL, round_page(mapsize),
 931             VM_PROT_ALL, 0, thread0.td_ucred);
 932         if (ef->object == NULL) {
 933                 error = ENOMEM;
 934                 goto out;
 935         }
 936 #if VM_NRESERVLEVEL > 0
 937         vm_object_color(ef->object, 0);
 938 #endif
 939
 940         /*
 941          * In order to satisfy amd64's architectural requirements on the
 942          * location of code and data in the kernel's address space, request a
 943          * mapping that is above the kernel.
 944          *
 945          * Protections will be restricted once relocations are applied.
 946          */
 947 #ifdef __amd64__
 948         mapbase = KERNBASE;
 949 #else
 950         mapbase = VM_MIN_KERNEL_ADDRESS;
 951 #endif
 952         error = vm_map_find(kernel_map, ef->object, 0, &mapbase,
 953             round_page(mapsize), 0, VMFS_OPTIMAL_SPACE, VM_PROT_ALL,
 954             VM_PROT_ALL, 0);
 955         if (error != KERN_SUCCESS) {
 956                 vm_object_deallocate(ef->object);
 957                 ef->object = NULL;
 958                 error = ENOMEM;
 959                 goto out;
 960         }
 961
 962         /* Wire the pages */
 963         error = vm_map_wire(kernel_map, mapbase,
 964             mapbase + round_page(mapsize),
 965             VM_MAP_WIRE_SYSTEM|VM_MAP_WIRE_NOHOLES);
 966         if (error != KERN_SUCCESS) {
 967                 error = ENOMEM;
 968                 goto out;
 969         }
 970
 971         /* Inform the kld system about the situation */
 972         lf->address = ef->address = (caddr_t)mapbase;
 973         lf->size = mapsize;
 974
 975         /*
 976          * Now load code/data(progbits), zero bss(nobits), allocate space for
 977          * and load relocs
 978          */
 979         pb = 0;
 980         rl = 0;
 981         ra = 0;
 982         alignmask = 0;
 983         for (i = 0; i < hdr->e_shnum; i++) {
 984                 if (shdr[i].sh_size == 0)
 985                         continue;
 986                 switch (shdr[i].sh_type) {
 987                 case SHT_PROGBITS:
 988                 case SHT_NOBITS:
 989 #ifdef __amd64__
 990                 case SHT_X86_64_UNWIND:
 991 #endif
 992                 case SHT_INIT_ARRAY:
 993                 case SHT_FINI_ARRAY:
 994                         if ((shdr[i].sh_flags & SHF_ALLOC) == 0)
 995                                 break;
 996                         alignmask = shdr[i].sh_addralign - 1;
 997                         mapbase += alignmask;
 998                         mapbase &= ~alignmask;
 999                         if (ef->shstrtab != NULL && shdr[i].sh_name != 0) {
1000                                 ef->progtab[pb].name =
1001                                     ef->shstrtab + shdr[i].sh_name;
1002                                 if (!strcmp(ef->progtab[pb].name, ".ctors") ||
1003                                     shdr[i].sh_type == SHT_INIT_ARRAY) {
1004                                         if (lf->ctors_addr != 0) {
1005                                                 printf(
1006                                     "%s: multiple ctor sections in %s\n",
1007                                                     __func__, filename);
1008                                         } else {
1009                                                 lf->ctors_addr =
1010                                                     (caddr_t)mapbase;
1011                                                 lf->ctors_size =
1012                                                     shdr[i].sh_size;
1013                                         }
1014                                 }
1015                         } else if (shdr[i].sh_type == SHT_PROGBITS)
1016                                 ef->progtab[pb].name = "<<PROGBITS>>";
1017 #ifdef __amd64__
1018                         else if (shdr[i].sh_type == SHT_X86_64_UNWIND)
1019                                 ef->progtab[pb].name = "<<UNWIND>>";
1020 #endif
1021                         else
1022                                 ef->progtab[pb].name = "<<NOBITS>>";
1023                         if (ef->progtab[pb].name != NULL &&
1024                             !strcmp(ef->progtab[pb].name, DPCPU_SETNAME)) {
1025                                 ef->progtab[pb].addr =
1026                                     dpcpu_alloc(shdr[i].sh_size);
1027                                 if (ef->progtab[pb].addr == NULL) {
1028                                         printf("%s: pcpu module space is out "
1029                                             "of space; cannot allocate %#jx "
1030                                             "for %s\n", __func__,
1031                                             (uintmax_t)shdr[i].sh_size,
1032                                             filename);
1033                                 }
1034                         }
1035 #ifdef VIMAGE
1036                         else if (ef->progtab[pb].name != NULL &&
1037                             !strcmp(ef->progtab[pb].name, VNET_SETNAME)) {
1038                                 ef->progtab[pb].addr =
1039                                     vnet_data_alloc(shdr[i].sh_size);
1040                                 if (ef->progtab[pb].addr == NULL) {
1041                                         printf("%s: vnet module space is out "
1042                                             "of space; cannot allocate %#jx "
1043                                             "for %s\n", __func__,
1044                                             (uintmax_t)shdr[i].sh_size,
1045                                             filename);
1046                                 }
1047                         }
1048 #endif
1049                         else
1050                                 ef->progtab[pb].addr =
1051                                     (void *)(uintptr_t)mapbase;
1052                         if (ef->progtab[pb].addr == NULL) {
1053                                 error = ENOSPC;
1054                                 goto out;
1055                         }
1056                         ef->progtab[pb].size = shdr[i].sh_size;
1057                         ef->progtab[pb].flags = shdr[i].sh_flags;
1058                         ef->progtab[pb].sec = i;
1059                         if (shdr[i].sh_type == SHT_PROGBITS
1060 #ifdef __amd64__
1061                             || shdr[i].sh_type == SHT_X86_64_UNWIND
1062 #endif
1063                             ) {
1064                                 error = vn_rdwr(UIO_READ, nd->ni_vp,
1065                                     ef->progtab[pb].addr,
1066                                     shdr[i].sh_size, shdr[i].sh_offset,
1067                                     UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred,
1068                                     NOCRED, &resid, td);
1069                                 if (error)
1070                                         goto out;
1071                                 if (resid != 0){
1072                                         error = EINVAL;
1073                                         goto out;
1074                                 }
1075                                 /* Initialize the per-cpu or vnet area. */
1076                                 if (ef->progtab[pb].addr != (void *)mapbase &&
1077                                     !strcmp(ef->progtab[pb].name, DPCPU_SETNAME))
1078                                         dpcpu_copy(ef->progtab[pb].addr,
1079                                             shdr[i].sh_size);
1080 #ifdef VIMAGE
1081                                 else if (ef->progtab[pb].addr !=
1082                                     (void *)mapbase &&
1083                                     !strcmp(ef->progtab[pb].name, VNET_SETNAME))
1084                                         vnet_data_copy(ef->progtab[pb].addr,
1085                                             shdr[i].sh_size);
1086 #endif
1087                         } else
1088                                 bzero(ef->progtab[pb].addr, shdr[i].sh_size);
1089
1090                         /* Update all symbol values with the offset. */
1091                         for (j = 0; j < ef->ddbsymcnt; j++) {
1092                                 es = &ef->ddbsymtab[j];
1093                                 if (es->st_shndx != i)
1094                                         continue;
1095                                 es->st_value += (Elf_Addr)ef->progtab[pb].addr;
1096                         }
1097                         mapbase += shdr[i].sh_size;
1098                         pb++;
1099                         break;
1100                 case SHT_REL:
1101                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
1102                                 break;
1103                         ef->reltab[rl].rel = malloc(shdr[i].sh_size, M_LINKER,
1104                             M_WAITOK);
1105                         ef->reltab[rl].nrel = shdr[i].sh_size / sizeof(Elf_Rel);
1106                         ef->reltab[rl].sec = shdr[i].sh_info;
1107                         error = vn_rdwr(UIO_READ, nd->ni_vp,
1108                             (void *)ef->reltab[rl].rel,
1109                             shdr[i].sh_size, shdr[i].sh_offset,
1110                             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
1111                             &resid, td);
1112                         if (error)
1113                                 goto out;
1114                         if (resid != 0){
1115                                 error = EINVAL;
1116                                 goto out;
1117                         }
1118                         rl++;
1119                         break;
1120                 case SHT_RELA:
1121                         if ((shdr[shdr[i].sh_info].sh_flags & SHF_ALLOC) == 0)
1122                                 break;
1123                         ef->relatab[ra].rela = malloc(shdr[i].sh_size, M_LINKER,
1124                             M_WAITOK);
1125                         ef->relatab[ra].nrela =
1126                             shdr[i].sh_size / sizeof(Elf_Rela);
1127                         ef->relatab[ra].sec = shdr[i].sh_info;
1128                         error = vn_rdwr(UIO_READ, nd->ni_vp,
1129                             (void *)ef->relatab[ra].rela,
1130                             shdr[i].sh_size, shdr[i].sh_offset,
1131                             UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED,
1132                             &resid, td);
1133                         if (error)
1134                                 goto out;
1135                         if (resid != 0){
1136                                 error = EINVAL;
1137                                 goto out;
1138                         }
1139                         ra++;
1140                         break;
1141                 }
1142         }
1143         if (pb != ef->nprogtab) {
1144                 link_elf_error(filename, "lost progbits");
1145                 error = ENOEXEC;
1146                 goto out;
1147         }
1148         if (rl != ef->nreltab) {
1149                 link_elf_error(filename, "lost reltab");
1150                 error = ENOEXEC;
1151                 goto out;
1152         }
1153         if (ra != ef->nrelatab) {
1154                 link_elf_error(filename, "lost relatab");
1155                 error = ENOEXEC;
1156                 goto out;
1157         }
1158         if (mapbase != (vm_offset_t)ef->address + mapsize) {
1159                 printf(
1160                     "%s: mapbase 0x%lx != address %p + mapsize 0x%lx (0x%lx)\n",
1161                     filename != NULL ? filename : "<none>",
1162                     (u_long)mapbase, ef->address, (u_long)mapsize,
1163                     (u_long)(vm_offset_t)ef->address + mapsize);
1164                 error = ENOMEM;
1165                 goto out;
1166         }
1167
1168         /* Local intra-module relocations */
1169         error = link_elf_reloc_local(lf, false);
1170         if (error != 0)
1171                 goto out;
1172
1173         /* Pull in dependencies */
1174         VOP_UNLOCK(nd->ni_vp);
1175         error = linker_load_dependencies(lf);
1176         vn_lock(nd->ni_vp, LK_EXCLUSIVE | LK_RETRY);
1177         if (error)
1178                 goto out;
1179
1180         /* External relocations */
1181         error = relocate_file(ef);
1182         if (error)
1183                 goto out;
1184
1185         /* Notify MD code that a module is being loaded. */
1186         error = elf_cpu_load_file(lf);
1187         if (error)
1188                 goto out;
1189
1190 #if defined(__i386__) || defined(__amd64__)
1191         /* Now ifuncs. */
1192         error = link_elf_reloc_local(lf, true);
1193         if (error != 0)
1194                 goto out;
1195 #endif
1196
1197         link_elf_protect(ef);
1198         link_elf_invoke_ctors(lf->ctors_addr, lf->ctors_size);
1199         *result = lf;
1200
1201 out:
1202         VOP_UNLOCK(nd->ni_vp);
1203         vn_close(nd->ni_vp, FREAD, td->td_ucred, td);
1204         free(nd, M_TEMP);
1205         if (error && lf)
1206                 linker_file_unload(lf, LINKER_UNLOAD_FORCE);
1207         free(hdr, M_LINKER);
1208
1209         return error;
1210 }
1211
1212 static void
1213 link_elf_unload_file(linker_file_t file)
1214 {
1215         elf_file_t ef = (elf_file_t) file;
1216         u_int i;
1217
1218         /* Notify MD code that a module is being unloaded. */
1219         elf_cpu_unload_file(file);
1220
1221         if (ef->progtab) {
1222                 for (i = 0; i < ef->nprogtab; i++) {
1223                         if (ef->progtab[i].size == 0)
1224                                 continue;
1225                         if (ef->progtab[i].name == NULL)
1226                                 continue;
1227                         if (!strcmp(ef->progtab[i].name, DPCPU_SETNAME))
1228                                 dpcpu_free(ef->progtab[i].addr,
1229                                     ef->progtab[i].size);
1230 #ifdef VIMAGE
1231                         else if (!strcmp(ef->progtab[i].name, VNET_SETNAME))
1232                                 vnet_data_free(ef->progtab[i].addr,
1233                                     ef->progtab[i].size);
1234 #endif
1235                 }
1236         }
1237         if (ef->preloaded) {
1238                 free(ef->reltab, M_LINKER);
1239                 free(ef->relatab, M_LINKER);
1240                 free(ef->progtab, M_LINKER);
1241                 free(ef->ctftab, M_LINKER);
1242                 free(ef->ctfoff, M_LINKER);
1243                 free(ef->typoff, M_LINKER);
1244                 if (file->pathname != NULL)
1245                         preload_delete_name(file->pathname);
1246                 return;
1247         }
1248
1249         for (i = 0; i < ef->nreltab; i++)
1250                 free(ef->reltab[i].rel, M_LINKER);
1251         for (i = 0; i < ef->nrelatab; i++)
1252                 free(ef->relatab[i].rela, M_LINKER);
1253         free(ef->reltab, M_LINKER);
1254         free(ef->relatab, M_LINKER);
1255         free(ef->progtab, M_LINKER);
1256
1257         if (ef->object != NULL)
1258                 vm_map_remove(kernel_map, (vm_offset_t)ef->address,
1259                     (vm_offset_t)ef->address + ptoa(ef->object->size));
1260         free(ef->e_shdr, M_LINKER);
1261         free(ef->ddbsymtab, M_LINKER);
1262         free(ef->ddbstrtab, M_LINKER);
1263         free(ef->shstrtab, M_LINKER);
1264         free(ef->ctftab, M_LINKER);
1265         free(ef->ctfoff, M_LINKER);
1266         free(ef->typoff, M_LINKER);
1267 }
1268
1269 static const char *
1270 symbol_name(elf_file_t ef, Elf_Size r_info)
1271 {
1272         const Elf_Sym *ref;
1273
1274         if (ELF_R_SYM(r_info)) {
1275                 ref = ef->ddbsymtab + ELF_R_SYM(r_info);
1276                 return ef->ddbstrtab + ref->st_name;
1277         } else
1278                 return NULL;
1279 }
1280
1281 static Elf_Addr
1282 findbase(elf_file_t ef, int sec)
1283 {
1284         int i;
1285         Elf_Addr base = 0;
1286
1287         for (i = 0; i < ef->nprogtab; i++) {
1288                 if (sec == ef->progtab[i].sec) {
1289                         base = (Elf_Addr)ef->progtab[i].addr;
1290                         break;
1291                 }
1292         }
1293         return base;
1294 }
1295
1296 static int
1297 relocate_file(elf_file_t ef)
1298 {
1299         const Elf_Rel *rellim;
1300         const Elf_Rel *rel;
1301         const Elf_Rela *relalim;
1302         const Elf_Rela *rela;
1303         const char *symname;
1304         const Elf_Sym *sym;
1305         int i;
1306         Elf_Size symidx;
1307         Elf_Addr base;
1308
1309         /* Perform relocations without addend if there are any: */
1310         for (i = 0; i < ef->nreltab; i++) {
1311                 rel = ef->reltab[i].rel;
1312                 if (rel == NULL) {
1313                         link_elf_error(ef->lf.filename, "lost a reltab!");
1314                         return (ENOEXEC);
1315                 }
1316                 rellim = rel + ef->reltab[i].nrel;
1317                 base = findbase(ef, ef->reltab[i].sec);
1318                 if (base == 0) {
1319                         link_elf_error(ef->lf.filename, "lost base for reltab");
1320                         return (ENOEXEC);
1321                 }
1322                 for ( ; rel < rellim; rel++) {
1323                         symidx = ELF_R_SYM(rel->r_info);
1324                         if (symidx >= ef->ddbsymcnt)
1325                                 continue;
1326                         sym = ef->ddbsymtab + symidx;
1327                         /* Local relocs are already done */
1328                         if (ELF_ST_BIND(sym->st_info) == STB_LOCAL)
1329                                 continue;
1330                         if (elf_reloc(&ef->lf, base, rel, ELF_RELOC_REL,
1331                             elf_obj_lookup)) {
1332                                 symname = symbol_name(ef, rel->r_info);
1333                                 printf("link_elf_obj: symbol %s undefined\n",
1334                                     symname);
1335                                 return (ENOENT);
1336                         }
1337                 }
1338         }
1339
1340         /* Perform relocations with addend if there are any: */
1341         for (i = 0; i < ef->nrelatab; i++) {
1342                 rela = ef->relatab[i].rela;
1343                 if (rela == NULL) {
1344                         link_elf_error(ef->lf.filename, "lost a relatab!");
1345                         return (ENOEXEC);
1346                 }
1347                 relalim = rela + ef->relatab[i].nrela;
1348                 base = findbase(ef, ef->relatab[i].sec);
1349                 if (base == 0) {
1350                         link_elf_error(ef->lf.filename,
1351                             "lost base for relatab");
1352                         return (ENOEXEC);
1353                 }
1354                 for ( ; rela < relalim; rela++) {
1355                         symidx = ELF_R_SYM(rela->r_info);
1356                         if (symidx >= ef->ddbsymcnt)
1357                                 continue;
1358                         sym = ef->ddbsymtab + symidx;
1359                         /* Local relocs are already done */
1360                         if (ELF_ST_BIND(sym->st_info) == STB_LOCAL)
1361                                 continue;
1362                         if (elf_reloc(&ef->lf, base, rela, ELF_RELOC_RELA,
1363                             elf_obj_lookup)) {
1364                                 symname = symbol_name(ef, rela->r_info);
1365                                 printf("link_elf_obj: symbol %s undefined\n",
1366                                     symname);
1367                                 return (ENOENT);
1368                         }
1369                 }
1370         }
1371
1372         /*
1373          * Only clean SHN_FBSD_CACHED for successful return.  If we
1374          * modified symbol table for the object but found an
1375          * unresolved symbol, there is no reason to roll back.
1376          */
1377         elf_obj_cleanup_globals_cache(ef);
1378
1379         return (0);
1380 }
1381
1382 static int
1383 link_elf_lookup_symbol(linker_file_t lf, const char *name, c_linker_sym_t *sym)
1384 {
1385         elf_file_t ef = (elf_file_t) lf;
1386         const Elf_Sym *symp;
1387         const char *strp;
1388         int i;
1389
1390         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1391                 strp = ef->ddbstrtab + symp->st_name;
1392                 if (symp->st_shndx != SHN_UNDEF && strcmp(name, strp) == 0) {
1393                         *sym = (c_linker_sym_t) symp;
1394                         return 0;
1395                 }
1396         }
1397         return ENOENT;
1398 }
1399
1400 static int
1401 link_elf_symbol_values(linker_file_t lf, c_linker_sym_t sym,
1402     linker_symval_t *symval)
1403 {
1404         elf_file_t ef;
1405         const Elf_Sym *es;
1406         caddr_t val;
1407
1408         ef = (elf_file_t) lf;
1409         es = (const Elf_Sym*) sym;
1410         val = (caddr_t)es->st_value;
1411         if (es >= ef->ddbsymtab && es < (ef->ddbsymtab + ef->ddbsymcnt)) {
1412                 symval->name = ef->ddbstrtab + es->st_name;
1413                 val = (caddr_t)es->st_value;
1414                 if (ELF_ST_TYPE(es->st_info) == STT_GNU_IFUNC)
1415                         val = ((caddr_t (*)(void))val)();
1416                 symval->value = val;
1417                 symval->size = es->st_size;
1418                 return 0;
1419         }
1420         return ENOENT;
1421 }
1422
1423 static int
1424 link_elf_search_symbol(linker_file_t lf, caddr_t value,
1425     c_linker_sym_t *sym, long *diffp)
1426 {
1427         elf_file_t ef = (elf_file_t) lf;
1428         u_long off = (uintptr_t) (void *) value;
1429         u_long diff = off;
1430         u_long st_value;
1431         const Elf_Sym *es;
1432         const Elf_Sym *best = NULL;
1433         int i;
1434
1435         for (i = 0, es = ef->ddbsymtab; i < ef->ddbsymcnt; i++, es++) {
1436                 if (es->st_name == 0)
1437                         continue;
1438                 st_value = es->st_value;
1439                 if (off >= st_value) {
1440                         if (off - st_value < diff) {
1441                                 diff = off - st_value;
1442                                 best = es;
1443                                 if (diff == 0)
1444                                         break;
1445                         } else if (off - st_value == diff) {
1446                                 best = es;
1447                         }
1448                 }
1449         }
1450         if (best == NULL)
1451                 *diffp = off;
1452         else
1453                 *diffp = diff;
1454         *sym = (c_linker_sym_t) best;
1455
1456         return 0;
1457 }
1458
1459 /*
1460  * Look up a linker set on an ELF system.
1461  */
1462 static int
1463 link_elf_lookup_set(linker_file_t lf, const char *name,
1464     void ***startp, void ***stopp, int *countp)
1465 {
1466         elf_file_t ef = (elf_file_t)lf;
1467         void **start, **stop;
1468         int i, count;
1469
1470         /* Relative to section number */
1471         for (i = 0; i < ef->nprogtab; i++) {
1472                 if ((strncmp(ef->progtab[i].name, "set_", 4) == 0) &&
1473                     strcmp(ef->progtab[i].name + 4, name) == 0) {
1474                         start  = (void **)ef->progtab[i].addr;
1475                         stop = (void **)((char *)ef->progtab[i].addr +
1476                             ef->progtab[i].size);
1477                         count = stop - start;
1478                         if (startp)
1479                                 *startp = start;
1480                         if (stopp)
1481                                 *stopp = stop;
1482                         if (countp)
1483                                 *countp = count;
1484                         return (0);
1485                 }
1486         }
1487         return (ESRCH);
1488 }
1489
1490 static int
1491 link_elf_each_function_name(linker_file_t file,
1492     int (*callback)(const char *, void *), void *opaque)
1493 {
1494         elf_file_t ef = (elf_file_t)file;
1495         const Elf_Sym *symp;
1496         int i, error;
1497
1498         /* Exhaustive search */
1499         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1500                 if (symp->st_value != 0 &&
1501                     (ELF_ST_TYPE(symp->st_info) == STT_FUNC ||
1502                     ELF_ST_TYPE(symp->st_info) == STT_GNU_IFUNC)) {
1503                         error = callback(ef->ddbstrtab + symp->st_name, opaque);
1504                         if (error)
1505                                 return (error);
1506                 }
1507         }
1508         return (0);
1509 }
1510
1511 static int
1512 link_elf_each_function_nameval(linker_file_t file,
1513     linker_function_nameval_callback_t callback, void *opaque)
1514 {
1515         linker_symval_t symval;
1516         elf_file_t ef = (elf_file_t)file;
1517         const Elf_Sym* symp;
1518         int i, error;
1519
1520         /* Exhaustive search */
1521         for (i = 0, symp = ef->ddbsymtab; i < ef->ddbsymcnt; i++, symp++) {
1522                 if (symp->st_value != 0 &&
1523                     (ELF_ST_TYPE(symp->st_info) == STT_FUNC ||
1524                     ELF_ST_TYPE(symp->st_info) == STT_GNU_IFUNC)) {
1525                         error = link_elf_symbol_values(file,
1526                             (c_linker_sym_t)symp, &symval);
1527                         if (error)
1528                                 return (error);
1529                         error = callback(file, i, &symval, opaque);
1530                         if (error)
1531                                 return (error);
1532                 }
1533         }
1534         return (0);
1535 }
1536
1537 static void
1538 elf_obj_cleanup_globals_cache(elf_file_t ef)
1539 {
1540         Elf_Sym *sym;
1541         Elf_Size i;
1542
1543         for (i = 0; i < ef->ddbsymcnt; i++) {
1544                 sym = ef->ddbsymtab + i;
1545                 if (sym->st_shndx == SHN_FBSD_CACHED) {
1546                         sym->st_shndx = SHN_UNDEF;
1547                         sym->st_value = 0;
1548                 }
1549         }
1550 }
1551
1552 /*
1553  * Symbol lookup function that can be used when the symbol index is known (ie
1554  * in relocations). It uses the symbol index instead of doing a fully fledged
1555  * hash table based lookup when such is valid. For example for local symbols.
1556  * This is not only more efficient, it's also more correct. It's not always
1557  * the case that the symbol can be found through the hash table.
1558  */
1559 static int
1560 elf_obj_lookup(linker_file_t lf, Elf_Size symidx, int deps, Elf_Addr *res)
1561 {
1562         elf_file_t ef = (elf_file_t)lf;
1563         Elf_Sym *sym;
1564         const char *symbol;
1565         Elf_Addr res1;
1566
1567         /* Don't even try to lookup the symbol if the index is bogus. */
1568         if (symidx >= ef->ddbsymcnt) {
1569                 *res = 0;
1570                 return (EINVAL);
1571         }
1572
1573         sym = ef->ddbsymtab + symidx;
1574
1575         /* Quick answer if there is a definition included. */
1576         if (sym->st_shndx != SHN_UNDEF) {
1577                 res1 = (Elf_Addr)sym->st_value;
1578                 if (ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC)
1579                         res1 = ((Elf_Addr (*)(void))res1)();
1580                 *res = res1;
1581                 return (0);
1582         }
1583
1584         /* If we get here, then it is undefined and needs a lookup. */
1585         switch (ELF_ST_BIND(sym->st_info)) {
1586         case STB_LOCAL:
1587                 /* Local, but undefined? huh? */
1588                 *res = 0;
1589                 return (EINVAL);
1590
1591         case STB_GLOBAL:
1592         case STB_WEAK:
1593                 /* Relative to Data or Function name */
1594                 symbol = ef->ddbstrtab + sym->st_name;
1595
1596                 /* Force a lookup failure if the symbol name is bogus. */
1597                 if (*symbol == 0) {
1598                         *res = 0;
1599                         return (EINVAL);
1600                 }
1601                 res1 = (Elf_Addr)linker_file_lookup_symbol(lf, symbol, deps);
1602
1603                 /*
1604                  * Cache global lookups during module relocation. The failure
1605                  * case is particularly expensive for callers, who must scan
1606                  * through the entire globals table doing strcmp(). Cache to
1607                  * avoid doing such work repeatedly.
1608                  *
1609                  * After relocation is complete, undefined globals will be
1610                  * restored to SHN_UNDEF in elf_obj_cleanup_globals_cache(),
1611                  * above.
1612                  */
1613                 if (res1 != 0) {
1614                         sym->st_shndx = SHN_FBSD_CACHED;
1615                         sym->st_value = res1;
1616                         *res = res1;
1617                         return (0);
1618                 } else if (ELF_ST_BIND(sym->st_info) == STB_WEAK) {
1619                         sym->st_value = 0;
1620                         *res = 0;
1621                         return (0);
1622                 }
1623                 return (EINVAL);
1624
1625         default:
1626                 return (EINVAL);
1627         }
1628 }
1629
1630 static void
1631 link_elf_fix_link_set(elf_file_t ef)
1632 {
1633         static const char startn[] = "__start_";
1634         static const char stopn[] = "__stop_";
1635         Elf_Sym *sym;
1636         const char *sym_name, *linkset_name;
1637         Elf_Addr startp, stopp;
1638         Elf_Size symidx;
1639         int start, i;
1640
1641         startp = stopp = 0;
1642         for (symidx = 1 /* zero entry is special */;
1643                 symidx < ef->ddbsymcnt; symidx++) {
1644                 sym = ef->ddbsymtab + symidx;
1645                 if (sym->st_shndx != SHN_UNDEF)
1646                         continue;
1647
1648                 sym_name = ef->ddbstrtab + sym->st_name;
1649                 if (strncmp(sym_name, startn, sizeof(startn) - 1) == 0) {
1650                         start = 1;
1651                         linkset_name = sym_name + sizeof(startn) - 1;
1652                 }
1653                 else if (strncmp(sym_name, stopn, sizeof(stopn) - 1) == 0) {
1654                         start = 0;
1655                         linkset_name = sym_name + sizeof(stopn) - 1;
1656                 }
1657                 else
1658                         continue;
1659
1660                 for (i = 0; i < ef->nprogtab; i++) {
1661                         if (strcmp(ef->progtab[i].name, linkset_name) == 0) {
1662                                 startp = (Elf_Addr)ef->progtab[i].addr;
1663                                 stopp = (Elf_Addr)(startp + ef->progtab[i].size);
1664                                 break;
1665                         }
1666                 }
1667                 if (i == ef->nprogtab)
1668                         continue;
1669
1670                 sym->st_value = start ? startp : stopp;
1671                 sym->st_shndx = i;
1672         }
1673 }
1674
1675 static int
1676 link_elf_reloc_local(linker_file_t lf, bool ifuncs)
1677 {
1678         elf_file_t ef = (elf_file_t)lf;
1679         const Elf_Rel *rellim;
1680         const Elf_Rel *rel;
1681         const Elf_Rela *relalim;
1682         const Elf_Rela *rela;
1683         const Elf_Sym *sym;
1684         Elf_Addr base;
1685         int i;
1686         Elf_Size symidx;
1687
1688         link_elf_fix_link_set(ef);
1689
1690         /* Perform relocations without addend if there are any: */
1691         for (i = 0; i < ef->nreltab; i++) {
1692                 rel = ef->reltab[i].rel;
1693                 if (rel == NULL) {
1694                         link_elf_error(ef->lf.filename, "lost a reltab");
1695                         return (ENOEXEC);
1696                 }
1697                 rellim = rel + ef->reltab[i].nrel;
1698                 base = findbase(ef, ef->reltab[i].sec);
1699                 if (base == 0) {
1700                         link_elf_error(ef->lf.filename, "lost base for reltab");
1701                         return (ENOEXEC);
1702                 }
1703                 for ( ; rel < rellim; rel++) {
1704                         symidx = ELF_R_SYM(rel->r_info);
1705                         if (symidx >= ef->ddbsymcnt)
1706                                 continue;
1707                         sym = ef->ddbsymtab + symidx;
1708                         /* Only do local relocs */
1709                         if (ELF_ST_BIND(sym->st_info) != STB_LOCAL)
1710                                 continue;
1711                         if ((ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC ||
1712                             elf_is_ifunc_reloc(rel->r_info)) != ifuncs)
1713                                 continue;
1714                         if (elf_reloc_local(lf, base, rel, ELF_RELOC_REL,
1715                             elf_obj_lookup) != 0)
1716                                 return (ENOEXEC);
1717                 }
1718         }
1719
1720         /* Perform relocations with addend if there are any: */
1721         for (i = 0; i < ef->nrelatab; i++) {
1722                 rela = ef->relatab[i].rela;
1723                 if (rela == NULL) {
1724                         link_elf_error(ef->lf.filename, "lost a relatab!");
1725                         return (ENOEXEC);
1726                 }
1727                 relalim = rela + ef->relatab[i].nrela;
1728                 base = findbase(ef, ef->relatab[i].sec);
1729                 if (base == 0) {
1730                         link_elf_error(ef->lf.filename, "lost base for reltab");
1731                         return (ENOEXEC);
1732                 }
1733                 for ( ; rela < relalim; rela++) {
1734                         symidx = ELF_R_SYM(rela->r_info);
1735                         if (symidx >= ef->ddbsymcnt)
1736                                 continue;
1737                         sym = ef->ddbsymtab + symidx;
1738                         /* Only do local relocs */
1739                         if (ELF_ST_BIND(sym->st_info) != STB_LOCAL)
1740                                 continue;
1741                         if ((ELF_ST_TYPE(sym->st_info) == STT_GNU_IFUNC ||
1742                             elf_is_ifunc_reloc(rela->r_info)) != ifuncs)
1743                                 continue;
1744                         if (elf_reloc_local(lf, base, rela, ELF_RELOC_RELA,
1745                             elf_obj_lookup) != 0)
1746                                 return (ENOEXEC);
1747                 }
1748         }
1749         return (0);
1750 }
1751
1752 static long
1753 link_elf_symtab_get(linker_file_t lf, const Elf_Sym **symtab)
1754 {
1755     elf_file_t ef = (elf_file_t)lf;
1756
1757     *symtab = ef->ddbsymtab;
1758
1759     if (*symtab == NULL)
1760         return (0);
1761
1762     return (ef->ddbsymcnt);
1763 }
1764
1765 static long
1766 link_elf_strtab_get(linker_file_t lf, caddr_t *strtab)
1767 {
1768     elf_file_t ef = (elf_file_t)lf;
1769
1770     *strtab = ef->ddbstrtab;
1771
1772     if (*strtab == NULL)
1773         return (0);
1774
1775     return (ef->ddbstrcnt);
1776 }