1 /*
2 * recordmcount.c: construct a table of the locations of calls to 'mcount'
3 * so that ftrace can find them quickly.
4 * Copyright 2009 John F. Reiser <jreiser@BitWagon.com>. All rights reserved.
5 * Licensed under the GNU General Public License, version 2 (GPLv2).
6 *
7 * Restructured to fit Linux format, as well as other updates:
8 * Copyright 2010 Steven Rostedt <srostedt@redhat.com>, Red Hat Inc.
9 */
10
11 /*
12 * Strategy: alter the .o file in-place.
13 *
14 * Append a new STRTAB that has the new section names, followed by a new array
15 * ElfXX_Shdr[] that has the new section headers, followed by the section
16 * contents for __mcount_loc and its relocations. The old shstrtab strings,
17 * and the old ElfXX_Shdr[] array, remain as "garbage" (commonly, a couple
18 * kilobytes.) Subsequent processing by /bin/ld (or the kernel module loader)
19 * will ignore the garbage regions, because they are not designated by the
20 * new .e_shoff nor the new ElfXX_Shdr[]. [In order to remove the garbage,
21 * then use "ld -r" to create a new file that omits the garbage.]
22 */
23
24 #include <sys/types.h>
25 #include <sys/mman.h>
26 #include <sys/stat.h>
27 #include <getopt.h>
28 #include <elf.h>
29 #include <fcntl.h>
30 #include <setjmp.h>
31 #include <stdio.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include <unistd.h>
35
36 /*
37 * glibc synced up and added the metag number but didn't add the relocations.
38 * Work around this in a crude manner for now.
39 */
40 #ifndef EM_METAG
41 #define EM_METAG 174
42 #endif
43 #ifndef R_METAG_ADDR32
44 #define R_METAG_ADDR32 2
45 #endif
46 #ifndef R_METAG_NONE
47 #define R_METAG_NONE 3
48 #endif
49
50 #ifndef EM_AARCH64
51 #define EM_AARCH64 183
52 #define R_AARCH64_NONE 0
53 #define R_AARCH64_ABS64 257
54 #endif
55
56 static int fd_map; /* File descriptor for file being modified. */
57 static int mmap_failed; /* Boolean flag. */
58 static char gpfx; /* prefix for global symbol name (sometimes '_') */
59 static struct stat sb; /* Remember .st_size, etc. */
60 static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */
61 static const char *altmcount; /* alternate mcount symbol name */
62 static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */
63 static void *file_map; /* pointer of the mapped file */
64 static void *file_end; /* pointer to the end of the mapped file */
65 static int file_updated; /* flag to state file was changed */
66 static void *file_ptr; /* current file pointer location */
67 static void *file_append; /* added to the end of the file */
68 static size_t file_append_size; /* how much is added to end of file */
69
70 /* setjmp() return values */
71 enum {
72 SJ_SETJMP = 0, /* hardwired first return */
73 SJ_FAIL,
74 SJ_SUCCEED
75 };
76
77 /* Per-file resource cleanup when multiple files. */
78 static void
cleanup(void)79 cleanup(void)
80 {
81 if (!mmap_failed)
82 munmap(file_map, sb.st_size);
83 else
84 free(file_map);
85 file_map = NULL;
86 free(file_append);
87 file_append = NULL;
88 file_append_size = 0;
89 file_updated = 0;
90 }
91
92 static void __attribute__((noreturn))
fail_file(void)93 fail_file(void)
94 {
95 cleanup();
96 longjmp(jmpenv, SJ_FAIL);
97 }
98
99 static void __attribute__((noreturn))
succeed_file(void)100 succeed_file(void)
101 {
102 cleanup();
103 longjmp(jmpenv, SJ_SUCCEED);
104 }
105
106 /* ulseek, uread, ...: Check return value for errors. */
107
108 static off_t
ulseek(int const fd,off_t const offset,int const whence)109 ulseek(int const fd, off_t const offset, int const whence)
110 {
111 switch (whence) {
112 case SEEK_SET:
113 file_ptr = file_map + offset;
114 break;
115 case SEEK_CUR:
116 file_ptr += offset;
117 break;
118 case SEEK_END:
119 file_ptr = file_map + (sb.st_size - offset);
120 break;
121 }
122 if (file_ptr < file_map) {
123 fprintf(stderr, "lseek: seek before file\n");
124 fail_file();
125 }
126 return file_ptr - file_map;
127 }
128
129 static size_t
uread(int const fd,void * const buf,size_t const count)130 uread(int const fd, void *const buf, size_t const count)
131 {
132 size_t const n = read(fd, buf, count);
133 if (n != count) {
134 perror("read");
135 fail_file();
136 }
137 return n;
138 }
139
140 static size_t
uwrite(int const fd,void const * const buf,size_t const count)141 uwrite(int const fd, void const *const buf, size_t const count)
142 {
143 size_t cnt = count;
144 off_t idx = 0;
145
146 file_updated = 1;
147
148 if (file_ptr + count >= file_end) {
149 off_t aoffset = (file_ptr + count) - file_end;
150
151 if (aoffset > file_append_size) {
152 file_append = realloc(file_append, aoffset);
153 file_append_size = aoffset;
154 }
155 if (!file_append) {
156 perror("write");
157 fail_file();
158 }
159 if (file_ptr < file_end) {
160 cnt = file_end - file_ptr;
161 } else {
162 cnt = 0;
163 idx = aoffset - count;
164 }
165 }
166
167 if (cnt)
168 memcpy(file_ptr, buf, cnt);
169
170 if (cnt < count)
171 memcpy(file_append + idx, buf + cnt, count - cnt);
172
173 file_ptr += count;
174 return count;
175 }
176
177 static void *
umalloc(size_t size)178 umalloc(size_t size)
179 {
180 void *const addr = malloc(size);
181 if (addr == 0) {
182 fprintf(stderr, "malloc failed: %zu bytes\n", size);
183 fail_file();
184 }
185 return addr;
186 }
187
188 static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
189 static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
190 static unsigned char *ideal_nop;
191
192 static char rel_type_nop;
193
194 static int (*make_nop)(void *map, size_t const offset);
195
make_nop_x86(void * map,size_t const offset)196 static int make_nop_x86(void *map, size_t const offset)
197 {
198 uint32_t *ptr;
199 unsigned char *op;
200
201 /* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */
202 ptr = map + offset;
203 if (*ptr != 0)
204 return -1;
205
206 op = map + offset - 1;
207 if (*op != 0xe8)
208 return -1;
209
210 /* convert to nop */
211 ulseek(fd_map, offset - 1, SEEK_SET);
212 uwrite(fd_map, ideal_nop, 5);
213 return 0;
214 }
215
216 static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
make_nop_arm64(void * map,size_t const offset)217 static int make_nop_arm64(void *map, size_t const offset)
218 {
219 uint32_t *ptr;
220
221 ptr = map + offset;
222 /* bl <_mcount> is 0x94000000 before relocation */
223 if (*ptr != 0x94000000)
224 return -1;
225
226 /* Convert to nop */
227 ulseek(fd_map, offset, SEEK_SET);
228 uwrite(fd_map, ideal_nop, 4);
229 return 0;
230 }
231
232 /*
233 * Get the whole file as a programming convenience in order to avoid
234 * malloc+lseek+read+free of many pieces. If successful, then mmap
235 * avoids copying unused pieces; else just read the whole file.
236 * Open for both read and write; new info will be appended to the file.
237 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
238 * do not propagate to the file until an explicit overwrite at the last.
239 * This preserves most aspects of consistency (all except .st_size)
240 * for simultaneous readers of the file while we are appending to it.
241 * However, multiple writers still are bad. We choose not to use
242 * locking because it is expensive and the use case of kernel build
243 * makes multiple writers unlikely.
244 */
mmap_file(char const * fname)245 static void *mmap_file(char const *fname)
246 {
247 fd_map = open(fname, O_RDONLY);
248 if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
249 perror(fname);
250 fail_file();
251 }
252 if (!S_ISREG(sb.st_mode)) {
253 fprintf(stderr, "not a regular file: %s\n", fname);
254 fail_file();
255 }
256 file_map = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
257 fd_map, 0);
258 mmap_failed = 0;
259 if (file_map == MAP_FAILED) {
260 mmap_failed = 1;
261 file_map = umalloc(sb.st_size);
262 uread(fd_map, file_map, sb.st_size);
263 }
264 close(fd_map);
265
266 file_end = file_map + sb.st_size;
267
268 return file_map;
269 }
270
write_file(const char * fname)271 static void write_file(const char *fname)
272 {
273 char tmp_file[strlen(fname) + 4];
274 size_t n;
275
276 if (!file_updated)
277 return;
278
279 sprintf(tmp_file, "%s.rc", fname);
280
281 /*
282 * After reading the entire file into memory, delete it
283 * and write it back, to prevent weird side effects of modifying
284 * an object file in place.
285 */
286 fd_map = open(tmp_file, O_WRONLY | O_TRUNC | O_CREAT, sb.st_mode);
287 if (fd_map < 0) {
288 perror(fname);
289 fail_file();
290 }
291 n = write(fd_map, file_map, sb.st_size);
292 if (n != sb.st_size) {
293 perror("write");
294 fail_file();
295 }
296 if (file_append_size) {
297 n = write(fd_map, file_append, file_append_size);
298 if (n != file_append_size) {
299 perror("write");
300 fail_file();
301 }
302 }
303 close(fd_map);
304 if (rename(tmp_file, fname) < 0) {
305 perror(fname);
306 fail_file();
307 }
308 }
309
310 /* w8rev, w8nat, ...: Handle endianness. */
311
w8rev(uint64_t const x)312 static uint64_t w8rev(uint64_t const x)
313 {
314 return ((0xff & (x >> (0 * 8))) << (7 * 8))
315 | ((0xff & (x >> (1 * 8))) << (6 * 8))
316 | ((0xff & (x >> (2 * 8))) << (5 * 8))
317 | ((0xff & (x >> (3 * 8))) << (4 * 8))
318 | ((0xff & (x >> (4 * 8))) << (3 * 8))
319 | ((0xff & (x >> (5 * 8))) << (2 * 8))
320 | ((0xff & (x >> (6 * 8))) << (1 * 8))
321 | ((0xff & (x >> (7 * 8))) << (0 * 8));
322 }
323
w4rev(uint32_t const x)324 static uint32_t w4rev(uint32_t const x)
325 {
326 return ((0xff & (x >> (0 * 8))) << (3 * 8))
327 | ((0xff & (x >> (1 * 8))) << (2 * 8))
328 | ((0xff & (x >> (2 * 8))) << (1 * 8))
329 | ((0xff & (x >> (3 * 8))) << (0 * 8));
330 }
331
w2rev(uint16_t const x)332 static uint32_t w2rev(uint16_t const x)
333 {
334 return ((0xff & (x >> (0 * 8))) << (1 * 8))
335 | ((0xff & (x >> (1 * 8))) << (0 * 8));
336 }
337
w8nat(uint64_t const x)338 static uint64_t w8nat(uint64_t const x)
339 {
340 return x;
341 }
342
w4nat(uint32_t const x)343 static uint32_t w4nat(uint32_t const x)
344 {
345 return x;
346 }
347
w2nat(uint16_t const x)348 static uint32_t w2nat(uint16_t const x)
349 {
350 return x;
351 }
352
353 static uint64_t (*w8)(uint64_t);
354 static uint32_t (*w)(uint32_t);
355 static uint32_t (*w2)(uint16_t);
356
357 /* Names of the sections that could contain calls to mcount. */
358 static int
is_mcounted_section_name(char const * const txtname)359 is_mcounted_section_name(char const *const txtname)
360 {
361 return strcmp(".text", txtname) == 0 ||
362 strcmp(".ref.text", txtname) == 0 ||
363 strcmp(".sched.text", txtname) == 0 ||
364 strcmp(".spinlock.text", txtname) == 0 ||
365 strcmp(".irqentry.text", txtname) == 0 ||
366 strcmp(".softirqentry.text", txtname) == 0 ||
367 strcmp(".kprobes.text", txtname) == 0 ||
368 strcmp(".cpuidle.text", txtname) == 0 ||
369 (strncmp(".text.", txtname, 6) == 0 &&
370 strcmp(".text..ftrace", txtname) != 0);
371 }
372
373 /* 32 bit and 64 bit are very similar */
374 #include "recordmcount.h"
375 #define RECORD_MCOUNT_64
376 #include "recordmcount.h"
377
378 /* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
379 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
380 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
381 * to imply the order of the members; the spec does not say so.
382 * typedef unsigned char Elf64_Byte;
383 * fails on MIPS64 because their <elf.h> already has it!
384 */
385
386 typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */
387
388 union mips_r_info {
389 Elf64_Xword r_info;
390 struct {
391 Elf64_Word r_sym; /* Symbol index. */
392 myElf64_Byte r_ssym; /* Special symbol. */
393 myElf64_Byte r_type3; /* Third relocation. */
394 myElf64_Byte r_type2; /* Second relocation. */
395 myElf64_Byte r_type; /* First relocation. */
396 } r_mips;
397 };
398
MIPS64_r_sym(Elf64_Rel const * rp)399 static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
400 {
401 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
402 }
403
MIPS64_r_info(Elf64_Rel * const rp,unsigned sym,unsigned type)404 static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
405 {
406 rp->r_info = ((union mips_r_info){
407 .r_mips = { .r_sym = w(sym), .r_type = type }
408 }).r_info;
409 }
410
411 static void
do_file(char const * const fname)412 do_file(char const *const fname)
413 {
414 Elf32_Ehdr *const ehdr = mmap_file(fname);
415 unsigned int reltype = 0;
416
417 w = w4nat;
418 w2 = w2nat;
419 w8 = w8nat;
420 switch (ehdr->e_ident[EI_DATA]) {
421 static unsigned int const endian = 1;
422 default:
423 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
424 ehdr->e_ident[EI_DATA], fname);
425 fail_file();
426 break;
427 case ELFDATA2LSB:
428 if (*(unsigned char const *)&endian != 1) {
429 /* main() is big endian, file.o is little endian. */
430 w = w4rev;
431 w2 = w2rev;
432 w8 = w8rev;
433 }
434 break;
435 case ELFDATA2MSB:
436 if (*(unsigned char const *)&endian != 0) {
437 /* main() is little endian, file.o is big endian. */
438 w = w4rev;
439 w2 = w2rev;
440 w8 = w8rev;
441 }
442 break;
443 } /* end switch */
444 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
445 || w2(ehdr->e_type) != ET_REL
446 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
447 fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
448 fail_file();
449 }
450
451 gpfx = 0;
452 switch (w2(ehdr->e_machine)) {
453 default:
454 fprintf(stderr, "unrecognized e_machine %d %s\n",
455 w2(ehdr->e_machine), fname);
456 fail_file();
457 break;
458 case EM_386:
459 reltype = R_386_32;
460 rel_type_nop = R_386_NONE;
461 make_nop = make_nop_x86;
462 ideal_nop = ideal_nop5_x86_32;
463 mcount_adjust_32 = -1;
464 break;
465 case EM_ARM: reltype = R_ARM_ABS32;
466 altmcount = "__gnu_mcount_nc";
467 break;
468 case EM_AARCH64:
469 reltype = R_AARCH64_ABS64;
470 make_nop = make_nop_arm64;
471 rel_type_nop = R_AARCH64_NONE;
472 ideal_nop = ideal_nop4_arm64;
473 gpfx = '_';
474 break;
475 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break;
476 case EM_METAG: reltype = R_METAG_ADDR32;
477 altmcount = "_mcount_wrapper";
478 rel_type_nop = R_METAG_NONE;
479 /* We happen to have the same requirement as MIPS */
480 is_fake_mcount32 = MIPS32_is_fake_mcount;
481 break;
482 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break;
483 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break;
484 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break;
485 case EM_S390: /* reltype: e_class */ gpfx = '_'; break;
486 case EM_SH: reltype = R_SH_DIR32; break;
487 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break;
488 case EM_X86_64:
489 make_nop = make_nop_x86;
490 ideal_nop = ideal_nop5_x86_64;
491 reltype = R_X86_64_64;
492 rel_type_nop = R_X86_64_NONE;
493 mcount_adjust_64 = -1;
494 break;
495 } /* end switch */
496
497 switch (ehdr->e_ident[EI_CLASS]) {
498 default:
499 fprintf(stderr, "unrecognized ELF class %d %s\n",
500 ehdr->e_ident[EI_CLASS], fname);
501 fail_file();
502 break;
503 case ELFCLASS32:
504 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
505 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
506 fprintf(stderr,
507 "unrecognized ET_REL file: %s\n", fname);
508 fail_file();
509 }
510 if (w2(ehdr->e_machine) == EM_MIPS) {
511 reltype = R_MIPS_32;
512 is_fake_mcount32 = MIPS32_is_fake_mcount;
513 }
514 do32(ehdr, fname, reltype);
515 break;
516 case ELFCLASS64: {
517 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
518 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
519 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
520 fprintf(stderr,
521 "unrecognized ET_REL file: %s\n", fname);
522 fail_file();
523 }
524 if (w2(ghdr->e_machine) == EM_S390) {
525 reltype = R_390_64;
526 mcount_adjust_64 = -14;
527 }
528 if (w2(ghdr->e_machine) == EM_MIPS) {
529 reltype = R_MIPS_64;
530 Elf64_r_sym = MIPS64_r_sym;
531 Elf64_r_info = MIPS64_r_info;
532 is_fake_mcount64 = MIPS64_is_fake_mcount;
533 }
534 do64(ghdr, fname, reltype);
535 break;
536 }
537 } /* end switch */
538
539 write_file(fname);
540 cleanup();
541 }
542
543 int
main(int argc,char * argv[])544 main(int argc, char *argv[])
545 {
546 const char ftrace[] = "/ftrace.o";
547 int ftrace_size = sizeof(ftrace) - 1;
548 int n_error = 0; /* gcc-4.3.0 false positive complaint */
549 int c;
550 int i;
551
552 while ((c = getopt(argc, argv, "w")) >= 0) {
553 switch (c) {
554 case 'w':
555 warn_on_notrace_sect = 1;
556 break;
557 default:
558 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
559 return 0;
560 }
561 }
562
563 if ((argc - optind) < 1) {
564 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
565 return 0;
566 }
567
568 /* Process each file in turn, allowing deep failure. */
569 for (i = optind; i < argc; i++) {
570 char *file = argv[i];
571 int const sjval = setjmp(jmpenv);
572 int len;
573
574 /*
575 * The file kernel/trace/ftrace.o references the mcount
576 * function but does not call it. Since ftrace.o should
577 * not be traced anyway, we just skip it.
578 */
579 len = strlen(file);
580 if (len >= ftrace_size &&
581 strcmp(file + (len - ftrace_size), ftrace) == 0)
582 continue;
583
584 switch (sjval) {
585 default:
586 fprintf(stderr, "internal error: %s\n", file);
587 exit(1);
588 break;
589 case SJ_SETJMP: /* normal sequence */
590 /* Avoid problems if early cleanup() */
591 fd_map = -1;
592 mmap_failed = 1;
593 file_map = NULL;
594 file_ptr = NULL;
595 file_updated = 0;
596 do_file(file);
597 break;
598 case SJ_FAIL: /* error in do_file or below */
599 fprintf(stderr, "%s: failed\n", file);
600 ++n_error;
601 break;
602 case SJ_SUCCEED: /* premature success */
603 /* do nothing */
604 break;
605 } /* end switch */
606 }
607 return !!n_error;
608 }
609