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).
7 * Restructured to fit Linux format, as well as other updates:
8 * Copyright 2010 Steven Rostedt <srostedt@redhat.com>, Red Hat Inc.
12 * Strategy: alter the .o file in-place.
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.]
24 #include <sys/types.h>
37 /* Remove this when these make it to the standard system elf.h. */
39 #define R_METAG_ADDR32 2
40 #define R_METAG_NONE 3
44 #define EM_AARCH64 183
45 #define R_AARCH64_NONE 0
46 #define R_AARCH64_ABS64 257
49 static int fd_map; /* File descriptor for file being modified. */
50 static int mmap_failed; /* Boolean flag. */
51 static void *ehdr_curr; /* current ElfXX_Ehdr * for resource cleanup */
52 static char gpfx; /* prefix for global symbol name (sometimes '_') */
53 static struct stat sb; /* Remember .st_size, etc. */
54 static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */
55 static const char *altmcount; /* alternate mcount symbol name */
56 static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */
58 /* setjmp() return values */
60 SJ_SETJMP = 0, /* hardwired first return */
65 /* Per-file resource cleanup when multiple files. */
70 munmap(ehdr_curr, sb.st_size);
76 static void __attribute__((noreturn))
80 longjmp(jmpenv, SJ_FAIL);
83 static void __attribute__((noreturn))
87 longjmp(jmpenv, SJ_SUCCEED);
90 /* ulseek, uread, ...: Check return value for errors. */
93 ulseek(int const fd, off_t const offset, int const whence)
95 off_t const w = lseek(fd, offset, whence);
104 uread(int const fd, void *const buf, size_t const count)
106 size_t const n = read(fd, buf, count);
115 uwrite(int const fd, void const *const buf, size_t const count)
117 size_t const n = write(fd, buf, count);
128 void *const addr = malloc(size);
130 fprintf(stderr, "malloc failed: %zu bytes\n", size);
136 static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
137 static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
138 static unsigned char *ideal_nop;
140 static char rel_type_nop;
142 static int (*make_nop)(void *map, size_t const offset);
144 static int make_nop_x86(void *map, size_t const offset)
149 /* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */
154 op = map + offset - 1;
159 ulseek(fd_map, offset - 1, SEEK_SET);
160 uwrite(fd_map, ideal_nop, 5);
164 static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
165 static int make_nop_arm64(void *map, size_t const offset)
170 /* bl <_mcount> is 0x94000000 before relocation */
171 if (*ptr != 0x94000000)
175 ulseek(fd_map, offset, SEEK_SET);
176 uwrite(fd_map, ideal_nop, 4);
181 * Get the whole file as a programming convenience in order to avoid
182 * malloc+lseek+read+free of many pieces. If successful, then mmap
183 * avoids copying unused pieces; else just read the whole file.
184 * Open for both read and write; new info will be appended to the file.
185 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
186 * do not propagate to the file until an explicit overwrite at the last.
187 * This preserves most aspects of consistency (all except .st_size)
188 * for simultaneous readers of the file while we are appending to it.
189 * However, multiple writers still are bad. We choose not to use
190 * locking because it is expensive and the use case of kernel build
191 * makes multiple writers unlikely.
193 static void *mmap_file(char const *fname)
197 fd_map = open(fname, O_RDWR);
198 if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
202 if (!S_ISREG(sb.st_mode)) {
203 fprintf(stderr, "not a regular file: %s\n", fname);
206 addr = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
209 if (addr == MAP_FAILED) {
211 addr = umalloc(sb.st_size);
212 uread(fd_map, addr, sb.st_size);
214 if (sb.st_nlink != 1) {
215 /* file is hard-linked, break the hard link */
217 if (unlink(fname) < 0) {
221 fd_map = open(fname, O_RDWR | O_CREAT, sb.st_mode);
226 uwrite(fd_map, addr, sb.st_size);
231 /* w8rev, w8nat, ...: Handle endianness. */
233 static uint64_t w8rev(uint64_t const x)
235 return ((0xff & (x >> (0 * 8))) << (7 * 8))
236 | ((0xff & (x >> (1 * 8))) << (6 * 8))
237 | ((0xff & (x >> (2 * 8))) << (5 * 8))
238 | ((0xff & (x >> (3 * 8))) << (4 * 8))
239 | ((0xff & (x >> (4 * 8))) << (3 * 8))
240 | ((0xff & (x >> (5 * 8))) << (2 * 8))
241 | ((0xff & (x >> (6 * 8))) << (1 * 8))
242 | ((0xff & (x >> (7 * 8))) << (0 * 8));
245 static uint32_t w4rev(uint32_t const x)
247 return ((0xff & (x >> (0 * 8))) << (3 * 8))
248 | ((0xff & (x >> (1 * 8))) << (2 * 8))
249 | ((0xff & (x >> (2 * 8))) << (1 * 8))
250 | ((0xff & (x >> (3 * 8))) << (0 * 8));
253 static uint32_t w2rev(uint16_t const x)
255 return ((0xff & (x >> (0 * 8))) << (1 * 8))
256 | ((0xff & (x >> (1 * 8))) << (0 * 8));
259 static uint64_t w8nat(uint64_t const x)
264 static uint32_t w4nat(uint32_t const x)
269 static uint32_t w2nat(uint16_t const x)
274 static uint64_t (*w8)(uint64_t);
275 static uint32_t (*w)(uint32_t);
276 static uint32_t (*w2)(uint16_t);
278 /* Names of the sections that could contain calls to mcount. */
280 is_mcounted_section_name(char const *const txtname)
282 return strcmp(".text", txtname) == 0 ||
283 strcmp(".ref.text", txtname) == 0 ||
284 strcmp(".sched.text", txtname) == 0 ||
285 strcmp(".spinlock.text", txtname) == 0 ||
286 strcmp(".irqentry.text", txtname) == 0 ||
287 strcmp(".kprobes.text", txtname) == 0 ||
288 strcmp(".text.unlikely", txtname) == 0;
291 /* 32 bit and 64 bit are very similar */
292 #include "recordmcount.h"
293 #define RECORD_MCOUNT_64
294 #include "recordmcount.h"
296 /* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
297 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
298 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
299 * to imply the order of the members; the spec does not say so.
300 * typedef unsigned char Elf64_Byte;
301 * fails on MIPS64 because their <elf.h> already has it!
304 typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */
309 Elf64_Word r_sym; /* Symbol index. */
310 myElf64_Byte r_ssym; /* Special symbol. */
311 myElf64_Byte r_type3; /* Third relocation. */
312 myElf64_Byte r_type2; /* Second relocation. */
313 myElf64_Byte r_type; /* First relocation. */
317 static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
319 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
322 static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
324 rp->r_info = ((union mips_r_info){
325 .r_mips = { .r_sym = w(sym), .r_type = type }
330 do_file(char const *const fname)
332 Elf32_Ehdr *const ehdr = mmap_file(fname);
333 unsigned int reltype = 0;
339 switch (ehdr->e_ident[EI_DATA]) {
340 static unsigned int const endian = 1;
342 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
343 ehdr->e_ident[EI_DATA], fname);
347 if (*(unsigned char const *)&endian != 1) {
348 /* main() is big endian, file.o is little endian. */
355 if (*(unsigned char const *)&endian != 0) {
356 /* main() is little endian, file.o is big endian. */
363 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
364 || w2(ehdr->e_type) != ET_REL
365 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
366 fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
371 switch (w2(ehdr->e_machine)) {
373 fprintf(stderr, "unrecognized e_machine %d %s\n",
374 w2(ehdr->e_machine), fname);
379 rel_type_nop = R_386_NONE;
380 make_nop = make_nop_x86;
381 ideal_nop = ideal_nop5_x86_32;
382 mcount_adjust_32 = -1;
384 case EM_ARM: reltype = R_ARM_ABS32;
385 altmcount = "__gnu_mcount_nc";
388 reltype = R_AARCH64_ABS64;
389 make_nop = make_nop_arm64;
390 rel_type_nop = R_AARCH64_NONE;
391 ideal_nop = ideal_nop4_arm64;
394 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break;
395 case EM_METAG: reltype = R_METAG_ADDR32;
396 altmcount = "_mcount_wrapper";
397 rel_type_nop = R_METAG_NONE;
398 /* We happen to have the same requirement as MIPS */
399 is_fake_mcount32 = MIPS32_is_fake_mcount;
401 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break;
402 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break;
403 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break;
404 case EM_S390: /* reltype: e_class */ gpfx = '_'; break;
405 case EM_SH: reltype = R_SH_DIR32; break;
406 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break;
408 make_nop = make_nop_x86;
409 ideal_nop = ideal_nop5_x86_64;
410 reltype = R_X86_64_64;
411 rel_type_nop = R_X86_64_NONE;
412 mcount_adjust_64 = -1;
416 switch (ehdr->e_ident[EI_CLASS]) {
418 fprintf(stderr, "unrecognized ELF class %d %s\n",
419 ehdr->e_ident[EI_CLASS], fname);
423 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
424 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
426 "unrecognized ET_REL file: %s\n", fname);
429 if (w2(ehdr->e_machine) == EM_MIPS) {
431 is_fake_mcount32 = MIPS32_is_fake_mcount;
433 do32(ehdr, fname, reltype);
436 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
437 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
438 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
440 "unrecognized ET_REL file: %s\n", fname);
443 if (w2(ghdr->e_machine) == EM_S390) {
445 mcount_adjust_64 = -14;
447 if (w2(ghdr->e_machine) == EM_MIPS) {
449 Elf64_r_sym = MIPS64_r_sym;
450 Elf64_r_info = MIPS64_r_info;
451 is_fake_mcount64 = MIPS64_is_fake_mcount;
453 do64(ghdr, fname, reltype);
462 main(int argc, char *argv[])
464 const char ftrace[] = "/ftrace.o";
465 int ftrace_size = sizeof(ftrace) - 1;
466 int n_error = 0; /* gcc-4.3.0 false positive complaint */
470 while ((c = getopt(argc, argv, "w")) >= 0) {
473 warn_on_notrace_sect = 1;
476 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
481 if ((argc - optind) < 1) {
482 fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
486 /* Process each file in turn, allowing deep failure. */
487 for (i = optind; i < argc; i++) {
488 char *file = argv[i];
489 int const sjval = setjmp(jmpenv);
493 * The file kernel/trace/ftrace.o references the mcount
494 * function but does not call it. Since ftrace.o should
495 * not be traced anyway, we just skip it.
498 if (len >= ftrace_size &&
499 strcmp(file + (len - ftrace_size), ftrace) == 0)
504 fprintf(stderr, "internal error: %s\n", file);
507 case SJ_SETJMP: /* normal sequence */
508 /* Avoid problems if early cleanup() */
514 case SJ_FAIL: /* error in do_file or below */
517 case SJ_SUCCEED: /* premature success */