linker.cpp revision c620059479c47a78d57086d73726c9adc2f337ad
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 <linux/auxvec.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/* Assume average path length of 64 and max 8 paths */
53#define LDPATH_BUFSIZE 512
54#define LDPATH_MAX 8
55
56#define LDPRELOAD_BUFSIZE 512
57#define LDPRELOAD_MAX 8
58
59/* >>> IMPORTANT NOTE - READ ME BEFORE MODIFYING <<<
60 *
61 * Do NOT use malloc() and friends or pthread_*() code here.
62 * Don't use printf() either; it's caused mysterious memory
63 * corruption in the past.
64 * The linker runs before we bring up libc and it's easiest
65 * to make sure it does not depend on any complex libc features
66 *
67 * open issues / todo:
68 *
69 * - are we doing everything we should for ARM_COPY relocations?
70 * - cleaner error reporting
71 * - after linking, set as much stuff as possible to READONLY
72 *   and NOEXEC
73 */
74
75static bool soinfo_link_image(soinfo* si);
76
77// We can't use malloc(3) in the dynamic linker. We use a linked list of anonymous
78// maps, each a single page in size. The pages are broken up into as many struct soinfo
79// objects as will fit, and they're all threaded together on a free list.
80#define SOINFO_PER_POOL ((PAGE_SIZE - sizeof(soinfo_pool_t*)) / sizeof(soinfo))
81struct soinfo_pool_t {
82  soinfo_pool_t* next;
83  soinfo info[SOINFO_PER_POOL];
84};
85static struct soinfo_pool_t* gSoInfoPools = NULL;
86static soinfo* gSoInfoFreeList = NULL;
87
88static soinfo* solist = &libdl_info;
89static soinfo* sonext = &libdl_info;
90static soinfo* somain; /* main process, always the one after libdl_info */
91
92static const char* const gSoPaths[] = {
93  "/vendor/lib",
94  "/system/lib",
95  NULL
96};
97
98static char gLdPathsBuffer[LDPATH_BUFSIZE];
99static const char* gLdPaths[LDPATH_MAX + 1];
100
101static char gLdPreloadsBuffer[LDPRELOAD_BUFSIZE];
102static const char* gLdPreloadNames[LDPRELOAD_MAX + 1];
103
104static soinfo* gLdPreloads[LDPRELOAD_MAX + 1];
105
106__LIBC_HIDDEN__ int gLdDebugVerbosity;
107
108__LIBC_HIDDEN__ abort_msg_t* gAbortMessage = NULL; // For debuggerd.
109
110enum RelocationKind {
111    kRelocAbsolute = 0,
112    kRelocRelative,
113    kRelocCopy,
114    kRelocSymbol,
115    kRelocMax
116};
117
118#if STATS
119struct linker_stats_t {
120    int count[kRelocMax];
121};
122
123static linker_stats_t linker_stats;
124
125static void count_relocation(RelocationKind kind) {
126    ++linker_stats.count[kind];
127}
128#else
129static void count_relocation(RelocationKind) {
130}
131#endif
132
133#if COUNT_PAGES
134static unsigned bitmask[4096];
135#define MARK(offset) \
136    do { \
137        bitmask[((offset) >> 12) >> 3] |= (1 << (((offset) >> 12) & 7)); \
138    } while(0)
139#else
140#define MARK(x) do {} while (0)
141#endif
142
143// You shouldn't try to call memory-allocating functions in the dynamic linker.
144// Guard against the most obvious ones.
145#define DISALLOW_ALLOCATION(return_type, name, ...) \
146    return_type name __VA_ARGS__ \
147    { \
148        const char* msg = "ERROR: " #name " called from the dynamic linker!\n"; \
149        __libc_format_log(ANDROID_LOG_FATAL, "linker", "%s", msg); \
150        write(2, msg, strlen(msg)); \
151        abort(); \
152    }
153#define UNUSED __attribute__((unused))
154DISALLOW_ALLOCATION(void*, malloc, (size_t u UNUSED));
155DISALLOW_ALLOCATION(void, free, (void* u UNUSED));
156DISALLOW_ALLOCATION(void*, realloc, (void* u1 UNUSED, size_t u2 UNUSED));
157DISALLOW_ALLOCATION(void*, calloc, (size_t u1 UNUSED, size_t u2 UNUSED));
158
159static char tmp_err_buf[768];
160static char __linker_dl_err_buf[768];
161
162char* linker_get_error_buffer() {
163  return &__linker_dl_err_buf[0];
164}
165
166size_t linker_get_error_buffer_size() {
167  return sizeof(__linker_dl_err_buf);
168}
169
170/*
171 * This function is an empty stub where GDB locates a breakpoint to get notified
172 * about linker activity.
173 */
174extern "C" void __attribute__((noinline)) __attribute__((visibility("default"))) rtld_db_dlactivity();
175
176static r_debug _r_debug = {1, NULL, &rtld_db_dlactivity, RT_CONSISTENT, 0};
177static link_map_t* r_debug_tail = 0;
178
179static pthread_mutex_t gDebugMutex = PTHREAD_MUTEX_INITIALIZER;
180
181static void insert_soinfo_into_debug_map(soinfo * info) {
182    // Copy the necessary fields into the debug structure.
183    link_map_t* map = &(info->link_map);
184    map->l_addr = info->base;
185    map->l_name = (char*) info->name;
186    map->l_ld = (uintptr_t)info->dynamic;
187
188    /* Stick the new library at the end of the list.
189     * gdb tends to care more about libc than it does
190     * about leaf libraries, and ordering it this way
191     * reduces the back-and-forth over the wire.
192     */
193    if (r_debug_tail) {
194        r_debug_tail->l_next = map;
195        map->l_prev = r_debug_tail;
196        map->l_next = 0;
197    } else {
198        _r_debug.r_map = map;
199        map->l_prev = 0;
200        map->l_next = 0;
201    }
202    r_debug_tail = map;
203}
204
205static void remove_soinfo_from_debug_map(soinfo* info) {
206    link_map_t* map = &(info->link_map);
207
208    if (r_debug_tail == map) {
209        r_debug_tail = map->l_prev;
210    }
211
212    if (map->l_prev) {
213        map->l_prev->l_next = map->l_next;
214    }
215    if (map->l_next) {
216        map->l_next->l_prev = map->l_prev;
217    }
218}
219
220static void notify_gdb_of_load(soinfo* info) {
221    if (info->flags & FLAG_EXE) {
222        // GDB already knows about the main executable
223        return;
224    }
225
226    ScopedPthreadMutexLocker locker(&gDebugMutex);
227
228    _r_debug.r_state = RT_ADD;
229    rtld_db_dlactivity();
230
231    insert_soinfo_into_debug_map(info);
232
233    _r_debug.r_state = RT_CONSISTENT;
234    rtld_db_dlactivity();
235}
236
237static void notify_gdb_of_unload(soinfo* info) {
238    if (info->flags & FLAG_EXE) {
239        // GDB already knows about the main executable
240        return;
241    }
242
243    ScopedPthreadMutexLocker locker(&gDebugMutex);
244
245    _r_debug.r_state = RT_DELETE;
246    rtld_db_dlactivity();
247
248    remove_soinfo_from_debug_map(info);
249
250    _r_debug.r_state = RT_CONSISTENT;
251    rtld_db_dlactivity();
252}
253
254void notify_gdb_of_libraries() {
255    _r_debug.r_state = RT_ADD;
256    rtld_db_dlactivity();
257    _r_debug.r_state = RT_CONSISTENT;
258    rtld_db_dlactivity();
259}
260
261static bool ensure_free_list_non_empty() {
262  if (gSoInfoFreeList != NULL) {
263    return true;
264  }
265
266  // Allocate a new pool.
267  soinfo_pool_t* pool = reinterpret_cast<soinfo_pool_t*>(mmap(NULL, sizeof(*pool),
268                                                              PROT_READ|PROT_WRITE,
269                                                              MAP_PRIVATE|MAP_ANONYMOUS, 0, 0));
270  if (pool == MAP_FAILED) {
271    return false;
272  }
273
274  // Add the pool to our list of pools.
275  pool->next = gSoInfoPools;
276  gSoInfoPools = pool;
277
278  // Chain the entries in the new pool onto the free list.
279  gSoInfoFreeList = &pool->info[0];
280  soinfo* next = NULL;
281  for (int i = SOINFO_PER_POOL - 1; i >= 0; --i) {
282    pool->info[i].next = next;
283    next = &pool->info[i];
284  }
285
286  return true;
287}
288
289static void set_soinfo_pool_protection(int protection) {
290  for (soinfo_pool_t* p = gSoInfoPools; p != NULL; p = p->next) {
291    if (mprotect(p, sizeof(*p), protection) == -1) {
292      abort(); // Can't happen.
293    }
294  }
295}
296
297static soinfo* soinfo_alloc(const char* name) {
298  if (strlen(name) >= SOINFO_NAME_LEN) {
299    DL_ERR("library name \"%s\" too long", name);
300    return NULL;
301  }
302
303  if (!ensure_free_list_non_empty()) {
304    DL_ERR("out of memory when loading \"%s\"", name);
305    return NULL;
306  }
307
308  // Take the head element off the free list.
309  soinfo* si = gSoInfoFreeList;
310  gSoInfoFreeList = gSoInfoFreeList->next;
311
312  // Initialize the new element.
313  memset(si, 0, sizeof(soinfo));
314  strlcpy(si->name, name, sizeof(si->name));
315  sonext->next = si;
316  sonext = si;
317
318  TRACE("name %s: allocated soinfo @ %p", name, si);
319  return si;
320}
321
322static void soinfo_free(soinfo* si)
323{
324    if (si == NULL) {
325        return;
326    }
327
328    soinfo *prev = NULL, *trav;
329
330    TRACE("name %s: freeing soinfo @ %p", si->name, si);
331
332    for (trav = solist; trav != NULL; trav = trav->next) {
333        if (trav == si)
334            break;
335        prev = trav;
336    }
337    if (trav == NULL) {
338        /* si was not in solist */
339        DL_ERR("name \"%s\" is not in solist!", si->name);
340        return;
341    }
342
343    /* prev will never be NULL, because the first entry in solist is
344       always the static libdl_info.
345    */
346    prev->next = si->next;
347    if (si == sonext) {
348        sonext = prev;
349    }
350    si->next = gSoInfoFreeList;
351    gSoInfoFreeList = si;
352}
353
354
355static void parse_path(const char* path, const char* delimiters,
356                       const char** array, char* buf, size_t buf_size, size_t max_count) {
357  if (path == NULL) {
358    return;
359  }
360
361  size_t len = strlcpy(buf, path, buf_size);
362
363  size_t i = 0;
364  char* buf_p = buf;
365  while (i < max_count && (array[i] = strsep(&buf_p, delimiters))) {
366    if (*array[i] != '\0') {
367      ++i;
368    }
369  }
370
371  // Forget the last path if we had to truncate; this occurs if the 2nd to
372  // last char isn't '\0' (i.e. wasn't originally a delimiter).
373  if (i > 0 && len >= buf_size && buf[buf_size - 2] != '\0') {
374    array[i - 1] = NULL;
375  } else {
376    array[i] = NULL;
377  }
378}
379
380static void parse_LD_LIBRARY_PATH(const char* path) {
381  parse_path(path, ":", gLdPaths,
382             gLdPathsBuffer, sizeof(gLdPathsBuffer), LDPATH_MAX);
383}
384
385static void parse_LD_PRELOAD(const char* path) {
386  // We have historically supported ':' as well as ' ' in LD_PRELOAD.
387  parse_path(path, " :", gLdPreloadNames,
388             gLdPreloadsBuffer, sizeof(gLdPreloadsBuffer), LDPRELOAD_MAX);
389}
390
391#ifdef ANDROID_ARM_LINKER
392
393/* For a given PC, find the .so that it belongs to.
394 * Returns the base address of the .ARM.exidx section
395 * for that .so, and the number of 8-byte entries
396 * in that section (via *pcount).
397 *
398 * Intended to be called by libc's __gnu_Unwind_Find_exidx().
399 *
400 * This function is exposed via dlfcn.cpp and libdl.so.
401 */
402_Unwind_Ptr dl_unwind_find_exidx(_Unwind_Ptr pc, int *pcount)
403{
404    soinfo *si;
405    unsigned addr = (unsigned)pc;
406
407    for (si = solist; si != 0; si = si->next){
408        if ((addr >= si->base) && (addr < (si->base + si->size))) {
409            *pcount = si->ARM_exidx_count;
410            return (_Unwind_Ptr)si->ARM_exidx;
411        }
412    }
413   *pcount = 0;
414    return NULL;
415}
416
417#endif
418
419/* Here, we only have to provide a callback to iterate across all the
420 * loaded libraries. gcc_eh does the rest. */
421int
422dl_iterate_phdr(int (*cb)(dl_phdr_info *info, size_t size, void *data),
423                void *data)
424{
425    int rv = 0;
426    for (soinfo* si = solist; si != NULL; si = si->next) {
427        dl_phdr_info dl_info;
428        dl_info.dlpi_addr = si->link_map.l_addr;
429        dl_info.dlpi_name = si->link_map.l_name;
430        dl_info.dlpi_phdr = si->phdr;
431        dl_info.dlpi_phnum = si->phnum;
432        rv = cb(&dl_info, sizeof(dl_phdr_info), data);
433        if (rv != 0) {
434            break;
435        }
436    }
437    return rv;
438}
439
440static Elf_Sym* soinfo_elf_lookup(soinfo* si, unsigned hash, const char* name) {
441    Elf_Sym* symtab = si->symtab;
442    const char* strtab = si->strtab;
443
444    TRACE_TYPE(LOOKUP, "SEARCH %s in %s@0x%08x %08x %zd",
445               name, si->name, si->base, hash, hash % si->nbucket);
446
447    for (unsigned n = si->bucket[hash % si->nbucket]; n != 0; n = si->chain[n]) {
448        Elf_Sym* s = symtab + n;
449        if (strcmp(strtab + s->st_name, name)) continue;
450
451            /* only concern ourselves with global and weak symbol definitions */
452        switch(ELF32_ST_BIND(s->st_info)){
453        case STB_GLOBAL:
454        case STB_WEAK:
455            if (s->st_shndx == SHN_UNDEF) {
456                continue;
457            }
458
459            TRACE_TYPE(LOOKUP, "FOUND %s in %s (%08x) %d",
460                       name, si->name, s->st_value, s->st_size);
461            return s;
462        }
463    }
464
465    return NULL;
466}
467
468static unsigned elfhash(const char* _name) {
469    const unsigned char* name = (const unsigned char*) _name;
470    unsigned h = 0, g;
471
472    while(*name) {
473        h = (h << 4) + *name++;
474        g = h & 0xf0000000;
475        h ^= g;
476        h ^= g >> 24;
477    }
478    return h;
479}
480
481static Elf_Sym* soinfo_do_lookup(soinfo* si, const char* name, soinfo** lsi, soinfo* needed[]) {
482    unsigned elf_hash = elfhash(name);
483    Elf_Sym* s = NULL;
484
485    if (si != NULL && somain != NULL) {
486
487        /*
488         * Local scope is executable scope. Just start looking into it right away
489         * for the shortcut.
490         */
491
492        if (si == somain) {
493            s = soinfo_elf_lookup(si, elf_hash, name);
494            if (s != NULL) {
495                *lsi = si;
496                goto done;
497            }
498        } else {
499            /* Order of symbol lookup is controlled by DT_SYMBOLIC flag */
500
501            /*
502             * If this object was built with symbolic relocations disabled, the
503             * first place to look to resolve external references is the main
504             * executable.
505             */
506
507            if (!si->has_DT_SYMBOLIC) {
508                DEBUG("%s: looking up %s in executable %s",
509                      si->name, name, somain->name);
510                s = soinfo_elf_lookup(somain, elf_hash, name);
511                if (s != NULL) {
512                    *lsi = somain;
513                    goto done;
514                }
515            }
516
517            /* Look for symbols in the local scope (the object who is
518             * searching). This happens with C++ templates on i386 for some
519             * reason.
520             *
521             * Notes on weak symbols:
522             * The ELF specs are ambiguous about treatment of weak definitions in
523             * dynamic linking.  Some systems return the first definition found
524             * and some the first non-weak definition.   This is system dependent.
525             * Here we return the first definition found for simplicity.  */
526
527            s = soinfo_elf_lookup(si, elf_hash, name);
528            if (s != NULL) {
529                *lsi = si;
530                goto done;
531            }
532
533            /*
534             * If this object was built with -Bsymbolic and symbol is not found
535             * in the local scope, try to find the symbol in the main executable.
536             */
537
538            if (si->has_DT_SYMBOLIC) {
539                DEBUG("%s: looking up %s in executable %s after local scope",
540                      si->name, name, somain->name);
541                s = soinfo_elf_lookup(somain, elf_hash, name);
542                if (s != NULL) {
543                    *lsi = somain;
544                    goto done;
545                }
546            }
547        }
548    }
549
550    /* Next, look for it in the preloads list */
551    for (int i = 0; gLdPreloads[i] != NULL; i++) {
552        s = soinfo_elf_lookup(gLdPreloads[i], elf_hash, name);
553        if (s != NULL) {
554            *lsi = gLdPreloads[i];
555            goto done;
556        }
557    }
558
559    for (int i = 0; needed[i] != NULL; i++) {
560        DEBUG("%s: looking up %s in %s",
561              si->name, name, needed[i]->name);
562        s = soinfo_elf_lookup(needed[i], elf_hash, name);
563        if (s != NULL) {
564            *lsi = needed[i];
565            goto done;
566        }
567    }
568
569done:
570    if (s != NULL) {
571        TRACE_TYPE(LOOKUP, "si %s sym %s s->st_value = 0x%08x, "
572                   "found in %s, base = 0x%08x, load bias = 0x%08x",
573                   si->name, name, s->st_value,
574                   (*lsi)->name, (*lsi)->base, (*lsi)->load_bias);
575        return s;
576    }
577
578    return NULL;
579}
580
581/* This is used by dlsym(3).  It performs symbol lookup only within the
582   specified soinfo object and not in any of its dependencies.
583
584   TODO: Only looking in the specified soinfo seems wrong. dlsym(3) says
585   that it should do a breadth first search through the dependency
586   tree. This agrees with the ELF spec (aka System V Application
587   Binary Interface) where in Chapter 5 it discuss resolving "Shared
588   Object Dependencies" in breadth first search order.
589 */
590Elf_Sym* dlsym_handle_lookup(soinfo* si, const char* name) {
591    return soinfo_elf_lookup(si, elfhash(name), name);
592}
593
594/* This is used by dlsym(3) to performs a global symbol lookup. If the
595   start value is null (for RTLD_DEFAULT), the search starts at the
596   beginning of the global solist. Otherwise the search starts at the
597   specified soinfo (for RTLD_NEXT).
598 */
599Elf_Sym* dlsym_linear_lookup(const char* name, soinfo** found, soinfo* start) {
600  unsigned elf_hash = elfhash(name);
601
602  if (start == NULL) {
603    start = solist;
604  }
605
606  Elf_Sym* s = NULL;
607  for (soinfo* si = start; (s == NULL) && (si != NULL); si = si->next) {
608    s = soinfo_elf_lookup(si, elf_hash, name);
609    if (s != NULL) {
610      *found = si;
611      break;
612    }
613  }
614
615  if (s != NULL) {
616    TRACE_TYPE(LOOKUP, "%s s->st_value = 0x%08x, found->base = 0x%08x",
617               name, s->st_value, (*found)->base);
618  }
619
620  return s;
621}
622
623soinfo* find_containing_library(const void* p) {
624  Elf_Addr address = reinterpret_cast<Elf_Addr>(p);
625  for (soinfo* si = solist; si != NULL; si = si->next) {
626    if (address >= si->base && address - si->base < si->size) {
627      return si;
628    }
629  }
630  return NULL;
631}
632
633Elf_Sym* dladdr_find_symbol(soinfo* si, const void* addr) {
634  Elf_Addr soaddr = reinterpret_cast<Elf_Addr>(addr) - si->base;
635
636  // Search the library's symbol table for any defined symbol which
637  // contains this address.
638  for (size_t i = 0; i < si->nchain; ++i) {
639    Elf_Sym* sym = &si->symtab[i];
640    if (sym->st_shndx != SHN_UNDEF &&
641        soaddr >= sym->st_value &&
642        soaddr < sym->st_value + sym->st_size) {
643      return sym;
644    }
645  }
646
647  return NULL;
648}
649
650#if 0
651static void dump(soinfo* si)
652{
653    Elf_Sym* s = si->symtab;
654    for (unsigned n = 0; n < si->nchain; n++) {
655        TRACE("%04d> %08x: %02x %04x %08x %08x %s", n, s,
656               s->st_info, s->st_shndx, s->st_value, s->st_size,
657               si->strtab + s->st_name);
658        s++;
659    }
660}
661#endif
662
663static int open_library_on_path(const char* name, const char* const paths[]) {
664  char buf[512];
665  for (size_t i = 0; paths[i] != NULL; ++i) {
666    int n = __libc_format_buffer(buf, sizeof(buf), "%s/%s", paths[i], name);
667    if (n < 0 || n >= static_cast<int>(sizeof(buf))) {
668      PRINT("Warning: ignoring very long library path: %s/%s", paths[i], name);
669      continue;
670    }
671    int fd = TEMP_FAILURE_RETRY(open(buf, O_RDONLY | O_CLOEXEC));
672    if (fd != -1) {
673      return fd;
674    }
675  }
676  return -1;
677}
678
679static int open_library(const char* name) {
680  TRACE("[ opening %s ]", name);
681
682  // If the name contains a slash, we should attempt to open it directly and not search the paths.
683  if (strchr(name, '/') != NULL) {
684    int fd = TEMP_FAILURE_RETRY(open(name, O_RDONLY | O_CLOEXEC));
685    if (fd != -1) {
686      return fd;
687    }
688    // ...but nvidia binary blobs (at least) rely on this behavior, so fall through for now.
689  }
690
691  // Otherwise we try LD_LIBRARY_PATH first, and fall back to the built-in well known paths.
692  int fd = open_library_on_path(name, gLdPaths);
693  if (fd == -1) {
694    fd = open_library_on_path(name, gSoPaths);
695  }
696  return fd;
697}
698
699static soinfo* load_library(const char* name) {
700    // Open the file.
701    int fd = open_library(name);
702    if (fd == -1) {
703        DL_ERR("library \"%s\" not found", name);
704        return NULL;
705    }
706
707    // Read the ELF header and load the segments.
708    ElfReader elf_reader(name, fd);
709    if (!elf_reader.Load()) {
710        return NULL;
711    }
712
713    const char* bname = strrchr(name, '/');
714    soinfo* si = soinfo_alloc(bname ? bname + 1 : name);
715    if (si == NULL) {
716        return NULL;
717    }
718    si->base = elf_reader.load_start();
719    si->size = elf_reader.load_size();
720    si->load_bias = elf_reader.load_bias();
721    si->flags = 0;
722    si->entry = 0;
723    si->dynamic = NULL;
724    si->phnum = elf_reader.phdr_count();
725    si->phdr = elf_reader.loaded_phdr();
726    return si;
727}
728
729static soinfo *find_loaded_library(const char *name)
730{
731    soinfo *si;
732    const char *bname;
733
734    // TODO: don't use basename only for determining libraries
735    // http://code.google.com/p/android/issues/detail?id=6670
736
737    bname = strrchr(name, '/');
738    bname = bname ? bname + 1 : name;
739
740    for (si = solist; si != NULL; si = si->next) {
741        if (!strcmp(bname, si->name)) {
742            return si;
743        }
744    }
745    return NULL;
746}
747
748static soinfo* find_library_internal(const char* name) {
749  if (name == NULL) {
750    return somain;
751  }
752
753  soinfo* si = find_loaded_library(name);
754  if (si != NULL) {
755    if (si->flags & FLAG_LINKED) {
756      return si;
757    }
758    DL_ERR("OOPS: recursive link to \"%s\"", si->name);
759    return NULL;
760  }
761
762  TRACE("[ '%s' has not been loaded yet.  Locating...]", name);
763  si = load_library(name);
764  if (si == NULL) {
765    return NULL;
766  }
767
768  // At this point we know that whatever is loaded @ base is a valid ELF
769  // shared library whose segments are properly mapped in.
770  TRACE("[ init_library base=0x%08x sz=0x%08x name='%s' ]",
771        si->base, si->size, si->name);
772
773  if (!soinfo_link_image(si)) {
774    munmap(reinterpret_cast<void*>(si->base), si->size);
775    soinfo_free(si);
776    return NULL;
777  }
778
779  return si;
780}
781
782static soinfo* find_library(const char* name) {
783  soinfo* si = find_library_internal(name);
784  if (si != NULL) {
785    si->ref_count++;
786  }
787  return si;
788}
789
790static int soinfo_unload(soinfo* si) {
791  if (si->ref_count == 1) {
792    TRACE("unloading '%s'", si->name);
793    si->CallDestructors();
794
795    for (Elf_Dyn* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
796      if (d->d_tag == DT_NEEDED) {
797        const char* library_name = si->strtab + d->d_un.d_val;
798        TRACE("%s needs to unload %s", si->name, library_name);
799        soinfo_unload(find_loaded_library(library_name));
800      }
801    }
802
803    munmap(reinterpret_cast<void*>(si->base), si->size);
804    notify_gdb_of_unload(si);
805    soinfo_free(si);
806    si->ref_count = 0;
807  } else {
808    si->ref_count--;
809    TRACE("not unloading '%s', decrementing ref_count to %zd", si->name, si->ref_count);
810  }
811  return 0;
812}
813
814void do_android_update_LD_LIBRARY_PATH(const char* ld_library_path) {
815  if (!get_AT_SECURE()) {
816    parse_LD_LIBRARY_PATH(ld_library_path);
817  }
818}
819
820soinfo* do_dlopen(const char* name, int flags) {
821  if ((flags & ~(RTLD_NOW|RTLD_LAZY|RTLD_LOCAL|RTLD_GLOBAL)) != 0) {
822    DL_ERR("invalid flags to dlopen: %x", flags);
823    return NULL;
824  }
825  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
826  soinfo* si = find_library(name);
827  if (si != NULL) {
828    si->CallConstructors();
829  }
830  set_soinfo_pool_protection(PROT_READ);
831  return si;
832}
833
834int do_dlclose(soinfo* si) {
835  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
836  int result = soinfo_unload(si);
837  set_soinfo_pool_protection(PROT_READ);
838  return result;
839}
840
841/* TODO: don't use unsigned for addrs below. It works, but is not
842 * ideal. They should probably be either uint32_t, Elf_Addr, or unsigned
843 * long.
844 */
845static int soinfo_relocate(soinfo* si, Elf_Rel* rel, unsigned count,
846                           soinfo* needed[])
847{
848    Elf_Sym* symtab = si->symtab;
849    const char* strtab = si->strtab;
850    Elf_Sym* s;
851    Elf_Rel* start = rel;
852    soinfo* lsi;
853
854    for (size_t idx = 0; idx < count; ++idx, ++rel) {
855        unsigned type = ELF32_R_TYPE(rel->r_info);
856        unsigned sym = ELF32_R_SYM(rel->r_info);
857        Elf_Addr reloc = static_cast<Elf_Addr>(rel->r_offset + si->load_bias);
858        Elf_Addr sym_addr = 0;
859        char* sym_name = NULL;
860
861        DEBUG("Processing '%s' relocation at index %zd", si->name, idx);
862        if (type == 0) { // R_*_NONE
863            continue;
864        }
865        if (sym != 0) {
866            sym_name = (char *)(strtab + symtab[sym].st_name);
867            s = soinfo_do_lookup(si, sym_name, &lsi, needed);
868            if (s == NULL) {
869                /* We only allow an undefined symbol if this is a weak
870                   reference..   */
871                s = &symtab[sym];
872                if (ELF32_ST_BIND(s->st_info) != STB_WEAK) {
873                    DL_ERR("cannot locate symbol \"%s\" referenced by \"%s\"...", sym_name, si->name);
874                    return -1;
875                }
876
877                /* IHI0044C AAELF 4.5.1.1:
878
879                   Libraries are not searched to resolve weak references.
880                   It is not an error for a weak reference to remain
881                   unsatisfied.
882
883                   During linking, the value of an undefined weak reference is:
884                   - Zero if the relocation type is absolute
885                   - The address of the place if the relocation is pc-relative
886                   - The address of nominal base address if the relocation
887                     type is base-relative.
888                  */
889
890                switch (type) {
891#if defined(ANDROID_ARM_LINKER)
892                case R_ARM_JUMP_SLOT:
893                case R_ARM_GLOB_DAT:
894                case R_ARM_ABS32:
895                case R_ARM_RELATIVE:    /* Don't care. */
896#elif defined(ANDROID_X86_LINKER)
897                case R_386_JMP_SLOT:
898                case R_386_GLOB_DAT:
899                case R_386_32:
900                case R_386_RELATIVE:    /* Dont' care. */
901#endif /* ANDROID_*_LINKER */
902                    /* sym_addr was initialized to be zero above or relocation
903                       code below does not care about value of sym_addr.
904                       No need to do anything.  */
905                    break;
906
907#if defined(ANDROID_X86_LINKER)
908                case R_386_PC32:
909                    sym_addr = reloc;
910                    break;
911#endif /* ANDROID_X86_LINKER */
912
913#if defined(ANDROID_ARM_LINKER)
914                case R_ARM_COPY:
915                    /* Fall through.  Can't really copy if weak symbol is
916                       not found in run-time.  */
917#endif /* ANDROID_ARM_LINKER */
918                default:
919                    DL_ERR("unknown weak reloc type %d @ %p (%d)",
920                                 type, rel, (int) (rel - start));
921                    return -1;
922                }
923            } else {
924                /* We got a definition.  */
925#if 0
926                if ((base == 0) && (si->base != 0)) {
927                        /* linking from libraries to main image is bad */
928                    DL_ERR("cannot locate \"%s\"...",
929                           strtab + symtab[sym].st_name);
930                    return -1;
931                }
932#endif
933                sym_addr = static_cast<Elf_Addr>(s->st_value + lsi->load_bias);
934            }
935            count_relocation(kRelocSymbol);
936        } else {
937            s = NULL;
938        }
939
940/* TODO: This is ugly. Split up the relocations by arch into
941 * different files.
942 */
943        switch(type){
944#if defined(ANDROID_ARM_LINKER)
945        case R_ARM_JUMP_SLOT:
946            count_relocation(kRelocAbsolute);
947            MARK(rel->r_offset);
948            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
949            *reinterpret_cast<Elf_Addr*>(reloc) = sym_addr;
950            break;
951        case R_ARM_GLOB_DAT:
952            count_relocation(kRelocAbsolute);
953            MARK(rel->r_offset);
954            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
955            *reinterpret_cast<Elf_Addr*>(reloc) = sym_addr;
956            break;
957        case R_ARM_ABS32:
958            count_relocation(kRelocAbsolute);
959            MARK(rel->r_offset);
960            TRACE_TYPE(RELO, "RELO ABS %08x <- %08x %s", reloc, sym_addr, sym_name);
961            *reinterpret_cast<Elf_Addr*>(reloc) += sym_addr;
962            break;
963        case R_ARM_REL32:
964            count_relocation(kRelocRelative);
965            MARK(rel->r_offset);
966            TRACE_TYPE(RELO, "RELO REL32 %08x <- %08x - %08x %s",
967                       reloc, sym_addr, rel->r_offset, sym_name);
968            *reinterpret_cast<Elf_Addr*>(reloc) += sym_addr - rel->r_offset;
969            break;
970#elif defined(ANDROID_X86_LINKER)
971        case R_386_JMP_SLOT:
972            count_relocation(kRelocAbsolute);
973            MARK(rel->r_offset);
974            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
975            *reinterpret_cast<Elf_Addr*>(reloc) = sym_addr;
976            break;
977        case R_386_GLOB_DAT:
978            count_relocation(kRelocAbsolute);
979            MARK(rel->r_offset);
980            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
981            *reinterpret_cast<Elf_Addr*>(reloc) = sym_addr;
982            break;
983#elif defined(ANDROID_MIPS_LINKER)
984    case R_MIPS_REL32:
985            count_relocation(kRelocAbsolute);
986            MARK(rel->r_offset);
987            TRACE_TYPE(RELO, "RELO REL32 %08x <- %08x %s",
988                       reloc, sym_addr, (sym_name) ? sym_name : "*SECTIONHDR*");
989            if (s) {
990                *reinterpret_cast<Elf_Addr*>(reloc) += sym_addr;
991            } else {
992                *reinterpret_cast<Elf_Addr*>(reloc) += si->base;
993            }
994            break;
995#endif /* ANDROID_*_LINKER */
996
997#if defined(ANDROID_ARM_LINKER)
998        case R_ARM_RELATIVE:
999#elif defined(ANDROID_X86_LINKER)
1000        case R_386_RELATIVE:
1001#endif /* ANDROID_*_LINKER */
1002            count_relocation(kRelocRelative);
1003            MARK(rel->r_offset);
1004            if (sym) {
1005                DL_ERR("odd RELATIVE form...");
1006                return -1;
1007            }
1008            TRACE_TYPE(RELO, "RELO RELATIVE %08x <- +%08x", reloc, si->base);
1009            *reinterpret_cast<Elf_Addr*>(reloc) += si->base;
1010            break;
1011
1012#if defined(ANDROID_X86_LINKER)
1013        case R_386_32:
1014            count_relocation(kRelocRelative);
1015            MARK(rel->r_offset);
1016
1017            TRACE_TYPE(RELO, "RELO R_386_32 %08x <- +%08x %s", reloc, sym_addr, sym_name);
1018            *reinterpret_cast<Elf_Addr*>(reloc) += sym_addr;
1019            break;
1020
1021        case R_386_PC32:
1022            count_relocation(kRelocRelative);
1023            MARK(rel->r_offset);
1024            TRACE_TYPE(RELO, "RELO R_386_PC32 %08x <- +%08x (%08x - %08x) %s",
1025                       reloc, (sym_addr - reloc), sym_addr, reloc, sym_name);
1026            *reinterpret_cast<Elf_Addr*>(reloc) += (sym_addr - reloc);
1027            break;
1028#endif /* ANDROID_X86_LINKER */
1029
1030#ifdef ANDROID_ARM_LINKER
1031        case R_ARM_COPY:
1032            if ((si->flags & FLAG_EXE) == 0) {
1033                /*
1034                 * http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044d/IHI0044D_aaelf.pdf
1035                 *
1036                 * Section 4.7.1.10 "Dynamic relocations"
1037                 * R_ARM_COPY may only appear in executable objects where e_type is
1038                 * set to ET_EXEC.
1039                 *
1040                 * TODO: FLAG_EXE is set for both ET_DYN and ET_EXEC executables.
1041                 * We should explicitly disallow ET_DYN executables from having
1042                 * R_ARM_COPY relocations.
1043                 */
1044                DL_ERR("%s R_ARM_COPY relocations only supported for ET_EXEC", si->name);
1045                return -1;
1046            }
1047            count_relocation(kRelocCopy);
1048            MARK(rel->r_offset);
1049            TRACE_TYPE(RELO, "RELO %08x <- %d @ %08x %s", reloc, s->st_size, sym_addr, sym_name);
1050            if (reloc == sym_addr) {
1051                Elf_Sym *src = soinfo_do_lookup(NULL, sym_name, &lsi, needed);
1052
1053                if (src == NULL) {
1054                    DL_ERR("%s R_ARM_COPY relocation source cannot be resolved", si->name);
1055                    return -1;
1056                }
1057                if (lsi->has_DT_SYMBOLIC) {
1058                    DL_ERR("%s invalid R_ARM_COPY relocation against DT_SYMBOLIC shared "
1059                           "library %s (built with -Bsymbolic?)", si->name, lsi->name);
1060                    return -1;
1061                }
1062                if (s->st_size < src->st_size) {
1063                    DL_ERR("%s R_ARM_COPY relocation size mismatch (%d < %d)",
1064                           si->name, s->st_size, src->st_size);
1065                    return -1;
1066                }
1067                memcpy((void*)reloc, (void*)(src->st_value + lsi->load_bias), src->st_size);
1068            } else {
1069                DL_ERR("%s R_ARM_COPY relocation target cannot be resolved", si->name);
1070                return -1;
1071            }
1072            break;
1073#endif /* ANDROID_ARM_LINKER */
1074
1075        default:
1076            DL_ERR("unknown reloc type %d @ %p (%d)",
1077                   type, rel, (int) (rel - start));
1078            return -1;
1079        }
1080    }
1081    return 0;
1082}
1083
1084#ifdef ANDROID_MIPS_LINKER
1085static bool mips_relocate_got(soinfo* si, soinfo* needed[]) {
1086    unsigned* got = si->plt_got;
1087    if (got == NULL) {
1088        return true;
1089    }
1090    unsigned local_gotno = si->mips_local_gotno;
1091    unsigned gotsym = si->mips_gotsym;
1092    unsigned symtabno = si->mips_symtabno;
1093    Elf_Sym* symtab = si->symtab;
1094
1095    /*
1096     * got[0] is address of lazy resolver function
1097     * got[1] may be used for a GNU extension
1098     * set it to a recognizable address in case someone calls it
1099     * (should be _rtld_bind_start)
1100     * FIXME: maybe this should be in a separate routine
1101     */
1102
1103    if ((si->flags & FLAG_LINKER) == 0) {
1104        size_t g = 0;
1105        got[g++] = 0xdeadbeef;
1106        if (got[g] & 0x80000000) {
1107            got[g++] = 0xdeadfeed;
1108        }
1109        /*
1110         * Relocate the local GOT entries need to be relocated
1111         */
1112        for (; g < local_gotno; g++) {
1113            got[g] += si->load_bias;
1114        }
1115    }
1116
1117    /* Now for the global GOT entries */
1118    Elf_Sym* sym = symtab + gotsym;
1119    got = si->plt_got + local_gotno;
1120    for (size_t g = gotsym; g < symtabno; g++, sym++, got++) {
1121        const char* sym_name;
1122        Elf_Sym* s;
1123        soinfo* lsi;
1124
1125        /* This is an undefined reference... try to locate it */
1126        sym_name = si->strtab + sym->st_name;
1127        s = soinfo_do_lookup(si, sym_name, &lsi, needed);
1128        if (s == NULL) {
1129            /* We only allow an undefined symbol if this is a weak
1130               reference..   */
1131            s = &symtab[g];
1132            if (ELF32_ST_BIND(s->st_info) != STB_WEAK) {
1133                DL_ERR("cannot locate \"%s\"...", sym_name);
1134                return false;
1135            }
1136            *got = 0;
1137        }
1138        else {
1139            /* FIXME: is this sufficient?
1140             * For reference see NetBSD link loader
1141             * 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
1142             */
1143             *got = lsi->load_bias + s->st_value;
1144        }
1145    }
1146    return true;
1147}
1148#endif
1149
1150void soinfo::CallArray(const char* array_name UNUSED, linker_function_t* functions, size_t count, bool reverse) {
1151  if (functions == NULL) {
1152    return;
1153  }
1154
1155  TRACE("[ Calling %s (size %zd) @ %p for '%s' ]", array_name, count, functions, name);
1156
1157  int begin = reverse ? (count - 1) : 0;
1158  int end = reverse ? -1 : count;
1159  int step = reverse ? -1 : 1;
1160
1161  for (int i = begin; i != end; i += step) {
1162    TRACE("[ %s[%d] == %p ]", array_name, i, functions[i]);
1163    CallFunction("function", functions[i]);
1164  }
1165
1166  TRACE("[ Done calling %s for '%s' ]", array_name, name);
1167}
1168
1169void soinfo::CallFunction(const char* function_name UNUSED, linker_function_t function) {
1170  if (function == NULL || reinterpret_cast<uintptr_t>(function) == static_cast<uintptr_t>(-1)) {
1171    return;
1172  }
1173
1174  TRACE("[ Calling %s @ %p for '%s' ]", function_name, function, name);
1175  function();
1176  TRACE("[ Done calling %s @ %p for '%s' ]", function_name, function, name);
1177
1178  // The function may have called dlopen(3) or dlclose(3), so we need to ensure our data structures
1179  // are still writable. This happens with our debug malloc (see http://b/7941716).
1180  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
1181}
1182
1183void soinfo::CallPreInitConstructors() {
1184  // DT_PREINIT_ARRAY functions are called before any other constructors for executables,
1185  // but ignored in a shared library.
1186  CallArray("DT_PREINIT_ARRAY", preinit_array, preinit_array_count, false);
1187}
1188
1189void soinfo::CallConstructors() {
1190  if (constructors_called) {
1191    return;
1192  }
1193
1194  // We set constructors_called before actually calling the constructors, otherwise it doesn't
1195  // protect against recursive constructor calls. One simple example of constructor recursion
1196  // is the libc debug malloc, which is implemented in libc_malloc_debug_leak.so:
1197  // 1. The program depends on libc, so libc's constructor is called here.
1198  // 2. The libc constructor calls dlopen() to load libc_malloc_debug_leak.so.
1199  // 3. dlopen() calls the constructors on the newly created
1200  //    soinfo for libc_malloc_debug_leak.so.
1201  // 4. The debug .so depends on libc, so CallConstructors is
1202  //    called again with the libc soinfo. If it doesn't trigger the early-
1203  //    out above, the libc constructor will be called again (recursively!).
1204  constructors_called = true;
1205
1206  if ((flags & FLAG_EXE) == 0 && preinit_array != NULL) {
1207    // The GNU dynamic linker silently ignores these, but we warn the developer.
1208    PRINT("\"%s\": ignoring %zd-entry DT_PREINIT_ARRAY in shared library!",
1209          name, preinit_array_count);
1210  }
1211
1212  if (dynamic != NULL) {
1213    for (Elf_Dyn* d = dynamic; d->d_tag != DT_NULL; ++d) {
1214      if (d->d_tag == DT_NEEDED) {
1215        const char* library_name = strtab + d->d_un.d_val;
1216        TRACE("\"%s\": calling constructors in DT_NEEDED \"%s\"", name, library_name);
1217        find_loaded_library(library_name)->CallConstructors();
1218      }
1219    }
1220  }
1221
1222  TRACE("\"%s\": calling constructors", name);
1223
1224  // DT_INIT should be called before DT_INIT_ARRAY if both are present.
1225  CallFunction("DT_INIT", init_func);
1226  CallArray("DT_INIT_ARRAY", init_array, init_array_count, false);
1227}
1228
1229void soinfo::CallDestructors() {
1230  TRACE("\"%s\": calling destructors", name);
1231
1232  // DT_FINI_ARRAY must be parsed in reverse order.
1233  CallArray("DT_FINI_ARRAY", fini_array, fini_array_count, true);
1234
1235  // DT_FINI should be called after DT_FINI_ARRAY if both are present.
1236  CallFunction("DT_FINI", fini_func);
1237}
1238
1239/* Force any of the closed stdin, stdout and stderr to be associated with
1240   /dev/null. */
1241static int nullify_closed_stdio() {
1242    int dev_null, i, status;
1243    int return_value = 0;
1244
1245    dev_null = TEMP_FAILURE_RETRY(open("/dev/null", O_RDWR));
1246    if (dev_null < 0) {
1247        DL_ERR("cannot open /dev/null: %s", strerror(errno));
1248        return -1;
1249    }
1250    TRACE("[ Opened /dev/null file-descriptor=%d]", dev_null);
1251
1252    /* If any of the stdio file descriptors is valid and not associated
1253       with /dev/null, dup /dev/null to it.  */
1254    for (i = 0; i < 3; i++) {
1255        /* If it is /dev/null already, we are done. */
1256        if (i == dev_null) {
1257            continue;
1258        }
1259
1260        TRACE("[ Nullifying stdio file descriptor %d]", i);
1261        status = TEMP_FAILURE_RETRY(fcntl(i, F_GETFL));
1262
1263        /* If file is opened, we are good. */
1264        if (status != -1) {
1265            continue;
1266        }
1267
1268        /* The only error we allow is that the file descriptor does not
1269           exist, in which case we dup /dev/null to it. */
1270        if (errno != EBADF) {
1271            DL_ERR("fcntl failed: %s", strerror(errno));
1272            return_value = -1;
1273            continue;
1274        }
1275
1276        /* Try dupping /dev/null to this stdio file descriptor and
1277           repeat if there is a signal.  Note that any errors in closing
1278           the stdio descriptor are lost.  */
1279        status = TEMP_FAILURE_RETRY(dup2(dev_null, i));
1280        if (status < 0) {
1281            DL_ERR("dup2 failed: %s", strerror(errno));
1282            return_value = -1;
1283            continue;
1284        }
1285    }
1286
1287    /* If /dev/null is not one of the stdio file descriptors, close it. */
1288    if (dev_null > 2) {
1289        TRACE("[ Closing /dev/null file-descriptor=%d]", dev_null);
1290        status = TEMP_FAILURE_RETRY(close(dev_null));
1291        if (status == -1) {
1292            DL_ERR("close failed: %s", strerror(errno));
1293            return_value = -1;
1294        }
1295    }
1296
1297    return return_value;
1298}
1299
1300static bool soinfo_link_image(soinfo* si) {
1301    /* "base" might wrap around UINT32_MAX. */
1302    Elf_Addr base = si->load_bias;
1303    const Elf_Phdr *phdr = si->phdr;
1304    int phnum = si->phnum;
1305    bool relocating_linker = (si->flags & FLAG_LINKER) != 0;
1306
1307    /* We can't debug anything until the linker is relocated */
1308    if (!relocating_linker) {
1309        INFO("[ linking %s ]", si->name);
1310        DEBUG("si->base = 0x%08x si->flags = 0x%08x", si->base, si->flags);
1311    }
1312
1313    /* Extract dynamic section */
1314    size_t dynamic_count;
1315    Elf_Word dynamic_flags;
1316    phdr_table_get_dynamic_section(phdr, phnum, base, &si->dynamic,
1317                                   &dynamic_count, &dynamic_flags);
1318    if (si->dynamic == NULL) {
1319        if (!relocating_linker) {
1320            DL_ERR("missing PT_DYNAMIC in \"%s\"", si->name);
1321        }
1322        return false;
1323    } else {
1324        if (!relocating_linker) {
1325            DEBUG("dynamic = %p", si->dynamic);
1326        }
1327    }
1328
1329#ifdef ANDROID_ARM_LINKER
1330    (void) phdr_table_get_arm_exidx(phdr, phnum, base,
1331                                    &si->ARM_exidx, &si->ARM_exidx_count);
1332#endif
1333
1334    // Extract useful information from dynamic section.
1335    uint32_t needed_count = 0;
1336    for (Elf_Dyn* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1337        DEBUG("d = %p, d[0](tag) = 0x%08x d[1](val) = 0x%08x", d, d->d_tag, d->d_un.d_val);
1338        switch(d->d_tag){
1339        case DT_HASH:
1340            si->nbucket = ((unsigned *) (base + d->d_un.d_ptr))[0];
1341            si->nchain = ((unsigned *) (base + d->d_un.d_ptr))[1];
1342            si->bucket = (unsigned *) (base + d->d_un.d_ptr + 8);
1343            si->chain = (unsigned *) (base + d->d_un.d_ptr + 8 + si->nbucket * 4);
1344            break;
1345        case DT_STRTAB:
1346            si->strtab = (const char *) (base + d->d_un.d_ptr);
1347            break;
1348        case DT_SYMTAB:
1349            si->symtab = (Elf_Sym *) (base + d->d_un.d_ptr);
1350            break;
1351        case DT_PLTREL:
1352            if (d->d_un.d_val != DT_REL) {
1353                DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1354                return false;
1355            }
1356            break;
1357        case DT_JMPREL:
1358            si->plt_rel = (Elf_Rel*) (base + d->d_un.d_ptr);
1359            break;
1360        case DT_PLTRELSZ:
1361            si->plt_rel_count = d->d_un.d_val / sizeof(Elf_Rel);
1362            break;
1363        case DT_REL:
1364            si->rel = (Elf_Rel*) (base + d->d_un.d_ptr);
1365            break;
1366        case DT_RELSZ:
1367            si->rel_count = d->d_un.d_val / sizeof(Elf_Rel);
1368            break;
1369        case DT_PLTGOT:
1370            /* Save this in case we decide to do lazy binding. We don't yet. */
1371            si->plt_got = (unsigned *)(base + d->d_un.d_ptr);
1372            break;
1373        case DT_DEBUG:
1374            // Set the DT_DEBUG entry to the address of _r_debug for GDB
1375            // if the dynamic table is writable
1376            if ((dynamic_flags & PF_W) != 0) {
1377                d->d_un.d_val = reinterpret_cast<uintptr_t>(&_r_debug);
1378            }
1379            break;
1380         case DT_RELA:
1381            DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1382            return false;
1383        case DT_INIT:
1384            si->init_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1385            DEBUG("%s constructors (DT_INIT) found at %p", si->name, si->init_func);
1386            break;
1387        case DT_FINI:
1388            si->fini_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1389            DEBUG("%s destructors (DT_FINI) found at %p", si->name, si->fini_func);
1390            break;
1391        case DT_INIT_ARRAY:
1392            si->init_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1393            DEBUG("%s constructors (DT_INIT_ARRAY) found at %p", si->name, si->init_array);
1394            break;
1395        case DT_INIT_ARRAYSZ:
1396            si->init_array_count = ((unsigned)d->d_un.d_val) / sizeof(Elf_Addr);
1397            break;
1398        case DT_FINI_ARRAY:
1399            si->fini_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1400            DEBUG("%s destructors (DT_FINI_ARRAY) found at %p", si->name, si->fini_array);
1401            break;
1402        case DT_FINI_ARRAYSZ:
1403            si->fini_array_count = ((unsigned)d->d_un.d_val) / sizeof(Elf_Addr);
1404            break;
1405        case DT_PREINIT_ARRAY:
1406            si->preinit_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1407            DEBUG("%s constructors (DT_PREINIT_ARRAY) found at %p", si->name, si->preinit_array);
1408            break;
1409        case DT_PREINIT_ARRAYSZ:
1410            si->preinit_array_count = ((unsigned)d->d_un.d_val) / sizeof(Elf_Addr);
1411            break;
1412        case DT_TEXTREL:
1413            si->has_text_relocations = true;
1414            break;
1415        case DT_SYMBOLIC:
1416            si->has_DT_SYMBOLIC = true;
1417            break;
1418        case DT_NEEDED:
1419            ++needed_count;
1420            break;
1421#if defined DT_FLAGS
1422        // TODO: why is DT_FLAGS not defined?
1423        case DT_FLAGS:
1424            if (d->d_un.d_val & DF_TEXTREL) {
1425                si->has_text_relocations = true;
1426            }
1427            if (d->d_un.d_val & DF_SYMBOLIC) {
1428                si->has_DT_SYMBOLIC = true;
1429            }
1430            break;
1431#endif
1432#if defined(ANDROID_MIPS_LINKER)
1433        case DT_STRSZ:
1434        case DT_SYMENT:
1435        case DT_RELENT:
1436             break;
1437        case DT_MIPS_RLD_MAP:
1438            // Set the DT_MIPS_RLD_MAP entry to the address of _r_debug for GDB.
1439            {
1440              r_debug** dp = (r_debug**) d->d_un.d_ptr;
1441              *dp = &_r_debug;
1442            }
1443            break;
1444        case DT_MIPS_RLD_VERSION:
1445        case DT_MIPS_FLAGS:
1446        case DT_MIPS_BASE_ADDRESS:
1447        case DT_MIPS_UNREFEXTNO:
1448            break;
1449
1450        case DT_MIPS_SYMTABNO:
1451            si->mips_symtabno = d->d_un.d_val;
1452            break;
1453
1454        case DT_MIPS_LOCAL_GOTNO:
1455            si->mips_local_gotno = d->d_un.d_val;
1456            break;
1457
1458        case DT_MIPS_GOTSYM:
1459            si->mips_gotsym = d->d_un.d_val;
1460            break;
1461
1462        default:
1463            DEBUG("Unused DT entry: type 0x%08x arg 0x%08x", d->d_tag, d->d_un.d_val);
1464            break;
1465#endif
1466        }
1467    }
1468
1469    DEBUG("si->base = 0x%08x, si->strtab = %p, si->symtab = %p",
1470          si->base, si->strtab, si->symtab);
1471
1472    // Sanity checks.
1473    if (relocating_linker && needed_count != 0) {
1474        DL_ERR("linker cannot have DT_NEEDED dependencies on other libraries");
1475        return false;
1476    }
1477    if (si->nbucket == 0) {
1478        DL_ERR("empty/missing DT_HASH in \"%s\" (built with --hash-style=gnu?)", si->name);
1479        return false;
1480    }
1481    if (si->strtab == 0) {
1482        DL_ERR("empty/missing DT_STRTAB in \"%s\"", si->name);
1483        return false;
1484    }
1485    if (si->symtab == 0) {
1486        DL_ERR("empty/missing DT_SYMTAB in \"%s\"", si->name);
1487        return false;
1488    }
1489
1490    // If this is the main executable, then load all of the libraries from LD_PRELOAD now.
1491    if (si->flags & FLAG_EXE) {
1492        memset(gLdPreloads, 0, sizeof(gLdPreloads));
1493        size_t preload_count = 0;
1494        for (size_t i = 0; gLdPreloadNames[i] != NULL; i++) {
1495            soinfo* lsi = find_library(gLdPreloadNames[i]);
1496            if (lsi != NULL) {
1497                gLdPreloads[preload_count++] = lsi;
1498            } else {
1499                // As with glibc, failure to load an LD_PRELOAD library is just a warning.
1500                DL_WARN("could not load library \"%s\" from LD_PRELOAD for \"%s\"; caused by %s",
1501                        gLdPreloadNames[i], si->name, linker_get_error_buffer());
1502            }
1503        }
1504    }
1505
1506    soinfo** needed = (soinfo**) alloca((1 + needed_count) * sizeof(soinfo*));
1507    soinfo** pneeded = needed;
1508
1509    for (Elf_Dyn* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1510        if (d->d_tag == DT_NEEDED) {
1511            const char* library_name = si->strtab + d->d_un.d_val;
1512            DEBUG("%s needs %s", si->name, library_name);
1513            soinfo* lsi = find_library(library_name);
1514            if (lsi == NULL) {
1515                strlcpy(tmp_err_buf, linker_get_error_buffer(), sizeof(tmp_err_buf));
1516                DL_ERR("could not load library \"%s\" needed by \"%s\"; caused by %s",
1517                       library_name, si->name, tmp_err_buf);
1518                return false;
1519            }
1520            *pneeded++ = lsi;
1521        }
1522    }
1523    *pneeded = NULL;
1524
1525    if (si->has_text_relocations) {
1526        /* Unprotect the segments, i.e. make them writable, to allow
1527         * text relocations to work properly. We will later call
1528         * phdr_table_protect_segments() after all of them are applied
1529         * and all constructors are run.
1530         */
1531        DL_WARN("%s has text relocations. This is wasting memory and is "
1532                "a security risk. Please fix.", si->name);
1533        if (phdr_table_unprotect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1534            DL_ERR("can't unprotect loadable segments for \"%s\": %s",
1535                   si->name, strerror(errno));
1536            return false;
1537        }
1538    }
1539
1540    if (si->plt_rel != NULL) {
1541        DEBUG("[ relocating %s plt ]", si->name );
1542        if (soinfo_relocate(si, si->plt_rel, si->plt_rel_count, needed)) {
1543            return false;
1544        }
1545    }
1546    if (si->rel != NULL) {
1547        DEBUG("[ relocating %s ]", si->name );
1548        if (soinfo_relocate(si, si->rel, si->rel_count, needed)) {
1549            return false;
1550        }
1551    }
1552
1553#ifdef ANDROID_MIPS_LINKER
1554    if (!mips_relocate_got(si, needed)) {
1555        return false;
1556    }
1557#endif
1558
1559    si->flags |= FLAG_LINKED;
1560    DEBUG("[ finished linking %s ]", si->name);
1561
1562    if (si->has_text_relocations) {
1563        /* All relocations are done, we can protect our segments back to
1564         * read-only. */
1565        if (phdr_table_protect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1566            DL_ERR("can't protect segments for \"%s\": %s",
1567                   si->name, strerror(errno));
1568            return false;
1569        }
1570    }
1571
1572    /* We can also turn on GNU RELRO protection */
1573    if (phdr_table_protect_gnu_relro(si->phdr, si->phnum, si->load_bias) < 0) {
1574        DL_ERR("can't enable GNU RELRO protection for \"%s\": %s",
1575               si->name, strerror(errno));
1576        return false;
1577    }
1578
1579    notify_gdb_of_load(si);
1580    return true;
1581}
1582
1583/*
1584 * This function add vdso to internal dso list.
1585 * It helps to stack unwinding through signal handlers.
1586 * Also, it makes bionic more like glibc.
1587 */
1588static void add_vdso(KernelArgumentBlock& args UNUSED) {
1589#ifdef AT_SYSINFO_EHDR
1590    Elf_Ehdr* ehdr_vdso = reinterpret_cast<Elf_Ehdr*>(args.getauxval(AT_SYSINFO_EHDR));
1591
1592    soinfo* si = soinfo_alloc("[vdso]");
1593    si->phdr = reinterpret_cast<Elf_Phdr*>(reinterpret_cast<char*>(ehdr_vdso) + ehdr_vdso->e_phoff);
1594    si->phnum = ehdr_vdso->e_phnum;
1595    si->link_map.l_name = si->name;
1596    for (size_t i = 0; i < si->phnum; ++i) {
1597        if (si->phdr[i].p_type == PT_LOAD) {
1598            si->link_map.l_addr = reinterpret_cast<Elf_Addr>(ehdr_vdso) - si->phdr[i].p_vaddr;
1599            break;
1600        }
1601    }
1602#endif
1603}
1604
1605/*
1606 * This code is called after the linker has linked itself and
1607 * fixed it's own GOT. It is safe to make references to externs
1608 * and other non-local data at this point.
1609 */
1610static Elf_Addr __linker_init_post_relocation(KernelArgumentBlock& args, Elf_Addr linker_base) {
1611    /* NOTE: we store the args pointer on a special location
1612     *       of the temporary TLS area in order to pass it to
1613     *       the C Library's runtime initializer.
1614     *
1615     *       The initializer must clear the slot and reset the TLS
1616     *       to point to a different location to ensure that no other
1617     *       shared library constructor can access it.
1618     */
1619  __libc_init_tls(args);
1620
1621#if TIMING
1622    struct timeval t0, t1;
1623    gettimeofday(&t0, 0);
1624#endif
1625
1626    // Initialize environment functions, and get to the ELF aux vectors table.
1627    linker_env_init(args);
1628
1629    // If this is a setuid/setgid program, close the security hole described in
1630    // ftp://ftp.freebsd.org/pub/FreeBSD/CERT/advisories/FreeBSD-SA-02:23.stdio.asc
1631    if (get_AT_SECURE()) {
1632        nullify_closed_stdio();
1633    }
1634
1635    debuggerd_init();
1636
1637    // Get a few environment variables.
1638    const char* LD_DEBUG = linker_env_get("LD_DEBUG");
1639    if (LD_DEBUG != NULL) {
1640      gLdDebugVerbosity = atoi(LD_DEBUG);
1641    }
1642
1643    // Normally, these are cleaned by linker_env_init, but the test
1644    // doesn't cost us anything.
1645    const char* ldpath_env = NULL;
1646    const char* ldpreload_env = NULL;
1647    if (!get_AT_SECURE()) {
1648      ldpath_env = linker_env_get("LD_LIBRARY_PATH");
1649      ldpreload_env = linker_env_get("LD_PRELOAD");
1650    }
1651
1652    INFO("[ android linker & debugger ]");
1653
1654    soinfo* si = soinfo_alloc(args.argv[0]);
1655    if (si == NULL) {
1656        exit(EXIT_FAILURE);
1657    }
1658
1659    /* bootstrap the link map, the main exe always needs to be first */
1660    si->flags |= FLAG_EXE;
1661    link_map_t* map = &(si->link_map);
1662
1663    map->l_addr = 0;
1664    map->l_name = args.argv[0];
1665    map->l_prev = NULL;
1666    map->l_next = NULL;
1667
1668    _r_debug.r_map = map;
1669    r_debug_tail = map;
1670
1671    /* gdb expects the linker to be in the debug shared object list.
1672     * Without this, gdb has trouble locating the linker's ".text"
1673     * and ".plt" sections. Gdb could also potentially use this to
1674     * relocate the offset of our exported 'rtld_db_dlactivity' symbol.
1675     * Don't use soinfo_alloc(), because the linker shouldn't
1676     * be on the soinfo list.
1677     */
1678    {
1679        static soinfo linker_soinfo;
1680        strlcpy(linker_soinfo.name, "/system/bin/linker", sizeof(linker_soinfo.name));
1681        linker_soinfo.flags = 0;
1682        linker_soinfo.base = linker_base;
1683
1684        /*
1685         * Set the dynamic field in the link map otherwise gdb will complain with
1686         * the following:
1687         *   warning: .dynamic section for "/system/bin/linker" is not at the
1688         *   expected address (wrong library or version mismatch?)
1689         */
1690        Elf_Ehdr *elf_hdr = (Elf_Ehdr *) linker_base;
1691        Elf_Phdr *phdr = (Elf_Phdr*)((unsigned char*) linker_base + elf_hdr->e_phoff);
1692        phdr_table_get_dynamic_section(phdr, elf_hdr->e_phnum, linker_base,
1693                                       &linker_soinfo.dynamic, NULL, NULL);
1694        insert_soinfo_into_debug_map(&linker_soinfo);
1695    }
1696
1697    // Extract information passed from the kernel.
1698    si->phdr = reinterpret_cast<Elf_Phdr*>(args.getauxval(AT_PHDR));
1699    si->phnum = args.getauxval(AT_PHNUM);
1700    si->entry = args.getauxval(AT_ENTRY);
1701
1702    /* Compute the value of si->base. We can't rely on the fact that
1703     * the first entry is the PHDR because this will not be true
1704     * for certain executables (e.g. some in the NDK unit test suite)
1705     */
1706    si->base = 0;
1707    si->size = phdr_table_get_load_size(si->phdr, si->phnum);
1708    si->load_bias = 0;
1709    for (size_t i = 0; i < si->phnum; ++i) {
1710      if (si->phdr[i].p_type == PT_PHDR) {
1711        si->load_bias = reinterpret_cast<Elf_Addr>(si->phdr) - si->phdr[i].p_vaddr;
1712        si->base = reinterpret_cast<Elf_Addr>(si->phdr) - si->phdr[i].p_offset;
1713        break;
1714      }
1715    }
1716    si->dynamic = NULL;
1717    si->ref_count = 1;
1718
1719    // Use LD_LIBRARY_PATH and LD_PRELOAD (but only if we aren't setuid/setgid).
1720    parse_LD_LIBRARY_PATH(ldpath_env);
1721    parse_LD_PRELOAD(ldpreload_env);
1722
1723    somain = si;
1724
1725    if (!soinfo_link_image(si)) {
1726        __libc_format_fd(2, "CANNOT LINK EXECUTABLE: %s\n", linker_get_error_buffer());
1727        exit(EXIT_FAILURE);
1728    }
1729
1730    add_vdso(args);
1731
1732    si->CallPreInitConstructors();
1733
1734    for (size_t i = 0; gLdPreloads[i] != NULL; ++i) {
1735        gLdPreloads[i]->CallConstructors();
1736    }
1737
1738    /* After the link_image, the si->load_bias is initialized.
1739     * For so lib, the map->l_addr will be updated in notify_gdb_of_load.
1740     * We need to update this value for so exe here. So Unwind_Backtrace
1741     * for some arch like x86 could work correctly within so exe.
1742     */
1743    map->l_addr = si->load_bias;
1744    si->CallConstructors();
1745
1746#if TIMING
1747    gettimeofday(&t1,NULL);
1748    PRINT("LINKER TIME: %s: %d microseconds", args.argv[0], (int) (
1749               (((long long)t1.tv_sec * 1000000LL) + (long long)t1.tv_usec) -
1750               (((long long)t0.tv_sec * 1000000LL) + (long long)t0.tv_usec)
1751               ));
1752#endif
1753#if STATS
1754    PRINT("RELO STATS: %s: %d abs, %d rel, %d copy, %d symbol", args.argv[0],
1755           linker_stats.count[kRelocAbsolute],
1756           linker_stats.count[kRelocRelative],
1757           linker_stats.count[kRelocCopy],
1758           linker_stats.count[kRelocSymbol]);
1759#endif
1760#if COUNT_PAGES
1761    {
1762        unsigned n;
1763        unsigned i;
1764        unsigned count = 0;
1765        for (n = 0; n < 4096; n++) {
1766            if (bitmask[n]) {
1767                unsigned x = bitmask[n];
1768                for (i = 0; i < 8; i++) {
1769                    if (x & 1) {
1770                        count++;
1771                    }
1772                    x >>= 1;
1773                }
1774            }
1775        }
1776        PRINT("PAGES MODIFIED: %s: %d (%dKB)", args.argv[0], count, count * 4);
1777    }
1778#endif
1779
1780#if TIMING || STATS || COUNT_PAGES
1781    fflush(stdout);
1782#endif
1783
1784    TRACE("[ Ready to execute '%s' @ 0x%08x ]", si->name, si->entry);
1785    return si->entry;
1786}
1787
1788/* Compute the load-bias of an existing executable. This shall only
1789 * be used to compute the load bias of an executable or shared library
1790 * that was loaded by the kernel itself.
1791 *
1792 * Input:
1793 *    elf    -> address of ELF header, assumed to be at the start of the file.
1794 * Return:
1795 *    load bias, i.e. add the value of any p_vaddr in the file to get
1796 *    the corresponding address in memory.
1797 */
1798static Elf_Addr get_elf_exec_load_bias(const Elf_Ehdr* elf) {
1799  Elf_Addr offset = elf->e_phoff;
1800  const Elf_Phdr* phdr_table = (const Elf_Phdr*)((char*)elf + offset);
1801  const Elf_Phdr* phdr_end = phdr_table + elf->e_phnum;
1802
1803  for (const Elf_Phdr* phdr = phdr_table; phdr < phdr_end; phdr++) {
1804    if (phdr->p_type == PT_LOAD) {
1805      return reinterpret_cast<Elf_Addr>(elf) + phdr->p_offset - phdr->p_vaddr;
1806    }
1807  }
1808  return 0;
1809}
1810
1811/*
1812 * This is the entry point for the linker, called from begin.S. This
1813 * method is responsible for fixing the linker's own relocations, and
1814 * then calling __linker_init_post_relocation().
1815 *
1816 * Because this method is called before the linker has fixed it's own
1817 * relocations, any attempt to reference an extern variable, extern
1818 * function, or other GOT reference will generate a segfault.
1819 */
1820extern "C" Elf_Addr __linker_init(void* raw_args) {
1821  KernelArgumentBlock args(raw_args);
1822
1823  Elf_Addr linker_addr = args.getauxval(AT_BASE);
1824
1825  Elf_Ehdr* elf_hdr = reinterpret_cast<Elf_Ehdr*>(linker_addr);
1826  Elf_Phdr* phdr = (Elf_Phdr*)((unsigned char*) linker_addr + elf_hdr->e_phoff);
1827
1828  soinfo linker_so;
1829  memset(&linker_so, 0, sizeof(soinfo));
1830
1831  linker_so.base = linker_addr;
1832  linker_so.size = phdr_table_get_load_size(phdr, elf_hdr->e_phnum);
1833  linker_so.load_bias = get_elf_exec_load_bias(elf_hdr);
1834  linker_so.dynamic = NULL;
1835  linker_so.phdr = phdr;
1836  linker_so.phnum = elf_hdr->e_phnum;
1837  linker_so.flags |= FLAG_LINKER;
1838
1839  if (!soinfo_link_image(&linker_so)) {
1840    // It would be nice to print an error message, but if the linker
1841    // can't link itself, there's no guarantee that we'll be able to
1842    // call write() (because it involves a GOT reference).
1843    //
1844    // This situation should never occur unless the linker itself
1845    // is corrupt.
1846    exit(EXIT_FAILURE);
1847  }
1848
1849  // We have successfully fixed our own relocations. It's safe to run
1850  // the main part of the linker now.
1851  args.abort_message_ptr = &gAbortMessage;
1852  Elf_Addr start_address = __linker_init_post_relocation(args, linker_addr);
1853
1854  set_soinfo_pool_protection(PROT_READ);
1855
1856  // Return the address that the calling assembly stub should jump to.
1857  return start_address;
1858}
1859