linker.cpp revision 812fd4263a005b88f3b4222baa910114f938d594
1/*
2 * Copyright (C) 2008, 2009 The Android Open Source Project
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 *  * Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 *  * Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in
12 *    the documentation and/or other materials provided with the
13 *    distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
18 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
19 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
22 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
25 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29#include <dlfcn.h>
30#include <errno.h>
31#include <fcntl.h>
32#include <inttypes.h>
33#include <pthread.h>
34#include <stdio.h>
35#include <stdlib.h>
36#include <string.h>
37#include <sys/atomics.h>
38#include <sys/mman.h>
39#include <sys/stat.h>
40#include <unistd.h>
41
42// Private C library headers.
43#include "private/bionic_tls.h"
44#include "private/KernelArgumentBlock.h"
45#include "private/ScopedPthreadMutexLocker.h"
46
47#include "linker.h"
48#include "linker_debug.h"
49#include "linker_environ.h"
50#include "linker_phdr.h"
51
52/* >>> IMPORTANT NOTE - READ ME BEFORE MODIFYING <<<
53 *
54 * Do NOT use malloc() and friends or pthread_*() code here.
55 * Don't use printf() either; it's caused mysterious memory
56 * corruption in the past.
57 * The linker runs before we bring up libc and it's easiest
58 * to make sure it does not depend on any complex libc features
59 *
60 * open issues / todo:
61 *
62 * - are we doing everything we should for ARM_COPY relocations?
63 * - cleaner error reporting
64 * - after linking, set as much stuff as possible to READONLY
65 *   and NOEXEC
66 */
67
68static bool soinfo_link_image(soinfo* si);
69static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf);
70
71// We can't use malloc(3) in the dynamic linker. We use a linked list of anonymous
72// maps, each a single page in size. The pages are broken up into as many struct soinfo
73// objects as will fit, and they're all threaded together on a free list.
74#define SOINFO_PER_POOL ((PAGE_SIZE - sizeof(soinfo_pool_t*)) / sizeof(soinfo))
75struct soinfo_pool_t {
76  soinfo_pool_t* next;
77  soinfo info[SOINFO_PER_POOL];
78};
79static struct soinfo_pool_t* gSoInfoPools = NULL;
80static soinfo* gSoInfoFreeList = NULL;
81
82static soinfo* solist = &libdl_info;
83static soinfo* sonext = &libdl_info;
84static soinfo* somain; /* main process, always the one after libdl_info */
85
86static const char* const gDefaultLdPaths[] = {
87#if defined(__LP64__)
88  "/vendor/lib64",
89  "/system/lib64",
90#else
91  "/vendor/lib",
92  "/system/lib",
93#endif
94  NULL
95};
96
97#define LDPATH_BUFSIZE (LDPATH_MAX*64)
98#define LDPATH_MAX 8
99
100#define LDPRELOAD_BUFSIZE (LDPRELOAD_MAX*64)
101#define LDPRELOAD_MAX 8
102
103static char gLdPathsBuffer[LDPATH_BUFSIZE];
104static const char* gLdPaths[LDPATH_MAX + 1];
105
106static char gLdPreloadsBuffer[LDPRELOAD_BUFSIZE];
107static const char* gLdPreloadNames[LDPRELOAD_MAX + 1];
108
109static soinfo* gLdPreloads[LDPRELOAD_MAX + 1];
110
111__LIBC_HIDDEN__ int gLdDebugVerbosity;
112
113__LIBC_HIDDEN__ abort_msg_t* gAbortMessage = NULL; // For debuggerd.
114
115enum RelocationKind {
116    kRelocAbsolute = 0,
117    kRelocRelative,
118    kRelocCopy,
119    kRelocSymbol,
120    kRelocMax
121};
122
123#if STATS
124struct linker_stats_t {
125    int count[kRelocMax];
126};
127
128static linker_stats_t linker_stats;
129
130static void count_relocation(RelocationKind kind) {
131    ++linker_stats.count[kind];
132}
133#else
134static void count_relocation(RelocationKind) {
135}
136#endif
137
138#if COUNT_PAGES
139static unsigned bitmask[4096];
140#if defined(__LP64__)
141#define MARK(offset) \
142    do { \
143        if ((((offset) >> 12) >> 5) < 4096) \
144            bitmask[((offset) >> 12) >> 5] |= (1 << (((offset) >> 12) & 31)); \
145    } while (0)
146#else
147#define MARK(offset) \
148    do { \
149        bitmask[((offset) >> 12) >> 3] |= (1 << (((offset) >> 12) & 7)); \
150    } while (0)
151#endif
152#else
153#define MARK(x) do {} while (0)
154#endif
155
156// You shouldn't try to call memory-allocating functions in the dynamic linker.
157// Guard against the most obvious ones.
158#define DISALLOW_ALLOCATION(return_type, name, ...) \
159    return_type name __VA_ARGS__ \
160    { \
161        const char* msg = "ERROR: " #name " called from the dynamic linker!\n"; \
162        __libc_format_log(ANDROID_LOG_FATAL, "linker", "%s", msg); \
163        write(2, msg, strlen(msg)); \
164        abort(); \
165    }
166DISALLOW_ALLOCATION(void*, malloc, (size_t u __unused));
167DISALLOW_ALLOCATION(void, free, (void* u __unused));
168DISALLOW_ALLOCATION(void*, realloc, (void* u1 __unused, size_t u2 __unused));
169DISALLOW_ALLOCATION(void*, calloc, (size_t u1 __unused, size_t u2 __unused));
170
171static char tmp_err_buf[768];
172static char __linker_dl_err_buf[768];
173
174char* linker_get_error_buffer() {
175  return &__linker_dl_err_buf[0];
176}
177
178size_t linker_get_error_buffer_size() {
179  return sizeof(__linker_dl_err_buf);
180}
181
182/*
183 * This function is an empty stub where GDB locates a breakpoint to get notified
184 * about linker activity.
185 */
186extern "C" void __attribute__((noinline)) __attribute__((visibility("default"))) rtld_db_dlactivity();
187
188static r_debug _r_debug = {1, NULL, reinterpret_cast<uintptr_t>(&rtld_db_dlactivity), r_debug::RT_CONSISTENT, 0};
189static link_map* r_debug_tail = 0;
190
191static pthread_mutex_t gDebugMutex = PTHREAD_MUTEX_INITIALIZER;
192
193static void insert_soinfo_into_debug_map(soinfo* info) {
194    // Copy the necessary fields into the debug structure.
195    link_map* map = &(info->link_map_head);
196    map->l_addr = info->load_bias;
197    map->l_name = reinterpret_cast<char*>(info->name);
198    map->l_ld = info->dynamic;
199
200    /* Stick the new library at the end of the list.
201     * gdb tends to care more about libc than it does
202     * about leaf libraries, and ordering it this way
203     * reduces the back-and-forth over the wire.
204     */
205    if (r_debug_tail) {
206        r_debug_tail->l_next = map;
207        map->l_prev = r_debug_tail;
208        map->l_next = 0;
209    } else {
210        _r_debug.r_map = map;
211        map->l_prev = 0;
212        map->l_next = 0;
213    }
214    r_debug_tail = map;
215}
216
217static void remove_soinfo_from_debug_map(soinfo* info) {
218    link_map* map = &(info->link_map_head);
219
220    if (r_debug_tail == map) {
221        r_debug_tail = map->l_prev;
222    }
223
224    if (map->l_prev) {
225        map->l_prev->l_next = map->l_next;
226    }
227    if (map->l_next) {
228        map->l_next->l_prev = map->l_prev;
229    }
230}
231
232static void notify_gdb_of_load(soinfo* info) {
233    if (info->flags & FLAG_EXE) {
234        // GDB already knows about the main executable
235        return;
236    }
237
238    ScopedPthreadMutexLocker locker(&gDebugMutex);
239
240    _r_debug.r_state = r_debug::RT_ADD;
241    rtld_db_dlactivity();
242
243    insert_soinfo_into_debug_map(info);
244
245    _r_debug.r_state = r_debug::RT_CONSISTENT;
246    rtld_db_dlactivity();
247}
248
249static void notify_gdb_of_unload(soinfo* info) {
250    if (info->flags & FLAG_EXE) {
251        // GDB already knows about the main executable
252        return;
253    }
254
255    ScopedPthreadMutexLocker locker(&gDebugMutex);
256
257    _r_debug.r_state = r_debug::RT_DELETE;
258    rtld_db_dlactivity();
259
260    remove_soinfo_from_debug_map(info);
261
262    _r_debug.r_state = r_debug::RT_CONSISTENT;
263    rtld_db_dlactivity();
264}
265
266void notify_gdb_of_libraries() {
267  _r_debug.r_state = r_debug::RT_ADD;
268  rtld_db_dlactivity();
269  _r_debug.r_state = r_debug::RT_CONSISTENT;
270  rtld_db_dlactivity();
271}
272
273static bool ensure_free_list_non_empty() {
274  if (gSoInfoFreeList != NULL) {
275    return true;
276  }
277
278  // Allocate a new pool.
279  soinfo_pool_t* pool = reinterpret_cast<soinfo_pool_t*>(mmap(NULL, sizeof(*pool),
280                                                              PROT_READ|PROT_WRITE,
281                                                              MAP_PRIVATE|MAP_ANONYMOUS, 0, 0));
282  if (pool == MAP_FAILED) {
283    return false;
284  }
285
286  // Add the pool to our list of pools.
287  pool->next = gSoInfoPools;
288  gSoInfoPools = pool;
289
290  // Chain the entries in the new pool onto the free list.
291  gSoInfoFreeList = &pool->info[0];
292  soinfo* next = NULL;
293  for (int i = SOINFO_PER_POOL - 1; i >= 0; --i) {
294    pool->info[i].next = next;
295    next = &pool->info[i];
296  }
297
298  return true;
299}
300
301static void set_soinfo_pool_protection(int protection) {
302  for (soinfo_pool_t* p = gSoInfoPools; p != NULL; p = p->next) {
303    if (mprotect(p, sizeof(*p), protection) == -1) {
304      abort(); // Can't happen.
305    }
306  }
307}
308
309static soinfo* soinfo_alloc(const char* name) {
310  if (strlen(name) >= SOINFO_NAME_LEN) {
311    DL_ERR("library name \"%s\" too long", name);
312    return NULL;
313  }
314
315  if (!ensure_free_list_non_empty()) {
316    DL_ERR("out of memory when loading \"%s\"", name);
317    return NULL;
318  }
319
320  // Take the head element off the free list.
321  soinfo* si = gSoInfoFreeList;
322  gSoInfoFreeList = gSoInfoFreeList->next;
323
324  // Initialize the new element.
325  memset(si, 0, sizeof(soinfo));
326  strlcpy(si->name, name, sizeof(si->name));
327  sonext->next = si;
328  sonext = si;
329
330  TRACE("name %s: allocated soinfo @ %p", name, si);
331  return si;
332}
333
334static void soinfo_free(soinfo* si) {
335    if (si == NULL) {
336        return;
337    }
338
339    soinfo *prev = NULL, *trav;
340
341    TRACE("name %s: freeing soinfo @ %p", si->name, si);
342
343    for (trav = solist; trav != NULL; trav = trav->next) {
344        if (trav == si)
345            break;
346        prev = trav;
347    }
348    if (trav == NULL) {
349        /* si was not in solist */
350        DL_ERR("name \"%s\" is not in solist!", si->name);
351        return;
352    }
353
354    /* prev will never be NULL, because the first entry in solist is
355       always the static libdl_info.
356    */
357    prev->next = si->next;
358    if (si == sonext) {
359        sonext = prev;
360    }
361    si->next = gSoInfoFreeList;
362    gSoInfoFreeList = si;
363}
364
365
366static void parse_path(const char* path, const char* delimiters,
367                       const char** array, char* buf, size_t buf_size, size_t max_count) {
368  if (path == NULL) {
369    return;
370  }
371
372  size_t len = strlcpy(buf, path, buf_size);
373
374  size_t i = 0;
375  char* buf_p = buf;
376  while (i < max_count && (array[i] = strsep(&buf_p, delimiters))) {
377    if (*array[i] != '\0') {
378      ++i;
379    }
380  }
381
382  // Forget the last path if we had to truncate; this occurs if the 2nd to
383  // last char isn't '\0' (i.e. wasn't originally a delimiter).
384  if (i > 0 && len >= buf_size && buf[buf_size - 2] != '\0') {
385    array[i - 1] = NULL;
386  } else {
387    array[i] = NULL;
388  }
389}
390
391static void parse_LD_LIBRARY_PATH(const char* path) {
392  parse_path(path, ":", gLdPaths,
393             gLdPathsBuffer, sizeof(gLdPathsBuffer), LDPATH_MAX);
394}
395
396static void parse_LD_PRELOAD(const char* path) {
397  // We have historically supported ':' as well as ' ' in LD_PRELOAD.
398  parse_path(path, " :", gLdPreloadNames,
399             gLdPreloadsBuffer, sizeof(gLdPreloadsBuffer), LDPRELOAD_MAX);
400}
401
402#if defined(__arm__)
403
404/* For a given PC, find the .so that it belongs to.
405 * Returns the base address of the .ARM.exidx section
406 * for that .so, and the number of 8-byte entries
407 * in that section (via *pcount).
408 *
409 * Intended to be called by libc's __gnu_Unwind_Find_exidx().
410 *
411 * This function is exposed via dlfcn.cpp and libdl.so.
412 */
413_Unwind_Ptr dl_unwind_find_exidx(_Unwind_Ptr pc, int* pcount) {
414    unsigned addr = (unsigned)pc;
415
416    for (soinfo* si = solist; si != 0; si = si->next) {
417        if ((addr >= si->base) && (addr < (si->base + si->size))) {
418            *pcount = si->ARM_exidx_count;
419            return (_Unwind_Ptr)si->ARM_exidx;
420        }
421    }
422    *pcount = 0;
423    return NULL;
424}
425
426#endif
427
428/* Here, we only have to provide a callback to iterate across all the
429 * loaded libraries. gcc_eh does the rest. */
430int dl_iterate_phdr(int (*cb)(dl_phdr_info* info, size_t size, void* data), void* data) {
431    int rv = 0;
432    for (soinfo* si = solist; si != NULL; si = si->next) {
433        dl_phdr_info dl_info;
434        dl_info.dlpi_addr = si->link_map_head.l_addr;
435        dl_info.dlpi_name = si->link_map_head.l_name;
436        dl_info.dlpi_phdr = si->phdr;
437        dl_info.dlpi_phnum = si->phnum;
438        rv = cb(&dl_info, sizeof(dl_phdr_info), data);
439        if (rv != 0) {
440            break;
441        }
442    }
443    return rv;
444}
445
446static ElfW(Sym)* soinfo_elf_lookup(soinfo* si, unsigned hash, const char* name) {
447  ElfW(Sym)* symtab = si->symtab;
448  const char* strtab = si->strtab;
449
450  TRACE_TYPE(LOOKUP, "SEARCH %s in %s@%p %x %zd",
451             name, si->name, reinterpret_cast<void*>(si->base), hash, hash % si->nbucket);
452
453  for (unsigned n = si->bucket[hash % si->nbucket]; n != 0; n = si->chain[n]) {
454    ElfW(Sym)* s = symtab + n;
455    if (strcmp(strtab + s->st_name, name)) continue;
456
457    /* only concern ourselves with global and weak symbol definitions */
458    switch (ELF_ST_BIND(s->st_info)) {
459      case STB_GLOBAL:
460      case STB_WEAK:
461        if (s->st_shndx == SHN_UNDEF) {
462        continue;
463      }
464
465      TRACE_TYPE(LOOKUP, "FOUND %s in %s (%p) %zd",
466                 name, si->name, reinterpret_cast<void*>(s->st_value),
467                 static_cast<size_t>(s->st_size));
468      return s;
469    }
470  }
471
472  return NULL;
473}
474
475static unsigned elfhash(const char* _name) {
476    const unsigned char* name = reinterpret_cast<const unsigned char*>(_name);
477    unsigned h = 0, g;
478
479    while (*name) {
480        h = (h << 4) + *name++;
481        g = h & 0xf0000000;
482        h ^= g;
483        h ^= g >> 24;
484    }
485    return h;
486}
487
488static ElfW(Sym)* soinfo_do_lookup(soinfo* si, const char* name, soinfo** lsi, soinfo* needed[]) {
489    unsigned elf_hash = elfhash(name);
490    ElfW(Sym)* s = NULL;
491
492    if (si != NULL && somain != NULL) {
493        /*
494         * Local scope is executable scope. Just start looking into it right away
495         * for the shortcut.
496         */
497
498        if (si == somain) {
499            s = soinfo_elf_lookup(si, elf_hash, name);
500            if (s != NULL) {
501                *lsi = si;
502                goto done;
503            }
504        } else {
505            /* Order of symbol lookup is controlled by DT_SYMBOLIC flag */
506
507            /*
508             * If this object was built with symbolic relocations disabled, the
509             * first place to look to resolve external references is the main
510             * executable.
511             */
512
513            if (!si->has_DT_SYMBOLIC) {
514                DEBUG("%s: looking up %s in executable %s",
515                      si->name, name, somain->name);
516                s = soinfo_elf_lookup(somain, elf_hash, name);
517                if (s != NULL) {
518                    *lsi = somain;
519                    goto done;
520                }
521            }
522
523            /* Look for symbols in the local scope (the object who is
524             * searching). This happens with C++ templates on x86 for some
525             * reason.
526             *
527             * Notes on weak symbols:
528             * The ELF specs are ambiguous about treatment of weak definitions in
529             * dynamic linking.  Some systems return the first definition found
530             * and some the first non-weak definition.   This is system dependent.
531             * Here we return the first definition found for simplicity.  */
532
533            s = soinfo_elf_lookup(si, elf_hash, name);
534            if (s != NULL) {
535                *lsi = si;
536                goto done;
537            }
538
539            /*
540             * If this object was built with -Bsymbolic and symbol is not found
541             * in the local scope, try to find the symbol in the main executable.
542             */
543
544            if (si->has_DT_SYMBOLIC) {
545                DEBUG("%s: looking up %s in executable %s after local scope",
546                      si->name, name, somain->name);
547                s = soinfo_elf_lookup(somain, elf_hash, name);
548                if (s != NULL) {
549                    *lsi = somain;
550                    goto done;
551                }
552            }
553        }
554    }
555
556    /* Next, look for it in the preloads list */
557    for (int i = 0; gLdPreloads[i] != NULL; i++) {
558        s = soinfo_elf_lookup(gLdPreloads[i], elf_hash, name);
559        if (s != NULL) {
560            *lsi = gLdPreloads[i];
561            goto done;
562        }
563    }
564
565    for (int i = 0; needed[i] != NULL; i++) {
566        DEBUG("%s: looking up %s in %s",
567              si->name, name, needed[i]->name);
568        s = soinfo_elf_lookup(needed[i], elf_hash, name);
569        if (s != NULL) {
570            *lsi = needed[i];
571            goto done;
572        }
573    }
574
575done:
576    if (s != NULL) {
577        TRACE_TYPE(LOOKUP, "si %s sym %s s->st_value = %p, "
578                   "found in %s, base = %p, load bias = %p",
579                   si->name, name, reinterpret_cast<void*>(s->st_value),
580                   (*lsi)->name, reinterpret_cast<void*>((*lsi)->base),
581                   reinterpret_cast<void*>((*lsi)->load_bias));
582        return s;
583    }
584
585    return NULL;
586}
587
588/* This is used by dlsym(3).  It performs symbol lookup only within the
589   specified soinfo object and not in any of its dependencies.
590
591   TODO: Only looking in the specified soinfo seems wrong. dlsym(3) says
592   that it should do a breadth first search through the dependency
593   tree. This agrees with the ELF spec (aka System V Application
594   Binary Interface) where in Chapter 5 it discuss resolving "Shared
595   Object Dependencies" in breadth first search order.
596 */
597ElfW(Sym)* dlsym_handle_lookup(soinfo* si, const char* name) {
598    return soinfo_elf_lookup(si, elfhash(name), name);
599}
600
601/* This is used by dlsym(3) to performs a global symbol lookup. If the
602   start value is null (for RTLD_DEFAULT), the search starts at the
603   beginning of the global solist. Otherwise the search starts at the
604   specified soinfo (for RTLD_NEXT).
605 */
606ElfW(Sym)* dlsym_linear_lookup(const char* name, soinfo** found, soinfo* start) {
607  unsigned elf_hash = elfhash(name);
608
609  if (start == NULL) {
610    start = solist;
611  }
612
613  ElfW(Sym)* s = NULL;
614  for (soinfo* si = start; (s == NULL) && (si != NULL); si = si->next) {
615    s = soinfo_elf_lookup(si, elf_hash, name);
616    if (s != NULL) {
617      *found = si;
618      break;
619    }
620  }
621
622  if (s != NULL) {
623    TRACE_TYPE(LOOKUP, "%s s->st_value = %p, found->base = %p",
624               name, reinterpret_cast<void*>(s->st_value), reinterpret_cast<void*>((*found)->base));
625  }
626
627  return s;
628}
629
630soinfo* find_containing_library(const void* p) {
631  ElfW(Addr) address = reinterpret_cast<ElfW(Addr)>(p);
632  for (soinfo* si = solist; si != NULL; si = si->next) {
633    if (address >= si->base && address - si->base < si->size) {
634      return si;
635    }
636  }
637  return NULL;
638}
639
640ElfW(Sym)* dladdr_find_symbol(soinfo* si, const void* addr) {
641  ElfW(Addr) soaddr = reinterpret_cast<ElfW(Addr)>(addr) - si->base;
642
643  // Search the library's symbol table for any defined symbol which
644  // contains this address.
645  for (size_t i = 0; i < si->nchain; ++i) {
646    ElfW(Sym)* sym = &si->symtab[i];
647    if (sym->st_shndx != SHN_UNDEF &&
648        soaddr >= sym->st_value &&
649        soaddr < sym->st_value + sym->st_size) {
650      return sym;
651    }
652  }
653
654  return NULL;
655}
656
657static int open_library_on_path(const char* name, const char* const paths[]) {
658  char buf[512];
659  for (size_t i = 0; paths[i] != NULL; ++i) {
660    int n = __libc_format_buffer(buf, sizeof(buf), "%s/%s", paths[i], name);
661    if (n < 0 || n >= static_cast<int>(sizeof(buf))) {
662      PRINT("Warning: ignoring very long library path: %s/%s", paths[i], name);
663      continue;
664    }
665    int fd = TEMP_FAILURE_RETRY(open(buf, O_RDONLY | O_CLOEXEC));
666    if (fd != -1) {
667      return fd;
668    }
669  }
670  return -1;
671}
672
673static int open_library(const char* name) {
674  TRACE("[ opening %s ]", name);
675
676  // If the name contains a slash, we should attempt to open it directly and not search the paths.
677  if (strchr(name, '/') != NULL) {
678    int fd = TEMP_FAILURE_RETRY(open(name, O_RDONLY | O_CLOEXEC));
679    if (fd != -1) {
680      return fd;
681    }
682    // ...but nvidia binary blobs (at least) rely on this behavior, so fall through for now.
683  }
684
685  // Otherwise we try LD_LIBRARY_PATH first, and fall back to the built-in well known paths.
686  int fd = open_library_on_path(name, gLdPaths);
687  if (fd == -1) {
688    fd = open_library_on_path(name, gDefaultLdPaths);
689  }
690  return fd;
691}
692
693static soinfo* load_library(const char* name) {
694    // Open the file.
695    int fd = open_library(name);
696    if (fd == -1) {
697        DL_ERR("library \"%s\" not found", name);
698        return NULL;
699    }
700
701    // Read the ELF header and load the segments.
702    ElfReader elf_reader(name, fd);
703    if (!elf_reader.Load()) {
704        return NULL;
705    }
706
707    const char* bname = strrchr(name, '/');
708    soinfo* si = soinfo_alloc(bname ? bname + 1 : name);
709    if (si == NULL) {
710        return NULL;
711    }
712    si->base = elf_reader.load_start();
713    si->size = elf_reader.load_size();
714    si->load_bias = elf_reader.load_bias();
715    si->flags = 0;
716    si->entry = 0;
717    si->dynamic = NULL;
718    si->phnum = elf_reader.phdr_count();
719    si->phdr = elf_reader.loaded_phdr();
720    return si;
721}
722
723static soinfo *find_loaded_library(const char* name) {
724    // TODO: don't use basename only for determining libraries
725    // http://code.google.com/p/android/issues/detail?id=6670
726
727    const char* bname = strrchr(name, '/');
728    bname = bname ? bname + 1 : name;
729
730    for (soinfo* si = solist; si != NULL; si = si->next) {
731        if (!strcmp(bname, si->name)) {
732            return si;
733        }
734    }
735    return NULL;
736}
737
738static soinfo* find_library_internal(const char* name) {
739  if (name == NULL) {
740    return somain;
741  }
742
743  soinfo* si = find_loaded_library(name);
744  if (si != NULL) {
745    if (si->flags & FLAG_LINKED) {
746      return si;
747    }
748    DL_ERR("OOPS: recursive link to \"%s\"", si->name);
749    return NULL;
750  }
751
752  TRACE("[ '%s' has not been loaded yet.  Locating...]", name);
753  si = load_library(name);
754  if (si == NULL) {
755    return NULL;
756  }
757
758  // At this point we know that whatever is loaded @ base is a valid ELF
759  // shared library whose segments are properly mapped in.
760  TRACE("[ find_library_internal base=%p size=%zu name='%s' ]",
761        reinterpret_cast<void*>(si->base), si->size, si->name);
762
763  if (!soinfo_link_image(si)) {
764    munmap(reinterpret_cast<void*>(si->base), si->size);
765    soinfo_free(si);
766    return NULL;
767  }
768
769  return si;
770}
771
772static soinfo* find_library(const char* name) {
773  soinfo* si = find_library_internal(name);
774  if (si != NULL) {
775    si->ref_count++;
776  }
777  return si;
778}
779
780static int soinfo_unload(soinfo* si) {
781  if (si->ref_count == 1) {
782    TRACE("unloading '%s'", si->name);
783    si->CallDestructors();
784
785    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
786      if (d->d_tag == DT_NEEDED) {
787        const char* library_name = si->strtab + d->d_un.d_val;
788        TRACE("%s needs to unload %s", si->name, library_name);
789        soinfo_unload(find_loaded_library(library_name));
790      }
791    }
792
793    munmap(reinterpret_cast<void*>(si->base), si->size);
794    notify_gdb_of_unload(si);
795    soinfo_free(si);
796    si->ref_count = 0;
797  } else {
798    si->ref_count--;
799    TRACE("not unloading '%s', decrementing ref_count to %zd", si->name, si->ref_count);
800  }
801  return 0;
802}
803
804void do_android_get_LD_LIBRARY_PATH(char* buffer, size_t buffer_size) {
805  snprintf(buffer, buffer_size, "%s:%s", gDefaultLdPaths[0], gDefaultLdPaths[1]);
806}
807
808void do_android_update_LD_LIBRARY_PATH(const char* ld_library_path) {
809  if (!get_AT_SECURE()) {
810    parse_LD_LIBRARY_PATH(ld_library_path);
811  }
812}
813
814soinfo* do_dlopen(const char* name, int flags) {
815  if ((flags & ~(RTLD_NOW|RTLD_LAZY|RTLD_LOCAL|RTLD_GLOBAL)) != 0) {
816    DL_ERR("invalid flags to dlopen: %x", flags);
817    return NULL;
818  }
819  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
820  soinfo* si = find_library(name);
821  if (si != NULL) {
822    si->CallConstructors();
823  }
824  set_soinfo_pool_protection(PROT_READ);
825  return si;
826}
827
828int do_dlclose(soinfo* si) {
829  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
830  int result = soinfo_unload(si);
831  set_soinfo_pool_protection(PROT_READ);
832  return result;
833}
834
835#if defined(USE_RELA)
836static int soinfo_relocate(soinfo* si, ElfW(Rela)* rela, unsigned count, soinfo* needed[]) {
837  ElfW(Sym)* s;
838  soinfo* lsi;
839
840  for (size_t idx = 0; idx < count; ++idx, ++rela) {
841    unsigned type = ELFW(R_TYPE)(rela->r_info);
842    unsigned sym = ELFW(R_SYM)(rela->r_info);
843    ElfW(Addr) reloc = static_cast<ElfW(Addr)>(rela->r_offset + si->load_bias);
844    ElfW(Addr) sym_addr = 0;
845    const char* sym_name = NULL;
846
847    DEBUG("Processing '%s' relocation at index %zd", si->name, idx);
848    if (type == 0) { // R_*_NONE
849      continue;
850    }
851    if (sym != 0) {
852      sym_name = reinterpret_cast<const char*>(si->strtab + si->symtab[sym].st_name);
853      s = soinfo_do_lookup(si, sym_name, &lsi, needed);
854      if (s == NULL) {
855        // We only allow an undefined symbol if this is a weak reference...
856        s = &si->symtab[sym];
857        if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
858          DL_ERR("cannot locate symbol \"%s\" referenced by \"%s\"...", sym_name, si->name);
859          return -1;
860        }
861
862        /* IHI0044C AAELF 4.5.1.1:
863
864           Libraries are not searched to resolve weak references.
865           It is not an error for a weak reference to remain unsatisfied.
866
867           During linking, the value of an undefined weak reference is:
868           - Zero if the relocation type is absolute
869           - The address of the place if the relocation is pc-relative
870           - The address of nominal base address if the relocation
871             type is base-relative.
872         */
873
874        switch (type) {
875#if defined(__aarch64__)
876        case R_AARCH64_JUMP_SLOT:
877        case R_AARCH64_GLOB_DAT:
878        case R_AARCH64_ABS64:
879        case R_AARCH64_ABS32:
880        case R_AARCH64_ABS16:
881        case R_AARCH64_RELATIVE:
882          /*
883           * The sym_addr was initialized to be zero above, or the relocation
884           * code below does not care about value of sym_addr.
885           * No need to do anything.
886           */
887          break;
888#elif defined(__x86_64__)
889        case R_X86_64_JUMP_SLOT:
890        case R_X86_64_GLOB_DAT:
891        case R_X86_64_32:
892        case R_X86_64_RELATIVE:
893          // No need to do anything.
894          break;
895        case R_X86_64_PC32:
896          sym_addr = reloc;
897          break;
898#endif
899        default:
900          DL_ERR("unknown weak reloc type %d @ %p (%zu)", type, rela, idx);
901          return -1;
902        }
903      } else {
904        // We got a definition.
905        sym_addr = static_cast<ElfW(Addr)>(s->st_value + lsi->load_bias);
906      }
907      count_relocation(kRelocSymbol);
908    } else {
909      s = NULL;
910    }
911
912    switch (type) {
913#if defined(__aarch64__)
914    case R_AARCH64_JUMP_SLOT:
915        count_relocation(kRelocAbsolute);
916        MARK(rela->r_offset);
917        TRACE_TYPE(RELO, "RELO JMP_SLOT %16llx <- %16llx %s\n",
918                   reloc, (sym_addr + rela->r_addend), sym_name);
919        *reinterpret_cast<ElfW(Addr)*>(reloc) = (sym_addr + rela->r_addend);
920        break;
921    case R_AARCH64_GLOB_DAT:
922        count_relocation(kRelocAbsolute);
923        MARK(rela->r_offset);
924        TRACE_TYPE(RELO, "RELO GLOB_DAT %16llx <- %16llx %s\n",
925                   reloc, (sym_addr + rela->r_addend), sym_name);
926        *reinterpret_cast<ElfW(Addr)*>(reloc) = (sym_addr + rela->r_addend);
927        break;
928    case R_AARCH64_ABS64:
929        count_relocation(kRelocAbsolute);
930        MARK(rela->r_offset);
931        TRACE_TYPE(RELO, "RELO ABS64 %16llx <- %16llx %s\n",
932                   reloc, (sym_addr + rela->r_addend), sym_name);
933        *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
934        break;
935    case R_AARCH64_ABS32:
936        count_relocation(kRelocAbsolute);
937        MARK(rela->r_offset);
938        TRACE_TYPE(RELO, "RELO ABS32 %16llx <- %16llx %s\n",
939                   reloc, (sym_addr + rela->r_addend), sym_name);
940        if ((static_cast<ElfW(Addr)>(INT32_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend))) &&
941            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)) <= static_cast<ElfW(Addr)>(UINT32_MAX))) {
942            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
943        } else {
944            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
945                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)),
946                   static_cast<ElfW(Addr)>(INT32_MIN),
947                   static_cast<ElfW(Addr)>(UINT32_MAX));
948            return -1;
949        }
950        break;
951    case R_AARCH64_ABS16:
952        count_relocation(kRelocAbsolute);
953        MARK(rela->r_offset);
954        TRACE_TYPE(RELO, "RELO ABS16 %16llx <- %16llx %s\n",
955                   reloc, (sym_addr + rela->r_addend), sym_name);
956        if ((static_cast<ElfW(Addr)>(INT16_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend))) &&
957            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)) <= static_cast<ElfW(Addr)>(UINT16_MAX))) {
958            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
959        } else {
960            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
961                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)),
962                   static_cast<ElfW(Addr)>(INT16_MIN),
963                   static_cast<ElfW(Addr)>(UINT16_MAX));
964            return -1;
965        }
966        break;
967    case R_AARCH64_PREL64:
968        count_relocation(kRelocRelative);
969        MARK(rela->r_offset);
970        TRACE_TYPE(RELO, "RELO REL64 %16llx <- %16llx - %16llx %s\n",
971                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
972        *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend) - rela->r_offset;
973        break;
974    case R_AARCH64_PREL32:
975        count_relocation(kRelocRelative);
976        MARK(rela->r_offset);
977        TRACE_TYPE(RELO, "RELO REL32 %16llx <- %16llx - %16llx %s\n",
978                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
979        if ((static_cast<ElfW(Addr)>(INT32_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset))) &&
980            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)) <= static_cast<ElfW(Addr)>(UINT32_MAX))) {
981            *reinterpret_cast<ElfW(Addr)*>(reloc) += ((sym_addr + rela->r_addend) - rela->r_offset);
982        } else {
983            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
984                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)),
985                   static_cast<ElfW(Addr)>(INT32_MIN),
986                   static_cast<ElfW(Addr)>(UINT32_MAX));
987            return -1;
988        }
989        break;
990    case R_AARCH64_PREL16:
991        count_relocation(kRelocRelative);
992        MARK(rela->r_offset);
993        TRACE_TYPE(RELO, "RELO REL16 %16llx <- %16llx - %16llx %s\n",
994                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
995        if ((static_cast<ElfW(Addr)>(INT16_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset))) &&
996            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)) <= static_cast<ElfW(Addr)>(UINT16_MAX))) {
997            *reinterpret_cast<ElfW(Addr)*>(reloc) += ((sym_addr + rela->r_addend) - rela->r_offset);
998        } else {
999            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
1000                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)),
1001                   static_cast<ElfW(Addr)>(INT16_MIN),
1002                   static_cast<ElfW(Addr)>(UINT16_MAX));
1003            return -1;
1004        }
1005        break;
1006
1007    case R_AARCH64_RELATIVE:
1008        count_relocation(kRelocRelative);
1009        MARK(rela->r_offset);
1010        if (sym) {
1011            DL_ERR("odd RELATIVE form...");
1012            return -1;
1013        }
1014        TRACE_TYPE(RELO, "RELO RELATIVE %16llx <- %16llx\n",
1015                   reloc, (si->base + rela->r_addend));
1016        *reinterpret_cast<ElfW(Addr)*>(reloc) = (si->base + rela->r_addend);
1017        break;
1018
1019    case R_AARCH64_COPY:
1020        if ((si->flags & FLAG_EXE) == 0) {
1021            /*
1022              * http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044d/IHI0044D_aaelf.pdf
1023              *
1024              * Section 4.7.1.10 "Dynamic relocations"
1025              * R_AARCH64_COPY may only appear in executable objects where e_type is
1026              * set to ET_EXEC.
1027              *
1028              * FLAG_EXE is set for both ET_DYN and ET_EXEC executables.
1029              * We should explicitly disallow ET_DYN executables from having
1030              * R_AARCH64_COPY relocations.
1031              */
1032            DL_ERR("%s R_AARCH64_COPY relocations only supported for ET_EXEC", si->name);
1033            return -1;
1034        }
1035        count_relocation(kRelocCopy);
1036        MARK(rela->r_offset);
1037        TRACE_TYPE(RELO, "RELO COPY %16llx <- %lld @ %16llx %s\n",
1038                   reloc,
1039                   s->st_size,
1040                   (sym_addr + rela->r_addend),
1041                   sym_name);
1042        if (reloc == (sym_addr + rela->r_addend)) {
1043            ElfW(Sym)* src = soinfo_do_lookup(NULL, sym_name, &lsi, needed);
1044
1045            if (src == NULL) {
1046                DL_ERR("%s R_AARCH64_COPY relocation source cannot be resolved", si->name);
1047                return -1;
1048            }
1049            if (lsi->has_DT_SYMBOLIC) {
1050                DL_ERR("%s invalid R_AARCH64_COPY relocation against DT_SYMBOLIC shared "
1051                       "library %s (built with -Bsymbolic?)", si->name, lsi->name);
1052                return -1;
1053            }
1054            if (s->st_size < src->st_size) {
1055                DL_ERR("%s R_AARCH64_COPY relocation size mismatch (%lld < %lld)",
1056                       si->name, s->st_size, src->st_size);
1057                return -1;
1058            }
1059            memcpy(reinterpret_cast<void*>(reloc),
1060                   reinterpret_cast<void*>(src->st_value + lsi->load_bias), src->st_size);
1061        } else {
1062            DL_ERR("%s R_AARCH64_COPY relocation target cannot be resolved", si->name);
1063            return -1;
1064        }
1065        break;
1066    case R_AARCH64_TLS_TPREL64:
1067        TRACE_TYPE(RELO, "RELO TLS_TPREL64 *** %16llx <- %16llx - %16llx\n",
1068                   reloc, (sym_addr + rela->r_addend), rela->r_offset);
1069        break;
1070    case R_AARCH64_TLS_DTPREL32:
1071        TRACE_TYPE(RELO, "RELO TLS_DTPREL32 *** %16llx <- %16llx - %16llx\n",
1072                   reloc, (sym_addr + rela->r_addend), rela->r_offset);
1073        break;
1074#elif defined(__x86_64__)
1075    case R_X86_64_JUMP_SLOT:
1076      count_relocation(kRelocAbsolute);
1077      MARK(rela->r_offset);
1078      TRACE_TYPE(RELO, "RELO JMP_SLOT %08zx <- %08zx %s", static_cast<size_t>(reloc),
1079                 static_cast<size_t>(sym_addr + rela->r_addend), sym_name);
1080      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1081      break;
1082    case R_X86_64_GLOB_DAT:
1083      count_relocation(kRelocAbsolute);
1084      MARK(rela->r_offset);
1085      TRACE_TYPE(RELO, "RELO GLOB_DAT %08zx <- %08zx %s", static_cast<size_t>(reloc),
1086                 static_cast<size_t>(sym_addr + rela->r_addend), sym_name);
1087      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1088      break;
1089    case R_X86_64_RELATIVE:
1090      count_relocation(kRelocRelative);
1091      MARK(rela->r_offset);
1092      if (sym) {
1093        DL_ERR("odd RELATIVE form...");
1094        return -1;
1095      }
1096      TRACE_TYPE(RELO, "RELO RELATIVE %08zx <- +%08zx", static_cast<size_t>(reloc),
1097                 static_cast<size_t>(si->base));
1098      *reinterpret_cast<ElfW(Addr)*>(reloc) = si->base + rela->r_addend;
1099      break;
1100    case R_X86_64_32:
1101      count_relocation(kRelocRelative);
1102      MARK(rela->r_offset);
1103      TRACE_TYPE(RELO, "RELO R_X86_64_32 %08zx <- +%08zx %s", static_cast<size_t>(reloc),
1104                 static_cast<size_t>(sym_addr), sym_name);
1105      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1106      break;
1107    case R_X86_64_64:
1108      count_relocation(kRelocRelative);
1109      MARK(rela->r_offset);
1110      TRACE_TYPE(RELO, "RELO R_X86_64_64 %08zx <- +%08zx %s", static_cast<size_t>(reloc),
1111                 static_cast<size_t>(sym_addr), sym_name);
1112      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1113      break;
1114    case R_X86_64_PC32:
1115      count_relocation(kRelocRelative);
1116      MARK(rela->r_offset);
1117      TRACE_TYPE(RELO, "RELO R_X86_64_PC32 %08zx <- +%08zx (%08zx - %08zx) %s",
1118                 static_cast<size_t>(reloc), static_cast<size_t>(sym_addr - reloc),
1119                 static_cast<size_t>(sym_addr), static_cast<size_t>(reloc), sym_name);
1120      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend - reloc;
1121      break;
1122#endif
1123
1124    default:
1125      DL_ERR("unknown reloc type %d @ %p (%zu)", type, rela, idx);
1126      return -1;
1127    }
1128  }
1129  return 0;
1130}
1131
1132#else // REL, not RELA.
1133
1134static int soinfo_relocate(soinfo* si, ElfW(Rel)* rel, unsigned count, soinfo* needed[]) {
1135    ElfW(Sym)* s;
1136    soinfo* lsi;
1137
1138    for (size_t idx = 0; idx < count; ++idx, ++rel) {
1139        unsigned type = ELFW(R_TYPE)(rel->r_info);
1140        // TODO: don't use unsigned for 'sym'. Use uint32_t or ElfW(Addr) instead.
1141        unsigned sym = ELFW(R_SYM)(rel->r_info);
1142        ElfW(Addr) reloc = static_cast<ElfW(Addr)>(rel->r_offset + si->load_bias);
1143        ElfW(Addr) sym_addr = 0;
1144        const char* sym_name = NULL;
1145
1146        DEBUG("Processing '%s' relocation at index %zd", si->name, idx);
1147        if (type == 0) { // R_*_NONE
1148            continue;
1149        }
1150        if (sym != 0) {
1151            sym_name = reinterpret_cast<const char*>(si->strtab + si->symtab[sym].st_name);
1152            s = soinfo_do_lookup(si, sym_name, &lsi, needed);
1153            if (s == NULL) {
1154                // We only allow an undefined symbol if this is a weak reference...
1155                s = &si->symtab[sym];
1156                if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
1157                    DL_ERR("cannot locate symbol \"%s\" referenced by \"%s\"...", sym_name, si->name);
1158                    return -1;
1159                }
1160
1161                /* IHI0044C AAELF 4.5.1.1:
1162
1163                   Libraries are not searched to resolve weak references.
1164                   It is not an error for a weak reference to remain
1165                   unsatisfied.
1166
1167                   During linking, the value of an undefined weak reference is:
1168                   - Zero if the relocation type is absolute
1169                   - The address of the place if the relocation is pc-relative
1170                   - The address of nominal base address if the relocation
1171                     type is base-relative.
1172                  */
1173
1174                switch (type) {
1175#if defined(__arm__)
1176                case R_ARM_JUMP_SLOT:
1177                case R_ARM_GLOB_DAT:
1178                case R_ARM_ABS32:
1179                case R_ARM_RELATIVE:    /* Don't care. */
1180                    // sym_addr was initialized to be zero above or relocation
1181                    // code below does not care about value of sym_addr.
1182                    // No need to do anything.
1183                    break;
1184#elif defined(__i386__)
1185                case R_386_JMP_SLOT:
1186                case R_386_GLOB_DAT:
1187                case R_386_32:
1188                case R_386_RELATIVE:    /* Don't care. */
1189                    // sym_addr was initialized to be zero above or relocation
1190                    // code below does not care about value of sym_addr.
1191                    // No need to do anything.
1192                    break;
1193                case R_386_PC32:
1194                    sym_addr = reloc;
1195                    break;
1196#endif
1197
1198#if defined(__arm__)
1199                case R_ARM_COPY:
1200                    // Fall through. Can't really copy if weak symbol is not found at run-time.
1201#endif
1202                default:
1203                    DL_ERR("unknown weak reloc type %d @ %p (%zu)", type, rel, idx);
1204                    return -1;
1205                }
1206            } else {
1207                // We got a definition.
1208                sym_addr = static_cast<ElfW(Addr)>(s->st_value + lsi->load_bias);
1209            }
1210            count_relocation(kRelocSymbol);
1211        } else {
1212            s = NULL;
1213        }
1214
1215        switch (type) {
1216#if defined(__arm__)
1217        case R_ARM_JUMP_SLOT:
1218            count_relocation(kRelocAbsolute);
1219            MARK(rel->r_offset);
1220            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
1221            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1222            break;
1223        case R_ARM_GLOB_DAT:
1224            count_relocation(kRelocAbsolute);
1225            MARK(rel->r_offset);
1226            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
1227            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1228            break;
1229        case R_ARM_ABS32:
1230            count_relocation(kRelocAbsolute);
1231            MARK(rel->r_offset);
1232            TRACE_TYPE(RELO, "RELO ABS %08x <- %08x %s", reloc, sym_addr, sym_name);
1233            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1234            break;
1235        case R_ARM_REL32:
1236            count_relocation(kRelocRelative);
1237            MARK(rel->r_offset);
1238            TRACE_TYPE(RELO, "RELO REL32 %08x <- %08x - %08x %s",
1239                       reloc, sym_addr, rel->r_offset, sym_name);
1240            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr - rel->r_offset;
1241            break;
1242        case R_ARM_COPY:
1243            if ((si->flags & FLAG_EXE) == 0) {
1244                /*
1245                 * http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044d/IHI0044D_aaelf.pdf
1246                 *
1247                 * Section 4.7.1.10 "Dynamic relocations"
1248                 * R_ARM_COPY may only appear in executable objects where e_type is
1249                 * set to ET_EXEC.
1250                 *
1251                 * TODO: FLAG_EXE is set for both ET_DYN and ET_EXEC executables.
1252                 * We should explicitly disallow ET_DYN executables from having
1253                 * R_ARM_COPY relocations.
1254                 */
1255                DL_ERR("%s R_ARM_COPY relocations only supported for ET_EXEC", si->name);
1256                return -1;
1257            }
1258            count_relocation(kRelocCopy);
1259            MARK(rel->r_offset);
1260            TRACE_TYPE(RELO, "RELO %08x <- %d @ %08x %s", reloc, s->st_size, sym_addr, sym_name);
1261            if (reloc == sym_addr) {
1262                ElfW(Sym)* src = soinfo_do_lookup(NULL, sym_name, &lsi, needed);
1263
1264                if (src == NULL) {
1265                    DL_ERR("%s R_ARM_COPY relocation source cannot be resolved", si->name);
1266                    return -1;
1267                }
1268                if (lsi->has_DT_SYMBOLIC) {
1269                    DL_ERR("%s invalid R_ARM_COPY relocation against DT_SYMBOLIC shared "
1270                           "library %s (built with -Bsymbolic?)", si->name, lsi->name);
1271                    return -1;
1272                }
1273                if (s->st_size < src->st_size) {
1274                    DL_ERR("%s R_ARM_COPY relocation size mismatch (%d < %d)",
1275                           si->name, s->st_size, src->st_size);
1276                    return -1;
1277                }
1278                memcpy(reinterpret_cast<void*>(reloc),
1279                       reinterpret_cast<void*>(src->st_value + lsi->load_bias), src->st_size);
1280            } else {
1281                DL_ERR("%s R_ARM_COPY relocation target cannot be resolved", si->name);
1282                return -1;
1283            }
1284            break;
1285#elif defined(__i386__)
1286        case R_386_JMP_SLOT:
1287            count_relocation(kRelocAbsolute);
1288            MARK(rel->r_offset);
1289            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
1290            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1291            break;
1292        case R_386_GLOB_DAT:
1293            count_relocation(kRelocAbsolute);
1294            MARK(rel->r_offset);
1295            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
1296            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1297            break;
1298        case R_386_32:
1299            count_relocation(kRelocRelative);
1300            MARK(rel->r_offset);
1301            TRACE_TYPE(RELO, "RELO R_386_32 %08x <- +%08x %s", reloc, sym_addr, sym_name);
1302            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1303            break;
1304        case R_386_PC32:
1305            count_relocation(kRelocRelative);
1306            MARK(rel->r_offset);
1307            TRACE_TYPE(RELO, "RELO R_386_PC32 %08x <- +%08x (%08x - %08x) %s",
1308                       reloc, (sym_addr - reloc), sym_addr, reloc, sym_name);
1309            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr - reloc);
1310            break;
1311#elif defined(__mips__)
1312        case R_MIPS_REL32:
1313#if defined(__LP64__)
1314            // MIPS Elf64_Rel entries contain compound relocations
1315            // We only handle the R_MIPS_NONE|R_MIPS_64|R_MIPS_REL32 case
1316            if (ELF64_R_TYPE2(rel->r_info) != R_MIPS_64 ||
1317                ELF64_R_TYPE3(rel->r_info) != R_MIPS_NONE) {
1318                DL_ERR("Unexpected compound relocation type:%d type2:%d type3:%d @ %p (%zu)",
1319                       type, (unsigned)ELF64_R_TYPE2(rel->r_info),
1320                       (unsigned)ELF64_R_TYPE3(rel->r_info), rel, idx);
1321                return -1;
1322            }
1323#endif
1324            count_relocation(kRelocAbsolute);
1325            MARK(rel->r_offset);
1326            TRACE_TYPE(RELO, "RELO REL32 %08zx <- %08zx %s", static_cast<size_t>(reloc),
1327                       static_cast<size_t>(sym_addr), sym_name ? sym_name : "*SECTIONHDR*");
1328            if (s) {
1329                *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1330            } else {
1331                *reinterpret_cast<ElfW(Addr)*>(reloc) += si->base;
1332            }
1333            break;
1334#endif
1335
1336#if defined(__arm__)
1337        case R_ARM_RELATIVE:
1338#elif defined(__i386__)
1339        case R_386_RELATIVE:
1340#endif
1341            count_relocation(kRelocRelative);
1342            MARK(rel->r_offset);
1343            if (sym) {
1344                DL_ERR("odd RELATIVE form...");
1345                return -1;
1346            }
1347            TRACE_TYPE(RELO, "RELO RELATIVE %p <- +%p",
1348                       reinterpret_cast<void*>(reloc), reinterpret_cast<void*>(si->base));
1349            *reinterpret_cast<ElfW(Addr)*>(reloc) += si->base;
1350            break;
1351
1352        default:
1353            DL_ERR("unknown reloc type %d @ %p (%zu)", type, rel, idx);
1354            return -1;
1355        }
1356    }
1357    return 0;
1358}
1359#endif
1360
1361#if defined(__mips__)
1362static bool mips_relocate_got(soinfo* si, soinfo* needed[]) {
1363    ElfW(Addr)** got = si->plt_got;
1364    if (got == NULL) {
1365        return true;
1366    }
1367    unsigned local_gotno = si->mips_local_gotno;
1368    unsigned gotsym = si->mips_gotsym;
1369    unsigned symtabno = si->mips_symtabno;
1370    ElfW(Sym)* symtab = si->symtab;
1371
1372    // got[0] is the address of the lazy resolver function.
1373    // got[1] may be used for a GNU extension.
1374    // Set it to a recognizable address in case someone calls it (should be _rtld_bind_start).
1375    // FIXME: maybe this should be in a separate routine?
1376    if ((si->flags & FLAG_LINKER) == 0) {
1377        size_t g = 0;
1378        got[g++] = reinterpret_cast<ElfW(Addr)*>(0xdeadbeef);
1379        if (reinterpret_cast<intptr_t>(got[g]) < 0) {
1380            got[g++] = reinterpret_cast<ElfW(Addr)*>(0xdeadfeed);
1381        }
1382        // Relocate the local GOT entries.
1383        for (; g < local_gotno; g++) {
1384            got[g] = reinterpret_cast<ElfW(Addr)*>(reinterpret_cast<uintptr_t>(got[g]) + si->load_bias);
1385        }
1386    }
1387
1388    // Now for the global GOT entries...
1389    ElfW(Sym)* sym = symtab + gotsym;
1390    got = si->plt_got + local_gotno;
1391    for (size_t g = gotsym; g < symtabno; g++, sym++, got++) {
1392        // This is an undefined reference... try to locate it.
1393        const char* sym_name = si->strtab + sym->st_name;
1394        soinfo* lsi;
1395        ElfW(Sym)* s = soinfo_do_lookup(si, sym_name, &lsi, needed);
1396        if (s == NULL) {
1397            // We only allow an undefined symbol if this is a weak reference.
1398            s = &symtab[g];
1399            if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
1400                DL_ERR("cannot locate \"%s\"...", sym_name);
1401                return false;
1402            }
1403            *got = 0;
1404        } else {
1405            // FIXME: is this sufficient?
1406            // For reference see NetBSD link loader
1407            // http://cvsweb.netbsd.org/bsdweb.cgi/src/libexec/ld.elf_so/arch/mips/mips_reloc.c?rev=1.53&content-type=text/x-cvsweb-markup
1408            *got = reinterpret_cast<ElfW(Addr)*>(lsi->load_bias + s->st_value);
1409        }
1410    }
1411    return true;
1412}
1413#endif
1414
1415void soinfo::CallArray(const char* array_name __unused, linker_function_t* functions, size_t count, bool reverse) {
1416  if (functions == NULL) {
1417    return;
1418  }
1419
1420  TRACE("[ Calling %s (size %zd) @ %p for '%s' ]", array_name, count, functions, name);
1421
1422  int begin = reverse ? (count - 1) : 0;
1423  int end = reverse ? -1 : count;
1424  int step = reverse ? -1 : 1;
1425
1426  for (int i = begin; i != end; i += step) {
1427    TRACE("[ %s[%d] == %p ]", array_name, i, functions[i]);
1428    CallFunction("function", functions[i]);
1429  }
1430
1431  TRACE("[ Done calling %s for '%s' ]", array_name, name);
1432}
1433
1434void soinfo::CallFunction(const char* function_name __unused, linker_function_t function) {
1435  if (function == NULL || reinterpret_cast<uintptr_t>(function) == static_cast<uintptr_t>(-1)) {
1436    return;
1437  }
1438
1439  TRACE("[ Calling %s @ %p for '%s' ]", function_name, function, name);
1440  function();
1441  TRACE("[ Done calling %s @ %p for '%s' ]", function_name, function, name);
1442
1443  // The function may have called dlopen(3) or dlclose(3), so we need to ensure our data structures
1444  // are still writable. This happens with our debug malloc (see http://b/7941716).
1445  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
1446}
1447
1448void soinfo::CallPreInitConstructors() {
1449  // DT_PREINIT_ARRAY functions are called before any other constructors for executables,
1450  // but ignored in a shared library.
1451  CallArray("DT_PREINIT_ARRAY", preinit_array, preinit_array_count, false);
1452}
1453
1454void soinfo::CallConstructors() {
1455  if (constructors_called) {
1456    return;
1457  }
1458
1459  // We set constructors_called before actually calling the constructors, otherwise it doesn't
1460  // protect against recursive constructor calls. One simple example of constructor recursion
1461  // is the libc debug malloc, which is implemented in libc_malloc_debug_leak.so:
1462  // 1. The program depends on libc, so libc's constructor is called here.
1463  // 2. The libc constructor calls dlopen() to load libc_malloc_debug_leak.so.
1464  // 3. dlopen() calls the constructors on the newly created
1465  //    soinfo for libc_malloc_debug_leak.so.
1466  // 4. The debug .so depends on libc, so CallConstructors is
1467  //    called again with the libc soinfo. If it doesn't trigger the early-
1468  //    out above, the libc constructor will be called again (recursively!).
1469  constructors_called = true;
1470
1471  if ((flags & FLAG_EXE) == 0 && preinit_array != NULL) {
1472    // The GNU dynamic linker silently ignores these, but we warn the developer.
1473    PRINT("\"%s\": ignoring %zd-entry DT_PREINIT_ARRAY in shared library!",
1474          name, preinit_array_count);
1475  }
1476
1477  if (dynamic != NULL) {
1478    for (ElfW(Dyn)* d = dynamic; d->d_tag != DT_NULL; ++d) {
1479      if (d->d_tag == DT_NEEDED) {
1480        const char* library_name = strtab + d->d_un.d_val;
1481        TRACE("\"%s\": calling constructors in DT_NEEDED \"%s\"", name, library_name);
1482        find_loaded_library(library_name)->CallConstructors();
1483      }
1484    }
1485  }
1486
1487  TRACE("\"%s\": calling constructors", name);
1488
1489  // DT_INIT should be called before DT_INIT_ARRAY if both are present.
1490  CallFunction("DT_INIT", init_func);
1491  CallArray("DT_INIT_ARRAY", init_array, init_array_count, false);
1492}
1493
1494void soinfo::CallDestructors() {
1495  TRACE("\"%s\": calling destructors", name);
1496
1497  // DT_FINI_ARRAY must be parsed in reverse order.
1498  CallArray("DT_FINI_ARRAY", fini_array, fini_array_count, true);
1499
1500  // DT_FINI should be called after DT_FINI_ARRAY if both are present.
1501  CallFunction("DT_FINI", fini_func);
1502}
1503
1504/* Force any of the closed stdin, stdout and stderr to be associated with
1505   /dev/null. */
1506static int nullify_closed_stdio() {
1507    int dev_null, i, status;
1508    int return_value = 0;
1509
1510    dev_null = TEMP_FAILURE_RETRY(open("/dev/null", O_RDWR));
1511    if (dev_null < 0) {
1512        DL_ERR("cannot open /dev/null: %s", strerror(errno));
1513        return -1;
1514    }
1515    TRACE("[ Opened /dev/null file-descriptor=%d]", dev_null);
1516
1517    /* If any of the stdio file descriptors is valid and not associated
1518       with /dev/null, dup /dev/null to it.  */
1519    for (i = 0; i < 3; i++) {
1520        /* If it is /dev/null already, we are done. */
1521        if (i == dev_null) {
1522            continue;
1523        }
1524
1525        TRACE("[ Nullifying stdio file descriptor %d]", i);
1526        status = TEMP_FAILURE_RETRY(fcntl(i, F_GETFL));
1527
1528        /* If file is opened, we are good. */
1529        if (status != -1) {
1530            continue;
1531        }
1532
1533        /* The only error we allow is that the file descriptor does not
1534           exist, in which case we dup /dev/null to it. */
1535        if (errno != EBADF) {
1536            DL_ERR("fcntl failed: %s", strerror(errno));
1537            return_value = -1;
1538            continue;
1539        }
1540
1541        /* Try dupping /dev/null to this stdio file descriptor and
1542           repeat if there is a signal.  Note that any errors in closing
1543           the stdio descriptor are lost.  */
1544        status = TEMP_FAILURE_RETRY(dup2(dev_null, i));
1545        if (status < 0) {
1546            DL_ERR("dup2 failed: %s", strerror(errno));
1547            return_value = -1;
1548            continue;
1549        }
1550    }
1551
1552    /* If /dev/null is not one of the stdio file descriptors, close it. */
1553    if (dev_null > 2) {
1554        TRACE("[ Closing /dev/null file-descriptor=%d]", dev_null);
1555        status = TEMP_FAILURE_RETRY(close(dev_null));
1556        if (status == -1) {
1557            DL_ERR("close failed: %s", strerror(errno));
1558            return_value = -1;
1559        }
1560    }
1561
1562    return return_value;
1563}
1564
1565static bool soinfo_link_image(soinfo* si) {
1566    /* "base" might wrap around UINT32_MAX. */
1567    ElfW(Addr) base = si->load_bias;
1568    const ElfW(Phdr)* phdr = si->phdr;
1569    int phnum = si->phnum;
1570    bool relocating_linker = (si->flags & FLAG_LINKER) != 0;
1571
1572    /* We can't debug anything until the linker is relocated */
1573    if (!relocating_linker) {
1574        INFO("[ linking %s ]", si->name);
1575        DEBUG("si->base = %p si->flags = 0x%08x", reinterpret_cast<void*>(si->base), si->flags);
1576    }
1577
1578    /* Extract dynamic section */
1579    size_t dynamic_count;
1580    ElfW(Word) dynamic_flags;
1581    phdr_table_get_dynamic_section(phdr, phnum, base, &si->dynamic,
1582                                   &dynamic_count, &dynamic_flags);
1583    if (si->dynamic == NULL) {
1584        if (!relocating_linker) {
1585            DL_ERR("missing PT_DYNAMIC in \"%s\"", si->name);
1586        }
1587        return false;
1588    } else {
1589        if (!relocating_linker) {
1590            DEBUG("dynamic = %p", si->dynamic);
1591        }
1592    }
1593
1594#if defined(__arm__)
1595    (void) phdr_table_get_arm_exidx(phdr, phnum, base,
1596                                    &si->ARM_exidx, &si->ARM_exidx_count);
1597#endif
1598
1599    // Extract useful information from dynamic section.
1600    uint32_t needed_count = 0;
1601    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1602        DEBUG("d = %p, d[0](tag) = %p d[1](val) = %p",
1603              d, reinterpret_cast<void*>(d->d_tag), reinterpret_cast<void*>(d->d_un.d_val));
1604        switch (d->d_tag) {
1605        case DT_HASH:
1606            si->nbucket = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr)[0];
1607            si->nchain = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr)[1];
1608            si->bucket = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr + 8);
1609            si->chain = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr + 8 + si->nbucket * 4);
1610            break;
1611        case DT_STRTAB:
1612            si->strtab = reinterpret_cast<const char*>(base + d->d_un.d_ptr);
1613            break;
1614        case DT_SYMTAB:
1615            si->symtab = reinterpret_cast<ElfW(Sym)*>(base + d->d_un.d_ptr);
1616            break;
1617#if !defined(__LP64__)
1618        case DT_PLTREL:
1619            if (d->d_un.d_val != DT_REL) {
1620                DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1621                return false;
1622            }
1623            break;
1624#endif
1625        case DT_JMPREL:
1626#if defined(USE_RELA)
1627            si->plt_rela = reinterpret_cast<ElfW(Rela)*>(base + d->d_un.d_ptr);
1628#else
1629            si->plt_rel = reinterpret_cast<ElfW(Rel)*>(base + d->d_un.d_ptr);
1630#endif
1631            break;
1632        case DT_PLTRELSZ:
1633#if defined(USE_RELA)
1634            si->plt_rela_count = d->d_un.d_val / sizeof(ElfW(Rela));
1635#else
1636            si->plt_rel_count = d->d_un.d_val / sizeof(ElfW(Rel));
1637#endif
1638            break;
1639#if defined(__mips__)
1640        case DT_PLTGOT:
1641            // Used by mips and mips64.
1642            si->plt_got = reinterpret_cast<ElfW(Addr)**>(base + d->d_un.d_ptr);
1643            break;
1644#endif
1645        case DT_DEBUG:
1646            // Set the DT_DEBUG entry to the address of _r_debug for GDB
1647            // if the dynamic table is writable
1648// FIXME: not working currently for N64
1649// The flags for the LOAD and DYNAMIC program headers do not agree.
1650// The LOAD section containng the dynamic table has been mapped as
1651// read-only, but the DYNAMIC header claims it is writable.
1652#if !(defined(__mips__) && defined(__LP64__))
1653            if ((dynamic_flags & PF_W) != 0) {
1654                d->d_un.d_val = reinterpret_cast<uintptr_t>(&_r_debug);
1655            }
1656            break;
1657#endif
1658#if defined(USE_RELA)
1659         case DT_RELA:
1660            si->rela = reinterpret_cast<ElfW(Rela)*>(base + d->d_un.d_ptr);
1661            break;
1662         case DT_RELASZ:
1663            si->rela_count = d->d_un.d_val / sizeof(ElfW(Rela));
1664            break;
1665        case DT_REL:
1666            DL_ERR("unsupported DT_REL in \"%s\"", si->name);
1667            return false;
1668        case DT_RELSZ:
1669            DL_ERR("unsupported DT_RELSZ in \"%s\"", si->name);
1670            return false;
1671#else
1672        case DT_REL:
1673            si->rel = reinterpret_cast<ElfW(Rel)*>(base + d->d_un.d_ptr);
1674            break;
1675        case DT_RELSZ:
1676            si->rel_count = d->d_un.d_val / sizeof(ElfW(Rel));
1677            break;
1678         case DT_RELA:
1679            DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1680            return false;
1681#endif
1682        case DT_INIT:
1683            si->init_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1684            DEBUG("%s constructors (DT_INIT) found at %p", si->name, si->init_func);
1685            break;
1686        case DT_FINI:
1687            si->fini_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1688            DEBUG("%s destructors (DT_FINI) found at %p", si->name, si->fini_func);
1689            break;
1690        case DT_INIT_ARRAY:
1691            si->init_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1692            DEBUG("%s constructors (DT_INIT_ARRAY) found at %p", si->name, si->init_array);
1693            break;
1694        case DT_INIT_ARRAYSZ:
1695            si->init_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1696            break;
1697        case DT_FINI_ARRAY:
1698            si->fini_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1699            DEBUG("%s destructors (DT_FINI_ARRAY) found at %p", si->name, si->fini_array);
1700            break;
1701        case DT_FINI_ARRAYSZ:
1702            si->fini_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1703            break;
1704        case DT_PREINIT_ARRAY:
1705            si->preinit_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1706            DEBUG("%s constructors (DT_PREINIT_ARRAY) found at %p", si->name, si->preinit_array);
1707            break;
1708        case DT_PREINIT_ARRAYSZ:
1709            si->preinit_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1710            break;
1711        case DT_TEXTREL:
1712#if defined(__LP64__)
1713            DL_ERR("text relocations (DT_TEXTREL) found in 64-bit ELF file \"%s\"", si->name);
1714            return false;
1715#else
1716            si->has_text_relocations = true;
1717            break;
1718#endif
1719        case DT_SYMBOLIC:
1720            si->has_DT_SYMBOLIC = true;
1721            break;
1722        case DT_NEEDED:
1723            ++needed_count;
1724            break;
1725        case DT_FLAGS:
1726            if (d->d_un.d_val & DF_TEXTREL) {
1727#if defined(__LP64__)
1728                DL_ERR("text relocations (DF_TEXTREL) found in 64-bit ELF file \"%s\"", si->name);
1729                return false;
1730#else
1731                si->has_text_relocations = true;
1732#endif
1733            }
1734            if (d->d_un.d_val & DF_SYMBOLIC) {
1735                si->has_DT_SYMBOLIC = true;
1736            }
1737            break;
1738#if defined(__mips__)
1739        case DT_STRSZ:
1740        case DT_SYMENT:
1741        case DT_RELENT:
1742             break;
1743        case DT_MIPS_RLD_MAP:
1744            // Set the DT_MIPS_RLD_MAP entry to the address of _r_debug for GDB.
1745            {
1746              r_debug** dp = reinterpret_cast<r_debug**>(base + d->d_un.d_ptr);
1747              *dp = &_r_debug;
1748            }
1749            break;
1750        case DT_MIPS_RLD_VERSION:
1751        case DT_MIPS_FLAGS:
1752        case DT_MIPS_BASE_ADDRESS:
1753        case DT_MIPS_UNREFEXTNO:
1754            break;
1755
1756        case DT_MIPS_SYMTABNO:
1757            si->mips_symtabno = d->d_un.d_val;
1758            break;
1759
1760        case DT_MIPS_LOCAL_GOTNO:
1761            si->mips_local_gotno = d->d_un.d_val;
1762            break;
1763
1764        case DT_MIPS_GOTSYM:
1765            si->mips_gotsym = d->d_un.d_val;
1766            break;
1767#endif
1768
1769        default:
1770            DEBUG("Unused DT entry: type %p arg %p",
1771                  reinterpret_cast<void*>(d->d_tag), reinterpret_cast<void*>(d->d_un.d_val));
1772            break;
1773        }
1774    }
1775
1776    DEBUG("si->base = %p, si->strtab = %p, si->symtab = %p",
1777          reinterpret_cast<void*>(si->base), si->strtab, si->symtab);
1778
1779    // Sanity checks.
1780    if (relocating_linker && needed_count != 0) {
1781        DL_ERR("linker cannot have DT_NEEDED dependencies on other libraries");
1782        return false;
1783    }
1784    if (si->nbucket == 0) {
1785        DL_ERR("empty/missing DT_HASH in \"%s\" (built with --hash-style=gnu?)", si->name);
1786        return false;
1787    }
1788    if (si->strtab == 0) {
1789        DL_ERR("empty/missing DT_STRTAB in \"%s\"", si->name);
1790        return false;
1791    }
1792    if (si->symtab == 0) {
1793        DL_ERR("empty/missing DT_SYMTAB in \"%s\"", si->name);
1794        return false;
1795    }
1796
1797    // If this is the main executable, then load all of the libraries from LD_PRELOAD now.
1798    if (si->flags & FLAG_EXE) {
1799        memset(gLdPreloads, 0, sizeof(gLdPreloads));
1800        size_t preload_count = 0;
1801        for (size_t i = 0; gLdPreloadNames[i] != NULL; i++) {
1802            soinfo* lsi = find_library(gLdPreloadNames[i]);
1803            if (lsi != NULL) {
1804                gLdPreloads[preload_count++] = lsi;
1805            } else {
1806                // As with glibc, failure to load an LD_PRELOAD library is just a warning.
1807                DL_WARN("could not load library \"%s\" from LD_PRELOAD for \"%s\"; caused by %s",
1808                        gLdPreloadNames[i], si->name, linker_get_error_buffer());
1809            }
1810        }
1811    }
1812
1813    soinfo** needed = reinterpret_cast<soinfo**>(alloca((1 + needed_count) * sizeof(soinfo*)));
1814    soinfo** pneeded = needed;
1815
1816    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1817        if (d->d_tag == DT_NEEDED) {
1818            const char* library_name = si->strtab + d->d_un.d_val;
1819            DEBUG("%s needs %s", si->name, library_name);
1820            soinfo* lsi = find_library(library_name);
1821            if (lsi == NULL) {
1822                strlcpy(tmp_err_buf, linker_get_error_buffer(), sizeof(tmp_err_buf));
1823                DL_ERR("could not load library \"%s\" needed by \"%s\"; caused by %s",
1824                       library_name, si->name, tmp_err_buf);
1825                return false;
1826            }
1827            *pneeded++ = lsi;
1828        }
1829    }
1830    *pneeded = NULL;
1831
1832#if !defined(__LP64__)
1833    if (si->has_text_relocations) {
1834        // Make segments writable to allow text relocations to work properly. We will later call
1835        // phdr_table_protect_segments() after all of them are applied and all constructors are run.
1836        DL_WARN("%s has text relocations. This is wasting memory and prevents "
1837                "security hardening. Please fix.", si->name);
1838        if (phdr_table_unprotect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1839            DL_ERR("can't unprotect loadable segments for \"%s\": %s",
1840                   si->name, strerror(errno));
1841            return false;
1842        }
1843    }
1844#endif
1845
1846#if defined(USE_RELA)
1847    if (si->plt_rela != NULL) {
1848        DEBUG("[ relocating %s plt ]\n", si->name);
1849        if (soinfo_relocate(si, si->plt_rela, si->plt_rela_count, needed)) {
1850            return false;
1851        }
1852    }
1853    if (si->rela != NULL) {
1854        DEBUG("[ relocating %s ]\n", si->name);
1855        if (soinfo_relocate(si, si->rela, si->rela_count, needed)) {
1856            return false;
1857        }
1858    }
1859#else
1860    if (si->plt_rel != NULL) {
1861        DEBUG("[ relocating %s plt ]", si->name);
1862        if (soinfo_relocate(si, si->plt_rel, si->plt_rel_count, needed)) {
1863            return false;
1864        }
1865    }
1866    if (si->rel != NULL) {
1867        DEBUG("[ relocating %s ]", si->name);
1868        if (soinfo_relocate(si, si->rel, si->rel_count, needed)) {
1869            return false;
1870        }
1871    }
1872#endif
1873
1874#if defined(__mips__)
1875    if (!mips_relocate_got(si, needed)) {
1876        return false;
1877    }
1878#endif
1879
1880    si->flags |= FLAG_LINKED;
1881    DEBUG("[ finished linking %s ]", si->name);
1882
1883#if !defined(__LP64__)
1884    if (si->has_text_relocations) {
1885        // All relocations are done, we can protect our segments back to read-only.
1886        if (phdr_table_protect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1887            DL_ERR("can't protect segments for \"%s\": %s",
1888                   si->name, strerror(errno));
1889            return false;
1890        }
1891    }
1892#endif
1893
1894    /* We can also turn on GNU RELRO protection */
1895    if (phdr_table_protect_gnu_relro(si->phdr, si->phnum, si->load_bias) < 0) {
1896        DL_ERR("can't enable GNU RELRO protection for \"%s\": %s",
1897               si->name, strerror(errno));
1898        return false;
1899    }
1900
1901    notify_gdb_of_load(si);
1902    return true;
1903}
1904
1905/*
1906 * This function add vdso to internal dso list.
1907 * It helps to stack unwinding through signal handlers.
1908 * Also, it makes bionic more like glibc.
1909 */
1910static void add_vdso(KernelArgumentBlock& args __unused) {
1911#if defined(AT_SYSINFO_EHDR)
1912  ElfW(Ehdr)* ehdr_vdso = reinterpret_cast<ElfW(Ehdr)*>(args.getauxval(AT_SYSINFO_EHDR));
1913  if (ehdr_vdso == NULL) {
1914    return;
1915  }
1916
1917  soinfo* si = soinfo_alloc("[vdso]");
1918
1919  si->phdr = reinterpret_cast<ElfW(Phdr)*>(reinterpret_cast<char*>(ehdr_vdso) + ehdr_vdso->e_phoff);
1920  si->phnum = ehdr_vdso->e_phnum;
1921  si->base = reinterpret_cast<ElfW(Addr)>(ehdr_vdso);
1922  si->size = phdr_table_get_load_size(si->phdr, si->phnum);
1923  si->flags = 0;
1924  si->load_bias = get_elf_exec_load_bias(ehdr_vdso);
1925
1926  soinfo_link_image(si);
1927#endif
1928}
1929
1930/*
1931 * This code is called after the linker has linked itself and
1932 * fixed it's own GOT. It is safe to make references to externs
1933 * and other non-local data at this point.
1934 */
1935static ElfW(Addr) __linker_init_post_relocation(KernelArgumentBlock& args, ElfW(Addr) linker_base) {
1936    /* NOTE: we store the args pointer on a special location
1937     *       of the temporary TLS area in order to pass it to
1938     *       the C Library's runtime initializer.
1939     *
1940     *       The initializer must clear the slot and reset the TLS
1941     *       to point to a different location to ensure that no other
1942     *       shared library constructor can access it.
1943     */
1944  __libc_init_tls(args);
1945
1946#if TIMING
1947    struct timeval t0, t1;
1948    gettimeofday(&t0, 0);
1949#endif
1950
1951    // Initialize environment functions, and get to the ELF aux vectors table.
1952    linker_env_init(args);
1953
1954    // If this is a setuid/setgid program, close the security hole described in
1955    // ftp://ftp.freebsd.org/pub/FreeBSD/CERT/advisories/FreeBSD-SA-02:23.stdio.asc
1956    if (get_AT_SECURE()) {
1957        nullify_closed_stdio();
1958    }
1959
1960    debuggerd_init();
1961
1962    // Get a few environment variables.
1963    const char* LD_DEBUG = linker_env_get("LD_DEBUG");
1964    if (LD_DEBUG != NULL) {
1965      gLdDebugVerbosity = atoi(LD_DEBUG);
1966    }
1967
1968    // Normally, these are cleaned by linker_env_init, but the test
1969    // doesn't cost us anything.
1970    const char* ldpath_env = NULL;
1971    const char* ldpreload_env = NULL;
1972    if (!get_AT_SECURE()) {
1973      ldpath_env = linker_env_get("LD_LIBRARY_PATH");
1974      ldpreload_env = linker_env_get("LD_PRELOAD");
1975    }
1976
1977    INFO("[ android linker & debugger ]");
1978
1979    soinfo* si = soinfo_alloc(args.argv[0]);
1980    if (si == NULL) {
1981        exit(EXIT_FAILURE);
1982    }
1983
1984    /* bootstrap the link map, the main exe always needs to be first */
1985    si->flags |= FLAG_EXE;
1986    link_map* map = &(si->link_map_head);
1987
1988    map->l_addr = 0;
1989    map->l_name = args.argv[0];
1990    map->l_prev = NULL;
1991    map->l_next = NULL;
1992
1993    _r_debug.r_map = map;
1994    r_debug_tail = map;
1995
1996    /* gdb expects the linker to be in the debug shared object list.
1997     * Without this, gdb has trouble locating the linker's ".text"
1998     * and ".plt" sections. Gdb could also potentially use this to
1999     * relocate the offset of our exported 'rtld_db_dlactivity' symbol.
2000     * Don't use soinfo_alloc(), because the linker shouldn't
2001     * be on the soinfo list.
2002     */
2003    {
2004        static soinfo linker_soinfo;
2005#if defined(__LP64__)
2006        strlcpy(linker_soinfo.name, "/system/bin/linker64", sizeof(linker_soinfo.name));
2007#else
2008        strlcpy(linker_soinfo.name, "/system/bin/linker", sizeof(linker_soinfo.name));
2009#endif
2010        linker_soinfo.flags = 0;
2011        linker_soinfo.base = linker_base;
2012
2013        /*
2014         * Set the dynamic field in the link map otherwise gdb will complain with
2015         * the following:
2016         *   warning: .dynamic section for "/system/bin/linker" is not at the
2017         *   expected address (wrong library or version mismatch?)
2018         */
2019        ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_base);
2020        ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_base + elf_hdr->e_phoff);
2021        phdr_table_get_dynamic_section(phdr, elf_hdr->e_phnum, linker_base,
2022                                       &linker_soinfo.dynamic, NULL, NULL);
2023        insert_soinfo_into_debug_map(&linker_soinfo);
2024    }
2025
2026    // Extract information passed from the kernel.
2027    si->phdr = reinterpret_cast<ElfW(Phdr)*>(args.getauxval(AT_PHDR));
2028    si->phnum = args.getauxval(AT_PHNUM);
2029    si->entry = args.getauxval(AT_ENTRY);
2030
2031    /* Compute the value of si->base. We can't rely on the fact that
2032     * the first entry is the PHDR because this will not be true
2033     * for certain executables (e.g. some in the NDK unit test suite)
2034     */
2035    si->base = 0;
2036    si->size = phdr_table_get_load_size(si->phdr, si->phnum);
2037    si->load_bias = 0;
2038    for (size_t i = 0; i < si->phnum; ++i) {
2039      if (si->phdr[i].p_type == PT_PHDR) {
2040        si->load_bias = reinterpret_cast<ElfW(Addr)>(si->phdr) - si->phdr[i].p_vaddr;
2041        si->base = reinterpret_cast<ElfW(Addr)>(si->phdr) - si->phdr[i].p_offset;
2042        break;
2043      }
2044    }
2045    si->dynamic = NULL;
2046    si->ref_count = 1;
2047
2048    // Use LD_LIBRARY_PATH and LD_PRELOAD (but only if we aren't setuid/setgid).
2049    parse_LD_LIBRARY_PATH(ldpath_env);
2050    parse_LD_PRELOAD(ldpreload_env);
2051
2052    somain = si;
2053
2054    if (!soinfo_link_image(si)) {
2055        __libc_format_fd(2, "CANNOT LINK EXECUTABLE: %s\n", linker_get_error_buffer());
2056        exit(EXIT_FAILURE);
2057    }
2058
2059    add_vdso(args);
2060
2061    si->CallPreInitConstructors();
2062
2063    for (size_t i = 0; gLdPreloads[i] != NULL; ++i) {
2064        gLdPreloads[i]->CallConstructors();
2065    }
2066
2067    /* After the link_image, the si->load_bias is initialized.
2068     * For so lib, the map->l_addr will be updated in notify_gdb_of_load.
2069     * We need to update this value for so exe here. So Unwind_Backtrace
2070     * for some arch like x86 could work correctly within so exe.
2071     */
2072    map->l_addr = si->load_bias;
2073    si->CallConstructors();
2074
2075#if TIMING
2076    gettimeofday(&t1, NULL);
2077    PRINT("LINKER TIME: %s: %d microseconds", args.argv[0], (int) (
2078               (((long long)t1.tv_sec * 1000000LL) + (long long)t1.tv_usec) -
2079               (((long long)t0.tv_sec * 1000000LL) + (long long)t0.tv_usec)));
2080#endif
2081#if STATS
2082    PRINT("RELO STATS: %s: %d abs, %d rel, %d copy, %d symbol", args.argv[0],
2083           linker_stats.count[kRelocAbsolute],
2084           linker_stats.count[kRelocRelative],
2085           linker_stats.count[kRelocCopy],
2086           linker_stats.count[kRelocSymbol]);
2087#endif
2088#if COUNT_PAGES
2089    {
2090        unsigned n;
2091        unsigned i;
2092        unsigned count = 0;
2093        for (n = 0; n < 4096; n++) {
2094            if (bitmask[n]) {
2095                unsigned x = bitmask[n];
2096#if defined(__LP64__)
2097                for (i = 0; i < 32; i++) {
2098#else
2099                for (i = 0; i < 8; i++) {
2100#endif
2101                    if (x & 1) {
2102                        count++;
2103                    }
2104                    x >>= 1;
2105                }
2106            }
2107        }
2108        PRINT("PAGES MODIFIED: %s: %d (%dKB)", args.argv[0], count, count * 4);
2109    }
2110#endif
2111
2112#if TIMING || STATS || COUNT_PAGES
2113    fflush(stdout);
2114#endif
2115
2116    TRACE("[ Ready to execute '%s' @ %p ]", si->name, reinterpret_cast<void*>(si->entry));
2117    return si->entry;
2118}
2119
2120/* Compute the load-bias of an existing executable. This shall only
2121 * be used to compute the load bias of an executable or shared library
2122 * that was loaded by the kernel itself.
2123 *
2124 * Input:
2125 *    elf    -> address of ELF header, assumed to be at the start of the file.
2126 * Return:
2127 *    load bias, i.e. add the value of any p_vaddr in the file to get
2128 *    the corresponding address in memory.
2129 */
2130static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf) {
2131  ElfW(Addr) offset = elf->e_phoff;
2132  const ElfW(Phdr)* phdr_table = reinterpret_cast<const ElfW(Phdr)*>(reinterpret_cast<uintptr_t>(elf) + offset);
2133  const ElfW(Phdr)* phdr_end = phdr_table + elf->e_phnum;
2134
2135  for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_end; phdr++) {
2136    if (phdr->p_type == PT_LOAD) {
2137      return reinterpret_cast<ElfW(Addr)>(elf) + phdr->p_offset - phdr->p_vaddr;
2138    }
2139  }
2140  return 0;
2141}
2142
2143/*
2144 * This is the entry point for the linker, called from begin.S. This
2145 * method is responsible for fixing the linker's own relocations, and
2146 * then calling __linker_init_post_relocation().
2147 *
2148 * Because this method is called before the linker has fixed it's own
2149 * relocations, any attempt to reference an extern variable, extern
2150 * function, or other GOT reference will generate a segfault.
2151 */
2152extern "C" ElfW(Addr) __linker_init(void* raw_args) {
2153  KernelArgumentBlock args(raw_args);
2154
2155  ElfW(Addr) linker_addr = args.getauxval(AT_BASE);
2156  ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_addr);
2157  ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_addr + elf_hdr->e_phoff);
2158
2159  soinfo linker_so;
2160  memset(&linker_so, 0, sizeof(soinfo));
2161
2162  strcpy(linker_so.name, "[dynamic linker]");
2163  linker_so.base = linker_addr;
2164  linker_so.size = phdr_table_get_load_size(phdr, elf_hdr->e_phnum);
2165  linker_so.load_bias = get_elf_exec_load_bias(elf_hdr);
2166  linker_so.dynamic = NULL;
2167  linker_so.phdr = phdr;
2168  linker_so.phnum = elf_hdr->e_phnum;
2169  linker_so.flags |= FLAG_LINKER;
2170
2171  if (!soinfo_link_image(&linker_so)) {
2172    // It would be nice to print an error message, but if the linker
2173    // can't link itself, there's no guarantee that we'll be able to
2174    // call write() (because it involves a GOT reference). We may as
2175    // well try though...
2176    const char* msg = "CANNOT LINK EXECUTABLE: ";
2177    write(2, msg, strlen(msg));
2178    write(2, __linker_dl_err_buf, strlen(__linker_dl_err_buf));
2179    write(2, "\n", 1);
2180    _exit(EXIT_FAILURE);
2181  }
2182
2183  // We have successfully fixed our own relocations. It's safe to run
2184  // the main part of the linker now.
2185  args.abort_message_ptr = &gAbortMessage;
2186  ElfW(Addr) start_address = __linker_init_post_relocation(args, linker_addr);
2187
2188  set_soinfo_pool_protection(PROT_READ);
2189
2190  // Return the address that the calling assembly stub should jump to.
2191  return start_address;
2192}
2193