Loading...
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
56static int fd_map; /* File descriptor for file being modified. */
57static int mmap_failed; /* Boolean flag. */
58static char gpfx; /* prefix for global symbol name (sometimes '_') */
59static struct stat sb; /* Remember .st_size, etc. */
60static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */
61static const char *altmcount; /* alternate mcount symbol name */
62static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */
63static void *file_map; /* pointer of the mapped file */
64static void *file_end; /* pointer to the end of the mapped file */
65static int file_updated; /* flag to state file was changed */
66static void *file_ptr; /* current file pointer location */
67static void *file_append; /* added to the end of the file */
68static size_t file_append_size; /* how much is added to end of file */
69
70/* setjmp() return values */
71enum {
72 SJ_SETJMP = 0, /* hardwired first return */
73 SJ_FAIL,
74 SJ_SUCCEED
75};
76
77/* Per-file resource cleanup when multiple files. */
78static void
79cleanup(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
92static void __attribute__((noreturn))
93fail_file(void)
94{
95 cleanup();
96 longjmp(jmpenv, SJ_FAIL);
97}
98
99static void __attribute__((noreturn))
100succeed_file(void)
101{
102 cleanup();
103 longjmp(jmpenv, SJ_SUCCEED);
104}
105
106/* ulseek, uread, ...: Check return value for errors. */
107
108static off_t
109ulseek(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
129static size_t
130uread(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
140static size_t
141uwrite(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
177static void *
178umalloc(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
188static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
189static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
190static unsigned char *ideal_nop;
191
192static char rel_type_nop;
193
194static int (*make_nop)(void *map, size_t const offset);
195
196static 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
216static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 }; /* mov r0, r0 */
217static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 }; /* mov r0, r0 */
218static unsigned char *ideal_nop4_arm;
219
220static unsigned char bl_mcount_arm_le[4] = { 0xfe, 0xff, 0xff, 0xeb }; /* bl */
221static unsigned char bl_mcount_arm_be[4] = { 0xeb, 0xff, 0xff, 0xfe }; /* bl */
222static unsigned char *bl_mcount_arm;
223
224static unsigned char push_arm_le[4] = { 0x04, 0xe0, 0x2d, 0xe5 }; /* push {lr} */
225static unsigned char push_arm_be[4] = { 0xe5, 0x2d, 0xe0, 0x04 }; /* push {lr} */
226static unsigned char *push_arm;
227
228static unsigned char ideal_nop2_thumb_le[2] = { 0x00, 0xbf }; /* nop */
229static unsigned char ideal_nop2_thumb_be[2] = { 0xbf, 0x00 }; /* nop */
230static unsigned char *ideal_nop2_thumb;
231
232static unsigned char push_bl_mcount_thumb_le[6] = { 0x00, 0xb5, 0xff, 0xf7, 0xfe, 0xff }; /* push {lr}, bl */
233static unsigned char push_bl_mcount_thumb_be[6] = { 0xb5, 0x00, 0xf7, 0xff, 0xff, 0xfe }; /* push {lr}, bl */
234static unsigned char *push_bl_mcount_thumb;
235
236static int make_nop_arm(void *map, size_t const offset)
237{
238 char *ptr;
239 int cnt = 1;
240 int nop_size;
241 size_t off = offset;
242
243 ptr = map + offset;
244 if (memcmp(ptr, bl_mcount_arm, 4) == 0) {
245 if (memcmp(ptr - 4, push_arm, 4) == 0) {
246 off -= 4;
247 cnt = 2;
248 }
249 ideal_nop = ideal_nop4_arm;
250 nop_size = 4;
251 } else if (memcmp(ptr - 2, push_bl_mcount_thumb, 6) == 0) {
252 cnt = 3;
253 nop_size = 2;
254 off -= 2;
255 ideal_nop = ideal_nop2_thumb;
256 } else
257 return -1;
258
259 /* Convert to nop */
260 ulseek(fd_map, off, SEEK_SET);
261
262 do {
263 uwrite(fd_map, ideal_nop, nop_size);
264 } while (--cnt > 0);
265
266 return 0;
267}
268
269static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
270static int make_nop_arm64(void *map, size_t const offset)
271{
272 uint32_t *ptr;
273
274 ptr = map + offset;
275 /* bl <_mcount> is 0x94000000 before relocation */
276 if (*ptr != 0x94000000)
277 return -1;
278
279 /* Convert to nop */
280 ulseek(fd_map, offset, SEEK_SET);
281 uwrite(fd_map, ideal_nop, 4);
282 return 0;
283}
284
285/*
286 * Get the whole file as a programming convenience in order to avoid
287 * malloc+lseek+read+free of many pieces. If successful, then mmap
288 * avoids copying unused pieces; else just read the whole file.
289 * Open for both read and write; new info will be appended to the file.
290 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
291 * do not propagate to the file until an explicit overwrite at the last.
292 * This preserves most aspects of consistency (all except .st_size)
293 * for simultaneous readers of the file while we are appending to it.
294 * However, multiple writers still are bad. We choose not to use
295 * locking because it is expensive and the use case of kernel build
296 * makes multiple writers unlikely.
297 */
298static void *mmap_file(char const *fname)
299{
300 fd_map = open(fname, O_RDONLY);
301 if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
302 perror(fname);
303 fail_file();
304 }
305 if (!S_ISREG(sb.st_mode)) {
306 fprintf(stderr, "not a regular file: %s\n", fname);
307 fail_file();
308 }
309 file_map = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
310 fd_map, 0);
311 mmap_failed = 0;
312 if (file_map == MAP_FAILED) {
313 mmap_failed = 1;
314 file_map = umalloc(sb.st_size);
315 uread(fd_map, file_map, sb.st_size);
316 }
317 close(fd_map);
318
319 file_end = file_map + sb.st_size;
320
321 return file_map;
322}
323
324static void write_file(const char *fname)
325{
326 char tmp_file[strlen(fname) + 4];
327 size_t n;
328
329 if (!file_updated)
330 return;
331
332 sprintf(tmp_file, "%s.rc", fname);
333
334 /*
335 * After reading the entire file into memory, delete it
336 * and write it back, to prevent weird side effects of modifying
337 * an object file in place.
338 */
339 fd_map = open(tmp_file, O_WRONLY | O_TRUNC | O_CREAT, sb.st_mode);
340 if (fd_map < 0) {
341 perror(fname);
342 fail_file();
343 }
344 n = write(fd_map, file_map, sb.st_size);
345 if (n != sb.st_size) {
346 perror("write");
347 fail_file();
348 }
349 if (file_append_size) {
350 n = write(fd_map, file_append, file_append_size);
351 if (n != file_append_size) {
352 perror("write");
353 fail_file();
354 }
355 }
356 close(fd_map);
357 if (rename(tmp_file, fname) < 0) {
358 perror(fname);
359 fail_file();
360 }
361}
362
363/* w8rev, w8nat, ...: Handle endianness. */
364
365static uint64_t w8rev(uint64_t const x)
366{
367 return ((0xff & (x >> (0 * 8))) << (7 * 8))
368 | ((0xff & (x >> (1 * 8))) << (6 * 8))
369 | ((0xff & (x >> (2 * 8))) << (5 * 8))
370 | ((0xff & (x >> (3 * 8))) << (4 * 8))
371 | ((0xff & (x >> (4 * 8))) << (3 * 8))
372 | ((0xff & (x >> (5 * 8))) << (2 * 8))
373 | ((0xff & (x >> (6 * 8))) << (1 * 8))
374 | ((0xff & (x >> (7 * 8))) << (0 * 8));
375}
376
377static uint32_t w4rev(uint32_t const x)
378{
379 return ((0xff & (x >> (0 * 8))) << (3 * 8))
380 | ((0xff & (x >> (1 * 8))) << (2 * 8))
381 | ((0xff & (x >> (2 * 8))) << (1 * 8))
382 | ((0xff & (x >> (3 * 8))) << (0 * 8));
383}
384
385static uint32_t w2rev(uint16_t const x)
386{
387 return ((0xff & (x >> (0 * 8))) << (1 * 8))
388 | ((0xff & (x >> (1 * 8))) << (0 * 8));
389}
390
391static uint64_t w8nat(uint64_t const x)
392{
393 return x;
394}
395
396static uint32_t w4nat(uint32_t const x)
397{
398 return x;
399}
400
401static uint32_t w2nat(uint16_t const x)
402{
403 return x;
404}
405
406static uint64_t (*w8)(uint64_t);
407static uint32_t (*w)(uint32_t);
408static uint32_t (*w2)(uint16_t);
409
410/* Names of the sections that could contain calls to mcount. */
411static int
412is_mcounted_section_name(char const *const txtname)
413{
414 return strcmp(".text", txtname) == 0 ||
415 strcmp(".ref.text", txtname) == 0 ||
416 strcmp(".sched.text", txtname) == 0 ||
417 strcmp(".spinlock.text", txtname) == 0 ||
418 strcmp(".irqentry.text", txtname) == 0 ||
419 strcmp(".softirqentry.text", txtname) == 0 ||
420 strcmp(".kprobes.text", txtname) == 0 ||
421 strcmp(".cpuidle.text", txtname) == 0 ||
422 strcmp(".text.unlikely", txtname) == 0;
423}
424
425/* 32 bit and 64 bit are very similar */
426#include "recordmcount.h"
427#define RECORD_MCOUNT_64
428#include "recordmcount.h"
429
430/* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
431 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
432 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
433 * to imply the order of the members; the spec does not say so.
434 * typedef unsigned char Elf64_Byte;
435 * fails on MIPS64 because their <elf.h> already has it!
436 */
437
438typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */
439
440union mips_r_info {
441 Elf64_Xword r_info;
442 struct {
443 Elf64_Word r_sym; /* Symbol index. */
444 myElf64_Byte r_ssym; /* Special symbol. */
445 myElf64_Byte r_type3; /* Third relocation. */
446 myElf64_Byte r_type2; /* Second relocation. */
447 myElf64_Byte r_type; /* First relocation. */
448 } r_mips;
449};
450
451static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
452{
453 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
454}
455
456static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
457{
458 rp->r_info = ((union mips_r_info){
459 .r_mips = { .r_sym = w(sym), .r_type = type }
460 }).r_info;
461}
462
463static void
464do_file(char const *const fname)
465{
466 Elf32_Ehdr *const ehdr = mmap_file(fname);
467 unsigned int reltype = 0;
468
469 w = w4nat;
470 w2 = w2nat;
471 w8 = w8nat;
472 switch (ehdr->e_ident[EI_DATA]) {
473 static unsigned int const endian = 1;
474 default:
475 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
476 ehdr->e_ident[EI_DATA], fname);
477 fail_file();
478 break;
479 case ELFDATA2LSB:
480 if (*(unsigned char const *)&endian != 1) {
481 /* main() is big endian, file.o is little endian. */
482 w = w4rev;
483 w2 = w2rev;
484 w8 = w8rev;
485 }
486 ideal_nop4_arm = ideal_nop4_arm_le;
487 bl_mcount_arm = bl_mcount_arm_le;
488 push_arm = push_arm_le;
489 ideal_nop2_thumb = ideal_nop2_thumb_le;
490 push_bl_mcount_thumb = push_bl_mcount_thumb_le;
491 break;
492 case ELFDATA2MSB:
493 if (*(unsigned char const *)&endian != 0) {
494 /* main() is little endian, file.o is big endian. */
495 w = w4rev;
496 w2 = w2rev;
497 w8 = w8rev;
498 }
499 ideal_nop4_arm = ideal_nop4_arm_be;
500 bl_mcount_arm = bl_mcount_arm_be;
501 push_arm = push_arm_be;
502 ideal_nop2_thumb = ideal_nop2_thumb_be;
503 push_bl_mcount_thumb = push_bl_mcount_thumb_be;
504 break;
505 } /* end switch */
506 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
507 || w2(ehdr->e_type) != ET_REL
508 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
509 fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
510 fail_file();
511 }
512
513 gpfx = 0;
514 switch (w2(ehdr->e_machine)) {
515 default:
516 fprintf(stderr, "unrecognized e_machine %d %s\n",
517 w2(ehdr->e_machine), fname);
518 fail_file();
519 break;
520 case EM_386:
521 reltype = R_386_32;
522 rel_type_nop = R_386_NONE;
523 make_nop = make_nop_x86;
524 ideal_nop = ideal_nop5_x86_32;
525 mcount_adjust_32 = -1;
526 break;
527 case EM_ARM: reltype = R_ARM_ABS32;
528 altmcount = "__gnu_mcount_nc";
529 make_nop = make_nop_arm;
530 rel_type_nop = R_ARM_NONE;
531 break;
532 case EM_AARCH64:
533 reltype = R_AARCH64_ABS64;
534 make_nop = make_nop_arm64;
535 rel_type_nop = R_AARCH64_NONE;
536 ideal_nop = ideal_nop4_arm64;
537 gpfx = '_';
538 break;
539 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break;
540 case EM_METAG: reltype = R_METAG_ADDR32;
541 altmcount = "_mcount_wrapper";
542 rel_type_nop = R_METAG_NONE;
543 /* We happen to have the same requirement as MIPS */
544 is_fake_mcount32 = MIPS32_is_fake_mcount;
545 break;
546 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break;
547 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break;
548 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break;
549 case EM_S390: /* reltype: e_class */ gpfx = '_'; break;
550 case EM_SH: reltype = R_SH_DIR32; break;
551 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break;
552 case EM_X86_64:
553 make_nop = make_nop_x86;
554 ideal_nop = ideal_nop5_x86_64;
555 reltype = R_X86_64_64;
556 rel_type_nop = R_X86_64_NONE;
557 mcount_adjust_64 = -1;
558 break;
559 } /* end switch */
560
561 switch (ehdr->e_ident[EI_CLASS]) {
562 default:
563 fprintf(stderr, "unrecognized ELF class %d %s\n",
564 ehdr->e_ident[EI_CLASS], fname);
565 fail_file();
566 break;
567 case ELFCLASS32:
568 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
569 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
570 fprintf(stderr,
571 "unrecognized ET_REL file: %s\n", fname);
572 fail_file();
573 }
574 if (w2(ehdr->e_machine) == EM_MIPS) {
575 reltype = R_MIPS_32;
576 is_fake_mcount32 = MIPS32_is_fake_mcount;
577 }
578 do32(ehdr, fname, reltype);
579 break;
580 case ELFCLASS64: {
581 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
582 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
583 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
584 fprintf(stderr,
585 "unrecognized ET_REL file: %s\n", fname);
586 fail_file();
587 }
588 if (w2(ghdr->e_machine) == EM_S390) {
589 reltype = R_390_64;
590 mcount_adjust_64 = -14;
591 }
592 if (w2(ghdr->e_machine) == EM_MIPS) {
593 reltype = R_MIPS_64;
594 Elf64_r_sym = MIPS64_r_sym;
595 Elf64_r_info = MIPS64_r_info;
596 is_fake_mcount64 = MIPS64_is_fake_mcount;
597 }
598 do64(ghdr, fname, reltype);
599 break;
600 }
601 } /* end switch */
602
603 write_file(fname);
604 cleanup();
605}
606
607int
608main(int argc, char *argv[])
609{
610 const char ftrace[] = "/ftrace.o";
611 int ftrace_size = sizeof(ftrace) - 1;
612 int n_error = 0; /* gcc-4.3.0 false positive complaint */
613 int c;
614 int i;
615
616 while ((c = getopt(argc, argv, "w")) >= 0) {
617 switch (c) {
618 case 'w':
619 warn_on_notrace_sect = 1;
620 break;
621 default:
622 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
623 return 0;
624 }
625 }
626
627 if ((argc - optind) < 1) {
628 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
629 return 0;
630 }
631
632 /* Process each file in turn, allowing deep failure. */
633 for (i = optind; i < argc; i++) {
634 char *file = argv[i];
635 int const sjval = setjmp(jmpenv);
636 int len;
637
638 /*
639 * The file kernel/trace/ftrace.o references the mcount
640 * function but does not call it. Since ftrace.o should
641 * not be traced anyway, we just skip it.
642 */
643 len = strlen(file);
644 if (len >= ftrace_size &&
645 strcmp(file + (len - ftrace_size), ftrace) == 0)
646 continue;
647
648 switch (sjval) {
649 default:
650 fprintf(stderr, "internal error: %s\n", file);
651 exit(1);
652 break;
653 case SJ_SETJMP: /* normal sequence */
654 /* Avoid problems if early cleanup() */
655 fd_map = -1;
656 mmap_failed = 1;
657 file_map = NULL;
658 file_ptr = NULL;
659 file_updated = 0;
660 do_file(file);
661 break;
662 case SJ_FAIL: /* error in do_file or below */
663 fprintf(stderr, "%s: failed\n", file);
664 ++n_error;
665 break;
666 case SJ_SUCCEED: /* premature success */
667 /* do nothing */
668 break;
669 } /* end switch */
670 }
671 return !!n_error;
672}
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
36static int fd_map; /* File descriptor for file being modified. */
37static int mmap_failed; /* Boolean flag. */
38static void *ehdr_curr; /* current ElfXX_Ehdr * for resource cleanup */
39static char gpfx; /* prefix for global symbol name (sometimes '_') */
40static struct stat sb; /* Remember .st_size, etc. */
41static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */
42static const char *altmcount; /* alternate mcount symbol name */
43static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */
44
45/* setjmp() return values */
46enum {
47 SJ_SETJMP = 0, /* hardwired first return */
48 SJ_FAIL,
49 SJ_SUCCEED
50};
51
52/* Per-file resource cleanup when multiple files. */
53static void
54cleanup(void)
55{
56 if (!mmap_failed)
57 munmap(ehdr_curr, sb.st_size);
58 else
59 free(ehdr_curr);
60 close(fd_map);
61}
62
63static void __attribute__((noreturn))
64fail_file(void)
65{
66 cleanup();
67 longjmp(jmpenv, SJ_FAIL);
68}
69
70static void __attribute__((noreturn))
71succeed_file(void)
72{
73 cleanup();
74 longjmp(jmpenv, SJ_SUCCEED);
75}
76
77/* ulseek, uread, ...: Check return value for errors. */
78
79static off_t
80ulseek(int const fd, off_t const offset, int const whence)
81{
82 off_t const w = lseek(fd, offset, whence);
83 if (w == (off_t)-1) {
84 perror("lseek");
85 fail_file();
86 }
87 return w;
88}
89
90static size_t
91uread(int const fd, void *const buf, size_t const count)
92{
93 size_t const n = read(fd, buf, count);
94 if (n != count) {
95 perror("read");
96 fail_file();
97 }
98 return n;
99}
100
101static size_t
102uwrite(int const fd, void const *const buf, size_t const count)
103{
104 size_t const n = write(fd, buf, count);
105 if (n != count) {
106 perror("write");
107 fail_file();
108 }
109 return n;
110}
111
112static void *
113umalloc(size_t size)
114{
115 void *const addr = malloc(size);
116 if (addr == 0) {
117 fprintf(stderr, "malloc failed: %zu bytes\n", size);
118 fail_file();
119 }
120 return addr;
121}
122
123static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
124static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
125static unsigned char *ideal_nop;
126
127static char rel_type_nop;
128
129static int (*make_nop)(void *map, size_t const offset);
130
131static int make_nop_x86(void *map, size_t const offset)
132{
133 uint32_t *ptr;
134 unsigned char *op;
135
136 /* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */
137 ptr = map + offset;
138 if (*ptr != 0)
139 return -1;
140
141 op = map + offset - 1;
142 if (*op != 0xe8)
143 return -1;
144
145 /* convert to nop */
146 ulseek(fd_map, offset - 1, SEEK_SET);
147 uwrite(fd_map, ideal_nop, 5);
148 return 0;
149}
150
151/*
152 * Get the whole file as a programming convenience in order to avoid
153 * malloc+lseek+read+free of many pieces. If successful, then mmap
154 * avoids copying unused pieces; else just read the whole file.
155 * Open for both read and write; new info will be appended to the file.
156 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
157 * do not propagate to the file until an explicit overwrite at the last.
158 * This preserves most aspects of consistency (all except .st_size)
159 * for simultaneous readers of the file while we are appending to it.
160 * However, multiple writers still are bad. We choose not to use
161 * locking because it is expensive and the use case of kernel build
162 * makes multiple writers unlikely.
163 */
164static void *mmap_file(char const *fname)
165{
166 void *addr;
167
168 fd_map = open(fname, O_RDWR);
169 if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
170 perror(fname);
171 fail_file();
172 }
173 if (!S_ISREG(sb.st_mode)) {
174 fprintf(stderr, "not a regular file: %s\n", fname);
175 fail_file();
176 }
177 addr = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
178 fd_map, 0);
179 mmap_failed = 0;
180 if (addr == MAP_FAILED) {
181 mmap_failed = 1;
182 addr = umalloc(sb.st_size);
183 uread(fd_map, addr, sb.st_size);
184 }
185 return addr;
186}
187
188/* w8rev, w8nat, ...: Handle endianness. */
189
190static uint64_t w8rev(uint64_t const x)
191{
192 return ((0xff & (x >> (0 * 8))) << (7 * 8))
193 | ((0xff & (x >> (1 * 8))) << (6 * 8))
194 | ((0xff & (x >> (2 * 8))) << (5 * 8))
195 | ((0xff & (x >> (3 * 8))) << (4 * 8))
196 | ((0xff & (x >> (4 * 8))) << (3 * 8))
197 | ((0xff & (x >> (5 * 8))) << (2 * 8))
198 | ((0xff & (x >> (6 * 8))) << (1 * 8))
199 | ((0xff & (x >> (7 * 8))) << (0 * 8));
200}
201
202static uint32_t w4rev(uint32_t const x)
203{
204 return ((0xff & (x >> (0 * 8))) << (3 * 8))
205 | ((0xff & (x >> (1 * 8))) << (2 * 8))
206 | ((0xff & (x >> (2 * 8))) << (1 * 8))
207 | ((0xff & (x >> (3 * 8))) << (0 * 8));
208}
209
210static uint32_t w2rev(uint16_t const x)
211{
212 return ((0xff & (x >> (0 * 8))) << (1 * 8))
213 | ((0xff & (x >> (1 * 8))) << (0 * 8));
214}
215
216static uint64_t w8nat(uint64_t const x)
217{
218 return x;
219}
220
221static uint32_t w4nat(uint32_t const x)
222{
223 return x;
224}
225
226static uint32_t w2nat(uint16_t const x)
227{
228 return x;
229}
230
231static uint64_t (*w8)(uint64_t);
232static uint32_t (*w)(uint32_t);
233static uint32_t (*w2)(uint16_t);
234
235/* Names of the sections that could contain calls to mcount. */
236static int
237is_mcounted_section_name(char const *const txtname)
238{
239 return strcmp(".text", txtname) == 0 ||
240 strcmp(".ref.text", txtname) == 0 ||
241 strcmp(".sched.text", txtname) == 0 ||
242 strcmp(".spinlock.text", txtname) == 0 ||
243 strcmp(".irqentry.text", txtname) == 0 ||
244 strcmp(".kprobes.text", txtname) == 0 ||
245 strcmp(".text.unlikely", txtname) == 0;
246}
247
248/* 32 bit and 64 bit are very similar */
249#include "recordmcount.h"
250#define RECORD_MCOUNT_64
251#include "recordmcount.h"
252
253/* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
254 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
255 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
256 * to imply the order of the members; the spec does not say so.
257 * typedef unsigned char Elf64_Byte;
258 * fails on MIPS64 because their <elf.h> already has it!
259 */
260
261typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */
262
263union mips_r_info {
264 Elf64_Xword r_info;
265 struct {
266 Elf64_Word r_sym; /* Symbol index. */
267 myElf64_Byte r_ssym; /* Special symbol. */
268 myElf64_Byte r_type3; /* Third relocation. */
269 myElf64_Byte r_type2; /* Second relocation. */
270 myElf64_Byte r_type; /* First relocation. */
271 } r_mips;
272};
273
274static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
275{
276 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
277}
278
279static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
280{
281 rp->r_info = ((union mips_r_info){
282 .r_mips = { .r_sym = w(sym), .r_type = type }
283 }).r_info;
284}
285
286static void
287do_file(char const *const fname)
288{
289 Elf32_Ehdr *const ehdr = mmap_file(fname);
290 unsigned int reltype = 0;
291
292 ehdr_curr = ehdr;
293 w = w4nat;
294 w2 = w2nat;
295 w8 = w8nat;
296 switch (ehdr->e_ident[EI_DATA]) {
297 static unsigned int const endian = 1;
298 default:
299 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
300 ehdr->e_ident[EI_DATA], fname);
301 fail_file();
302 break;
303 case ELFDATA2LSB:
304 if (*(unsigned char const *)&endian != 1) {
305 /* main() is big endian, file.o is little endian. */
306 w = w4rev;
307 w2 = w2rev;
308 w8 = w8rev;
309 }
310 break;
311 case ELFDATA2MSB:
312 if (*(unsigned char const *)&endian != 0) {
313 /* main() is little endian, file.o is big endian. */
314 w = w4rev;
315 w2 = w2rev;
316 w8 = w8rev;
317 }
318 break;
319 } /* end switch */
320 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
321 || w2(ehdr->e_type) != ET_REL
322 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
323 fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
324 fail_file();
325 }
326
327 gpfx = 0;
328 switch (w2(ehdr->e_machine)) {
329 default:
330 fprintf(stderr, "unrecognized e_machine %d %s\n",
331 w2(ehdr->e_machine), fname);
332 fail_file();
333 break;
334 case EM_386:
335 reltype = R_386_32;
336 make_nop = make_nop_x86;
337 ideal_nop = ideal_nop5_x86_32;
338 mcount_adjust_32 = -1;
339 break;
340 case EM_ARM: reltype = R_ARM_ABS32;
341 altmcount = "__gnu_mcount_nc";
342 break;
343 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break;
344 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break;
345 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break;
346 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break;
347 case EM_S390: /* reltype: e_class */ gpfx = '_'; break;
348 case EM_SH: reltype = R_SH_DIR32; break;
349 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break;
350 case EM_X86_64:
351 make_nop = make_nop_x86;
352 ideal_nop = ideal_nop5_x86_64;
353 reltype = R_X86_64_64;
354 mcount_adjust_64 = -1;
355 break;
356 } /* end switch */
357
358 switch (ehdr->e_ident[EI_CLASS]) {
359 default:
360 fprintf(stderr, "unrecognized ELF class %d %s\n",
361 ehdr->e_ident[EI_CLASS], fname);
362 fail_file();
363 break;
364 case ELFCLASS32:
365 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
366 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
367 fprintf(stderr,
368 "unrecognized ET_REL file: %s\n", fname);
369 fail_file();
370 }
371 if (w2(ehdr->e_machine) == EM_S390) {
372 reltype = R_390_32;
373 mcount_adjust_32 = -4;
374 }
375 if (w2(ehdr->e_machine) == EM_MIPS) {
376 reltype = R_MIPS_32;
377 is_fake_mcount32 = MIPS32_is_fake_mcount;
378 }
379 do32(ehdr, fname, reltype);
380 break;
381 case ELFCLASS64: {
382 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
383 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
384 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
385 fprintf(stderr,
386 "unrecognized ET_REL file: %s\n", fname);
387 fail_file();
388 }
389 if (w2(ghdr->e_machine) == EM_S390) {
390 reltype = R_390_64;
391 mcount_adjust_64 = -8;
392 }
393 if (w2(ghdr->e_machine) == EM_MIPS) {
394 reltype = R_MIPS_64;
395 Elf64_r_sym = MIPS64_r_sym;
396 Elf64_r_info = MIPS64_r_info;
397 is_fake_mcount64 = MIPS64_is_fake_mcount;
398 }
399 do64(ghdr, fname, reltype);
400 break;
401 }
402 } /* end switch */
403
404 cleanup();
405}
406
407int
408main(int argc, char *argv[])
409{
410 const char ftrace[] = "/ftrace.o";
411 int ftrace_size = sizeof(ftrace) - 1;
412 int n_error = 0; /* gcc-4.3.0 false positive complaint */
413 int c;
414 int i;
415
416 while ((c = getopt(argc, argv, "w")) >= 0) {
417 switch (c) {
418 case 'w':
419 warn_on_notrace_sect = 1;
420 break;
421 default:
422 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
423 return 0;
424 }
425 }
426
427 if ((argc - optind) < 1) {
428 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
429 return 0;
430 }
431
432 /* Process each file in turn, allowing deep failure. */
433 for (i = optind; i < argc; i++) {
434 char *file = argv[i];
435 int const sjval = setjmp(jmpenv);
436 int len;
437
438 /*
439 * The file kernel/trace/ftrace.o references the mcount
440 * function but does not call it. Since ftrace.o should
441 * not be traced anyway, we just skip it.
442 */
443 len = strlen(file);
444 if (len >= ftrace_size &&
445 strcmp(file + (len - ftrace_size), ftrace) == 0)
446 continue;
447
448 switch (sjval) {
449 default:
450 fprintf(stderr, "internal error: %s\n", file);
451 exit(1);
452 break;
453 case SJ_SETJMP: /* normal sequence */
454 /* Avoid problems if early cleanup() */
455 fd_map = -1;
456 ehdr_curr = NULL;
457 mmap_failed = 1;
458 do_file(file);
459 break;
460 case SJ_FAIL: /* error in do_file or below */
461 ++n_error;
462 break;
463 case SJ_SUCCEED: /* premature success */
464 /* do nothing */
465 break;
466 } /* end switch */
467 }
468 return !!n_error;
469}
470
471