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