6468f203cb69b1dbfb6c5731c4e3f7c650546c7e
[oweals/musl.git] / ldso / dynlink.c
1 #define _GNU_SOURCE
2 #define SYSCALL_NO_TLS 1
3 #include <stdio.h>
4 #include <stdlib.h>
5 #include <stdarg.h>
6 #include <stddef.h>
7 #include <string.h>
8 #include <unistd.h>
9 #include <stdint.h>
10 #include <elf.h>
11 #include <sys/mman.h>
12 #include <limits.h>
13 #include <fcntl.h>
14 #include <sys/stat.h>
15 #include <errno.h>
16 #include <link.h>
17 #include <setjmp.h>
18 #include <pthread.h>
19 #include <ctype.h>
20 #include <dlfcn.h>
21 #include <semaphore.h>
22 #include <sys/membarrier.h>
23 #include "pthread_impl.h"
24 #include "libc.h"
25 #include "dynlink.h"
26 #include "malloc_impl.h"
27
28 static void error(const char *, ...);
29
30 #define MAXP2(a,b) (-(-(a)&-(b)))
31 #define ALIGN(x,y) ((x)+(y)-1 & -(y))
32
33 #define container_of(p,t,m) ((t*)((char *)(p)-offsetof(t,m)))
34 #define countof(a) ((sizeof (a))/(sizeof (a)[0]))
35
36 struct debug {
37         int ver;
38         void *head;
39         void (*bp)(void);
40         int state;
41         void *base;
42 };
43
44 struct td_index {
45         size_t args[2];
46         struct td_index *next;
47 };
48
49 struct dso {
50 #if DL_FDPIC
51         struct fdpic_loadmap *loadmap;
52 #else
53         unsigned char *base;
54 #endif
55         char *name;
56         size_t *dynv;
57         struct dso *next, *prev;
58
59         Phdr *phdr;
60         int phnum;
61         size_t phentsize;
62         Sym *syms;
63         Elf_Symndx *hashtab;
64         uint32_t *ghashtab;
65         int16_t *versym;
66         char *strings;
67         struct dso *syms_next, *lazy_next;
68         size_t *lazy, lazy_cnt;
69         unsigned char *map;
70         size_t map_len;
71         dev_t dev;
72         ino_t ino;
73         char relocated;
74         char constructed;
75         char kernel_mapped;
76         char mark;
77         char bfs_built;
78         char runtime_loaded;
79         struct dso **deps, *needed_by;
80         size_t ndeps_direct;
81         size_t next_dep;
82         int ctor_visitor;
83         char *rpath_orig, *rpath;
84         struct tls_module tls;
85         size_t tls_id;
86         size_t relro_start, relro_end;
87         uintptr_t *new_dtv;
88         unsigned char *new_tls;
89         struct td_index *td_index;
90         struct dso *fini_next;
91         char *shortname;
92 #if DL_FDPIC
93         unsigned char *base;
94 #else
95         struct fdpic_loadmap *loadmap;
96 #endif
97         struct funcdesc {
98                 void *addr;
99                 size_t *got;
100         } *funcdescs;
101         size_t *got;
102         char buf[];
103 };
104
105 struct symdef {
106         Sym *sym;
107         struct dso *dso;
108 };
109
110 typedef void (*stage3_func)(size_t *, size_t *);
111
112 static struct builtin_tls {
113         char c;
114         struct pthread pt;
115         void *space[16];
116 } builtin_tls[1];
117 #define MIN_TLS_ALIGN offsetof(struct builtin_tls, pt)
118
119 #define ADDEND_LIMIT 4096
120 static size_t *saved_addends, *apply_addends_to;
121
122 static struct dso ldso;
123 static struct dso *head, *tail, *fini_head, *syms_tail, *lazy_head;
124 static char *env_path, *sys_path;
125 static unsigned long long gencnt;
126 static int runtime;
127 static int ldd_mode;
128 static int ldso_fail;
129 static int noload;
130 static int shutting_down;
131 static jmp_buf *rtld_fail;
132 static pthread_rwlock_t lock;
133 static struct debug debug;
134 static struct tls_module *tls_tail;
135 static size_t tls_cnt, tls_offset, tls_align = MIN_TLS_ALIGN;
136 static size_t static_tls_cnt;
137 static pthread_mutex_t init_fini_lock;
138 static pthread_cond_t ctor_cond;
139 static struct dso *builtin_deps[2];
140 static struct dso *const no_deps[1];
141 static struct dso *builtin_ctor_queue[4];
142 static struct dso **main_ctor_queue;
143 static struct fdpic_loadmap *app_loadmap;
144 static struct fdpic_dummy_loadmap app_dummy_loadmap;
145
146 struct debug *_dl_debug_addr = &debug;
147
148 extern hidden int __malloc_replaced;
149
150 hidden void (*const __init_array_start)(void)=0, (*const __fini_array_start)(void)=0;
151
152 extern hidden void (*const __init_array_end)(void), (*const __fini_array_end)(void);
153
154 weak_alias(__init_array_start, __init_array_end);
155 weak_alias(__fini_array_start, __fini_array_end);
156
157 static int dl_strcmp(const char *l, const char *r)
158 {
159         for (; *l==*r && *l; l++, r++);
160         return *(unsigned char *)l - *(unsigned char *)r;
161 }
162 #define strcmp(l,r) dl_strcmp(l,r)
163
164 /* Compute load address for a virtual address in a given dso. */
165 #if DL_FDPIC
166 static void *laddr(const struct dso *p, size_t v)
167 {
168         size_t j=0;
169         if (!p->loadmap) return p->base + v;
170         for (j=0; v-p->loadmap->segs[j].p_vaddr >= p->loadmap->segs[j].p_memsz; j++);
171         return (void *)(v - p->loadmap->segs[j].p_vaddr + p->loadmap->segs[j].addr);
172 }
173 static void *laddr_pg(const struct dso *p, size_t v)
174 {
175         size_t j=0;
176         size_t pgsz = PAGE_SIZE;
177         if (!p->loadmap) return p->base + v;
178         for (j=0; ; j++) {
179                 size_t a = p->loadmap->segs[j].p_vaddr;
180                 size_t b = a + p->loadmap->segs[j].p_memsz;
181                 a &= -pgsz;
182                 b += pgsz-1;
183                 b &= -pgsz;
184                 if (v-a<b-a) break;
185         }
186         return (void *)(v - p->loadmap->segs[j].p_vaddr + p->loadmap->segs[j].addr);
187 }
188 static void (*fdbarrier(void *p))()
189 {
190         void (*fd)();
191         __asm__("" : "=r"(fd) : "0"(p));
192         return fd;
193 }
194 #define fpaddr(p, v) fdbarrier((&(struct funcdesc){ \
195         laddr(p, v), (p)->got }))
196 #else
197 #define laddr(p, v) (void *)((p)->base + (v))
198 #define laddr_pg(p, v) laddr(p, v)
199 #define fpaddr(p, v) ((void (*)())laddr(p, v))
200 #endif
201
202 static void decode_vec(size_t *v, size_t *a, size_t cnt)
203 {
204         size_t i;
205         for (i=0; i<cnt; i++) a[i] = 0;
206         for (; v[0]; v+=2) if (v[0]-1<cnt-1) {
207                 a[0] |= 1UL<<v[0];
208                 a[v[0]] = v[1];
209         }
210 }
211
212 static int search_vec(size_t *v, size_t *r, size_t key)
213 {
214         for (; v[0]!=key; v+=2)
215                 if (!v[0]) return 0;
216         *r = v[1];
217         return 1;
218 }
219
220 static uint32_t sysv_hash(const char *s0)
221 {
222         const unsigned char *s = (void *)s0;
223         uint_fast32_t h = 0;
224         while (*s) {
225                 h = 16*h + *s++;
226                 h ^= h>>24 & 0xf0;
227         }
228         return h & 0xfffffff;
229 }
230
231 static uint32_t gnu_hash(const char *s0)
232 {
233         const unsigned char *s = (void *)s0;
234         uint_fast32_t h = 5381;
235         for (; *s; s++)
236                 h += h*32 + *s;
237         return h;
238 }
239
240 static Sym *sysv_lookup(const char *s, uint32_t h, struct dso *dso)
241 {
242         size_t i;
243         Sym *syms = dso->syms;
244         Elf_Symndx *hashtab = dso->hashtab;
245         char *strings = dso->strings;
246         for (i=hashtab[2+h%hashtab[0]]; i; i=hashtab[2+hashtab[0]+i]) {
247                 if ((!dso->versym || dso->versym[i] >= 0)
248                     && (!strcmp(s, strings+syms[i].st_name)))
249                         return syms+i;
250         }
251         return 0;
252 }
253
254 static Sym *gnu_lookup(uint32_t h1, uint32_t *hashtab, struct dso *dso, const char *s)
255 {
256         uint32_t nbuckets = hashtab[0];
257         uint32_t *buckets = hashtab + 4 + hashtab[2]*(sizeof(size_t)/4);
258         uint32_t i = buckets[h1 % nbuckets];
259
260         if (!i) return 0;
261
262         uint32_t *hashval = buckets + nbuckets + (i - hashtab[1]);
263
264         for (h1 |= 1; ; i++) {
265                 uint32_t h2 = *hashval++;
266                 if ((h1 == (h2|1)) && (!dso->versym || dso->versym[i] >= 0)
267                     && !strcmp(s, dso->strings + dso->syms[i].st_name))
268                         return dso->syms+i;
269                 if (h2 & 1) break;
270         }
271
272         return 0;
273 }
274
275 static Sym *gnu_lookup_filtered(uint32_t h1, uint32_t *hashtab, struct dso *dso, const char *s, uint32_t fofs, size_t fmask)
276 {
277         const size_t *bloomwords = (const void *)(hashtab+4);
278         size_t f = bloomwords[fofs & (hashtab[2]-1)];
279         if (!(f & fmask)) return 0;
280
281         f >>= (h1 >> hashtab[3]) % (8 * sizeof f);
282         if (!(f & 1)) return 0;
283
284         return gnu_lookup(h1, hashtab, dso, s);
285 }
286
287 #define OK_TYPES (1<<STT_NOTYPE | 1<<STT_OBJECT | 1<<STT_FUNC | 1<<STT_COMMON | 1<<STT_TLS)
288 #define OK_BINDS (1<<STB_GLOBAL | 1<<STB_WEAK | 1<<STB_GNU_UNIQUE)
289
290 #ifndef ARCH_SYM_REJECT_UND
291 #define ARCH_SYM_REJECT_UND(s) 0
292 #endif
293
294 #if defined(__GNUC__)
295 __attribute__((always_inline))
296 #endif
297 static inline struct symdef find_sym2(struct dso *dso, const char *s, int need_def, int use_deps)
298 {
299         uint32_t h = 0, gh = gnu_hash(s), gho = gh / (8*sizeof(size_t)), *ght;
300         size_t ghm = 1ul << gh % (8*sizeof(size_t));
301         struct symdef def = {0};
302         struct dso **deps = use_deps ? dso->deps : 0;
303         for (; dso; dso=use_deps ? *deps++ : dso->syms_next) {
304                 Sym *sym;
305                 if ((ght = dso->ghashtab)) {
306                         sym = gnu_lookup_filtered(gh, ght, dso, s, gho, ghm);
307                 } else {
308                         if (!h) h = sysv_hash(s);
309                         sym = sysv_lookup(s, h, dso);
310                 }
311                 if (!sym) continue;
312                 if (!sym->st_shndx)
313                         if (need_def || (sym->st_info&0xf) == STT_TLS
314                             || ARCH_SYM_REJECT_UND(sym))
315                                 continue;
316                 if (!sym->st_value)
317                         if ((sym->st_info&0xf) != STT_TLS)
318                                 continue;
319                 if (!(1<<(sym->st_info&0xf) & OK_TYPES)) continue;
320                 if (!(1<<(sym->st_info>>4) & OK_BINDS)) continue;
321                 def.sym = sym;
322                 def.dso = dso;
323                 break;
324         }
325         return def;
326 }
327
328 static struct symdef find_sym(struct dso *dso, const char *s, int need_def)
329 {
330         return find_sym2(dso, s, need_def, 0);
331 }
332
333 static void do_relocs(struct dso *dso, size_t *rel, size_t rel_size, size_t stride)
334 {
335         unsigned char *base = dso->base;
336         Sym *syms = dso->syms;
337         char *strings = dso->strings;
338         Sym *sym;
339         const char *name;
340         void *ctx;
341         int type;
342         int sym_index;
343         struct symdef def;
344         size_t *reloc_addr;
345         size_t sym_val;
346         size_t tls_val;
347         size_t addend;
348         int skip_relative = 0, reuse_addends = 0, save_slot = 0;
349
350         if (dso == &ldso) {
351                 /* Only ldso's REL table needs addend saving/reuse. */
352                 if (rel == apply_addends_to)
353                         reuse_addends = 1;
354                 skip_relative = 1;
355         }
356
357         for (; rel_size; rel+=stride, rel_size-=stride*sizeof(size_t)) {
358                 if (skip_relative && IS_RELATIVE(rel[1], dso->syms)) continue;
359                 type = R_TYPE(rel[1]);
360                 if (type == REL_NONE) continue;
361                 reloc_addr = laddr(dso, rel[0]);
362
363                 if (stride > 2) {
364                         addend = rel[2];
365                 } else if (type==REL_GOT || type==REL_PLT|| type==REL_COPY) {
366                         addend = 0;
367                 } else if (reuse_addends) {
368                         /* Save original addend in stage 2 where the dso
369                          * chain consists of just ldso; otherwise read back
370                          * saved addend since the inline one was clobbered. */
371                         if (head==&ldso)
372                                 saved_addends[save_slot] = *reloc_addr;
373                         addend = saved_addends[save_slot++];
374                 } else {
375                         addend = *reloc_addr;
376                 }
377
378                 sym_index = R_SYM(rel[1]);
379                 if (sym_index) {
380                         sym = syms + sym_index;
381                         name = strings + sym->st_name;
382                         ctx = type==REL_COPY ? head->syms_next : head;
383                         def = (sym->st_info>>4) == STB_LOCAL
384                                 ? (struct symdef){ .dso = dso, .sym = sym }
385                                 : find_sym(ctx, name, type==REL_PLT);
386                         if (!def.sym && (sym->st_shndx != SHN_UNDEF
387                             || sym->st_info>>4 != STB_WEAK)) {
388                                 if (dso->lazy && (type==REL_PLT || type==REL_GOT)) {
389                                         dso->lazy[3*dso->lazy_cnt+0] = rel[0];
390                                         dso->lazy[3*dso->lazy_cnt+1] = rel[1];
391                                         dso->lazy[3*dso->lazy_cnt+2] = addend;
392                                         dso->lazy_cnt++;
393                                         continue;
394                                 }
395                                 error("Error relocating %s: %s: symbol not found",
396                                         dso->name, name);
397                                 if (runtime) longjmp(*rtld_fail, 1);
398                                 continue;
399                         }
400                 } else {
401                         sym = 0;
402                         def.sym = 0;
403                         def.dso = dso;
404                 }
405
406                 sym_val = def.sym ? (size_t)laddr(def.dso, def.sym->st_value) : 0;
407                 tls_val = def.sym ? def.sym->st_value : 0;
408
409                 if ((type == REL_TPOFF || type == REL_TPOFF_NEG)
410                     && def.dso->tls_id > static_tls_cnt) {
411                         error("Error relocating %s: %s: initial-exec TLS "
412                                 "resolves to dynamic definition in %s",
413                                 dso->name, name, def.dso->name);
414                         longjmp(*rtld_fail, 1);
415                 }
416
417                 switch(type) {
418                 case REL_OFFSET:
419                         addend -= (size_t)reloc_addr;
420                 case REL_SYMBOLIC:
421                 case REL_GOT:
422                 case REL_PLT:
423                         *reloc_addr = sym_val + addend;
424                         break;
425                 case REL_USYMBOLIC:
426                         memcpy(reloc_addr, &(size_t){sym_val + addend}, sizeof(size_t));
427                         break;
428                 case REL_RELATIVE:
429                         *reloc_addr = (size_t)base + addend;
430                         break;
431                 case REL_SYM_OR_REL:
432                         if (sym) *reloc_addr = sym_val + addend;
433                         else *reloc_addr = (size_t)base + addend;
434                         break;
435                 case REL_COPY:
436                         memcpy(reloc_addr, (void *)sym_val, sym->st_size);
437                         break;
438                 case REL_OFFSET32:
439                         *(uint32_t *)reloc_addr = sym_val + addend
440                                 - (size_t)reloc_addr;
441                         break;
442                 case REL_FUNCDESC:
443                         *reloc_addr = def.sym ? (size_t)(def.dso->funcdescs
444                                 + (def.sym - def.dso->syms)) : 0;
445                         break;
446                 case REL_FUNCDESC_VAL:
447                         if ((sym->st_info&0xf) == STT_SECTION) *reloc_addr += sym_val;
448                         else *reloc_addr = sym_val;
449                         reloc_addr[1] = def.sym ? (size_t)def.dso->got : 0;
450                         break;
451                 case REL_DTPMOD:
452                         *reloc_addr = def.dso->tls_id;
453                         break;
454                 case REL_DTPOFF:
455                         *reloc_addr = tls_val + addend - DTP_OFFSET;
456                         break;
457 #ifdef TLS_ABOVE_TP
458                 case REL_TPOFF:
459                         *reloc_addr = tls_val + def.dso->tls.offset + TPOFF_K + addend;
460                         break;
461 #else
462                 case REL_TPOFF:
463                         *reloc_addr = tls_val - def.dso->tls.offset + addend;
464                         break;
465                 case REL_TPOFF_NEG:
466                         *reloc_addr = def.dso->tls.offset - tls_val + addend;
467                         break;
468 #endif
469                 case REL_TLSDESC:
470                         if (stride<3) addend = reloc_addr[1];
471                         if (def.dso->tls_id > static_tls_cnt) {
472                                 struct td_index *new = malloc(sizeof *new);
473                                 if (!new) {
474                                         error(
475                                         "Error relocating %s: cannot allocate TLSDESC for %s",
476                                         dso->name, sym ? name : "(local)" );
477                                         longjmp(*rtld_fail, 1);
478                                 }
479                                 new->next = dso->td_index;
480                                 dso->td_index = new;
481                                 new->args[0] = def.dso->tls_id;
482                                 new->args[1] = tls_val + addend - DTP_OFFSET;
483                                 reloc_addr[0] = (size_t)__tlsdesc_dynamic;
484                                 reloc_addr[1] = (size_t)new;
485                         } else {
486                                 reloc_addr[0] = (size_t)__tlsdesc_static;
487 #ifdef TLS_ABOVE_TP
488                                 reloc_addr[1] = tls_val + def.dso->tls.offset
489                                         + TPOFF_K + addend;
490 #else
491                                 reloc_addr[1] = tls_val - def.dso->tls.offset
492                                         + addend;
493 #endif
494                         }
495 #ifdef TLSDESC_BACKWARDS
496                         /* Some archs (32-bit ARM at least) invert the order of
497                          * the descriptor members. Fix them up here. */
498                         size_t tmp = reloc_addr[0];
499                         reloc_addr[0] = reloc_addr[1];
500                         reloc_addr[1] = tmp;
501 #endif
502                         break;
503                 default:
504                         error("Error relocating %s: unsupported relocation type %d",
505                                 dso->name, type);
506                         if (runtime) longjmp(*rtld_fail, 1);
507                         continue;
508                 }
509         }
510 }
511
512 static void redo_lazy_relocs()
513 {
514         struct dso *p = lazy_head, *next;
515         lazy_head = 0;
516         for (; p; p=next) {
517                 next = p->lazy_next;
518                 size_t size = p->lazy_cnt*3*sizeof(size_t);
519                 p->lazy_cnt = 0;
520                 do_relocs(p, p->lazy, size, 3);
521                 if (p->lazy_cnt) {
522                         p->lazy_next = lazy_head;
523                         lazy_head = p;
524                 } else {
525                         free(p->lazy);
526                         p->lazy = 0;
527                         p->lazy_next = 0;
528                 }
529         }
530 }
531
532 /* A huge hack: to make up for the wastefulness of shared libraries
533  * needing at least a page of dirty memory even if they have no global
534  * data, we reclaim the gaps at the beginning and end of writable maps
535  * and "donate" them to the heap. */
536
537 static void reclaim(struct dso *dso, size_t start, size_t end)
538 {
539         if (start >= dso->relro_start && start < dso->relro_end) start = dso->relro_end;
540         if (end   >= dso->relro_start && end   < dso->relro_end) end = dso->relro_start;
541         if (start >= end) return;
542         char *base = laddr_pg(dso, start);
543         __malloc_donate(base, base+(end-start));
544 }
545
546 static void reclaim_gaps(struct dso *dso)
547 {
548         Phdr *ph = dso->phdr;
549         size_t phcnt = dso->phnum;
550
551         for (; phcnt--; ph=(void *)((char *)ph+dso->phentsize)) {
552                 if (ph->p_type!=PT_LOAD) continue;
553                 if ((ph->p_flags&(PF_R|PF_W))!=(PF_R|PF_W)) continue;
554                 reclaim(dso, ph->p_vaddr & -PAGE_SIZE, ph->p_vaddr);
555                 reclaim(dso, ph->p_vaddr+ph->p_memsz,
556                         ph->p_vaddr+ph->p_memsz+PAGE_SIZE-1 & -PAGE_SIZE);
557         }
558 }
559
560 static void *mmap_fixed(void *p, size_t n, int prot, int flags, int fd, off_t off)
561 {
562         static int no_map_fixed;
563         char *q;
564         if (!no_map_fixed) {
565                 q = mmap(p, n, prot, flags|MAP_FIXED, fd, off);
566                 if (!DL_NOMMU_SUPPORT || q != MAP_FAILED || errno != EINVAL)
567                         return q;
568                 no_map_fixed = 1;
569         }
570         /* Fallbacks for MAP_FIXED failure on NOMMU kernels. */
571         if (flags & MAP_ANONYMOUS) {
572                 memset(p, 0, n);
573                 return p;
574         }
575         ssize_t r;
576         if (lseek(fd, off, SEEK_SET) < 0) return MAP_FAILED;
577         for (q=p; n; q+=r, off+=r, n-=r) {
578                 r = read(fd, q, n);
579                 if (r < 0 && errno != EINTR) return MAP_FAILED;
580                 if (!r) {
581                         memset(q, 0, n);
582                         break;
583                 }
584         }
585         return p;
586 }
587
588 static void unmap_library(struct dso *dso)
589 {
590         if (dso->loadmap) {
591                 size_t i;
592                 for (i=0; i<dso->loadmap->nsegs; i++) {
593                         if (!dso->loadmap->segs[i].p_memsz)
594                                 continue;
595                         munmap((void *)dso->loadmap->segs[i].addr,
596                                 dso->loadmap->segs[i].p_memsz);
597                 }
598                 free(dso->loadmap);
599         } else if (dso->map && dso->map_len) {
600                 munmap(dso->map, dso->map_len);
601         }
602 }
603
604 static void *map_library(int fd, struct dso *dso)
605 {
606         Ehdr buf[(896+sizeof(Ehdr))/sizeof(Ehdr)];
607         void *allocated_buf=0;
608         size_t phsize;
609         size_t addr_min=SIZE_MAX, addr_max=0, map_len;
610         size_t this_min, this_max;
611         size_t nsegs = 0;
612         off_t off_start;
613         Ehdr *eh;
614         Phdr *ph, *ph0;
615         unsigned prot;
616         unsigned char *map=MAP_FAILED, *base;
617         size_t dyn=0;
618         size_t tls_image=0;
619         size_t i;
620
621         ssize_t l = read(fd, buf, sizeof buf);
622         eh = buf;
623         if (l<0) return 0;
624         if (l<sizeof *eh || (eh->e_type != ET_DYN && eh->e_type != ET_EXEC))
625                 goto noexec;
626         phsize = eh->e_phentsize * eh->e_phnum;
627         if (phsize > sizeof buf - sizeof *eh) {
628                 allocated_buf = malloc(phsize);
629                 if (!allocated_buf) return 0;
630                 l = pread(fd, allocated_buf, phsize, eh->e_phoff);
631                 if (l < 0) goto error;
632                 if (l != phsize) goto noexec;
633                 ph = ph0 = allocated_buf;
634         } else if (eh->e_phoff + phsize > l) {
635                 l = pread(fd, buf+1, phsize, eh->e_phoff);
636                 if (l < 0) goto error;
637                 if (l != phsize) goto noexec;
638                 ph = ph0 = (void *)(buf + 1);
639         } else {
640                 ph = ph0 = (void *)((char *)buf + eh->e_phoff);
641         }
642         for (i=eh->e_phnum; i; i--, ph=(void *)((char *)ph+eh->e_phentsize)) {
643                 if (ph->p_type == PT_DYNAMIC) {
644                         dyn = ph->p_vaddr;
645                 } else if (ph->p_type == PT_TLS) {
646                         tls_image = ph->p_vaddr;
647                         dso->tls.align = ph->p_align;
648                         dso->tls.len = ph->p_filesz;
649                         dso->tls.size = ph->p_memsz;
650                 } else if (ph->p_type == PT_GNU_RELRO) {
651                         dso->relro_start = ph->p_vaddr & -PAGE_SIZE;
652                         dso->relro_end = (ph->p_vaddr + ph->p_memsz) & -PAGE_SIZE;
653                 } else if (ph->p_type == PT_GNU_STACK) {
654                         if (!runtime && ph->p_memsz > __default_stacksize) {
655                                 __default_stacksize =
656                                         ph->p_memsz < DEFAULT_STACK_MAX ?
657                                         ph->p_memsz : DEFAULT_STACK_MAX;
658                         }
659                 }
660                 if (ph->p_type != PT_LOAD) continue;
661                 nsegs++;
662                 if (ph->p_vaddr < addr_min) {
663                         addr_min = ph->p_vaddr;
664                         off_start = ph->p_offset;
665                         prot = (((ph->p_flags&PF_R) ? PROT_READ : 0) |
666                                 ((ph->p_flags&PF_W) ? PROT_WRITE: 0) |
667                                 ((ph->p_flags&PF_X) ? PROT_EXEC : 0));
668                 }
669                 if (ph->p_vaddr+ph->p_memsz > addr_max) {
670                         addr_max = ph->p_vaddr+ph->p_memsz;
671                 }
672         }
673         if (!dyn) goto noexec;
674         if (DL_FDPIC && !(eh->e_flags & FDPIC_CONSTDISP_FLAG)) {
675                 dso->loadmap = calloc(1, sizeof *dso->loadmap
676                         + nsegs * sizeof *dso->loadmap->segs);
677                 if (!dso->loadmap) goto error;
678                 dso->loadmap->nsegs = nsegs;
679                 for (ph=ph0, i=0; i<nsegs; ph=(void *)((char *)ph+eh->e_phentsize)) {
680                         if (ph->p_type != PT_LOAD) continue;
681                         prot = (((ph->p_flags&PF_R) ? PROT_READ : 0) |
682                                 ((ph->p_flags&PF_W) ? PROT_WRITE: 0) |
683                                 ((ph->p_flags&PF_X) ? PROT_EXEC : 0));
684                         map = mmap(0, ph->p_memsz + (ph->p_vaddr & PAGE_SIZE-1),
685                                 prot, MAP_PRIVATE,
686                                 fd, ph->p_offset & -PAGE_SIZE);
687                         if (map == MAP_FAILED) {
688                                 unmap_library(dso);
689                                 goto error;
690                         }
691                         dso->loadmap->segs[i].addr = (size_t)map +
692                                 (ph->p_vaddr & PAGE_SIZE-1);
693                         dso->loadmap->segs[i].p_vaddr = ph->p_vaddr;
694                         dso->loadmap->segs[i].p_memsz = ph->p_memsz;
695                         i++;
696                         if (prot & PROT_WRITE) {
697                                 size_t brk = (ph->p_vaddr & PAGE_SIZE-1)
698                                         + ph->p_filesz;
699                                 size_t pgbrk = brk + PAGE_SIZE-1 & -PAGE_SIZE;
700                                 size_t pgend = brk + ph->p_memsz - ph->p_filesz
701                                         + PAGE_SIZE-1 & -PAGE_SIZE;
702                                 if (pgend > pgbrk && mmap_fixed(map+pgbrk,
703                                         pgend-pgbrk, prot,
704                                         MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS,
705                                         -1, off_start) == MAP_FAILED)
706                                         goto error;
707                                 memset(map + brk, 0, pgbrk-brk);
708                         }
709                 }
710                 map = (void *)dso->loadmap->segs[0].addr;
711                 map_len = 0;
712                 goto done_mapping;
713         }
714         addr_max += PAGE_SIZE-1;
715         addr_max &= -PAGE_SIZE;
716         addr_min &= -PAGE_SIZE;
717         off_start &= -PAGE_SIZE;
718         map_len = addr_max - addr_min + off_start;
719         /* The first time, we map too much, possibly even more than
720          * the length of the file. This is okay because we will not
721          * use the invalid part; we just need to reserve the right
722          * amount of virtual address space to map over later. */
723         map = DL_NOMMU_SUPPORT
724                 ? mmap((void *)addr_min, map_len, PROT_READ|PROT_WRITE|PROT_EXEC,
725                         MAP_PRIVATE|MAP_ANONYMOUS, -1, 0)
726                 : mmap((void *)addr_min, map_len, prot,
727                         MAP_PRIVATE, fd, off_start);
728         if (map==MAP_FAILED) goto error;
729         dso->map = map;
730         dso->map_len = map_len;
731         /* If the loaded file is not relocatable and the requested address is
732          * not available, then the load operation must fail. */
733         if (eh->e_type != ET_DYN && addr_min && map!=(void *)addr_min) {
734                 errno = EBUSY;
735                 goto error;
736         }
737         base = map - addr_min;
738         dso->phdr = 0;
739         dso->phnum = 0;
740         for (ph=ph0, i=eh->e_phnum; i; i--, ph=(void *)((char *)ph+eh->e_phentsize)) {
741                 if (ph->p_type != PT_LOAD) continue;
742                 /* Check if the programs headers are in this load segment, and
743                  * if so, record the address for use by dl_iterate_phdr. */
744                 if (!dso->phdr && eh->e_phoff >= ph->p_offset
745                     && eh->e_phoff+phsize <= ph->p_offset+ph->p_filesz) {
746                         dso->phdr = (void *)(base + ph->p_vaddr
747                                 + (eh->e_phoff-ph->p_offset));
748                         dso->phnum = eh->e_phnum;
749                         dso->phentsize = eh->e_phentsize;
750                 }
751                 this_min = ph->p_vaddr & -PAGE_SIZE;
752                 this_max = ph->p_vaddr+ph->p_memsz+PAGE_SIZE-1 & -PAGE_SIZE;
753                 off_start = ph->p_offset & -PAGE_SIZE;
754                 prot = (((ph->p_flags&PF_R) ? PROT_READ : 0) |
755                         ((ph->p_flags&PF_W) ? PROT_WRITE: 0) |
756                         ((ph->p_flags&PF_X) ? PROT_EXEC : 0));
757                 /* Reuse the existing mapping for the lowest-address LOAD */
758                 if ((ph->p_vaddr & -PAGE_SIZE) != addr_min || DL_NOMMU_SUPPORT)
759                         if (mmap_fixed(base+this_min, this_max-this_min, prot, MAP_PRIVATE|MAP_FIXED, fd, off_start) == MAP_FAILED)
760                                 goto error;
761                 if (ph->p_memsz > ph->p_filesz && (ph->p_flags&PF_W)) {
762                         size_t brk = (size_t)base+ph->p_vaddr+ph->p_filesz;
763                         size_t pgbrk = brk+PAGE_SIZE-1 & -PAGE_SIZE;
764                         memset((void *)brk, 0, pgbrk-brk & PAGE_SIZE-1);
765                         if (pgbrk-(size_t)base < this_max && mmap_fixed((void *)pgbrk, (size_t)base+this_max-pgbrk, prot, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) == MAP_FAILED)
766                                 goto error;
767                 }
768         }
769         for (i=0; ((size_t *)(base+dyn))[i]; i+=2)
770                 if (((size_t *)(base+dyn))[i]==DT_TEXTREL) {
771                         if (mprotect(map, map_len, PROT_READ|PROT_WRITE|PROT_EXEC)
772                             && errno != ENOSYS)
773                                 goto error;
774                         break;
775                 }
776 done_mapping:
777         dso->base = base;
778         dso->dynv = laddr(dso, dyn);
779         if (dso->tls.size) dso->tls.image = laddr(dso, tls_image);
780         free(allocated_buf);
781         return map;
782 noexec:
783         errno = ENOEXEC;
784 error:
785         if (map!=MAP_FAILED) unmap_library(dso);
786         free(allocated_buf);
787         return 0;
788 }
789
790 static int path_open(const char *name, const char *s, char *buf, size_t buf_size)
791 {
792         size_t l;
793         int fd;
794         for (;;) {
795                 s += strspn(s, ":\n");
796                 l = strcspn(s, ":\n");
797                 if (l-1 >= INT_MAX) return -1;
798                 if (snprintf(buf, buf_size, "%.*s/%s", (int)l, s, name) < buf_size) {
799                         if ((fd = open(buf, O_RDONLY|O_CLOEXEC))>=0) return fd;
800                         switch (errno) {
801                         case ENOENT:
802                         case ENOTDIR:
803                         case EACCES:
804                         case ENAMETOOLONG:
805                                 break;
806                         default:
807                                 /* Any negative value but -1 will inhibit
808                                  * futher path search. */
809                                 return -2;
810                         }
811                 }
812                 s += l;
813         }
814 }
815
816 static int fixup_rpath(struct dso *p, char *buf, size_t buf_size)
817 {
818         size_t n, l;
819         const char *s, *t, *origin;
820         char *d;
821         if (p->rpath || !p->rpath_orig) return 0;
822         if (!strchr(p->rpath_orig, '$')) {
823                 p->rpath = p->rpath_orig;
824                 return 0;
825         }
826         n = 0;
827         s = p->rpath_orig;
828         while ((t=strchr(s, '$'))) {
829                 if (strncmp(t, "$ORIGIN", 7) && strncmp(t, "${ORIGIN}", 9))
830                         return 0;
831                 s = t+1;
832                 n++;
833         }
834         if (n > SSIZE_MAX/PATH_MAX) return 0;
835
836         if (p->kernel_mapped) {
837                 /* $ORIGIN searches cannot be performed for the main program
838                  * when it is suid/sgid/AT_SECURE. This is because the
839                  * pathname is under the control of the caller of execve.
840                  * For libraries, however, $ORIGIN can be processed safely
841                  * since the library's pathname came from a trusted source
842                  * (either system paths or a call to dlopen). */
843                 if (libc.secure)
844                         return 0;
845                 l = readlink("/proc/self/exe", buf, buf_size);
846                 if (l == -1) switch (errno) {
847                 case ENOENT:
848                 case ENOTDIR:
849                 case EACCES:
850                         break;
851                 default:
852                         return -1;
853                 }
854                 if (l >= buf_size)
855                         return 0;
856                 buf[l] = 0;
857                 origin = buf;
858         } else {
859                 origin = p->name;
860         }
861         t = strrchr(origin, '/');
862         if (t) {
863                 l = t-origin;
864         } else {
865                 /* Normally p->name will always be an absolute or relative
866                  * pathname containing at least one '/' character, but in the
867                  * case where ldso was invoked as a command to execute a
868                  * program in the working directory, app.name may not. Fix. */
869                 origin = ".";
870                 l = 1;
871         }
872         /* Disallow non-absolute origins for suid/sgid/AT_SECURE. */
873         if (libc.secure && *origin != '/')
874                 return 0;
875         p->rpath = malloc(strlen(p->rpath_orig) + n*l + 1);
876         if (!p->rpath) return -1;
877
878         d = p->rpath;
879         s = p->rpath_orig;
880         while ((t=strchr(s, '$'))) {
881                 memcpy(d, s, t-s);
882                 d += t-s;
883                 memcpy(d, origin, l);
884                 d += l;
885                 /* It was determined previously that the '$' is followed
886                  * either by "ORIGIN" or "{ORIGIN}". */
887                 s = t + 7 + 2*(t[1]=='{');
888         }
889         strcpy(d, s);
890         return 0;
891 }
892
893 static void decode_dyn(struct dso *p)
894 {
895         size_t dyn[DYN_CNT];
896         decode_vec(p->dynv, dyn, DYN_CNT);
897         p->syms = laddr(p, dyn[DT_SYMTAB]);
898         p->strings = laddr(p, dyn[DT_STRTAB]);
899         if (dyn[0]&(1<<DT_HASH))
900                 p->hashtab = laddr(p, dyn[DT_HASH]);
901         if (dyn[0]&(1<<DT_RPATH))
902                 p->rpath_orig = p->strings + dyn[DT_RPATH];
903         if (dyn[0]&(1<<DT_RUNPATH))
904                 p->rpath_orig = p->strings + dyn[DT_RUNPATH];
905         if (dyn[0]&(1<<DT_PLTGOT))
906                 p->got = laddr(p, dyn[DT_PLTGOT]);
907         if (search_vec(p->dynv, dyn, DT_GNU_HASH))
908                 p->ghashtab = laddr(p, *dyn);
909         if (search_vec(p->dynv, dyn, DT_VERSYM))
910                 p->versym = laddr(p, *dyn);
911 }
912
913 static size_t count_syms(struct dso *p)
914 {
915         if (p->hashtab) return p->hashtab[1];
916
917         size_t nsym, i;
918         uint32_t *buckets = p->ghashtab + 4 + (p->ghashtab[2]*sizeof(size_t)/4);
919         uint32_t *hashval;
920         for (i = nsym = 0; i < p->ghashtab[0]; i++) {
921                 if (buckets[i] > nsym)
922                         nsym = buckets[i];
923         }
924         if (nsym) {
925                 hashval = buckets + p->ghashtab[0] + (nsym - p->ghashtab[1]);
926                 do nsym++;
927                 while (!(*hashval++ & 1));
928         }
929         return nsym;
930 }
931
932 static void *dl_mmap(size_t n)
933 {
934         void *p;
935         int prot = PROT_READ|PROT_WRITE, flags = MAP_ANONYMOUS|MAP_PRIVATE;
936 #ifdef SYS_mmap2
937         p = (void *)__syscall(SYS_mmap2, 0, n, prot, flags, -1, 0);
938 #else
939         p = (void *)__syscall(SYS_mmap, 0, n, prot, flags, -1, 0);
940 #endif
941         return (unsigned long)p > -4096UL ? 0 : p;
942 }
943
944 static void makefuncdescs(struct dso *p)
945 {
946         static int self_done;
947         size_t nsym = count_syms(p);
948         size_t i, size = nsym * sizeof(*p->funcdescs);
949
950         if (!self_done) {
951                 p->funcdescs = dl_mmap(size);
952                 self_done = 1;
953         } else {
954                 p->funcdescs = malloc(size);
955         }
956         if (!p->funcdescs) {
957                 if (!runtime) a_crash();
958                 error("Error allocating function descriptors for %s", p->name);
959                 longjmp(*rtld_fail, 1);
960         }
961         for (i=0; i<nsym; i++) {
962                 if ((p->syms[i].st_info&0xf)==STT_FUNC && p->syms[i].st_shndx) {
963                         p->funcdescs[i].addr = laddr(p, p->syms[i].st_value);
964                         p->funcdescs[i].got = p->got;
965                 } else {
966                         p->funcdescs[i].addr = 0;
967                         p->funcdescs[i].got = 0;
968                 }
969         }
970 }
971
972 static struct dso *load_library(const char *name, struct dso *needed_by)
973 {
974         char buf[2*NAME_MAX+2];
975         const char *pathname;
976         unsigned char *map;
977         struct dso *p, temp_dso = {0};
978         int fd;
979         struct stat st;
980         size_t alloc_size;
981         int n_th = 0;
982         int is_self = 0;
983
984         if (!*name) {
985                 errno = EINVAL;
986                 return 0;
987         }
988
989         /* Catch and block attempts to reload the implementation itself */
990         if (name[0]=='l' && name[1]=='i' && name[2]=='b') {
991                 static const char reserved[] =
992                         "c.pthread.rt.m.dl.util.xnet.";
993                 const char *rp, *next;
994                 for (rp=reserved; *rp; rp=next) {
995                         next = strchr(rp, '.') + 1;
996                         if (strncmp(name+3, rp, next-rp) == 0)
997                                 break;
998                 }
999                 if (*rp) {
1000                         if (ldd_mode) {
1001                                 /* Track which names have been resolved
1002                                  * and only report each one once. */
1003                                 static unsigned reported;
1004                                 unsigned mask = 1U<<(rp-reserved);
1005                                 if (!(reported & mask)) {
1006                                         reported |= mask;
1007                                         dprintf(1, "\t%s => %s (%p)\n",
1008                                                 name, ldso.name,
1009                                                 ldso.base);
1010                                 }
1011                         }
1012                         is_self = 1;
1013                 }
1014         }
1015         if (!strcmp(name, ldso.name)) is_self = 1;
1016         if (is_self) {
1017                 if (!ldso.prev) {
1018                         tail->next = &ldso;
1019                         ldso.prev = tail;
1020                         tail = &ldso;
1021                 }
1022                 return &ldso;
1023         }
1024         if (strchr(name, '/')) {
1025                 pathname = name;
1026                 fd = open(name, O_RDONLY|O_CLOEXEC);
1027         } else {
1028                 /* Search for the name to see if it's already loaded */
1029                 for (p=head->next; p; p=p->next) {
1030                         if (p->shortname && !strcmp(p->shortname, name)) {
1031                                 return p;
1032                         }
1033                 }
1034                 if (strlen(name) > NAME_MAX) return 0;
1035                 fd = -1;
1036                 if (env_path) fd = path_open(name, env_path, buf, sizeof buf);
1037                 for (p=needed_by; fd == -1 && p; p=p->needed_by) {
1038                         if (fixup_rpath(p, buf, sizeof buf) < 0)
1039                                 fd = -2; /* Inhibit further search. */
1040                         if (p->rpath)
1041                                 fd = path_open(name, p->rpath, buf, sizeof buf);
1042                 }
1043                 if (fd == -1) {
1044                         if (!sys_path) {
1045                                 char *prefix = 0;
1046                                 size_t prefix_len;
1047                                 if (ldso.name[0]=='/') {
1048                                         char *s, *t, *z;
1049                                         for (s=t=z=ldso.name; *s; s++)
1050                                                 if (*s=='/') z=t, t=s;
1051                                         prefix_len = z-ldso.name;
1052                                         if (prefix_len < PATH_MAX)
1053                                                 prefix = ldso.name;
1054                                 }
1055                                 if (!prefix) {
1056                                         prefix = "";
1057                                         prefix_len = 0;
1058                                 }
1059                                 char etc_ldso_path[prefix_len + 1
1060                                         + sizeof "/etc/ld-musl-" LDSO_ARCH ".path"];
1061                                 snprintf(etc_ldso_path, sizeof etc_ldso_path,
1062                                         "%.*s/etc/ld-musl-" LDSO_ARCH ".path",
1063                                         (int)prefix_len, prefix);
1064                                 FILE *f = fopen(etc_ldso_path, "rbe");
1065                                 if (f) {
1066                                         if (getdelim(&sys_path, (size_t[1]){0}, 0, f) <= 0) {
1067                                                 free(sys_path);
1068                                                 sys_path = "";
1069                                         }
1070                                         fclose(f);
1071                                 } else if (errno != ENOENT) {
1072                                         sys_path = "";
1073                                 }
1074                         }
1075                         if (!sys_path) sys_path = "/lib:/usr/local/lib:/usr/lib";
1076                         fd = path_open(name, sys_path, buf, sizeof buf);
1077                 }
1078                 pathname = buf;
1079         }
1080         if (fd < 0) return 0;
1081         if (fstat(fd, &st) < 0) {
1082                 close(fd);
1083                 return 0;
1084         }
1085         for (p=head->next; p; p=p->next) {
1086                 if (p->dev == st.st_dev && p->ino == st.st_ino) {
1087                         /* If this library was previously loaded with a
1088                          * pathname but a search found the same inode,
1089                          * setup its shortname so it can be found by name. */
1090                         if (!p->shortname && pathname != name)
1091                                 p->shortname = strrchr(p->name, '/')+1;
1092                         close(fd);
1093                         return p;
1094                 }
1095         }
1096         map = noload ? 0 : map_library(fd, &temp_dso);
1097         close(fd);
1098         if (!map) return 0;
1099
1100         /* Avoid the danger of getting two versions of libc mapped into the
1101          * same process when an absolute pathname was used. The symbols
1102          * checked are chosen to catch both musl and glibc, and to avoid
1103          * false positives from interposition-hack libraries. */
1104         decode_dyn(&temp_dso);
1105         if (find_sym(&temp_dso, "__libc_start_main", 1).sym &&
1106             find_sym(&temp_dso, "stdin", 1).sym) {
1107                 unmap_library(&temp_dso);
1108                 return load_library("libc.so", needed_by);
1109         }
1110         /* Past this point, if we haven't reached runtime yet, ldso has
1111          * committed either to use the mapped library or to abort execution.
1112          * Unmapping is not possible, so we can safely reclaim gaps. */
1113         if (!runtime) reclaim_gaps(&temp_dso);
1114
1115         /* Allocate storage for the new DSO. When there is TLS, this
1116          * storage must include a reservation for all pre-existing
1117          * threads to obtain copies of both the new TLS, and an
1118          * extended DTV capable of storing an additional slot for
1119          * the newly-loaded DSO. */
1120         alloc_size = sizeof *p + strlen(pathname) + 1;
1121         if (runtime && temp_dso.tls.image) {
1122                 size_t per_th = temp_dso.tls.size + temp_dso.tls.align
1123                         + sizeof(void *) * (tls_cnt+3);
1124                 n_th = libc.threads_minus_1 + 1;
1125                 if (n_th > SSIZE_MAX / per_th) alloc_size = SIZE_MAX;
1126                 else alloc_size += n_th * per_th;
1127         }
1128         p = calloc(1, alloc_size);
1129         if (!p) {
1130                 unmap_library(&temp_dso);
1131                 return 0;
1132         }
1133         memcpy(p, &temp_dso, sizeof temp_dso);
1134         p->dev = st.st_dev;
1135         p->ino = st.st_ino;
1136         p->needed_by = needed_by;
1137         p->name = p->buf;
1138         p->runtime_loaded = runtime;
1139         strcpy(p->name, pathname);
1140         /* Add a shortname only if name arg was not an explicit pathname. */
1141         if (pathname != name) p->shortname = strrchr(p->name, '/')+1;
1142         if (p->tls.image) {
1143                 p->tls_id = ++tls_cnt;
1144                 tls_align = MAXP2(tls_align, p->tls.align);
1145 #ifdef TLS_ABOVE_TP
1146                 p->tls.offset = tls_offset + ( (p->tls.align-1) &
1147                         (-tls_offset + (uintptr_t)p->tls.image) );
1148                 tls_offset = p->tls.offset + p->tls.size;
1149 #else
1150                 tls_offset += p->tls.size + p->tls.align - 1;
1151                 tls_offset -= (tls_offset + (uintptr_t)p->tls.image)
1152                         & (p->tls.align-1);
1153                 p->tls.offset = tls_offset;
1154 #endif
1155                 p->new_dtv = (void *)(-sizeof(size_t) &
1156                         (uintptr_t)(p->name+strlen(p->name)+sizeof(size_t)));
1157                 p->new_tls = (void *)(p->new_dtv + n_th*(tls_cnt+1));
1158                 if (tls_tail) tls_tail->next = &p->tls;
1159                 else libc.tls_head = &p->tls;
1160                 tls_tail = &p->tls;
1161         }
1162
1163         tail->next = p;
1164         p->prev = tail;
1165         tail = p;
1166
1167         if (DL_FDPIC) makefuncdescs(p);
1168
1169         if (ldd_mode) dprintf(1, "\t%s => %s (%p)\n", name, pathname, p->base);
1170
1171         return p;
1172 }
1173
1174 static void load_direct_deps(struct dso *p)
1175 {
1176         size_t i, cnt=0;
1177
1178         if (p->deps) return;
1179         /* For head, all preloads are direct pseudo-dependencies.
1180          * Count and include them now to avoid realloc later. */
1181         if (p==head) for (struct dso *q=p->next; q; q=q->next)
1182                 cnt++;
1183         for (i=0; p->dynv[i]; i+=2)
1184                 if (p->dynv[i] == DT_NEEDED) cnt++;
1185         /* Use builtin buffer for apps with no external deps, to
1186          * preserve property of no runtime failure paths. */
1187         p->deps = (p==head && cnt<2) ? builtin_deps :
1188                 calloc(cnt+1, sizeof *p->deps);
1189         if (!p->deps) {
1190                 error("Error loading dependencies for %s", p->name);
1191                 if (runtime) longjmp(*rtld_fail, 1);
1192         }
1193         cnt=0;
1194         if (p==head) for (struct dso *q=p->next; q; q=q->next)
1195                 p->deps[cnt++] = q;
1196         for (i=0; p->dynv[i]; i+=2) {
1197                 if (p->dynv[i] != DT_NEEDED) continue;
1198                 struct dso *dep = load_library(p->strings + p->dynv[i+1], p);
1199                 if (!dep) {
1200                         error("Error loading shared library %s: %m (needed by %s)",
1201                                 p->strings + p->dynv[i+1], p->name);
1202                         if (runtime) longjmp(*rtld_fail, 1);
1203                         continue;
1204                 }
1205                 p->deps[cnt++] = dep;
1206         }
1207         p->deps[cnt] = 0;
1208         p->ndeps_direct = cnt;
1209 }
1210
1211 static void load_deps(struct dso *p)
1212 {
1213         if (p->deps) return;
1214         for (; p; p=p->next)
1215                 load_direct_deps(p);
1216 }
1217
1218 static void extend_bfs_deps(struct dso *p)
1219 {
1220         size_t i, j, cnt, ndeps_all;
1221         struct dso **tmp;
1222
1223         /* Can't use realloc if the original p->deps was allocated at
1224          * program entry and malloc has been replaced, or if it's
1225          * the builtin non-allocated trivial main program deps array. */
1226         int no_realloc = (__malloc_replaced && !p->runtime_loaded)
1227                 || p->deps == builtin_deps;
1228
1229         if (p->bfs_built) return;
1230         ndeps_all = p->ndeps_direct;
1231
1232         /* Mark existing (direct) deps so they won't be duplicated. */
1233         for (i=0; p->deps[i]; i++)
1234                 p->deps[i]->mark = 1;
1235
1236         /* For each dependency already in the list, copy its list of direct
1237          * dependencies to the list, excluding any items already in the
1238          * list. Note that the list this loop iterates over will grow during
1239          * the loop, but since duplicates are excluded, growth is bounded. */
1240         for (i=0; p->deps[i]; i++) {
1241                 struct dso *dep = p->deps[i];
1242                 for (j=cnt=0; j<dep->ndeps_direct; j++)
1243                         if (!dep->deps[j]->mark) cnt++;
1244                 tmp = no_realloc ? 
1245                         malloc(sizeof(*tmp) * (ndeps_all+cnt+1)) :
1246                         realloc(p->deps, sizeof(*tmp) * (ndeps_all+cnt+1));
1247                 if (!tmp) {
1248                         error("Error recording dependencies for %s", p->name);
1249                         if (runtime) longjmp(*rtld_fail, 1);
1250                         continue;
1251                 }
1252                 if (no_realloc) {
1253                         memcpy(tmp, p->deps, sizeof(*tmp) * (ndeps_all+1));
1254                         no_realloc = 0;
1255                 }
1256                 p->deps = tmp;
1257                 for (j=0; j<dep->ndeps_direct; j++) {
1258                         if (dep->deps[j]->mark) continue;
1259                         dep->deps[j]->mark = 1;
1260                         p->deps[ndeps_all++] = dep->deps[j];
1261                 }
1262                 p->deps[ndeps_all] = 0;
1263         }
1264         p->bfs_built = 1;
1265         for (p=head; p; p=p->next)
1266                 p->mark = 0;
1267 }
1268
1269 static void load_preload(char *s)
1270 {
1271         int tmp;
1272         char *z;
1273         for (z=s; *z; s=z) {
1274                 for (   ; *s && (isspace(*s) || *s==':'); s++);
1275                 for (z=s; *z && !isspace(*z) && *z!=':'; z++);
1276                 tmp = *z;
1277                 *z = 0;
1278                 load_library(s, 0);
1279                 *z = tmp;
1280         }
1281 }
1282
1283 static void add_syms(struct dso *p)
1284 {
1285         if (!p->syms_next && syms_tail != p) {
1286                 syms_tail->syms_next = p;
1287                 syms_tail = p;
1288         }
1289 }
1290
1291 static void revert_syms(struct dso *old_tail)
1292 {
1293         struct dso *p, *next;
1294         /* Chop off the tail of the list of dsos that participate in
1295          * the global symbol table, reverting them to RTLD_LOCAL. */
1296         for (p=old_tail; p; p=next) {
1297                 next = p->syms_next;
1298                 p->syms_next = 0;
1299         }
1300         syms_tail = old_tail;
1301 }
1302
1303 static void do_mips_relocs(struct dso *p, size_t *got)
1304 {
1305         size_t i, j, rel[2];
1306         unsigned char *base = p->base;
1307         i=0; search_vec(p->dynv, &i, DT_MIPS_LOCAL_GOTNO);
1308         if (p==&ldso) {
1309                 got += i;
1310         } else {
1311                 while (i--) *got++ += (size_t)base;
1312         }
1313         j=0; search_vec(p->dynv, &j, DT_MIPS_GOTSYM);
1314         i=0; search_vec(p->dynv, &i, DT_MIPS_SYMTABNO);
1315         Sym *sym = p->syms + j;
1316         rel[0] = (unsigned char *)got - base;
1317         for (i-=j; i; i--, sym++, rel[0]+=sizeof(size_t)) {
1318                 rel[1] = R_INFO(sym-p->syms, R_MIPS_JUMP_SLOT);
1319                 do_relocs(p, rel, sizeof rel, 2);
1320         }
1321 }
1322
1323 static void reloc_all(struct dso *p)
1324 {
1325         size_t dyn[DYN_CNT];
1326         for (; p; p=p->next) {
1327                 if (p->relocated) continue;
1328                 decode_vec(p->dynv, dyn, DYN_CNT);
1329                 if (NEED_MIPS_GOT_RELOCS)
1330                         do_mips_relocs(p, laddr(p, dyn[DT_PLTGOT]));
1331                 do_relocs(p, laddr(p, dyn[DT_JMPREL]), dyn[DT_PLTRELSZ],
1332                         2+(dyn[DT_PLTREL]==DT_RELA));
1333                 do_relocs(p, laddr(p, dyn[DT_REL]), dyn[DT_RELSZ], 2);
1334                 do_relocs(p, laddr(p, dyn[DT_RELA]), dyn[DT_RELASZ], 3);
1335
1336                 if (head != &ldso && p->relro_start != p->relro_end &&
1337                     mprotect(laddr(p, p->relro_start), p->relro_end-p->relro_start, PROT_READ)
1338                     && errno != ENOSYS) {
1339                         error("Error relocating %s: RELRO protection failed: %m",
1340                                 p->name);
1341                         if (runtime) longjmp(*rtld_fail, 1);
1342                 }
1343
1344                 p->relocated = 1;
1345         }
1346 }
1347
1348 static void kernel_mapped_dso(struct dso *p)
1349 {
1350         size_t min_addr = -1, max_addr = 0, cnt;
1351         Phdr *ph = p->phdr;
1352         for (cnt = p->phnum; cnt--; ph = (void *)((char *)ph + p->phentsize)) {
1353                 if (ph->p_type == PT_DYNAMIC) {
1354                         p->dynv = laddr(p, ph->p_vaddr);
1355                 } else if (ph->p_type == PT_GNU_RELRO) {
1356                         p->relro_start = ph->p_vaddr & -PAGE_SIZE;
1357                         p->relro_end = (ph->p_vaddr + ph->p_memsz) & -PAGE_SIZE;
1358                 } else if (ph->p_type == PT_GNU_STACK) {
1359                         if (!runtime && ph->p_memsz > __default_stacksize) {
1360                                 __default_stacksize =
1361                                         ph->p_memsz < DEFAULT_STACK_MAX ?
1362                                         ph->p_memsz : DEFAULT_STACK_MAX;
1363                         }
1364                 }
1365                 if (ph->p_type != PT_LOAD) continue;
1366                 if (ph->p_vaddr < min_addr)
1367                         min_addr = ph->p_vaddr;
1368                 if (ph->p_vaddr+ph->p_memsz > max_addr)
1369                         max_addr = ph->p_vaddr+ph->p_memsz;
1370         }
1371         min_addr &= -PAGE_SIZE;
1372         max_addr = (max_addr + PAGE_SIZE-1) & -PAGE_SIZE;
1373         p->map = p->base + min_addr;
1374         p->map_len = max_addr - min_addr;
1375         p->kernel_mapped = 1;
1376 }
1377
1378 void __libc_exit_fini()
1379 {
1380         struct dso *p;
1381         size_t dyn[DYN_CNT];
1382         int self = __pthread_self()->tid;
1383
1384         /* Take both locks before setting shutting_down, so that
1385          * either lock is sufficient to read its value. The lock
1386          * order matches that in dlopen to avoid deadlock. */
1387         pthread_rwlock_wrlock(&lock);
1388         pthread_mutex_lock(&init_fini_lock);
1389         shutting_down = 1;
1390         pthread_rwlock_unlock(&lock);
1391         for (p=fini_head; p; p=p->fini_next) {
1392                 while (p->ctor_visitor && p->ctor_visitor!=self)
1393                         pthread_cond_wait(&ctor_cond, &init_fini_lock);
1394                 if (!p->constructed) continue;
1395                 decode_vec(p->dynv, dyn, DYN_CNT);
1396                 if (dyn[0] & (1<<DT_FINI_ARRAY)) {
1397                         size_t n = dyn[DT_FINI_ARRAYSZ]/sizeof(size_t);
1398                         size_t *fn = (size_t *)laddr(p, dyn[DT_FINI_ARRAY])+n;
1399                         while (n--) ((void (*)(void))*--fn)();
1400                 }
1401 #ifndef NO_LEGACY_INITFINI
1402                 if ((dyn[0] & (1<<DT_FINI)) && dyn[DT_FINI])
1403                         fpaddr(p, dyn[DT_FINI])();
1404 #endif
1405         }
1406 }
1407
1408 static struct dso **queue_ctors(struct dso *dso)
1409 {
1410         size_t cnt, qpos, spos, i;
1411         struct dso *p, **queue, **stack;
1412
1413         if (ldd_mode) return 0;
1414
1415         /* Bound on queue size is the total number of indirect deps.
1416          * If a bfs deps list was built, we can use it. Otherwise,
1417          * bound by the total number of DSOs, which is always safe and
1418          * is reasonable we use it (for main app at startup). */
1419         if (dso->bfs_built) {
1420                 for (cnt=0; dso->deps[cnt]; cnt++)
1421                         dso->deps[cnt]->mark = 0;
1422                 cnt++; /* self, not included in deps */
1423         } else {
1424                 for (cnt=0, p=head; p; cnt++, p=p->next)
1425                         p->mark = 0;
1426         }
1427         cnt++; /* termination slot */
1428         if (dso==head && cnt <= countof(builtin_ctor_queue))
1429                 queue = builtin_ctor_queue;
1430         else
1431                 queue = calloc(cnt, sizeof *queue);
1432
1433         if (!queue) {
1434                 error("Error allocating constructor queue: %m\n");
1435                 if (runtime) longjmp(*rtld_fail, 1);
1436                 return 0;
1437         }
1438
1439         /* Opposite ends of the allocated buffer serve as an output queue
1440          * and a working stack. Setup initial stack with just the argument
1441          * dso and initial queue empty... */
1442         stack = queue;
1443         qpos = 0;
1444         spos = cnt;
1445         stack[--spos] = dso;
1446         dso->next_dep = 0;
1447         dso->mark = 1;
1448
1449         /* Then perform pseudo-DFS sort, but ignoring circular deps. */
1450         while (spos<cnt) {
1451                 p = stack[spos++];
1452                 while (p->next_dep < p->ndeps_direct) {
1453                         if (p->deps[p->next_dep]->mark) {
1454                                 p->next_dep++;
1455                         } else {
1456                                 stack[--spos] = p;
1457                                 p = p->deps[p->next_dep];
1458                                 p->next_dep = 0;
1459                                 p->mark = 1;
1460                         }
1461                 }
1462                 queue[qpos++] = p;
1463         }
1464         queue[qpos] = 0;
1465         for (i=0; i<qpos; i++) queue[i]->mark = 0;
1466
1467         return queue;
1468 }
1469
1470 static void do_init_fini(struct dso **queue)
1471 {
1472         struct dso *p;
1473         size_t dyn[DYN_CNT], i;
1474         int self = __pthread_self()->tid;
1475
1476         pthread_mutex_lock(&init_fini_lock);
1477         for (i=0; (p=queue[i]); i++) {
1478                 while ((p->ctor_visitor && p->ctor_visitor!=self) || shutting_down)
1479                         pthread_cond_wait(&ctor_cond, &init_fini_lock);
1480                 if (p->ctor_visitor || p->constructed)
1481                         continue;
1482                 p->ctor_visitor = self;
1483                 
1484                 decode_vec(p->dynv, dyn, DYN_CNT);
1485                 if (dyn[0] & ((1<<DT_FINI) | (1<<DT_FINI_ARRAY))) {
1486                         p->fini_next = fini_head;
1487                         fini_head = p;
1488                 }
1489
1490                 pthread_mutex_unlock(&init_fini_lock);
1491
1492 #ifndef NO_LEGACY_INITFINI
1493                 if ((dyn[0] & (1<<DT_INIT)) && dyn[DT_INIT])
1494                         fpaddr(p, dyn[DT_INIT])();
1495 #endif
1496                 if (dyn[0] & (1<<DT_INIT_ARRAY)) {
1497                         size_t n = dyn[DT_INIT_ARRAYSZ]/sizeof(size_t);
1498                         size_t *fn = laddr(p, dyn[DT_INIT_ARRAY]);
1499                         while (n--) ((void (*)(void))*fn++)();
1500                 }
1501
1502                 pthread_mutex_lock(&init_fini_lock);
1503                 p->ctor_visitor = 0;
1504                 p->constructed = 1;
1505                 pthread_cond_broadcast(&ctor_cond);
1506         }
1507         pthread_mutex_unlock(&init_fini_lock);
1508 }
1509
1510 void __libc_start_init(void)
1511 {
1512         do_init_fini(main_ctor_queue);
1513         if (!__malloc_replaced && main_ctor_queue != builtin_ctor_queue)
1514                 free(main_ctor_queue);
1515         main_ctor_queue = 0;
1516 }
1517
1518 static void dl_debug_state(void)
1519 {
1520 }
1521
1522 weak_alias(dl_debug_state, _dl_debug_state);
1523
1524 void __init_tls(size_t *auxv)
1525 {
1526 }
1527
1528 static void update_tls_size()
1529 {
1530         libc.tls_cnt = tls_cnt;
1531         libc.tls_align = tls_align;
1532         libc.tls_size = ALIGN(
1533                 (1+tls_cnt) * sizeof(void *) +
1534                 tls_offset +
1535                 sizeof(struct pthread) +
1536                 tls_align * 2,
1537         tls_align);
1538 }
1539
1540 static void install_new_tls(void)
1541 {
1542         sigset_t set;
1543         pthread_t self = __pthread_self(), td;
1544         struct dso *dtv_provider = container_of(tls_tail, struct dso, tls);
1545         uintptr_t (*newdtv)[tls_cnt+1] = (void *)dtv_provider->new_dtv;
1546         struct dso *p;
1547         size_t i, j;
1548         size_t old_cnt = self->dtv[0];
1549
1550         __block_app_sigs(&set);
1551         __tl_lock();
1552         /* Copy existing dtv contents from all existing threads. */
1553         for (i=0, td=self; !i || td!=self; i++, td=td->next) {
1554                 memcpy(newdtv+i, td->dtv,
1555                         (old_cnt+1)*sizeof(uintptr_t));
1556                 newdtv[i][0] = tls_cnt;
1557         }
1558         /* Install new dtls into the enlarged, uninstalled dtv copies. */
1559         for (p=head; ; p=p->next) {
1560                 if (p->tls_id <= old_cnt) continue;
1561                 unsigned char *mem = p->new_tls;
1562                 for (j=0; j<i; j++) {
1563                         unsigned char *new = mem;
1564                         new += ((uintptr_t)p->tls.image - (uintptr_t)mem)
1565                                 & (p->tls.align-1);
1566                         memcpy(new, p->tls.image, p->tls.len);
1567                         newdtv[j][p->tls_id] =
1568                                 (uintptr_t)new + DTP_OFFSET;
1569                         mem += p->tls.size + p->tls.align;
1570                 }
1571                 if (p->tls_id == tls_cnt) break;
1572         }
1573
1574         /* Broadcast barrier to ensure contents of new dtv is visible
1575          * if the new dtv pointer is. The __membarrier function has a
1576          * fallback emulation using signals for kernels that lack the
1577          * feature at the syscall level. */
1578
1579         __membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0);
1580
1581         /* Install new dtv for each thread. */
1582         for (j=0, td=self; !j || td!=self; j++, td=td->next) {
1583                 td->dtv = td->dtv_copy = newdtv[j];
1584         }
1585
1586         __tl_unlock();
1587         __restore_sigs(&set);
1588 }
1589
1590 /* Stage 1 of the dynamic linker is defined in dlstart.c. It calls the
1591  * following stage 2 and stage 3 functions via primitive symbolic lookup
1592  * since it does not have access to their addresses to begin with. */
1593
1594 /* Stage 2 of the dynamic linker is called after relative relocations 
1595  * have been processed. It can make function calls to static functions
1596  * and access string literals and static data, but cannot use extern
1597  * symbols. Its job is to perform symbolic relocations on the dynamic
1598  * linker itself, but some of the relocations performed may need to be
1599  * replaced later due to copy relocations in the main program. */
1600
1601 hidden void __dls2(unsigned char *base, size_t *sp)
1602 {
1603         size_t *auxv;
1604         for (auxv=sp+1+*sp+1; *auxv; auxv++);
1605         auxv++;
1606         if (DL_FDPIC) {
1607                 void *p1 = (void *)sp[-2];
1608                 void *p2 = (void *)sp[-1];
1609                 if (!p1) {
1610                         size_t aux[AUX_CNT];
1611                         decode_vec(auxv, aux, AUX_CNT);
1612                         if (aux[AT_BASE]) ldso.base = (void *)aux[AT_BASE];
1613                         else ldso.base = (void *)(aux[AT_PHDR] & -4096);
1614                 }
1615                 app_loadmap = p2 ? p1 : 0;
1616                 ldso.loadmap = p2 ? p2 : p1;
1617                 ldso.base = laddr(&ldso, 0);
1618         } else {
1619                 ldso.base = base;
1620         }
1621         Ehdr *ehdr = (void *)ldso.base;
1622         ldso.name = ldso.shortname = "libc.so";
1623         ldso.phnum = ehdr->e_phnum;
1624         ldso.phdr = laddr(&ldso, ehdr->e_phoff);
1625         ldso.phentsize = ehdr->e_phentsize;
1626         kernel_mapped_dso(&ldso);
1627         decode_dyn(&ldso);
1628
1629         if (DL_FDPIC) makefuncdescs(&ldso);
1630
1631         /* Prepare storage for to save clobbered REL addends so they
1632          * can be reused in stage 3. There should be very few. If
1633          * something goes wrong and there are a huge number, abort
1634          * instead of risking stack overflow. */
1635         size_t dyn[DYN_CNT];
1636         decode_vec(ldso.dynv, dyn, DYN_CNT);
1637         size_t *rel = laddr(&ldso, dyn[DT_REL]);
1638         size_t rel_size = dyn[DT_RELSZ];
1639         size_t symbolic_rel_cnt = 0;
1640         apply_addends_to = rel;
1641         for (; rel_size; rel+=2, rel_size-=2*sizeof(size_t))
1642                 if (!IS_RELATIVE(rel[1], ldso.syms)) symbolic_rel_cnt++;
1643         if (symbolic_rel_cnt >= ADDEND_LIMIT) a_crash();
1644         size_t addends[symbolic_rel_cnt+1];
1645         saved_addends = addends;
1646
1647         head = &ldso;
1648         reloc_all(&ldso);
1649
1650         ldso.relocated = 0;
1651
1652         /* Call dynamic linker stage-2b, __dls2b, looking it up
1653          * symbolically as a barrier against moving the address
1654          * load across the above relocation processing. */
1655         struct symdef dls2b_def = find_sym(&ldso, "__dls2b", 0);
1656         if (DL_FDPIC) ((stage3_func)&ldso.funcdescs[dls2b_def.sym-ldso.syms])(sp, auxv);
1657         else ((stage3_func)laddr(&ldso, dls2b_def.sym->st_value))(sp, auxv);
1658 }
1659
1660 /* Stage 2b sets up a valid thread pointer, which requires relocations
1661  * completed in stage 2, and on which stage 3 is permitted to depend.
1662  * This is done as a separate stage, with symbolic lookup as a barrier,
1663  * so that loads of the thread pointer and &errno can be pure/const and
1664  * thereby hoistable. */
1665
1666 void __dls2b(size_t *sp, size_t *auxv)
1667 {
1668         /* Setup early thread pointer in builtin_tls for ldso/libc itself to
1669          * use during dynamic linking. If possible it will also serve as the
1670          * thread pointer at runtime. */
1671         search_vec(auxv, &__hwcap, AT_HWCAP);
1672         libc.auxv = auxv;
1673         libc.tls_size = sizeof builtin_tls;
1674         libc.tls_align = tls_align;
1675         if (__init_tp(__copy_tls((void *)builtin_tls)) < 0) {
1676                 a_crash();
1677         }
1678
1679         struct symdef dls3_def = find_sym(&ldso, "__dls3", 0);
1680         if (DL_FDPIC) ((stage3_func)&ldso.funcdescs[dls3_def.sym-ldso.syms])(sp, auxv);
1681         else ((stage3_func)laddr(&ldso, dls3_def.sym->st_value))(sp, auxv);
1682 }
1683
1684 /* Stage 3 of the dynamic linker is called with the dynamic linker/libc
1685  * fully functional. Its job is to load (if not already loaded) and
1686  * process dependencies and relocations for the main application and
1687  * transfer control to its entry point. */
1688
1689 void __dls3(size_t *sp, size_t *auxv)
1690 {
1691         static struct dso app, vdso;
1692         size_t aux[AUX_CNT];
1693         size_t i;
1694         char *env_preload=0;
1695         char *replace_argv0=0;
1696         size_t vdso_base;
1697         int argc = *sp;
1698         char **argv = (void *)(sp+1);
1699         char **argv_orig = argv;
1700         char **envp = argv+argc+1;
1701
1702         /* Find aux vector just past environ[] and use it to initialize
1703          * global data that may be needed before we can make syscalls. */
1704         __environ = envp;
1705         decode_vec(auxv, aux, AUX_CNT);
1706         search_vec(auxv, &__sysinfo, AT_SYSINFO);
1707         __pthread_self()->sysinfo = __sysinfo;
1708         libc.page_size = aux[AT_PAGESZ];
1709         libc.secure = ((aux[0]&0x7800)!=0x7800 || aux[AT_UID]!=aux[AT_EUID]
1710                 || aux[AT_GID]!=aux[AT_EGID] || aux[AT_SECURE]);
1711
1712         /* Only trust user/env if kernel says we're not suid/sgid */
1713         if (!libc.secure) {
1714                 env_path = getenv("LD_LIBRARY_PATH");
1715                 env_preload = getenv("LD_PRELOAD");
1716         }
1717
1718         /* If the main program was already loaded by the kernel,
1719          * AT_PHDR will point to some location other than the dynamic
1720          * linker's program headers. */
1721         if (aux[AT_PHDR] != (size_t)ldso.phdr) {
1722                 size_t interp_off = 0;
1723                 size_t tls_image = 0;
1724                 /* Find load address of the main program, via AT_PHDR vs PT_PHDR. */
1725                 Phdr *phdr = app.phdr = (void *)aux[AT_PHDR];
1726                 app.phnum = aux[AT_PHNUM];
1727                 app.phentsize = aux[AT_PHENT];
1728                 for (i=aux[AT_PHNUM]; i; i--, phdr=(void *)((char *)phdr + aux[AT_PHENT])) {
1729                         if (phdr->p_type == PT_PHDR)
1730                                 app.base = (void *)(aux[AT_PHDR] - phdr->p_vaddr);
1731                         else if (phdr->p_type == PT_INTERP)
1732                                 interp_off = (size_t)phdr->p_vaddr;
1733                         else if (phdr->p_type == PT_TLS) {
1734                                 tls_image = phdr->p_vaddr;
1735                                 app.tls.len = phdr->p_filesz;
1736                                 app.tls.size = phdr->p_memsz;
1737                                 app.tls.align = phdr->p_align;
1738                         }
1739                 }
1740                 if (DL_FDPIC) app.loadmap = app_loadmap;
1741                 if (app.tls.size) app.tls.image = laddr(&app, tls_image);
1742                 if (interp_off) ldso.name = laddr(&app, interp_off);
1743                 if ((aux[0] & (1UL<<AT_EXECFN))
1744                     && strncmp((char *)aux[AT_EXECFN], "/proc/", 6))
1745                         app.name = (char *)aux[AT_EXECFN];
1746                 else
1747                         app.name = argv[0];
1748                 kernel_mapped_dso(&app);
1749         } else {
1750                 int fd;
1751                 char *ldname = argv[0];
1752                 size_t l = strlen(ldname);
1753                 if (l >= 3 && !strcmp(ldname+l-3, "ldd")) ldd_mode = 1;
1754                 argv++;
1755                 while (argv[0] && argv[0][0]=='-' && argv[0][1]=='-') {
1756                         char *opt = argv[0]+2;
1757                         *argv++ = (void *)-1;
1758                         if (!*opt) {
1759                                 break;
1760                         } else if (!memcmp(opt, "list", 5)) {
1761                                 ldd_mode = 1;
1762                         } else if (!memcmp(opt, "library-path", 12)) {
1763                                 if (opt[12]=='=') env_path = opt+13;
1764                                 else if (opt[12]) *argv = 0;
1765                                 else if (*argv) env_path = *argv++;
1766                         } else if (!memcmp(opt, "preload", 7)) {
1767                                 if (opt[7]=='=') env_preload = opt+8;
1768                                 else if (opt[7]) *argv = 0;
1769                                 else if (*argv) env_preload = *argv++;
1770                         } else if (!memcmp(opt, "argv0", 5)) {
1771                                 if (opt[5]=='=') replace_argv0 = opt+6;
1772                                 else if (opt[5]) *argv = 0;
1773                                 else if (*argv) replace_argv0 = *argv++;
1774                         } else {
1775                                 argv[0] = 0;
1776                         }
1777                 }
1778                 argv[-1] = (void *)(argc - (argv-argv_orig));
1779                 if (!argv[0]) {
1780                         dprintf(2, "musl libc (" LDSO_ARCH ")\n"
1781                                 "Version %s\n"
1782                                 "Dynamic Program Loader\n"
1783                                 "Usage: %s [options] [--] pathname%s\n",
1784                                 __libc_version, ldname,
1785                                 ldd_mode ? "" : " [args]");
1786                         _exit(1);
1787                 }
1788                 fd = open(argv[0], O_RDONLY);
1789                 if (fd < 0) {
1790                         dprintf(2, "%s: cannot load %s: %s\n", ldname, argv[0], strerror(errno));
1791                         _exit(1);
1792                 }
1793                 Ehdr *ehdr = (void *)map_library(fd, &app);
1794                 if (!ehdr) {
1795                         dprintf(2, "%s: %s: Not a valid dynamic program\n", ldname, argv[0]);
1796                         _exit(1);
1797                 }
1798                 close(fd);
1799                 ldso.name = ldname;
1800                 app.name = argv[0];
1801                 aux[AT_ENTRY] = (size_t)laddr(&app, ehdr->e_entry);
1802                 /* Find the name that would have been used for the dynamic
1803                  * linker had ldd not taken its place. */
1804                 if (ldd_mode) {
1805                         for (i=0; i<app.phnum; i++) {
1806                                 if (app.phdr[i].p_type == PT_INTERP)
1807                                         ldso.name = laddr(&app, app.phdr[i].p_vaddr);
1808                         }
1809                         dprintf(1, "\t%s (%p)\n", ldso.name, ldso.base);
1810                 }
1811         }
1812         if (app.tls.size) {
1813                 libc.tls_head = tls_tail = &app.tls;
1814                 app.tls_id = tls_cnt = 1;
1815 #ifdef TLS_ABOVE_TP
1816                 app.tls.offset = GAP_ABOVE_TP;
1817                 app.tls.offset += (-GAP_ABOVE_TP + (uintptr_t)app.tls.image)
1818                         & (app.tls.align-1);
1819                 tls_offset = app.tls.offset + app.tls.size;
1820 #else
1821                 tls_offset = app.tls.offset = app.tls.size
1822                         + ( -((uintptr_t)app.tls.image + app.tls.size)
1823                         & (app.tls.align-1) );
1824 #endif
1825                 tls_align = MAXP2(tls_align, app.tls.align);
1826         }
1827         decode_dyn(&app);
1828         if (DL_FDPIC) {
1829                 makefuncdescs(&app);
1830                 if (!app.loadmap) {
1831                         app.loadmap = (void *)&app_dummy_loadmap;
1832                         app.loadmap->nsegs = 1;
1833                         app.loadmap->segs[0].addr = (size_t)app.map;
1834                         app.loadmap->segs[0].p_vaddr = (size_t)app.map
1835                                 - (size_t)app.base;
1836                         app.loadmap->segs[0].p_memsz = app.map_len;
1837                 }
1838                 argv[-3] = (void *)app.loadmap;
1839         }
1840
1841         /* Initial dso chain consists only of the app. */
1842         head = tail = syms_tail = &app;
1843
1844         /* Donate unused parts of app and library mapping to malloc */
1845         reclaim_gaps(&app);
1846         reclaim_gaps(&ldso);
1847
1848         /* Load preload/needed libraries, add symbols to global namespace. */
1849         ldso.deps = (struct dso **)no_deps;
1850         if (env_preload) load_preload(env_preload);
1851         load_deps(&app);
1852         for (struct dso *p=head; p; p=p->next)
1853                 add_syms(p);
1854
1855         /* Attach to vdso, if provided by the kernel, last so that it does
1856          * not become part of the global namespace.  */
1857         if (search_vec(auxv, &vdso_base, AT_SYSINFO_EHDR) && vdso_base) {
1858                 Ehdr *ehdr = (void *)vdso_base;
1859                 Phdr *phdr = vdso.phdr = (void *)(vdso_base + ehdr->e_phoff);
1860                 vdso.phnum = ehdr->e_phnum;
1861                 vdso.phentsize = ehdr->e_phentsize;
1862                 for (i=ehdr->e_phnum; i; i--, phdr=(void *)((char *)phdr + ehdr->e_phentsize)) {
1863                         if (phdr->p_type == PT_DYNAMIC)
1864                                 vdso.dynv = (void *)(vdso_base + phdr->p_offset);
1865                         if (phdr->p_type == PT_LOAD)
1866                                 vdso.base = (void *)(vdso_base - phdr->p_vaddr + phdr->p_offset);
1867                 }
1868                 vdso.name = "";
1869                 vdso.shortname = "linux-gate.so.1";
1870                 vdso.relocated = 1;
1871                 vdso.deps = (struct dso **)no_deps;
1872                 decode_dyn(&vdso);
1873                 vdso.prev = tail;
1874                 tail->next = &vdso;
1875                 tail = &vdso;
1876         }
1877
1878         for (i=0; app.dynv[i]; i+=2) {
1879                 if (!DT_DEBUG_INDIRECT && app.dynv[i]==DT_DEBUG)
1880                         app.dynv[i+1] = (size_t)&debug;
1881                 if (DT_DEBUG_INDIRECT && app.dynv[i]==DT_DEBUG_INDIRECT) {
1882                         size_t *ptr = (size_t *) app.dynv[i+1];
1883                         *ptr = (size_t)&debug;
1884                 }
1885         }
1886
1887         /* This must be done before final relocations, since it calls
1888          * malloc, which may be provided by the application. Calling any
1889          * application code prior to the jump to its entry point is not
1890          * valid in our model and does not work with FDPIC, where there
1891          * are additional relocation-like fixups that only the entry point
1892          * code can see to perform. */
1893         main_ctor_queue = queue_ctors(&app);
1894
1895         /* Initial TLS must also be allocated before final relocations
1896          * might result in calloc being a call to application code. */
1897         update_tls_size();
1898         void *initial_tls = builtin_tls;
1899         if (libc.tls_size > sizeof builtin_tls || tls_align > MIN_TLS_ALIGN) {
1900                 initial_tls = calloc(libc.tls_size, 1);
1901                 if (!initial_tls) {
1902                         dprintf(2, "%s: Error getting %zu bytes thread-local storage: %m\n",
1903                                 argv[0], libc.tls_size);
1904                         _exit(127);
1905                 }
1906         }
1907         static_tls_cnt = tls_cnt;
1908
1909         /* The main program must be relocated LAST since it may contain
1910          * copy relocations which depend on libraries' relocations. */
1911         reloc_all(app.next);
1912         reloc_all(&app);
1913
1914         /* Actual copying to new TLS needs to happen after relocations,
1915          * since the TLS images might have contained relocated addresses. */
1916         if (initial_tls != builtin_tls) {
1917                 if (__init_tp(__copy_tls(initial_tls)) < 0) {
1918                         a_crash();
1919                 }
1920         } else {
1921                 size_t tmp_tls_size = libc.tls_size;
1922                 pthread_t self = __pthread_self();
1923                 /* Temporarily set the tls size to the full size of
1924                  * builtin_tls so that __copy_tls will use the same layout
1925                  * as it did for before. Then check, just to be safe. */
1926                 libc.tls_size = sizeof builtin_tls;
1927                 if (__copy_tls((void*)builtin_tls) != self) a_crash();
1928                 libc.tls_size = tmp_tls_size;
1929         }
1930
1931         if (ldso_fail) _exit(127);
1932         if (ldd_mode) _exit(0);
1933
1934         /* Determine if malloc was interposed by a replacement implementation
1935          * so that calloc and the memalign family can harden against the
1936          * possibility of incomplete replacement. */
1937         if (find_sym(head, "malloc", 1).dso != &ldso)
1938                 __malloc_replaced = 1;
1939
1940         /* Switch to runtime mode: any further failures in the dynamic
1941          * linker are a reportable failure rather than a fatal startup
1942          * error. */
1943         runtime = 1;
1944
1945         debug.ver = 1;
1946         debug.bp = dl_debug_state;
1947         debug.head = head;
1948         debug.base = ldso.base;
1949         debug.state = 0;
1950         _dl_debug_state();
1951
1952         if (replace_argv0) argv[0] = replace_argv0;
1953
1954         errno = 0;
1955
1956         CRTJMP((void *)aux[AT_ENTRY], argv-1);
1957         for(;;);
1958 }
1959
1960 static void prepare_lazy(struct dso *p)
1961 {
1962         size_t dyn[DYN_CNT], n, flags1=0;
1963         decode_vec(p->dynv, dyn, DYN_CNT);
1964         search_vec(p->dynv, &flags1, DT_FLAGS_1);
1965         if (dyn[DT_BIND_NOW] || (dyn[DT_FLAGS] & DF_BIND_NOW) || (flags1 & DF_1_NOW))
1966                 return;
1967         n = dyn[DT_RELSZ]/2 + dyn[DT_RELASZ]/3 + dyn[DT_PLTRELSZ]/2 + 1;
1968         if (NEED_MIPS_GOT_RELOCS) {
1969                 size_t j=0; search_vec(p->dynv, &j, DT_MIPS_GOTSYM);
1970                 size_t i=0; search_vec(p->dynv, &i, DT_MIPS_SYMTABNO);
1971                 n += i-j;
1972         }
1973         p->lazy = calloc(n, 3*sizeof(size_t));
1974         if (!p->lazy) {
1975                 error("Error preparing lazy relocation for %s: %m", p->name);
1976                 longjmp(*rtld_fail, 1);
1977         }
1978         p->lazy_next = lazy_head;
1979         lazy_head = p;
1980 }
1981
1982 void *dlopen(const char *file, int mode)
1983 {
1984         struct dso *volatile p, *orig_tail, *orig_syms_tail, *orig_lazy_head, *next;
1985         struct tls_module *orig_tls_tail;
1986         size_t orig_tls_cnt, orig_tls_offset, orig_tls_align;
1987         size_t i;
1988         int cs;
1989         jmp_buf jb;
1990         struct dso **volatile ctor_queue = 0;
1991
1992         if (!file) return head;
1993
1994         pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cs);
1995         pthread_rwlock_wrlock(&lock);
1996         __inhibit_ptc();
1997
1998         p = 0;
1999         if (shutting_down) {
2000                 error("Cannot dlopen while program is exiting.");
2001                 goto end;
2002         }
2003         orig_tls_tail = tls_tail;
2004         orig_tls_cnt = tls_cnt;
2005         orig_tls_offset = tls_offset;
2006         orig_tls_align = tls_align;
2007         orig_lazy_head = lazy_head;
2008         orig_syms_tail = syms_tail;
2009         orig_tail = tail;
2010         noload = mode & RTLD_NOLOAD;
2011
2012         rtld_fail = &jb;
2013         if (setjmp(*rtld_fail)) {
2014                 /* Clean up anything new that was (partially) loaded */
2015                 revert_syms(orig_syms_tail);
2016                 for (p=orig_tail->next; p; p=next) {
2017                         next = p->next;
2018                         while (p->td_index) {
2019                                 void *tmp = p->td_index->next;
2020                                 free(p->td_index);
2021                                 p->td_index = tmp;
2022                         }
2023                         free(p->funcdescs);
2024                         if (p->rpath != p->rpath_orig)
2025                                 free(p->rpath);
2026                         free(p->deps);
2027                         unmap_library(p);
2028                         free(p);
2029                 }
2030                 free(ctor_queue);
2031                 ctor_queue = 0;
2032                 if (!orig_tls_tail) libc.tls_head = 0;
2033                 tls_tail = orig_tls_tail;
2034                 if (tls_tail) tls_tail->next = 0;
2035                 tls_cnt = orig_tls_cnt;
2036                 tls_offset = orig_tls_offset;
2037                 tls_align = orig_tls_align;
2038                 lazy_head = orig_lazy_head;
2039                 tail = orig_tail;
2040                 tail->next = 0;
2041                 p = 0;
2042                 goto end;
2043         } else p = load_library(file, head);
2044
2045         if (!p) {
2046                 error(noload ?
2047                         "Library %s is not already loaded" :
2048                         "Error loading shared library %s: %m",
2049                         file);
2050                 goto end;
2051         }
2052
2053         /* First load handling */
2054         load_deps(p);
2055         extend_bfs_deps(p);
2056         pthread_mutex_lock(&init_fini_lock);
2057         if (!p->constructed) ctor_queue = queue_ctors(p);
2058         pthread_mutex_unlock(&init_fini_lock);
2059         if (!p->relocated && (mode & RTLD_LAZY)) {
2060                 prepare_lazy(p);
2061                 for (i=0; p->deps[i]; i++)
2062                         if (!p->deps[i]->relocated)
2063                                 prepare_lazy(p->deps[i]);
2064         }
2065         if (!p->relocated || (mode & RTLD_GLOBAL)) {
2066                 /* Make new symbols global, at least temporarily, so we can do
2067                  * relocations. If not RTLD_GLOBAL, this is reverted below. */
2068                 add_syms(p);
2069                 for (i=0; p->deps[i]; i++)
2070                         add_syms(p->deps[i]);
2071         }
2072         if (!p->relocated) {
2073                 reloc_all(p);
2074         }
2075
2076         /* If RTLD_GLOBAL was not specified, undo any new additions
2077          * to the global symbol table. This is a nop if the library was
2078          * previously loaded and already global. */
2079         if (!(mode & RTLD_GLOBAL))
2080                 revert_syms(orig_syms_tail);
2081
2082         /* Processing of deferred lazy relocations must not happen until
2083          * the new libraries are committed; otherwise we could end up with
2084          * relocations resolved to symbol definitions that get removed. */
2085         redo_lazy_relocs();
2086
2087         update_tls_size();
2088         if (tls_cnt != orig_tls_cnt)
2089                 install_new_tls();
2090         _dl_debug_state();
2091         orig_tail = tail;
2092 end:
2093         __release_ptc();
2094         if (p) gencnt++;
2095         pthread_rwlock_unlock(&lock);
2096         if (ctor_queue) {
2097                 do_init_fini(ctor_queue);
2098                 free(ctor_queue);
2099         }
2100         pthread_setcancelstate(cs, 0);
2101         return p;
2102 }
2103
2104 hidden int __dl_invalid_handle(void *h)
2105 {
2106         struct dso *p;
2107         for (p=head; p; p=p->next) if (h==p) return 0;
2108         error("Invalid library handle %p", (void *)h);
2109         return 1;
2110 }
2111
2112 static void *addr2dso(size_t a)
2113 {
2114         struct dso *p;
2115         size_t i;
2116         if (DL_FDPIC) for (p=head; p; p=p->next) {
2117                 i = count_syms(p);
2118                 if (a-(size_t)p->funcdescs < i*sizeof(*p->funcdescs))
2119                         return p;
2120         }
2121         for (p=head; p; p=p->next) {
2122                 if (DL_FDPIC && p->loadmap) {
2123                         for (i=0; i<p->loadmap->nsegs; i++) {
2124                                 if (a-p->loadmap->segs[i].p_vaddr
2125                                     < p->loadmap->segs[i].p_memsz)
2126                                         return p;
2127                         }
2128                 } else {
2129                         Phdr *ph = p->phdr;
2130                         size_t phcnt = p->phnum;
2131                         size_t entsz = p->phentsize;
2132                         size_t base = (size_t)p->base;
2133                         for (; phcnt--; ph=(void *)((char *)ph+entsz)) {
2134                                 if (ph->p_type != PT_LOAD) continue;
2135                                 if (a-base-ph->p_vaddr < ph->p_memsz)
2136                                         return p;
2137                         }
2138                         if (a-(size_t)p->map < p->map_len)
2139                                 return 0;
2140                 }
2141         }
2142         return 0;
2143 }
2144
2145 static void *do_dlsym(struct dso *p, const char *s, void *ra)
2146 {
2147         int use_deps = 0;
2148         if (p == head || p == RTLD_DEFAULT) {
2149                 p = head;
2150         } else if (p == RTLD_NEXT) {
2151                 p = addr2dso((size_t)ra);
2152                 if (!p) p=head;
2153                 p = p->next;
2154         } else if (__dl_invalid_handle(p)) {
2155                 return 0;
2156         } else
2157                 use_deps = 1;
2158         struct symdef def = find_sym2(p, s, 0, use_deps);
2159         if (!def.sym) {
2160                 error("Symbol not found: %s", s);
2161                 return 0;
2162         }
2163         if ((def.sym->st_info&0xf) == STT_TLS)
2164                 return __tls_get_addr((tls_mod_off_t []){def.dso->tls_id, def.sym->st_value-DTP_OFFSET});
2165         if (DL_FDPIC && (def.sym->st_info&0xf) == STT_FUNC)
2166                 return def.dso->funcdescs + (def.sym - def.dso->syms);
2167         return laddr(def.dso, def.sym->st_value);
2168 }
2169
2170 int dladdr(const void *addr_arg, Dl_info *info)
2171 {
2172         size_t addr = (size_t)addr_arg;
2173         struct dso *p;
2174         Sym *sym, *bestsym;
2175         uint32_t nsym;
2176         char *strings;
2177         size_t best = 0;
2178         size_t besterr = -1;
2179
2180         pthread_rwlock_rdlock(&lock);
2181         p = addr2dso(addr);
2182         pthread_rwlock_unlock(&lock);
2183
2184         if (!p) return 0;
2185
2186         sym = p->syms;
2187         strings = p->strings;
2188         nsym = count_syms(p);
2189
2190         if (DL_FDPIC) {
2191                 size_t idx = (addr-(size_t)p->funcdescs)
2192                         / sizeof(*p->funcdescs);
2193                 if (idx < nsym && (sym[idx].st_info&0xf) == STT_FUNC) {
2194                         best = (size_t)(p->funcdescs + idx);
2195                         bestsym = sym + idx;
2196                         besterr = 0;
2197                 }
2198         }
2199
2200         if (!best) for (; nsym; nsym--, sym++) {
2201                 if (sym->st_value
2202                  && (1<<(sym->st_info&0xf) & OK_TYPES)
2203                  && (1<<(sym->st_info>>4) & OK_BINDS)) {
2204                         size_t symaddr = (size_t)laddr(p, sym->st_value);
2205                         if (symaddr > addr || symaddr <= best)
2206                                 continue;
2207                         best = symaddr;
2208                         bestsym = sym;
2209                         besterr = addr - symaddr;
2210                         if (addr == symaddr)
2211                                 break;
2212                 }
2213         }
2214
2215         if (best && besterr > bestsym->st_size-1) {
2216                 best = 0;
2217                 bestsym = 0;
2218         }
2219
2220         info->dli_fname = p->name;
2221         info->dli_fbase = p->map;
2222
2223         if (!best) {
2224                 info->dli_sname = 0;
2225                 info->dli_saddr = 0;
2226                 return 1;
2227         }
2228
2229         if (DL_FDPIC && (bestsym->st_info&0xf) == STT_FUNC)
2230                 best = (size_t)(p->funcdescs + (bestsym - p->syms));
2231         info->dli_sname = strings + bestsym->st_name;
2232         info->dli_saddr = (void *)best;
2233
2234         return 1;
2235 }
2236
2237 hidden void *__dlsym(void *restrict p, const char *restrict s, void *restrict ra)
2238 {
2239         void *res;
2240         pthread_rwlock_rdlock(&lock);
2241         res = do_dlsym(p, s, ra);
2242         pthread_rwlock_unlock(&lock);
2243         return res;
2244 }
2245
2246 hidden void *__dlsym_redir_time64(void *restrict p, const char *restrict s, void *restrict ra)
2247 {
2248 #if _REDIR_TIME64
2249         const char *suffix, *suffix2 = "";
2250         char redir[36];
2251
2252         /* Map the symbol name to a time64 version of itself according to the
2253          * pattern used for naming the redirected time64 symbols. */
2254         size_t l = strnlen(s, sizeof redir);
2255         if (l<4 || l==sizeof redir) goto no_redir;
2256         if (s[l-2]=='_' && s[l-1]=='r') {
2257                 l -= 2;
2258                 suffix2 = s+l;
2259         }
2260         if (l<4) goto no_redir;
2261         if (!strcmp(s+l-4, "time")) suffix = "64";
2262         else suffix = "_time64";
2263
2264         /* Use the presence of the remapped symbol name in libc to determine
2265          * whether it's one that requires time64 redirection; replace if so. */
2266         snprintf(redir, sizeof redir, "__%.*s%s%s", (int)l, s, suffix, suffix2);
2267         if (find_sym(&ldso, redir, 1).sym) s = redir;
2268 no_redir:
2269 #endif
2270         return __dlsym(p, s, ra);
2271 }
2272
2273 int dl_iterate_phdr(int(*callback)(struct dl_phdr_info *info, size_t size, void *data), void *data)
2274 {
2275         struct dso *current;
2276         struct dl_phdr_info info;
2277         int ret = 0;
2278         for(current = head; current;) {
2279                 info.dlpi_addr      = (uintptr_t)current->base;
2280                 info.dlpi_name      = current->name;
2281                 info.dlpi_phdr      = current->phdr;
2282                 info.dlpi_phnum     = current->phnum;
2283                 info.dlpi_adds      = gencnt;
2284                 info.dlpi_subs      = 0;
2285                 info.dlpi_tls_modid = current->tls_id;
2286                 info.dlpi_tls_data  = current->tls.image;
2287
2288                 ret = (callback)(&info, sizeof (info), data);
2289
2290                 if (ret != 0) break;
2291
2292                 pthread_rwlock_rdlock(&lock);
2293                 current = current->next;
2294                 pthread_rwlock_unlock(&lock);
2295         }
2296         return ret;
2297 }
2298
2299 static void error(const char *fmt, ...)
2300 {
2301         va_list ap;
2302         va_start(ap, fmt);
2303         if (!runtime) {
2304                 vdprintf(2, fmt, ap);
2305                 dprintf(2, "\n");
2306                 ldso_fail = 1;
2307                 va_end(ap);
2308                 return;
2309         }
2310         __dl_vseterr(fmt, ap);
2311         va_end(ap);
2312 }