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 
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 */
57 
58 /* setjmp() return values */
59 enum {
60 	SJ_SETJMP = 0,  /* hardwired first return */
61 	SJ_FAIL,
62 	SJ_SUCCEED
63 };
64 
65 /* Per-file resource cleanup when multiple files. */
66 static void
cleanup(void)67 cleanup(void)
68 {
69 	if (!mmap_failed)
70 		munmap(ehdr_curr, sb.st_size);
71 	else
72 		free(ehdr_curr);
73 	close(fd_map);
74 }
75 
76 static void __attribute__((noreturn))
fail_file(void)77 fail_file(void)
78 {
79 	cleanup();
80 	longjmp(jmpenv, SJ_FAIL);
81 }
82 
83 static void __attribute__((noreturn))
succeed_file(void)84 succeed_file(void)
85 {
86 	cleanup();
87 	longjmp(jmpenv, SJ_SUCCEED);
88 }
89 
90 /* ulseek, uread, ...:  Check return value for errors. */
91 
92 static off_t
ulseek(int const fd,off_t const offset,int const whence)93 ulseek(int const fd, off_t const offset, int const whence)
94 {
95 	off_t const w = lseek(fd, offset, whence);
96 	if (w == (off_t)-1) {
97 		perror("lseek");
98 		fail_file();
99 	}
100 	return w;
101 }
102 
103 static size_t
uread(int const fd,void * const buf,size_t const count)104 uread(int const fd, void *const buf, size_t const count)
105 {
106 	size_t const n = read(fd, buf, count);
107 	if (n != count) {
108 		perror("read");
109 		fail_file();
110 	}
111 	return n;
112 }
113 
114 static size_t
uwrite(int const fd,void const * const buf,size_t const count)115 uwrite(int const fd, void const *const buf, size_t const count)
116 {
117 	size_t const n = write(fd, buf, count);
118 	if (n != count) {
119 		perror("write");
120 		fail_file();
121 	}
122 	return n;
123 }
124 
125 static void *
umalloc(size_t size)126 umalloc(size_t size)
127 {
128 	void *const addr = malloc(size);
129 	if (addr == 0) {
130 		fprintf(stderr, "malloc failed: %zu bytes\n", size);
131 		fail_file();
132 	}
133 	return addr;
134 }
135 
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;
139 
140 static char rel_type_nop;
141 
142 static int (*make_nop)(void *map, size_t const offset);
143 
make_nop_x86(void * map,size_t const offset)144 static int make_nop_x86(void *map, size_t const offset)
145 {
146 	uint32_t *ptr;
147 	unsigned char *op;
148 
149 	/* Confirm we have 0xe8 0x0 0x0 0x0 0x0 */
150 	ptr = map + offset;
151 	if (*ptr != 0)
152 		return -1;
153 
154 	op = map + offset - 1;
155 	if (*op != 0xe8)
156 		return -1;
157 
158 	/* convert to nop */
159 	ulseek(fd_map, offset - 1, SEEK_SET);
160 	uwrite(fd_map, ideal_nop, 5);
161 	return 0;
162 }
163 
164 static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
make_nop_arm64(void * map,size_t const offset)165 static int make_nop_arm64(void *map, size_t const offset)
166 {
167 	uint32_t *ptr;
168 
169 	ptr = map + offset;
170 	/* bl <_mcount> is 0x94000000 before relocation */
171 	if (*ptr != 0x94000000)
172 		return -1;
173 
174 	/* Convert to nop */
175 	ulseek(fd_map, offset, SEEK_SET);
176 	uwrite(fd_map, ideal_nop, 4);
177 	return 0;
178 }
179 
180 /*
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.
192  */
mmap_file(char const * fname)193 static void *mmap_file(char const *fname)
194 {
195 	void *addr;
196 
197 	fd_map = open(fname, O_RDWR);
198 	if (fd_map < 0 || fstat(fd_map, &sb) < 0) {
199 		perror(fname);
200 		fail_file();
201 	}
202 	if (!S_ISREG(sb.st_mode)) {
203 		fprintf(stderr, "not a regular file: %s\n", fname);
204 		fail_file();
205 	}
206 	addr = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
207 		    fd_map, 0);
208 	mmap_failed = 0;
209 	if (addr == MAP_FAILED) {
210 		mmap_failed = 1;
211 		addr = umalloc(sb.st_size);
212 		uread(fd_map, addr, sb.st_size);
213 	}
214 	return addr;
215 }
216 
217 /* w8rev, w8nat, ...: Handle endianness. */
218 
w8rev(uint64_t const x)219 static uint64_t w8rev(uint64_t const x)
220 {
221 	return   ((0xff & (x >> (0 * 8))) << (7 * 8))
222 	       | ((0xff & (x >> (1 * 8))) << (6 * 8))
223 	       | ((0xff & (x >> (2 * 8))) << (5 * 8))
224 	       | ((0xff & (x >> (3 * 8))) << (4 * 8))
225 	       | ((0xff & (x >> (4 * 8))) << (3 * 8))
226 	       | ((0xff & (x >> (5 * 8))) << (2 * 8))
227 	       | ((0xff & (x >> (6 * 8))) << (1 * 8))
228 	       | ((0xff & (x >> (7 * 8))) << (0 * 8));
229 }
230 
w4rev(uint32_t const x)231 static uint32_t w4rev(uint32_t const x)
232 {
233 	return   ((0xff & (x >> (0 * 8))) << (3 * 8))
234 	       | ((0xff & (x >> (1 * 8))) << (2 * 8))
235 	       | ((0xff & (x >> (2 * 8))) << (1 * 8))
236 	       | ((0xff & (x >> (3 * 8))) << (0 * 8));
237 }
238 
w2rev(uint16_t const x)239 static uint32_t w2rev(uint16_t const x)
240 {
241 	return   ((0xff & (x >> (0 * 8))) << (1 * 8))
242 	       | ((0xff & (x >> (1 * 8))) << (0 * 8));
243 }
244 
w8nat(uint64_t const x)245 static uint64_t w8nat(uint64_t const x)
246 {
247 	return x;
248 }
249 
w4nat(uint32_t const x)250 static uint32_t w4nat(uint32_t const x)
251 {
252 	return x;
253 }
254 
w2nat(uint16_t const x)255 static uint32_t w2nat(uint16_t const x)
256 {
257 	return x;
258 }
259 
260 static uint64_t (*w8)(uint64_t);
261 static uint32_t (*w)(uint32_t);
262 static uint32_t (*w2)(uint16_t);
263 
264 /* Names of the sections that could contain calls to mcount. */
265 static int
is_mcounted_section_name(char const * const txtname)266 is_mcounted_section_name(char const *const txtname)
267 {
268 	return strcmp(".text",           txtname) == 0 ||
269 		strcmp(".ref.text",      txtname) == 0 ||
270 		strcmp(".sched.text",    txtname) == 0 ||
271 		strcmp(".spinlock.text", txtname) == 0 ||
272 		strcmp(".irqentry.text", txtname) == 0 ||
273 		strcmp(".kprobes.text", txtname) == 0 ||
274 		strcmp(".text.unlikely", txtname) == 0;
275 }
276 
277 /* 32 bit and 64 bit are very similar */
278 #include "recordmcount.h"
279 #define RECORD_MCOUNT_64
280 #include "recordmcount.h"
281 
282 /* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
283  * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
284  * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
285  * to imply the order of the members; the spec does not say so.
286  *	typedef unsigned char Elf64_Byte;
287  * fails on MIPS64 because their <elf.h> already has it!
288  */
289 
290 typedef uint8_t myElf64_Byte;		/* Type for a 8-bit quantity.  */
291 
292 union mips_r_info {
293 	Elf64_Xword r_info;
294 	struct {
295 		Elf64_Word r_sym;		/* Symbol index.  */
296 		myElf64_Byte r_ssym;		/* Special symbol.  */
297 		myElf64_Byte r_type3;		/* Third relocation.  */
298 		myElf64_Byte r_type2;		/* Second relocation.  */
299 		myElf64_Byte r_type;		/* First relocation.  */
300 	} r_mips;
301 };
302 
MIPS64_r_sym(Elf64_Rel const * rp)303 static uint64_t MIPS64_r_sym(Elf64_Rel const *rp)
304 {
305 	return w(((union mips_r_info){ .r_info = rp->r_info }).r_mips.r_sym);
306 }
307 
MIPS64_r_info(Elf64_Rel * const rp,unsigned sym,unsigned type)308 static void MIPS64_r_info(Elf64_Rel *const rp, unsigned sym, unsigned type)
309 {
310 	rp->r_info = ((union mips_r_info){
311 		.r_mips = { .r_sym = w(sym), .r_type = type }
312 	}).r_info;
313 }
314 
315 static void
do_file(char const * const fname)316 do_file(char const *const fname)
317 {
318 	Elf32_Ehdr *const ehdr = mmap_file(fname);
319 	unsigned int reltype = 0;
320 
321 	ehdr_curr = ehdr;
322 	w = w4nat;
323 	w2 = w2nat;
324 	w8 = w8nat;
325 	switch (ehdr->e_ident[EI_DATA]) {
326 		static unsigned int const endian = 1;
327 	default:
328 		fprintf(stderr, "unrecognized ELF data encoding %d: %s\n",
329 			ehdr->e_ident[EI_DATA], fname);
330 		fail_file();
331 		break;
332 	case ELFDATA2LSB:
333 		if (*(unsigned char const *)&endian != 1) {
334 			/* main() is big endian, file.o is little endian. */
335 			w = w4rev;
336 			w2 = w2rev;
337 			w8 = w8rev;
338 		}
339 		break;
340 	case ELFDATA2MSB:
341 		if (*(unsigned char const *)&endian != 0) {
342 			/* main() is little endian, file.o is big endian. */
343 			w = w4rev;
344 			w2 = w2rev;
345 			w8 = w8rev;
346 		}
347 		break;
348 	}  /* end switch */
349 	if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
350 	||  w2(ehdr->e_type) != ET_REL
351 	||  ehdr->e_ident[EI_VERSION] != EV_CURRENT) {
352 		fprintf(stderr, "unrecognized ET_REL file %s\n", fname);
353 		fail_file();
354 	}
355 
356 	gpfx = 0;
357 	switch (w2(ehdr->e_machine)) {
358 	default:
359 		fprintf(stderr, "unrecognized e_machine %d %s\n",
360 			w2(ehdr->e_machine), fname);
361 		fail_file();
362 		break;
363 	case EM_386:
364 		reltype = R_386_32;
365 		make_nop = make_nop_x86;
366 		ideal_nop = ideal_nop5_x86_32;
367 		mcount_adjust_32 = -1;
368 		break;
369 	case EM_ARM:	 reltype = R_ARM_ABS32;
370 			 altmcount = "__gnu_mcount_nc";
371 			 break;
372 	case EM_AARCH64:
373 			reltype = R_AARCH64_ABS64;
374 			make_nop = make_nop_arm64;
375 			rel_type_nop = R_AARCH64_NONE;
376 			ideal_nop = ideal_nop4_arm64;
377 			gpfx = '_';
378 			break;
379 	case EM_IA_64:	 reltype = R_IA64_IMM64;   gpfx = '_'; break;
380 	case EM_METAG:	 reltype = R_METAG_ADDR32;
381 			 altmcount = "_mcount_wrapper";
382 			 rel_type_nop = R_METAG_NONE;
383 			 /* We happen to have the same requirement as MIPS */
384 			 is_fake_mcount32 = MIPS32_is_fake_mcount;
385 			 break;
386 	case EM_MIPS:	 /* reltype: e_class    */ gpfx = '_'; break;
387 	case EM_PPC:	 reltype = R_PPC_ADDR32;   gpfx = '_'; break;
388 	case EM_PPC64:	 reltype = R_PPC64_ADDR64; gpfx = '_'; break;
389 	case EM_S390:    /* reltype: e_class    */ gpfx = '_'; break;
390 	case EM_SH:	 reltype = R_SH_DIR32;                 break;
391 	case EM_SPARCV9: reltype = R_SPARC_64;     gpfx = '_'; break;
392 	case EM_X86_64:
393 		make_nop = make_nop_x86;
394 		ideal_nop = ideal_nop5_x86_64;
395 		reltype = R_X86_64_64;
396 		mcount_adjust_64 = -1;
397 		break;
398 	}  /* end switch */
399 
400 	switch (ehdr->e_ident[EI_CLASS]) {
401 	default:
402 		fprintf(stderr, "unrecognized ELF class %d %s\n",
403 			ehdr->e_ident[EI_CLASS], fname);
404 		fail_file();
405 		break;
406 	case ELFCLASS32:
407 		if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
408 		||  w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr)) {
409 			fprintf(stderr,
410 				"unrecognized ET_REL file: %s\n", fname);
411 			fail_file();
412 		}
413 		if (w2(ehdr->e_machine) == EM_MIPS) {
414 			reltype = R_MIPS_32;
415 			is_fake_mcount32 = MIPS32_is_fake_mcount;
416 		}
417 		do32(ehdr, fname, reltype);
418 		break;
419 	case ELFCLASS64: {
420 		Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
421 		if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
422 		||  w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr)) {
423 			fprintf(stderr,
424 				"unrecognized ET_REL file: %s\n", fname);
425 			fail_file();
426 		}
427 		if (w2(ghdr->e_machine) == EM_S390) {
428 			reltype = R_390_64;
429 			mcount_adjust_64 = -14;
430 		}
431 		if (w2(ghdr->e_machine) == EM_MIPS) {
432 			reltype = R_MIPS_64;
433 			Elf64_r_sym = MIPS64_r_sym;
434 			Elf64_r_info = MIPS64_r_info;
435 			is_fake_mcount64 = MIPS64_is_fake_mcount;
436 		}
437 		do64(ghdr, fname, reltype);
438 		break;
439 	}
440 	}  /* end switch */
441 
442 	cleanup();
443 }
444 
445 int
main(int argc,char * argv[])446 main(int argc, char *argv[])
447 {
448 	const char ftrace[] = "/ftrace.o";
449 	int ftrace_size = sizeof(ftrace) - 1;
450 	int n_error = 0;  /* gcc-4.3.0 false positive complaint */
451 	int c;
452 	int i;
453 
454 	while ((c = getopt(argc, argv, "w")) >= 0) {
455 		switch (c) {
456 		case 'w':
457 			warn_on_notrace_sect = 1;
458 			break;
459 		default:
460 			fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
461 			return 0;
462 		}
463 	}
464 
465 	if ((argc - optind) < 1) {
466 		fprintf(stderr, "usage: recordmcount [-w] file.o...\n");
467 		return 0;
468 	}
469 
470 	/* Process each file in turn, allowing deep failure. */
471 	for (i = optind; i < argc; i++) {
472 		char *file = argv[i];
473 		int const sjval = setjmp(jmpenv);
474 		int len;
475 
476 		/*
477 		 * The file kernel/trace/ftrace.o references the mcount
478 		 * function but does not call it. Since ftrace.o should
479 		 * not be traced anyway, we just skip it.
480 		 */
481 		len = strlen(file);
482 		if (len >= ftrace_size &&
483 		    strcmp(file + (len - ftrace_size), ftrace) == 0)
484 			continue;
485 
486 		switch (sjval) {
487 		default:
488 			fprintf(stderr, "internal error: %s\n", file);
489 			exit(1);
490 			break;
491 		case SJ_SETJMP:    /* normal sequence */
492 			/* Avoid problems if early cleanup() */
493 			fd_map = -1;
494 			ehdr_curr = NULL;
495 			mmap_failed = 1;
496 			do_file(file);
497 			break;
498 		case SJ_FAIL:    /* error in do_file or below */
499 			++n_error;
500 			break;
501 		case SJ_SUCCEED:    /* premature success */
502 			/* do nothing */
503 			break;
504 		}  /* end switch */
505 	}
506 	return !!n_error;
507 }
508