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