| 1 | /* |
| 2 | * Dynamic function tracing support. |
| 3 | * |
| 4 | * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> |
| 5 | * |
| 6 | * Thanks goes to Ingo Molnar, for suggesting the idea. |
| 7 | * Mathieu Desnoyers, for suggesting postponing the modifications. |
| 8 | * Arjan van de Ven, for keeping me straight, and explaining to me |
| 9 | * the dangers of modifying code on the run. |
| 10 | */ |
| 11 | |
| 12 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 13 | |
| 14 | #include <linux/spinlock.h> |
| 15 | #include <linux/hardirq.h> |
| 16 | #include <linux/uaccess.h> |
| 17 | #include <linux/ftrace.h> |
| 18 | #include <linux/percpu.h> |
| 19 | #include <linux/sched.h> |
| 20 | #include <linux/slab.h> |
| 21 | #include <linux/init.h> |
| 22 | #include <linux/list.h> |
| 23 | #include <linux/module.h> |
| 24 | |
| 25 | #include <trace/syscall.h> |
| 26 | |
| 27 | #include <asm/cacheflush.h> |
| 28 | #include <asm/kprobes.h> |
| 29 | #include <asm/ftrace.h> |
| 30 | #include <asm/nops.h> |
| 31 | |
| 32 | #ifdef CONFIG_DYNAMIC_FTRACE |
| 33 | |
| 34 | int ftrace_arch_code_modify_prepare(void) |
| 35 | { |
| 36 | set_kernel_text_rw(); |
| 37 | set_all_modules_text_rw(); |
| 38 | return 0; |
| 39 | } |
| 40 | |
| 41 | int ftrace_arch_code_modify_post_process(void) |
| 42 | { |
| 43 | set_all_modules_text_ro(); |
| 44 | set_kernel_text_ro(); |
| 45 | return 0; |
| 46 | } |
| 47 | |
| 48 | union ftrace_code_union { |
| 49 | char code[MCOUNT_INSN_SIZE]; |
| 50 | struct { |
| 51 | unsigned char e8; |
| 52 | int offset; |
| 53 | } __attribute__((packed)); |
| 54 | }; |
| 55 | |
| 56 | static int ftrace_calc_offset(long ip, long addr) |
| 57 | { |
| 58 | return (int)(addr - ip); |
| 59 | } |
| 60 | |
| 61 | static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) |
| 62 | { |
| 63 | static union ftrace_code_union calc; |
| 64 | |
| 65 | calc.e8 = 0xe8; |
| 66 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
| 67 | |
| 68 | /* |
| 69 | * No locking needed, this must be called via kstop_machine |
| 70 | * which in essence is like running on a uniprocessor machine. |
| 71 | */ |
| 72 | return calc.code; |
| 73 | } |
| 74 | |
| 75 | static inline int |
| 76 | within(unsigned long addr, unsigned long start, unsigned long end) |
| 77 | { |
| 78 | return addr >= start && addr < end; |
| 79 | } |
| 80 | |
| 81 | static unsigned long text_ip_addr(unsigned long ip) |
| 82 | { |
| 83 | /* |
| 84 | * On x86_64, kernel text mappings are mapped read-only, so we use |
| 85 | * the kernel identity mapping instead of the kernel text mapping |
| 86 | * to modify the kernel text. |
| 87 | * |
| 88 | * For 32bit kernels, these mappings are same and we can use |
| 89 | * kernel identity mapping to modify code. |
| 90 | */ |
| 91 | if (within(ip, (unsigned long)_text, (unsigned long)_etext)) |
| 92 | ip = (unsigned long)__va(__pa_symbol(ip)); |
| 93 | |
| 94 | return ip; |
| 95 | } |
| 96 | |
| 97 | static const unsigned char *ftrace_nop_replace(void) |
| 98 | { |
| 99 | return ideal_nops[NOP_ATOMIC5]; |
| 100 | } |
| 101 | |
| 102 | static int |
| 103 | ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code, |
| 104 | unsigned const char *new_code) |
| 105 | { |
| 106 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
| 107 | |
| 108 | ftrace_expected = old_code; |
| 109 | |
| 110 | /* |
| 111 | * Note: |
| 112 | * We are paranoid about modifying text, as if a bug was to happen, it |
| 113 | * could cause us to read or write to someplace that could cause harm. |
| 114 | * Carefully read and modify the code with probe_kernel_*(), and make |
| 115 | * sure what we read is what we expected it to be before modifying it. |
| 116 | */ |
| 117 | |
| 118 | /* read the text we want to modify */ |
| 119 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
| 120 | return -EFAULT; |
| 121 | |
| 122 | /* Make sure it is what we expect it to be */ |
| 123 | if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) |
| 124 | return -EINVAL; |
| 125 | |
| 126 | ip = text_ip_addr(ip); |
| 127 | |
| 128 | /* replace the text with the new text */ |
| 129 | if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE)) |
| 130 | return -EPERM; |
| 131 | |
| 132 | sync_core(); |
| 133 | |
| 134 | return 0; |
| 135 | } |
| 136 | |
| 137 | int ftrace_make_nop(struct module *mod, |
| 138 | struct dyn_ftrace *rec, unsigned long addr) |
| 139 | { |
| 140 | unsigned const char *new, *old; |
| 141 | unsigned long ip = rec->ip; |
| 142 | |
| 143 | old = ftrace_call_replace(ip, addr); |
| 144 | new = ftrace_nop_replace(); |
| 145 | |
| 146 | /* |
| 147 | * On boot up, and when modules are loaded, the MCOUNT_ADDR |
| 148 | * is converted to a nop, and will never become MCOUNT_ADDR |
| 149 | * again. This code is either running before SMP (on boot up) |
| 150 | * or before the code will ever be executed (module load). |
| 151 | * We do not want to use the breakpoint version in this case, |
| 152 | * just modify the code directly. |
| 153 | */ |
| 154 | if (addr == MCOUNT_ADDR) |
| 155 | return ftrace_modify_code_direct(rec->ip, old, new); |
| 156 | |
| 157 | ftrace_expected = NULL; |
| 158 | |
| 159 | /* Normal cases use add_brk_on_nop */ |
| 160 | WARN_ONCE(1, "invalid use of ftrace_make_nop"); |
| 161 | return -EINVAL; |
| 162 | } |
| 163 | |
| 164 | int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) |
| 165 | { |
| 166 | unsigned const char *new, *old; |
| 167 | unsigned long ip = rec->ip; |
| 168 | |
| 169 | old = ftrace_nop_replace(); |
| 170 | new = ftrace_call_replace(ip, addr); |
| 171 | |
| 172 | /* Should only be called when module is loaded */ |
| 173 | return ftrace_modify_code_direct(rec->ip, old, new); |
| 174 | } |
| 175 | |
| 176 | /* |
| 177 | * The modifying_ftrace_code is used to tell the breakpoint |
| 178 | * handler to call ftrace_int3_handler(). If it fails to |
| 179 | * call this handler for a breakpoint added by ftrace, then |
| 180 | * the kernel may crash. |
| 181 | * |
| 182 | * As atomic_writes on x86 do not need a barrier, we do not |
| 183 | * need to add smp_mb()s for this to work. It is also considered |
| 184 | * that we can not read the modifying_ftrace_code before |
| 185 | * executing the breakpoint. That would be quite remarkable if |
| 186 | * it could do that. Here's the flow that is required: |
| 187 | * |
| 188 | * CPU-0 CPU-1 |
| 189 | * |
| 190 | * atomic_inc(mfc); |
| 191 | * write int3s |
| 192 | * <trap-int3> // implicit (r)mb |
| 193 | * if (atomic_read(mfc)) |
| 194 | * call ftrace_int3_handler() |
| 195 | * |
| 196 | * Then when we are finished: |
| 197 | * |
| 198 | * atomic_dec(mfc); |
| 199 | * |
| 200 | * If we hit a breakpoint that was not set by ftrace, it does not |
| 201 | * matter if ftrace_int3_handler() is called or not. It will |
| 202 | * simply be ignored. But it is crucial that a ftrace nop/caller |
| 203 | * breakpoint is handled. No other user should ever place a |
| 204 | * breakpoint on an ftrace nop/caller location. It must only |
| 205 | * be done by this code. |
| 206 | */ |
| 207 | atomic_t modifying_ftrace_code __read_mostly; |
| 208 | |
| 209 | static int |
| 210 | ftrace_modify_code(unsigned long ip, unsigned const char *old_code, |
| 211 | unsigned const char *new_code); |
| 212 | |
| 213 | /* |
| 214 | * Should never be called: |
| 215 | * As it is only called by __ftrace_replace_code() which is called by |
| 216 | * ftrace_replace_code() that x86 overrides, and by ftrace_update_code() |
| 217 | * which is called to turn mcount into nops or nops into function calls |
| 218 | * but not to convert a function from not using regs to one that uses |
| 219 | * regs, which ftrace_modify_call() is for. |
| 220 | */ |
| 221 | int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr, |
| 222 | unsigned long addr) |
| 223 | { |
| 224 | WARN_ON(1); |
| 225 | ftrace_expected = NULL; |
| 226 | return -EINVAL; |
| 227 | } |
| 228 | |
| 229 | static unsigned long ftrace_update_func; |
| 230 | |
| 231 | static int update_ftrace_func(unsigned long ip, void *new) |
| 232 | { |
| 233 | unsigned char old[MCOUNT_INSN_SIZE]; |
| 234 | int ret; |
| 235 | |
| 236 | memcpy(old, (void *)ip, MCOUNT_INSN_SIZE); |
| 237 | |
| 238 | ftrace_update_func = ip; |
| 239 | /* Make sure the breakpoints see the ftrace_update_func update */ |
| 240 | smp_wmb(); |
| 241 | |
| 242 | /* See comment above by declaration of modifying_ftrace_code */ |
| 243 | atomic_inc(&modifying_ftrace_code); |
| 244 | |
| 245 | ret = ftrace_modify_code(ip, old, new); |
| 246 | |
| 247 | atomic_dec(&modifying_ftrace_code); |
| 248 | |
| 249 | return ret; |
| 250 | } |
| 251 | |
| 252 | int ftrace_update_ftrace_func(ftrace_func_t func) |
| 253 | { |
| 254 | unsigned long ip = (unsigned long)(&ftrace_call); |
| 255 | unsigned char *new; |
| 256 | int ret; |
| 257 | |
| 258 | new = ftrace_call_replace(ip, (unsigned long)func); |
| 259 | ret = update_ftrace_func(ip, new); |
| 260 | |
| 261 | /* Also update the regs callback function */ |
| 262 | if (!ret) { |
| 263 | ip = (unsigned long)(&ftrace_regs_call); |
| 264 | new = ftrace_call_replace(ip, (unsigned long)func); |
| 265 | ret = update_ftrace_func(ip, new); |
| 266 | } |
| 267 | |
| 268 | return ret; |
| 269 | } |
| 270 | |
| 271 | static int is_ftrace_caller(unsigned long ip) |
| 272 | { |
| 273 | if (ip == ftrace_update_func) |
| 274 | return 1; |
| 275 | |
| 276 | return 0; |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * A breakpoint was added to the code address we are about to |
| 281 | * modify, and this is the handle that will just skip over it. |
| 282 | * We are either changing a nop into a trace call, or a trace |
| 283 | * call to a nop. While the change is taking place, we treat |
| 284 | * it just like it was a nop. |
| 285 | */ |
| 286 | int ftrace_int3_handler(struct pt_regs *regs) |
| 287 | { |
| 288 | unsigned long ip; |
| 289 | |
| 290 | if (WARN_ON_ONCE(!regs)) |
| 291 | return 0; |
| 292 | |
| 293 | ip = regs->ip - 1; |
| 294 | if (!ftrace_location(ip) && !is_ftrace_caller(ip)) |
| 295 | return 0; |
| 296 | |
| 297 | regs->ip += MCOUNT_INSN_SIZE - 1; |
| 298 | |
| 299 | return 1; |
| 300 | } |
| 301 | |
| 302 | static int ftrace_write(unsigned long ip, const char *val, int size) |
| 303 | { |
| 304 | ip = text_ip_addr(ip); |
| 305 | |
| 306 | if (probe_kernel_write((void *)ip, val, size)) |
| 307 | return -EPERM; |
| 308 | |
| 309 | return 0; |
| 310 | } |
| 311 | |
| 312 | static int add_break(unsigned long ip, const char *old) |
| 313 | { |
| 314 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
| 315 | unsigned char brk = BREAKPOINT_INSTRUCTION; |
| 316 | |
| 317 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
| 318 | return -EFAULT; |
| 319 | |
| 320 | ftrace_expected = old; |
| 321 | |
| 322 | /* Make sure it is what we expect it to be */ |
| 323 | if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0) |
| 324 | return -EINVAL; |
| 325 | |
| 326 | return ftrace_write(ip, &brk, 1); |
| 327 | } |
| 328 | |
| 329 | static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr) |
| 330 | { |
| 331 | unsigned const char *old; |
| 332 | unsigned long ip = rec->ip; |
| 333 | |
| 334 | old = ftrace_call_replace(ip, addr); |
| 335 | |
| 336 | return add_break(rec->ip, old); |
| 337 | } |
| 338 | |
| 339 | |
| 340 | static int add_brk_on_nop(struct dyn_ftrace *rec) |
| 341 | { |
| 342 | unsigned const char *old; |
| 343 | |
| 344 | old = ftrace_nop_replace(); |
| 345 | |
| 346 | return add_break(rec->ip, old); |
| 347 | } |
| 348 | |
| 349 | static int add_breakpoints(struct dyn_ftrace *rec, int enable) |
| 350 | { |
| 351 | unsigned long ftrace_addr; |
| 352 | int ret; |
| 353 | |
| 354 | ftrace_addr = ftrace_get_addr_curr(rec); |
| 355 | |
| 356 | ret = ftrace_test_record(rec, enable); |
| 357 | |
| 358 | switch (ret) { |
| 359 | case FTRACE_UPDATE_IGNORE: |
| 360 | return 0; |
| 361 | |
| 362 | case FTRACE_UPDATE_MAKE_CALL: |
| 363 | /* converting nop to call */ |
| 364 | return add_brk_on_nop(rec); |
| 365 | |
| 366 | case FTRACE_UPDATE_MODIFY_CALL: |
| 367 | case FTRACE_UPDATE_MAKE_NOP: |
| 368 | /* converting a call to a nop */ |
| 369 | return add_brk_on_call(rec, ftrace_addr); |
| 370 | } |
| 371 | return 0; |
| 372 | } |
| 373 | |
| 374 | /* |
| 375 | * On error, we need to remove breakpoints. This needs to |
| 376 | * be done caefully. If the address does not currently have a |
| 377 | * breakpoint, we know we are done. Otherwise, we look at the |
| 378 | * remaining 4 bytes of the instruction. If it matches a nop |
| 379 | * we replace the breakpoint with the nop. Otherwise we replace |
| 380 | * it with the call instruction. |
| 381 | */ |
| 382 | static int remove_breakpoint(struct dyn_ftrace *rec) |
| 383 | { |
| 384 | unsigned char ins[MCOUNT_INSN_SIZE]; |
| 385 | unsigned char brk = BREAKPOINT_INSTRUCTION; |
| 386 | const unsigned char *nop; |
| 387 | unsigned long ftrace_addr; |
| 388 | unsigned long ip = rec->ip; |
| 389 | |
| 390 | /* If we fail the read, just give up */ |
| 391 | if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE)) |
| 392 | return -EFAULT; |
| 393 | |
| 394 | /* If this does not have a breakpoint, we are done */ |
| 395 | if (ins[0] != brk) |
| 396 | return 0; |
| 397 | |
| 398 | nop = ftrace_nop_replace(); |
| 399 | |
| 400 | /* |
| 401 | * If the last 4 bytes of the instruction do not match |
| 402 | * a nop, then we assume that this is a call to ftrace_addr. |
| 403 | */ |
| 404 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) { |
| 405 | /* |
| 406 | * For extra paranoidism, we check if the breakpoint is on |
| 407 | * a call that would actually jump to the ftrace_addr. |
| 408 | * If not, don't touch the breakpoint, we make just create |
| 409 | * a disaster. |
| 410 | */ |
| 411 | ftrace_addr = ftrace_get_addr_new(rec); |
| 412 | nop = ftrace_call_replace(ip, ftrace_addr); |
| 413 | |
| 414 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0) |
| 415 | goto update; |
| 416 | |
| 417 | /* Check both ftrace_addr and ftrace_old_addr */ |
| 418 | ftrace_addr = ftrace_get_addr_curr(rec); |
| 419 | nop = ftrace_call_replace(ip, ftrace_addr); |
| 420 | |
| 421 | ftrace_expected = nop; |
| 422 | |
| 423 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) |
| 424 | return -EINVAL; |
| 425 | } |
| 426 | |
| 427 | update: |
| 428 | return ftrace_write(ip, nop, 1); |
| 429 | } |
| 430 | |
| 431 | static int add_update_code(unsigned long ip, unsigned const char *new) |
| 432 | { |
| 433 | /* skip breakpoint */ |
| 434 | ip++; |
| 435 | new++; |
| 436 | return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1); |
| 437 | } |
| 438 | |
| 439 | static int add_update_call(struct dyn_ftrace *rec, unsigned long addr) |
| 440 | { |
| 441 | unsigned long ip = rec->ip; |
| 442 | unsigned const char *new; |
| 443 | |
| 444 | new = ftrace_call_replace(ip, addr); |
| 445 | return add_update_code(ip, new); |
| 446 | } |
| 447 | |
| 448 | static int add_update_nop(struct dyn_ftrace *rec) |
| 449 | { |
| 450 | unsigned long ip = rec->ip; |
| 451 | unsigned const char *new; |
| 452 | |
| 453 | new = ftrace_nop_replace(); |
| 454 | return add_update_code(ip, new); |
| 455 | } |
| 456 | |
| 457 | static int add_update(struct dyn_ftrace *rec, int enable) |
| 458 | { |
| 459 | unsigned long ftrace_addr; |
| 460 | int ret; |
| 461 | |
| 462 | ret = ftrace_test_record(rec, enable); |
| 463 | |
| 464 | ftrace_addr = ftrace_get_addr_new(rec); |
| 465 | |
| 466 | switch (ret) { |
| 467 | case FTRACE_UPDATE_IGNORE: |
| 468 | return 0; |
| 469 | |
| 470 | case FTRACE_UPDATE_MODIFY_CALL: |
| 471 | case FTRACE_UPDATE_MAKE_CALL: |
| 472 | /* converting nop to call */ |
| 473 | return add_update_call(rec, ftrace_addr); |
| 474 | |
| 475 | case FTRACE_UPDATE_MAKE_NOP: |
| 476 | /* converting a call to a nop */ |
| 477 | return add_update_nop(rec); |
| 478 | } |
| 479 | |
| 480 | return 0; |
| 481 | } |
| 482 | |
| 483 | static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr) |
| 484 | { |
| 485 | unsigned long ip = rec->ip; |
| 486 | unsigned const char *new; |
| 487 | |
| 488 | new = ftrace_call_replace(ip, addr); |
| 489 | |
| 490 | return ftrace_write(ip, new, 1); |
| 491 | } |
| 492 | |
| 493 | static int finish_update_nop(struct dyn_ftrace *rec) |
| 494 | { |
| 495 | unsigned long ip = rec->ip; |
| 496 | unsigned const char *new; |
| 497 | |
| 498 | new = ftrace_nop_replace(); |
| 499 | |
| 500 | return ftrace_write(ip, new, 1); |
| 501 | } |
| 502 | |
| 503 | static int finish_update(struct dyn_ftrace *rec, int enable) |
| 504 | { |
| 505 | unsigned long ftrace_addr; |
| 506 | int ret; |
| 507 | |
| 508 | ret = ftrace_update_record(rec, enable); |
| 509 | |
| 510 | ftrace_addr = ftrace_get_addr_new(rec); |
| 511 | |
| 512 | switch (ret) { |
| 513 | case FTRACE_UPDATE_IGNORE: |
| 514 | return 0; |
| 515 | |
| 516 | case FTRACE_UPDATE_MODIFY_CALL: |
| 517 | case FTRACE_UPDATE_MAKE_CALL: |
| 518 | /* converting nop to call */ |
| 519 | return finish_update_call(rec, ftrace_addr); |
| 520 | |
| 521 | case FTRACE_UPDATE_MAKE_NOP: |
| 522 | /* converting a call to a nop */ |
| 523 | return finish_update_nop(rec); |
| 524 | } |
| 525 | |
| 526 | return 0; |
| 527 | } |
| 528 | |
| 529 | static void do_sync_core(void *data) |
| 530 | { |
| 531 | sync_core(); |
| 532 | } |
| 533 | |
| 534 | static void run_sync(void) |
| 535 | { |
| 536 | int enable_irqs = irqs_disabled(); |
| 537 | |
| 538 | /* We may be called with interrupts disbled (on bootup). */ |
| 539 | if (enable_irqs) |
| 540 | local_irq_enable(); |
| 541 | on_each_cpu(do_sync_core, NULL, 1); |
| 542 | if (enable_irqs) |
| 543 | local_irq_disable(); |
| 544 | } |
| 545 | |
| 546 | void ftrace_replace_code(int enable) |
| 547 | { |
| 548 | struct ftrace_rec_iter *iter; |
| 549 | struct dyn_ftrace *rec; |
| 550 | const char *report = "adding breakpoints"; |
| 551 | int count = 0; |
| 552 | int ret; |
| 553 | |
| 554 | for_ftrace_rec_iter(iter) { |
| 555 | rec = ftrace_rec_iter_record(iter); |
| 556 | |
| 557 | ret = add_breakpoints(rec, enable); |
| 558 | if (ret) |
| 559 | goto remove_breakpoints; |
| 560 | count++; |
| 561 | } |
| 562 | |
| 563 | run_sync(); |
| 564 | |
| 565 | report = "updating code"; |
| 566 | count = 0; |
| 567 | |
| 568 | for_ftrace_rec_iter(iter) { |
| 569 | rec = ftrace_rec_iter_record(iter); |
| 570 | |
| 571 | ret = add_update(rec, enable); |
| 572 | if (ret) |
| 573 | goto remove_breakpoints; |
| 574 | count++; |
| 575 | } |
| 576 | |
| 577 | run_sync(); |
| 578 | |
| 579 | report = "removing breakpoints"; |
| 580 | count = 0; |
| 581 | |
| 582 | for_ftrace_rec_iter(iter) { |
| 583 | rec = ftrace_rec_iter_record(iter); |
| 584 | |
| 585 | ret = finish_update(rec, enable); |
| 586 | if (ret) |
| 587 | goto remove_breakpoints; |
| 588 | count++; |
| 589 | } |
| 590 | |
| 591 | run_sync(); |
| 592 | |
| 593 | return; |
| 594 | |
| 595 | remove_breakpoints: |
| 596 | pr_warn("Failed on %s (%d):\n", report, count); |
| 597 | ftrace_bug(ret, rec); |
| 598 | for_ftrace_rec_iter(iter) { |
| 599 | rec = ftrace_rec_iter_record(iter); |
| 600 | /* |
| 601 | * Breakpoints are handled only when this function is in |
| 602 | * progress. The system could not work with them. |
| 603 | */ |
| 604 | if (remove_breakpoint(rec)) |
| 605 | BUG(); |
| 606 | } |
| 607 | run_sync(); |
| 608 | } |
| 609 | |
| 610 | static int |
| 611 | ftrace_modify_code(unsigned long ip, unsigned const char *old_code, |
| 612 | unsigned const char *new_code) |
| 613 | { |
| 614 | int ret; |
| 615 | |
| 616 | ret = add_break(ip, old_code); |
| 617 | if (ret) |
| 618 | goto out; |
| 619 | |
| 620 | run_sync(); |
| 621 | |
| 622 | ret = add_update_code(ip, new_code); |
| 623 | if (ret) |
| 624 | goto fail_update; |
| 625 | |
| 626 | run_sync(); |
| 627 | |
| 628 | ret = ftrace_write(ip, new_code, 1); |
| 629 | /* |
| 630 | * The breakpoint is handled only when this function is in progress. |
| 631 | * The system could not work if we could not remove it. |
| 632 | */ |
| 633 | BUG_ON(ret); |
| 634 | out: |
| 635 | run_sync(); |
| 636 | return ret; |
| 637 | |
| 638 | fail_update: |
| 639 | /* Also here the system could not work with the breakpoint */ |
| 640 | if (ftrace_write(ip, old_code, 1)) |
| 641 | BUG(); |
| 642 | goto out; |
| 643 | } |
| 644 | |
| 645 | void arch_ftrace_update_code(int command) |
| 646 | { |
| 647 | /* See comment above by declaration of modifying_ftrace_code */ |
| 648 | atomic_inc(&modifying_ftrace_code); |
| 649 | |
| 650 | ftrace_modify_all_code(command); |
| 651 | |
| 652 | atomic_dec(&modifying_ftrace_code); |
| 653 | } |
| 654 | |
| 655 | int __init ftrace_dyn_arch_init(void) |
| 656 | { |
| 657 | return 0; |
| 658 | } |
| 659 | |
| 660 | #if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER) |
| 661 | static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr) |
| 662 | { |
| 663 | static union ftrace_code_union calc; |
| 664 | |
| 665 | /* Jmp not a call (ignore the .e8) */ |
| 666 | calc.e8 = 0xe9; |
| 667 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
| 668 | |
| 669 | /* |
| 670 | * ftrace external locks synchronize the access to the static variable. |
| 671 | */ |
| 672 | return calc.code; |
| 673 | } |
| 674 | #endif |
| 675 | |
| 676 | /* Currently only x86_64 supports dynamic trampolines */ |
| 677 | #ifdef CONFIG_X86_64 |
| 678 | |
| 679 | #ifdef CONFIG_MODULES |
| 680 | #include <linux/moduleloader.h> |
| 681 | /* Module allocation simplifies allocating memory for code */ |
| 682 | static inline void *alloc_tramp(unsigned long size) |
| 683 | { |
| 684 | return module_alloc(size); |
| 685 | } |
| 686 | static inline void tramp_free(void *tramp) |
| 687 | { |
| 688 | module_memfree(tramp); |
| 689 | } |
| 690 | #else |
| 691 | /* Trampolines can only be created if modules are supported */ |
| 692 | static inline void *alloc_tramp(unsigned long size) |
| 693 | { |
| 694 | return NULL; |
| 695 | } |
| 696 | static inline void tramp_free(void *tramp) { } |
| 697 | #endif |
| 698 | |
| 699 | /* Defined as markers to the end of the ftrace default trampolines */ |
| 700 | extern void ftrace_regs_caller_end(void); |
| 701 | extern void ftrace_epilogue(void); |
| 702 | extern void ftrace_caller_op_ptr(void); |
| 703 | extern void ftrace_regs_caller_op_ptr(void); |
| 704 | |
| 705 | /* movq function_trace_op(%rip), %rdx */ |
| 706 | /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */ |
| 707 | #define OP_REF_SIZE 7 |
| 708 | |
| 709 | /* |
| 710 | * The ftrace_ops is passed to the function callback. Since the |
| 711 | * trampoline only services a single ftrace_ops, we can pass in |
| 712 | * that ops directly. |
| 713 | * |
| 714 | * The ftrace_op_code_union is used to create a pointer to the |
| 715 | * ftrace_ops that will be passed to the callback function. |
| 716 | */ |
| 717 | union ftrace_op_code_union { |
| 718 | char code[OP_REF_SIZE]; |
| 719 | struct { |
| 720 | char op[3]; |
| 721 | int offset; |
| 722 | } __attribute__((packed)); |
| 723 | }; |
| 724 | |
| 725 | static unsigned long |
| 726 | create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size) |
| 727 | { |
| 728 | unsigned const char *jmp; |
| 729 | unsigned long start_offset; |
| 730 | unsigned long end_offset; |
| 731 | unsigned long op_offset; |
| 732 | unsigned long offset; |
| 733 | unsigned long size; |
| 734 | unsigned long ip; |
| 735 | unsigned long *ptr; |
| 736 | void *trampoline; |
| 737 | /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */ |
| 738 | unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 }; |
| 739 | union ftrace_op_code_union op_ptr; |
| 740 | int ret; |
| 741 | |
| 742 | if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { |
| 743 | start_offset = (unsigned long)ftrace_regs_caller; |
| 744 | end_offset = (unsigned long)ftrace_regs_caller_end; |
| 745 | op_offset = (unsigned long)ftrace_regs_caller_op_ptr; |
| 746 | } else { |
| 747 | start_offset = (unsigned long)ftrace_caller; |
| 748 | end_offset = (unsigned long)ftrace_epilogue; |
| 749 | op_offset = (unsigned long)ftrace_caller_op_ptr; |
| 750 | } |
| 751 | |
| 752 | size = end_offset - start_offset; |
| 753 | |
| 754 | /* |
| 755 | * Allocate enough size to store the ftrace_caller code, |
| 756 | * the jmp to ftrace_epilogue, as well as the address of |
| 757 | * the ftrace_ops this trampoline is used for. |
| 758 | */ |
| 759 | trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *)); |
| 760 | if (!trampoline) |
| 761 | return 0; |
| 762 | |
| 763 | *tramp_size = size + MCOUNT_INSN_SIZE + sizeof(void *); |
| 764 | |
| 765 | /* Copy ftrace_caller onto the trampoline memory */ |
| 766 | ret = probe_kernel_read(trampoline, (void *)start_offset, size); |
| 767 | if (WARN_ON(ret < 0)) { |
| 768 | tramp_free(trampoline); |
| 769 | return 0; |
| 770 | } |
| 771 | |
| 772 | ip = (unsigned long)trampoline + size; |
| 773 | |
| 774 | /* The trampoline ends with a jmp to ftrace_epilogue */ |
| 775 | jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_epilogue); |
| 776 | memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE); |
| 777 | |
| 778 | /* |
| 779 | * The address of the ftrace_ops that is used for this trampoline |
| 780 | * is stored at the end of the trampoline. This will be used to |
| 781 | * load the third parameter for the callback. Basically, that |
| 782 | * location at the end of the trampoline takes the place of |
| 783 | * the global function_trace_op variable. |
| 784 | */ |
| 785 | |
| 786 | ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE); |
| 787 | *ptr = (unsigned long)ops; |
| 788 | |
| 789 | op_offset -= start_offset; |
| 790 | memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE); |
| 791 | |
| 792 | /* Are we pointing to the reference? */ |
| 793 | if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) { |
| 794 | tramp_free(trampoline); |
| 795 | return 0; |
| 796 | } |
| 797 | |
| 798 | /* Load the contents of ptr into the callback parameter */ |
| 799 | offset = (unsigned long)ptr; |
| 800 | offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE; |
| 801 | |
| 802 | op_ptr.offset = offset; |
| 803 | |
| 804 | /* put in the new offset to the ftrace_ops */ |
| 805 | memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE); |
| 806 | |
| 807 | /* ALLOC_TRAMP flags lets us know we created it */ |
| 808 | ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP; |
| 809 | |
| 810 | return (unsigned long)trampoline; |
| 811 | } |
| 812 | |
| 813 | static unsigned long calc_trampoline_call_offset(bool save_regs) |
| 814 | { |
| 815 | unsigned long start_offset; |
| 816 | unsigned long call_offset; |
| 817 | |
| 818 | if (save_regs) { |
| 819 | start_offset = (unsigned long)ftrace_regs_caller; |
| 820 | call_offset = (unsigned long)ftrace_regs_call; |
| 821 | } else { |
| 822 | start_offset = (unsigned long)ftrace_caller; |
| 823 | call_offset = (unsigned long)ftrace_call; |
| 824 | } |
| 825 | |
| 826 | return call_offset - start_offset; |
| 827 | } |
| 828 | |
| 829 | void arch_ftrace_update_trampoline(struct ftrace_ops *ops) |
| 830 | { |
| 831 | ftrace_func_t func; |
| 832 | unsigned char *new; |
| 833 | unsigned long offset; |
| 834 | unsigned long ip; |
| 835 | unsigned int size; |
| 836 | int ret; |
| 837 | |
| 838 | if (ops->trampoline) { |
| 839 | /* |
| 840 | * The ftrace_ops caller may set up its own trampoline. |
| 841 | * In such a case, this code must not modify it. |
| 842 | */ |
| 843 | if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) |
| 844 | return; |
| 845 | } else { |
| 846 | ops->trampoline = create_trampoline(ops, &size); |
| 847 | if (!ops->trampoline) |
| 848 | return; |
| 849 | ops->trampoline_size = size; |
| 850 | } |
| 851 | |
| 852 | offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); |
| 853 | ip = ops->trampoline + offset; |
| 854 | |
| 855 | func = ftrace_ops_get_func(ops); |
| 856 | |
| 857 | /* Do a safe modify in case the trampoline is executing */ |
| 858 | new = ftrace_call_replace(ip, (unsigned long)func); |
| 859 | ret = update_ftrace_func(ip, new); |
| 860 | |
| 861 | /* The update should never fail */ |
| 862 | WARN_ON(ret); |
| 863 | } |
| 864 | |
| 865 | /* Return the address of the function the trampoline calls */ |
| 866 | static void *addr_from_call(void *ptr) |
| 867 | { |
| 868 | union ftrace_code_union calc; |
| 869 | int ret; |
| 870 | |
| 871 | ret = probe_kernel_read(&calc, ptr, MCOUNT_INSN_SIZE); |
| 872 | if (WARN_ON_ONCE(ret < 0)) |
| 873 | return NULL; |
| 874 | |
| 875 | /* Make sure this is a call */ |
| 876 | if (WARN_ON_ONCE(calc.e8 != 0xe8)) { |
| 877 | pr_warn("Expected e8, got %x\n", calc.e8); |
| 878 | return NULL; |
| 879 | } |
| 880 | |
| 881 | return ptr + MCOUNT_INSN_SIZE + calc.offset; |
| 882 | } |
| 883 | |
| 884 | void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent, |
| 885 | unsigned long frame_pointer); |
| 886 | |
| 887 | /* |
| 888 | * If the ops->trampoline was not allocated, then it probably |
| 889 | * has a static trampoline func, or is the ftrace caller itself. |
| 890 | */ |
| 891 | static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) |
| 892 | { |
| 893 | unsigned long offset; |
| 894 | bool save_regs = rec->flags & FTRACE_FL_REGS_EN; |
| 895 | void *ptr; |
| 896 | |
| 897 | if (ops && ops->trampoline) { |
| 898 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER |
| 899 | /* |
| 900 | * We only know about function graph tracer setting as static |
| 901 | * trampoline. |
| 902 | */ |
| 903 | if (ops->trampoline == FTRACE_GRAPH_ADDR) |
| 904 | return (void *)prepare_ftrace_return; |
| 905 | #endif |
| 906 | return NULL; |
| 907 | } |
| 908 | |
| 909 | offset = calc_trampoline_call_offset(save_regs); |
| 910 | |
| 911 | if (save_regs) |
| 912 | ptr = (void *)FTRACE_REGS_ADDR + offset; |
| 913 | else |
| 914 | ptr = (void *)FTRACE_ADDR + offset; |
| 915 | |
| 916 | return addr_from_call(ptr); |
| 917 | } |
| 918 | |
| 919 | void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) |
| 920 | { |
| 921 | unsigned long offset; |
| 922 | |
| 923 | /* If we didn't allocate this trampoline, consider it static */ |
| 924 | if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) |
| 925 | return static_tramp_func(ops, rec); |
| 926 | |
| 927 | offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); |
| 928 | return addr_from_call((void *)ops->trampoline + offset); |
| 929 | } |
| 930 | |
| 931 | void arch_ftrace_trampoline_free(struct ftrace_ops *ops) |
| 932 | { |
| 933 | if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) |
| 934 | return; |
| 935 | |
| 936 | tramp_free((void *)ops->trampoline); |
| 937 | ops->trampoline = 0; |
| 938 | } |
| 939 | |
| 940 | #endif /* CONFIG_X86_64 */ |
| 941 | #endif /* CONFIG_DYNAMIC_FTRACE */ |
| 942 | |
| 943 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER |
| 944 | |
| 945 | #ifdef CONFIG_DYNAMIC_FTRACE |
| 946 | extern void ftrace_graph_call(void); |
| 947 | |
| 948 | static int ftrace_mod_jmp(unsigned long ip, void *func) |
| 949 | { |
| 950 | unsigned char *new; |
| 951 | |
| 952 | new = ftrace_jmp_replace(ip, (unsigned long)func); |
| 953 | |
| 954 | return update_ftrace_func(ip, new); |
| 955 | } |
| 956 | |
| 957 | int ftrace_enable_ftrace_graph_caller(void) |
| 958 | { |
| 959 | unsigned long ip = (unsigned long)(&ftrace_graph_call); |
| 960 | |
| 961 | return ftrace_mod_jmp(ip, &ftrace_graph_caller); |
| 962 | } |
| 963 | |
| 964 | int ftrace_disable_ftrace_graph_caller(void) |
| 965 | { |
| 966 | unsigned long ip = (unsigned long)(&ftrace_graph_call); |
| 967 | |
| 968 | return ftrace_mod_jmp(ip, &ftrace_stub); |
| 969 | } |
| 970 | |
| 971 | #endif /* !CONFIG_DYNAMIC_FTRACE */ |
| 972 | |
| 973 | /* |
| 974 | * Hook the return address and push it in the stack of return addrs |
| 975 | * in current thread info. |
| 976 | */ |
| 977 | void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent, |
| 978 | unsigned long frame_pointer) |
| 979 | { |
| 980 | unsigned long old; |
| 981 | int faulted; |
| 982 | struct ftrace_graph_ent trace; |
| 983 | unsigned long return_hooker = (unsigned long) |
| 984 | &return_to_handler; |
| 985 | |
| 986 | if (unlikely(ftrace_graph_is_dead())) |
| 987 | return; |
| 988 | |
| 989 | if (unlikely(atomic_read(¤t->tracing_graph_pause))) |
| 990 | return; |
| 991 | |
| 992 | /* |
| 993 | * Protect against fault, even if it shouldn't |
| 994 | * happen. This tool is too much intrusive to |
| 995 | * ignore such a protection. |
| 996 | */ |
| 997 | asm volatile( |
| 998 | "1: " _ASM_MOV " (%[parent]), %[old]\n" |
| 999 | "2: " _ASM_MOV " %[return_hooker], (%[parent])\n" |
| 1000 | " movl $0, %[faulted]\n" |
| 1001 | "3:\n" |
| 1002 | |
| 1003 | ".section .fixup, \"ax\"\n" |
| 1004 | "4: movl $1, %[faulted]\n" |
| 1005 | " jmp 3b\n" |
| 1006 | ".previous\n" |
| 1007 | |
| 1008 | _ASM_EXTABLE(1b, 4b) |
| 1009 | _ASM_EXTABLE(2b, 4b) |
| 1010 | |
| 1011 | : [old] "=&r" (old), [faulted] "=r" (faulted) |
| 1012 | : [parent] "r" (parent), [return_hooker] "r" (return_hooker) |
| 1013 | : "memory" |
| 1014 | ); |
| 1015 | |
| 1016 | if (unlikely(faulted)) { |
| 1017 | ftrace_graph_stop(); |
| 1018 | WARN_ON(1); |
| 1019 | return; |
| 1020 | } |
| 1021 | |
| 1022 | trace.func = self_addr; |
| 1023 | trace.depth = current->curr_ret_stack + 1; |
| 1024 | |
| 1025 | /* Only trace if the calling function expects to */ |
| 1026 | if (!ftrace_graph_entry(&trace)) { |
| 1027 | *parent = old; |
| 1028 | return; |
| 1029 | } |
| 1030 | |
| 1031 | if (ftrace_push_return_trace(old, self_addr, &trace.depth, |
| 1032 | frame_pointer) == -EBUSY) { |
| 1033 | *parent = old; |
| 1034 | return; |
| 1035 | } |
| 1036 | } |
| 1037 | #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ |