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