The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1 | #include <stdio.h> |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 2 | #include <stdlib.h> |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 3 | #include <common.h> |
| 4 | #include <debug.h> |
| 5 | #include <libelf.h> |
| 6 | #include <libebl.h> |
| 7 | #ifdef ARM_SPECIFIC_HACKS |
| 8 | #include <libebl_arm.h> |
| 9 | #endif/*ARM_SPECIFIC_HACKS*/ |
| 10 | #include <elf.h> |
| 11 | #include <gelf.h> |
| 12 | #include <string.h> |
| 13 | #include <errno.h> |
| 14 | #include <string.h> |
| 15 | #include <sys/types.h> |
| 16 | #include <sys/stat.h> |
| 17 | #include <fcntl.h> |
| 18 | #include <unistd.h> |
| 19 | #include <hash.h> |
| 20 | #include <apriori.h> |
| 21 | #include <source.h> |
| 22 | #include <tweak.h> |
| 23 | #include <rangesort.h> |
| 24 | #include <prelink_info.h> |
| 25 | #include <prelinkmap.h> |
| 26 | #include <libgen.h> |
| 27 | |
| 28 | #ifndef ADJUST_ELF |
| 29 | #error "ADJUST_ELF must be defined!" |
| 30 | #endif |
| 31 | |
| 32 | /* When this macro is defined, apriori sets to ZERO those relocation values for |
| 33 | which it canot find the appropriate referent. |
| 34 | */ |
| 35 | #define PERMISSIVE |
| 36 | #define COPY_SECTION_DATA_BUFFER (0) |
| 37 | /* When this macro is set to a nonzero value, we replace calls to elf_strptr() |
| 38 | on the target ELF handle with code that extracts the strings directly from |
| 39 | the data buffers of that ELF handle. In this case, elf_strptr() does not |
| 40 | work as expected, as it tries to read the data buffer of the associated |
| 41 | string section directly from the file, and that buffer does not exist yet |
| 42 | in the file, since we haven't committed our changes yet. |
| 43 | */ |
| 44 | #define ELF_STRPTR_IS_BROKEN (1) |
| 45 | |
| 46 | /* When the macro below is defined, apriori does not mark for removal those |
| 47 | relocation sections that it fully handles. Instead, apriori just sets their |
| 48 | sizes to zero. This is more for debugging than of any actual use. |
| 49 | |
| 50 | This macro is meaningful only when ADJUST_ELF!=0 |
| 51 | */ |
| 52 | #define REMOVE_HANDLED_SECTIONS |
| 53 | |
| 54 | extern int verbose_flag; |
| 55 | |
| 56 | static source_t *sources = NULL; |
| 57 | |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 58 | /* Retouch data is a very concise representation of the resolved relocations. |
| 59 | This data is used to randomize the location of prelinked libraries at |
| 60 | update time, on the device. |
| 61 | */ |
| 62 | |
| 63 | // We will store retouch entries into this buffer, then dump them at the |
| 64 | // end of the .so file before setup_prelink_info(). |
Steve Block | e3425a0 | 2010-08-27 13:18:46 +0100 | [diff] [blame] | 65 | #define RETOUCH_MAX_SIZE 600000 |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 66 | static char *retouch_buf; |
| 67 | static unsigned int retouch_byte_cnt; |
| 68 | // Compression state. |
| 69 | static int32_t offs_prev; |
| 70 | static uint32_t cont_prev; |
| 71 | |
| 72 | #define false 0 |
| 73 | #define true 1 |
| 74 | |
| 75 | void retouch_init(void) { |
| 76 | offs_prev = 0; |
| 77 | cont_prev = 0; |
| 78 | retouch_byte_cnt = 0; |
| 79 | retouch_buf = malloc(RETOUCH_MAX_SIZE+12); |
| 80 | FAILIF(retouch_buf == NULL, |
| 81 | "Could not allocate %d bytes.\n", RETOUCH_MAX_SIZE+12); |
| 82 | } |
| 83 | |
| 84 | // |
| 85 | // We use three encoding schemes; this takes care of much of the redundancy |
| 86 | // inherent in lists of relocations: |
| 87 | // |
| 88 | // * two bytes, leading 1, 2b for d_offset ("s"), 13b for d_contents ("c") |
| 89 | // |
| 90 | // 76543210 76543210 |
| 91 | // 1ssccccc cccccccc |
| 92 | // |
| 93 | // * three bytes, leading 01, 2b for delta offset, 20b for delta contents |
| 94 | // |
| 95 | // 76543210 76543210 76543210 |
| 96 | // 01sscccc cccccccc cccccccc |
| 97 | // |
| 98 | // * eigth bytes, leading 00, 30b for offset, 32b for contents |
| 99 | // |
| 100 | // 76543210 76543210 76543210 76543210 |
| 101 | // 00ssssss ssssssss ssssssss ssssssss + 4 bytes contents |
| 102 | // |
| 103 | // NOTE 1: All deltas are w.r.t. the previous line in the list. |
| 104 | // NOTE 2: Two-bit ("ss") offsets mean: "00"=4, "01"=8, "10"=12, and "11"=16. |
| 105 | // NOTE 3: Delta contents are signed. To map back to a int32 we refill with 1s. |
| 106 | // NOTE 4: Special encoding for -1 offset. Extended back to 32b when decoded. |
| 107 | // |
| 108 | |
| 109 | void retouch_encode(int32_t offset, uint32_t contents) { |
| 110 | int64_t d_offs = offset-offs_prev; |
| 111 | int64_t d_cont = (int64_t)contents-(int64_t)cont_prev; |
| 112 | |
| 113 | uint8_t output[8]; |
| 114 | uint32_t output_size; |
| 115 | |
| 116 | if ((d_offs > 3) && |
| 117 | (d_offs % 4) == 0 && |
| 118 | (d_offs / 4) < 5 && |
| 119 | (d_cont < 4000) && |
| 120 | (d_cont > -4000)) { |
| 121 | // we can fit in 2 bytes |
| 122 | output[0] = |
| 123 | 0x80 | |
| 124 | (((d_offs/4)-1) << 5) | |
| 125 | (((uint64_t)d_cont & 0x1f00) >> 8); |
| 126 | output[1] = |
| 127 | ((uint64_t)d_cont & 0xff); |
| 128 | output_size = 2; |
| 129 | } else if ((d_offs > 3) && |
| 130 | (d_offs % 4) == 0 && |
| 131 | (d_offs / 4) < 5 && |
| 132 | (d_cont < 510000) && |
| 133 | (d_cont > -510000)) { |
| 134 | // fit in 3 bytes |
| 135 | output[0] = |
| 136 | 0x40 | |
| 137 | (((d_offs/4)-1) << 4) | |
| 138 | (((uint64_t)d_cont & 0xf0000) >> 16); |
| 139 | output[1] = |
| 140 | ((uint64_t)d_cont & 0xff00) >> 8; |
| 141 | output[2] = |
| 142 | ((uint64_t)d_cont & 0xff); |
| 143 | output_size = 3; |
| 144 | } else { |
| 145 | // fit in 8 bytes; we can't support files bigger than (1GB-1) |
| 146 | // with this encoding: no library is that big anyway.. |
| 147 | FAILIF(offset < -1 || offset > 0x3ffffffe, "Offset out of range.\n"); |
| 148 | output[0] = (offset & 0x3f000000) >> 24; |
| 149 | output[1] = (offset & 0xff0000) >> 16; |
| 150 | output[2] = (offset & 0xff00) >> 8; |
| 151 | output[3] = (offset & 0xff); |
| 152 | output[4] = (contents & 0xff000000) >> 24; |
| 153 | output[5] = (contents & 0xff0000) >> 16; |
| 154 | output[6] = (contents & 0xff00) >> 8; |
| 155 | output[7] = (contents & 0xff); |
| 156 | output_size = 8; |
| 157 | } |
| 158 | |
| 159 | // If this happens, the retouch buffer size can be bumped up. |
| 160 | // Currently, the largest case is libwebcore, at about 250K. |
| 161 | FAILIF((retouch_byte_cnt+output_size) > RETOUCH_MAX_SIZE, |
| 162 | "About to overflow retouch buffer.\n"); |
| 163 | |
| 164 | memcpy(retouch_buf+retouch_byte_cnt, output, output_size); |
| 165 | retouch_byte_cnt += output_size; |
| 166 | |
| 167 | offs_prev = offset; |
| 168 | cont_prev = contents; |
| 169 | } |
| 170 | |
| 171 | void retouch_dump(const char *fname, int elf_little, |
| 172 | unsigned int retouch_byte_cnt, char *retouch_buf) { |
| 173 | int fd = open(fname, O_WRONLY); |
| 174 | FAILIF(fd < 0, |
| 175 | "open(%s, O_WRONLY): %s (%d)\n" , |
| 176 | fname, strerror(errno), errno); |
| 177 | off_t sz = lseek(fd, 0, SEEK_END); |
| 178 | FAILIF(sz == (off_t)-1, |
| 179 | "lseek(%d, 0, SEEK_END): %s (%d)!\n", |
| 180 | fd, strerror(errno), errno); |
| 181 | |
| 182 | // The retouch blob ends with "RETOUCH XXXX", where XXXX is the 4-byte |
| 183 | // size of the retouch blob, in target endianness. |
| 184 | strncpy(retouch_buf+retouch_byte_cnt, "RETOUCH ", 8); |
| 185 | if (elf_little ^ is_host_little()) { |
| 186 | *(unsigned int *)(retouch_buf+retouch_byte_cnt+8) = |
| 187 | switch_endianness(retouch_byte_cnt); |
| 188 | } else { |
| 189 | *(unsigned int *)(retouch_buf+retouch_byte_cnt+8) = |
| 190 | retouch_byte_cnt; |
| 191 | } |
| 192 | |
| 193 | int num_written = write(fd, retouch_buf, retouch_byte_cnt+12); |
| 194 | FAILIF(num_written < 0, |
| 195 | "write(%d, &info, sizeof(info)): %s (%d)\n", |
| 196 | fd, strerror(errno), errno); |
| 197 | FAILIF((retouch_byte_cnt+12) != num_written, |
| 198 | "Could not write %d bytes as expected (wrote %d bytes instead)!\n", |
| 199 | retouch_byte_cnt, num_written); |
| 200 | FAILIF(close(fd) < 0, "close(%d): %s (%d)!\n", fd, strerror(errno), errno); |
| 201 | } |
| 202 | |
| 203 | /* End of retouch code. |
| 204 | */ |
| 205 | |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 206 | #if defined(DEBUG) && 0 |
| 207 | |
| 208 | static void print_shdr(source_t *source, Elf_Scn *scn) |
| 209 | { |
| 210 | GElf_Shdr shdr_mem, *shdr; |
| 211 | shdr = gelf_getshdr(scn, &shdr_mem); |
| 212 | Elf_Data *data = elf_getdata(scn, NULL); |
| 213 | INFO("\t%02d: data = %p, hdr = { offset = %8lld, size = %lld }, " |
| 214 | "data->d_buf = %p data->d_off = %lld, data->d_size = %d\n", |
| 215 | elf_ndxscn(scn), |
| 216 | data, |
| 217 | shdr->sh_offset, shdr->sh_size, |
| 218 | data->d_buf, data->d_off, data->d_size); |
| 219 | } |
| 220 | |
| 221 | static void print_shdr_idx(source_t *source, Elf *elf, int idx) |
| 222 | { |
| 223 | print_shdr(source, elf_getscn(elf, idx)); |
| 224 | } |
| 225 | |
| 226 | static void print_shdrs(source_t *source) { |
| 227 | Elf_Scn *scn = NULL; |
| 228 | INFO("section offset dump for new ELF\n"); |
| 229 | while ((scn = elf_nextscn (source->elf, scn)) != NULL) |
| 230 | print_shdr(source, scn); |
| 231 | |
| 232 | INFO("\nsection offset dump for original ELF\n"); |
| 233 | while ((scn = elf_nextscn (source->oldelf, scn)) != NULL) |
| 234 | print_shdr(source, scn); |
| 235 | |
| 236 | #if 0 |
| 237 | { |
| 238 | INFO("section offset dump for new ELF\n"); |
| 239 | int i = 0; |
| 240 | for (i = 0; i < source->shnum; i++) { |
| 241 | scn = elf_getscn(source->elf, i); |
| 242 | print_shdr(source, scn); |
| 243 | } |
| 244 | } |
| 245 | #endif |
| 246 | } |
| 247 | |
| 248 | #endif /* DEBUG */ |
| 249 | |
| 250 | static char * find_file(const char *libname, |
| 251 | char **lib_lookup_dirs, |
| 252 | int num_lib_lookup_dirs); |
| 253 | |
| 254 | static inline source_t* find_source(const char *name, |
| 255 | char **lib_lookup_dirs, |
| 256 | int num_lib_lookup_dirs) { |
| 257 | char *full = find_file(name, lib_lookup_dirs, num_lib_lookup_dirs); |
| 258 | if (full) { |
| 259 | source_t *trav = sources; |
| 260 | while (trav) { |
| 261 | if (!strcmp(trav->name, full)) |
| 262 | break; |
| 263 | trav = trav->next; |
| 264 | } |
| 265 | free(full); |
| 266 | return trav; |
| 267 | } |
| 268 | return NULL; |
| 269 | } |
| 270 | |
| 271 | static inline void add_to_sources(source_t *src) { |
| 272 | src->next = sources; |
| 273 | sources = src; |
| 274 | } |
| 275 | |
| 276 | static void handle_range_error(range_error_t err, |
| 277 | range_t *left, range_t *right) { |
| 278 | switch (err) { |
| 279 | case ERROR_CONTAINS: |
| 280 | ERROR("ERROR: section (%lld, %lld bytes) contains " |
| 281 | "section (%lld, %lld bytes)\n", |
| 282 | left->start, left->length, |
| 283 | right->start, right->length); |
| 284 | break; |
| 285 | case ERROR_OVERLAPS: |
| 286 | ERROR("ERROR: Section (%lld, %lld bytes) intersects " |
| 287 | "section (%lld, %lld bytes)\n", |
| 288 | left->start, left->length, |
| 289 | right->start, right->length); |
| 290 | break; |
| 291 | default: |
| 292 | ASSERT(!"Unknown range error code!"); |
| 293 | } |
| 294 | |
| 295 | FAILIF(1, "Range error.\n"); |
| 296 | } |
| 297 | |
| 298 | static void create_elf_sections(source_t *source, Elf *elf) |
| 299 | { |
| 300 | INFO("Creating new ELF sections.\n"); |
| 301 | ASSERT(elf == NULL || source->elf == NULL || source->elf == elf); |
| 302 | if (elf == NULL) { |
| 303 | ASSERT(source->elf != NULL); |
| 304 | elf = source->elf; |
| 305 | } |
| 306 | |
| 307 | int cnt = 1; |
| 308 | Elf_Scn *oldscn = NULL, *scn; |
| 309 | while ((oldscn = elf_nextscn (source->oldelf, oldscn)) != NULL) { |
| 310 | GElf_Shdr *oldshdr, oldshdr_mem; |
| 311 | |
| 312 | scn = elf_newscn(elf); |
| 313 | FAILIF_LIBELF(NULL == scn, elf_newscn); |
| 314 | |
| 315 | oldshdr = gelf_getshdr(oldscn, &oldshdr_mem); |
| 316 | FAILIF_LIBELF(NULL == oldshdr, gelf_getshdr); |
| 317 | /* Set the section header of the new section to be the same as the |
| 318 | headset of the old section by default. */ |
| 319 | gelf_update_shdr(scn, oldshdr); |
| 320 | |
| 321 | /* Copy the section data */ |
| 322 | Elf_Data *olddata = elf_getdata(oldscn, NULL); |
| 323 | FAILIF_LIBELF(NULL == olddata, elf_getdata); |
| 324 | |
| 325 | Elf_Data *data = elf_newdata(scn); |
| 326 | FAILIF_LIBELF(NULL == data, elf_newdata); |
| 327 | *data = *olddata; |
| 328 | #if COPY_SECTION_DATA_BUFFER |
| 329 | if (olddata->d_buf != NULL) { |
| 330 | data->d_buf = MALLOC(data->d_size); |
| 331 | memcpy(data->d_buf, olddata->d_buf, olddata->d_size); |
| 332 | } |
| 333 | #endif |
| 334 | |
| 335 | INFO("\tsection %02d: [%-30s] created\n", |
| 336 | cnt, |
| 337 | elf_strptr(source->oldelf, |
| 338 | source->shstrndx, |
| 339 | oldshdr->sh_name)); |
| 340 | |
| 341 | if (ADJUST_ELF) { |
| 342 | ASSERT(source->shdr_info != NULL); |
| 343 | /* Create a new section. */ |
| 344 | source->shdr_info[cnt].idx = cnt; |
| 345 | source->shdr_info[cnt].newscn = scn; |
| 346 | source->shdr_info[cnt].data = data; |
| 347 | source->shdr_info[cnt]. |
| 348 | use_old_shdr_for_relocation_calculations = 1; |
| 349 | INFO("\tsection [%s] (old offset %lld, old size %lld) " |
| 350 | "will have index %d (was %d).\n", |
| 351 | source->shdr_info[cnt].name, |
| 352 | source->shdr_info[cnt].old_shdr.sh_offset, |
| 353 | source->shdr_info[cnt].old_shdr.sh_size, |
| 354 | source->shdr_info[cnt].idx, |
| 355 | elf_ndxscn(source->shdr_info[cnt].scn)); |
| 356 | /* Same as the next assert */ |
| 357 | ASSERT(elf_ndxscn (source->shdr_info[cnt].newscn) == |
| 358 | source->shdr_info[cnt].idx); |
| 359 | } |
| 360 | |
| 361 | ASSERT(elf_ndxscn(scn) == (size_t)cnt); |
| 362 | cnt++; |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | /* This function sets up the shdr_info[] array of a source_t. We call it only |
| 367 | when ADJUST_ELF is non-zero (i.e., support for adjusting an ELF file for |
| 368 | changes in sizes and numbers of relocation sections is compiled in. Note |
| 369 | that setup_shdr_info() depends only on the information in source->oldelf, |
| 370 | not on source->elf. |
| 371 | */ |
| 372 | |
| 373 | static void setup_shdr_info(source_t *source) |
| 374 | { |
| 375 | if (ADJUST_ELF) |
| 376 | { |
| 377 | /* Allocate the section-header-info buffer. */ |
| 378 | INFO("Allocating section-header info structure (%d) bytes...\n", |
| 379 | source->shnum * sizeof (shdr_info_t)); |
| 380 | |
| 381 | source->shdr_info = (shdr_info_t *)CALLOC(source->shnum, |
| 382 | sizeof (shdr_info_t)); |
| 383 | |
| 384 | /* Mark the SHT_NULL section as handled. */ |
| 385 | source->shdr_info[0].idx = 2; |
| 386 | |
| 387 | int cnt = 1; |
| 388 | Elf_Scn *oldscn = NULL; |
| 389 | while ((oldscn = elf_nextscn (source->oldelf, oldscn)) != NULL) { |
| 390 | /* Copy the section header */ |
| 391 | ASSERT(elf_ndxscn(oldscn) == (size_t)cnt); |
| 392 | |
| 393 | /* Initialized the corresponding shdr_info entry */ |
| 394 | { |
| 395 | /* Mark the section with a non-zero index. Later, when we |
| 396 | decide to drop a section, we will set its idx to zero, and |
| 397 | assign section numbers to the remaining sections. |
| 398 | */ |
| 399 | source->shdr_info[cnt].idx = 1; |
| 400 | |
| 401 | source->shdr_info[cnt].scn = oldscn; |
| 402 | |
| 403 | /* NOTE: Here we pupulate the section-headset struct with the |
| 404 | same values as the original section's. After the |
| 405 | first run of prelink(), we will update the sh_size |
| 406 | fields of those sections that need resizing. |
| 407 | */ |
| 408 | FAILIF_LIBELF(NULL == |
| 409 | gelf_getshdr(oldscn, |
| 410 | &source->shdr_info[cnt].shdr), |
| 411 | gelf_getshdr); |
| 412 | |
| 413 | /* Get the name of the section. */ |
| 414 | source->shdr_info[cnt].name = |
| 415 | elf_strptr (source->oldelf, source->shstrndx, |
| 416 | source->shdr_info[cnt].shdr.sh_name); |
| 417 | |
| 418 | INFO("\tname: %s\n", source->shdr_info[cnt].name); |
| 419 | FAILIF(source->shdr_info[cnt].name == NULL, |
| 420 | "Malformed file: section %d name is null\n", |
| 421 | cnt); |
| 422 | |
| 423 | /* Remember the shdr.sh_link value. We need to remember this |
| 424 | value for those sections that refer to other sections. For |
| 425 | example, we need to remember it for relocation-entry |
| 426 | sections, because if we modify the symbol table that a |
| 427 | relocation-entry section is relative to, then we need to |
| 428 | patch the relocation section. By the time we get to |
| 429 | deciding whether we need to patch the relocation section, we |
| 430 | will have overwritten its header's sh_link field with a new |
| 431 | value. |
| 432 | */ |
| 433 | source->shdr_info[cnt].old_shdr = source->shdr_info[cnt].shdr; |
| 434 | INFO("\t\toriginal sh_link: %08d\n", |
| 435 | source->shdr_info[cnt].old_shdr.sh_link); |
| 436 | INFO("\t\toriginal sh_addr: %lld\n", |
| 437 | source->shdr_info[cnt].old_shdr.sh_addr); |
| 438 | INFO("\t\toriginal sh_offset: %lld\n", |
| 439 | source->shdr_info[cnt].old_shdr.sh_offset); |
| 440 | INFO("\t\toriginal sh_size: %lld\n", |
| 441 | source->shdr_info[cnt].old_shdr.sh_size); |
| 442 | |
| 443 | FAILIF(source->shdr_info[cnt].shdr.sh_type == SHT_SYMTAB_SHNDX, |
| 444 | "Cannot handle sh_type SHT_SYMTAB_SHNDX!\n"); |
| 445 | FAILIF(source->shdr_info[cnt].shdr.sh_type == SHT_GROUP, |
| 446 | "Cannot handle sh_type SHT_GROUP!\n"); |
| 447 | FAILIF(source->shdr_info[cnt].shdr.sh_type == SHT_GNU_versym, |
| 448 | "Cannot handle sh_type SHT_GNU_versym!\n"); |
| 449 | } |
| 450 | |
| 451 | cnt++; |
| 452 | } /* for each section */ |
| 453 | } /* if (ADJUST_ELF) */ |
| 454 | } |
| 455 | |
| 456 | static Elf * init_elf(source_t *source, bool create_new_sections) |
| 457 | { |
| 458 | Elf *elf; |
| 459 | if (source->output != NULL) { |
| 460 | if (source->output_is_dir) { |
| 461 | source->output_is_dir++; |
| 462 | char *dir = source->output; |
| 463 | int dirlen = strlen(dir); |
| 464 | /* The main() function maintains a pointer to source->output; it |
| 465 | frees the buffer after apriori() returns. |
| 466 | */ |
| 467 | source->output = MALLOC(dirlen + |
| 468 | 1 + /* slash */ |
| 469 | strlen(source->name) + |
| 470 | 1); /* null terminator */ |
| 471 | strcpy(source->output, dir); |
| 472 | source->output[dirlen] = '/'; |
| 473 | strcpy(source->output + dirlen + 1, |
| 474 | basename(source->name)); |
| 475 | } |
| 476 | |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 477 | /* Save some of the info; needed for retouching (ASLR). */ |
| 478 | retouch_init(); |
| 479 | |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 480 | source->newelf_fd = open(source->output, |
| 481 | O_RDWR | O_CREAT, |
| 482 | 0666); |
| 483 | FAILIF(source->newelf_fd < 0, "open(%s): %s (%d)\n", |
| 484 | source->output, |
| 485 | strerror(errno), |
| 486 | errno); |
| 487 | elf = elf_begin(source->newelf_fd, ELF_C_WRITE, NULL); |
| 488 | FAILIF_LIBELF(elf == NULL, elf_begin); |
| 489 | } else { |
| 490 | elf = elf_clone(source->oldelf, ELF_C_EMPTY); |
| 491 | FAILIF_LIBELF(elf == NULL, elf_clone); |
| 492 | } |
| 493 | |
| 494 | GElf_Ehdr *oldehdr = gelf_getehdr(source->oldelf, &source->old_ehdr_mem); |
| 495 | FAILIF_LIBELF(NULL == oldehdr, gelf_getehdr); |
| 496 | |
| 497 | /* Create new ELF and program headers for the elf file */ |
| 498 | INFO("Creating empty ELF and program headers...\n"); |
| 499 | FAILIF_LIBELF(gelf_newehdr (elf, gelf_getclass (source->oldelf)) == 0, |
| 500 | gelf_newehdr); |
| 501 | FAILIF_LIBELF(oldehdr->e_type != ET_REL |
| 502 | && gelf_newphdr (elf, |
| 503 | oldehdr->e_phnum) == 0, |
| 504 | gelf_newphdr); |
| 505 | |
| 506 | /* Copy the elf header */ |
| 507 | INFO("Copying ELF header...\n"); |
| 508 | GElf_Ehdr *ehdr = gelf_getehdr(elf, &source->ehdr_mem); |
| 509 | FAILIF_LIBELF(NULL == ehdr, gelf_getehdr); |
| 510 | memcpy(ehdr, oldehdr, sizeof(GElf_Ehdr)); |
| 511 | FAILIF_LIBELF(!gelf_update_ehdr(elf, ehdr), gelf_update_ehdr); |
| 512 | |
| 513 | /* Copy out the old program header: notice that if the ELF file does not |
| 514 | have a program header, this loop won't execute. |
| 515 | */ |
| 516 | INFO("Copying ELF program header...\n"); |
| 517 | { |
| 518 | int cnt; |
| 519 | source->phdr_info = (GElf_Phdr *)CALLOC(ehdr->e_phnum, |
| 520 | sizeof(GElf_Phdr)); |
| 521 | for (cnt = 0; cnt < ehdr->e_phnum; ++cnt) { |
| 522 | INFO("\tRetrieving entry %d\n", cnt); |
| 523 | FAILIF_LIBELF(NULL == |
| 524 | gelf_getphdr(source->oldelf, cnt, |
| 525 | source->phdr_info + cnt), |
| 526 | gelf_getphdr); |
| 527 | FAILIF_LIBELF(gelf_update_phdr (elf, cnt, |
| 528 | source->phdr_info + cnt) == 0, |
| 529 | gelf_update_phdr); |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | /* Copy the sections and the section headers. */ |
| 534 | if (create_new_sections) |
| 535 | { |
| 536 | create_elf_sections(source, elf); |
| 537 | } |
| 538 | |
| 539 | /* The ELF library better follows our layout when this is not a |
| 540 | relocatable object file. */ |
| 541 | elf_flagelf (elf, ELF_C_SET, (ehdr->e_type != ET_REL ? ELF_F_LAYOUT : 0)); |
| 542 | |
| 543 | return elf; |
| 544 | } |
| 545 | |
| 546 | static shdr_info_t *lookup_shdr_info_by_new_section( |
| 547 | source_t *source, |
| 548 | const char *sname, |
| 549 | Elf_Scn *newscn) |
| 550 | { |
| 551 | if (source->shdr_info == NULL) return NULL; |
| 552 | int cnt; |
| 553 | for (cnt = 0; cnt < source->shnum; cnt++) { |
| 554 | if (source->shdr_info[cnt].newscn == newscn) { |
| 555 | INFO("\t\tnew section at %p matches shdr_info[%d], " |
| 556 | "section [%s]!\n", |
| 557 | newscn, |
| 558 | cnt, |
| 559 | source->shdr_info[cnt].name); |
| 560 | FAILIF(strcmp(sname, source->shdr_info[cnt].name), |
| 561 | "Matched section's name [%s] does not match " |
| 562 | "looked-up section's name [%s]!\n", |
| 563 | source->shdr_info[cnt].name, |
| 564 | sname); |
| 565 | return source->shdr_info + cnt; |
| 566 | } |
| 567 | } |
| 568 | return NULL; |
| 569 | } |
| 570 | |
| 571 | static bool do_init_source(source_t *source, unsigned base) |
| 572 | { |
| 573 | /* Find various sections. */ |
| 574 | size_t scnidx; |
| 575 | Elf_Scn *scn; |
| 576 | GElf_Shdr *shdr, shdr_mem; |
| 577 | source->sorted_sections = init_range_list(); |
| 578 | INFO("Processing [%s]'s sections...\n", source->name); |
| 579 | for (scnidx = 1; scnidx < (size_t)source->shnum; scnidx++) { |
| 580 | INFO("\tGetting section index %d...\n", scnidx); |
| 581 | scn = elf_getscn(source->elf, scnidx); |
| 582 | if (NULL == scn) { |
| 583 | /* If we get an error from elf_getscn(), it means that a section |
| 584 | at the requested index does not exist. This may happen when |
| 585 | we remove sections. Since we do not update source->shnum |
| 586 | (we can't, since we need to know the original number of sections |
| 587 | to know source->shdr_info[]'s length), we will attempt to |
| 588 | retrieve a section for an index that no longer exists in the |
| 589 | new ELF file. */ |
| 590 | INFO("\tThere is no section at index %d anymore, continuing.\n", |
| 591 | scnidx); |
| 592 | continue; |
| 593 | } |
| 594 | shdr = gelf_getshdr(scn, &shdr_mem); |
| 595 | FAILIF_LIBELF(NULL == shdr, gelf_getshdr); |
| 596 | |
| 597 | /* We haven't modified the shstrtab section, and so shdr->sh_name |
| 598 | has the same value as before. Thus we look up the name based |
| 599 | on the old ELF handle. We cannot use shstrndx on the new ELF |
| 600 | handle because the index of the shstrtab section may have |
| 601 | changed (and calling elf_getshstrndx() returns the same section |
| 602 | index, so libelf can't handle thise ither). |
| 603 | */ |
| 604 | const char *sname = |
| 605 | elf_strptr(source->oldelf, source->shstrndx, shdr->sh_name); |
| 606 | ASSERT(sname); |
| 607 | |
| 608 | INFO("\tAdding [%s] (%lld, %lld)...\n", |
| 609 | sname, |
| 610 | shdr->sh_addr, |
| 611 | shdr->sh_addr + shdr->sh_size); |
| 612 | if ((shdr->sh_flags & SHF_ALLOC) == SHF_ALLOC) { |
| 613 | add_unique_range_nosort(source->sorted_sections, |
| 614 | shdr->sh_addr, |
| 615 | shdr->sh_size, |
| 616 | scn, |
| 617 | handle_range_error, |
| 618 | NULL); /* no user-data destructor */ |
| 619 | } |
| 620 | |
| 621 | if (shdr->sh_type == SHT_DYNSYM) { |
| 622 | source->symtab.scn = scn; |
| 623 | source->symtab.data = elf_getdata(scn, NULL); |
| 624 | FAILIF_LIBELF(NULL == source->symtab.data, elf_getdata); |
| 625 | memcpy(&source->symtab.shdr, shdr, sizeof(GElf_Shdr)); |
| 626 | source->symtab.info = lookup_shdr_info_by_new_section( |
| 627 | source, sname, scn); |
| 628 | ASSERT(source->shdr_info == NULL || source->symtab.info != NULL); |
| 629 | |
| 630 | /* The sh_link field of the section header of the symbol table |
| 631 | contains the index of the associated strings table. */ |
| 632 | source->strtab.scn = elf_getscn(source->elf, |
| 633 | source->symtab.shdr.sh_link); |
| 634 | FAILIF_LIBELF(NULL == source->strtab.scn, elf_getscn); |
| 635 | FAILIF_LIBELF(NULL == gelf_getshdr(source->strtab.scn, |
| 636 | &source->strtab.shdr), |
| 637 | gelf_getshdr); |
| 638 | source->strtab.data = elf_getdata(source->strtab.scn, NULL); |
| 639 | FAILIF_LIBELF(NULL == source->strtab.data, elf_getdata); |
| 640 | source->strtab.info = lookup_shdr_info_by_new_section( |
| 641 | source, |
| 642 | elf_strptr(source->oldelf, source->shstrndx, |
| 643 | source->strtab.shdr.sh_name), |
| 644 | source->strtab.scn); |
| 645 | ASSERT(source->shdr_info == NULL || source->strtab.info != NULL); |
| 646 | } else if (shdr->sh_type == SHT_DYNAMIC) { |
| 647 | source->dynamic.scn = scn; |
| 648 | source->dynamic.data = elf_getdata(scn, NULL); |
| 649 | FAILIF_LIBELF(NULL == source->dynamic.data, elf_getdata); |
| 650 | memcpy(&source->dynamic.shdr, shdr, sizeof(GElf_Shdr)); |
| 651 | source->dynamic.info = lookup_shdr_info_by_new_section( |
| 652 | source, sname, scn); |
| 653 | ASSERT(source->shdr_info == NULL || source->dynamic.info != NULL); |
| 654 | } else if (shdr->sh_type == SHT_HASH) { |
| 655 | source->hash.scn = scn; |
| 656 | source->hash.data = elf_getdata(scn, NULL); |
| 657 | FAILIF_LIBELF(NULL == source->hash.data, elf_getdata); |
| 658 | memcpy(&source->hash.shdr, shdr, sizeof(GElf_Shdr)); |
| 659 | source->hash.info = lookup_shdr_info_by_new_section( |
| 660 | source, sname, scn); |
| 661 | ASSERT(source->shdr_info == NULL || source->hash.info != NULL); |
| 662 | } else if (shdr->sh_type == SHT_REL || shdr->sh_type == SHT_RELA) { |
| 663 | if (source->num_relocation_sections == |
| 664 | source->relocation_sections_size) { |
| 665 | source->relocation_sections_size += 5; |
| 666 | source->relocation_sections = |
| 667 | (section_info_t *)REALLOC(source->relocation_sections, |
| 668 | source->relocation_sections_size * |
| 669 | sizeof(section_info_t)); |
| 670 | } |
| 671 | section_info_t *reloc = |
| 672 | source->relocation_sections + source->num_relocation_sections; |
| 673 | reloc->scn = scn; |
| 674 | reloc->info = lookup_shdr_info_by_new_section(source, sname, scn); |
| 675 | ASSERT(source->shdr_info == NULL || reloc->info != NULL); |
| 676 | reloc->data = elf_getdata(scn, NULL); |
| 677 | FAILIF_LIBELF(NULL == reloc->data, elf_getdata); |
| 678 | memcpy(&reloc->shdr, shdr, sizeof(GElf_Shdr)); |
| 679 | source->num_relocation_sections++; |
| 680 | } else if (!strcmp(sname, ".bss")) { |
| 681 | source->bss.scn = scn; |
| 682 | source->bss.data = elf_getdata(scn, NULL); |
| 683 | source->bss.info = lookup_shdr_info_by_new_section( |
| 684 | source, sname, scn); |
| 685 | ASSERT(source->shdr_info == NULL || source->bss.info != NULL); |
| 686 | /* The BSS section occupies no space in the ELF file. */ |
| 687 | FAILIF_LIBELF(NULL == source->bss.data, elf_getdata) |
| 688 | FAILIF(NULL != source->bss.data->d_buf, |
| 689 | "Enexpected: section [%s] has data!", |
| 690 | sname); |
| 691 | memcpy(&source->bss.shdr, shdr, sizeof(GElf_Shdr)); |
| 692 | } |
| 693 | } |
| 694 | sort_ranges(source->sorted_sections); |
| 695 | |
| 696 | source->unfinished = |
| 697 | (unfinished_relocation_t *)CALLOC(source->num_relocation_sections, |
| 698 | sizeof(unfinished_relocation_t)); |
| 699 | |
| 700 | if (source->dynamic.scn == NULL) { |
| 701 | INFO("File [%s] does not have a dynamic section!\n", source->name); |
| 702 | /* If this is a static executable, we won't update anything. */ |
| 703 | source->dry_run = 1; |
| 704 | return false; |
| 705 | } |
| 706 | |
| 707 | FAILIF(source->symtab.scn == NULL, |
| 708 | "File [%s] does not have a dynamic symbol table!\n", |
| 709 | source->name); |
| 710 | FAILIF(source->hash.scn == NULL, |
| 711 | "File [%s] does not have a hash table!\n", |
| 712 | source->name); |
| 713 | FAILIF(source->hash.shdr.sh_link != elf_ndxscn(source->symtab.scn), |
| 714 | "Hash points to section %d, not to %d as expected!\n", |
| 715 | source->hash.shdr.sh_link, |
| 716 | elf_ndxscn(source->symtab.scn)); |
| 717 | |
| 718 | /* Now, find out how many symbols we have and allocate the array of |
| 719 | satisfied symbols. |
| 720 | |
| 721 | NOTE: We don't count the number of undefined symbols here; we will |
| 722 | iterate over the symbol table later, and count them then, when it is |
| 723 | more convenient. |
| 724 | */ |
| 725 | size_t symsize = gelf_fsize (source->elf, |
| 726 | ELF_T_SYM, |
| 727 | 1, source->elf_hdr.e_version); |
| 728 | ASSERT(symsize); |
| 729 | |
| 730 | source->num_syms = source->symtab.data->d_size / symsize; |
| 731 | source->base = (source->oldelf_hdr.e_type == ET_DYN) ? base : 0; |
| 732 | INFO("Relink base for [%s]: 0x%lx\n", source->name, source->base); |
| 733 | FAILIF(source->base == -1, |
| 734 | "Can't prelink [%s]: it's a shared library and you did not " |
| 735 | "provide a prelink address!\n", |
| 736 | source->name); |
| 737 | #ifdef SUPPORT_ANDROID_PRELINK_TAGS |
| 738 | FAILIF(source->prelinked && source->base != source->prelink_base, |
| 739 | "ERROR: file [%s] has already been prelinked for 0x%08lx. " |
| 740 | "Cannot change to 0x%08lx!\n", |
| 741 | source->name, |
| 742 | source->prelink_base, |
| 743 | source->base); |
| 744 | #endif/*SUPPORT_ANDROID_PRELINK_TAGS*/ |
| 745 | |
| 746 | return true; |
| 747 | } |
| 748 | |
| 749 | static source_t* init_source(const char *full_path, |
| 750 | const char *output, int is_file, |
| 751 | int base, int dry_run) |
| 752 | { |
| 753 | source_t *source = (source_t *)CALLOC(1, sizeof(source_t)); |
| 754 | |
| 755 | ASSERT(full_path); |
| 756 | source->name = full_path; |
| 757 | source->output = output; |
| 758 | source->output_is_dir = !is_file; |
| 759 | |
| 760 | source->newelf_fd = -1; |
| 761 | source->elf_fd = -1; |
| 762 | INFO("Opening %s...\n", full_path); |
| 763 | source->elf_fd = |
| 764 | open(full_path, ((dry_run || output != NULL) ? O_RDONLY : O_RDWR)); |
| 765 | FAILIF(source->elf_fd < 0, "open(%s): %s (%d)\n", |
| 766 | full_path, |
| 767 | strerror(errno), |
| 768 | errno); |
| 769 | |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 770 | FAILIF(fstat(source->elf_fd, &source->elf_file_info) < 0, |
| 771 | "fstat(%s(fd %d)): %s (%d)\n", |
| 772 | source->name, |
| 773 | source->elf_fd, |
| 774 | strerror(errno), |
| 775 | errno); |
| 776 | INFO("File [%s]'s size is %lld bytes!\n", |
| 777 | source->name, |
| 778 | source->elf_file_info.st_size); |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 779 | |
| 780 | INFO("Calling elf_begin(%s)...\n", full_path); |
| 781 | |
| 782 | source->oldelf = |
| 783 | elf_begin(source->elf_fd, |
| 784 | (dry_run || output != NULL) ? ELF_C_READ : ELF_C_RDWR, |
| 785 | NULL); |
| 786 | FAILIF_LIBELF(source->oldelf == NULL, elf_begin); |
| 787 | |
| 788 | /* libelf can recognize COFF and A.OUT formats, but we handle only ELF. */ |
| 789 | if(elf_kind(source->oldelf) != ELF_K_ELF) { |
| 790 | ERROR("Input file %s is not in ELF format!\n", full_path); |
| 791 | return NULL; |
| 792 | } |
| 793 | |
| 794 | /* Make sure this is a shared library or an executable. */ |
| 795 | { |
| 796 | INFO("Making sure %s is a shared library or an executable...\n", |
| 797 | full_path); |
| 798 | FAILIF_LIBELF(0 == gelf_getehdr(source->oldelf, &source->oldelf_hdr), |
| 799 | gelf_getehdr); |
| 800 | FAILIF(source->oldelf_hdr.e_type != ET_DYN && |
| 801 | source->oldelf_hdr.e_type != ET_EXEC, |
| 802 | "%s must be a shared library (elf type is %d, expecting %d).\n", |
| 803 | full_path, |
| 804 | source->oldelf_hdr.e_type, |
| 805 | ET_DYN); |
| 806 | } |
| 807 | |
| 808 | #ifdef SUPPORT_ANDROID_PRELINK_TAGS |
| 809 | /* First, check to see if the file has been prelinked. */ |
| 810 | source->prelinked = |
| 811 | check_prelinked(source->name, |
| 812 | source->oldelf_hdr.e_ident[EI_DATA] == ELFDATA2LSB, |
| 813 | &source->prelink_base); |
| 814 | /* Note that in the INFO() below we need to use oldelf_hdr because we |
| 815 | haven't cloned the ELF file yet, and source->elf_hdr is not defined. */ |
| 816 | if (source->prelinked) { |
| 817 | PRINT("%s [%s] is already prelinked at 0x%08lx!\n", |
| 818 | (source->oldelf_hdr.e_type == ET_EXEC ? |
| 819 | "Executable" : "Shared library"), |
| 820 | source->name, |
| 821 | source->prelink_base); |
| 822 | /* Force a dry run when the file has already been prelinked */ |
| 823 | source->dry_run = dry_run = 1; |
| 824 | } |
| 825 | else { |
| 826 | INFO("%s [%s] is not prelinked!\n", |
| 827 | (source->oldelf_hdr.e_type == ET_EXEC ? |
| 828 | "Executable" : "Shared library"), |
| 829 | source->name); |
| 830 | source->dry_run = dry_run; |
| 831 | } |
| 832 | #endif/*SUPPORT_ANDROID_PRELINK_TAGS*/ |
| 833 | |
| 834 | /* Get the index of the section-header-strings-table section. */ |
| 835 | FAILIF_LIBELF(elf_getshstrndx (source->oldelf, &source->shstrndx) < 0, |
| 836 | elf_getshstrndx); |
| 837 | |
| 838 | FAILIF_LIBELF(elf_getshnum (source->oldelf, (size_t *)&source->shnum) < 0, |
| 839 | elf_getshnum); |
| 840 | |
| 841 | /* When we have a dry run, or when ADJUST_ELF is enabled, we use |
| 842 | source->oldelf for source->elf, because the former is mmapped privately, |
| 843 | so changes to it have no effect. With ADJUST_ELF, the first run of |
| 844 | prelink() is a dry run. We will reopen the elf file for write access |
| 845 | after that dry run, before we call adjust_elf. */ |
| 846 | |
| 847 | source->elf = (ADJUST_ELF || source->dry_run) ? |
| 848 | source->oldelf : init_elf(source, ADJUST_ELF == 0); |
| 849 | |
| 850 | FAILIF_LIBELF(0 == gelf_getehdr(source->elf, &source->elf_hdr), |
| 851 | gelf_getehdr); |
| 852 | #ifdef DEBUG |
| 853 | ASSERT(!memcmp(&source->oldelf_hdr, |
| 854 | &source->elf_hdr, |
| 855 | sizeof(source->elf_hdr))); |
| 856 | #endif |
| 857 | |
| 858 | /* Get the EBL handling. The -g option is currently the only reason |
| 859 | we need EBL so dont open the backend unless necessary. */ |
| 860 | source->ebl = ebl_openbackend (source->elf); |
| 861 | FAILIF_LIBELF(NULL == source->ebl, ebl_openbackend); |
| 862 | #ifdef ARM_SPECIFIC_HACKS |
| 863 | FAILIF_LIBELF(0 != arm_init(source->elf, source->elf_hdr.e_machine, |
| 864 | source->ebl, sizeof(Ebl)), |
| 865 | arm_init); |
| 866 | #endif/*ARM_SPECIFIC_HACKS*/ |
| 867 | |
| 868 | add_to_sources(source); |
| 869 | if (do_init_source(source, base) == false) return NULL; |
| 870 | return source; |
| 871 | } |
| 872 | |
| 873 | /* complements do_init_source() */ |
| 874 | static void do_destroy_source(source_t *source) |
| 875 | { |
| 876 | int cnt; |
| 877 | destroy_range_list(source->sorted_sections); |
| 878 | source->sorted_sections = NULL; |
| 879 | for (cnt = 0; cnt < source->num_relocation_sections; cnt++) { |
| 880 | FREEIF(source->unfinished[cnt].rels); |
| 881 | source->unfinished[cnt].rels = NULL; |
| 882 | source->unfinished[cnt].num_rels = 0; |
| 883 | source->unfinished[cnt].rels_size = 0; |
| 884 | } |
| 885 | if (source->jmprel.sections != NULL) { |
| 886 | destroy_range_list(source->jmprel.sections); |
| 887 | source->jmprel.sections = NULL; |
| 888 | } |
| 889 | if (source->rel.sections != NULL) { |
| 890 | destroy_range_list(source->rel.sections); |
| 891 | source->rel.sections = NULL; |
| 892 | } |
| 893 | FREE(source->unfinished); /* do_init_source() */ |
| 894 | source->unfinished = NULL; |
| 895 | FREE(source->relocation_sections); /* do_init_source() */ |
| 896 | source->relocation_sections = NULL; |
| 897 | source->num_relocation_sections = source->relocation_sections_size = 0; |
| 898 | } |
| 899 | |
| 900 | static void destroy_source(source_t *source) |
| 901 | { |
| 902 | /* Is this a little-endian ELF file? */ |
| 903 | if (source->oldelf != source->elf) { |
| 904 | /* If it's a dynamic executable, this must not be a dry run. */ |
| 905 | if (!source->dry_run && source->dynamic.scn != NULL) |
| 906 | { |
| 907 | FAILIF_LIBELF(elf_update(source->elf, ELF_C_WRITE) == -1, |
| 908 | elf_update); |
| 909 | } |
| 910 | FAILIF_LIBELF(elf_end(source->oldelf), elf_end); |
| 911 | } |
| 912 | ebl_closebackend(source->ebl); |
| 913 | FAILIF_LIBELF(elf_end(source->elf), elf_end); |
| 914 | FAILIF(close(source->elf_fd) < 0, "Could not close file %s: %s (%d)!\n", |
| 915 | source->name, strerror(errno), errno); |
| 916 | FAILIF((source->newelf_fd >= 0) && (close(source->newelf_fd) < 0), |
| 917 | "Could not close output file: %s (%d)!\n", strerror(errno), errno); |
| 918 | |
| 919 | #ifdef SUPPORT_ANDROID_PRELINK_TAGS |
| 920 | if (!source->dry_run) { |
| 921 | if (source->dynamic.scn != NULL && |
| 922 | source->elf_hdr.e_type != ET_EXEC) |
| 923 | { |
| 924 | /* For some reason, trying to write directly to source->elf_fd |
| 925 | causes a "bad file descriptor" error because of something libelf |
| 926 | does. We just close the file descriptor and open a new one in |
| 927 | function setup_prelink_info() below. */ |
| 928 | INFO("%s: setting up prelink tag at end of file.\n", |
| 929 | source->output ? source->output : source->name); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 930 | retouch_encode(-1, source->base); |
| 931 | retouch_dump(source->output ? source->output : source->name, |
| 932 | source->elf_hdr.e_ident[EI_DATA] == ELFDATA2LSB, |
| 933 | retouch_byte_cnt, |
| 934 | retouch_buf); |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 935 | setup_prelink_info(source->output ? source->output : source->name, |
| 936 | source->elf_hdr.e_ident[EI_DATA] == ELFDATA2LSB, |
| 937 | source->base); |
| 938 | } |
| 939 | else INFO("%s: executable, NOT setting up prelink tag.\n", |
| 940 | source->name); |
| 941 | } |
| 942 | #endif/*SUPPORT_ANDROID_PRELINK_TAGS*/ |
| 943 | |
| 944 | do_destroy_source(source); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 945 | if (retouch_buf != NULL) { free(retouch_buf); retouch_buf = NULL; } |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 946 | |
| 947 | if (source->shstrtab_data != NULL) |
| 948 | FREEIF(source->shstrtab_data->d_buf); /* adjust_elf */ |
| 949 | |
| 950 | FREE(source->lib_deps); /* list of library dependencies (process_file()) */ |
| 951 | FREEIF(source->shdr_info); /* setup_shdr_info() */ |
| 952 | FREEIF(source->phdr_info); /* init_elf() */ |
| 953 | FREE(source->name); /* assigned to by init_source() */ |
| 954 | /* If the output is a directory, in init_elf() we allocate a buffer where |
| 955 | we copy the directory, a slash, and the file name. Here we free that |
| 956 | buffer. |
| 957 | */ |
| 958 | if (source->output_is_dir > 1) { |
| 959 | FREE(source->output); |
| 960 | } |
| 961 | FREE(source); /* init_source() */ |
| 962 | } |
| 963 | |
| 964 | static void reinit_source(source_t *source) |
| 965 | { |
| 966 | do_destroy_source(source); |
| 967 | do_init_source(source, source->base); |
| 968 | |
| 969 | { |
| 970 | /* We've gathered all the DT_DYNAMIC entries; now we need to figure |
| 971 | out which relocation sections fit in which range as described by |
| 972 | the entries. Before we do so, however, we will populate the |
| 973 | jmprel and rel members of source, as well as their sizes. |
| 974 | */ |
| 975 | |
| 976 | size_t dynidx, numdyn; |
| 977 | GElf_Dyn *dyn, dyn_mem; |
| 978 | |
| 979 | numdyn = source->dynamic.shdr.sh_size / |
| 980 | source->dynamic.shdr.sh_entsize; |
| 981 | |
| 982 | source->rel.idx = source->rel.sz_idx = -1; |
| 983 | source->jmprel.idx = source->jmprel.sz_idx = -1; |
| 984 | for (dynidx = 0; dynidx < numdyn; dynidx++) { |
| 985 | dyn = gelf_getdyn (source->dynamic.data, |
| 986 | dynidx, |
| 987 | &dyn_mem); |
| 988 | FAILIF_LIBELF(NULL == dyn, gelf_getdyn); |
| 989 | switch (dyn->d_tag) |
| 990 | { |
| 991 | case DT_NEEDED: |
| 992 | break; |
| 993 | case DT_JMPREL: |
| 994 | INFO("reinit_source: DT_JMPREL is at index %d, 0x%08llx.\n", |
| 995 | dynidx, dyn->d_un.d_ptr); |
| 996 | source->jmprel.idx = dynidx; |
| 997 | source->jmprel.addr = dyn->d_un.d_ptr; |
| 998 | break; |
| 999 | case DT_PLTRELSZ: |
| 1000 | INFO("reinit_source: DT_PLTRELSZ is at index %d, 0x%08llx.\n", |
| 1001 | dynidx, dyn->d_un.d_val); |
| 1002 | source->jmprel.sz_idx = dynidx; |
| 1003 | source->jmprel.size = dyn->d_un.d_val; |
| 1004 | break; |
| 1005 | case DT_REL: |
| 1006 | INFO("reinit_source: DT_REL is at index %d, 0x%08llx.\n", |
| 1007 | dynidx, dyn->d_un.d_ptr); |
| 1008 | source->rel.idx = dynidx; |
| 1009 | source->rel.addr = dyn->d_un.d_ptr; |
| 1010 | break; |
| 1011 | case DT_RELSZ: |
| 1012 | INFO("reinit_source: DT_RELSZ is at index %d, 0x%08llx.\n", |
| 1013 | dynidx, dyn->d_un.d_val); |
| 1014 | source->rel.sz_idx = dynidx; |
| 1015 | source->rel.size = dyn->d_un.d_val; |
| 1016 | break; |
| 1017 | case DT_RELA: |
| 1018 | case DT_RELASZ: |
| 1019 | FAILIF(1, "Can't handle DT_RELA and DT_RELASZ entries!\n"); |
| 1020 | break; |
| 1021 | } /* switch */ |
| 1022 | } /* for each dynamic entry... */ |
| 1023 | } |
| 1024 | } |
| 1025 | |
| 1026 | static GElf_Sym *hash_lookup_global_or_weak_symbol(source_t *lib, |
| 1027 | const char *symname, |
| 1028 | GElf_Sym *lib_sym_mem) |
| 1029 | { |
| 1030 | int lib_symidx = hash_lookup(lib->elf, |
| 1031 | lib->hash.data, |
| 1032 | lib->symtab.data, |
| 1033 | lib->strtab.data, |
| 1034 | symname); |
| 1035 | |
| 1036 | GElf_Sym sym_mem; |
| 1037 | if (SHN_UNDEF != lib_symidx) { |
| 1038 | /* We found the symbol--now check to see if it is global |
| 1039 | or weak. If this is the case, then the symbol satisfies |
| 1040 | the dependency. */ |
| 1041 | GElf_Sym *lib_sym = gelf_getsymshndx(lib->symtab.data, |
| 1042 | NULL, |
| 1043 | lib_symidx, |
| 1044 | &sym_mem, |
| 1045 | NULL); |
| 1046 | FAILIF_LIBELF(NULL == lib_sym, gelf_getsymshndx); |
| 1047 | #if ELF_STRPTR_IS_BROKEN |
| 1048 | ASSERT(!strcmp( |
| 1049 | symname, |
| 1050 | ((char *)elf_getdata(elf_getscn(lib->elf, |
| 1051 | lib->symtab.shdr.sh_link), |
| 1052 | NULL)->d_buf) + |
| 1053 | lib_sym->st_name)); |
| 1054 | #else |
| 1055 | ASSERT(!strcmp( |
| 1056 | symname, |
| 1057 | elf_strptr(lib->elf, lib->symtab.shdr.sh_link, |
| 1058 | lib_sym->st_name))); |
| 1059 | #endif |
| 1060 | if (lib_sym->st_shndx != SHN_UNDEF && |
| 1061 | (GELF_ST_BIND(lib_sym->st_info) == STB_GLOBAL || |
| 1062 | GELF_ST_BIND(lib_sym->st_info) == STB_WEAK)) { |
| 1063 | memcpy(lib_sym_mem, &sym_mem, sizeof(GElf_Sym)); |
| 1064 | return lib_sym; |
| 1065 | } |
| 1066 | } |
| 1067 | |
| 1068 | return NULL; |
| 1069 | } |
| 1070 | |
| 1071 | static source_t *lookup_symbol_in_dependencies(source_t *source, |
| 1072 | const char *symname, |
| 1073 | GElf_Sym *found_sym) |
| 1074 | { |
| 1075 | source_t *sym_source = NULL; /* return value */ |
| 1076 | |
| 1077 | /* This is an undefined symbol. Go over the list of libraries |
| 1078 | and look it up. */ |
| 1079 | size_t libidx; |
| 1080 | int found = 0; |
| 1081 | source_t *last_found = NULL; |
| 1082 | for (libidx = 0; libidx < (size_t)source->num_lib_deps; libidx++) { |
| 1083 | source_t *lib = source->lib_deps[libidx]; |
| 1084 | if (hash_lookup_global_or_weak_symbol(lib, symname, found_sym) != NULL) |
| 1085 | { |
| 1086 | sym_source = lib; |
| 1087 | if (found) { |
| 1088 | if (found == 1) { |
| 1089 | found++; |
| 1090 | ERROR("ERROR: multiple definitions found for [%s:%s]!\n", |
| 1091 | source->name, symname); |
| 1092 | ERROR("\tthis definition [%s]\n", lib->name); |
| 1093 | } |
| 1094 | ERROR("\tprevious definition [%s]\n", last_found->name); |
| 1095 | } |
| 1096 | last_found = lib; |
| 1097 | if (!found) found = 1; |
| 1098 | } |
| 1099 | } |
| 1100 | |
| 1101 | #if ELF_STRPTR_IS_BROKEN |
| 1102 | ASSERT(!sym_source || |
| 1103 | !strcmp(symname, |
| 1104 | (char *)(elf_getdata(elf_getscn( |
| 1105 | sym_source->elf, |
| 1106 | sym_source->symtab.shdr.sh_link), |
| 1107 | NULL)->d_buf) + |
| 1108 | found_sym->st_name)); |
| 1109 | #else |
| 1110 | ASSERT(!sym_source || |
| 1111 | !strcmp(symname, |
| 1112 | elf_strptr(sym_source->elf, |
| 1113 | sym_source->symtab.shdr.sh_link, |
| 1114 | found_sym->st_name))); |
| 1115 | #endif |
| 1116 | |
| 1117 | return sym_source; |
| 1118 | } |
| 1119 | |
| 1120 | static int do_prelink(source_t *source, |
| 1121 | Elf_Data *reloc_scn_data, |
| 1122 | int reloc_scn_entry_size, |
| 1123 | unfinished_relocation_t *unfinished, |
| 1124 | int locals_only, |
| 1125 | bool dry_run, |
| 1126 | char **lib_lookup_dirs, int num_lib_lookup_dirs, |
| 1127 | char **default_libs, int num_default_libs, |
| 1128 | int *num_unfinished_relocs) |
| 1129 | { |
| 1130 | int num_relocations = 0; |
| 1131 | |
| 1132 | size_t num_rels; |
| 1133 | num_rels = reloc_scn_data->d_size / reloc_scn_entry_size; |
| 1134 | |
| 1135 | INFO("\tThere are %d relocations.\n", num_rels); |
| 1136 | |
| 1137 | int rel_idx; |
| 1138 | for (rel_idx = 0; rel_idx < (size_t)num_rels; rel_idx++) { |
| 1139 | GElf_Rel *rel, rel_mem; |
| 1140 | |
| 1141 | //INFO("\tHandling relocation %d/%d\n", rel_idx, num_rels); |
| 1142 | |
| 1143 | rel = gelf_getrel(reloc_scn_data, rel_idx, &rel_mem); |
| 1144 | FAILIF_LIBELF(rel == NULL, gelf_getrel); |
| 1145 | GElf_Sym *sym = NULL, sym_mem; |
| 1146 | unsigned sym_idx = GELF_R_SYM(rel->r_info); |
| 1147 | source_t *sym_source = NULL; |
| 1148 | /* found_sym points to found_sym_mem, when sym_source != NULL, and |
| 1149 | to sym, when the sybmol is locally defined. If the symbol is |
| 1150 | not locally defined and sym_source == NULL, then sym is not |
| 1151 | defined either. */ |
| 1152 | GElf_Sym *found_sym = NULL, found_sym_mem; |
| 1153 | const char *symname = NULL; |
| 1154 | int sym_is_local = 1; |
| 1155 | if (sym_idx) { |
| 1156 | sym = gelf_getsymshndx(source->symtab.data, |
| 1157 | NULL, |
| 1158 | sym_idx, |
| 1159 | &sym_mem, |
| 1160 | NULL); |
| 1161 | FAILIF_LIBELF(NULL == sym, gelf_getsymshndx); |
| 1162 | #if ELF_STRPTR_IS_BROKEN |
| 1163 | symname = |
| 1164 | ((char *)source->strtab.data->d_buf) + |
| 1165 | sym->st_name; |
| 1166 | #else |
| 1167 | symname = elf_strptr(source->elf, |
| 1168 | elf_ndxscn(source->strtab.scn), |
| 1169 | sym->st_name); |
| 1170 | #endif |
| 1171 | |
| 1172 | /* If the symbol is defined and is either not in the BSS |
| 1173 | section, or if it is in the BSS then the relocation is |
| 1174 | not a copy relocation, then the symbol's source is this |
| 1175 | library (i.e., it is locally-defined). Otherwise, the |
| 1176 | symbol is imported. |
| 1177 | */ |
| 1178 | |
| 1179 | sym_is_local = 0; |
| 1180 | if (sym->st_shndx != SHN_UNDEF && |
| 1181 | (source->bss.scn == NULL || |
| 1182 | sym->st_shndx != elf_ndxscn(source->bss.scn) || |
| 1183 | #ifdef ARM_SPECIFIC_HACKS |
| 1184 | GELF_R_TYPE(rel->r_info) != R_ARM_COPY |
| 1185 | #else |
| 1186 | 1 |
| 1187 | #endif |
| 1188 | )) |
| 1189 | { |
| 1190 | sym_is_local = 1; |
| 1191 | } |
| 1192 | |
| 1193 | if (sym_is_local) { |
| 1194 | INFO("\t\tSymbol [%s:%s] is defined locally.\n", |
| 1195 | source->name, |
| 1196 | symname); |
| 1197 | sym_source = source; |
| 1198 | found_sym = sym; |
| 1199 | } |
| 1200 | else if (!locals_only) { |
| 1201 | sym_source = lookup_symbol_in_dependencies(source, |
| 1202 | symname, |
| 1203 | &found_sym_mem); |
| 1204 | |
| 1205 | /* The symbol was not in the list of dependencies, which by |
| 1206 | itself is an error: it means either that the symbol does |
| 1207 | not exist anywhere, or that the library which has the symbol |
| 1208 | has not been listed as a dependency in this library or |
| 1209 | executable. It could also mean (for a library) that the |
| 1210 | symbol is defined in the executable that links agsinst it, |
| 1211 | which is obviously not a good thing. These are bad things, |
| 1212 | but they do happen, which is why we have the ability to |
| 1213 | provide a list of default dependencies, including |
| 1214 | executables. Here we check to see if the symbol has been |
| 1215 | defined in any of them. |
| 1216 | */ |
| 1217 | if (NULL == sym_source) { |
| 1218 | INFO("\t\tChecking default dependencies...\n"); |
| 1219 | int i; |
| 1220 | source_t *lib, *old_sym_source = NULL; |
| 1221 | int printed_initial_error = 0; |
| 1222 | for (i = 0; i < num_default_libs; i++) { |
| 1223 | INFO("\tChecking in [%s].\n", default_libs[i]); |
| 1224 | lib = find_source(default_libs[i], |
| 1225 | lib_lookup_dirs, |
| 1226 | num_lib_lookup_dirs); |
| 1227 | FAILIF(NULL == lib, |
| 1228 | "Can't find default library [%s]!\n", |
| 1229 | default_libs[i]); |
| 1230 | if (hash_lookup_global_or_weak_symbol(lib, |
| 1231 | symname, |
| 1232 | &found_sym_mem)) { |
| 1233 | found_sym = &found_sym_mem; |
| 1234 | sym_source = lib; |
| 1235 | #if ELF_STRPTR_IS_BROKEN |
| 1236 | ASSERT(!strcmp(symname, |
| 1237 | (char *)(elf_getdata( |
| 1238 | elf_getscn( |
| 1239 | sym_source->elf, |
| 1240 | sym_source->symtab. |
| 1241 | shdr.sh_link), |
| 1242 | NULL)->d_buf) + |
| 1243 | found_sym->st_name)); |
| 1244 | #else |
| 1245 | ASSERT(!strcmp(symname, |
| 1246 | elf_strptr(sym_source->elf, |
| 1247 | sym_source->symtab.shdr.sh_link, |
| 1248 | found_sym->st_name))); |
| 1249 | |
| 1250 | #endif |
| 1251 | INFO("\tFound symbol [%s] in [%s]!\n", |
| 1252 | symname, lib->name); |
| 1253 | if (old_sym_source) { |
| 1254 | if (printed_initial_error == 0) { |
| 1255 | printed_initial_error = 1; |
| 1256 | ERROR("Multiple definition of [%s]:\n" |
| 1257 | "\t[%s]\n", |
| 1258 | symname, |
| 1259 | old_sym_source->name); |
| 1260 | } |
| 1261 | ERROR("\t[%s]\n", sym_source->name); |
| 1262 | } |
| 1263 | old_sym_source = sym_source; |
| 1264 | } else { |
| 1265 | INFO("\tCould not find symbol [%s] in default " |
| 1266 | "lib [%s]!\n", symname, lib->name); |
| 1267 | } |
| 1268 | } |
| 1269 | if (sym_source) { |
| 1270 | ERROR("ERROR: Could not find [%s:%s] in dependent " |
| 1271 | "libraries (but found in default [%s])!\n", |
| 1272 | source->name, |
| 1273 | symname, |
| 1274 | sym_source->name); |
| 1275 | } |
| 1276 | } else { |
| 1277 | found_sym = &found_sym_mem; |
| 1278 | /* We found the symbol in a dependency library. */ |
| 1279 | INFO("\t\tSymbol [%s:%s, value %lld] is imported from [%s]\n", |
| 1280 | source->name, |
| 1281 | symname, |
| 1282 | found_sym->st_value, |
| 1283 | sym_source->name); |
| 1284 | } |
| 1285 | } /* if symbol is defined in this library... */ |
| 1286 | |
| 1287 | if (!locals_only) { |
| 1288 | /* If a symbol is weak and we haven't found it, then report |
| 1289 | an error. We really need to find a way to set its value |
| 1290 | to zero. The problem is that it needs to refer to some |
| 1291 | section. */ |
| 1292 | |
| 1293 | FAILIF(NULL == sym_source && |
| 1294 | GELF_ST_BIND(sym->st_info) == STB_WEAK, |
| 1295 | "Cannot handle weak symbols yet (%s:%s <- %s).\n", |
| 1296 | source->name, |
| 1297 | symname, |
| 1298 | sym_source->name); |
| 1299 | #ifdef PERMISSIVE |
| 1300 | if (GELF_ST_BIND(sym->st_info) != STB_WEAK && |
| 1301 | NULL == sym_source) { |
| 1302 | ERROR("ERROR: Can't find symbol [%s:%s] in dependent or " |
| 1303 | "default libraries!\n", source->name, symname); |
| 1304 | } |
| 1305 | #else |
| 1306 | FAILIF(GELF_ST_BIND(sym->st_info) != STB_WEAK && |
| 1307 | NULL == sym_source, |
| 1308 | "Can't find symbol [%s:%s] in dependent or default " |
| 1309 | "libraries!\n", |
| 1310 | source->name, |
| 1311 | symname); |
| 1312 | #endif |
| 1313 | } /* if (!locals_only) */ |
| 1314 | } |
| 1315 | #if 0 // too chatty |
| 1316 | else |
| 1317 | INFO("\t\tno symbol is associated with this relocation\n"); |
| 1318 | #endif |
| 1319 | |
| 1320 | |
| 1321 | // We prelink only local symbols when locals_only == 1. |
| 1322 | |
| 1323 | bool can_relocate = true; |
| 1324 | if (!sym_is_local && |
| 1325 | (symname[0] == 'd' && symname[1] == 'l' && symname[2] != '\0' && |
| 1326 | (!strcmp(symname + 2, "open") || |
| 1327 | !strcmp(symname + 2, "close") || |
| 1328 | !strcmp(symname + 2, "sym") || |
| 1329 | !strcmp(symname + 2, "error")))) { |
| 1330 | INFO("********* NOT RELOCATING LIBDL SYMBOL [%s]\n", symname); |
| 1331 | can_relocate = false; |
| 1332 | } |
| 1333 | |
| 1334 | if (can_relocate && (sym_is_local || !locals_only)) |
| 1335 | { |
| 1336 | GElf_Shdr shdr_mem; Elf_Scn *scn; Elf_Data *data; |
| 1337 | find_section(source, rel->r_offset, &scn, &shdr_mem, &data); |
| 1338 | unsigned *dest = |
| 1339 | (unsigned*)(((char *)data->d_buf) + |
| 1340 | (rel->r_offset - shdr_mem.sh_addr)); |
| 1341 | unsigned rel_type = GELF_R_TYPE(rel->r_info); |
| 1342 | char buf[64]; |
| 1343 | INFO("\t\t%-15s ", |
| 1344 | ebl_reloc_type_name(source->ebl, |
| 1345 | GELF_R_TYPE(rel->r_info), |
| 1346 | buf, |
| 1347 | sizeof(buf))); |
| 1348 | |
| 1349 | /* Section-name offsets do not change, so we use oldelf to get the |
| 1350 | strings. This makes a difference in the second pass of the |
| 1351 | perlinker, after the call to adjust_elf, because |
| 1352 | source->shstrndx no longer contains the index of the |
| 1353 | section-header-strings table. |
| 1354 | */ |
| 1355 | const char *sname = elf_strptr( |
| 1356 | source->oldelf, source->shstrndx, shdr_mem.sh_name); |
| 1357 | |
| 1358 | switch (rel_type) { |
| 1359 | case R_ARM_JUMP_SLOT: |
| 1360 | case R_ARM_GLOB_DAT: |
| 1361 | case R_ARM_ABS32: |
| 1362 | ASSERT(data->d_buf != NULL); |
| 1363 | ASSERT(data->d_size >= rel->r_offset - shdr_mem.sh_addr); |
| 1364 | #ifdef PERMISSIVE |
| 1365 | if (sym_source == NULL) { |
| 1366 | ERROR("ERROR: Permissive relocation " |
| 1367 | "[%-15s] [%s:%s]: [0x%llx] = ZERO\n", |
| 1368 | ebl_reloc_type_name(source->ebl, |
| 1369 | GELF_R_TYPE(rel->r_info), |
| 1370 | buf, |
| 1371 | sizeof(buf)), |
| 1372 | sname, |
| 1373 | symname, |
| 1374 | rel->r_offset); |
| 1375 | if (!dry_run) |
| 1376 | *dest = 0; |
| 1377 | } else |
| 1378 | #endif |
| 1379 | { |
| 1380 | ASSERT(sym_source); |
| 1381 | INFO("[%s:%s]: [0x%llx] = 0x%llx + 0x%lx\n", |
| 1382 | sname, |
| 1383 | symname, |
| 1384 | rel->r_offset, |
| 1385 | found_sym->st_value, |
| 1386 | sym_source->base); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1387 | if (!dry_run) { |
| 1388 | PRINT("WARNING: Relocation type not supported " |
| 1389 | "for retouching!"); |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1390 | *dest = found_sym->st_value + sym_source->base; |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1391 | } |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1392 | } |
| 1393 | num_relocations++; |
| 1394 | break; |
| 1395 | case R_ARM_RELATIVE: |
| 1396 | ASSERT(data->d_buf != NULL); |
| 1397 | ASSERT(data->d_size >= rel->r_offset - shdr_mem.sh_addr); |
| 1398 | FAILIF(sym != NULL, |
| 1399 | "Unsupported RELATIVE form (symbol != 0)...\n"); |
| 1400 | INFO("[%s:%s]: [0x%llx] = 0x%x + 0x%lx\n", |
| 1401 | sname, |
| 1402 | symname ?: "(symbol has no name)", |
| 1403 | rel->r_offset, *dest, source->base); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1404 | if (!dry_run) { |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1405 | *dest += source->base; |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1406 | |
| 1407 | /* Output an entry for the ASLR touch-up process. */ |
| 1408 | retouch_encode(rel->r_offset |
| 1409 | -shdr_mem.sh_addr |
| 1410 | +shdr_mem.sh_offset, |
| 1411 | *dest); |
| 1412 | } |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1413 | num_relocations++; |
| 1414 | break; |
| 1415 | case R_ARM_COPY: |
| 1416 | #ifdef PERMISSIVE |
| 1417 | if (sym_source == NULL) { |
| 1418 | ERROR("ERROR: Permissive relocation " |
| 1419 | "[%-15s] [%s:%s]: NOT PERFORMING\n", |
| 1420 | ebl_reloc_type_name(source->ebl, |
| 1421 | GELF_R_TYPE(rel->r_info), |
| 1422 | buf, |
| 1423 | sizeof(buf)), |
| 1424 | sname, |
| 1425 | symname); |
| 1426 | } else |
| 1427 | #endif |
| 1428 | { |
| 1429 | ASSERT(sym); |
| 1430 | ASSERT(sym_source); |
| 1431 | GElf_Shdr src_shdr_mem; |
| 1432 | Elf_Scn *src_scn; |
| 1433 | Elf_Data *src_data; |
| 1434 | find_section(sym_source, found_sym->st_value, |
| 1435 | &src_scn, |
| 1436 | &src_shdr_mem, |
| 1437 | &src_data); |
| 1438 | INFO("Found [%s:%s (%lld)] in section [%s] .\n", |
| 1439 | sym_source->name, |
| 1440 | symname, |
| 1441 | found_sym->st_value, |
| 1442 | #if ELF_STRPTR_IS_BROKEN |
| 1443 | (((char *)elf_getdata( |
| 1444 | elf_getscn(sym_source->elf, |
| 1445 | sym_source->shstrndx), |
| 1446 | NULL)->d_buf) + src_shdr_mem.sh_name) |
| 1447 | #else |
| 1448 | elf_strptr(sym_source->elf, |
| 1449 | sym_source->shstrndx, |
| 1450 | src_shdr_mem.sh_name) |
| 1451 | #endif |
| 1452 | ); |
| 1453 | |
| 1454 | unsigned *src = NULL; |
| 1455 | if (src_data->d_buf == NULL) |
| 1456 | { |
| 1457 | #ifdef PERMISSIVE |
| 1458 | if (sym_source->bss.scn == NULL || |
| 1459 | elf_ndxscn(src_scn) != |
| 1460 | elf_ndxscn(sym_source->bss.scn)) { |
| 1461 | ERROR("ERROR: Permissive relocation (NULL source " |
| 1462 | "not from .bss) [%-15s] [%s:%s]: " |
| 1463 | "NOT PERFORMING\n", |
| 1464 | ebl_reloc_type_name(source->ebl, |
| 1465 | GELF_R_TYPE(rel->r_info), |
| 1466 | buf, |
| 1467 | sizeof(buf)), |
| 1468 | sname, |
| 1469 | symname); |
| 1470 | } |
| 1471 | #endif |
| 1472 | } |
| 1473 | else { |
| 1474 | ASSERT(src_data->d_size >= |
| 1475 | found_sym->st_value - src_shdr_mem.sh_addr); |
| 1476 | src = (unsigned*)(((char *)src_data->d_buf) + |
| 1477 | (found_sym->st_value - |
| 1478 | src_shdr_mem.sh_addr)); |
| 1479 | } |
| 1480 | ASSERT(symname); |
| 1481 | INFO("[%s:%s]: [0x%llx] <- [0x%llx] size %lld\n", |
| 1482 | sname, |
| 1483 | symname, rel->r_offset, |
| 1484 | found_sym->st_value, |
| 1485 | found_sym->st_size); |
| 1486 | |
| 1487 | #ifdef PERMISSIVE |
| 1488 | if (src_data->d_buf != NULL || |
| 1489 | (sym_source->bss.scn != NULL && |
| 1490 | elf_ndxscn(src_scn) == |
| 1491 | elf_ndxscn(sym_source->bss.scn))) |
| 1492 | #endif/*PERMISSIVE*/ |
| 1493 | { |
| 1494 | if (data->d_buf == NULL) { |
| 1495 | INFO("Incomplete relocation [%-15s] of [%s:%s].\n", |
| 1496 | ebl_reloc_type_name(source->ebl, |
| 1497 | GELF_R_TYPE(rel->r_info), |
| 1498 | buf, |
| 1499 | sizeof(buf)), |
| 1500 | sname, |
| 1501 | symname); |
| 1502 | FAILIF(unfinished == NULL, |
| 1503 | "You passed unfinished as NULL expecting " |
| 1504 | "to handle all relocations, " |
| 1505 | "but at least one cannot be handled!\n"); |
| 1506 | if (unfinished->num_rels == unfinished->rels_size) { |
| 1507 | unfinished->rels_size += 10; |
| 1508 | unfinished->rels = (GElf_Rel *)REALLOC( |
| 1509 | unfinished->rels, |
| 1510 | unfinished->rels_size * |
| 1511 | sizeof(GElf_Rel)); |
| 1512 | } |
| 1513 | unfinished->rels[unfinished->num_rels++] = *rel; |
| 1514 | num_relocations--; |
| 1515 | (*num_unfinished_relocs)++; |
| 1516 | } |
| 1517 | else { |
| 1518 | if (src_data->d_buf != NULL) |
| 1519 | { |
| 1520 | ASSERT(data->d_buf != NULL); |
| 1521 | ASSERT(data->d_size >= rel->r_offset - |
| 1522 | shdr_mem.sh_addr); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1523 | if (!dry_run) { |
| 1524 | PRINT("WARNING: Relocation type not supported " |
| 1525 | "for retouching!"); |
| 1526 | memcpy(dest, src, found_sym->st_size); |
| 1527 | } |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1528 | } |
| 1529 | else { |
| 1530 | ASSERT(src == NULL); |
| 1531 | ASSERT(elf_ndxscn(src_scn) == |
| 1532 | elf_ndxscn(sym_source->bss.scn)); |
Hristo Bojinov | 96be720 | 2010-08-02 10:26:17 -0700 | [diff] [blame] | 1533 | if (!dry_run) { |
| 1534 | PRINT("WARNING: Relocation type not supported " |
| 1535 | "for retouching!"); |
| 1536 | memset(dest, 0, found_sym->st_size); |
| 1537 | } |
The Android Open Source Project | 88b6079 | 2009-03-03 19:28:42 -0800 | [diff] [blame] | 1538 | } |
| 1539 | } |
| 1540 | } |
| 1541 | num_relocations++; |
| 1542 | } |
| 1543 | break; |
| 1544 | default: |
| 1545 | FAILIF(1, "Unknown relocation type %d!\n", rel_type); |
| 1546 | } // switch |
| 1547 | } // relocate |
| 1548 | else { |
| 1549 | INFO("\t\tNot relocating symbol [%s]%s\n", |
| 1550 | symname, |
| 1551 | (can_relocate ? ", relocating only locals" : |
| 1552 | ", which is a libdl symbol")); |
| 1553 | FAILIF(unfinished == NULL, |
| 1554 | "You passed unfinished as NULL expecting to handle all " |
| 1555 | "relocations, but at least one cannot be handled!\n"); |
| 1556 | if (unfinished->num_rels == unfinished->rels_size) { |
| 1557 | unfinished->rels_size += 10; |
| 1558 | unfinished->rels = (GElf_Rel *)REALLOC( |
| 1559 | unfinished->rels, |
| 1560 | unfinished->rels_size * |
| 1561 | sizeof(GElf_Rel)); |
| 1562 | } |
| 1563 | unfinished->rels[unfinished->num_rels++] = *rel; |
| 1564 | (*num_unfinished_relocs)++; |
| 1565 | } |
| 1566 | } // for each relocation entry |
| 1567 | |
| 1568 | return num_relocations; |
| 1569 | } |
| 1570 | |
| 1571 | static int prelink(source_t *source, |
| 1572 | int locals_only, |
| 1573 | bool dry_run, |
| 1574 | char **lib_lookup_dirs, int num_lib_lookup_dirs, |
| 1575 | char **default_libs, int num_default_libs, |
| 1576 | int *num_unfinished_relocs) |
| 1577 | { |
| 1578 | INFO("Prelinking [%s] (number of relocation sections: %d)%s...\n", |
| 1579 | source->name, source->num_relocation_sections, |
| 1580 | (dry_run ? " (dry run)" : "")); |
| 1581 | int num_relocations = 0; |
| 1582 | int rel_scn_idx; |
| 1583 | for (rel_scn_idx = 0; rel_scn_idx < source->num_relocation_sections; |
| 1584 | rel_scn_idx++) |
| 1585 | { |
| 1586 | section_info_t *reloc_scn = source->relocation_sections + rel_scn_idx; |
| 1587 | unfinished_relocation_t *unfinished = source->unfinished + rel_scn_idx; |
| 1588 | |
| 1589 | /* We haven't modified the shstrtab section, and so shdr->sh_name has |
| 1590 | the same value as before. Thus we look up the name based on the old |
| 1591 | ELF handle. We cannot use shstrndx on the new ELF handle because |
| 1592 | the index of the shstrtab section may have changed (and calling |
| 1593 | elf_getshstrndx() returns the same section index, so libelf can't |
| 1594 | handle thise ither). |
| 1595 | |
| 1596 | If reloc_scn->info is available, we can assert that the |
| 1597 | section-name has not changed. If this assertion fails, |
| 1598 | then we cannot use the elf_strptr() trick below to get |
| 1599 | the section name. One solution would be to save it in |
| 1600 | the section_info_t structure. |
| 1601 | */ |
| 1602 | ASSERT(reloc_scn->info == NULL || |
| 1603 | reloc_scn->shdr.sh_name == reloc_scn->info->old_shdr.sh_name); |
| 1604 | const char *sname = |
| 1605 | elf_strptr(source->oldelf, |
| 1606 | source->shstrndx, |
| 1607 | reloc_scn->shdr.sh_name); |
| 1608 | ASSERT(sname != NULL); |
| 1609 | |
| 1610 | INFO("\n\tIterating relocation section [%s]...\n", sname); |
| 1611 | |
| 1612 | /* In general, the new size of the section differs from the original |
| 1613 | size of the section, because we can handle some of the relocations. |
| 1614 | This was communicated to adjust_elf, which modified the ELF file |
| 1615 | according to the new section sizes. Now, when prelink() does the |
| 1616 | actual work of prelinking, it needs to know the original size of the |
| 1617 | relocation section so that it can see all of the original relocation |
| 1618 | entries! |
| 1619 | */ |
| 1620 | size_t d_size = reloc_scn->data->d_size; |
| 1621 | if (reloc_scn->info != NULL && |
| 1622 | reloc_scn->data->d_size != reloc_scn->info->old_shdr.sh_size) |
| 1623 | { |
| 1624 | INFO("Setting size of section [%s] to from new size %d to old " |
| 1625 | "size %lld temporarily (so prelinker can see all " |
| 1626 | "relocations).\n", |
| 1627 | reloc_scn->info->name, |
| 1628 | d_size, |
| 1629 | reloc_scn->info->old_shdr.sh_size); |
| 1630 | reloc_scn->data->d_size = reloc_scn->info->old_shdr.sh_size; |
| 1631 | } |
| 1632 | |
| 1633 | num_relocations += |
| 1634 | do_prelink(source, |
| 1635 | reloc_scn->data, reloc_scn->shdr.sh_entsize, |
| 1636 | unfinished, |
| 1637 | locals_only, dry_run, |
| 1638 | lib_lookup_dirs, num_lib_lookup_dirs, |
| 1639 | default_libs, num_default_libs, |
| 1640 | num_unfinished_relocs); |
| 1641 | |
| 1642 | if (reloc_scn->data->d_size != d_size) |
| 1643 | { |
| 1644 | ASSERT(reloc_scn->info != NULL); |
| 1645 | INFO("Resetting size of section [%s] to %d\n", |
| 1646 | reloc_scn->info->name, |
| 1647 | d_size); |
| 1648 | reloc_scn->data->d_size = d_size; |
| 1649 | } |
| 1650 | } |
| 1651 | |
| 1652 | /* Now prelink those relocation sections which were fully handled, and |
| 1653 | therefore removed. They are not a part of the |
| 1654 | source->relocation_sections[] array anymore, but we can find them by |
| 1655 | scanning source->shdr_info[] and looking for sections with idx == 0. |
| 1656 | */ |
| 1657 | |
| 1658 | if (ADJUST_ELF && source->shdr_info != NULL) { |
| 1659 | /* Walk over the shdr_info[] array to see if we've removed any |
| 1660 | relocation sections. prelink() those sections as well. |
| 1661 | */ |
| 1662 | int i; |
| 1663 | for (i = 0; i < source->shnum; i++) { |
| 1664 | shdr_info_t *info = source->shdr_info + i; |
| 1665 | if (info->idx == 0 && |
| 1666 | (info->shdr.sh_type == SHT_REL || |
| 1667 | info->shdr.sh_type == SHT_RELA)) { |
| 1668 | |
| 1669 | Elf_Data *data = elf_getdata(info->scn, NULL); |
| 1670 | ASSERT(data->d_size == 0); |
| 1671 | data->d_size = info->old_shdr.sh_size; |
| 1672 | |
| 1673 | INFO("\n\tIterating relocation section [%s], which was " |
| 1674 | "discarded (size %d, entry size %lld).\n", |
| 1675 | info->name, |
| 1676 | data->d_size, |
| 1677 | info->old_shdr.sh_entsize); |
| 1678 | |
| 1679 | num_relocations += |
| 1680 | do_prelink(source, |
| 1681 | data, info->old_shdr.sh_entsize, |
| 1682 | NULL, /* the section was fully handled */ |
| 1683 | locals_only, dry_run, |
| 1684 | lib_lookup_dirs, num_lib_lookup_dirs, |
| 1685 | default_libs, num_default_libs, |
| 1686 | num_unfinished_relocs); |
| 1687 | |
| 1688 | data->d_size = 0; |
| 1689 | } |
| 1690 | } |
| 1691 | } |
| 1692 | return num_relocations; |
| 1693 | } |
| 1694 | |
| 1695 | static char * find_file(const char *libname, |
| 1696 | char **lib_lookup_dirs, |
| 1697 | int num_lib_lookup_dirs) { |
| 1698 | if (libname[0] == '/') { |
| 1699 | /* This is an absolute path name--just return it. */ |
| 1700 | /* INFO("ABSOLUTE PATH: [%s].\n", libname); */ |
| 1701 | return strdup(libname); |
| 1702 | } else { |
| 1703 | /* First try the working directory. */ |
| 1704 | int fd; |
| 1705 | if ((fd = open(libname, O_RDONLY)) > 0) { |
| 1706 | close(fd); |
| 1707 | /* INFO("FOUND IN CURRENT DIR: [%s].\n", libname); */ |
| 1708 | return strdup(libname); |
| 1709 | } else { |
| 1710 | /* Iterate over all library paths. For each path, append the file |
| 1711 | name and see if there is a file at that place. If that fails, |
| 1712 | bail out. */ |
| 1713 | |
| 1714 | char *name; |
| 1715 | while (num_lib_lookup_dirs--) { |
| 1716 | size_t lib_len = strlen(*lib_lookup_dirs); |
| 1717 | /* one extra character for the slash, and another for the |
| 1718 | terminating NULL. */ |
| 1719 | name = (char *)MALLOC(lib_len + strlen(libname) + 2); |
| 1720 | strcpy(name, *lib_lookup_dirs); |
| 1721 | name[lib_len] = '/'; |
| 1722 | strcpy(name + lib_len + 1, libname); |
| 1723 | if ((fd = open(name, O_RDONLY)) > 0) { |
| 1724 | close(fd); |
| 1725 | /* INFO("FOUND: [%s] in [%s].\n", libname, name); */ |
| 1726 | return name; |
| 1727 | } |
| 1728 | INFO("NOT FOUND: [%s] in [%s].\n", libname, name); |
| 1729 | free(name); |
| 1730 | } |
| 1731 | } |
| 1732 | } |
| 1733 | return NULL; |
| 1734 | } |
| 1735 | |
| 1736 | static void adjust_dynamic_segment_entry_size(source_t *source, |
| 1737 | dt_rel_info_t *dyn) |
| 1738 | { |
| 1739 | /* Update the size entry in the DT_DYNAMIC segment. */ |
| 1740 | GElf_Dyn *dyn_entry, dyn_entry_mem; |
| 1741 | dyn_entry = gelf_getdyn(source->dynamic.data, |
| 1742 | dyn->sz_idx, |
| 1743 | &dyn_entry_mem); |
| 1744 | FAILIF_LIBELF(NULL == dyn_entry, gelf_getdyn); |
| 1745 | /* If we are calling this function to adjust the size of the dynamic entry, |
| 1746 | then there should be some unfinished relocations remaining. If there |
| 1747 | are none, then we should remove the entry from the dynamic section |
| 1748 | altogether. |
| 1749 | */ |
| 1750 | ASSERT(dyn->num_unfinished_relocs); |
| 1751 | |
| 1752 | size_t relsize = gelf_fsize(source->elf, |
| 1753 | ELF_T_REL, |
| 1754 | 1, |
| 1755 | source->elf_hdr.e_version); |
| 1756 | |
| 1757 | if (unlikely(verbose_flag)) { |
| 1758 | char buf[64]; |
| 1759 | INFO("Updating entry %d: [%-10s], %08llx --> %08x\n", |
| 1760 | dyn->sz_idx, |
| 1761 | ebl_dynamic_tag_name (source->ebl, dyn_entry->d_tag, |
| 1762 | buf, sizeof (buf)), |
| 1763 | dyn_entry->d_un.d_val, |
| 1764 | dyn->num_unfinished_relocs * relsize); |
| 1765 | } |
| 1766 | |
| 1767 | dyn_entry->d_un.d_val = dyn->num_unfinished_relocs * relsize; |
| 1768 | |
| 1769 | FAILIF_LIBELF(!gelf_update_dyn(source->dynamic.data, |
| 1770 | dyn->sz_idx, |
| 1771 | dyn_entry), |
| 1772 | gelf_update_dyn); |
| 1773 | } |
| 1774 | |
| 1775 | static void adjust_dynamic_segment_entries(source_t *source) |
| 1776 | { |
| 1777 | /* This function many remove entries from the dynamic segment, but it won't |
| 1778 | resize the relevant section. It'll just fill the remainted with empty |
| 1779 | DT entries. |
| 1780 | |
| 1781 | FIXME: This is not guaranteed right now. If a dynamic segment does not |
| 1782 | end with null DT entries, I think this will break. |
| 1783 | */ |
| 1784 | FAILIF(source->rel.processed, |
| 1785 | "More than one section matches DT_REL entry in dynamic segment!\n"); |
| 1786 | FAILIF(source->jmprel.processed, |
| 1787 | "More than one section matches DT_JMPREL entry in " |
| 1788 | "dynamic segment!\n"); |
| 1789 | source->rel.processed = |
| 1790 | source->jmprel.processed = 1; |
| 1791 | |
| 1792 | if (source->rel.num_unfinished_relocs > 0) |
| 1793 | adjust_dynamic_segment_entry_size(source, &source->rel); |
| 1794 | |
| 1795 | if (source->jmprel.num_unfinished_relocs > 0) |
| 1796 | adjust_dynamic_segment_entry_size(source, &source->jmprel); |
| 1797 | |
| 1798 | /* If at least one of the entries is empty, then we need to remove it. We |
| 1799 | have already adjusted the size of the other. |
| 1800 | */ |
| 1801 | if (source->rel.num_unfinished_relocs == 0 || |
| 1802 | source->jmprel.num_unfinished_relocs == 0) |
| 1803 | { |
| 1804 | /* We need to delete the DT_REL/DT_RELSZ and DT_PLTREL/DT_PLTRELSZ |
| 1805 | entries from the dynamic segment. */ |
| 1806 | |
| 1807 | GElf_Dyn *dyn_entry, dyn_entry_mem; |
| 1808 | size_t dynidx, updateidx; |
| 1809 | |
| 1810 | size_t numdyn = |
| 1811 | source->dynamic.shdr.sh_size / |
| 1812 | source->dynamic.shdr.sh_entsize; |
| 1813 | |
| 1814 | for (updateidx = dynidx = 0; dynidx < numdyn; dynidx++) |
| 1815 | { |
| 1816 | dyn_entry = gelf_getdyn(source->dynamic.data, |
| 1817 | dynidx, |
| 1818 | &dyn_entry_mem); |
| 1819 | FAILIF_LIBELF(NULL == dyn_entry, gelf_getdyn); |
| 1820 | if ((source->rel.num_unfinished_relocs == 0 && |
| 1821 | (dynidx == source->rel.idx || |
| 1822 | dynidx == source->rel.sz_idx)) || |
| 1823 | (source->jmprel.num_unfinished_relocs == 0 && |
| 1824 | (dynidx == source->jmprel.idx || |
| 1825 | dynidx == source->jmprel.sz_idx))) |
| 1826 | { |
| 1827 | if (unlikely(verbose_flag)) { |
| 1828 | char buf[64]; |
| 1829 | INFO("\t(!)\tRemoving entry %02d: [%-10s], %08llx\n", |
| 1830 | dynidx, |
| 1831 | ebl_dynamic_tag_name (source->ebl, dyn_entry->d_tag, |
| 1832 | buf, sizeof (buf)), |
| 1833 | dyn_entry->d_un.d_val); |
| 1834 | } |
| 1835 | continue; |
| 1836 | } |
| 1837 | |
| 1838 | if (unlikely(verbose_flag)) { |
| 1839 | char buf[64]; |
| 1840 | INFO("\t\tKeeping entry %02d: [%-10s], %08llx\n", |
| 1841 | dynidx, |
| 1842 | ebl_dynamic_tag_name (source->ebl, dyn_entry->d_tag, |
| 1843 | buf, sizeof (buf)), |
| 1844 | dyn_entry->d_un.d_val); |
| 1845 | } |
| 1846 | |
| 1847 | gelf_update_dyn(source->dynamic.data, |
| 1848 | updateidx, |
| 1849 | &dyn_entry_mem); |
| 1850 | updateidx++; |
| 1851 | } |
| 1852 | } |
| 1853 | } /* adjust_dynamic_segment_entries */ |
| 1854 | |
| 1855 | static bool adjust_dynamic_segment_for(source_t *source, |
| 1856 | dt_rel_info_t *dyn, |
| 1857 | bool adjust_section_size_only) |
| 1858 | { |
| 1859 | bool dropped_sections = false; |
| 1860 | |
| 1861 | /* Go over the sections that belong to this dynamic range. */ |
| 1862 | dyn->num_unfinished_relocs = 0; |
| 1863 | if (dyn->sections) { |
| 1864 | int num_scns, idx; |
| 1865 | range_t *scns = get_sorted_ranges(dyn->sections, &num_scns); |
| 1866 | |
| 1867 | INFO("\tdynamic range %s:[%lld, %lld) contains %d sections.\n", |
| 1868 | source->name, |
| 1869 | dyn->addr, |
| 1870 | dyn->addr + dyn->size, |
| 1871 | num_scns); |
| 1872 | |
| 1873 | ASSERT(scns); |
| 1874 | int next_idx = 0, next_rel_off = 0; |
| 1875 | /* The total number of unfinished relocations for this dynamic |
| 1876 | * entry. */ |
| 1877 | section_info_t *next = (section_info_t *)scns[next_idx].user; |
| 1878 | section_info_t *first = next; |
| 1879 | ASSERT(first); |
| 1880 | for (idx = 0; idx < num_scns; idx++) { |
| 1881 | section_info_t *reloc_scn = (section_info_t *)scns[idx].user; |
| 1882 | size_t rel_scn_idx = reloc_scn - source->relocation_sections; |
| 1883 | ASSERT(rel_scn_idx < (size_t)source->num_relocation_sections); |
| 1884 | unfinished_relocation_t *unfinished = |
| 1885 | &source->unfinished[rel_scn_idx]; |
| 1886 | int unf_idx; |
| 1887 | |
| 1888 | ASSERT(reloc_scn->info == NULL || |
| 1889 | reloc_scn->shdr.sh_name == |
| 1890 | reloc_scn->info->old_shdr.sh_name); |
| 1891 | const char *sname = |
| 1892 | elf_strptr(source->oldelf, |
| 1893 | source->shstrndx, |
| 1894 | reloc_scn->shdr.sh_name); |
| 1895 | |
| 1896 | INFO("\tsection [%s] contains %d unfinished relocs.\n", |
| 1897 | sname, |
| 1898 | unfinished->num_rels); |
| 1899 | |
| 1900 | for (unf_idx = 0; unf_idx < unfinished->num_rels; unf_idx++) |
| 1901 | { |
| 1902 | /* There are unfinished relocations. Copy them forward to the |
| 1903 | lowest section we can. */ |
| 1904 | |
| 1905 | while (next_rel_off == |
| 1906 | (int)(next->shdr.sh_size/next->shdr.sh_entsize)) |
| 1907 | { |
| 1908 | INFO("\tsection [%s] has filled up with %d unfinished " |
| 1909 | "relocs.\n", |
| 1910 | sname, |
| 1911 | next_rel_off); |
| 1912 | |
| 1913 | next_idx++; |
| 1914 | ASSERT(next_idx <= idx); |
| 1915 | next = (section_info_t *)scns[next_idx].user; |
| 1916 | next_rel_off = 0; |
| 1917 | } |
| 1918 | |
| 1919 | if (!adjust_section_size_only) { |
| 1920 | INFO("\t\tmoving unfinished relocation %2d to [%s:%d]\n", |
| 1921 | unf_idx, |
| 1922 | sname, |
| 1923 | next_rel_off); |
| 1924 | FAILIF_LIBELF(0 == |
| 1925 | gelf_update_rel(next->data, |
| 1926 | next_rel_off, |
| 1927 | &unfinished->rels[unf_idx]), |
| 1928 | gelf_update_rel); |
| 1929 | } |
| 1930 | |
| 1931 | next_rel_off++; |
| 1932 | dyn->num_unfinished_relocs++; |
| 1933 | } |
| 1934 | } /* for */ |
| 1935 | |
| 1936 | /* Set the size of the last section, and mark all subsequent |
| 1937 | sections for removal. At this point, next is the section |
| 1938 | to which we last wrote data, next_rel_off is the offset before |
| 1939 | which we wrote the last relocation, and so next_rel_off * |
| 1940 | relsize is the new size of the section. |
| 1941 | */ |
| 1942 | |
| 1943 | bool adjust_file = ADJUST_ELF && source->elf_hdr.e_type != ET_EXEC; |
| 1944 | if (adjust_file && !source->dry_run) |
| 1945 | { |
| 1946 | size_t relsize = gelf_fsize(source->elf, |
| 1947 | ELF_T_REL, |
| 1948 | 1, |
| 1949 | source->elf_hdr.e_version); |
| 1950 | |
| 1951 | ASSERT(next->info == NULL || |
| 1952 | next->shdr.sh_name == next->info->old_shdr.sh_name); |
| 1953 | const char *sname = |
| 1954 | elf_strptr(source->oldelf, |
| 1955 | source->shstrndx, |
| 1956 | next->shdr.sh_name); |
| 1957 | |
| 1958 | INFO("\tsection [%s] (index %d) has %d unfinished relocs, " |
| 1959 | "changing its size to %ld bytes (from %ld bytes).\n", |
| 1960 | sname, |
| 1961 | elf_ndxscn(next->scn), |
| 1962 | next_rel_off, |
| 1963 | (long)(next_rel_off * relsize), |
| 1964 | (long)(next->shdr.sh_size)); |
| 1965 | |
| 1966 | /* source->shdr_info[] must be allocated prior to calling this |
| 1967 | function. This is in fact done in process_file(), by calling |
| 1968 | setup_shdr_info() just before we call adjust_dynamic_segment(). |
| 1969 | */ |
| 1970 | ASSERT(source->shdr_info != NULL); |
| 1971 | |
| 1972 | /* We do not update the data field of shdr_info[], because it does |
| 1973 | not exist yet (with ADJUST_ELF != 0). We create the new section |
| 1974 | and section data after the first call to prelink(). For now, we |
| 1975 | save the results of our analysis by modifying the sh_size field |
| 1976 | of the section header. When we create the new sections' data, |
| 1977 | we set the size of the data from the sh_size fields of the |
| 1978 | section headers. |
| 1979 | |
| 1980 | NOTE: The assertion applies only to the first call of |
| 1981 | adjust_dynamic_segment (which calls this function). By |
| 1982 | the second call, we've already created the data for the |
| 1983 | new sections. The only sections for which we haven't |
| 1984 | created data are the relocation sections we are removing. |
| 1985 | */ |
| 1986 | #ifdef DEBUG |
| 1987 | ASSERT((!adjust_section_size_only && |
| 1988 | (source->shdr_info[elf_ndxscn(next->scn)].idx > 0)) || |
| 1989 | source->shdr_info[elf_ndxscn(next->scn)].data == NULL); |
| 1990 | #endif |
| 1991 | |
| 1992 | //FIXME: what else do we need to do here? Do we need to update |
| 1993 | // another copy of the shdr so that it's picked up when we |
| 1994 | // commit the file? |
| 1995 | next->shdr.sh_size = next_rel_off * relsize; |
| 1996 | source->shdr_info[elf_ndxscn(next->scn)].shdr.sh_size = |
| 1997 | next->shdr.sh_size; |
| 1998 | if (next_rel_off * relsize == 0) { |
| 1999 | #ifdef REMOVE_HANDLED_SECTIONS |
| 2000 | INFO("\tsection [%s] (index %d) is now empty, marking for " |
| 2001 | "removal.\n", |
| 2002 | sname, |
| 2003 | elf_ndxscn(next->scn)); |
| 2004 | source->shdr_info[elf_ndxscn(next->scn)].idx = 0; |
| 2005 | dropped_sections = true; |
| 2006 | #endif |
| 2007 | } |
| 2008 | |
| 2009 | while (++next_idx < num_scns) { |
| 2010 | next = (section_info_t *)scns[next_idx].user; |
| 2011 | #ifdef REMOVE_HANDLED_SECTIONS |
| 2012 | ASSERT(next->info == NULL || |
| 2013 | next->shdr.sh_name == next->info->old_shdr.sh_name); |
| 2014 | const char *sname = |
| 2015 | elf_strptr(source->oldelf, |
| 2016 | source->shstrndx, |
| 2017 | next->shdr.sh_name); |
| 2018 | INFO("\tsection [%s] (index %d) is now empty, marking for " |
| 2019 | "removal.\n", |
| 2020 | sname, |
| 2021 | elf_ndxscn(next->scn)); |
| 2022 | /* mark for removal */ |
| 2023 | source->shdr_info[elf_ndxscn(next->scn)].idx = 0; |
| 2024 | dropped_sections = true; |
| 2025 | #endif |
| 2026 | } |
| 2027 | } |
| 2028 | |
| 2029 | } /* if (dyn->sections) */ |
| 2030 | else { |
| 2031 | /* The dynamic entry won't have any sections when it itself doesn't |
| 2032 | exist. This could happen when we remove all relocation sections |
| 2033 | from a dynamic entry because we have managed to handle all |
| 2034 | relocations in them. |
| 2035 | */ |
| 2036 | INFO("\tNo section for dynamic entry!\n"); |
| 2037 | } |
| 2038 | |
| 2039 | return dropped_sections; |
| 2040 | } |
| 2041 | |
| 2042 | static bool adjust_dynamic_segment(source_t *source, |
| 2043 | bool adjust_section_size_only) |
| 2044 | { |
| 2045 | bool dropped_section; |
| 2046 | INFO("Adjusting dynamic segment%s.\n", |
| 2047 | (adjust_section_size_only ? " (section sizes only)" : "")); |
| 2048 | INFO("\tadjusting dynamic segment REL.\n"); |
| 2049 | dropped_section = |
| 2050 | adjust_dynamic_segment_for(source, &source->rel, |
| 2051 | adjust_section_size_only); |
| 2052 | INFO("\tadjusting dynamic segment JMPREL.\n"); |
| 2053 | dropped_section = |
| 2054 | adjust_dynamic_segment_for(source, &source->jmprel, |
| 2055 | adjust_section_size_only) || |
| 2056 | dropped_section; |
| 2057 | if (!adjust_section_size_only) |
| 2058 | adjust_dynamic_segment_entries(source); |
| 2059 | return dropped_section; |
| 2060 | } |
| 2061 | |
| 2062 | static void match_relocation_sections_to_dynamic_ranges(source_t *source) |
| 2063 | { |
| 2064 | /* We've gathered all the DT_DYNAMIC entries; now we need to figure out |
| 2065 | which relocation sections fit in which range as described by the |
| 2066 | entries. |
| 2067 | */ |
| 2068 | |
| 2069 | int relidx; |
| 2070 | for (relidx = 0; relidx < source->num_relocation_sections; relidx++) { |
| 2071 | section_info_t *reloc_scn = &source->relocation_sections[relidx]; |
| 2072 | |
| 2073 | int index = elf_ndxscn(reloc_scn->scn); |
| 2074 | |
| 2075 | ASSERT(reloc_scn->info == NULL || |
| 2076 | reloc_scn->shdr.sh_name == reloc_scn->info->old_shdr.sh_name); |
| 2077 | const char *sname = |
| 2078 | elf_strptr(source->oldelf, |
| 2079 | source->shstrndx, |
| 2080 | reloc_scn->shdr.sh_name); |
| 2081 | |
| 2082 | INFO("Checking section [%s], index %d, for match to dynamic ranges\n", |
| 2083 | sname, index); |
| 2084 | if (source->shdr_info == NULL || reloc_scn->info->idx > 0) { |
| 2085 | if (source->rel.addr && |
| 2086 | source->rel.addr <= reloc_scn->shdr.sh_addr && |
| 2087 | reloc_scn->shdr.sh_addr < source->rel.addr + source->rel.size) |
| 2088 | { |
| 2089 | /* The entire section must fit in the dynamic range. */ |
| 2090 | if((reloc_scn->shdr.sh_addr + reloc_scn->shdr.sh_size) > |
| 2091 | (source->rel.addr + source->rel.size)) |
| 2092 | { |
| 2093 | PRINT("WARNING: In [%s], section %s:[%lld,%lld) " |
| 2094 | "is not fully contained in dynamic range " |
| 2095 | "[%lld,%lld)!\n", |
| 2096 | source->name, |
| 2097 | sname, |
| 2098 | reloc_scn->shdr.sh_addr, |
| 2099 | reloc_scn->shdr.sh_addr + |
| 2100 | reloc_scn->shdr.sh_size, |
| 2101 | source->rel.addr, |
| 2102 | source->rel.addr + source->rel.size); |
| 2103 | } |
| 2104 | |
| 2105 | if (NULL == source->rel.sections) { |
| 2106 | source->rel.sections = init_range_list(); |
| 2107 | ASSERT(source->rel.sections); |
| 2108 | } |
| 2109 | add_unique_range_nosort(source->rel.sections, |
| 2110 | reloc_scn->shdr.sh_addr, |
| 2111 | reloc_scn->shdr.sh_size, |
| 2112 | reloc_scn, |
| 2113 | NULL, |
| 2114 | NULL); |
| 2115 | INFO("\tSection [%s] matches dynamic range REL.\n", |
| 2116 | sname); |
| 2117 | } |
| 2118 | else if (source->jmprel.addr && |
| 2119 | source->jmprel.addr <= reloc_scn->shdr.sh_addr && |
| 2120 | reloc_scn->shdr.sh_addr <= source->jmprel.addr + |
| 2121 | source->jmprel.size) |
| 2122 | { |
| 2123 | if((reloc_scn->shdr.sh_addr + reloc_scn->shdr.sh_size) > |
| 2124 | (source->jmprel.addr + source->jmprel.size)) |
| 2125 | { |
| 2126 | PRINT("WARNING: In [%s], section %s:[%lld,%lld) " |
| 2127 | "is not fully " |
| 2128 | "contained in dynamic range [%lld,%lld)!\n", |
| 2129 | source->name, |
| 2130 | sname, |
| 2131 | reloc_scn->shdr.sh_addr, |
| 2132 | reloc_scn->shdr.sh_addr + |
| 2133 | reloc_scn->shdr.sh_size, |
| 2134 | source->jmprel.addr, |
| 2135 | source->jmprel.addr + source->jmprel.size); |
| 2136 | } |
| 2137 | |
| 2138 | if (NULL == source->jmprel.sections) { |
| 2139 | source->jmprel.sections = init_range_list(); |
| 2140 | ASSERT(source->jmprel.sections); |
| 2141 | } |
| 2142 | add_unique_range_nosort(source->jmprel.sections, |
| 2143 | reloc_scn->shdr.sh_addr, |
| 2144 | reloc_scn->shdr.sh_size, |
| 2145 | reloc_scn, |
| 2146 | NULL, |
| 2147 | NULL); |
| 2148 | INFO("\tSection [%s] matches dynamic range JMPREL.\n", |
| 2149 | sname); |
| 2150 | } |
| 2151 | else |
| 2152 | PRINT("WARNING: Relocation section [%s:%s] does not match " |
| 2153 | "any DT_ entry.\n", |
| 2154 | source->name, |
| 2155 | sname); |
| 2156 | } |
| 2157 | else { |
| 2158 | INFO("Section [%s] was removed, not matching it to dynamic " |
| 2159 | "ranges.\n", |
| 2160 | sname); |
| 2161 | } |
| 2162 | } /* for ... */ |
| 2163 | |
| 2164 | if (source->rel.sections) sort_ranges(source->rel.sections); |
| 2165 | if (source->jmprel.sections) sort_ranges(source->jmprel.sections); |
| 2166 | } |
| 2167 | |
| 2168 | static void drop_sections(source_t *source) |
| 2169 | { |
| 2170 | INFO("We are dropping some sections from [%s]--creating section entries " |
| 2171 | "only for remaining sections.\n", |
| 2172 | source->name); |
| 2173 | /* Renumber the sections. The numbers for the sections after those we are |
| 2174 | dropping will be shifted back by the number of dropped sections. */ |
| 2175 | int cnt, idx; |
| 2176 | for (cnt = idx = 1; cnt < source->shnum; ++cnt) { |
| 2177 | if (source->shdr_info[cnt].idx > 0) { |
| 2178 | source->shdr_info[cnt].idx = idx++; |
| 2179 | |
| 2180 | /* Create a new section. */ |
| 2181 | FAILIF_LIBELF((source->shdr_info[cnt].newscn = |
| 2182 | elf_newscn(source->elf)) == NULL, elf_newscn); |
| 2183 | ASSERT(elf_ndxscn (source->shdr_info[cnt].newscn) == |
| 2184 | source->shdr_info[cnt].idx); |
| 2185 | |
| 2186 | /* Copy the section data */ |
| 2187 | Elf_Data *olddata = |
| 2188 | elf_getdata(source->shdr_info[cnt].scn, // old section |
| 2189 | NULL); |
| 2190 | FAILIF_LIBELF(NULL == olddata, elf_getdata); |
| 2191 | Elf_Data *data = |
| 2192 | elf_newdata(source->shdr_info[cnt].newscn); |
| 2193 | FAILIF_LIBELF(NULL == data, elf_newdata); |
| 2194 | *data = *olddata; |
| 2195 | #if COPY_SECTION_DATA_BUFFER |
| 2196 | if (olddata->d_buf != NULL) { |
| 2197 | data->d_buf = MALLOC(data->d_size); |
| 2198 | memcpy(data->d_buf, olddata->d_buf, olddata->d_size); |
| 2199 | } |
| 2200 | #endif |
| 2201 | source->shdr_info[cnt].data = data; |
| 2202 | |
| 2203 | if (data->d_size != |
| 2204 | source->shdr_info[cnt].shdr.sh_size) { |
| 2205 | INFO("Trimming new-section data from %d to %lld bytes " |
| 2206 | "(as calculated by adjust_dynamic_segment()).\n", |
| 2207 | data->d_size, |
| 2208 | source->shdr_info[cnt].shdr.sh_size); |
| 2209 | data->d_size = |
| 2210 | source->shdr_info[cnt].shdr.sh_size; |
| 2211 | } |
| 2212 | |
| 2213 | INFO("\tsection [%s] (old offset %lld, old size %lld) " |
| 2214 | "will have index %d (was %d), new size %d\n", |
| 2215 | source->shdr_info[cnt].name, |
| 2216 | source->shdr_info[cnt].old_shdr.sh_offset, |
| 2217 | source->shdr_info[cnt].old_shdr.sh_size, |
| 2218 | source->shdr_info[cnt].idx, |
| 2219 | elf_ndxscn(source->shdr_info[cnt].scn), |
| 2220 | data->d_size); |
| 2221 | } else { |
| 2222 | INFO("\tIgnoring section [%s] (offset %lld, size %lld, index %d), " |
| 2223 | "it will be discarded.\n", |
| 2224 | source->shdr_info[cnt].name, |
| 2225 | source->shdr_info[cnt].shdr.sh_offset, |
| 2226 | source->shdr_info[cnt].shdr.sh_size, |
| 2227 | elf_ndxscn(source->shdr_info[cnt].scn)); |
| 2228 | } |
| 2229 | |
| 2230 | /* NOTE: We mark use_old_shdr_for_relocation_calculations even for the |
| 2231 | sections we are removing. adjust_elf has an assertion that makes |
| 2232 | sure that if the values for the size of a section according to its |
| 2233 | header and its data structure differ, then we are using explicitly |
| 2234 | the old section header for calculations, and that the section in |
| 2235 | question is a relocation section. |
| 2236 | */ |
| 2237 | source->shdr_info[cnt].use_old_shdr_for_relocation_calculations = true; |
| 2238 | } /* for */ |
| 2239 | } |
| 2240 | |
| 2241 | static source_t* process_file(const char *filename, |
| 2242 | const char *output, int is_file, |
| 2243 | void (*report_library_size_in_memory)( |
| 2244 | const char *name, off_t fsize), |
| 2245 | unsigned (*get_next_link_address)( |
| 2246 | const char *name), |
| 2247 | int locals_only, |
| 2248 | char **lib_lookup_dirs, |
| 2249 | int num_lib_lookup_dirs, |
| 2250 | char **default_libs, |
| 2251 | int num_default_libs, |
| 2252 | int dry_run, |
| 2253 | int *total_num_handled_relocs, |
| 2254 | int *total_num_unhandled_relocs) |
| 2255 | { |
| 2256 | /* Look up the file in the list of already-handles files, which are |
| 2257 | represented by source_t structs. If we do not find the file, then we |
| 2258 | haven't prelinked it yet. If we find it, then we have, so we do |
| 2259 | nothing. Keep in mind that apriori operates on an entire collection |
| 2260 | of files, and if application A used library L, and so does application |
| 2261 | B, if we process A first, then by the time we get to B we will have |
| 2262 | prelinked L already; that's why we check first to see if a library has |
| 2263 | been prelinked. |
| 2264 | */ |
| 2265 | source_t *source = |
| 2266 | find_source(filename, lib_lookup_dirs, num_lib_lookup_dirs); |
| 2267 | if (NULL == source) { |
| 2268 | /* If we could not find the source, then it hasn't been processed yet, |
| 2269 | so we go ahead and process it! */ |
| 2270 | INFO("Processing [%s].\n", filename); |
| 2271 | char *full = find_file(filename, lib_lookup_dirs, num_lib_lookup_dirs); |
| 2272 | FAILIF(NULL == full, |
| 2273 | "Could not find [%s] in the current directory or in any of " |
| 2274 | "the search paths!\n", filename); |
| 2275 | |
| 2276 | unsigned base = get_next_link_address(full); |
| 2277 | |
| 2278 | source = init_source(full, output, is_file, base, dry_run); |
| 2279 | |
| 2280 | if (source == NULL) { |
| 2281 | INFO("File [%s] is a static executable.\n", filename); |
| 2282 | return NULL; |
| 2283 | } |
| 2284 | ASSERT(source->dynamic.scn != NULL); |
| 2285 | |
| 2286 | /* We need to increment the next prelink address only when the file we |
| 2287 | are currently handing is a shared library. Executables do not need |
| 2288 | to be prelinked at a different address, they are always at address |
| 2289 | zero. |
| 2290 | |
| 2291 | Also, if we are prelinking locals only, then we are handling a |
| 2292 | single file per invokation of apriori, so there is no need to |
| 2293 | increment the prelink address unless there is a global prelink map, |
| 2294 | in which case we do need to check to see if the library isn't |
| 2295 | running into its neighbouts in the prelink map. |
| 2296 | */ |
| 2297 | if (source->oldelf_hdr.e_type != ET_EXEC && |
| 2298 | (!locals_only || |
| 2299 | report_library_size_in_memory == |
| 2300 | pm_report_library_size_in_memory)) { |
| 2301 | /* This sets the next link address only if an increment was not |
| 2302 | specified by the user. If an address increment was specified, |
| 2303 | then we just check to make sure that the file size is less than |
| 2304 | the increment. |
| 2305 | |
| 2306 | NOTE: The file size is the absolute highest number of bytes that |
| 2307 | the file may occupy in memory, if the entire file is loaded, but |
| 2308 | this is almost next the case. A file will often have sections |
| 2309 | which are not loaded, which could add a lot of size. That's why |
| 2310 | we start off with the file size and then subtract the size of |
| 2311 | the biggest sections that will not get loaded, which are the |
| 2312 | varios DWARF sections, all of which of which are named starting |
| 2313 | with ".debug_". |
| 2314 | |
| 2315 | We could do better than this (by caculating exactly how many |
| 2316 | bytes from that file will be loaded), but that's an overkill. |
| 2317 | Unless the prelink-address increment becomes too small, the file |
| 2318 | size after subtracting the sizes of the DWARF section will be a |
| 2319 | good-enough upper bound. |
| 2320 | */ |
| 2321 | |
| 2322 | unsigned long fsize = source->elf_file_info.st_size; |
| 2323 | INFO("Calculating loadable file size for next link address. " |
| 2324 | "Starting with %ld.\n", fsize); |
| 2325 | if (true) { |
| 2326 | Elf_Scn *scn = NULL; |
| 2327 | GElf_Shdr shdr_mem, *shdr; |
| 2328 | const char *scn_name; |
| 2329 | while ((scn = elf_nextscn (source->oldelf, scn)) != NULL) { |
| 2330 | shdr = gelf_getshdr(scn, &shdr_mem); |
| 2331 | FAILIF_LIBELF(NULL == shdr, gelf_getshdr); |
| 2332 | scn_name = elf_strptr (source->oldelf, |
| 2333 | source->shstrndx, shdr->sh_name); |
| 2334 | ASSERT(scn_name != NULL); |
| 2335 | |
| 2336 | if (!(shdr->sh_flags & SHF_ALLOC)) { |
| 2337 | INFO("\tDecrementing by %lld on account of section " |
| 2338 | "[%s].\n", |
| 2339 | shdr->sh_size, |
| 2340 | scn_name); |
| 2341 | fsize -= shdr->sh_size; |
| 2342 | } |
| 2343 | } |
| 2344 | } |
| 2345 | INFO("Done calculating loadable file size for next link address: " |
| 2346 | "Final value is %ld.\n", fsize); |
| 2347 | report_library_size_in_memory(source->name, fsize); |
| 2348 | } |
| 2349 | |
| 2350 | /* Identify the dynamic segment and process it. Specifically, we find |
| 2351 | out what dependencies, if any, this file has. Whenever we encounter |
| 2352 | such a dependency, we process it recursively; we find out where the |
| 2353 | various relocation information sections are stored. */ |
| 2354 | |
| 2355 | size_t dynidx; |
| 2356 | GElf_Dyn *dyn, dyn_mem; |
| 2357 | size_t numdyn = |
| 2358 | source->dynamic.shdr.sh_size / |
| 2359 | source->dynamic.shdr.sh_entsize; |
| 2360 | ASSERT(source->dynamic.shdr.sh_size == source->dynamic.data->d_size); |
| 2361 | |
| 2362 | source->rel.idx = source->rel.sz_idx = -1; |
| 2363 | source->jmprel.idx = source->jmprel.sz_idx = -1; |
| 2364 | |
| 2365 | for (dynidx = 0; dynidx < numdyn; dynidx++) { |
| 2366 | dyn = gelf_getdyn (source->dynamic.data, |
| 2367 | dynidx, |
| 2368 | &dyn_mem); |
| 2369 | FAILIF_LIBELF(NULL == dyn, gelf_getdyn); |
| 2370 | /* When we are processing only the local relocations in a file, |
| 2371 | we don't need to handle any of the dependencies. It won't |
| 2372 | hurt if we do, but we will be doing unnecessary work. |
| 2373 | */ |
| 2374 | switch (dyn->d_tag) |
| 2375 | { |
| 2376 | case DT_NEEDED: |
| 2377 | if (!locals_only) { |
| 2378 | /* Process the needed library recursively. |
| 2379 | */ |
| 2380 | const char *dep_lib = |
| 2381 | #if ELF_STRPTR_IS_BROKEN |
| 2382 | (((char *)elf_getdata( |
| 2383 | elf_getscn(source->elf, |
| 2384 | source->dynamic.shdr.sh_link), |
| 2385 | NULL)->d_buf) + dyn->d_un.d_val); |
| 2386 | #else |
| 2387 | elf_strptr (source->elf, |
| 2388 | source->dynamic.shdr.sh_link, |
| 2389 | dyn->d_un.d_val); |
| 2390 | #endif |
| 2391 | ASSERT(dep_lib != NULL); |
| 2392 | INFO("[%s] depends on [%s].\n", filename, dep_lib); |
| 2393 | ASSERT(output == NULL || is_file == 0); |
| 2394 | source_t *dep = process_file(dep_lib, |
| 2395 | output, is_file, |
| 2396 | report_library_size_in_memory, |
| 2397 | get_next_link_address, |
| 2398 | locals_only, |
| 2399 | lib_lookup_dirs, |
| 2400 | num_lib_lookup_dirs, |
| 2401 | default_libs, |
| 2402 | num_default_libs, |
| 2403 | dry_run, |
| 2404 | total_num_handled_relocs, |
| 2405 | total_num_unhandled_relocs); |
| 2406 | |
| 2407 | /* Add the library to the dependency list. */ |
| 2408 | if (source->num_lib_deps == source->lib_deps_size) { |
| 2409 | source->lib_deps_size += 10; |
| 2410 | source->lib_deps = REALLOC(source->lib_deps, |
| 2411 | source->lib_deps_size * |
| 2412 | sizeof(source_t *)); |
| 2413 | } |
| 2414 | source->lib_deps[source->num_lib_deps++] = dep; |
| 2415 | } |
| 2416 | break; |
| 2417 | case DT_JMPREL: |
| 2418 | source->jmprel.idx = dynidx; |
| 2419 | source->jmprel.addr = dyn->d_un.d_ptr; |
| 2420 | break; |
| 2421 | case DT_PLTRELSZ: |
| 2422 | source->jmprel.sz_idx = dynidx; |
| 2423 | source->jmprel.size = dyn->d_un.d_val; |
| 2424 | break; |
| 2425 | case DT_REL: |
| 2426 | source->rel.idx = dynidx; |
| 2427 | source->rel.addr = dyn->d_un.d_ptr; |
| 2428 | break; |
| 2429 | case DT_RELSZ: |
| 2430 | source->rel.sz_idx = dynidx; |
| 2431 | source->rel.size = dyn->d_un.d_val; |
| 2432 | break; |
| 2433 | case DT_RELA: |
| 2434 | case DT_RELASZ: |
| 2435 | FAILIF(1, "Can't handle DT_RELA and DT_RELASZ entries!\n"); |
| 2436 | break; |
| 2437 | } /* switch */ |
| 2438 | } /* for each dynamic entry... */ |
| 2439 | |
| 2440 | INFO("Handling [%s].\n", filename); |
| 2441 | |
| 2442 | #ifdef SUPPORT_ANDROID_PRELINK_TAGS |
| 2443 | if (!source->prelinked) |
| 2444 | #endif |
| 2445 | { |
| 2446 | /* When ADJUST_ELF is defined, this call to prelink is a dry run |
| 2447 | intended to calculate the number of relocations that could not |
| 2448 | be handled. This, in turn, allows us to calculate the amount by |
| 2449 | which we can shrink the various relocation sections before we |
| 2450 | call adjust_elf. After we've adjusted the sections, we will |
| 2451 | call prelink() one more time to do the actual work. |
| 2452 | |
| 2453 | NOTE: Even when ADJUST_ELF != 0, we cannot adjust an ELF file |
| 2454 | that is an executabe, because an executable is not PIC. |
| 2455 | */ |
| 2456 | |
| 2457 | int num_unfinished_relocs = 0; |
| 2458 | bool adjust_file = ADJUST_ELF && source->elf_hdr.e_type != ET_EXEC; |
| 2459 | INFO("\n\n\tPRELINKING %s\n\n", |
| 2460 | adjust_file ? |
| 2461 | "(CALCULATE NUMBER OF HANDLED RELOCATIONS)" : |
| 2462 | "(ACTUAL)"); |
| 2463 | int num_relocs = prelink(source, locals_only, |
| 2464 | adjust_file || dry_run, |
| 2465 | lib_lookup_dirs, num_lib_lookup_dirs, |
| 2466 | default_libs, num_default_libs, |
| 2467 | &num_unfinished_relocs); |
| 2468 | INFO("[%s]: (calculate changes) handled %d, could not handle %d " |
| 2469 | "relocations.\n", |
| 2470 | source->name, |
| 2471 | num_relocs, |
| 2472 | num_unfinished_relocs); |
| 2473 | |
| 2474 | if (adjust_file && !dry_run) |
| 2475 | { |
| 2476 | /* Find out the new section sizes of the relocation sections, |
| 2477 | but do not move any relocations around, because adjust_elf |
| 2478 | needs to know about all relocations in order to adjust the |
| 2479 | file correctly. |
| 2480 | */ |
| 2481 | match_relocation_sections_to_dynamic_ranges(source); |
| 2482 | |
| 2483 | /* We haven't set up source->shdr_info[] yet, so we do it now. |
| 2484 | |
| 2485 | NOTE: setup_shdr_info() depends only on source->oldelf, not |
| 2486 | on source->elf! source->elf is not even defined yet. We |
| 2487 | initialize source->shdr_info[] based on the section |
| 2488 | information of the unmodified ELF file, and then make our |
| 2489 | modifications in the call to adjust_dynamic_segment() based |
| 2490 | on this information. adjust_dynamic_segment() will |
| 2491 | rearrange the unhandled relocations in the beginning of |
| 2492 | their relocation sections, and adjust the size of those |
| 2493 | relocation sections. In the case when a relocation section |
| 2494 | is completely handled, adjust_dynamic_segment() will mark it |
| 2495 | for removal by function adjust_elf. |
| 2496 | */ |
| 2497 | |
| 2498 | ASSERT(source->elf == source->oldelf); |
| 2499 | ASSERT(source->shdr_info == NULL); |
| 2500 | setup_shdr_info(source); |
| 2501 | ASSERT(source->shdr_info != NULL); |
| 2502 | |
| 2503 | INFO("\n\n\tADJUSTING DYNAMIC SEGMENT " |
| 2504 | "(CALCULATE CHANGES)\n\n"); |
| 2505 | bool drop_some_sections = adjust_dynamic_segment(source, true); |
| 2506 | |
| 2507 | /* Reopen the elf file! Note that we are not doing a dry run |
| 2508 | (the if statement above makes sure of that.) |
| 2509 | |
| 2510 | NOTE: We call init_elf() after we called |
| 2511 | adjust_dynamic_segment() in order to have |
| 2512 | adjust_dynamic_segment() refer to source->oldelf when |
| 2513 | it refers to source->elf. Since |
| 2514 | adjust_dynamic_segment doesn't actually write to the |
| 2515 | ELF file, this is OK. adjust_dynamic_segment() |
| 2516 | updates the sh_size fields of saved section headers |
| 2517 | and optionally marks sections for removal. |
| 2518 | |
| 2519 | Having adjust_dynamic_segment() refer to |
| 2520 | source->oldelf means that we'll have access to |
| 2521 | section-name strings so we can print them out in our |
| 2522 | logging and debug output. |
| 2523 | */ |
| 2524 | source->elf = init_elf(source, false); |
| 2525 | |
| 2526 | /* This is the same code as in init_source() after the call to |
| 2527 | * init_elf(). */ |
| 2528 | ASSERT(source->elf != source->oldelf); |
| 2529 | ebl_closebackend(source->ebl); |
| 2530 | source->ebl = ebl_openbackend (source->elf); |
| 2531 | FAILIF_LIBELF(NULL == source->ebl, ebl_openbackend); |
| 2532 | #ifdef ARM_SPECIFIC_HACKS |
| 2533 | FAILIF_LIBELF(0 != arm_init(source->elf, |
| 2534 | source->elf_hdr.e_machine, |
| 2535 | source->ebl, sizeof(Ebl)), |
| 2536 | arm_init); |
| 2537 | #endif/*ARM_SPECIFIC_HACKS*/ |
| 2538 | |
| 2539 | if (drop_some_sections) |
| 2540 | drop_sections(source); |
| 2541 | else { |
| 2542 | INFO("All sections remain in [%s]--we are changing at " |
| 2543 | "most section sizes.\n", source->name); |
| 2544 | create_elf_sections(source, NULL); |
| 2545 | int cnt, idx; |
| 2546 | for (cnt = idx = 1; cnt < source->shnum; ++cnt) { |
| 2547 | Elf_Data *data = elf_getdata( |
| 2548 | source->shdr_info[cnt].newscn, // new section |
| 2549 | NULL); |
| 2550 | if (data->d_size != |
| 2551 | source->shdr_info[cnt].shdr.sh_size) { |
| 2552 | INFO("Trimming new-section data from %d to %lld " |
| 2553 | "bytes (as calculated by " |
| 2554 | "adjust_dynamic_segment()).\n", |
| 2555 | data->d_size, |
| 2556 | source->shdr_info[cnt].shdr.sh_size); |
| 2557 | data->d_size = source->shdr_info[cnt].shdr.sh_size; |
| 2558 | } |
| 2559 | } |
| 2560 | } |
| 2561 | |
| 2562 | /* Shrink it! */ |
| 2563 | INFO("\n\n\tADJUSTING ELF\n\n"); |
| 2564 | adjust_elf( |
| 2565 | source->oldelf, source->name, |
| 2566 | source->elf, source->name, |
| 2567 | source->ebl, |
| 2568 | &source->old_ehdr_mem, |
| 2569 | NULL, 0, // no symbol filter |
| 2570 | source->shdr_info, // information on how to adjust the ELF |
| 2571 | source->shnum, // length of source->shdr_info[] |
| 2572 | source->phdr_info, // program-header info |
| 2573 | source->shnum, // irrelevant--we're not rebuilding shstrtab |
| 2574 | source->shnum, // number of sections in file |
| 2575 | source->shstrndx, // index of shstrtab (both in |
| 2576 | // shdr_info[] and as a section index) |
| 2577 | NULL, // irrelevant, since we are not rebuilding shstrtab |
| 2578 | drop_some_sections, // some sections are being dropped |
| 2579 | elf_ndxscn(source->dynamic.scn), // index of .dynamic |
| 2580 | elf_ndxscn(source->symtab.scn), // index of .dynsym |
| 2581 | 1, // allow shady business |
| 2582 | &source->shstrtab_data, |
| 2583 | true, |
| 2584 | false); // do not rebuild shstrtab |
| 2585 | |
| 2586 | INFO("\n\n\tREINITIALIZING STRUCTURES " |
| 2587 | "(TO CONTAIN ADJUSTMENTS)\n\n"); |
| 2588 | reinit_source(source); |
| 2589 | |
| 2590 | INFO("\n\n\tPRELINKING (ACTUAL)\n\n"); |
| 2591 | #ifdef DEBUG |
| 2592 | int old_num_unfinished_relocs = num_unfinished_relocs; |
| 2593 | #endif |
| 2594 | num_unfinished_relocs = 0; |
| 2595 | #ifdef DEBUG |
| 2596 | int num_relocs_take_two = |
| 2597 | #endif |
| 2598 | prelink(source, locals_only, |
| 2599 | false, /* not a dry run */ |
| 2600 | lib_lookup_dirs, num_lib_lookup_dirs, |
| 2601 | default_libs, num_default_libs, |
| 2602 | &num_unfinished_relocs); |
| 2603 | |
| 2604 | /* The numbers for the total number of relocations and the |
| 2605 | number of unhandled relocations between the first and second |
| 2606 | invokationof prelink() must be the same! The first time we |
| 2607 | ran prelink() just to calculate the numbers so that we could |
| 2608 | calculate the adjustments to pass to adjust_elf, and the |
| 2609 | second time we actually carry out the prelinking; the |
| 2610 | numbers must stay the same! |
| 2611 | */ |
| 2612 | ASSERT(num_relocs == num_relocs_take_two); |
| 2613 | ASSERT(old_num_unfinished_relocs == num_unfinished_relocs); |
| 2614 | |
| 2615 | INFO("[%s]: (actual prelink) handled %d, could not " |
| 2616 | "handle %d relocations.\n", |
| 2617 | source->name, |
| 2618 | num_relocs, |
| 2619 | num_unfinished_relocs); |
| 2620 | } /* if (adjust_elf && !dry_run) */ |
| 2621 | |
| 2622 | *total_num_handled_relocs += num_relocs; |
| 2623 | *total_num_unhandled_relocs += num_unfinished_relocs; |
| 2624 | |
| 2625 | if(num_unfinished_relocs != 0 && |
| 2626 | source->elf_hdr.e_type != ET_EXEC && |
| 2627 | !locals_only) |
| 2628 | { |
| 2629 | /* One reason you could have unfinished relocations in an |
| 2630 | executable file is if this file used dlopen() and friends. |
| 2631 | We do not adjust relocation entries to those symbols, |
| 2632 | because libdl is a dummy only--the real functions are |
| 2633 | provided for by the dynamic linker itsef. |
| 2634 | |
| 2635 | NOTE FIXME HACK: This is specific to the Android dynamic |
| 2636 | linker, and may not be true in other cases. |
| 2637 | */ |
| 2638 | PRINT("WARNING: Expecting to have unhandled relocations only " |
| 2639 | "for executables (%s is not an executable)!\n", |
| 2640 | source->name); |
| 2641 | } |
| 2642 | |
| 2643 | match_relocation_sections_to_dynamic_ranges(source); |
| 2644 | |
| 2645 | /* Now, for each relocation section, check to see if its address |
| 2646 | matches one of the DT_DYNAMIC relocation pointers. If so, then |
| 2647 | if the section has no unhandled relocations, simply set the |
| 2648 | associated DT_DYNAMIC entry's size to zero. If the section does |
| 2649 | have unhandled entries, then lump them all together at the front |
| 2650 | of the respective section and update the size of the respective |
| 2651 | DT_DYNAMIC entry to the new size of the section. A better |
| 2652 | approach would be do delete a relocation section if it has been |
| 2653 | fully relocated and to remove its entry from the DT_DYNAMIC |
| 2654 | array, and for relocation entries that still have some |
| 2655 | relocations in them, we should shrink the section if that won't |
| 2656 | violate relative offsets. This is more work, however, and for |
| 2657 | the speed improvement we expect from a prelinker, just patching |
| 2658 | up DT_DYNAMIC will suffice. |
| 2659 | |
| 2660 | Note: adjust_dynamic_segment() will modify source->shdr_info[] |
| 2661 | to denote any change in a relocation section's size. This |
| 2662 | will be picked up by adjust_elf, which will rearrange the |
| 2663 | file to eliminate the gap created by the decrease in size |
| 2664 | of the relocation section. We do not need to do this, but |
| 2665 | the relocation section could be large, and reduced |
| 2666 | drastically by the prelinking process, so it pays to |
| 2667 | adjust the file. |
| 2668 | */ |
| 2669 | |
| 2670 | INFO("\n\n\tADJUSTING DYNAMIC SEGMENT (ACTUAL)\n\n"); |
| 2671 | adjust_dynamic_segment(source, false); |
| 2672 | } |
| 2673 | #ifdef SUPPORT_ANDROID_PRELINK_TAGS |
| 2674 | else INFO("[%s] is already prelinked at 0x%08lx.\n", |
| 2675 | filename, |
| 2676 | source->prelink_base); |
| 2677 | #endif |
| 2678 | } else INFO("[%s] has been processed already.\n", filename); |
| 2679 | |
| 2680 | return source; |
| 2681 | } |
| 2682 | |
| 2683 | void apriori(char **execs, int num_execs, |
| 2684 | char *output, |
| 2685 | void (*report_library_size_in_memory)( |
| 2686 | const char *name, off_t fsize), |
| 2687 | int (*get_next_link_address)(const char *name), |
| 2688 | int locals_only, |
| 2689 | int dry_run, |
| 2690 | char **lib_lookup_dirs, int num_lib_lookup_dirs, |
| 2691 | char **default_libs, int num_default_libs, |
| 2692 | char *mapfile) |
| 2693 | { |
| 2694 | source_t *source; /* for general usage */ |
| 2695 | int input_idx; |
| 2696 | |
| 2697 | ASSERT(report_library_size_in_memory != NULL); |
| 2698 | ASSERT(get_next_link_address != NULL); |
| 2699 | |
| 2700 | /* Process and prelink each executable and object file. Function |
| 2701 | process_file() is called for each executable in the loop below. |
| 2702 | It calls itself recursively for each library. We prelink each library |
| 2703 | after prelinking its dependencies. */ |
| 2704 | int total_num_handled_relocs = 0, total_num_unhandled_relocs = 0; |
| 2705 | for (input_idx = 0; input_idx < num_execs; input_idx++) { |
| 2706 | INFO("executable: [%s]\n", execs[input_idx]); |
| 2707 | /* Here process_file() is actually processing the top-level |
| 2708 | executable files. */ |
| 2709 | process_file(execs[input_idx], output, num_execs == 1, |
| 2710 | report_library_size_in_memory, |
| 2711 | get_next_link_address, /* executables get a link address |
| 2712 | of zero, regardless of this |
| 2713 | value */ |
| 2714 | locals_only, |
| 2715 | lib_lookup_dirs, num_lib_lookup_dirs, |
| 2716 | default_libs, num_default_libs, |
| 2717 | dry_run, |
| 2718 | &total_num_handled_relocs, |
| 2719 | &total_num_unhandled_relocs); |
| 2720 | /* if source is NULL, then the respective executable is static */ |
| 2721 | /* Mark the source as an executable */ |
| 2722 | } /* for each input executable... */ |
| 2723 | |
| 2724 | PRINT("Handled %d relocations.\n", total_num_handled_relocs); |
| 2725 | PRINT("Could not handle %d relocations.\n", total_num_unhandled_relocs); |
| 2726 | |
| 2727 | /* We are done! Since the end result of our calculations is a set of |
| 2728 | symbols for each library that other libraries or executables link |
| 2729 | against, we iterate over the set of libraries one last time, and for |
| 2730 | each symbol that is marked as satisfying some dependence, we emit |
| 2731 | a line with the symbol's name to a text file derived from the library's |
| 2732 | name by appending the suffix .syms to it. */ |
| 2733 | |
| 2734 | if (mapfile != NULL) { |
| 2735 | const char *mapfile_name = mapfile; |
| 2736 | FILE *fp; |
| 2737 | if (*mapfile == '+') { |
| 2738 | mapfile_name = mapfile + 1; |
| 2739 | INFO("Opening map file %s for append/write.\n", |
| 2740 | mapfile_name); |
| 2741 | fp = fopen(mapfile_name, "a"); |
| 2742 | } |
| 2743 | else fp = fopen(mapfile_name, "w"); |
| 2744 | |
| 2745 | FAILIF(fp == NULL, "Cannot open file [%s]: %s (%d)!\n", |
| 2746 | mapfile_name, |
| 2747 | strerror(errno), |
| 2748 | errno); |
| 2749 | source = sources; |
| 2750 | while (source) { |
| 2751 | /* If it's a library, print the results. */ |
| 2752 | if (source->elf_hdr.e_type == ET_DYN) { |
| 2753 | /* Add to the memory map file. */ |
| 2754 | fprintf(fp, "%s 0x%08lx %lld\n", |
| 2755 | basename(source->name), |
| 2756 | source->base, |
| 2757 | source->elf_file_info.st_size); |
| 2758 | } |
| 2759 | source = source->next; |
| 2760 | } |
| 2761 | fclose(fp); |
| 2762 | } |
| 2763 | |
| 2764 | /* Free the resources--you can't do it in the loop above because function |
| 2765 | print_symbol_references() accesses nodes other than the one being |
| 2766 | iterated over. |
| 2767 | */ |
| 2768 | source = sources; |
| 2769 | while (source) { |
| 2770 | source_t *old = source; |
| 2771 | source = source->next; |
| 2772 | /* Destroy the evidence. */ |
| 2773 | destroy_source(old); |
| 2774 | } |
| 2775 | } |