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 #define R_ARM_PC24 1
57 #define R_ARM_THM_CALL 10
58 #define R_ARM_CALL 28
59
60 static int fd_map; /* File descriptor for file being modified. */
61 static int mmap_failed; /* Boolean flag. */
62 static char gpfx; /* prefix for global symbol name (sometimes '_') */
63 static struct stat sb; /* Remember .st_size, etc. */
64 static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */
65 static const char *altmcount; /* alternate mcount symbol name */
66 static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */
67 static void *file_map; /* pointer of the mapped file */
68 static void *file_end; /* pointer to the end of the mapped file */
69 static int file_updated; /* flag to state file was changed */
70 static void *file_ptr; /* current file pointer location */
71 static void *file_append; /* added to the end of the file */
72 static size_t file_append_size; /* how much is added to end of file */
73
74 /* setjmp() return values */
75 enum {
76 SJ_SETJMP = 0, /* hardwired first return */
77 SJ_FAIL,
78 SJ_SUCCEED
79 };
80
81 /* Per-file resource cleanup when multiple files. */
82 static void
cleanup(void)83 cleanup(void)
84 {
85 if (!mmap_failed)
86 munmap(file_map, sb.st_size);
87 else
88 free(file_map);
89 file_map = NULL;
90 free(file_append);
91 file_append = NULL;
92 file_append_size = 0;
93 file_updated = 0;
94 }
95
96 static void __attribute__((noreturn))
fail_file(void)97 fail_file(void)
98 {
99 cleanup();
100 longjmp(jmpenv, SJ_FAIL);
101 }
102
103 static void __attribute__((noreturn))
succeed_file(void)104 succeed_file(void)
105 {
106 cleanup();
107 longjmp(jmpenv, SJ_SUCCEED);
108 }
109
110 /* ulseek, uread, ...: Check return value for errors. */
111
112 static off_t
ulseek(int const fd,off_t const offset,int const whence)113 ulseek(int const fd, off_t const offset, int const whence)
114 {
115 switch (whence) {
116 case SEEK_SET:
117 file_ptr = file_map + offset;
118 break;
119 case SEEK_CUR:
120 file_ptr += offset;
121 break;
122 case SEEK_END:
123 file_ptr = file_map + (sb.st_size - offset);
124 break;
125 }
126 if (file_ptr < file_map) {
127 fprintf(stderr, "lseek: seek before file\n");
128 fail_file();
129 }
130 return file_ptr - file_map;
131 }
132
133 static size_t
uread(int const fd,void * const buf,size_t const count)134 uread(int const fd, void *const buf, size_t const count)
135 {
136 size_t const n = read(fd, buf, count);
137 if (n != count) {
138 perror("read");
139 fail_file();
140 }
141 return n;
142 }
143
144 static size_t
uwrite(int const fd,void const * const buf,size_t const count)145 uwrite(int const fd, void const *const buf, size_t const count)
146 {
147 size_t cnt = count;
148 off_t idx = 0;
149
150 file_updated = 1;
151
152 if (file_ptr + count >= file_end) {
153 off_t aoffset = (file_ptr + count) - file_end;
154
155 if (aoffset > file_append_size) {
156 file_append = realloc(file_append, aoffset);
157 file_append_size = aoffset;
158 }
159 if (!file_append) {
160 perror("write");
161 fail_file();
162 }
163 if (file_ptr < file_end) {
164 cnt = file_end - file_ptr;
165 } else {
166 cnt = 0;
167 idx = aoffset - count;
168 }
169 }
170
171 if (cnt)
172 memcpy(file_ptr, buf, cnt);
173
174 if (cnt < count)
175 memcpy(file_append + idx, buf + cnt, count - cnt);
176
177 file_ptr += count;
178 return count;
179 }
180
181 static void *
umalloc(size_t size)182 umalloc(size_t size)
183 {
184 void *const addr = malloc(size);
185 if (addr == 0) {
186 fprintf(stderr, "malloc failed: %zu bytes\n", size);
187 fail_file();
188 }
189 return addr;
190 }
191
192 static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
193 static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
194 static unsigned char *ideal_nop;
195
196 static char rel_type_nop;
197
198 static int (*make_nop)(void *map, size_t const offset);
199
make_nop_x86(void * map,size_t const offset)200 static int make_nop_x86(void *map, size_t const offset)
201 {
202 uint32_t *ptr;
203 unsigned char *op;
204
205 /* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */
206 ptr = map + offset;
207 if (*ptr != 0)
208 return -1;
209
210 op = map + offset - 1;
211 if (*op != 0xe8)
212 return -1;
213
214 /* convert to nop */
215 ulseek(fd_map, offset - 1, SEEK_SET);
216 uwrite(fd_map, ideal_nop, 5);
217 return 0;
218 }
219
220 static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
make_nop_arm64(void * map,size_t const offset)221 static int make_nop_arm64(void *map, size_t const offset)
222 {
223 uint32_t *ptr;
224
225 ptr = map + offset;
226 /* bl <_mcount> is 0x94000000 before relocation */
227 if (*ptr != 0x94000000)
228 return -1;
229
230 /* Convert to nop */
231 ulseek(fd_map, offset, SEEK_SET);
232 uwrite(fd_map, ideal_nop, 4);
233 return 0;
234 }
235
236 /*
237 * Get the whole file as a programming convenience in order to avoid
238 * malloc+lseek+read+free of many pieces. If successful, then mmap
239 * avoids copying unused pieces; else just read the whole file.
240 * Open for both read and write; new info will be appended to the file.
241 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
242 * do not propagate to the file until an explicit overwrite at the last.
243 * This preserves most aspects of consistency (all except .st_size)
244 * for simultaneous readers of the file while we are appending to it.
245 * However, multiple writers still are bad. We choose not to use
246 * locking because it is expensive and the use case of kernel build
247 * makes multiple writers unlikely.
248 */
mmap_file(char const * fname)249 static void *mmap_file(char const *fname)
250 {
251 fd_map = open(fname, O_RDONLY);
252 if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
253 perror(fname);
254 fail_file();
255 }
256 if (!S_ISREG(sb.st_mode)) {
257 fprintf(stderr, "not a regular file: %s\n", fname);
258 fail_file();
259 }
260 file_map = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
261 fd_map, 0);
262 mmap_failed = 0;
263 if (file_map == MAP_FAILED) {
264 mmap_failed = 1;
265 file_map = umalloc(sb.st_size);
266 uread(fd_map, file_map, sb.st_size);
267 }
268 close(fd_map);
269
270 file_end = file_map + sb.st_size;
271
272 return file_map;
273 }
274
write_file(const char * fname)275 static void write_file(const char *fname)
276 {
277 char tmp_file[strlen(fname) + 4];
278 size_t n;
279
280 if (!file_updated)
281 return;
282
283 sprintf(tmp_file, "%s.rc", fname);
284
285 /*
286 * After reading the entire file into memory, delete it
287 * and write it back, to prevent weird side effects of modifying
288 * an object file in place.
289 */
290 fd_map = open(tmp_file, O_WRONLY | O_TRUNC | O_CREAT, sb.st_mode);
291 if (fd_map < 0) {
292 perror(fname);
293 fail_file();
294 }
295 n = write(fd_map, file_map, sb.st_size);
296 if (n != sb.st_size) {
297 perror("write");
298 fail_file();
299 }
300 if (file_append_size) {
301 n = write(fd_map, file_append, file_append_size);
302 if (n != file_append_size) {
303 perror("write");
304 fail_file();
305 }
306 }
307 close(fd_map);
308 if (rename(tmp_file, fname) < 0) {
309 perror(fname);
310 fail_file();
311 }
312 }
313
314 /* w8rev, w8nat, ...: Handle endianness. */
315
w8rev(uint64_t const x)316 static uint64_t w8rev(uint64_t const x)
317 {
318 return ((0xff & (x >> (0 * 8))) << (7 * 8))
319 | ((0xff & (x >> (1 * 8))) << (6 * 8))
320 | ((0xff & (x >> (2 * 8))) << (5 * 8))
321 | ((0xff & (x >> (3 * 8))) << (4 * 8))
322 | ((0xff & (x >> (4 * 8))) << (3 * 8))
323 | ((0xff & (x >> (5 * 8))) << (2 * 8))
324 | ((0xff & (x >> (6 * 8))) << (1 * 8))
325 | ((0xff & (x >> (7 * 8))) << (0 * 8));
326 }
327
w4rev(uint32_t const x)328 static uint32_t w4rev(uint32_t const x)
329 {
330 return ((0xff & (x >> (0 * 8))) << (3 * 8))
331 | ((0xff & (x >> (1 * 8))) << (2 * 8))
332 | ((0xff & (x >> (2 * 8))) << (1 * 8))
333 | ((0xff & (x >> (3 * 8))) << (0 * 8));
334 }
335
w2rev(uint16_t const x)336 static uint32_t w2rev(uint16_t const x)
337 {
338 return ((0xff & (x >> (0 * 8))) << (1 * 8))
339 | ((0xff & (x >> (1 * 8))) << (0 * 8));
340 }
341
w8nat(uint64_t const x)342 static uint64_t w8nat(uint64_t const x)
343 {
344 return x;
345 }
346
w4nat(uint32_t const x)347 static uint32_t w4nat(uint32_t const x)
348 {
349 return x;
350 }
351
w2nat(uint16_t const x)352 static uint32_t w2nat(uint16_t const x)
353 {
354 return x;
355 }
356
357 static uint64_t (*w8)(uint64_t);
358 static uint32_t (*w)(uint32_t);
359 static uint32_t (*w2)(uint16_t);
360
361 /* Names of the sections that could contain calls to mcount. */
362 static int
is_mcounted_section_name(char const * const txtname)363 is_mcounted_section_name(char const *const txtname)
364 {
365 return strncmp(".text", txtname, 5) == 0 ||
366 strcmp(".ref.text", txtname) == 0 ||
367 strcmp(".sched.text", txtname) == 0 ||
368 strcmp(".spinlock.text", txtname) == 0 ||
369 strcmp(".irqentry.text", txtname) == 0 ||
370 strcmp(".kprobes.text", txtname) == 0 ||
371 strcmp(".text.unlikely", txtname) == 0;
372 }
373
374 /* 32 bit and 64 bit are very similar */
375 #include "recordmcount.h"
376 #define RECORD_MCOUNT_64
377 #include "recordmcount.h"
378
arm_is_fake_mcount(Elf32_Rel const * rp)379 static int arm_is_fake_mcount(Elf32_Rel const *rp)
380 {
381 switch (ELF32_R_TYPE(w(rp->r_info))) {
382 case R_ARM_THM_CALL:
383 case R_ARM_CALL:
384 case R_ARM_PC24:
385 return 0;
386 }
387
388 return 1;
389 }
390
391 /* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
392 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
393 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
394 * to imply the order of the members; the spec does not say so.
395 * typedef unsigned char Elf64_Byte;
396 * fails on MIPS64 because their <elf.h> already has it!
397 */
398
399 typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */
400
401 union mips_r_info {
402 Elf64_Xword r_info;
403 struct {
404 Elf64_Word r_sym; /* Symbol index. */
405 myElf64_Byte r_ssym; /* Special symbol. */
406 myElf64_Byte r_type3; /* Third relocation. */
407 myElf64_Byte r_type2; /* Second relocation. */
408 myElf64_Byte r_type; /* First relocation. */
409 } r_mips;
410 };
411
MIPS64_r_sym(Elf64_Rel const * rp)412 static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
413 {
414 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
415 }
416
MIPS64_r_info(Elf64_Rel * const rp,unsigned sym,unsigned type)417 static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
418 {
419 rp->r_info = ((union mips_r_info){
420 .r_mips = { .r_sym = w(sym), .r_type = type }
421 }).r_info;
422 }
423
424 static void
do_file(char const * const fname)425 do_file(char const *const fname)
426 {
427 Elf32_Ehdr *const ehdr = mmap_file(fname);
428 unsigned int reltype = 0;
429
430 w = w4nat;
431 w2 = w2nat;
432 w8 = w8nat;
433 switch (ehdr->e_ident[EI_DATA]) {
434 static unsigned int const endian = 1;
435 default:
436 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
437 ehdr->e_ident[EI_DATA], fname);
438 fail_file();
439 break;
440 case ELFDATA2LSB:
441 if (*(unsigned char const *)&endian != 1) {
442 /* main() is big endian, file.o is little endian. */
443 w = w4rev;
444 w2 = w2rev;
445 w8 = w8rev;
446 }
447 break;
448 case ELFDATA2MSB:
449 if (*(unsigned char const *)&endian != 0) {
450 /* main() is little endian, file.o is big endian. */
451 w = w4rev;
452 w2 = w2rev;
453 w8 = w8rev;
454 }
455 break;
456 } /* end switch */
457 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
458 || w2(ehdr->e_type) != ET_REL
459 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
460 fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
461 fail_file();
462 }
463
464 gpfx = 0;
465 switch (w2(ehdr->e_machine)) {
466 default:
467 fprintf(stderr, "unrecognized e_machine %d %s\n",
468 w2(ehdr->e_machine), fname);
469 fail_file();
470 break;
471 case EM_386:
472 reltype = R_386_32;
473 rel_type_nop = R_386_NONE;
474 make_nop = make_nop_x86;
475 ideal_nop = ideal_nop5_x86_32;
476 mcount_adjust_32 = -1;
477 break;
478 case EM_ARM: reltype = R_ARM_ABS32;
479 altmcount = "__gnu_mcount_nc";
480 is_fake_mcount32 = arm_is_fake_mcount;
481 break;
482 case EM_AARCH64:
483 reltype = R_AARCH64_ABS64;
484 make_nop = make_nop_arm64;
485 rel_type_nop = R_AARCH64_NONE;
486 ideal_nop = ideal_nop4_arm64;
487 gpfx = '_';
488 break;
489 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break;
490 case EM_METAG: reltype = R_METAG_ADDR32;
491 altmcount = "_mcount_wrapper";
492 rel_type_nop = R_METAG_NONE;
493 /* We happen to have the same requirement as MIPS */
494 is_fake_mcount32 = MIPS32_is_fake_mcount;
495 break;
496 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break;
497 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break;
498 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break;
499 case EM_S390: /* reltype: e_class */ gpfx = '_'; break;
500 case EM_SH: reltype = R_SH_DIR32; break;
501 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break;
502 case EM_X86_64:
503 make_nop = make_nop_x86;
504 ideal_nop = ideal_nop5_x86_64;
505 reltype = R_X86_64_64;
506 rel_type_nop = R_X86_64_NONE;
507 mcount_adjust_64 = -1;
508 break;
509 } /* end switch */
510
511 switch (ehdr->e_ident[EI_CLASS]) {
512 default:
513 fprintf(stderr, "unrecognized ELF class %d %s\n",
514 ehdr->e_ident[EI_CLASS], fname);
515 fail_file();
516 break;
517 case ELFCLASS32:
518 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
519 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
520 fprintf(stderr,
521 "unrecognized ET_REL file: %s\n", fname);
522 fail_file();
523 }
524 if (w2(ehdr->e_machine) == EM_MIPS) {
525 reltype = R_MIPS_32;
526 is_fake_mcount32 = MIPS32_is_fake_mcount;
527 }
528 do32(ehdr, fname, reltype);
529 break;
530 case ELFCLASS64: {
531 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
532 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
533 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
534 fprintf(stderr,
535 "unrecognized ET_REL file: %s\n", fname);
536 fail_file();
537 }
538 if (w2(ghdr->e_machine) == EM_S390) {
539 reltype = R_390_64;
540 mcount_adjust_64 = -14;
541 }
542 if (w2(ghdr->e_machine) == EM_MIPS) {
543 reltype = R_MIPS_64;
544 Elf64_r_sym = MIPS64_r_sym;
545 Elf64_r_info = MIPS64_r_info;
546 is_fake_mcount64 = MIPS64_is_fake_mcount;
547 }
548 do64(ghdr, fname, reltype);
549 break;
550 }
551 } /* end switch */
552
553 write_file(fname);
554 cleanup();
555 }
556
557 int
main(int argc,char * argv[])558 main(int argc, char *argv[])
559 {
560 const char ftrace[] = "/ftrace.o";
561 int ftrace_size = sizeof(ftrace) - 1;
562 int n_error = 0; /* gcc-4.3.0 false positive complaint */
563 int c;
564 int i;
565
566 while ((c = getopt(argc, argv, "w")) >= 0) {
567 switch (c) {
568 case 'w':
569 warn_on_notrace_sect = 1;
570 break;
571 default:
572 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
573 return 0;
574 }
575 }
576
577 if ((argc - optind) < 1) {
578 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
579 return 0;
580 }
581
582 /* Process each file in turn, allowing deep failure. */
583 for (i = optind; i < argc; i++) {
584 char *file = argv[i];
585 int const sjval = setjmp(jmpenv);
586 int len;
587
588 /*
589 * The file kernel/trace/ftrace.o references the mcount
590 * function but does not call it. Since ftrace.o should
591 * not be traced anyway, we just skip it.
592 */
593 len = strlen(file);
594 if (len >= ftrace_size &&
595 strcmp(file + (len - ftrace_size), ftrace) == 0)
596 continue;
597
598 switch (sjval) {
599 default:
600 fprintf(stderr, "internal error: %s\n", file);
601 exit(1);
602 break;
603 case SJ_SETJMP: /* normal sequence */
604 /* Avoid problems if early cleanup() */
605 fd_map = -1;
606 mmap_failed = 1;
607 file_map = NULL;
608 file_ptr = NULL;
609 file_updated = 0;
610 do_file(file);
611 break;
612 case SJ_FAIL: /* error in do_file or below */
613 fprintf(stderr, "%s: failed\n", file);
614 ++n_error;
615 break;
616 case SJ_SUCCEED: /* premature success */
617 /* do nothing */
618 break;
619 } /* end switch */
620 }
621 return !!n_error;
622 }
623