1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Dynamic function tracing support. 4 * 5 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> 6 * 7 * Thanks goes to Ingo Molnar, for suggesting the idea. 8 * Mathieu Desnoyers, for suggesting postponing the modifications. 9 * Arjan van de Ven, for keeping me straight, and explaining to me 10 * the dangers of modifying code on the run. 11 */ 12 13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 14 15 #include <linux/spinlock.h> 16 #include <linux/hardirq.h> 17 #include <linux/uaccess.h> 18 #include <linux/ftrace.h> 19 #include <linux/percpu.h> 20 #include <linux/sched.h> 21 #include <linux/slab.h> 22 #include <linux/init.h> 23 #include <linux/list.h> 24 #include <linux/module.h> 25 #include <linux/memory.h> 26 #include <linux/vmalloc.h> 27 28 #include <trace/syscall.h> 29 30 #include <asm/set_memory.h> 31 #include <asm/kprobes.h> 32 #include <asm/ftrace.h> 33 #include <asm/nops.h> 34 #include <asm/text-patching.h> 35 36 #ifdef CONFIG_DYNAMIC_FTRACE 37 38 static int ftrace_poke_late = 0; 39 40 void ftrace_arch_code_modify_prepare(void) 41 __acquires(&text_mutex) 42 { 43 /* 44 * Need to grab text_mutex to prevent a race from module loading 45 * and live kernel patching from changing the text permissions while 46 * ftrace has it set to "read/write". 47 */ 48 mutex_lock(&text_mutex); 49 ftrace_poke_late = 1; 50 } 51 52 void ftrace_arch_code_modify_post_process(void) 53 __releases(&text_mutex) 54 { 55 /* 56 * ftrace_make_{call,nop}() may be called during 57 * module load, and we need to finish the text_poke_queue() 58 * that they do, here. 59 */ 60 text_poke_finish(); 61 ftrace_poke_late = 0; 62 mutex_unlock(&text_mutex); 63 } 64 65 static const char *ftrace_nop_replace(void) 66 { 67 return x86_nops[5]; 68 } 69 70 static const char *ftrace_call_replace(unsigned long ip, unsigned long addr) 71 { 72 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr); 73 } 74 75 static int ftrace_verify_code(unsigned long ip, const char *old_code) 76 { 77 char cur_code[MCOUNT_INSN_SIZE]; 78 79 /* 80 * Note: 81 * We are paranoid about modifying text, as if a bug was to happen, it 82 * could cause us to read or write to someplace that could cause harm. 83 * Carefully read and modify the code with probe_kernel_*(), and make 84 * sure what we read is what we expected it to be before modifying it. 85 */ 86 /* read the text we want to modify */ 87 if (copy_from_kernel_nofault(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) { 88 WARN_ON(1); 89 return -EFAULT; 90 } 91 92 /* Make sure it is what we expect it to be */ 93 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) { 94 WARN_ON(1); 95 return -EINVAL; 96 } 97 98 return 0; 99 } 100 101 /* 102 * Marked __ref because it calls text_poke_early() which is .init.text. That is 103 * ok because that call will happen early, during boot, when .init sections are 104 * still present. 105 */ 106 static int __ref 107 ftrace_modify_code_direct(unsigned long ip, const char *old_code, 108 const char *new_code) 109 { 110 int ret = ftrace_verify_code(ip, old_code); 111 if (ret) 112 return ret; 113 114 /* replace the text with the new text */ 115 if (ftrace_poke_late) 116 text_poke_queue((void *)ip, new_code, MCOUNT_INSN_SIZE, NULL); 117 else 118 text_poke_early((void *)ip, new_code, MCOUNT_INSN_SIZE); 119 return 0; 120 } 121 122 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr) 123 { 124 unsigned long ip = rec->ip; 125 const char *new, *old; 126 127 old = ftrace_call_replace(ip, addr); 128 new = ftrace_nop_replace(); 129 130 /* 131 * On boot up, and when modules are loaded, the MCOUNT_ADDR 132 * is converted to a nop, and will never become MCOUNT_ADDR 133 * again. This code is either running before SMP (on boot up) 134 * or before the code will ever be executed (module load). 135 * We do not want to use the breakpoint version in this case, 136 * just modify the code directly. 137 */ 138 if (addr == MCOUNT_ADDR) 139 return ftrace_modify_code_direct(ip, old, new); 140 141 /* 142 * x86 overrides ftrace_replace_code -- this function will never be used 143 * in this case. 144 */ 145 WARN_ONCE(1, "invalid use of ftrace_make_nop"); 146 return -EINVAL; 147 } 148 149 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) 150 { 151 unsigned long ip = rec->ip; 152 const char *new, *old; 153 154 old = ftrace_nop_replace(); 155 new = ftrace_call_replace(ip, addr); 156 157 /* Should only be called when module is loaded */ 158 return ftrace_modify_code_direct(rec->ip, old, new); 159 } 160 161 /* 162 * Should never be called: 163 * As it is only called by __ftrace_replace_code() which is called by 164 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code() 165 * which is called to turn mcount into nops or nops into function calls 166 * but not to convert a function from not using regs to one that uses 167 * regs, which ftrace_modify_call() is for. 168 */ 169 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr, 170 unsigned long addr) 171 { 172 WARN_ON(1); 173 return -EINVAL; 174 } 175 176 int ftrace_update_ftrace_func(ftrace_func_t func) 177 { 178 unsigned long ip; 179 const char *new; 180 181 ip = (unsigned long)(&ftrace_call); 182 new = ftrace_call_replace(ip, (unsigned long)func); 183 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 184 185 ip = (unsigned long)(&ftrace_regs_call); 186 new = ftrace_call_replace(ip, (unsigned long)func); 187 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 188 189 return 0; 190 } 191 192 void ftrace_replace_code(int enable) 193 { 194 struct ftrace_rec_iter *iter; 195 struct dyn_ftrace *rec; 196 const char *new, *old; 197 int ret; 198 199 for_ftrace_rec_iter(iter) { 200 rec = ftrace_rec_iter_record(iter); 201 202 switch (ftrace_test_record(rec, enable)) { 203 case FTRACE_UPDATE_IGNORE: 204 default: 205 continue; 206 207 case FTRACE_UPDATE_MAKE_CALL: 208 old = ftrace_nop_replace(); 209 break; 210 211 case FTRACE_UPDATE_MODIFY_CALL: 212 case FTRACE_UPDATE_MAKE_NOP: 213 old = ftrace_call_replace(rec->ip, ftrace_get_addr_curr(rec)); 214 break; 215 } 216 217 ret = ftrace_verify_code(rec->ip, old); 218 if (ret) { 219 ftrace_bug(ret, rec); 220 return; 221 } 222 } 223 224 for_ftrace_rec_iter(iter) { 225 rec = ftrace_rec_iter_record(iter); 226 227 switch (ftrace_test_record(rec, enable)) { 228 case FTRACE_UPDATE_IGNORE: 229 default: 230 continue; 231 232 case FTRACE_UPDATE_MAKE_CALL: 233 case FTRACE_UPDATE_MODIFY_CALL: 234 new = ftrace_call_replace(rec->ip, ftrace_get_addr_new(rec)); 235 break; 236 237 case FTRACE_UPDATE_MAKE_NOP: 238 new = ftrace_nop_replace(); 239 break; 240 } 241 242 text_poke_queue((void *)rec->ip, new, MCOUNT_INSN_SIZE, NULL); 243 ftrace_update_record(rec, enable); 244 } 245 text_poke_finish(); 246 } 247 248 void arch_ftrace_update_code(int command) 249 { 250 ftrace_modify_all_code(command); 251 } 252 253 /* Currently only x86_64 supports dynamic trampolines */ 254 #ifdef CONFIG_X86_64 255 256 #ifdef CONFIG_MODULES 257 #include <linux/moduleloader.h> 258 /* Module allocation simplifies allocating memory for code */ 259 static inline void *alloc_tramp(unsigned long size) 260 { 261 return module_alloc(size); 262 } 263 static inline void tramp_free(void *tramp) 264 { 265 module_memfree(tramp); 266 } 267 #else 268 /* Trampolines can only be created if modules are supported */ 269 static inline void *alloc_tramp(unsigned long size) 270 { 271 return NULL; 272 } 273 static inline void tramp_free(void *tramp) { } 274 #endif 275 276 /* Defined as markers to the end of the ftrace default trampolines */ 277 extern void ftrace_regs_caller_end(void); 278 extern void ftrace_regs_caller_ret(void); 279 extern void ftrace_caller_end(void); 280 extern void ftrace_caller_op_ptr(void); 281 extern void ftrace_regs_caller_op_ptr(void); 282 extern void ftrace_regs_caller_jmp(void); 283 284 /* movq function_trace_op(%rip), %rdx */ 285 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */ 286 #define OP_REF_SIZE 7 287 288 /* 289 * The ftrace_ops is passed to the function callback. Since the 290 * trampoline only services a single ftrace_ops, we can pass in 291 * that ops directly. 292 * 293 * The ftrace_op_code_union is used to create a pointer to the 294 * ftrace_ops that will be passed to the callback function. 295 */ 296 union ftrace_op_code_union { 297 char code[OP_REF_SIZE]; 298 struct { 299 char op[3]; 300 int offset; 301 } __attribute__((packed)); 302 }; 303 304 #define RET_SIZE 1 + IS_ENABLED(CONFIG_SLS) 305 306 static unsigned long 307 create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size) 308 { 309 unsigned long start_offset; 310 unsigned long end_offset; 311 unsigned long op_offset; 312 unsigned long call_offset; 313 unsigned long jmp_offset; 314 unsigned long offset; 315 unsigned long npages; 316 unsigned long size; 317 unsigned long *ptr; 318 void *trampoline; 319 void *ip; 320 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */ 321 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 }; 322 unsigned const char retq[] = { RET_INSN_OPCODE, INT3_INSN_OPCODE }; 323 union ftrace_op_code_union op_ptr; 324 int ret; 325 326 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 327 start_offset = (unsigned long)ftrace_regs_caller; 328 end_offset = (unsigned long)ftrace_regs_caller_end; 329 op_offset = (unsigned long)ftrace_regs_caller_op_ptr; 330 call_offset = (unsigned long)ftrace_regs_call; 331 jmp_offset = (unsigned long)ftrace_regs_caller_jmp; 332 } else { 333 start_offset = (unsigned long)ftrace_caller; 334 end_offset = (unsigned long)ftrace_caller_end; 335 op_offset = (unsigned long)ftrace_caller_op_ptr; 336 call_offset = (unsigned long)ftrace_call; 337 jmp_offset = 0; 338 } 339 340 size = end_offset - start_offset; 341 342 /* 343 * Allocate enough size to store the ftrace_caller code, 344 * the iret , as well as the address of the ftrace_ops this 345 * trampoline is used for. 346 */ 347 trampoline = alloc_tramp(size + RET_SIZE + sizeof(void *)); 348 if (!trampoline) 349 return 0; 350 351 *tramp_size = size + RET_SIZE + sizeof(void *); 352 npages = DIV_ROUND_UP(*tramp_size, PAGE_SIZE); 353 354 /* Copy ftrace_caller onto the trampoline memory */ 355 ret = copy_from_kernel_nofault(trampoline, (void *)start_offset, size); 356 if (WARN_ON(ret < 0)) 357 goto fail; 358 359 ip = trampoline + size; 360 memcpy(ip, retq, RET_SIZE); 361 362 /* No need to test direct calls on created trampolines */ 363 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 364 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */ 365 ip = trampoline + (jmp_offset - start_offset); 366 if (WARN_ON(*(char *)ip != 0x75)) 367 goto fail; 368 ret = copy_from_kernel_nofault(ip, x86_nops[2], 2); 369 if (ret < 0) 370 goto fail; 371 } 372 373 /* 374 * The address of the ftrace_ops that is used for this trampoline 375 * is stored at the end of the trampoline. This will be used to 376 * load the third parameter for the callback. Basically, that 377 * location at the end of the trampoline takes the place of 378 * the global function_trace_op variable. 379 */ 380 381 ptr = (unsigned long *)(trampoline + size + RET_SIZE); 382 *ptr = (unsigned long)ops; 383 384 op_offset -= start_offset; 385 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE); 386 387 /* Are we pointing to the reference? */ 388 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) 389 goto fail; 390 391 /* Load the contents of ptr into the callback parameter */ 392 offset = (unsigned long)ptr; 393 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE; 394 395 op_ptr.offset = offset; 396 397 /* put in the new offset to the ftrace_ops */ 398 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE); 399 400 /* put in the call to the function */ 401 mutex_lock(&text_mutex); 402 call_offset -= start_offset; 403 memcpy(trampoline + call_offset, 404 text_gen_insn(CALL_INSN_OPCODE, 405 trampoline + call_offset, 406 ftrace_ops_get_func(ops)), CALL_INSN_SIZE); 407 mutex_unlock(&text_mutex); 408 409 /* ALLOC_TRAMP flags lets us know we created it */ 410 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP; 411 412 set_vm_flush_reset_perms(trampoline); 413 414 if (likely(system_state != SYSTEM_BOOTING)) 415 set_memory_ro((unsigned long)trampoline, npages); 416 set_memory_x((unsigned long)trampoline, npages); 417 return (unsigned long)trampoline; 418 fail: 419 tramp_free(trampoline); 420 return 0; 421 } 422 423 void set_ftrace_ops_ro(void) 424 { 425 struct ftrace_ops *ops; 426 unsigned long start_offset; 427 unsigned long end_offset; 428 unsigned long npages; 429 unsigned long size; 430 431 do_for_each_ftrace_op(ops, ftrace_ops_list) { 432 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 433 continue; 434 435 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 436 start_offset = (unsigned long)ftrace_regs_caller; 437 end_offset = (unsigned long)ftrace_regs_caller_end; 438 } else { 439 start_offset = (unsigned long)ftrace_caller; 440 end_offset = (unsigned long)ftrace_caller_end; 441 } 442 size = end_offset - start_offset; 443 size = size + RET_SIZE + sizeof(void *); 444 npages = DIV_ROUND_UP(size, PAGE_SIZE); 445 set_memory_ro((unsigned long)ops->trampoline, npages); 446 } while_for_each_ftrace_op(ops); 447 } 448 449 static unsigned long calc_trampoline_call_offset(bool save_regs) 450 { 451 unsigned long start_offset; 452 unsigned long call_offset; 453 454 if (save_regs) { 455 start_offset = (unsigned long)ftrace_regs_caller; 456 call_offset = (unsigned long)ftrace_regs_call; 457 } else { 458 start_offset = (unsigned long)ftrace_caller; 459 call_offset = (unsigned long)ftrace_call; 460 } 461 462 return call_offset - start_offset; 463 } 464 465 void arch_ftrace_update_trampoline(struct ftrace_ops *ops) 466 { 467 ftrace_func_t func; 468 unsigned long offset; 469 unsigned long ip; 470 unsigned int size; 471 const char *new; 472 473 if (!ops->trampoline) { 474 ops->trampoline = create_trampoline(ops, &size); 475 if (!ops->trampoline) 476 return; 477 ops->trampoline_size = size; 478 return; 479 } 480 481 /* 482 * The ftrace_ops caller may set up its own trampoline. 483 * In such a case, this code must not modify it. 484 */ 485 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 486 return; 487 488 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); 489 ip = ops->trampoline + offset; 490 func = ftrace_ops_get_func(ops); 491 492 mutex_lock(&text_mutex); 493 /* Do a safe modify in case the trampoline is executing */ 494 new = ftrace_call_replace(ip, (unsigned long)func); 495 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 496 mutex_unlock(&text_mutex); 497 } 498 499 /* Return the address of the function the trampoline calls */ 500 static void *addr_from_call(void *ptr) 501 { 502 union text_poke_insn call; 503 int ret; 504 505 ret = copy_from_kernel_nofault(&call, ptr, CALL_INSN_SIZE); 506 if (WARN_ON_ONCE(ret < 0)) 507 return NULL; 508 509 /* Make sure this is a call */ 510 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) { 511 pr_warn("Expected E8, got %x\n", call.opcode); 512 return NULL; 513 } 514 515 return ptr + CALL_INSN_SIZE + call.disp; 516 } 517 518 void prepare_ftrace_return(unsigned long ip, unsigned long *parent, 519 unsigned long frame_pointer); 520 521 /* 522 * If the ops->trampoline was not allocated, then it probably 523 * has a static trampoline func, or is the ftrace caller itself. 524 */ 525 static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) 526 { 527 unsigned long offset; 528 bool save_regs = rec->flags & FTRACE_FL_REGS_EN; 529 void *ptr; 530 531 if (ops && ops->trampoline) { 532 #if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \ 533 defined(CONFIG_FUNCTION_GRAPH_TRACER) 534 /* 535 * We only know about function graph tracer setting as static 536 * trampoline. 537 */ 538 if (ops->trampoline == FTRACE_GRAPH_ADDR) 539 return (void *)prepare_ftrace_return; 540 #endif 541 return NULL; 542 } 543 544 offset = calc_trampoline_call_offset(save_regs); 545 546 if (save_regs) 547 ptr = (void *)FTRACE_REGS_ADDR + offset; 548 else 549 ptr = (void *)FTRACE_ADDR + offset; 550 551 return addr_from_call(ptr); 552 } 553 554 void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) 555 { 556 unsigned long offset; 557 558 /* If we didn't allocate this trampoline, consider it static */ 559 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 560 return static_tramp_func(ops, rec); 561 562 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); 563 return addr_from_call((void *)ops->trampoline + offset); 564 } 565 566 void arch_ftrace_trampoline_free(struct ftrace_ops *ops) 567 { 568 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 569 return; 570 571 tramp_free((void *)ops->trampoline); 572 ops->trampoline = 0; 573 } 574 575 #endif /* CONFIG_X86_64 */ 576 #endif /* CONFIG_DYNAMIC_FTRACE */ 577 578 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 579 580 #if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) 581 extern void ftrace_graph_call(void); 582 static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr) 583 { 584 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr); 585 } 586 587 static int ftrace_mod_jmp(unsigned long ip, void *func) 588 { 589 const char *new; 590 591 new = ftrace_jmp_replace(ip, (unsigned long)func); 592 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 593 return 0; 594 } 595 596 int ftrace_enable_ftrace_graph_caller(void) 597 { 598 unsigned long ip = (unsigned long)(&ftrace_graph_call); 599 600 return ftrace_mod_jmp(ip, &ftrace_graph_caller); 601 } 602 603 int ftrace_disable_ftrace_graph_caller(void) 604 { 605 unsigned long ip = (unsigned long)(&ftrace_graph_call); 606 607 return ftrace_mod_jmp(ip, &ftrace_stub); 608 } 609 #endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */ 610 611 /* 612 * Hook the return address and push it in the stack of return addrs 613 * in current thread info. 614 */ 615 void prepare_ftrace_return(unsigned long ip, unsigned long *parent, 616 unsigned long frame_pointer) 617 { 618 unsigned long return_hooker = (unsigned long)&return_to_handler; 619 int bit; 620 621 /* 622 * When resuming from suspend-to-ram, this function can be indirectly 623 * called from early CPU startup code while the CPU is in real mode, 624 * which would fail miserably. Make sure the stack pointer is a 625 * virtual address. 626 * 627 * This check isn't as accurate as virt_addr_valid(), but it should be 628 * good enough for this purpose, and it's fast. 629 */ 630 if (unlikely((long)__builtin_frame_address(0) >= 0)) 631 return; 632 633 if (unlikely(ftrace_graph_is_dead())) 634 return; 635 636 if (unlikely(atomic_read(¤t->tracing_graph_pause))) 637 return; 638 639 bit = ftrace_test_recursion_trylock(ip, *parent); 640 if (bit < 0) 641 return; 642 643 if (!function_graph_enter(*parent, ip, frame_pointer, parent)) 644 *parent = return_hooker; 645 646 ftrace_test_recursion_unlock(bit); 647 } 648 649 #ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS 650 void ftrace_graph_func(unsigned long ip, unsigned long parent_ip, 651 struct ftrace_ops *op, struct ftrace_regs *fregs) 652 { 653 struct pt_regs *regs = &fregs->regs; 654 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs); 655 656 prepare_ftrace_return(ip, (unsigned long *)stack, 0); 657 } 658 #endif 659 660 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ 661