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 static int fd_map; /* File descriptor for file being modified. */ 44 static int mmap_failed; /* Boolean flag. */ 45 static void *ehdr_curr; /* current ElfXX_Ehdr * for resource cleanup */ 46 static char gpfx; /* prefix for global symbol name (sometimes '_') */ 47 static struct stat sb; /* Remember .st_size, etc. */ 48 static jmp_buf jmpenv; /* setjmp/longjmp per-file error escape */ 49 static const char *altmcount; /* alternate mcount symbol name */ 50 static int warn_on_notrace_sect; /* warn when section has mcount not being recorded */ 51 52 /* setjmp() return values */ 53 enum { 54 SJ_SETJMP = 0, /* hardwired first return */ 55 SJ_FAIL, 56 SJ_SUCCEED 57 }; 58 59 /* Per-file resource cleanup when multiple files. */ 60 static void 61 cleanup(void) 62 { 63 if (!mmap_failed) 64 munmap(ehdr_curr, sb.st_size); 65 else 66 free(ehdr_curr); 67 close(fd_map); 68 } 69 70 static void __attribute__((noreturn)) 71 fail_file(void) 72 { 73 cleanup(); 74 longjmp(jmpenv, SJ_FAIL); 75 } 76 77 static void __attribute__((noreturn)) 78 succeed_file(void) 79 { 80 cleanup(); 81 longjmp(jmpenv, SJ_SUCCEED); 82 } 83 84 /* ulseek, uread, ...: Check return value for errors. */ 85 86 static off_t 87 ulseek(int const fd, off_t const offset, int const whence) 88 { 89 off_t const w = lseek(fd, offset, whence); 90 if (w == (off_t)-1) { 91 perror("lseek"); 92 fail_file(); 93 } 94 return w; 95 } 96 97 static size_t 98 uread(int const fd, void *const buf, size_t const count) 99 { 100 size_t const n = read(fd, buf, count); 101 if (n != count) { 102 perror("read"); 103 fail_file(); 104 } 105 return n; 106 } 107 108 static size_t 109 uwrite(int const fd, void const *const buf, size_t const count) 110 { 111 size_t const n = write(fd, buf, count); 112 if (n != count) { 113 perror("write"); 114 fail_file(); 115 } 116 return n; 117 } 118 119 static void * 120 umalloc(size_t size) 121 { 122 void *const addr = malloc(size); 123 if (addr == 0) { 124 fprintf(stderr, "malloc failed: %zu bytes\n", size); 125 fail_file(); 126 } 127 return addr; 128 } 129 130 static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 }; 131 static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 }; 132 static unsigned char *ideal_nop; 133 134 static char rel_type_nop; 135 136 static int (*make_nop)(void *map, size_t const offset); 137 138 static int make_nop_x86(void *map, size_t const offset) 139 { 140 uint32_t *ptr; 141 unsigned char *op; 142 143 /* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */ 144 ptr = map + offset; 145 if (*ptr != 0) 146 return -1; 147 148 op = map + offset - 1; 149 if (*op != 0xe8) 150 return -1; 151 152 /* convert to nop */ 153 ulseek(fd_map, offset - 1, SEEK_SET); 154 uwrite(fd_map, ideal_nop, 5); 155 return 0; 156 } 157 158 /* 159 * Get the whole file as a programming convenience in order to avoid 160 * malloc+lseek+read+free of many pieces. If successful, then mmap 161 * avoids copying unused pieces; else just read the whole file. 162 * Open for both read and write; new info will be appended to the file. 163 * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr 164 * do not propagate to the file until an explicit overwrite at the last. 165 * This preserves most aspects of consistency (all except .st_size) 166 * for simultaneous readers of the file while we are appending to it. 167 * However, multiple writers still are bad. We choose not to use 168 * locking because it is expensive and the use case of kernel build 169 * makes multiple writers unlikely. 170 */ 171 static void *mmap_file(char const *fname) 172 { 173 void *addr; 174 175 fd_map = open(fname, O_RDWR); 176 if (fd_map < 0 || fstat(fd_map, &sb) < 0) { 177 perror(fname); 178 fail_file(); 179 } 180 if (!S_ISREG(sb.st_mode)) { 181 fprintf(stderr, "not a regular file: %s\n", fname); 182 fail_file(); 183 } 184 addr = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE, 185 fd_map, 0); 186 mmap_failed = 0; 187 if (addr == MAP_FAILED) { 188 mmap_failed = 1; 189 addr = umalloc(sb.st_size); 190 uread(fd_map, addr, sb.st_size); 191 } 192 return addr; 193 } 194 195 /* w8rev, w8nat, ...: Handle endianness. */ 196 197 static uint64_t w8rev(uint64_t const x) 198 { 199 return ((0xff & (x >> (0 * 8))) << (7 * 8)) 200 | ((0xff & (x >> (1 * 8))) << (6 * 8)) 201 | ((0xff & (x >> (2 * 8))) << (5 * 8)) 202 | ((0xff & (x >> (3 * 8))) << (4 * 8)) 203 | ((0xff & (x >> (4 * 8))) << (3 * 8)) 204 | ((0xff & (x >> (5 * 8))) << (2 * 8)) 205 | ((0xff & (x >> (6 * 8))) << (1 * 8)) 206 | ((0xff & (x >> (7 * 8))) << (0 * 8)); 207 } 208 209 static uint32_t w4rev(uint32_t const x) 210 { 211 return ((0xff & (x >> (0 * 8))) << (3 * 8)) 212 | ((0xff & (x >> (1 * 8))) << (2 * 8)) 213 | ((0xff & (x >> (2 * 8))) << (1 * 8)) 214 | ((0xff & (x >> (3 * 8))) << (0 * 8)); 215 } 216 217 static uint32_t w2rev(uint16_t const x) 218 { 219 return ((0xff & (x >> (0 * 8))) << (1 * 8)) 220 | ((0xff & (x >> (1 * 8))) << (0 * 8)); 221 } 222 223 static uint64_t w8nat(uint64_t const x) 224 { 225 return x; 226 } 227 228 static uint32_t w4nat(uint32_t const x) 229 { 230 return x; 231 } 232 233 static uint32_t w2nat(uint16_t const x) 234 { 235 return x; 236 } 237 238 static uint64_t (*w8)(uint64_t); 239 static uint32_t (*w)(uint32_t); 240 static uint32_t (*w2)(uint16_t); 241 242 /* Names of the sections that could contain calls to mcount. */ 243 static int 244 is_mcounted_section_name(char const *const txtname) 245 { 246 return strcmp(".text", txtname) == 0 || 247 strcmp(".ref.text", txtname) == 0 || 248 strcmp(".sched.text", txtname) == 0 || 249 strcmp(".spinlock.text", txtname) == 0 || 250 strcmp(".irqentry.text", txtname) == 0 || 251 strcmp(".kprobes.text", txtname) == 0 || 252 strcmp(".text.unlikely", txtname) == 0; 253 } 254 255 /* 32 bit and 64 bit are very similar */ 256 #include "recordmcount.h" 257 #define RECORD_MCOUNT_64 258 #include "recordmcount.h" 259 260 /* 64-bit EM_MIPS has weird ELF64_Rela.r_info. 261 * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf 262 * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40] 263 * to imply the order of the members; the spec does not say so. 264 * typedef unsigned char Elf64_Byte; 265 * fails on MIPS64 because their <elf.h> already has it! 266 */ 267 268 typedef uint8_t myElf64_Byte; /* Type for a 8-bit quantity. */ 269 270 union mips_r_info { 271 Elf64_Xword r_info; 272 struct { 273 Elf64_Word r_sym; /* Symbol index. */ 274 myElf64_Byte r_ssym; /* Special symbol. */ 275 myElf64_Byte r_type3; /* Third relocation. */ 276 myElf64_Byte r_type2; /* Second relocation. */ 277 myElf64_Byte r_type; /* First relocation. */ 278 } r_mips; 279 }; 280 281 static uint64_t MIPS64_r_sym(Elf64_Rel const *rp) 282 { 283 return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym); 284 } 285 286 static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type) 287 { 288 rp->r_info = ((union mips_r_info){ 289 .r_mips = { .r_sym = w(sym), .r_type = type } 290 }).r_info; 291 } 292 293 static void 294 do_file(char const *const fname) 295 { 296 Elf32_Ehdr *const ehdr = mmap_file(fname); 297 unsigned int reltype = 0; 298 299 ehdr_curr = ehdr; 300 w = w4nat; 301 w2 = w2nat; 302 w8 = w8nat; 303 switch (ehdr->e_ident[EI_DATA]) { 304 static unsigned int const endian = 1; 305 default: 306 fprintf(stderr, "unrecognized ELF data encoding %d: %s\n", 307 ehdr->e_ident[EI_DATA], fname); 308 fail_file(); 309 break; 310 case ELFDATA2LSB: 311 if (*(unsigned char const *)&endian != 1) { 312 /* main() is big endian, file.o is little endian. */ 313 w = w4rev; 314 w2 = w2rev; 315 w8 = w8rev; 316 } 317 break; 318 case ELFDATA2MSB: 319 if (*(unsigned char const *)&endian != 0) { 320 /* main() is little endian, file.o is big endian. */ 321 w = w4rev; 322 w2 = w2rev; 323 w8 = w8rev; 324 } 325 break; 326 } /* end switch */ 327 if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0 328 || w2(ehdr->e_type) != ET_REL 329 || ehdr->e_ident[EI_VERSION] != EV_CURRENT) { 330 fprintf(stderr, "unrecognized ET_REL file %s\n", fname); 331 fail_file(); 332 } 333 334 gpfx = 0; 335 switch (w2(ehdr->e_machine)) { 336 default: 337 fprintf(stderr, "unrecognized e_machine %d %s\n", 338 w2(ehdr->e_machine), fname); 339 fail_file(); 340 break; 341 case EM_386: 342 reltype = R_386_32; 343 make_nop = make_nop_x86; 344 ideal_nop = ideal_nop5_x86_32; 345 mcount_adjust_32 = -1; 346 break; 347 case EM_ARM: reltype = R_ARM_ABS32; 348 altmcount = "__gnu_mcount_nc"; 349 break; 350 case EM_IA_64: reltype = R_IA64_IMM64; gpfx = '_'; break; 351 case EM_METAG: reltype = R_METAG_ADDR32; 352 altmcount = "_mcount_wrapper"; 353 rel_type_nop = R_METAG_NONE; 354 /* We happen to have the same requirement as MIPS */ 355 is_fake_mcount32 = MIPS32_is_fake_mcount; 356 break; 357 case EM_MIPS: /* reltype: e_class */ gpfx = '_'; break; 358 case EM_PPC: reltype = R_PPC_ADDR32; gpfx = '_'; break; 359 case EM_PPC64: reltype = R_PPC64_ADDR64; gpfx = '_'; break; 360 case EM_S390: /* reltype: e_class */ gpfx = '_'; break; 361 case EM_SH: reltype = R_SH_DIR32; break; 362 case EM_SPARCV9: reltype = R_SPARC_64; gpfx = '_'; break; 363 case EM_X86_64: 364 make_nop = make_nop_x86; 365 ideal_nop = ideal_nop5_x86_64; 366 reltype = R_X86_64_64; 367 mcount_adjust_64 = -1; 368 break; 369 } /* end switch */ 370 371 switch (ehdr->e_ident[EI_CLASS]) { 372 default: 373 fprintf(stderr, "unrecognized ELF class %d %s\n", 374 ehdr->e_ident[EI_CLASS], fname); 375 fail_file(); 376 break; 377 case ELFCLASS32: 378 if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr) 379 || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) { 380 fprintf(stderr, 381 "unrecognized ET_REL file: %s\n", fname); 382 fail_file(); 383 } 384 if (w2(ehdr->e_machine) == EM_S390) { 385 reltype = R_390_32; 386 mcount_adjust_32 = -4; 387 } 388 if (w2(ehdr->e_machine) == EM_MIPS) { 389 reltype = R_MIPS_32; 390 is_fake_mcount32 = MIPS32_is_fake_mcount; 391 } 392 do32(ehdr, fname, reltype); 393 break; 394 case ELFCLASS64: { 395 Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr; 396 if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr) 397 || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) { 398 fprintf(stderr, 399 "unrecognized ET_REL file: %s\n", fname); 400 fail_file(); 401 } 402 if (w2(ghdr->e_machine) == EM_S390) { 403 reltype = R_390_64; 404 mcount_adjust_64 = -8; 405 } 406 if (w2(ghdr->e_machine) == EM_MIPS) { 407 reltype = R_MIPS_64; 408 Elf64_r_sym = MIPS64_r_sym; 409 Elf64_r_info = MIPS64_r_info; 410 is_fake_mcount64 = MIPS64_is_fake_mcount; 411 } 412 do64(ghdr, fname, reltype); 413 break; 414 } 415 } /* end switch */ 416 417 cleanup(); 418 } 419 420 int 421 main(int argc, char *argv[]) 422 { 423 const char ftrace[] = "/ftrace.o"; 424 int ftrace_size = sizeof(ftrace) - 1; 425 int n_error = 0; /* gcc-4.3.0 false positive complaint */ 426 int c; 427 int i; 428 429 while ((c = getopt(argc, argv, "w")) >= 0) { 430 switch (c) { 431 case 'w': 432 warn_on_notrace_sect = 1; 433 break; 434 default: 435 fprintf(stderr, "usage: recordmcount [-w] file.o...\n"); 436 return 0; 437 } 438 } 439 440 if ((argc - optind) < 1) { 441 fprintf(stderr, "usage: recordmcount [-w] file.o...\n"); 442 return 0; 443 } 444 445 /* Process each file in turn, allowing deep failure. */ 446 for (i = optind; i < argc; i++) { 447 char *file = argv[i]; 448 int const sjval = setjmp(jmpenv); 449 int len; 450 451 /* 452 * The file kernel/trace/ftrace.o references the mcount 453 * function but does not call it. Since ftrace.o should 454 * not be traced anyway, we just skip it. 455 */ 456 len = strlen(file); 457 if (len >= ftrace_size && 458 strcmp(file + (len - ftrace_size), ftrace) == 0) 459 continue; 460 461 switch (sjval) { 462 default: 463 fprintf(stderr, "internal error: %s\n", file); 464 exit(1); 465 break; 466 case SJ_SETJMP: /* normal sequence */ 467 /* Avoid problems if early cleanup() */ 468 fd_map = -1; 469 ehdr_curr = NULL; 470 mmap_failed = 1; 471 do_file(file); 472 break; 473 case SJ_FAIL: /* error in do_file or below */ 474 ++n_error; 475 break; 476 case SJ_SUCCEED: /* premature success */ 477 /* do nothing */ 478 break; 479 } /* end switch */ 480 } 481 return !!n_error; 482 } 483 484 485