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