1 /* Machine-dependent ELF dynamic relocation functions. PowerPC version.
2 Copyright (C) 1995-2006, 2008, 2011 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <http://www.gnu.org/licenses/>. */
21 #include <sys/param.h>
24 #include <elf/dynamic-link.h>
25 #include <dl-machine.h>
28 /* The value __cache_line_size is defined in dl-sysdep.c and is initialised
29 by _dl_sysdep_start via DL_PLATFORM_INIT. */
30 extern int __cache_line_size attribute_hidden
;
32 /* Because ld.so is now versioned, these functions can be in their own file;
33 no relocations need to be done to call them.
34 Of course, if ld.so is not versioned... */
35 #if defined SHARED && !(DO_VERSIONING - 0)
36 #error This will not work with versioning turned off, sorry.
40 /* Stuff for the PLT. */
41 #define PLT_INITIAL_ENTRY_WORDS 18
42 #define PLT_LONGBRANCH_ENTRY_WORDS 0
43 #define PLT_TRAMPOLINE_ENTRY_WORDS 6
44 #define PLT_DOUBLE_SIZE (1<<13)
45 #define PLT_ENTRY_START_WORDS(entry_number) \
46 (PLT_INITIAL_ENTRY_WORDS + (entry_number)*2 \
47 + ((entry_number) > PLT_DOUBLE_SIZE \
48 ? ((entry_number) - PLT_DOUBLE_SIZE)*2 \
50 #define PLT_DATA_START_WORDS(num_entries) PLT_ENTRY_START_WORDS(num_entries)
52 /* Macros to build PowerPC opcode words. */
53 #define OPCODE_ADDI(rd,ra,simm) \
54 (0x38000000 | (rd) << 21 | (ra) << 16 | ((simm) & 0xffff))
55 #define OPCODE_ADDIS(rd,ra,simm) \
56 (0x3c000000 | (rd) << 21 | (ra) << 16 | ((simm) & 0xffff))
57 #define OPCODE_ADD(rd,ra,rb) \
58 (0x7c000214 | (rd) << 21 | (ra) << 16 | (rb) << 11)
59 #define OPCODE_B(target) (0x48000000 | ((target) & 0x03fffffc))
60 #define OPCODE_BA(target) (0x48000002 | ((target) & 0x03fffffc))
61 #define OPCODE_BCTR() 0x4e800420
62 #define OPCODE_LWZ(rd,d,ra) \
63 (0x80000000 | (rd) << 21 | (ra) << 16 | ((d) & 0xffff))
64 #define OPCODE_LWZU(rd,d,ra) \
65 (0x84000000 | (rd) << 21 | (ra) << 16 | ((d) & 0xffff))
66 #define OPCODE_MTCTR(rd) (0x7C0903A6 | (rd) << 21)
67 #define OPCODE_RLWINM(ra,rs,sh,mb,me) \
68 (0x54000000 | (rs) << 21 | (ra) << 16 | (sh) << 11 | (mb) << 6 | (me) << 1)
70 #define OPCODE_LI(rd,simm) OPCODE_ADDI(rd,0,simm)
71 #define OPCODE_ADDIS_HI(rd,ra,value) \
72 OPCODE_ADDIS(rd,ra,((value) + 0x8000) >> 16)
73 #define OPCODE_LIS_HI(rd,value) OPCODE_ADDIS_HI(rd,0,value)
74 #define OPCODE_SLWI(ra,rs,sh) OPCODE_RLWINM(ra,rs,sh,0,31-sh)
77 #define PPC_DCBST(where) asm volatile ("dcbst 0,%0" : : "r"(where) : "memory")
78 #define PPC_SYNC asm volatile ("sync" : : : "memory")
79 #define PPC_ISYNC asm volatile ("sync; isync" : : : "memory")
80 #define PPC_ICBI(where) asm volatile ("icbi 0,%0" : : "r"(where) : "memory")
81 #define PPC_DIE asm volatile ("tweq 0,0")
83 /* Use this when you've modified some code, but it won't be in the
84 instruction fetch queue (or when it doesn't matter if it is). */
85 #define MODIFIED_CODE_NOQUEUE(where) \
86 do { PPC_DCBST(where); PPC_SYNC; PPC_ICBI(where); } while (0)
87 /* Use this when it might be in the instruction queue. */
88 #define MODIFIED_CODE(where) \
89 do { PPC_DCBST(where); PPC_SYNC; PPC_ICBI(where); PPC_ISYNC; } while (0)
92 /* The idea here is that to conform to the ABI, we are supposed to try
93 to load dynamic objects between 0x10000 (we actually use 0x40000 as
94 the lower bound, to increase the chance of a memory reference from
95 a null pointer giving a segfault) and the program's load address;
96 this may allow us to use a branch instruction in the PLT rather
97 than a computed jump. The address is only used as a preference for
98 mmap, so if we get it wrong the worst that happens is that it gets
99 mapped somewhere else. */
102 __elf_preferred_address (struct link_map
*loader
, size_t maplength
,
103 ElfW(Addr
) mapstartpref
)
105 ElfW(Addr
) low
, high
;
109 /* If the object has a preference, load it there! */
110 if (mapstartpref
!= 0)
113 /* Otherwise, quickly look for a suitable gap between 0x3FFFF and
114 0x70000000. 0x3FFFF is so that references off NULL pointers will
115 cause a segfault, 0x70000000 is just paranoia (it should always
116 be superceded by the program's load address). */
119 for (nsid
= 0; nsid
< DL_NNS
; ++nsid
)
120 for (l
= GL(dl_ns
)[nsid
]._ns_loaded
; l
; l
= l
->l_next
)
122 ElfW(Addr
) mapstart
, mapend
;
123 mapstart
= l
->l_map_start
& ~(GLRO(dl_pagesize
) - 1);
124 mapend
= l
->l_map_end
| (GLRO(dl_pagesize
) - 1);
125 assert (mapend
> mapstart
);
127 /* Prefer gaps below the main executable, note that l ==
128 _dl_loaded does not work for static binaries loading
130 if ((mapend
>= high
|| l
->l_type
== lt_executable
)
133 else if (mapend
>= low
&& low
>= mapstart
)
135 else if (high
>= mapend
&& mapstart
>= low
)
137 if (high
- mapend
>= mapstart
- low
)
144 high
-= 0x10000; /* Allow some room between objects. */
145 maplength
= (maplength
| (GLRO(dl_pagesize
) - 1)) + 1;
146 if (high
<= low
|| high
- low
< maplength
)
148 return high
- maplength
; /* Both high and maplength are page-aligned. */
151 /* Set up the loaded object described by L so its unrelocated PLT
152 entries will jump to the on-demand fixup code in dl-runtime.c.
153 Also install a small trampoline to be used by entries that have
154 been relocated to an address too far away for a single branch. */
156 /* There are many kinds of PLT entries:
158 (1) A direct jump to the actual routine, either a relative or
159 absolute branch. These are set up in __elf_machine_fixup_plt.
161 (2) Short lazy entries. These cover the first 8192 slots in
162 the PLT, and look like (where 'index' goes from 0 to 8191):
165 b &plt[PLT_TRAMPOLINE_ENTRY_WORDS+1]
167 (3) Short indirect jumps. These replace (2) when a direct jump
168 wouldn't reach. They look the same except that the branch
169 is 'b &plt[PLT_LONGBRANCH_ENTRY_WORDS]'.
171 (4) Long lazy entries. These cover the slots when a short entry
172 won't fit ('index*4' overflows its field), and look like:
174 lis %r11, %hi(index*4 + &plt[PLT_DATA_START_WORDS])
175 lwzu %r12, %r11, %lo(index*4 + &plt[PLT_DATA_START_WORDS])
176 b &plt[PLT_TRAMPOLINE_ENTRY_WORDS]
179 (5) Long indirect jumps. These replace (4) when a direct jump
180 wouldn't reach. They look like:
182 lis %r11, %hi(index*4 + &plt[PLT_DATA_START_WORDS])
183 lwz %r12, %r11, %lo(index*4 + &plt[PLT_DATA_START_WORDS])
187 (6) Long direct jumps. These are used when thread-safety is not
188 required. They look like:
190 lis %r12, %hi(finaladdr)
191 addi %r12, %r12, %lo(finaladdr)
196 The lazy entries, (2) and (4), are set up here in
197 __elf_machine_runtime_setup. (1), (3), and (5) are set up in
198 __elf_machine_fixup_plt. (1), (3), and (6) can also be constructed
199 in __process_machine_rela.
201 The reason for the somewhat strange construction of the long
202 entries, (4) and (5), is that we need to ensure thread-safety. For
203 (1) and (3), this is obvious because only one instruction is
204 changed and the PPC architecture guarantees that aligned stores are
205 atomic. For (5), this is more tricky. When changing (4) to (5),
206 the `b' instruction is first changed to `mtctr'; this is safe
207 and is why the `lwzu' instruction is not just a simple `addi'.
208 Once this is done, and is visible to all processors, the `lwzu' can
209 safely be changed to a `lwz'. */
211 __elf_machine_runtime_setup (struct link_map
*map
, int lazy
, int profile
)
213 if (map
->l_info
[DT_JMPREL
])
216 Elf32_Word
*plt
= (Elf32_Word
*) D_PTR (map
, l_info
[DT_PLTGOT
]);
217 Elf32_Word num_plt_entries
= (map
->l_info
[DT_PLTRELSZ
]->d_un
.d_val
218 / sizeof (Elf32_Rela
));
219 Elf32_Word rel_offset_words
= PLT_DATA_START_WORDS (num_plt_entries
);
220 Elf32_Word data_words
= (Elf32_Word
) (plt
+ rel_offset_words
);
221 Elf32_Word size_modified
;
223 extern void _dl_runtime_resolve (void);
224 extern void _dl_prof_resolve (void);
226 /* Convert the index in r11 into an actual address, and get the
227 word at that address. */
228 plt
[PLT_LONGBRANCH_ENTRY_WORDS
] = OPCODE_ADDIS_HI (11, 11, data_words
);
229 plt
[PLT_LONGBRANCH_ENTRY_WORDS
+ 1] = OPCODE_LWZ (11, data_words
, 11);
231 /* Call the procedure at that address. */
232 plt
[PLT_LONGBRANCH_ENTRY_WORDS
+ 2] = OPCODE_MTCTR (11);
233 plt
[PLT_LONGBRANCH_ENTRY_WORDS
+ 3] = OPCODE_BCTR ();
237 Elf32_Word
*tramp
= plt
+ PLT_TRAMPOLINE_ENTRY_WORDS
;
238 Elf32_Word dlrr
= (Elf32_Word
)(profile
240 : _dl_runtime_resolve
);
243 if (profile
&& GLRO(dl_profile
) != NULL
244 && _dl_name_match_p (GLRO(dl_profile
), map
))
245 /* This is the object we are looking for. Say that we really
246 want profiling and the timers are started. */
247 GL(dl_profile_map
) = map
;
249 /* For the long entries, subtract off data_words. */
250 tramp
[0] = OPCODE_ADDIS_HI (11, 11, -data_words
);
251 tramp
[1] = OPCODE_ADDI (11, 11, -data_words
);
253 /* Multiply index of entry by 3 (in r11). */
254 tramp
[2] = OPCODE_SLWI (12, 11, 1);
255 tramp
[3] = OPCODE_ADD (11, 12, 11);
256 if (dlrr
<= 0x01fffffc || dlrr
>= 0xfe000000)
258 /* Load address of link map in r12. */
259 tramp
[4] = OPCODE_LI (12, (Elf32_Word
) map
);
260 tramp
[5] = OPCODE_ADDIS_HI (12, 12, (Elf32_Word
) map
);
262 /* Call _dl_runtime_resolve. */
263 tramp
[6] = OPCODE_BA (dlrr
);
267 /* Get address of _dl_runtime_resolve in CTR. */
268 tramp
[4] = OPCODE_LI (12, dlrr
);
269 tramp
[5] = OPCODE_ADDIS_HI (12, 12, dlrr
);
270 tramp
[6] = OPCODE_MTCTR (12);
272 /* Load address of link map in r12. */
273 tramp
[7] = OPCODE_LI (12, (Elf32_Word
) map
);
274 tramp
[8] = OPCODE_ADDIS_HI (12, 12, (Elf32_Word
) map
);
276 /* Call _dl_runtime_resolve. */
277 tramp
[9] = OPCODE_BCTR ();
280 /* Set up the lazy PLT entries. */
281 offset
= PLT_INITIAL_ENTRY_WORDS
;
283 while (i
< num_plt_entries
&& i
< PLT_DOUBLE_SIZE
)
285 plt
[offset
] = OPCODE_LI (11, i
* 4);
286 plt
[offset
+1] = OPCODE_B ((PLT_TRAMPOLINE_ENTRY_WORDS
+ 2
292 while (i
< num_plt_entries
)
294 plt
[offset
] = OPCODE_LIS_HI (11, i
* 4 + data_words
);
295 plt
[offset
+1] = OPCODE_LWZU (12, i
* 4 + data_words
, 11);
296 plt
[offset
+2] = OPCODE_B ((PLT_TRAMPOLINE_ENTRY_WORDS
299 plt
[offset
+3] = OPCODE_BCTR ();
305 /* Now, we've modified code. We need to write the changes from
306 the data cache to a second-level unified cache, then make
307 sure that stale data in the instruction cache is removed.
308 (In a multiprocessor system, the effect is more complex.)
309 Most of the PLT shouldn't be in the instruction cache, but
310 there may be a little overlap at the start and the end.
312 Assumes that dcbst and icbi apply to lines of 16 bytes or
313 more. Current known line sizes are 16, 32, and 128 bytes.
314 The following gets the __cache_line_size, when available. */
316 /* Default minimum 4 words per cache line. */
317 int line_size_words
= 4;
319 if (lazy
&& __cache_line_size
!= 0)
320 /* Convert bytes to words. */
321 line_size_words
= __cache_line_size
/ 4;
323 size_modified
= lazy
? rel_offset_words
: 6;
324 for (i
= 0; i
< size_modified
; i
+= line_size_words
)
326 PPC_DCBST (plt
+ size_modified
- 1);
329 for (i
= 0; i
< size_modified
; i
+= line_size_words
)
331 PPC_ICBI (plt
+ size_modified
- 1);
339 __elf_machine_fixup_plt (struct link_map
*map
,
340 Elf32_Addr
*reloc_addr
, Elf32_Addr finaladdr
)
342 Elf32_Sword delta
= finaladdr
- (Elf32_Word
) reloc_addr
;
343 if (delta
<< 6 >> 6 == delta
)
344 *reloc_addr
= OPCODE_B (delta
);
345 else if (finaladdr
<= 0x01fffffc || finaladdr
>= 0xfe000000)
346 *reloc_addr
= OPCODE_BA (finaladdr
);
349 Elf32_Word
*plt
, *data_words
;
350 Elf32_Word index
, offset
, num_plt_entries
;
352 num_plt_entries
= (map
->l_info
[DT_PLTRELSZ
]->d_un
.d_val
353 / sizeof(Elf32_Rela
));
354 plt
= (Elf32_Word
*) D_PTR (map
, l_info
[DT_PLTGOT
]);
355 offset
= reloc_addr
- plt
;
356 index
= (offset
- PLT_INITIAL_ENTRY_WORDS
)/2;
357 data_words
= plt
+ PLT_DATA_START_WORDS (num_plt_entries
);
361 if (index
< PLT_DOUBLE_SIZE
)
363 data_words
[index
] = finaladdr
;
365 *reloc_addr
= OPCODE_B ((PLT_LONGBRANCH_ENTRY_WORDS
- (offset
+1))
370 index
-= (index
- PLT_DOUBLE_SIZE
)/2;
372 data_words
[index
] = finaladdr
;
375 reloc_addr
[1] = OPCODE_MTCTR (12);
376 MODIFIED_CODE_NOQUEUE (reloc_addr
+ 1);
379 reloc_addr
[0] = OPCODE_LWZ (12,
380 (Elf32_Word
) (data_words
+ index
), 11);
383 MODIFIED_CODE (reloc_addr
);
388 _dl_reloc_overflow (struct link_map
*map
,
390 Elf32_Addr
*const reloc_addr
,
391 const Elf32_Sym
*refsym
)
395 t
= stpcpy (buffer
, name
);
396 t
= stpcpy (t
, " relocation at 0x00000000");
397 _itoa_word ((unsigned) reloc_addr
, t
, 16, 0);
402 strtab
= (const void *) D_PTR (map
, l_info
[DT_STRTAB
]);
403 t
= stpcpy (t
, " for symbol `");
404 t
= stpcpy (t
, strtab
+ refsym
->st_name
);
407 t
= stpcpy (t
, " out of range");
408 _dl_signal_error (0, map
->l_name
, NULL
, buffer
);
412 __process_machine_rela (struct link_map
*map
,
413 const Elf32_Rela
*reloc
,
414 struct link_map
*sym_map
,
415 const Elf32_Sym
*sym
,
416 const Elf32_Sym
*refsym
,
417 Elf32_Addr
*const reloc_addr
,
418 Elf32_Addr
const finaladdr
,
429 *reloc_addr
= finaladdr
;
432 case R_PPC_IRELATIVE
:
433 *reloc_addr
= ((Elf32_Addr (*) (void)) finaladdr
) ();
437 ((char *) reloc_addr
)[0] = finaladdr
>> 24;
438 ((char *) reloc_addr
)[1] = finaladdr
>> 16;
439 ((char *) reloc_addr
)[2] = finaladdr
>> 8;
440 ((char *) reloc_addr
)[3] = finaladdr
;
444 if (__builtin_expect (finaladdr
> 0x01fffffc && finaladdr
< 0xfe000000, 0))
445 _dl_reloc_overflow (map
, "R_PPC_ADDR24", reloc_addr
, refsym
);
446 *reloc_addr
= (*reloc_addr
& 0xfc000003) | (finaladdr
& 0x3fffffc);
450 if (__builtin_expect (finaladdr
> 0x7fff && finaladdr
< 0xffff8000, 0))
451 _dl_reloc_overflow (map
, "R_PPC_ADDR16", reloc_addr
, refsym
);
452 *(Elf32_Half
*) reloc_addr
= finaladdr
;
456 if (__builtin_expect (finaladdr
> 0x7fff && finaladdr
< 0xffff8000, 0))
457 _dl_reloc_overflow (map
, "R_PPC_UADDR16", reloc_addr
, refsym
);
458 ((char *) reloc_addr
)[0] = finaladdr
>> 8;
459 ((char *) reloc_addr
)[1] = finaladdr
;
462 case R_PPC_ADDR16_LO
:
463 *(Elf32_Half
*) reloc_addr
= finaladdr
;
466 case R_PPC_ADDR16_HI
:
467 *(Elf32_Half
*) reloc_addr
= finaladdr
>> 16;
470 case R_PPC_ADDR16_HA
:
471 *(Elf32_Half
*) reloc_addr
= (finaladdr
+ 0x8000) >> 16;
475 case R_PPC_ADDR14_BRTAKEN
:
476 case R_PPC_ADDR14_BRNTAKEN
:
477 if (__builtin_expect (finaladdr
> 0x7fff && finaladdr
< 0xffff8000, 0))
478 _dl_reloc_overflow (map
, "R_PPC_ADDR14", reloc_addr
, refsym
);
479 *reloc_addr
= (*reloc_addr
& 0xffff0003) | (finaladdr
& 0xfffc);
480 if (rinfo
!= R_PPC_ADDR14
)
481 *reloc_addr
= ((*reloc_addr
& 0xffdfffff)
482 | ((rinfo
== R_PPC_ADDR14_BRTAKEN
)
483 ^ (finaladdr
>> 31)) << 21);
488 Elf32_Sword delta
= finaladdr
- (Elf32_Word
) reloc_addr
;
489 if (delta
<< 6 >> 6 != delta
)
490 _dl_reloc_overflow (map
, "R_PPC_REL24", reloc_addr
, refsym
);
491 *reloc_addr
= (*reloc_addr
& 0xfc000003) | (delta
& 0x3fffffc);
497 /* This can happen in trace mode when an object could not be
500 if (sym
->st_size
> refsym
->st_size
501 || (GLRO(dl_verbose
) && sym
->st_size
< refsym
->st_size
))
505 strtab
= (const void *) D_PTR (map
, l_info
[DT_STRTAB
]);
507 %s: Symbol `%s' has different size in shared object, consider re-linking\n",
508 rtld_progname
?: "<program name unknown>",
509 strtab
+ refsym
->st_name
);
511 memcpy (reloc_addr
, (char *) finaladdr
, MIN (sym
->st_size
,
516 *reloc_addr
= finaladdr
- (Elf32_Word
) reloc_addr
;
520 /* It used to be that elf_machine_fixup_plt was used here,
521 but that doesn't work when ld.so relocates itself
522 for the second time. On the bright side, there's
523 no need to worry about thread-safety here. */
525 Elf32_Sword delta
= finaladdr
- (Elf32_Word
) reloc_addr
;
526 if (delta
<< 6 >> 6 == delta
)
527 *reloc_addr
= OPCODE_B (delta
);
528 else if (finaladdr
<= 0x01fffffc || finaladdr
>= 0xfe000000)
529 *reloc_addr
= OPCODE_BA (finaladdr
);
532 Elf32_Word
*plt
, *data_words
;
533 Elf32_Word index
, offset
, num_plt_entries
;
535 plt
= (Elf32_Word
*) D_PTR (map
, l_info
[DT_PLTGOT
]);
536 offset
= reloc_addr
- plt
;
538 if (offset
< PLT_DOUBLE_SIZE
*2 + PLT_INITIAL_ENTRY_WORDS
)
540 index
= (offset
- PLT_INITIAL_ENTRY_WORDS
)/2;
541 num_plt_entries
= (map
->l_info
[DT_PLTRELSZ
]->d_un
.d_val
542 / sizeof(Elf32_Rela
));
543 data_words
= plt
+ PLT_DATA_START_WORDS (num_plt_entries
);
544 data_words
[index
] = finaladdr
;
545 reloc_addr
[0] = OPCODE_LI (11, index
* 4);
546 reloc_addr
[1] = OPCODE_B ((PLT_LONGBRANCH_ENTRY_WORDS
549 MODIFIED_CODE_NOQUEUE (reloc_addr
+ 1);
553 reloc_addr
[0] = OPCODE_LIS_HI (12, finaladdr
);
554 reloc_addr
[1] = OPCODE_ADDI (12, 12, finaladdr
);
555 reloc_addr
[2] = OPCODE_MTCTR (12);
556 reloc_addr
[3] = OPCODE_BCTR ();
557 MODIFIED_CODE_NOQUEUE (reloc_addr
+ 3);
563 #define DO_TLS_RELOC(suffix) \
564 case R_PPC_DTPREL##suffix: \
565 /* During relocation all TLS symbols are defined and used. \
566 Therefore the offset is already correct. */ \
567 if (sym_map != NULL) \
568 do_reloc##suffix ("R_PPC_DTPREL"#suffix, \
569 TLS_DTPREL_VALUE (sym, reloc)); \
571 case R_PPC_TPREL##suffix: \
572 if (sym_map != NULL) \
574 CHECK_STATIC_TLS (map, sym_map); \
575 do_reloc##suffix ("R_PPC_TPREL"#suffix, \
576 TLS_TPREL_VALUE (sym_map, sym, reloc)); \
580 inline void do_reloc16 (const char *r_name
, Elf32_Addr value
)
582 if (__builtin_expect (value
> 0x7fff && value
< 0xffff8000, 0))
583 _dl_reloc_overflow (map
, r_name
, reloc_addr
, refsym
);
584 *(Elf32_Half
*) reloc_addr
= value
;
586 inline void do_reloc16_LO (const char *r_name
, Elf32_Addr value
)
588 *(Elf32_Half
*) reloc_addr
= value
;
590 inline void do_reloc16_HI (const char *r_name
, Elf32_Addr value
)
592 *(Elf32_Half
*) reloc_addr
= value
>> 16;
594 inline void do_reloc16_HA (const char *r_name
, Elf32_Addr value
)
596 *(Elf32_Half
*) reloc_addr
= (value
+ 0x8000) >> 16;
604 _dl_reloc_bad_type (map
, rinfo
, 0);
608 MODIFIED_CODE_NOQUEUE (reloc_addr
);