linker.cpp revision faf05bacd45719291b371f24b1b89543881b37f6
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    }
166#define UNUSED __attribute__((unused))
167DISALLOW_ALLOCATION(void*, malloc, (size_t u UNUSED));
168DISALLOW_ALLOCATION(void, free, (void* u UNUSED));
169DISALLOW_ALLOCATION(void*, realloc, (void* u1 UNUSED, size_t u2 UNUSED));
170DISALLOW_ALLOCATION(void*, calloc, (size_t u1 UNUSED, size_t u2 UNUSED));
171
172static char tmp_err_buf[768];
173static char __linker_dl_err_buf[768];
174
175char* linker_get_error_buffer() {
176  return &__linker_dl_err_buf[0];
177}
178
179size_t linker_get_error_buffer_size() {
180  return sizeof(__linker_dl_err_buf);
181}
182
183/*
184 * This function is an empty stub where GDB locates a breakpoint to get notified
185 * about linker activity.
186 */
187extern "C" void __attribute__((noinline)) __attribute__((visibility("default"))) rtld_db_dlactivity();
188
189static r_debug _r_debug = {1, NULL, reinterpret_cast<uintptr_t>(&rtld_db_dlactivity), r_debug::RT_CONSISTENT, 0};
190static link_map* r_debug_tail = 0;
191
192static pthread_mutex_t gDebugMutex = PTHREAD_MUTEX_INITIALIZER;
193
194static void insert_soinfo_into_debug_map(soinfo* info) {
195    // Copy the necessary fields into the debug structure.
196    link_map* map = &(info->link_map_head);
197    map->l_addr = info->load_bias;
198    map->l_name = reinterpret_cast<char*>(info->name);
199    map->l_ld = info->dynamic;
200
201    /* Stick the new library at the end of the list.
202     * gdb tends to care more about libc than it does
203     * about leaf libraries, and ordering it this way
204     * reduces the back-and-forth over the wire.
205     */
206    if (r_debug_tail) {
207        r_debug_tail->l_next = map;
208        map->l_prev = r_debug_tail;
209        map->l_next = 0;
210    } else {
211        _r_debug.r_map = map;
212        map->l_prev = 0;
213        map->l_next = 0;
214    }
215    r_debug_tail = map;
216}
217
218static void remove_soinfo_from_debug_map(soinfo* info) {
219    link_map* map = &(info->link_map_head);
220
221    if (r_debug_tail == map) {
222        r_debug_tail = map->l_prev;
223    }
224
225    if (map->l_prev) {
226        map->l_prev->l_next = map->l_next;
227    }
228    if (map->l_next) {
229        map->l_next->l_prev = map->l_prev;
230    }
231}
232
233static void notify_gdb_of_load(soinfo* info) {
234    if (info->flags & FLAG_EXE) {
235        // GDB already knows about the main executable
236        return;
237    }
238
239    ScopedPthreadMutexLocker locker(&gDebugMutex);
240
241    _r_debug.r_state = r_debug::RT_ADD;
242    rtld_db_dlactivity();
243
244    insert_soinfo_into_debug_map(info);
245
246    _r_debug.r_state = r_debug::RT_CONSISTENT;
247    rtld_db_dlactivity();
248}
249
250static void notify_gdb_of_unload(soinfo* info) {
251    if (info->flags & FLAG_EXE) {
252        // GDB already knows about the main executable
253        return;
254    }
255
256    ScopedPthreadMutexLocker locker(&gDebugMutex);
257
258    _r_debug.r_state = r_debug::RT_DELETE;
259    rtld_db_dlactivity();
260
261    remove_soinfo_from_debug_map(info);
262
263    _r_debug.r_state = r_debug::RT_CONSISTENT;
264    rtld_db_dlactivity();
265}
266
267void notify_gdb_of_libraries() {
268  _r_debug.r_state = r_debug::RT_ADD;
269  rtld_db_dlactivity();
270  _r_debug.r_state = r_debug::RT_CONSISTENT;
271  rtld_db_dlactivity();
272}
273
274static bool ensure_free_list_non_empty() {
275  if (gSoInfoFreeList != NULL) {
276    return true;
277  }
278
279  // Allocate a new pool.
280  soinfo_pool_t* pool = reinterpret_cast<soinfo_pool_t*>(mmap(NULL, sizeof(*pool),
281                                                              PROT_READ|PROT_WRITE,
282                                                              MAP_PRIVATE|MAP_ANONYMOUS, 0, 0));
283  if (pool == MAP_FAILED) {
284    return false;
285  }
286
287  // Add the pool to our list of pools.
288  pool->next = gSoInfoPools;
289  gSoInfoPools = pool;
290
291  // Chain the entries in the new pool onto the free list.
292  gSoInfoFreeList = &pool->info[0];
293  soinfo* next = NULL;
294  for (int i = SOINFO_PER_POOL - 1; i >= 0; --i) {
295    pool->info[i].next = next;
296    next = &pool->info[i];
297  }
298
299  return true;
300}
301
302static void set_soinfo_pool_protection(int protection) {
303  for (soinfo_pool_t* p = gSoInfoPools; p != NULL; p = p->next) {
304    if (mprotect(p, sizeof(*p), protection) == -1) {
305      abort(); // Can't happen.
306    }
307  }
308}
309
310static soinfo* soinfo_alloc(const char* name) {
311  if (strlen(name) >= SOINFO_NAME_LEN) {
312    DL_ERR("library name \"%s\" too long", name);
313    return NULL;
314  }
315
316  if (!ensure_free_list_non_empty()) {
317    DL_ERR("out of memory when loading \"%s\"", name);
318    return NULL;
319  }
320
321  // Take the head element off the free list.
322  soinfo* si = gSoInfoFreeList;
323  gSoInfoFreeList = gSoInfoFreeList->next;
324
325  // Initialize the new element.
326  memset(si, 0, sizeof(soinfo));
327  strlcpy(si->name, name, sizeof(si->name));
328  sonext->next = si;
329  sonext = si;
330
331  TRACE("name %s: allocated soinfo @ %p", name, si);
332  return si;
333}
334
335static void soinfo_free(soinfo* si) {
336    if (si == NULL) {
337        return;
338    }
339
340    soinfo *prev = NULL, *trav;
341
342    TRACE("name %s: freeing soinfo @ %p", si->name, si);
343
344    for (trav = solist; trav != NULL; trav = trav->next) {
345        if (trav == si)
346            break;
347        prev = trav;
348    }
349    if (trav == NULL) {
350        /* si was not in solist */
351        DL_ERR("name \"%s\" is not in solist!", si->name);
352        return;
353    }
354
355    /* prev will never be NULL, because the first entry in solist is
356       always the static libdl_info.
357    */
358    prev->next = si->next;
359    if (si == sonext) {
360        sonext = prev;
361    }
362    si->next = gSoInfoFreeList;
363    gSoInfoFreeList = si;
364}
365
366
367static void parse_path(const char* path, const char* delimiters,
368                       const char** array, char* buf, size_t buf_size, size_t max_count) {
369  if (path == NULL) {
370    return;
371  }
372
373  size_t len = strlcpy(buf, path, buf_size);
374
375  size_t i = 0;
376  char* buf_p = buf;
377  while (i < max_count && (array[i] = strsep(&buf_p, delimiters))) {
378    if (*array[i] != '\0') {
379      ++i;
380    }
381  }
382
383  // Forget the last path if we had to truncate; this occurs if the 2nd to
384  // last char isn't '\0' (i.e. wasn't originally a delimiter).
385  if (i > 0 && len >= buf_size && buf[buf_size - 2] != '\0') {
386    array[i - 1] = NULL;
387  } else {
388    array[i] = NULL;
389  }
390}
391
392static void parse_LD_LIBRARY_PATH(const char* path) {
393  parse_path(path, ":", gLdPaths,
394             gLdPathsBuffer, sizeof(gLdPathsBuffer), LDPATH_MAX);
395}
396
397static void parse_LD_PRELOAD(const char* path) {
398  // We have historically supported ':' as well as ' ' in LD_PRELOAD.
399  parse_path(path, " :", gLdPreloadNames,
400             gLdPreloadsBuffer, sizeof(gLdPreloadsBuffer), LDPRELOAD_MAX);
401}
402
403#if defined(__arm__)
404
405/* For a given PC, find the .so that it belongs to.
406 * Returns the base address of the .ARM.exidx section
407 * for that .so, and the number of 8-byte entries
408 * in that section (via *pcount).
409 *
410 * Intended to be called by libc's __gnu_Unwind_Find_exidx().
411 *
412 * This function is exposed via dlfcn.cpp and libdl.so.
413 */
414_Unwind_Ptr dl_unwind_find_exidx(_Unwind_Ptr pc, int* pcount) {
415    unsigned addr = (unsigned)pc;
416
417    for (soinfo* si = solist; si != 0; si = si->next) {
418        if ((addr >= si->base) && (addr < (si->base + si->size))) {
419            *pcount = si->ARM_exidx_count;
420            return (_Unwind_Ptr)si->ARM_exidx;
421        }
422    }
423    *pcount = 0;
424    return NULL;
425}
426
427#endif
428
429/* Here, we only have to provide a callback to iterate across all the
430 * loaded libraries. gcc_eh does the rest. */
431int dl_iterate_phdr(int (*cb)(dl_phdr_info* info, size_t size, void* data), void* data) {
432    int rv = 0;
433    for (soinfo* si = solist; si != NULL; si = si->next) {
434        dl_phdr_info dl_info;
435        dl_info.dlpi_addr = si->link_map_head.l_addr;
436        dl_info.dlpi_name = si->link_map_head.l_name;
437        dl_info.dlpi_phdr = si->phdr;
438        dl_info.dlpi_phnum = si->phnum;
439        rv = cb(&dl_info, sizeof(dl_phdr_info), data);
440        if (rv != 0) {
441            break;
442        }
443    }
444    return rv;
445}
446
447static ElfW(Sym)* soinfo_elf_lookup(soinfo* si, unsigned hash, const char* name) {
448  ElfW(Sym)* symtab = si->symtab;
449  const char* strtab = si->strtab;
450
451  TRACE_TYPE(LOOKUP, "SEARCH %s in %s@%p %x %zd",
452             name, si->name, reinterpret_cast<void*>(si->base), hash, hash % si->nbucket);
453
454  for (unsigned n = si->bucket[hash % si->nbucket]; n != 0; n = si->chain[n]) {
455    ElfW(Sym)* s = symtab + n;
456    if (strcmp(strtab + s->st_name, name)) continue;
457
458    /* only concern ourselves with global and weak symbol definitions */
459    switch (ELF_ST_BIND(s->st_info)) {
460      case STB_GLOBAL:
461      case STB_WEAK:
462        if (s->st_shndx == SHN_UNDEF) {
463        continue;
464      }
465
466      TRACE_TYPE(LOOKUP, "FOUND %s in %s (%p) %zd",
467                 name, si->name, reinterpret_cast<void*>(s->st_value),
468                 static_cast<size_t>(s->st_size));
469      return s;
470    }
471  }
472
473  return NULL;
474}
475
476static unsigned elfhash(const char* _name) {
477    const unsigned char* name = reinterpret_cast<const unsigned char*>(_name);
478    unsigned h = 0, g;
479
480    while (*name) {
481        h = (h << 4) + *name++;
482        g = h & 0xf0000000;
483        h ^= g;
484        h ^= g >> 24;
485    }
486    return h;
487}
488
489static ElfW(Sym)* soinfo_do_lookup(soinfo* si, const char* name, soinfo** lsi, soinfo* needed[]) {
490    unsigned elf_hash = elfhash(name);
491    ElfW(Sym)* s = NULL;
492
493    if (si != NULL && somain != NULL) {
494        /*
495         * Local scope is executable scope. Just start looking into it right away
496         * for the shortcut.
497         */
498
499        if (si == somain) {
500            s = soinfo_elf_lookup(si, elf_hash, name);
501            if (s != NULL) {
502                *lsi = si;
503                goto done;
504            }
505        } else {
506            /* Order of symbol lookup is controlled by DT_SYMBOLIC flag */
507
508            /*
509             * If this object was built with symbolic relocations disabled, the
510             * first place to look to resolve external references is the main
511             * executable.
512             */
513
514            if (!si->has_DT_SYMBOLIC) {
515                DEBUG("%s: looking up %s in executable %s",
516                      si->name, name, somain->name);
517                s = soinfo_elf_lookup(somain, elf_hash, name);
518                if (s != NULL) {
519                    *lsi = somain;
520                    goto done;
521                }
522            }
523
524            /* Look for symbols in the local scope (the object who is
525             * searching). This happens with C++ templates on x86 for some
526             * reason.
527             *
528             * Notes on weak symbols:
529             * The ELF specs are ambiguous about treatment of weak definitions in
530             * dynamic linking.  Some systems return the first definition found
531             * and some the first non-weak definition.   This is system dependent.
532             * Here we return the first definition found for simplicity.  */
533
534            s = soinfo_elf_lookup(si, elf_hash, name);
535            if (s != NULL) {
536                *lsi = si;
537                goto done;
538            }
539
540            /*
541             * If this object was built with -Bsymbolic and symbol is not found
542             * in the local scope, try to find the symbol in the main executable.
543             */
544
545            if (si->has_DT_SYMBOLIC) {
546                DEBUG("%s: looking up %s in executable %s after local scope",
547                      si->name, name, somain->name);
548                s = soinfo_elf_lookup(somain, elf_hash, name);
549                if (s != NULL) {
550                    *lsi = somain;
551                    goto done;
552                }
553            }
554        }
555    }
556
557    /* Next, look for it in the preloads list */
558    for (int i = 0; gLdPreloads[i] != NULL; i++) {
559        s = soinfo_elf_lookup(gLdPreloads[i], elf_hash, name);
560        if (s != NULL) {
561            *lsi = gLdPreloads[i];
562            goto done;
563        }
564    }
565
566    for (int i = 0; needed[i] != NULL; i++) {
567        DEBUG("%s: looking up %s in %s",
568              si->name, name, needed[i]->name);
569        s = soinfo_elf_lookup(needed[i], elf_hash, name);
570        if (s != NULL) {
571            *lsi = needed[i];
572            goto done;
573        }
574    }
575
576done:
577    if (s != NULL) {
578        TRACE_TYPE(LOOKUP, "si %s sym %s s->st_value = %p, "
579                   "found in %s, base = %p, load bias = %p",
580                   si->name, name, reinterpret_cast<void*>(s->st_value),
581                   (*lsi)->name, reinterpret_cast<void*>((*lsi)->base),
582                   reinterpret_cast<void*>((*lsi)->load_bias));
583        return s;
584    }
585
586    return NULL;
587}
588
589/* This is used by dlsym(3).  It performs symbol lookup only within the
590   specified soinfo object and not in any of its dependencies.
591
592   TODO: Only looking in the specified soinfo seems wrong. dlsym(3) says
593   that it should do a breadth first search through the dependency
594   tree. This agrees with the ELF spec (aka System V Application
595   Binary Interface) where in Chapter 5 it discuss resolving "Shared
596   Object Dependencies" in breadth first search order.
597 */
598ElfW(Sym)* dlsym_handle_lookup(soinfo* si, const char* name) {
599    return soinfo_elf_lookup(si, elfhash(name), name);
600}
601
602/* This is used by dlsym(3) to performs a global symbol lookup. If the
603   start value is null (for RTLD_DEFAULT), the search starts at the
604   beginning of the global solist. Otherwise the search starts at the
605   specified soinfo (for RTLD_NEXT).
606 */
607ElfW(Sym)* dlsym_linear_lookup(const char* name, soinfo** found, soinfo* start) {
608  unsigned elf_hash = elfhash(name);
609
610  if (start == NULL) {
611    start = solist;
612  }
613
614  ElfW(Sym)* s = NULL;
615  for (soinfo* si = start; (s == NULL) && (si != NULL); si = si->next) {
616    s = soinfo_elf_lookup(si, elf_hash, name);
617    if (s != NULL) {
618      *found = si;
619      break;
620    }
621  }
622
623  if (s != NULL) {
624    TRACE_TYPE(LOOKUP, "%s s->st_value = %p, found->base = %p",
625               name, reinterpret_cast<void*>(s->st_value), reinterpret_cast<void*>((*found)->base));
626  }
627
628  return s;
629}
630
631soinfo* find_containing_library(const void* p) {
632  ElfW(Addr) address = reinterpret_cast<ElfW(Addr)>(p);
633  for (soinfo* si = solist; si != NULL; si = si->next) {
634    if (address >= si->base && address - si->base < si->size) {
635      return si;
636    }
637  }
638  return NULL;
639}
640
641ElfW(Sym)* dladdr_find_symbol(soinfo* si, const void* addr) {
642  ElfW(Addr) soaddr = reinterpret_cast<ElfW(Addr)>(addr) - si->base;
643
644  // Search the library's symbol table for any defined symbol which
645  // contains this address.
646  for (size_t i = 0; i < si->nchain; ++i) {
647    ElfW(Sym)* sym = &si->symtab[i];
648    if (sym->st_shndx != SHN_UNDEF &&
649        soaddr >= sym->st_value &&
650        soaddr < sym->st_value + sym->st_size) {
651      return sym;
652    }
653  }
654
655  return NULL;
656}
657
658static int open_library_on_path(const char* name, const char* const paths[]) {
659  char buf[512];
660  for (size_t i = 0; paths[i] != NULL; ++i) {
661    int n = __libc_format_buffer(buf, sizeof(buf), "%s/%s", paths[i], name);
662    if (n < 0 || n >= static_cast<int>(sizeof(buf))) {
663      PRINT("Warning: ignoring very long library path: %s/%s", paths[i], name);
664      continue;
665    }
666    int fd = TEMP_FAILURE_RETRY(open(buf, O_RDONLY | O_CLOEXEC));
667    if (fd != -1) {
668      return fd;
669    }
670  }
671  return -1;
672}
673
674static int open_library(const char* name) {
675  TRACE("[ opening %s ]", name);
676
677  // If the name contains a slash, we should attempt to open it directly and not search the paths.
678  if (strchr(name, '/') != NULL) {
679    int fd = TEMP_FAILURE_RETRY(open(name, O_RDONLY | O_CLOEXEC));
680    if (fd != -1) {
681      return fd;
682    }
683    // ...but nvidia binary blobs (at least) rely on this behavior, so fall through for now.
684  }
685
686  // Otherwise we try LD_LIBRARY_PATH first, and fall back to the built-in well known paths.
687  int fd = open_library_on_path(name, gLdPaths);
688  if (fd == -1) {
689    fd = open_library_on_path(name, gDefaultLdPaths);
690  }
691  return fd;
692}
693
694static soinfo* load_library(const char* name) {
695    // Open the file.
696    int fd = open_library(name);
697    if (fd == -1) {
698        DL_ERR("library \"%s\" not found", name);
699        return NULL;
700    }
701
702    // Read the ELF header and load the segments.
703    ElfReader elf_reader(name, fd);
704    if (!elf_reader.Load()) {
705        return NULL;
706    }
707
708    const char* bname = strrchr(name, '/');
709    soinfo* si = soinfo_alloc(bname ? bname + 1 : name);
710    if (si == NULL) {
711        return NULL;
712    }
713    si->base = elf_reader.load_start();
714    si->size = elf_reader.load_size();
715    si->load_bias = elf_reader.load_bias();
716    si->flags = 0;
717    si->entry = 0;
718    si->dynamic = NULL;
719    si->phnum = elf_reader.phdr_count();
720    si->phdr = elf_reader.loaded_phdr();
721    return si;
722}
723
724static soinfo *find_loaded_library(const char* name) {
725    // TODO: don't use basename only for determining libraries
726    // http://code.google.com/p/android/issues/detail?id=6670
727
728    const char* bname = strrchr(name, '/');
729    bname = bname ? bname + 1 : name;
730
731    for (soinfo* si = solist; si != NULL; si = si->next) {
732        if (!strcmp(bname, si->name)) {
733            return si;
734        }
735    }
736    return NULL;
737}
738
739static soinfo* find_library_internal(const char* name) {
740  if (name == NULL) {
741    return somain;
742  }
743
744  soinfo* si = find_loaded_library(name);
745  if (si != NULL) {
746    if (si->flags & FLAG_LINKED) {
747      return si;
748    }
749    DL_ERR("OOPS: recursive link to \"%s\"", si->name);
750    return NULL;
751  }
752
753  TRACE("[ '%s' has not been loaded yet.  Locating...]", name);
754  si = load_library(name);
755  if (si == NULL) {
756    return NULL;
757  }
758
759  // At this point we know that whatever is loaded @ base is a valid ELF
760  // shared library whose segments are properly mapped in.
761  TRACE("[ find_library_internal base=%p size=%zu name='%s' ]",
762        reinterpret_cast<void*>(si->base), si->size, si->name);
763
764  if (!soinfo_link_image(si)) {
765    munmap(reinterpret_cast<void*>(si->base), si->size);
766    soinfo_free(si);
767    return NULL;
768  }
769
770  return si;
771}
772
773static soinfo* find_library(const char* name) {
774  soinfo* si = find_library_internal(name);
775  if (si != NULL) {
776    si->ref_count++;
777  }
778  return si;
779}
780
781static int soinfo_unload(soinfo* si) {
782  if (si->ref_count == 1) {
783    TRACE("unloading '%s'", si->name);
784    si->CallDestructors();
785
786    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
787      if (d->d_tag == DT_NEEDED) {
788        const char* library_name = si->strtab + d->d_un.d_val;
789        TRACE("%s needs to unload %s", si->name, library_name);
790        soinfo_unload(find_loaded_library(library_name));
791      }
792    }
793
794    munmap(reinterpret_cast<void*>(si->base), si->size);
795    notify_gdb_of_unload(si);
796    soinfo_free(si);
797    si->ref_count = 0;
798  } else {
799    si->ref_count--;
800    TRACE("not unloading '%s', decrementing ref_count to %zd", si->name, si->ref_count);
801  }
802  return 0;
803}
804
805void do_android_get_LD_LIBRARY_PATH(char* buffer, size_t buffer_size) {
806  snprintf(buffer, buffer_size, "%s:%s", gDefaultLdPaths[0], gDefaultLdPaths[1]);
807}
808
809void do_android_update_LD_LIBRARY_PATH(const char* ld_library_path) {
810  if (!get_AT_SECURE()) {
811    parse_LD_LIBRARY_PATH(ld_library_path);
812  }
813}
814
815soinfo* do_dlopen(const char* name, int flags) {
816  if ((flags & ~(RTLD_NOW|RTLD_LAZY|RTLD_LOCAL|RTLD_GLOBAL)) != 0) {
817    DL_ERR("invalid flags to dlopen: %x", flags);
818    return NULL;
819  }
820  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
821  soinfo* si = find_library(name);
822  if (si != NULL) {
823    si->CallConstructors();
824  }
825  set_soinfo_pool_protection(PROT_READ);
826  return si;
827}
828
829int do_dlclose(soinfo* si) {
830  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
831  int result = soinfo_unload(si);
832  set_soinfo_pool_protection(PROT_READ);
833  return result;
834}
835
836#if defined(USE_RELA)
837static int soinfo_relocate(soinfo* si, ElfW(Rela)* rela, unsigned count, soinfo* needed[]) {
838  ElfW(Sym)* symtab = si->symtab;
839  const char* strtab = si->strtab;
840  ElfW(Sym)* s;
841  ElfW(Rela)* start = rela;
842  soinfo* lsi;
843
844  for (size_t idx = 0; idx < count; ++idx, ++rela) {
845    unsigned type = ELFW(R_TYPE)(rela->r_info);
846    unsigned sym = ELFW(R_SYM)(rela->r_info);
847    ElfW(Addr) reloc = static_cast<ElfW(Addr)>(rela->r_offset + si->load_bias);
848    ElfW(Addr) sym_addr = 0;
849    const char* sym_name = NULL;
850
851    DEBUG("Processing '%s' relocation at index %zd", si->name, idx);
852    if (type == 0) { // R_*_NONE
853      continue;
854    }
855    if (sym != 0) {
856      sym_name = reinterpret_cast<const char*>(strtab + symtab[sym].st_name);
857      s = soinfo_do_lookup(si, sym_name, &lsi, needed);
858      if (s == NULL) {
859        // We only allow an undefined symbol if this is a weak reference...
860        s = &symtab[sym];
861        if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
862          DL_ERR("cannot locate symbol \"%s\" referenced by \"%s\"...", sym_name, si->name);
863          return -1;
864        }
865
866        /* IHI0044C AAELF 4.5.1.1:
867
868           Libraries are not searched to resolve weak references.
869           It is not an error for a weak reference to remain unsatisfied.
870
871           During linking, the value of an undefined weak reference is:
872           - Zero if the relocation type is absolute
873           - The address of the place if the relocation is pc-relative
874           - The address of nominal base address if the relocation
875             type is base-relative.
876         */
877
878        switch (type) {
879#if defined(__aarch64__)
880        case R_AARCH64_JUMP_SLOT:
881        case R_AARCH64_GLOB_DAT:
882        case R_AARCH64_ABS64:
883        case R_AARCH64_ABS32:
884        case R_AARCH64_ABS16:
885        case R_AARCH64_RELATIVE:
886          /*
887           * The sym_addr was initialized to be zero above, or the relocation
888           * code below does not care about value of sym_addr.
889           * No need to do anything.
890           */
891          break;
892#elif defined(__x86_64__)
893        case R_X86_64_JUMP_SLOT:
894        case R_X86_64_GLOB_DAT:
895        case R_X86_64_32:
896        case R_X86_64_RELATIVE:
897          // No need to do anything.
898          break;
899        case R_X86_64_PC32:
900          sym_addr = reloc;
901          break;
902#endif
903        default:
904          DL_ERR("unknown weak reloc type %d @ %p (%zu)", type, rela, idx);
905          return -1;
906        }
907      } else {
908        // We got a definition.
909        sym_addr = static_cast<ElfW(Addr)>(s->st_value + lsi->load_bias);
910      }
911      count_relocation(kRelocSymbol);
912    } else {
913      s = NULL;
914    }
915
916    switch (type) {
917#if defined(__aarch64__)
918    case R_AARCH64_JUMP_SLOT:
919        count_relocation(kRelocAbsolute);
920        MARK(rela->r_offset);
921        TRACE_TYPE(RELO, "RELO JMP_SLOT %16llx <- %16llx %s\n",
922                   reloc, (sym_addr + rela->r_addend), sym_name);
923        *reinterpret_cast<ElfW(Addr)*>(reloc) = (sym_addr + rela->r_addend);
924        break;
925    case R_AARCH64_GLOB_DAT:
926        count_relocation(kRelocAbsolute);
927        MARK(rela->r_offset);
928        TRACE_TYPE(RELO, "RELO GLOB_DAT %16llx <- %16llx %s\n",
929                   reloc, (sym_addr + rela->r_addend), sym_name);
930        *reinterpret_cast<ElfW(Addr)*>(reloc) = (sym_addr + rela->r_addend);
931        break;
932    case R_AARCH64_ABS64:
933        count_relocation(kRelocAbsolute);
934        MARK(rela->r_offset);
935        TRACE_TYPE(RELO, "RELO ABS64 %16llx <- %16llx %s\n",
936                   reloc, (sym_addr + rela->r_addend), sym_name);
937        *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
938        break;
939    case R_AARCH64_ABS32:
940        count_relocation(kRelocAbsolute);
941        MARK(rela->r_offset);
942        TRACE_TYPE(RELO, "RELO ABS32 %16llx <- %16llx %s\n",
943                   reloc, (sym_addr + rela->r_addend), sym_name);
944        if ((static_cast<ElfW(Addr)>(INT32_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend))) &&
945            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)) <= static_cast<ElfW(Addr)>(UINT32_MAX))) {
946            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
947        } else {
948            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
949                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)),
950                   static_cast<ElfW(Addr)>(INT32_MIN),
951                   static_cast<ElfW(Addr)>(UINT32_MAX));
952            return -1;
953        }
954        break;
955    case R_AARCH64_ABS16:
956        count_relocation(kRelocAbsolute);
957        MARK(rela->r_offset);
958        TRACE_TYPE(RELO, "RELO ABS16 %16llx <- %16llx %s\n",
959                   reloc, (sym_addr + rela->r_addend), sym_name);
960        if ((static_cast<ElfW(Addr)>(INT16_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend))) &&
961            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)) <= static_cast<ElfW(Addr)>(UINT16_MAX))) {
962            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend);
963        } else {
964            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
965                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + (sym_addr + rela->r_addend)),
966                   static_cast<ElfW(Addr)>(INT16_MIN),
967                   static_cast<ElfW(Addr)>(UINT16_MAX));
968            return -1;
969        }
970        break;
971    case R_AARCH64_PREL64:
972        count_relocation(kRelocRelative);
973        MARK(rela->r_offset);
974        TRACE_TYPE(RELO, "RELO REL64 %16llx <- %16llx - %16llx %s\n",
975                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
976        *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr + rela->r_addend) - rela->r_offset;
977        break;
978    case R_AARCH64_PREL32:
979        count_relocation(kRelocRelative);
980        MARK(rela->r_offset);
981        TRACE_TYPE(RELO, "RELO REL32 %16llx <- %16llx - %16llx %s\n",
982                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
983        if ((static_cast<ElfW(Addr)>(INT32_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset))) &&
984            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)) <= static_cast<ElfW(Addr)>(UINT32_MAX))) {
985            *reinterpret_cast<ElfW(Addr)*>(reloc) += ((sym_addr + rela->r_addend) - rela->r_offset);
986        } else {
987            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
988                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)),
989                   static_cast<ElfW(Addr)>(INT32_MIN),
990                   static_cast<ElfW(Addr)>(UINT32_MAX));
991            return -1;
992        }
993        break;
994    case R_AARCH64_PREL16:
995        count_relocation(kRelocRelative);
996        MARK(rela->r_offset);
997        TRACE_TYPE(RELO, "RELO REL16 %16llx <- %16llx - %16llx %s\n",
998                   reloc, (sym_addr + rela->r_addend), rela->r_offset, sym_name);
999        if ((static_cast<ElfW(Addr)>(INT16_MIN) <= (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset))) &&
1000            ((*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)) <= static_cast<ElfW(Addr)>(UINT16_MAX))) {
1001            *reinterpret_cast<ElfW(Addr)*>(reloc) += ((sym_addr + rela->r_addend) - rela->r_offset);
1002        } else {
1003            DL_ERR("0x%016llx out of range 0x%016llx to 0x%016llx",
1004                   (*reinterpret_cast<ElfW(Addr)*>(reloc) + ((sym_addr + rela->r_addend) - rela->r_offset)),
1005                   static_cast<ElfW(Addr)>(INT16_MIN),
1006                   static_cast<ElfW(Addr)>(UINT16_MAX));
1007            return -1;
1008        }
1009        break;
1010
1011    case R_AARCH64_RELATIVE:
1012        count_relocation(kRelocRelative);
1013        MARK(rela->r_offset);
1014        if (sym) {
1015            DL_ERR("odd RELATIVE form...");
1016            return -1;
1017        }
1018        TRACE_TYPE(RELO, "RELO RELATIVE %16llx <- %16llx\n",
1019                   reloc, (si->base + rela->r_addend));
1020        *reinterpret_cast<ElfW(Addr)*>(reloc) = (si->base + rela->r_addend);
1021        break;
1022
1023    case R_AARCH64_COPY:
1024        if ((si->flags & FLAG_EXE) == 0) {
1025            /*
1026              * http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044d/IHI0044D_aaelf.pdf
1027              *
1028              * Section 4.7.1.10 "Dynamic relocations"
1029              * R_AARCH64_COPY may only appear in executable objects where e_type is
1030              * set to ET_EXEC.
1031              *
1032              * FLAG_EXE is set for both ET_DYN and ET_EXEC executables.
1033              * We should explicitly disallow ET_DYN executables from having
1034              * R_AARCH64_COPY relocations.
1035              */
1036            DL_ERR("%s R_AARCH64_COPY relocations only supported for ET_EXEC", si->name);
1037            return -1;
1038        }
1039        count_relocation(kRelocCopy);
1040        MARK(rela->r_offset);
1041        TRACE_TYPE(RELO, "RELO COPY %16llx <- %lld @ %16llx %s\n",
1042                   reloc,
1043                   s->st_size,
1044                   (sym_addr + rela->r_addend),
1045                   sym_name);
1046        if (reloc == (sym_addr + rela->r_addend)) {
1047            ElfW(Sym)* src = soinfo_do_lookup(NULL, sym_name, &lsi, needed);
1048
1049            if (src == NULL) {
1050                DL_ERR("%s R_AARCH64_COPY relocation source cannot be resolved", si->name);
1051                return -1;
1052            }
1053            if (lsi->has_DT_SYMBOLIC) {
1054                DL_ERR("%s invalid R_AARCH64_COPY relocation against DT_SYMBOLIC shared "
1055                       "library %s (built with -Bsymbolic?)", si->name, lsi->name);
1056                return -1;
1057            }
1058            if (s->st_size < src->st_size) {
1059                DL_ERR("%s R_AARCH64_COPY relocation size mismatch (%lld < %lld)",
1060                       si->name, s->st_size, src->st_size);
1061                return -1;
1062            }
1063            memcpy(reinterpret_cast<void*>(reloc),
1064                   reinterpret_cast<void*>(src->st_value + lsi->load_bias), src->st_size);
1065        } else {
1066            DL_ERR("%s R_AARCH64_COPY relocation target cannot be resolved", si->name);
1067            return -1;
1068        }
1069        break;
1070    case R_AARCH64_TLS_TPREL64:
1071        TRACE_TYPE(RELO, "RELO TLS_TPREL64 *** %16llx <- %16llx - %16llx\n",
1072                   reloc, (sym_addr + rela->r_addend), rela->r_offset);
1073        break;
1074    case R_AARCH64_TLS_DTPREL32:
1075        TRACE_TYPE(RELO, "RELO TLS_DTPREL32 *** %16llx <- %16llx - %16llx\n",
1076                   reloc, (sym_addr + rela->r_addend), rela->r_offset);
1077        break;
1078#elif defined(__x86_64__)
1079    case R_X86_64_JUMP_SLOT:
1080      count_relocation(kRelocAbsolute);
1081      MARK(rela->r_offset);
1082      TRACE_TYPE(RELO, "RELO JMP_SLOT %08zx <- %08zx %s", static_cast<size_t>(reloc),
1083                 static_cast<size_t>(sym_addr + rela->r_addend), sym_name);
1084      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1085      break;
1086    case R_X86_64_GLOB_DAT:
1087      count_relocation(kRelocAbsolute);
1088      MARK(rela->r_offset);
1089      TRACE_TYPE(RELO, "RELO GLOB_DAT %08zx <- %08zx %s", static_cast<size_t>(reloc),
1090                 static_cast<size_t>(sym_addr + rela->r_addend), sym_name);
1091      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1092      break;
1093    case R_X86_64_RELATIVE:
1094      count_relocation(kRelocRelative);
1095      MARK(rela->r_offset);
1096      if (sym) {
1097        DL_ERR("odd RELATIVE form...");
1098        return -1;
1099      }
1100      TRACE_TYPE(RELO, "RELO RELATIVE %08zx <- +%08zx", static_cast<size_t>(reloc),
1101                 static_cast<size_t>(si->base));
1102      *reinterpret_cast<ElfW(Addr)*>(reloc) = si->base + rela->r_addend;
1103      break;
1104    case R_X86_64_32:
1105      count_relocation(kRelocRelative);
1106      MARK(rela->r_offset);
1107      TRACE_TYPE(RELO, "RELO R_X86_64_32 %08zx <- +%08zx %s", static_cast<size_t>(reloc),
1108                 static_cast<size_t>(sym_addr), sym_name);
1109      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1110      break;
1111    case R_X86_64_64:
1112      count_relocation(kRelocRelative);
1113      MARK(rela->r_offset);
1114      TRACE_TYPE(RELO, "RELO R_X86_64_64 %08zx <- +%08zx %s", static_cast<size_t>(reloc),
1115                 static_cast<size_t>(sym_addr), sym_name);
1116      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend;
1117      break;
1118    case R_X86_64_PC32:
1119      count_relocation(kRelocRelative);
1120      MARK(rela->r_offset);
1121      TRACE_TYPE(RELO, "RELO R_X86_64_PC32 %08zx <- +%08zx (%08zx - %08zx) %s",
1122                 static_cast<size_t>(reloc), static_cast<size_t>(sym_addr - reloc),
1123                 static_cast<size_t>(sym_addr), static_cast<size_t>(reloc), sym_name);
1124      *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr + rela->r_addend - reloc;
1125      break;
1126#endif
1127
1128    default:
1129      DL_ERR("unknown reloc type %d @ %p (%zu)", type, rela, idx);
1130      return -1;
1131    }
1132  }
1133  return 0;
1134}
1135
1136#else // REL, not RELA.
1137
1138static int soinfo_relocate(soinfo* si, ElfW(Rel)* rel, unsigned count, soinfo* needed[]) {
1139    ElfW(Sym)* symtab = si->symtab;
1140    const char* strtab = si->strtab;
1141    ElfW(Sym)* s;
1142    ElfW(Rel)* start = rel;
1143    soinfo* lsi;
1144
1145    for (size_t idx = 0; idx < count; ++idx, ++rel) {
1146        unsigned type = ELFW(R_TYPE)(rel->r_info);
1147        // TODO: don't use unsigned for 'sym'. Use uint32_t or ElfW(Addr) instead.
1148        unsigned sym = ELFW(R_SYM)(rel->r_info);
1149        ElfW(Addr) reloc = static_cast<ElfW(Addr)>(rel->r_offset + si->load_bias);
1150        ElfW(Addr) sym_addr = 0;
1151        const char* sym_name = NULL;
1152
1153        DEBUG("Processing '%s' relocation at index %zd", si->name, idx);
1154        if (type == 0) { // R_*_NONE
1155            continue;
1156        }
1157        if (sym != 0) {
1158            sym_name = reinterpret_cast<const char*>(strtab + symtab[sym].st_name);
1159            s = soinfo_do_lookup(si, sym_name, &lsi, needed);
1160            if (s == NULL) {
1161                // We only allow an undefined symbol if this is a weak reference...
1162                s = &symtab[sym];
1163                if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
1164                    DL_ERR("cannot locate symbol \"%s\" referenced by \"%s\"...", sym_name, si->name);
1165                    return -1;
1166                }
1167
1168                /* IHI0044C AAELF 4.5.1.1:
1169
1170                   Libraries are not searched to resolve weak references.
1171                   It is not an error for a weak reference to remain
1172                   unsatisfied.
1173
1174                   During linking, the value of an undefined weak reference is:
1175                   - Zero if the relocation type is absolute
1176                   - The address of the place if the relocation is pc-relative
1177                   - The address of nominal base address if the relocation
1178                     type is base-relative.
1179                  */
1180
1181                switch (type) {
1182#if defined(__arm__)
1183                case R_ARM_JUMP_SLOT:
1184                case R_ARM_GLOB_DAT:
1185                case R_ARM_ABS32:
1186                case R_ARM_RELATIVE:    /* Don't care. */
1187                    // sym_addr was initialized to be zero above or relocation
1188                    // code below does not care about value of sym_addr.
1189                    // No need to do anything.
1190                    break;
1191#elif defined(__i386__)
1192                case R_386_JMP_SLOT:
1193                case R_386_GLOB_DAT:
1194                case R_386_32:
1195                case R_386_RELATIVE:    /* Don't care. */
1196                    // sym_addr was initialized to be zero above or relocation
1197                    // code below does not care about value of sym_addr.
1198                    // No need to do anything.
1199                    break;
1200                case R_386_PC32:
1201                    sym_addr = reloc;
1202                    break;
1203#endif
1204
1205#if defined(__arm__)
1206                case R_ARM_COPY:
1207                    // Fall through. Can't really copy if weak symbol is not found at run-time.
1208#endif
1209                default:
1210                    DL_ERR("unknown weak reloc type %d @ %p (%zu)", type, rel, idx);
1211                    return -1;
1212                }
1213            } else {
1214                // We got a definition.
1215                sym_addr = static_cast<ElfW(Addr)>(s->st_value + lsi->load_bias);
1216            }
1217            count_relocation(kRelocSymbol);
1218        } else {
1219            s = NULL;
1220        }
1221
1222        switch (type) {
1223#if defined(__arm__)
1224        case R_ARM_JUMP_SLOT:
1225            count_relocation(kRelocAbsolute);
1226            MARK(rel->r_offset);
1227            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
1228            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1229            break;
1230        case R_ARM_GLOB_DAT:
1231            count_relocation(kRelocAbsolute);
1232            MARK(rel->r_offset);
1233            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
1234            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1235            break;
1236        case R_ARM_ABS32:
1237            count_relocation(kRelocAbsolute);
1238            MARK(rel->r_offset);
1239            TRACE_TYPE(RELO, "RELO ABS %08x <- %08x %s", reloc, sym_addr, sym_name);
1240            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1241            break;
1242        case R_ARM_REL32:
1243            count_relocation(kRelocRelative);
1244            MARK(rel->r_offset);
1245            TRACE_TYPE(RELO, "RELO REL32 %08x <- %08x - %08x %s",
1246                       reloc, sym_addr, rel->r_offset, sym_name);
1247            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr - rel->r_offset;
1248            break;
1249        case R_ARM_COPY:
1250            if ((si->flags & FLAG_EXE) == 0) {
1251                /*
1252                 * http://infocenter.arm.com/help/topic/com.arm.doc.ihi0044d/IHI0044D_aaelf.pdf
1253                 *
1254                 * Section 4.7.1.10 "Dynamic relocations"
1255                 * R_ARM_COPY may only appear in executable objects where e_type is
1256                 * set to ET_EXEC.
1257                 *
1258                 * TODO: FLAG_EXE is set for both ET_DYN and ET_EXEC executables.
1259                 * We should explicitly disallow ET_DYN executables from having
1260                 * R_ARM_COPY relocations.
1261                 */
1262                DL_ERR("%s R_ARM_COPY relocations only supported for ET_EXEC", si->name);
1263                return -1;
1264            }
1265            count_relocation(kRelocCopy);
1266            MARK(rel->r_offset);
1267            TRACE_TYPE(RELO, "RELO %08x <- %d @ %08x %s", reloc, s->st_size, sym_addr, sym_name);
1268            if (reloc == sym_addr) {
1269                ElfW(Sym)* src = soinfo_do_lookup(NULL, sym_name, &lsi, needed);
1270
1271                if (src == NULL) {
1272                    DL_ERR("%s R_ARM_COPY relocation source cannot be resolved", si->name);
1273                    return -1;
1274                }
1275                if (lsi->has_DT_SYMBOLIC) {
1276                    DL_ERR("%s invalid R_ARM_COPY relocation against DT_SYMBOLIC shared "
1277                           "library %s (built with -Bsymbolic?)", si->name, lsi->name);
1278                    return -1;
1279                }
1280                if (s->st_size < src->st_size) {
1281                    DL_ERR("%s R_ARM_COPY relocation size mismatch (%d < %d)",
1282                           si->name, s->st_size, src->st_size);
1283                    return -1;
1284                }
1285                memcpy(reinterpret_cast<void*>(reloc),
1286                       reinterpret_cast<void*>(src->st_value + lsi->load_bias), src->st_size);
1287            } else {
1288                DL_ERR("%s R_ARM_COPY relocation target cannot be resolved", si->name);
1289                return -1;
1290            }
1291            break;
1292#elif defined(__i386__)
1293        case R_386_JMP_SLOT:
1294            count_relocation(kRelocAbsolute);
1295            MARK(rel->r_offset);
1296            TRACE_TYPE(RELO, "RELO JMP_SLOT %08x <- %08x %s", reloc, sym_addr, sym_name);
1297            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1298            break;
1299        case R_386_GLOB_DAT:
1300            count_relocation(kRelocAbsolute);
1301            MARK(rel->r_offset);
1302            TRACE_TYPE(RELO, "RELO GLOB_DAT %08x <- %08x %s", reloc, sym_addr, sym_name);
1303            *reinterpret_cast<ElfW(Addr)*>(reloc) = sym_addr;
1304            break;
1305        case R_386_32:
1306            count_relocation(kRelocRelative);
1307            MARK(rel->r_offset);
1308            TRACE_TYPE(RELO, "RELO R_386_32 %08x <- +%08x %s", reloc, sym_addr, sym_name);
1309            *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1310            break;
1311        case R_386_PC32:
1312            count_relocation(kRelocRelative);
1313            MARK(rel->r_offset);
1314            TRACE_TYPE(RELO, "RELO R_386_PC32 %08x <- +%08x (%08x - %08x) %s",
1315                       reloc, (sym_addr - reloc), sym_addr, reloc, sym_name);
1316            *reinterpret_cast<ElfW(Addr)*>(reloc) += (sym_addr - reloc);
1317            break;
1318#elif defined(__mips__)
1319        case R_MIPS_REL32:
1320#if defined(__LP64__)
1321            // MIPS Elf64_Rel entries contain compound relocations
1322            // We only handle the R_MIPS_NONE|R_MIPS_64|R_MIPS_REL32 case
1323            if (ELF64_R_TYPE2(rel->r_info) != R_MIPS_64 ||
1324                ELF64_R_TYPE3(rel->r_info) != R_MIPS_NONE) {
1325                DL_ERR("Unexpected compound relocation type:%d type2:%d type3:%d @ %p (%zu)",
1326                       type, (unsigned)ELF64_R_TYPE2(rel->r_info),
1327                       (unsigned)ELF64_R_TYPE3(rel->r_info), rel, idx);
1328                return -1;
1329            }
1330#endif
1331            count_relocation(kRelocAbsolute);
1332            MARK(rel->r_offset);
1333            TRACE_TYPE(RELO, "RELO REL32 %08zx <- %08zx %s", static_cast<size_t>(reloc),
1334                       static_cast<size_t>(sym_addr), sym_name ? sym_name : "*SECTIONHDR*");
1335            if (s) {
1336                *reinterpret_cast<ElfW(Addr)*>(reloc) += sym_addr;
1337            } else {
1338                *reinterpret_cast<ElfW(Addr)*>(reloc) += si->base;
1339            }
1340            break;
1341#endif
1342
1343#if defined(__arm__)
1344        case R_ARM_RELATIVE:
1345#elif defined(__i386__)
1346        case R_386_RELATIVE:
1347#endif
1348            count_relocation(kRelocRelative);
1349            MARK(rel->r_offset);
1350            if (sym) {
1351                DL_ERR("odd RELATIVE form...");
1352                return -1;
1353            }
1354            TRACE_TYPE(RELO, "RELO RELATIVE %p <- +%p",
1355                       reinterpret_cast<void*>(reloc), reinterpret_cast<void*>(si->base));
1356            *reinterpret_cast<ElfW(Addr)*>(reloc) += si->base;
1357            break;
1358
1359        default:
1360            DL_ERR("unknown reloc type %d @ %p (%zu)", type, rel, idx);
1361            return -1;
1362        }
1363    }
1364    return 0;
1365}
1366#endif
1367
1368#if defined(__mips__)
1369static bool mips_relocate_got(soinfo* si, soinfo* needed[]) {
1370    ElfW(Addr)** got = si->plt_got;
1371    if (got == NULL) {
1372        return true;
1373    }
1374    unsigned local_gotno = si->mips_local_gotno;
1375    unsigned gotsym = si->mips_gotsym;
1376    unsigned symtabno = si->mips_symtabno;
1377    ElfW(Sym)* symtab = si->symtab;
1378
1379    // got[0] is the address of the lazy resolver function.
1380    // got[1] may be used for a GNU extension.
1381    // Set it to a recognizable address in case someone calls it (should be _rtld_bind_start).
1382    // FIXME: maybe this should be in a separate routine?
1383    if ((si->flags & FLAG_LINKER) == 0) {
1384        size_t g = 0;
1385        got[g++] = reinterpret_cast<ElfW(Addr)*>(0xdeadbeef);
1386        if (reinterpret_cast<intptr_t>(got[g]) < 0) {
1387            got[g++] = reinterpret_cast<ElfW(Addr)*>(0xdeadfeed);
1388        }
1389        // Relocate the local GOT entries.
1390        for (; g < local_gotno; g++) {
1391            got[g] = reinterpret_cast<ElfW(Addr)*>(reinterpret_cast<uintptr_t>(got[g]) + si->load_bias);
1392        }
1393    }
1394
1395    // Now for the global GOT entries...
1396    ElfW(Sym)* sym = symtab + gotsym;
1397    got = si->plt_got + local_gotno;
1398    for (size_t g = gotsym; g < symtabno; g++, sym++, got++) {
1399        // This is an undefined reference... try to locate it.
1400        const char* sym_name = si->strtab + sym->st_name;
1401        soinfo* lsi;
1402        ElfW(Sym)* s = soinfo_do_lookup(si, sym_name, &lsi, needed);
1403        if (s == NULL) {
1404            // We only allow an undefined symbol if this is a weak reference.
1405            s = &symtab[g];
1406            if (ELF_ST_BIND(s->st_info) != STB_WEAK) {
1407                DL_ERR("cannot locate \"%s\"...", sym_name);
1408                return false;
1409            }
1410            *got = 0;
1411        } else {
1412            // FIXME: is this sufficient?
1413            // For reference see NetBSD link loader
1414            // 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
1415            *got = reinterpret_cast<ElfW(Addr)*>(lsi->load_bias + s->st_value);
1416        }
1417    }
1418    return true;
1419}
1420#endif
1421
1422void soinfo::CallArray(const char* array_name UNUSED, linker_function_t* functions, size_t count, bool reverse) {
1423  if (functions == NULL) {
1424    return;
1425  }
1426
1427  TRACE("[ Calling %s (size %zd) @ %p for '%s' ]", array_name, count, functions, name);
1428
1429  int begin = reverse ? (count - 1) : 0;
1430  int end = reverse ? -1 : count;
1431  int step = reverse ? -1 : 1;
1432
1433  for (int i = begin; i != end; i += step) {
1434    TRACE("[ %s[%d] == %p ]", array_name, i, functions[i]);
1435    CallFunction("function", functions[i]);
1436  }
1437
1438  TRACE("[ Done calling %s for '%s' ]", array_name, name);
1439}
1440
1441void soinfo::CallFunction(const char* function_name UNUSED, linker_function_t function) {
1442  if (function == NULL || reinterpret_cast<uintptr_t>(function) == static_cast<uintptr_t>(-1)) {
1443    return;
1444  }
1445
1446  TRACE("[ Calling %s @ %p for '%s' ]", function_name, function, name);
1447  function();
1448  TRACE("[ Done calling %s @ %p for '%s' ]", function_name, function, name);
1449
1450  // The function may have called dlopen(3) or dlclose(3), so we need to ensure our data structures
1451  // are still writable. This happens with our debug malloc (see http://b/7941716).
1452  set_soinfo_pool_protection(PROT_READ | PROT_WRITE);
1453}
1454
1455void soinfo::CallPreInitConstructors() {
1456  // DT_PREINIT_ARRAY functions are called before any other constructors for executables,
1457  // but ignored in a shared library.
1458  CallArray("DT_PREINIT_ARRAY", preinit_array, preinit_array_count, false);
1459}
1460
1461void soinfo::CallConstructors() {
1462  if (constructors_called) {
1463    return;
1464  }
1465
1466  // We set constructors_called before actually calling the constructors, otherwise it doesn't
1467  // protect against recursive constructor calls. One simple example of constructor recursion
1468  // is the libc debug malloc, which is implemented in libc_malloc_debug_leak.so:
1469  // 1. The program depends on libc, so libc's constructor is called here.
1470  // 2. The libc constructor calls dlopen() to load libc_malloc_debug_leak.so.
1471  // 3. dlopen() calls the constructors on the newly created
1472  //    soinfo for libc_malloc_debug_leak.so.
1473  // 4. The debug .so depends on libc, so CallConstructors is
1474  //    called again with the libc soinfo. If it doesn't trigger the early-
1475  //    out above, the libc constructor will be called again (recursively!).
1476  constructors_called = true;
1477
1478  if ((flags & FLAG_EXE) == 0 && preinit_array != NULL) {
1479    // The GNU dynamic linker silently ignores these, but we warn the developer.
1480    PRINT("\"%s\": ignoring %zd-entry DT_PREINIT_ARRAY in shared library!",
1481          name, preinit_array_count);
1482  }
1483
1484  if (dynamic != NULL) {
1485    for (ElfW(Dyn)* d = dynamic; d->d_tag != DT_NULL; ++d) {
1486      if (d->d_tag == DT_NEEDED) {
1487        const char* library_name = strtab + d->d_un.d_val;
1488        TRACE("\"%s\": calling constructors in DT_NEEDED \"%s\"", name, library_name);
1489        find_loaded_library(library_name)->CallConstructors();
1490      }
1491    }
1492  }
1493
1494  TRACE("\"%s\": calling constructors", name);
1495
1496  // DT_INIT should be called before DT_INIT_ARRAY if both are present.
1497  CallFunction("DT_INIT", init_func);
1498  CallArray("DT_INIT_ARRAY", init_array, init_array_count, false);
1499}
1500
1501void soinfo::CallDestructors() {
1502  TRACE("\"%s\": calling destructors", name);
1503
1504  // DT_FINI_ARRAY must be parsed in reverse order.
1505  CallArray("DT_FINI_ARRAY", fini_array, fini_array_count, true);
1506
1507  // DT_FINI should be called after DT_FINI_ARRAY if both are present.
1508  CallFunction("DT_FINI", fini_func);
1509}
1510
1511/* Force any of the closed stdin, stdout and stderr to be associated with
1512   /dev/null. */
1513static int nullify_closed_stdio() {
1514    int dev_null, i, status;
1515    int return_value = 0;
1516
1517    dev_null = TEMP_FAILURE_RETRY(open("/dev/null", O_RDWR));
1518    if (dev_null < 0) {
1519        DL_ERR("cannot open /dev/null: %s", strerror(errno));
1520        return -1;
1521    }
1522    TRACE("[ Opened /dev/null file-descriptor=%d]", dev_null);
1523
1524    /* If any of the stdio file descriptors is valid and not associated
1525       with /dev/null, dup /dev/null to it.  */
1526    for (i = 0; i < 3; i++) {
1527        /* If it is /dev/null already, we are done. */
1528        if (i == dev_null) {
1529            continue;
1530        }
1531
1532        TRACE("[ Nullifying stdio file descriptor %d]", i);
1533        status = TEMP_FAILURE_RETRY(fcntl(i, F_GETFL));
1534
1535        /* If file is opened, we are good. */
1536        if (status != -1) {
1537            continue;
1538        }
1539
1540        /* The only error we allow is that the file descriptor does not
1541           exist, in which case we dup /dev/null to it. */
1542        if (errno != EBADF) {
1543            DL_ERR("fcntl failed: %s", strerror(errno));
1544            return_value = -1;
1545            continue;
1546        }
1547
1548        /* Try dupping /dev/null to this stdio file descriptor and
1549           repeat if there is a signal.  Note that any errors in closing
1550           the stdio descriptor are lost.  */
1551        status = TEMP_FAILURE_RETRY(dup2(dev_null, i));
1552        if (status < 0) {
1553            DL_ERR("dup2 failed: %s", strerror(errno));
1554            return_value = -1;
1555            continue;
1556        }
1557    }
1558
1559    /* If /dev/null is not one of the stdio file descriptors, close it. */
1560    if (dev_null > 2) {
1561        TRACE("[ Closing /dev/null file-descriptor=%d]", dev_null);
1562        status = TEMP_FAILURE_RETRY(close(dev_null));
1563        if (status == -1) {
1564            DL_ERR("close failed: %s", strerror(errno));
1565            return_value = -1;
1566        }
1567    }
1568
1569    return return_value;
1570}
1571
1572static bool soinfo_link_image(soinfo* si) {
1573    /* "base" might wrap around UINT32_MAX. */
1574    ElfW(Addr) base = si->load_bias;
1575    const ElfW(Phdr)* phdr = si->phdr;
1576    int phnum = si->phnum;
1577    bool relocating_linker = (si->flags & FLAG_LINKER) != 0;
1578
1579    /* We can't debug anything until the linker is relocated */
1580    if (!relocating_linker) {
1581        INFO("[ linking %s ]", si->name);
1582        DEBUG("si->base = %p si->flags = 0x%08x", reinterpret_cast<void*>(si->base), si->flags);
1583    }
1584
1585    /* Extract dynamic section */
1586    size_t dynamic_count;
1587    ElfW(Word) dynamic_flags;
1588    phdr_table_get_dynamic_section(phdr, phnum, base, &si->dynamic,
1589                                   &dynamic_count, &dynamic_flags);
1590    if (si->dynamic == NULL) {
1591        if (!relocating_linker) {
1592            DL_ERR("missing PT_DYNAMIC in \"%s\"", si->name);
1593        }
1594        return false;
1595    } else {
1596        if (!relocating_linker) {
1597            DEBUG("dynamic = %p", si->dynamic);
1598        }
1599    }
1600
1601#if defined(__arm__)
1602    (void) phdr_table_get_arm_exidx(phdr, phnum, base,
1603                                    &si->ARM_exidx, &si->ARM_exidx_count);
1604#endif
1605
1606    // Extract useful information from dynamic section.
1607    uint32_t needed_count = 0;
1608    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1609        DEBUG("d = %p, d[0](tag) = %p d[1](val) = %p",
1610              d, reinterpret_cast<void*>(d->d_tag), reinterpret_cast<void*>(d->d_un.d_val));
1611        switch (d->d_tag) {
1612        case DT_HASH:
1613            si->nbucket = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr)[0];
1614            si->nchain = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr)[1];
1615            si->bucket = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr + 8);
1616            si->chain = reinterpret_cast<uint32_t*>(base + d->d_un.d_ptr + 8 + si->nbucket * 4);
1617            break;
1618        case DT_STRTAB:
1619            si->strtab = reinterpret_cast<const char*>(base + d->d_un.d_ptr);
1620            break;
1621        case DT_SYMTAB:
1622            si->symtab = reinterpret_cast<ElfW(Sym)*>(base + d->d_un.d_ptr);
1623            break;
1624#if !defined(__LP64__)
1625        case DT_PLTREL:
1626            if (d->d_un.d_val != DT_REL) {
1627                DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1628                return false;
1629            }
1630            break;
1631#endif
1632        case DT_JMPREL:
1633#if defined(USE_RELA)
1634            si->plt_rela = reinterpret_cast<ElfW(Rela)*>(base + d->d_un.d_ptr);
1635#else
1636            si->plt_rel = reinterpret_cast<ElfW(Rel)*>(base + d->d_un.d_ptr);
1637#endif
1638            break;
1639        case DT_PLTRELSZ:
1640#if defined(USE_RELA)
1641            si->plt_rela_count = d->d_un.d_val / sizeof(ElfW(Rela));
1642#else
1643            si->plt_rel_count = d->d_un.d_val / sizeof(ElfW(Rel));
1644#endif
1645            break;
1646#if defined(__mips__)
1647        case DT_PLTGOT:
1648            // Used by mips and mips64.
1649            si->plt_got = reinterpret_cast<ElfW(Addr)**>(base + d->d_un.d_ptr);
1650            break;
1651#endif
1652        case DT_DEBUG:
1653            // Set the DT_DEBUG entry to the address of _r_debug for GDB
1654            // if the dynamic table is writable
1655// FIXME: not working currently for N64
1656// The flags for the LOAD and DYNAMIC program headers do not agree.
1657// The LOAD section containng the dynamic table has been mapped as
1658// read-only, but the DYNAMIC header claims it is writable.
1659#if !(defined(__mips__) && defined(__LP64__))
1660            if ((dynamic_flags & PF_W) != 0) {
1661                d->d_un.d_val = reinterpret_cast<uintptr_t>(&_r_debug);
1662            }
1663            break;
1664#endif
1665#if defined(USE_RELA)
1666         case DT_RELA:
1667            si->rela = reinterpret_cast<ElfW(Rela)*>(base + d->d_un.d_ptr);
1668            break;
1669         case DT_RELASZ:
1670            si->rela_count = d->d_un.d_val / sizeof(ElfW(Rela));
1671            break;
1672        case DT_REL:
1673            DL_ERR("unsupported DT_REL in \"%s\"", si->name);
1674            return false;
1675        case DT_RELSZ:
1676            DL_ERR("unsupported DT_RELSZ in \"%s\"", si->name);
1677            return false;
1678#else
1679        case DT_REL:
1680            si->rel = reinterpret_cast<ElfW(Rel)*>(base + d->d_un.d_ptr);
1681            break;
1682        case DT_RELSZ:
1683            si->rel_count = d->d_un.d_val / sizeof(ElfW(Rel));
1684            break;
1685         case DT_RELA:
1686            DL_ERR("unsupported DT_RELA in \"%s\"", si->name);
1687            return false;
1688#endif
1689        case DT_INIT:
1690            si->init_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1691            DEBUG("%s constructors (DT_INIT) found at %p", si->name, si->init_func);
1692            break;
1693        case DT_FINI:
1694            si->fini_func = reinterpret_cast<linker_function_t>(base + d->d_un.d_ptr);
1695            DEBUG("%s destructors (DT_FINI) found at %p", si->name, si->fini_func);
1696            break;
1697        case DT_INIT_ARRAY:
1698            si->init_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1699            DEBUG("%s constructors (DT_INIT_ARRAY) found at %p", si->name, si->init_array);
1700            break;
1701        case DT_INIT_ARRAYSZ:
1702            si->init_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1703            break;
1704        case DT_FINI_ARRAY:
1705            si->fini_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1706            DEBUG("%s destructors (DT_FINI_ARRAY) found at %p", si->name, si->fini_array);
1707            break;
1708        case DT_FINI_ARRAYSZ:
1709            si->fini_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1710            break;
1711        case DT_PREINIT_ARRAY:
1712            si->preinit_array = reinterpret_cast<linker_function_t*>(base + d->d_un.d_ptr);
1713            DEBUG("%s constructors (DT_PREINIT_ARRAY) found at %p", si->name, si->preinit_array);
1714            break;
1715        case DT_PREINIT_ARRAYSZ:
1716            si->preinit_array_count = ((unsigned)d->d_un.d_val) / sizeof(ElfW(Addr));
1717            break;
1718        case DT_TEXTREL:
1719#if defined(__LP64__)
1720            DL_ERR("text relocations (DT_TEXTREL) found in 64-bit ELF file \"%s\"", si->name);
1721            return false;
1722#else
1723            si->has_text_relocations = true;
1724            break;
1725#endif
1726        case DT_SYMBOLIC:
1727            si->has_DT_SYMBOLIC = true;
1728            break;
1729        case DT_NEEDED:
1730            ++needed_count;
1731            break;
1732        case DT_FLAGS:
1733            if (d->d_un.d_val & DF_TEXTREL) {
1734#if defined(__LP64__)
1735                DL_ERR("text relocations (DF_TEXTREL) found in 64-bit ELF file \"%s\"", si->name);
1736                return false;
1737#else
1738                si->has_text_relocations = true;
1739#endif
1740            }
1741            if (d->d_un.d_val & DF_SYMBOLIC) {
1742                si->has_DT_SYMBOLIC = true;
1743            }
1744            break;
1745#if defined(__mips__)
1746        case DT_STRSZ:
1747        case DT_SYMENT:
1748        case DT_RELENT:
1749             break;
1750        case DT_MIPS_RLD_MAP:
1751            // Set the DT_MIPS_RLD_MAP entry to the address of _r_debug for GDB.
1752            {
1753              r_debug** dp = reinterpret_cast<r_debug**>(d->d_un.d_ptr);
1754              *dp = &_r_debug;
1755            }
1756            break;
1757        case DT_MIPS_RLD_VERSION:
1758        case DT_MIPS_FLAGS:
1759        case DT_MIPS_BASE_ADDRESS:
1760        case DT_MIPS_UNREFEXTNO:
1761            break;
1762
1763        case DT_MIPS_SYMTABNO:
1764            si->mips_symtabno = d->d_un.d_val;
1765            break;
1766
1767        case DT_MIPS_LOCAL_GOTNO:
1768            si->mips_local_gotno = d->d_un.d_val;
1769            break;
1770
1771        case DT_MIPS_GOTSYM:
1772            si->mips_gotsym = d->d_un.d_val;
1773            break;
1774#endif
1775
1776        default:
1777            DEBUG("Unused DT entry: type %p arg %p",
1778                  reinterpret_cast<void*>(d->d_tag), reinterpret_cast<void*>(d->d_un.d_val));
1779            break;
1780        }
1781    }
1782
1783    DEBUG("si->base = %p, si->strtab = %p, si->symtab = %p",
1784          reinterpret_cast<void*>(si->base), si->strtab, si->symtab);
1785
1786    // Sanity checks.
1787    if (relocating_linker && needed_count != 0) {
1788        DL_ERR("linker cannot have DT_NEEDED dependencies on other libraries");
1789        return false;
1790    }
1791    if (si->nbucket == 0) {
1792        DL_ERR("empty/missing DT_HASH in \"%s\" (built with --hash-style=gnu?)", si->name);
1793        return false;
1794    }
1795    if (si->strtab == 0) {
1796        DL_ERR("empty/missing DT_STRTAB in \"%s\"", si->name);
1797        return false;
1798    }
1799    if (si->symtab == 0) {
1800        DL_ERR("empty/missing DT_SYMTAB in \"%s\"", si->name);
1801        return false;
1802    }
1803
1804    // If this is the main executable, then load all of the libraries from LD_PRELOAD now.
1805    if (si->flags & FLAG_EXE) {
1806        memset(gLdPreloads, 0, sizeof(gLdPreloads));
1807        size_t preload_count = 0;
1808        for (size_t i = 0; gLdPreloadNames[i] != NULL; i++) {
1809            soinfo* lsi = find_library(gLdPreloadNames[i]);
1810            if (lsi != NULL) {
1811                gLdPreloads[preload_count++] = lsi;
1812            } else {
1813                // As with glibc, failure to load an LD_PRELOAD library is just a warning.
1814                DL_WARN("could not load library \"%s\" from LD_PRELOAD for \"%s\"; caused by %s",
1815                        gLdPreloadNames[i], si->name, linker_get_error_buffer());
1816            }
1817        }
1818    }
1819
1820    soinfo** needed = reinterpret_cast<soinfo**>(alloca((1 + needed_count) * sizeof(soinfo*)));
1821    soinfo** pneeded = needed;
1822
1823    for (ElfW(Dyn)* d = si->dynamic; d->d_tag != DT_NULL; ++d) {
1824        if (d->d_tag == DT_NEEDED) {
1825            const char* library_name = si->strtab + d->d_un.d_val;
1826            DEBUG("%s needs %s", si->name, library_name);
1827            soinfo* lsi = find_library(library_name);
1828            if (lsi == NULL) {
1829                strlcpy(tmp_err_buf, linker_get_error_buffer(), sizeof(tmp_err_buf));
1830                DL_ERR("could not load library \"%s\" needed by \"%s\"; caused by %s",
1831                       library_name, si->name, tmp_err_buf);
1832                return false;
1833            }
1834            *pneeded++ = lsi;
1835        }
1836    }
1837    *pneeded = NULL;
1838
1839#if !defined(__LP64__)
1840    if (si->has_text_relocations) {
1841        // Make segments writable to allow text relocations to work properly. We will later call
1842        // phdr_table_protect_segments() after all of them are applied and all constructors are run.
1843        DL_WARN("%s has text relocations. This is wasting memory and prevents "
1844                "security hardening. Please fix.", si->name);
1845        if (phdr_table_unprotect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1846            DL_ERR("can't unprotect loadable segments for \"%s\": %s",
1847                   si->name, strerror(errno));
1848            return false;
1849        }
1850    }
1851#endif
1852
1853#if defined(USE_RELA)
1854    if (si->plt_rela != NULL) {
1855        DEBUG("[ relocating %s plt ]\n", si->name);
1856        if (soinfo_relocate(si, si->plt_rela, si->plt_rela_count, needed)) {
1857            return false;
1858        }
1859    }
1860    if (si->rela != NULL) {
1861        DEBUG("[ relocating %s ]\n", si->name);
1862        if (soinfo_relocate(si, si->rela, si->rela_count, needed)) {
1863            return false;
1864        }
1865    }
1866#else
1867    if (si->plt_rel != NULL) {
1868        DEBUG("[ relocating %s plt ]", si->name);
1869        if (soinfo_relocate(si, si->plt_rel, si->plt_rel_count, needed)) {
1870            return false;
1871        }
1872    }
1873    if (si->rel != NULL) {
1874        DEBUG("[ relocating %s ]", si->name);
1875        if (soinfo_relocate(si, si->rel, si->rel_count, needed)) {
1876            return false;
1877        }
1878    }
1879#endif
1880
1881#if defined(__mips__)
1882    if (!mips_relocate_got(si, needed)) {
1883        return false;
1884    }
1885#endif
1886
1887    si->flags |= FLAG_LINKED;
1888    DEBUG("[ finished linking %s ]", si->name);
1889
1890#if !defined(__LP64__)
1891    if (si->has_text_relocations) {
1892        // All relocations are done, we can protect our segments back to read-only.
1893        if (phdr_table_protect_segments(si->phdr, si->phnum, si->load_bias) < 0) {
1894            DL_ERR("can't protect segments for \"%s\": %s",
1895                   si->name, strerror(errno));
1896            return false;
1897        }
1898    }
1899#endif
1900
1901    /* We can also turn on GNU RELRO protection */
1902    if (phdr_table_protect_gnu_relro(si->phdr, si->phnum, si->load_bias) < 0) {
1903        DL_ERR("can't enable GNU RELRO protection for \"%s\": %s",
1904               si->name, strerror(errno));
1905        return false;
1906    }
1907
1908    notify_gdb_of_load(si);
1909    return true;
1910}
1911
1912/*
1913 * This function add vdso to internal dso list.
1914 * It helps to stack unwinding through signal handlers.
1915 * Also, it makes bionic more like glibc.
1916 */
1917static void add_vdso(KernelArgumentBlock& args UNUSED) {
1918#if defined(AT_SYSINFO_EHDR)
1919  ElfW(Ehdr)* ehdr_vdso = reinterpret_cast<ElfW(Ehdr)*>(args.getauxval(AT_SYSINFO_EHDR));
1920  if (ehdr_vdso == NULL) {
1921    return;
1922  }
1923
1924  soinfo* si = soinfo_alloc("[vdso]");
1925
1926  si->phdr = reinterpret_cast<ElfW(Phdr)*>(reinterpret_cast<char*>(ehdr_vdso) + ehdr_vdso->e_phoff);
1927  si->phnum = ehdr_vdso->e_phnum;
1928  si->base = reinterpret_cast<ElfW(Addr)>(ehdr_vdso);
1929  si->size = phdr_table_get_load_size(si->phdr, si->phnum);
1930  si->flags = 0;
1931  si->load_bias = get_elf_exec_load_bias(ehdr_vdso);
1932
1933  soinfo_link_image(si);
1934#endif
1935}
1936
1937/*
1938 * This code is called after the linker has linked itself and
1939 * fixed it's own GOT. It is safe to make references to externs
1940 * and other non-local data at this point.
1941 */
1942static ElfW(Addr) __linker_init_post_relocation(KernelArgumentBlock& args, ElfW(Addr) linker_base) {
1943    /* NOTE: we store the args pointer on a special location
1944     *       of the temporary TLS area in order to pass it to
1945     *       the C Library's runtime initializer.
1946     *
1947     *       The initializer must clear the slot and reset the TLS
1948     *       to point to a different location to ensure that no other
1949     *       shared library constructor can access it.
1950     */
1951  __libc_init_tls(args);
1952
1953#if TIMING
1954    struct timeval t0, t1;
1955    gettimeofday(&t0, 0);
1956#endif
1957
1958    // Initialize environment functions, and get to the ELF aux vectors table.
1959    linker_env_init(args);
1960
1961    // If this is a setuid/setgid program, close the security hole described in
1962    // ftp://ftp.freebsd.org/pub/FreeBSD/CERT/advisories/FreeBSD-SA-02:23.stdio.asc
1963    if (get_AT_SECURE()) {
1964        nullify_closed_stdio();
1965    }
1966
1967    debuggerd_init();
1968
1969    // Get a few environment variables.
1970    const char* LD_DEBUG = linker_env_get("LD_DEBUG");
1971    if (LD_DEBUG != NULL) {
1972      gLdDebugVerbosity = atoi(LD_DEBUG);
1973    }
1974
1975    // Normally, these are cleaned by linker_env_init, but the test
1976    // doesn't cost us anything.
1977    const char* ldpath_env = NULL;
1978    const char* ldpreload_env = NULL;
1979    if (!get_AT_SECURE()) {
1980      ldpath_env = linker_env_get("LD_LIBRARY_PATH");
1981      ldpreload_env = linker_env_get("LD_PRELOAD");
1982    }
1983
1984    INFO("[ android linker & debugger ]");
1985
1986    soinfo* si = soinfo_alloc(args.argv[0]);
1987    if (si == NULL) {
1988        exit(EXIT_FAILURE);
1989    }
1990
1991    /* bootstrap the link map, the main exe always needs to be first */
1992    si->flags |= FLAG_EXE;
1993    link_map* map = &(si->link_map_head);
1994
1995    map->l_addr = 0;
1996    map->l_name = args.argv[0];
1997    map->l_prev = NULL;
1998    map->l_next = NULL;
1999
2000    _r_debug.r_map = map;
2001    r_debug_tail = map;
2002
2003    /* gdb expects the linker to be in the debug shared object list.
2004     * Without this, gdb has trouble locating the linker's ".text"
2005     * and ".plt" sections. Gdb could also potentially use this to
2006     * relocate the offset of our exported 'rtld_db_dlactivity' symbol.
2007     * Don't use soinfo_alloc(), because the linker shouldn't
2008     * be on the soinfo list.
2009     */
2010    {
2011        static soinfo linker_soinfo;
2012#if defined(__LP64__)
2013        strlcpy(linker_soinfo.name, "/system/bin/linker64", sizeof(linker_soinfo.name));
2014#else
2015        strlcpy(linker_soinfo.name, "/system/bin/linker", sizeof(linker_soinfo.name));
2016#endif
2017        linker_soinfo.flags = 0;
2018        linker_soinfo.base = linker_base;
2019
2020        /*
2021         * Set the dynamic field in the link map otherwise gdb will complain with
2022         * the following:
2023         *   warning: .dynamic section for "/system/bin/linker" is not at the
2024         *   expected address (wrong library or version mismatch?)
2025         */
2026        ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_base);
2027        ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_base + elf_hdr->e_phoff);
2028        phdr_table_get_dynamic_section(phdr, elf_hdr->e_phnum, linker_base,
2029                                       &linker_soinfo.dynamic, NULL, NULL);
2030        insert_soinfo_into_debug_map(&linker_soinfo);
2031    }
2032
2033    // Extract information passed from the kernel.
2034    si->phdr = reinterpret_cast<ElfW(Phdr)*>(args.getauxval(AT_PHDR));
2035    si->phnum = args.getauxval(AT_PHNUM);
2036    si->entry = args.getauxval(AT_ENTRY);
2037
2038    /* Compute the value of si->base. We can't rely on the fact that
2039     * the first entry is the PHDR because this will not be true
2040     * for certain executables (e.g. some in the NDK unit test suite)
2041     */
2042    si->base = 0;
2043    si->size = phdr_table_get_load_size(si->phdr, si->phnum);
2044    si->load_bias = 0;
2045    for (size_t i = 0; i < si->phnum; ++i) {
2046      if (si->phdr[i].p_type == PT_PHDR) {
2047        si->load_bias = reinterpret_cast<ElfW(Addr)>(si->phdr) - si->phdr[i].p_vaddr;
2048        si->base = reinterpret_cast<ElfW(Addr)>(si->phdr) - si->phdr[i].p_offset;
2049        break;
2050      }
2051    }
2052    si->dynamic = NULL;
2053    si->ref_count = 1;
2054
2055    // Use LD_LIBRARY_PATH and LD_PRELOAD (but only if we aren't setuid/setgid).
2056    parse_LD_LIBRARY_PATH(ldpath_env);
2057    parse_LD_PRELOAD(ldpreload_env);
2058
2059    somain = si;
2060
2061    if (!soinfo_link_image(si)) {
2062        __libc_format_fd(2, "CANNOT LINK EXECUTABLE: %s\n", linker_get_error_buffer());
2063        exit(EXIT_FAILURE);
2064    }
2065
2066    add_vdso(args);
2067
2068    si->CallPreInitConstructors();
2069
2070    for (size_t i = 0; gLdPreloads[i] != NULL; ++i) {
2071        gLdPreloads[i]->CallConstructors();
2072    }
2073
2074    /* After the link_image, the si->load_bias is initialized.
2075     * For so lib, the map->l_addr will be updated in notify_gdb_of_load.
2076     * We need to update this value for so exe here. So Unwind_Backtrace
2077     * for some arch like x86 could work correctly within so exe.
2078     */
2079    map->l_addr = si->load_bias;
2080    si->CallConstructors();
2081
2082#if TIMING
2083    gettimeofday(&t1, NULL);
2084    PRINT("LINKER TIME: %s: %d microseconds", args.argv[0], (int) (
2085               (((long long)t1.tv_sec * 1000000LL) + (long long)t1.tv_usec) -
2086               (((long long)t0.tv_sec * 1000000LL) + (long long)t0.tv_usec)));
2087#endif
2088#if STATS
2089    PRINT("RELO STATS: %s: %d abs, %d rel, %d copy, %d symbol", args.argv[0],
2090           linker_stats.count[kRelocAbsolute],
2091           linker_stats.count[kRelocRelative],
2092           linker_stats.count[kRelocCopy],
2093           linker_stats.count[kRelocSymbol]);
2094#endif
2095#if COUNT_PAGES
2096    {
2097        unsigned n;
2098        unsigned i;
2099        unsigned count = 0;
2100        for (n = 0; n < 4096; n++) {
2101            if (bitmask[n]) {
2102                unsigned x = bitmask[n];
2103#if defined(__LP64__)
2104                for (i = 0; i < 32; i++) {
2105#else
2106                for (i = 0; i < 8; i++) {
2107#endif
2108                    if (x & 1) {
2109                        count++;
2110                    }
2111                    x >>= 1;
2112                }
2113            }
2114        }
2115        PRINT("PAGES MODIFIED: %s: %d (%dKB)", args.argv[0], count, count * 4);
2116    }
2117#endif
2118
2119#if TIMING || STATS || COUNT_PAGES
2120    fflush(stdout);
2121#endif
2122
2123    TRACE("[ Ready to execute '%s' @ %p ]", si->name, reinterpret_cast<void*>(si->entry));
2124    return si->entry;
2125}
2126
2127/* Compute the load-bias of an existing executable. This shall only
2128 * be used to compute the load bias of an executable or shared library
2129 * that was loaded by the kernel itself.
2130 *
2131 * Input:
2132 *    elf    -> address of ELF header, assumed to be at the start of the file.
2133 * Return:
2134 *    load bias, i.e. add the value of any p_vaddr in the file to get
2135 *    the corresponding address in memory.
2136 */
2137static ElfW(Addr) get_elf_exec_load_bias(const ElfW(Ehdr)* elf) {
2138  ElfW(Addr) offset = elf->e_phoff;
2139  const ElfW(Phdr)* phdr_table = reinterpret_cast<const ElfW(Phdr)*>(reinterpret_cast<uintptr_t>(elf) + offset);
2140  const ElfW(Phdr)* phdr_end = phdr_table + elf->e_phnum;
2141
2142  for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_end; phdr++) {
2143    if (phdr->p_type == PT_LOAD) {
2144      return reinterpret_cast<ElfW(Addr)>(elf) + phdr->p_offset - phdr->p_vaddr;
2145    }
2146  }
2147  return 0;
2148}
2149
2150/*
2151 * This is the entry point for the linker, called from begin.S. This
2152 * method is responsible for fixing the linker's own relocations, and
2153 * then calling __linker_init_post_relocation().
2154 *
2155 * Because this method is called before the linker has fixed it's own
2156 * relocations, any attempt to reference an extern variable, extern
2157 * function, or other GOT reference will generate a segfault.
2158 */
2159extern "C" ElfW(Addr) __linker_init(void* raw_args) {
2160  KernelArgumentBlock args(raw_args);
2161
2162  ElfW(Addr) linker_addr = args.getauxval(AT_BASE);
2163  ElfW(Ehdr)* elf_hdr = reinterpret_cast<ElfW(Ehdr)*>(linker_addr);
2164  ElfW(Phdr)* phdr = reinterpret_cast<ElfW(Phdr)*>(linker_addr + elf_hdr->e_phoff);
2165
2166  soinfo linker_so;
2167  memset(&linker_so, 0, sizeof(soinfo));
2168
2169  strcpy(linker_so.name, "[dynamic linker]");
2170  linker_so.base = linker_addr;
2171  linker_so.size = phdr_table_get_load_size(phdr, elf_hdr->e_phnum);
2172  linker_so.load_bias = get_elf_exec_load_bias(elf_hdr);
2173  linker_so.dynamic = NULL;
2174  linker_so.phdr = phdr;
2175  linker_so.phnum = elf_hdr->e_phnum;
2176  linker_so.flags |= FLAG_LINKER;
2177
2178  if (!soinfo_link_image(&linker_so)) {
2179    // It would be nice to print an error message, but if the linker
2180    // can't link itself, there's no guarantee that we'll be able to
2181    // call write() (because it involves a GOT reference). We may as
2182    // well try though...
2183    const char* msg = "CANNOT LINK EXECUTABLE: ";
2184    write(2, msg, strlen(msg));
2185    write(2, __linker_dl_err_buf, strlen(__linker_dl_err_buf));
2186    write(2, "\n", 1);
2187    _exit(EXIT_FAILURE);
2188  }
2189
2190  // We have successfully fixed our own relocations. It's safe to run
2191  // the main part of the linker now.
2192  args.abort_message_ptr = &gAbortMessage;
2193  ElfW(Addr) start_address = __linker_init_post_relocation(args, linker_addr);
2194
2195  set_soinfo_pool_protection(PROT_READ);
2196
2197  // Return the address that the calling assembly stub should jump to.
2198  return start_address;
2199}
2200