1 /* Create descriptor for processing file.
2    Copyright (C) 1998-2010, 2012, 2014, 2015 Red Hat, Inc.
3    This file is part of elfutils.
4    Written by Ulrich Drepper <drepper@redhat.com>, 1998.
5 
6    This file is free software; you can redistribute it and/or modify
7    it under the terms of either
8 
9      * the GNU Lesser General Public License as published by the Free
10        Software Foundation; either version 3 of the License, or (at
11        your option) any later version
12 
13    or
14 
15      * the GNU General Public License as published by the Free
16        Software Foundation; either version 2 of the License, or (at
17        your option) any later version
18 
19    or both in parallel, as here.
20 
21    elfutils is distributed in the hope that it will be useful, but
22    WITHOUT ANY WARRANTY; without even the implied warranty of
23    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
24    General Public License for more details.
25 
26    You should have received copies of the GNU General Public License and
27    the GNU Lesser General Public License along with this program.  If
28    not, see <http://www.gnu.org/licenses/>.  */
29 
30 #ifdef HAVE_CONFIG_H
31 # include <config.h>
32 #endif
33 
34 #include <assert.h>
35 #include <ctype.h>
36 #include <errno.h>
37 #include <fcntl.h>
38 #include <stdbool.h>
39 #include <stddef.h>
40 #include <string.h>
41 #include <unistd.h>
42 #include <sys/mman.h>
43 #include <sys/param.h>
44 #include <sys/stat.h>
45 
46 #include <system.h>
47 #include "libelfP.h"
48 #include "common.h"
49 
50 
51 /* Create descriptor for archive in memory.  */
52 static inline Elf *
file_read_ar(int fildes,void * map_address,off_t offset,size_t maxsize,Elf_Cmd cmd,Elf * parent)53 file_read_ar (int fildes, void *map_address, off_t offset, size_t maxsize,
54 	      Elf_Cmd cmd, Elf *parent)
55 {
56   Elf *elf;
57 
58   /* Create a descriptor.  */
59   elf = allocate_elf (fildes, map_address, offset, maxsize, cmd, parent,
60                       ELF_K_AR, 0);
61   if (elf != NULL)
62     {
63       /* We don't read all the symbol tables in advance.  All this will
64 	 happen on demand.  */
65       elf->state.ar.offset = offset + SARMAG;
66 
67       elf->state.ar.elf_ar_hdr.ar_rawname = elf->state.ar.raw_name;
68     }
69 
70   return elf;
71 }
72 
73 
74 static size_t
get_shnum(void * map_address,unsigned char * e_ident,int fildes,off_t offset,size_t maxsize)75 get_shnum (void *map_address, unsigned char *e_ident, int fildes, off_t offset,
76 	   size_t maxsize)
77 {
78   size_t result;
79   union
80   {
81     Elf32_Ehdr *e32;
82     Elf64_Ehdr *e64;
83     void *p;
84   } ehdr;
85   union
86   {
87     Elf32_Ehdr e32;
88     Elf64_Ehdr e64;
89   } ehdr_mem;
90   bool is32 = e_ident[EI_CLASS] == ELFCLASS32;
91 
92   /* Make the ELF header available.  */
93   if (e_ident[EI_DATA] == MY_ELFDATA
94       && (ALLOW_UNALIGNED
95 	  || (((size_t) e_ident
96 	       & ((is32 ? __alignof__ (Elf32_Ehdr) : __alignof__ (Elf64_Ehdr))
97 		  - 1)) == 0)))
98     ehdr.p = e_ident;
99   else
100     {
101       /* We already read the ELF header.  We have to copy the header
102 	 since we possibly modify the data here and the caller
103 	 expects the memory it passes in to be preserved.  */
104       ehdr.p = &ehdr_mem;
105 
106       if (is32)
107 	{
108 	  if (ALLOW_UNALIGNED)
109 	    {
110 	      ehdr_mem.e32.e_shnum = ((Elf32_Ehdr *) e_ident)->e_shnum;
111 	      ehdr_mem.e32.e_shoff = ((Elf32_Ehdr *) e_ident)->e_shoff;
112 	    }
113 	  else
114 	    memcpy (&ehdr_mem, e_ident, sizeof (Elf32_Ehdr));
115 
116 	  if (e_ident[EI_DATA] != MY_ELFDATA)
117 	    {
118 	      CONVERT (ehdr_mem.e32.e_shnum);
119 	      CONVERT (ehdr_mem.e32.e_shoff);
120 	    }
121 	}
122       else
123 	{
124 	  if (ALLOW_UNALIGNED)
125 	    {
126 	      ehdr_mem.e64.e_shnum = ((Elf64_Ehdr *) e_ident)->e_shnum;
127 	      ehdr_mem.e64.e_shoff = ((Elf64_Ehdr *) e_ident)->e_shoff;
128 	    }
129 	  else
130 	    memcpy (&ehdr_mem, e_ident, sizeof (Elf64_Ehdr));
131 
132 	  if (e_ident[EI_DATA] != MY_ELFDATA)
133 	    {
134 	      CONVERT (ehdr_mem.e64.e_shnum);
135 	      CONVERT (ehdr_mem.e64.e_shoff);
136 	    }
137 	}
138     }
139 
140   if (is32)
141     {
142       /* Get the number of sections from the ELF header.  */
143       result = ehdr.e32->e_shnum;
144 
145       if (unlikely (result == 0) && ehdr.e32->e_shoff != 0)
146 	{
147 	  if (unlikely (ehdr.e32->e_shoff >= maxsize)
148 	      || unlikely (maxsize - ehdr.e32->e_shoff < sizeof (Elf32_Shdr)))
149 	    /* Cannot read the first section header.  */
150 	    return 0;
151 
152 	  if (likely (map_address != NULL) && e_ident[EI_DATA] == MY_ELFDATA
153 	      && (ALLOW_UNALIGNED
154 		  || (((size_t) ((char *) map_address + ehdr.e32->e_shoff))
155 		      & (__alignof__ (Elf32_Shdr) - 1)) == 0))
156 	    /* We can directly access the memory.  */
157 	    result = ((Elf32_Shdr *) ((char *) map_address + ehdr.e32->e_shoff
158 				      + offset))->sh_size;
159 	  else
160 	    {
161 	      Elf32_Word size;
162 
163 	      if (likely (map_address != NULL))
164 		/* gcc will optimize the memcpy to a simple memory
165 		   access while taking care of alignment issues.  */
166 		memcpy (&size, &((Elf32_Shdr *) ((char *) map_address
167 						 + ehdr.e32->e_shoff
168 						 + offset))->sh_size,
169 			sizeof (Elf32_Word));
170 	      else
171 		if (unlikely (pread_retry (fildes, &size, sizeof (Elf32_Word),
172 					   offset + ehdr.e32->e_shoff
173 					   + offsetof (Elf32_Shdr, sh_size))
174 			      != sizeof (Elf32_Word)))
175 		  return (size_t) -1l;
176 
177 	      if (e_ident[EI_DATA] != MY_ELFDATA)
178 		CONVERT (size);
179 
180 	      result = size;
181 	    }
182 	}
183 
184       /* If the section headers were truncated, pretend none were there.  */
185       if (ehdr.e32->e_shoff > maxsize
186 	  || maxsize - ehdr.e32->e_shoff < sizeof (Elf32_Shdr) * result)
187 	result = 0;
188     }
189   else
190     {
191       /* Get the number of sections from the ELF header.  */
192       result = ehdr.e64->e_shnum;
193 
194       if (unlikely (result == 0) && ehdr.e64->e_shoff != 0)
195 	{
196 	  if (unlikely (ehdr.e64->e_shoff >= maxsize)
197 	      || unlikely (ehdr.e64->e_shoff + sizeof (Elf64_Shdr) > maxsize))
198 	    /* Cannot read the first section header.  */
199 	    return 0;
200 
201 	  Elf64_Xword size;
202 	  if (likely (map_address != NULL) && e_ident[EI_DATA] == MY_ELFDATA
203 	      && (ALLOW_UNALIGNED
204 		  || (((size_t) ((char *) map_address + ehdr.e64->e_shoff))
205 		      & (__alignof__ (Elf64_Shdr) - 1)) == 0))
206 	    /* We can directly access the memory.  */
207 	    size = ((Elf64_Shdr *) ((char *) map_address + ehdr.e64->e_shoff
208 				    + offset))->sh_size;
209 	  else
210 	    {
211 	      if (likely (map_address != NULL))
212 		/* gcc will optimize the memcpy to a simple memory
213 		   access while taking care of alignment issues.  */
214 		memcpy (&size, &((Elf64_Shdr *) ((char *) map_address
215 						 + ehdr.e64->e_shoff
216 						 + offset))->sh_size,
217 			sizeof (Elf64_Xword));
218 	      else
219 		if (unlikely (pread_retry (fildes, &size, sizeof (Elf64_Xword),
220 					   offset + ehdr.e64->e_shoff
221 					   + offsetof (Elf64_Shdr, sh_size))
222 			      != sizeof (Elf64_Xword)))
223 		  return (size_t) -1l;
224 
225 	      if (e_ident[EI_DATA] != MY_ELFDATA)
226 		CONVERT (size);
227 	    }
228 
229 	  if (size > ~((GElf_Word) 0))
230 	    /* Invalid value, it is too large.  */
231 	    return (size_t) -1l;
232 
233 	  result = size;
234 	}
235 
236       /* If the section headers were truncated, pretend none were there.  */
237       if (ehdr.e64->e_shoff > maxsize
238 	  || maxsize - ehdr.e64->e_shoff < sizeof (Elf64_Shdr) * result)
239 	result = 0;
240     }
241 
242   return result;
243 }
244 
245 
246 /* Create descriptor for ELF file in memory.  */
247 static Elf *
file_read_elf(int fildes,void * map_address,unsigned char * e_ident,off_t offset,size_t maxsize,Elf_Cmd cmd,Elf * parent)248 file_read_elf (int fildes, void *map_address, unsigned char *e_ident,
249 	       off_t offset, size_t maxsize, Elf_Cmd cmd, Elf *parent)
250 {
251   /* Verify the binary is of the class we can handle.  */
252   if (unlikely ((e_ident[EI_CLASS] != ELFCLASS32
253 		 && e_ident[EI_CLASS] != ELFCLASS64)
254 		/* We also can only handle two encodings.  */
255 		|| (e_ident[EI_DATA] != ELFDATA2LSB
256 		    && e_ident[EI_DATA] != ELFDATA2MSB)))
257     {
258       /* Cannot handle this.  */
259       __libelf_seterrno (ELF_E_INVALID_FILE);
260       return NULL;
261     }
262 
263   /* Determine the number of sections.  */
264   size_t scncnt = get_shnum (map_address, e_ident, fildes, offset, maxsize);
265   if (scncnt == (size_t) -1l)
266     /* Could not determine the number of sections.  */
267     return NULL;
268 
269   /* Check for too many sections.  */
270   if (e_ident[EI_CLASS] == ELFCLASS32)
271     {
272       if (scncnt > SIZE_MAX / (sizeof (Elf_Scn) + sizeof (Elf32_Shdr)))
273 	return NULL;
274     }
275   else if (scncnt > SIZE_MAX / (sizeof (Elf_Scn) + sizeof (Elf64_Shdr)))
276     return NULL;
277 
278   /* We can now allocate the memory.  Even if there are no section headers,
279      we allocate space for a zeroth section in case we need it later.  */
280   const size_t scnmax = (scncnt ?: (cmd == ELF_C_RDWR || cmd == ELF_C_RDWR_MMAP)
281 			 ? 1 : 0);
282   Elf *elf = allocate_elf (fildes, map_address, offset, maxsize, cmd, parent,
283 			   ELF_K_ELF, scnmax * sizeof (Elf_Scn));
284   if (elf == NULL)
285     /* Not enough memory.  */
286     return NULL;
287 
288   assert ((unsigned int) scncnt == scncnt);
289   assert (offsetof (struct Elf, state.elf32.scns)
290 	  == offsetof (struct Elf, state.elf64.scns));
291   elf->state.elf32.scns.cnt = scncnt;
292   elf->state.elf32.scns.max = scnmax;
293 
294   /* Some more or less arbitrary value.  */
295   elf->state.elf.scnincr = 10;
296 
297   /* Make the class easily available.  */
298   elf->class = e_ident[EI_CLASS];
299 
300   if (e_ident[EI_CLASS] == ELFCLASS32)
301     {
302       /* This pointer might not be directly usable if the alignment is
303 	 not sufficient for the architecture.  */
304       Elf32_Ehdr *ehdr = (Elf32_Ehdr *) ((char *) map_address + offset);
305 
306       /* This is a 32-bit binary.  */
307       if (map_address != NULL && e_ident[EI_DATA] == MY_ELFDATA
308 	  && (ALLOW_UNALIGNED
309 	      || (((uintptr_t) ehdr) & (__alignof__ (Elf32_Ehdr) - 1)) == 0))
310 	{
311 	  /* We can use the mmapped memory.  */
312 	  elf->state.elf32.ehdr = ehdr;
313 	}
314       else
315 	{
316 	  /* Copy the ELF header.  */
317 	  elf->state.elf32.ehdr = memcpy (&elf->state.elf32.ehdr_mem, e_ident,
318 					  sizeof (Elf32_Ehdr));
319 
320 	  if (e_ident[EI_DATA] != MY_ELFDATA)
321 	    {
322 	      CONVERT (elf->state.elf32.ehdr_mem.e_type);
323 	      CONVERT (elf->state.elf32.ehdr_mem.e_machine);
324 	      CONVERT (elf->state.elf32.ehdr_mem.e_version);
325 	      CONVERT (elf->state.elf32.ehdr_mem.e_entry);
326 	      CONVERT (elf->state.elf32.ehdr_mem.e_phoff);
327 	      CONVERT (elf->state.elf32.ehdr_mem.e_shoff);
328 	      CONVERT (elf->state.elf32.ehdr_mem.e_flags);
329 	      CONVERT (elf->state.elf32.ehdr_mem.e_ehsize);
330 	      CONVERT (elf->state.elf32.ehdr_mem.e_phentsize);
331 	      CONVERT (elf->state.elf32.ehdr_mem.e_phnum);
332 	      CONVERT (elf->state.elf32.ehdr_mem.e_shentsize);
333 	      CONVERT (elf->state.elf32.ehdr_mem.e_shnum);
334 	      CONVERT (elf->state.elf32.ehdr_mem.e_shstrndx);
335 	    }
336 	}
337 
338       /* Don't precache the phdr pointer here.
339 	 elf32_getphdr will validate it against the size when asked.  */
340 
341       Elf32_Off e_shoff = elf->state.elf32.ehdr->e_shoff;
342       if (map_address != NULL && e_ident[EI_DATA] == MY_ELFDATA
343 	  && cmd != ELF_C_READ_MMAP /* We need a copy to be able to write.  */
344 	  && (ALLOW_UNALIGNED
345 	      || (((uintptr_t) ((char *) ehdr + e_shoff)
346 		   & (__alignof__ (Elf32_Shdr) - 1)) == 0)))
347 	{
348 	  if (unlikely (e_shoff >= maxsize)
349 	      || unlikely (maxsize - e_shoff
350 			   < scncnt * sizeof (Elf32_Shdr)))
351 	    {
352 	    free_and_out:
353 	      free (elf);
354 	      __libelf_seterrno (ELF_E_INVALID_FILE);
355 	      return NULL;
356 	    }
357 	  elf->state.elf32.shdr
358 	    = (Elf32_Shdr *) ((char *) ehdr + e_shoff);
359 
360 	  for (size_t cnt = 0; cnt < scncnt; ++cnt)
361 	    {
362 	      elf->state.elf32.scns.data[cnt].index = cnt;
363 	      elf->state.elf32.scns.data[cnt].elf = elf;
364 	      elf->state.elf32.scns.data[cnt].shdr.e32 =
365 		&elf->state.elf32.shdr[cnt];
366 	      if (likely (elf->state.elf32.shdr[cnt].sh_offset < maxsize)
367 		  && likely (elf->state.elf32.shdr[cnt].sh_size
368 			     <= maxsize - elf->state.elf32.shdr[cnt].sh_offset))
369 		elf->state.elf32.scns.data[cnt].rawdata_base =
370 		  elf->state.elf32.scns.data[cnt].data_base =
371 		  ((char *) map_address + offset
372 		   + elf->state.elf32.shdr[cnt].sh_offset);
373 	      elf->state.elf32.scns.data[cnt].list = &elf->state.elf32.scns;
374 
375 	      /* If this is a section with an extended index add a
376 		 reference in the section which uses the extended
377 		 index.  */
378 	      if (elf->state.elf32.shdr[cnt].sh_type == SHT_SYMTAB_SHNDX
379 		  && elf->state.elf32.shdr[cnt].sh_link < scncnt)
380 		elf->state.elf32.scns.data[elf->state.elf32.shdr[cnt].sh_link].shndx_index
381 		  = cnt;
382 
383 	      /* Set the own shndx_index field in case it has not yet
384 		 been set.  */
385 	      if (elf->state.elf32.scns.data[cnt].shndx_index == 0)
386 		elf->state.elf32.scns.data[cnt].shndx_index = -1;
387 	    }
388 	}
389       else
390 	{
391 	  for (size_t cnt = 0; cnt < scncnt; ++cnt)
392 	    {
393 	      elf->state.elf32.scns.data[cnt].index = cnt;
394 	      elf->state.elf32.scns.data[cnt].elf = elf;
395 	      elf->state.elf32.scns.data[cnt].list = &elf->state.elf32.scns;
396 	    }
397 	}
398 
399       /* So far only one block with sections.  */
400       elf->state.elf32.scns_last = &elf->state.elf32.scns;
401     }
402   else
403     {
404       /* This pointer might not be directly usable if the alignment is
405 	 not sufficient for the architecture.  */
406       Elf64_Ehdr *ehdr = (Elf64_Ehdr *) ((char *) map_address + offset);
407 
408       /* This is a 64-bit binary.  */
409       if (map_address != NULL && e_ident[EI_DATA] == MY_ELFDATA
410 	  && (ALLOW_UNALIGNED
411 	      || (((uintptr_t) ehdr) & (__alignof__ (Elf64_Ehdr) - 1)) == 0))
412 	{
413 	  /* We can use the mmapped memory.  */
414 	  elf->state.elf64.ehdr = ehdr;
415 	}
416       else
417 	{
418 	  /* Copy the ELF header.  */
419 	  elf->state.elf64.ehdr = memcpy (&elf->state.elf64.ehdr_mem, e_ident,
420 					  sizeof (Elf64_Ehdr));
421 
422 	  if (e_ident[EI_DATA] != MY_ELFDATA)
423 	    {
424 	      CONVERT (elf->state.elf64.ehdr_mem.e_type);
425 	      CONVERT (elf->state.elf64.ehdr_mem.e_machine);
426 	      CONVERT (elf->state.elf64.ehdr_mem.e_version);
427 	      CONVERT (elf->state.elf64.ehdr_mem.e_entry);
428 	      CONVERT (elf->state.elf64.ehdr_mem.e_phoff);
429 	      CONVERT (elf->state.elf64.ehdr_mem.e_shoff);
430 	      CONVERT (elf->state.elf64.ehdr_mem.e_flags);
431 	      CONVERT (elf->state.elf64.ehdr_mem.e_ehsize);
432 	      CONVERT (elf->state.elf64.ehdr_mem.e_phentsize);
433 	      CONVERT (elf->state.elf64.ehdr_mem.e_phnum);
434 	      CONVERT (elf->state.elf64.ehdr_mem.e_shentsize);
435 	      CONVERT (elf->state.elf64.ehdr_mem.e_shnum);
436 	      CONVERT (elf->state.elf64.ehdr_mem.e_shstrndx);
437 	    }
438 	}
439 
440       /* Don't precache the phdr pointer here.
441 	 elf64_getphdr will validate it against the size when asked.  */
442 
443       Elf64_Off e_shoff = elf->state.elf64.ehdr->e_shoff;
444       if (map_address != NULL && e_ident[EI_DATA] == MY_ELFDATA
445 	  && cmd != ELF_C_READ_MMAP /* We need a copy to be able to write.  */
446 	  && (ALLOW_UNALIGNED
447 	      || (((uintptr_t) ((char *) ehdr + e_shoff)
448 		   & (__alignof__ (Elf64_Shdr) - 1)) == 0)))
449 	{
450 	  if (unlikely (e_shoff >= maxsize)
451 	      || unlikely (maxsize - e_shoff
452 			   < scncnt * sizeof (Elf64_Shdr)))
453 	    goto free_and_out;
454 	  elf->state.elf64.shdr
455 	    = (Elf64_Shdr *) ((char *) ehdr + e_shoff);
456 
457 	  for (size_t cnt = 0; cnt < scncnt; ++cnt)
458 	    {
459 	      elf->state.elf64.scns.data[cnt].index = cnt;
460 	      elf->state.elf64.scns.data[cnt].elf = elf;
461 	      elf->state.elf64.scns.data[cnt].shdr.e64 =
462 		&elf->state.elf64.shdr[cnt];
463 	      if (likely (elf->state.elf64.shdr[cnt].sh_offset < maxsize)
464 		  && likely (elf->state.elf64.shdr[cnt].sh_size
465 			     <= maxsize - elf->state.elf64.shdr[cnt].sh_offset))
466 		elf->state.elf64.scns.data[cnt].rawdata_base =
467 		  elf->state.elf64.scns.data[cnt].data_base =
468 		  ((char *) map_address + offset
469 		   + elf->state.elf64.shdr[cnt].sh_offset);
470 	      elf->state.elf64.scns.data[cnt].list = &elf->state.elf64.scns;
471 
472 	      /* If this is a section with an extended index add a
473 		 reference in the section which uses the extended
474 		 index.  */
475 	      if (elf->state.elf64.shdr[cnt].sh_type == SHT_SYMTAB_SHNDX
476 		  && elf->state.elf64.shdr[cnt].sh_link < scncnt)
477 		elf->state.elf64.scns.data[elf->state.elf64.shdr[cnt].sh_link].shndx_index
478 		  = cnt;
479 
480 	      /* Set the own shndx_index field in case it has not yet
481 		 been set.  */
482 	      if (elf->state.elf64.scns.data[cnt].shndx_index == 0)
483 		elf->state.elf64.scns.data[cnt].shndx_index = -1;
484 	    }
485 	}
486       else
487 	{
488 	  for (size_t cnt = 0; cnt < scncnt; ++cnt)
489 	    {
490 	      elf->state.elf64.scns.data[cnt].index = cnt;
491 	      elf->state.elf64.scns.data[cnt].elf = elf;
492 	      elf->state.elf64.scns.data[cnt].list = &elf->state.elf64.scns;
493 	    }
494 	}
495 
496       /* So far only one block with sections.  */
497       elf->state.elf64.scns_last = &elf->state.elf64.scns;
498     }
499 
500   return elf;
501 }
502 
503 
504 Elf *
505 internal_function
__libelf_read_mmaped_file(int fildes,void * map_address,off_t offset,size_t maxsize,Elf_Cmd cmd,Elf * parent)506 __libelf_read_mmaped_file (int fildes, void *map_address,  off_t offset,
507 			   size_t maxsize, Elf_Cmd cmd, Elf *parent)
508 {
509   /* We have to find out what kind of file this is.  We handle ELF
510      files and archives.  To find out what we have we must look at the
511      header.  The header for an ELF file is EI_NIDENT bytes in size,
512      the header for an archive file SARMAG bytes long.  */
513   unsigned char *e_ident = (unsigned char *) map_address + offset;
514 
515   /* See what kind of object we have here.  */
516   Elf_Kind kind = determine_kind (e_ident, maxsize);
517 
518   switch (kind)
519     {
520     case ELF_K_ELF:
521       return file_read_elf (fildes, map_address, e_ident, offset, maxsize,
522 			    cmd, parent);
523 
524     case ELF_K_AR:
525       return file_read_ar (fildes, map_address, offset, maxsize, cmd, parent);
526 
527     default:
528       break;
529     }
530 
531   /* This case is easy.  Since we cannot do anything with this file
532      create a dummy descriptor.  */
533   return allocate_elf (fildes, map_address, offset, maxsize, cmd, parent,
534 		       ELF_K_NONE, 0);
535 }
536 
537 
538 static Elf *
read_unmmaped_file(int fildes,off_t offset,size_t maxsize,Elf_Cmd cmd,Elf * parent)539 read_unmmaped_file (int fildes, off_t offset, size_t maxsize, Elf_Cmd cmd,
540 		    Elf *parent)
541 {
542   /* We have to find out what kind of file this is.  We handle ELF
543      files and archives.  To find out what we have we must read the
544      header.  The identification header for an ELF file is EI_NIDENT
545      bytes in size, but we read the whole ELF header since we will
546      need it anyway later.  For archives the header in SARMAG bytes
547      long.  Read the maximum of these numbers.
548 
549      XXX We have to change this for the extended `ar' format some day.
550 
551      Use a union to ensure alignment.  We might later access the
552      memory as a ElfXX_Ehdr.  */
553   union
554   {
555     Elf64_Ehdr ehdr;
556     unsigned char header[MAX (sizeof (Elf64_Ehdr), SARMAG)];
557   } mem;
558 
559   /* Read the head of the file.  */
560   ssize_t nread = pread_retry (fildes, mem.header,
561 			       MIN (MAX (sizeof (Elf64_Ehdr), SARMAG),
562 				    maxsize),
563 			       offset);
564   if (unlikely (nread == -1))
565     {
566       /* We cannot even read the head of the file.  Maybe FILDES is associated
567 	 with an unseekable device.  This is nothing we can handle.  */
568       __libelf_seterrno (ELF_E_INVALID_FILE);
569       return NULL;
570     }
571 
572   /* See what kind of object we have here.  */
573   Elf_Kind kind = determine_kind (mem.header, nread);
574 
575   switch (kind)
576     {
577     case ELF_K_AR:
578       return file_read_ar (fildes, NULL, offset, maxsize, cmd, parent);
579 
580     case ELF_K_ELF:
581       /* Make sure at least the ELF header is contained in the file.  */
582       if ((size_t) nread >= (mem.header[EI_CLASS] == ELFCLASS32
583 			     ? sizeof (Elf32_Ehdr) : sizeof (Elf64_Ehdr)))
584 	return file_read_elf (fildes, NULL, mem.header, offset, maxsize, cmd,
585 			      parent);
586       /* FALLTHROUGH */
587 
588     default:
589       break;
590     }
591 
592   /* This case is easy.  Since we cannot do anything with this file
593      create a dummy descriptor.  */
594   return allocate_elf (fildes, NULL, offset, maxsize, cmd, parent,
595 		       ELF_K_NONE, 0);
596 }
597 
598 
599 /* Open a file for reading.  If possible we will try to mmap() the file.  */
600 static struct Elf *
read_file(int fildes,off_t offset,size_t maxsize,Elf_Cmd cmd,Elf * parent)601 read_file (int fildes, off_t offset, size_t maxsize,
602 	   Elf_Cmd cmd, Elf *parent)
603 {
604   void *map_address = NULL;
605   int use_mmap = (cmd == ELF_C_READ_MMAP || cmd == ELF_C_RDWR_MMAP
606 		  || cmd == ELF_C_WRITE_MMAP
607 		  || cmd == ELF_C_READ_MMAP_PRIVATE);
608 
609   if (use_mmap)
610     {
611       if (parent == NULL)
612 	{
613 	  if (maxsize == ~((size_t) 0))
614 	    {
615 	      /* We don't know in the moment how large the file is.
616 		 Determine it now.  */
617 	      struct stat st;
618 
619 	      if (fstat (fildes, &st) == 0
620 		  && (sizeof (size_t) >= sizeof (st.st_size)
621 		      || st.st_size <= ~((size_t) 0)))
622 		maxsize = (size_t) st.st_size;
623 	    }
624 
625 	  /* We try to map the file ourself.  */
626 	  map_address = mmap (NULL, maxsize, (cmd == ELF_C_READ_MMAP
627 					      ? PROT_READ
628 					      : PROT_READ|PROT_WRITE),
629 			      cmd == ELF_C_READ_MMAP_PRIVATE
630 			      || cmd == ELF_C_READ_MMAP
631 			      ? MAP_PRIVATE : MAP_SHARED,
632 			      fildes, offset);
633 
634 	  if (map_address == MAP_FAILED)
635 	    map_address = NULL;
636 	}
637       else
638 	{
639 	  /* The parent is already loaded.  Use it.  */
640 	  assert (maxsize != ~((size_t) 0));
641 
642 	  map_address = parent->map_address;
643 	}
644     }
645 
646   /* If we have the file in memory optimize the access.  */
647   if (map_address != NULL)
648     {
649       assert (map_address != MAP_FAILED);
650 
651       struct Elf *result = __libelf_read_mmaped_file (fildes, map_address,
652 						      offset, maxsize, cmd,
653 						      parent);
654 
655       /* If something went wrong during the initialization unmap the
656 	 memory if we mmaped here.  */
657       if (result == NULL
658 	  && (parent == NULL
659 	      || parent->map_address != map_address))
660 	munmap (map_address, maxsize);
661       else if (parent == NULL)
662 	/* Remember that we mmap()ed the memory.  */
663 	result->flags |= ELF_F_MMAPPED;
664 
665       return result;
666     }
667 
668   /* Otherwise we have to do it the hard way.  We read as much as necessary
669      from the file whenever we need information which is not available.  */
670   return read_unmmaped_file (fildes, offset, maxsize, cmd, parent);
671 }
672 
673 
674 /* Find the entry with the long names for the content of this archive.  */
675 static const char *
read_long_names(Elf * elf)676 read_long_names (Elf *elf)
677 {
678   off_t offset = SARMAG;	/* This is the first entry.  */
679   struct ar_hdr hdrm;
680   struct ar_hdr *hdr;
681   char *newp;
682   size_t len;
683 
684   while (1)
685     {
686       if (elf->map_address != NULL)
687 	{
688 	  if ((size_t) offset > elf->maximum_size
689 	      || elf->maximum_size - offset < sizeof (struct ar_hdr))
690 	    return NULL;
691 
692 	  /* The data is mapped.  */
693 	  hdr = (struct ar_hdr *) (elf->map_address + offset);
694 	}
695       else
696 	{
697 	  /* Read the header from the file.  */
698 	  if (unlikely (pread_retry (elf->fildes, &hdrm, sizeof (hdrm),
699 				     elf->start_offset + offset)
700 			!= sizeof (hdrm)))
701 	    return NULL;
702 
703 	  hdr = &hdrm;
704 	}
705 
706       len = atol (hdr->ar_size);
707 
708       if (memcmp (hdr->ar_name, "//              ", 16) == 0)
709 	break;
710 
711       offset += sizeof (struct ar_hdr) + ((len + 1) & ~1l);
712     }
713 
714   /* Due to the stupid format of the long name table entry (which are not
715      NUL terminted) we have to provide an appropriate representation anyhow.
716      Therefore we always make a copy which has the appropriate form.  */
717   newp = (char *) malloc (len);
718   if (newp != NULL)
719     {
720       char *runp;
721 
722       if (elf->map_address != NULL)
723 	{
724 	  if (len > elf->maximum_size - offset - sizeof (struct ar_hdr))
725 	    goto too_much;
726 	  /* Simply copy it over.  */
727 	  elf->state.ar.long_names = (char *) memcpy (newp,
728 						      elf->map_address + offset
729 						      + sizeof (struct ar_hdr),
730 						      len);
731 	}
732       else
733 	{
734 	  if (unlikely ((size_t) pread_retry (elf->fildes, newp, len,
735 					      elf->start_offset + offset
736 					      + sizeof (struct ar_hdr))
737 			!= len))
738 	    {
739 	    too_much:
740 	      /* We were not able to read all data.  */
741 	      free (newp);
742 	      elf->state.ar.long_names = NULL;
743 	      return NULL;
744 	    }
745 	  elf->state.ar.long_names = newp;
746 	}
747 
748       elf->state.ar.long_names_len = len;
749 
750       /* Now NUL-terminate the strings.  */
751       runp = newp;
752       while (1)
753         {
754 	  char *startp = runp;
755 	  runp = (char *) memchr (runp, '/', newp + len - runp);
756 	  if (runp == NULL)
757 	    {
758 	      /* This was the last entry.  Clear any left overs.  */
759 	      memset (startp, '\0', newp + len - startp);
760 	      break;
761 	    }
762 
763 	  /* NUL-terminate the string.  */
764 	  *runp++ = '\0';
765 
766 	  /* A sanity check.  Somebody might have generated invalid
767 	     archive.  */
768 	  if (runp >= newp + len)
769 	    break;
770 	}
771     }
772 
773   return newp;
774 }
775 
776 
777 /* Read the next archive header.  */
778 int
779 internal_function
__libelf_next_arhdr_wrlock(Elf * elf)780 __libelf_next_arhdr_wrlock (Elf *elf)
781 {
782   struct ar_hdr *ar_hdr;
783   Elf_Arhdr *elf_ar_hdr;
784 
785   if (elf->map_address != NULL)
786     {
787       /* See whether this entry is in the file.  */
788       if (unlikely ((size_t) elf->state.ar.offset
789 		    > elf->start_offset + elf->maximum_size
790 		    || (elf->start_offset + elf->maximum_size
791 			- elf->state.ar.offset) < sizeof (struct ar_hdr)))
792 	{
793 	  /* This record is not anymore in the file.  */
794 	  __libelf_seterrno (ELF_E_RANGE);
795 	  return -1;
796 	}
797       ar_hdr = (struct ar_hdr *) (elf->map_address + elf->state.ar.offset);
798     }
799   else
800     {
801       ar_hdr = &elf->state.ar.ar_hdr;
802 
803       if (unlikely (pread_retry (elf->fildes, ar_hdr, sizeof (struct ar_hdr),
804 				 elf->state.ar.offset)
805 		    != sizeof (struct ar_hdr)))
806 	{
807 	  /* Something went wrong while reading the file.  */
808 	  __libelf_seterrno (ELF_E_RANGE);
809 	  return -1;
810 	}
811     }
812 
813   /* One little consistency check.  */
814   if (unlikely (memcmp (ar_hdr->ar_fmag, ARFMAG, 2) != 0))
815     {
816       /* This is no valid archive.  */
817       __libelf_seterrno (ELF_E_ARCHIVE_FMAG);
818       return -1;
819     }
820 
821   /* Copy the raw name over to a NUL terminated buffer.  */
822   *((char *) mempcpy (elf->state.ar.raw_name, ar_hdr->ar_name, 16)) = '\0';
823 
824   elf_ar_hdr = &elf->state.ar.elf_ar_hdr;
825 
826   /* Now convert the `struct ar_hdr' into `Elf_Arhdr'.
827      Determine whether this is a special entry.  */
828   if (ar_hdr->ar_name[0] == '/')
829     {
830       if (ar_hdr->ar_name[1] == ' '
831 	  && memcmp (ar_hdr->ar_name, "/               ", 16) == 0)
832 	/* This is the index.  */
833 	elf_ar_hdr->ar_name = memcpy (elf->state.ar.ar_name, "/", 2);
834       else if (ar_hdr->ar_name[1] == 'S'
835 	       && memcmp (ar_hdr->ar_name, "/SYM64/         ", 16) == 0)
836 	/* 64-bit index.  */
837 	elf_ar_hdr->ar_name = memcpy (elf->state.ar.ar_name, "/SYM64/", 8);
838       else if (ar_hdr->ar_name[1] == '/'
839 	       && memcmp (ar_hdr->ar_name, "//              ", 16) == 0)
840 	/* This is the array with the long names.  */
841 	elf_ar_hdr->ar_name = memcpy (elf->state.ar.ar_name, "//", 3);
842       else if (likely  (isdigit (ar_hdr->ar_name[1])))
843 	{
844 	  size_t offset;
845 
846 	  /* This is a long name.  First we have to read the long name
847 	     table, if this hasn't happened already.  */
848 	  if (unlikely (elf->state.ar.long_names == NULL
849 			&& read_long_names (elf) == NULL))
850 	    {
851 	      /* No long name table although it is reference.  The archive is
852 		 broken.  */
853 	      __libelf_seterrno (ELF_E_INVALID_ARCHIVE);
854 	      return -1;
855 	    }
856 
857 	  offset = atol (ar_hdr->ar_name + 1);
858 	  if (unlikely (offset >= elf->state.ar.long_names_len))
859 	    {
860 	      /* The index in the long name table is larger than the table.  */
861 	      __libelf_seterrno (ELF_E_INVALID_ARCHIVE);
862 	      return -1;
863 	    }
864 	  elf_ar_hdr->ar_name = elf->state.ar.long_names + offset;
865 	}
866       else
867 	{
868 	  /* This is none of the known special entries.  */
869 	  __libelf_seterrno (ELF_E_INVALID_ARCHIVE);
870 	  return -1;
871 	}
872     }
873   else
874     {
875       char *endp;
876 
877       /* It is a normal entry.  Copy over the name.  */
878       endp = (char *) memccpy (elf->state.ar.ar_name, ar_hdr->ar_name,
879 			       '/', 16);
880       if (endp != NULL)
881 	endp[-1] = '\0';
882       else
883 	{
884 	  /* In the old BSD style of archive, there is no / terminator.
885 	     Instead, there is space padding at the end of the name.  */
886 	  size_t i = 15;
887 	  do
888 	    elf->state.ar.ar_name[i] = '\0';
889 	  while (i > 0 && elf->state.ar.ar_name[--i] == ' ');
890 	}
891 
892       elf_ar_hdr->ar_name = elf->state.ar.ar_name;
893     }
894 
895   if (unlikely (ar_hdr->ar_size[0] == ' '))
896     /* Something is really wrong.  We cannot live without a size for
897        the member since it will not be possible to find the next
898        archive member.  */
899     {
900       __libelf_seterrno (ELF_E_INVALID_ARCHIVE);
901       return -1;
902     }
903 
904   /* Since there are no specialized functions to convert ASCII to
905      time_t, uid_t, gid_t, mode_t, and off_t we use either atol or
906      atoll depending on the size of the types.  We are also prepared
907      for the case where the whole field in the `struct ar_hdr' is
908      filled in which case we cannot simply use atol/l but instead have
909      to create a temporary copy.  */
910 
911 #define INT_FIELD(FIELD)						      \
912   do									      \
913     {									      \
914       char buf[sizeof (ar_hdr->FIELD) + 1];				      \
915       const char *string = ar_hdr->FIELD;				      \
916       if (ar_hdr->FIELD[sizeof (ar_hdr->FIELD) - 1] != ' ')		      \
917 	{								      \
918 	  *((char *) mempcpy (buf, ar_hdr->FIELD, sizeof (ar_hdr->FIELD)))  \
919 	    = '\0';							      \
920 	  string = buf;							      \
921 	}								      \
922       if (sizeof (elf_ar_hdr->FIELD) <= sizeof (long int))		      \
923 	elf_ar_hdr->FIELD = (__typeof (elf_ar_hdr->FIELD)) atol (string);     \
924       else								      \
925 	elf_ar_hdr->FIELD = (__typeof (elf_ar_hdr->FIELD)) atoll (string);    \
926     }									      \
927   while (0)
928 
929   INT_FIELD (ar_date);
930   INT_FIELD (ar_uid);
931   INT_FIELD (ar_gid);
932   INT_FIELD (ar_mode);
933   INT_FIELD (ar_size);
934 
935   if (elf_ar_hdr->ar_size < 0)
936     {
937       __libelf_seterrno (ELF_E_INVALID_ARCHIVE);
938       return -1;
939     }
940 
941   /* Truncated file?  */
942   size_t maxsize;
943   maxsize = (elf->start_offset + elf->maximum_size
944 	     - elf->state.ar.offset - sizeof (struct ar_hdr));
945   if ((size_t) elf_ar_hdr->ar_size > maxsize)
946     elf_ar_hdr->ar_size = maxsize;
947 
948   return 0;
949 }
950 
951 
952 /* We were asked to return a clone of an existing descriptor.  This
953    function must be called with the lock on the parent descriptor
954    being held. */
955 static Elf *
dup_elf(int fildes,Elf_Cmd cmd,Elf * ref)956 dup_elf (int fildes, Elf_Cmd cmd, Elf *ref)
957 {
958   struct Elf *result;
959 
960   if (fildes == -1)
961     /* Allow the user to pass -1 as the file descriptor for the new file.  */
962     fildes = ref->fildes;
963   /* The file descriptor better should be the same.  If it was disconnected
964      already (using `elf_cntl') we do not test it.  */
965   else if (unlikely (ref->fildes != -1 && fildes != ref->fildes))
966     {
967       __libelf_seterrno (ELF_E_FD_MISMATCH);
968       return NULL;
969     }
970 
971   /* The mode must allow reading.  I.e., a descriptor creating with a
972      command different then ELF_C_READ, ELF_C_WRITE and ELF_C_RDWR is
973      not allowed.  */
974   if (unlikely (ref->cmd != ELF_C_READ && ref->cmd != ELF_C_READ_MMAP
975 		&& ref->cmd != ELF_C_WRITE && ref->cmd != ELF_C_WRITE_MMAP
976 		&& ref->cmd != ELF_C_RDWR && ref->cmd != ELF_C_RDWR_MMAP
977 		&& ref->cmd != ELF_C_READ_MMAP_PRIVATE))
978     {
979       __libelf_seterrno (ELF_E_INVALID_OP);
980       return NULL;
981     }
982 
983   /* Now it is time to distinguish between reading normal files and
984      archives.  Normal files can easily be handled be incrementing the
985      reference counter and return the same descriptor.  */
986   if (ref->kind != ELF_K_AR)
987     {
988       ++ref->ref_count;
989       return ref;
990     }
991 
992   /* This is an archive.  We must create a descriptor for the archive
993      member the internal pointer of the archive file desriptor is
994      pointing to.  First read the header of the next member if this
995      has not happened already.  */
996   if (ref->state.ar.elf_ar_hdr.ar_name == NULL
997       && __libelf_next_arhdr_wrlock (ref) != 0)
998     /* Something went wrong.  Maybe there is no member left.  */
999     return NULL;
1000 
1001   /* We have all the information we need about the next archive member.
1002      Now create a descriptor for it.  */
1003   result = read_file (fildes, ref->state.ar.offset + sizeof (struct ar_hdr),
1004 		      ref->state.ar.elf_ar_hdr.ar_size, cmd, ref);
1005 
1006   /* Enlist this new descriptor in the list of children.  */
1007   if (result != NULL)
1008     {
1009       result->next = ref->state.ar.children;
1010       ref->state.ar.children = result;
1011     }
1012 
1013   return result;
1014 }
1015 
1016 
1017 /* Return desriptor for empty file ready for writing.  */
1018 static struct Elf *
write_file(int fd,Elf_Cmd cmd)1019 write_file (int fd, Elf_Cmd cmd)
1020 {
1021   /* We simply create an empty `Elf' structure.  */
1022 #define NSCNSALLOC	10
1023   Elf *result = allocate_elf (fd, NULL, 0, 0, cmd, NULL, ELF_K_ELF,
1024 			      NSCNSALLOC * sizeof (Elf_Scn));
1025 
1026   if (result != NULL)
1027     {
1028       /* We have to write to the file in any case.  */
1029       result->flags = ELF_F_DIRTY;
1030 
1031       /* Some more or less arbitrary value.  */
1032       result->state.elf.scnincr = NSCNSALLOC;
1033 
1034       /* We have allocated room for some sections.  */
1035       assert (offsetof (struct Elf, state.elf32.scns)
1036 	      == offsetof (struct Elf, state.elf64.scns));
1037       result->state.elf.scns_last = &result->state.elf32.scns;
1038       result->state.elf32.scns.max = NSCNSALLOC;
1039     }
1040 
1041   return result;
1042 }
1043 
1044 /* Lock if necessary before dup an archive.  */
1045 static inline Elf *
lock_dup_elf(int fildes,Elf_Cmd cmd,Elf * ref)1046 lock_dup_elf (int fildes, Elf_Cmd cmd, Elf *ref)
1047 {
1048   /* We need wrlock to dup an archive.  */
1049   if (ref->kind == ELF_K_AR)
1050     {
1051       rwlock_unlock (ref->lock);
1052       rwlock_wrlock (ref->lock);
1053     }
1054     /* Duplicate the descriptor.  */
1055   return dup_elf (fildes, cmd, ref);
1056 }
1057 
1058 /* Return a descriptor for the file belonging to FILDES.  */
1059 Elf *
elf_begin(int fildes,Elf_Cmd cmd,Elf * ref)1060 elf_begin (int fildes, Elf_Cmd cmd, Elf *ref)
1061 {
1062   Elf *retval;
1063 
1064   if (unlikely (! __libelf_version_initialized))
1065     {
1066       /* Version wasn't set so far.  */
1067       __libelf_seterrno (ELF_E_NO_VERSION);
1068       return NULL;
1069     }
1070 
1071   if (ref != NULL)
1072     /* Make sure the descriptor is not suddenly going away.  */
1073     rwlock_rdlock (ref->lock);
1074   else if (unlikely (fcntl (fildes, F_GETFL) == -1 && errno == EBADF))
1075     {
1076       /* We cannot do anything productive without a file descriptor.  */
1077       __libelf_seterrno (ELF_E_INVALID_FILE);
1078       return NULL;
1079     }
1080 
1081   switch (cmd)
1082     {
1083     case ELF_C_NULL:
1084       /* We simply return a NULL pointer.  */
1085       retval = NULL;
1086       break;
1087 
1088     case ELF_C_READ_MMAP_PRIVATE:
1089       /* If we have a reference it must also be opened this way.  */
1090       if (unlikely (ref != NULL && ref->cmd != ELF_C_READ_MMAP_PRIVATE))
1091 	{
1092 	  __libelf_seterrno (ELF_E_INVALID_CMD);
1093 	  retval = NULL;
1094 	  break;
1095 	}
1096       /* FALLTHROUGH */
1097 
1098     case ELF_C_READ:
1099     case ELF_C_READ_MMAP:
1100       if (ref != NULL)
1101 	retval = lock_dup_elf (fildes, cmd, ref);
1102       else
1103 	/* Create descriptor for existing file.  */
1104 	retval = read_file (fildes, 0, ~((size_t) 0), cmd, NULL);
1105       break;
1106 
1107     case ELF_C_RDWR:
1108     case ELF_C_RDWR_MMAP:
1109       /* If we have a REF object it must also be opened using this
1110 	 command.  */
1111       if (ref != NULL)
1112 	{
1113 	  if (unlikely (ref->cmd != ELF_C_RDWR && ref->cmd != ELF_C_RDWR_MMAP
1114 			&& ref->cmd != ELF_C_WRITE
1115 			&& ref->cmd != ELF_C_WRITE_MMAP))
1116 	    {
1117 	      /* This is not ok.  REF must also be opened for writing.  */
1118 	      __libelf_seterrno (ELF_E_INVALID_CMD);
1119 	      retval = NULL;
1120 	    }
1121 	  else
1122 	    retval = lock_dup_elf (fildes, cmd, ref);
1123 	}
1124       else
1125 	/* Create descriptor for existing file.  */
1126 	retval = read_file (fildes, 0, ~((size_t) 0), cmd, NULL);
1127       break;
1128 
1129     case ELF_C_WRITE:
1130     case ELF_C_WRITE_MMAP:
1131       /* We ignore REF and prepare a descriptor to write a new file.  */
1132       retval = write_file (fildes, cmd);
1133       break;
1134 
1135     default:
1136       __libelf_seterrno (ELF_E_INVALID_CMD);
1137       retval = NULL;
1138       break;
1139     }
1140 
1141   /* Release the lock.  */
1142   if (ref != NULL)
1143     rwlock_unlock (ref->lock);
1144 
1145   return retval;
1146 }
1147 INTDEF(elf_begin)
1148