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 int 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 return 0; 51 } 52 53 int ftrace_arch_code_modify_post_process(void) 54 __releases(&text_mutex) 55 { 56 /* 57 * ftrace_make_{call,nop}() may be called during 58 * module load, and we need to finish the text_poke_queue() 59 * that they do, here. 60 */ 61 text_poke_finish(); 62 ftrace_poke_late = 0; 63 mutex_unlock(&text_mutex); 64 return 0; 65 } 66 67 static const char *ftrace_nop_replace(void) 68 { 69 return x86_nops[5]; 70 } 71 72 static const char *ftrace_call_replace(unsigned long ip, unsigned long addr) 73 { 74 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr); 75 } 76 77 static int ftrace_verify_code(unsigned long ip, const char *old_code) 78 { 79 char cur_code[MCOUNT_INSN_SIZE]; 80 81 /* 82 * Note: 83 * We are paranoid about modifying text, as if a bug was to happen, it 84 * could cause us to read or write to someplace that could cause harm. 85 * Carefully read and modify the code with probe_kernel_*(), and make 86 * sure what we read is what we expected it to be before modifying it. 87 */ 88 /* read the text we want to modify */ 89 if (copy_from_kernel_nofault(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) { 90 WARN_ON(1); 91 return -EFAULT; 92 } 93 94 /* Make sure it is what we expect it to be */ 95 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) { 96 WARN_ON(1); 97 return -EINVAL; 98 } 99 100 return 0; 101 } 102 103 /* 104 * Marked __ref because it calls text_poke_early() which is .init.text. That is 105 * ok because that call will happen early, during boot, when .init sections are 106 * still present. 107 */ 108 static int __ref 109 ftrace_modify_code_direct(unsigned long ip, const char *old_code, 110 const char *new_code) 111 { 112 int ret = ftrace_verify_code(ip, old_code); 113 if (ret) 114 return ret; 115 116 /* replace the text with the new text */ 117 if (ftrace_poke_late) 118 text_poke_queue((void *)ip, new_code, MCOUNT_INSN_SIZE, NULL); 119 else 120 text_poke_early((void *)ip, new_code, MCOUNT_INSN_SIZE); 121 return 0; 122 } 123 124 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr) 125 { 126 unsigned long ip = rec->ip; 127 const char *new, *old; 128 129 old = ftrace_call_replace(ip, addr); 130 new = ftrace_nop_replace(); 131 132 /* 133 * On boot up, and when modules are loaded, the MCOUNT_ADDR 134 * is converted to a nop, and will never become MCOUNT_ADDR 135 * again. This code is either running before SMP (on boot up) 136 * or before the code will ever be executed (module load). 137 * We do not want to use the breakpoint version in this case, 138 * just modify the code directly. 139 */ 140 if (addr == MCOUNT_ADDR) 141 return ftrace_modify_code_direct(ip, old, new); 142 143 /* 144 * x86 overrides ftrace_replace_code -- this function will never be used 145 * in this case. 146 */ 147 WARN_ONCE(1, "invalid use of ftrace_make_nop"); 148 return -EINVAL; 149 } 150 151 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) 152 { 153 unsigned long ip = rec->ip; 154 const char *new, *old; 155 156 old = ftrace_nop_replace(); 157 new = ftrace_call_replace(ip, addr); 158 159 /* Should only be called when module is loaded */ 160 return ftrace_modify_code_direct(rec->ip, old, new); 161 } 162 163 /* 164 * Should never be called: 165 * As it is only called by __ftrace_replace_code() which is called by 166 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code() 167 * which is called to turn mcount into nops or nops into function calls 168 * but not to convert a function from not using regs to one that uses 169 * regs, which ftrace_modify_call() is for. 170 */ 171 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr, 172 unsigned long addr) 173 { 174 WARN_ON(1); 175 return -EINVAL; 176 } 177 178 int ftrace_update_ftrace_func(ftrace_func_t func) 179 { 180 unsigned long ip; 181 const char *new; 182 183 ip = (unsigned long)(&ftrace_call); 184 new = ftrace_call_replace(ip, (unsigned long)func); 185 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 186 187 ip = (unsigned long)(&ftrace_regs_call); 188 new = ftrace_call_replace(ip, (unsigned long)func); 189 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 190 191 return 0; 192 } 193 194 void ftrace_replace_code(int enable) 195 { 196 struct ftrace_rec_iter *iter; 197 struct dyn_ftrace *rec; 198 const char *new, *old; 199 int ret; 200 201 for_ftrace_rec_iter(iter) { 202 rec = ftrace_rec_iter_record(iter); 203 204 switch (ftrace_test_record(rec, enable)) { 205 case FTRACE_UPDATE_IGNORE: 206 default: 207 continue; 208 209 case FTRACE_UPDATE_MAKE_CALL: 210 old = ftrace_nop_replace(); 211 break; 212 213 case FTRACE_UPDATE_MODIFY_CALL: 214 case FTRACE_UPDATE_MAKE_NOP: 215 old = ftrace_call_replace(rec->ip, ftrace_get_addr_curr(rec)); 216 break; 217 } 218 219 ret = ftrace_verify_code(rec->ip, old); 220 if (ret) { 221 ftrace_bug(ret, rec); 222 return; 223 } 224 } 225 226 for_ftrace_rec_iter(iter) { 227 rec = ftrace_rec_iter_record(iter); 228 229 switch (ftrace_test_record(rec, enable)) { 230 case FTRACE_UPDATE_IGNORE: 231 default: 232 continue; 233 234 case FTRACE_UPDATE_MAKE_CALL: 235 case FTRACE_UPDATE_MODIFY_CALL: 236 new = ftrace_call_replace(rec->ip, ftrace_get_addr_new(rec)); 237 break; 238 239 case FTRACE_UPDATE_MAKE_NOP: 240 new = ftrace_nop_replace(); 241 break; 242 } 243 244 text_poke_queue((void *)rec->ip, new, MCOUNT_INSN_SIZE, NULL); 245 ftrace_update_record(rec, enable); 246 } 247 text_poke_finish(); 248 } 249 250 void arch_ftrace_update_code(int command) 251 { 252 ftrace_modify_all_code(command); 253 } 254 255 /* Currently only x86_64 supports dynamic trampolines */ 256 #ifdef CONFIG_X86_64 257 258 #ifdef CONFIG_MODULES 259 #include <linux/moduleloader.h> 260 /* Module allocation simplifies allocating memory for code */ 261 static inline void *alloc_tramp(unsigned long size) 262 { 263 return module_alloc(size); 264 } 265 static inline void tramp_free(void *tramp) 266 { 267 module_memfree(tramp); 268 } 269 #else 270 /* Trampolines can only be created if modules are supported */ 271 static inline void *alloc_tramp(unsigned long size) 272 { 273 return NULL; 274 } 275 static inline void tramp_free(void *tramp) { } 276 #endif 277 278 /* Defined as markers to the end of the ftrace default trampolines */ 279 extern void ftrace_regs_caller_end(void); 280 extern void ftrace_regs_caller_ret(void); 281 extern void ftrace_caller_end(void); 282 extern void ftrace_caller_op_ptr(void); 283 extern void ftrace_regs_caller_op_ptr(void); 284 extern void ftrace_regs_caller_jmp(void); 285 286 /* movq function_trace_op(%rip), %rdx */ 287 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */ 288 #define OP_REF_SIZE 7 289 290 /* 291 * The ftrace_ops is passed to the function callback. Since the 292 * trampoline only services a single ftrace_ops, we can pass in 293 * that ops directly. 294 * 295 * The ftrace_op_code_union is used to create a pointer to the 296 * ftrace_ops that will be passed to the callback function. 297 */ 298 union ftrace_op_code_union { 299 char code[OP_REF_SIZE]; 300 struct { 301 char op[3]; 302 int offset; 303 } __attribute__((packed)); 304 }; 305 306 #define RET_SIZE 1 + IS_ENABLED(CONFIG_SLS) 307 308 static unsigned long 309 create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size) 310 { 311 unsigned long start_offset; 312 unsigned long end_offset; 313 unsigned long op_offset; 314 unsigned long call_offset; 315 unsigned long jmp_offset; 316 unsigned long offset; 317 unsigned long npages; 318 unsigned long size; 319 unsigned long *ptr; 320 void *trampoline; 321 void *ip; 322 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */ 323 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 }; 324 unsigned const char retq[] = { RET_INSN_OPCODE, INT3_INSN_OPCODE }; 325 union ftrace_op_code_union op_ptr; 326 int ret; 327 328 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 329 start_offset = (unsigned long)ftrace_regs_caller; 330 end_offset = (unsigned long)ftrace_regs_caller_end; 331 op_offset = (unsigned long)ftrace_regs_caller_op_ptr; 332 call_offset = (unsigned long)ftrace_regs_call; 333 jmp_offset = (unsigned long)ftrace_regs_caller_jmp; 334 } else { 335 start_offset = (unsigned long)ftrace_caller; 336 end_offset = (unsigned long)ftrace_caller_end; 337 op_offset = (unsigned long)ftrace_caller_op_ptr; 338 call_offset = (unsigned long)ftrace_call; 339 jmp_offset = 0; 340 } 341 342 size = end_offset - start_offset; 343 344 /* 345 * Allocate enough size to store the ftrace_caller code, 346 * the iret , as well as the address of the ftrace_ops this 347 * trampoline is used for. 348 */ 349 trampoline = alloc_tramp(size + RET_SIZE + sizeof(void *)); 350 if (!trampoline) 351 return 0; 352 353 *tramp_size = size + RET_SIZE + sizeof(void *); 354 npages = DIV_ROUND_UP(*tramp_size, PAGE_SIZE); 355 356 /* Copy ftrace_caller onto the trampoline memory */ 357 ret = copy_from_kernel_nofault(trampoline, (void *)start_offset, size); 358 if (WARN_ON(ret < 0)) 359 goto fail; 360 361 ip = trampoline + size; 362 memcpy(ip, retq, RET_SIZE); 363 364 /* No need to test direct calls on created trampolines */ 365 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 366 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */ 367 ip = trampoline + (jmp_offset - start_offset); 368 if (WARN_ON(*(char *)ip != 0x75)) 369 goto fail; 370 ret = copy_from_kernel_nofault(ip, x86_nops[2], 2); 371 if (ret < 0) 372 goto fail; 373 } 374 375 /* 376 * The address of the ftrace_ops that is used for this trampoline 377 * is stored at the end of the trampoline. This will be used to 378 * load the third parameter for the callback. Basically, that 379 * location at the end of the trampoline takes the place of 380 * the global function_trace_op variable. 381 */ 382 383 ptr = (unsigned long *)(trampoline + size + RET_SIZE); 384 *ptr = (unsigned long)ops; 385 386 op_offset -= start_offset; 387 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE); 388 389 /* Are we pointing to the reference? */ 390 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) 391 goto fail; 392 393 /* Load the contents of ptr into the callback parameter */ 394 offset = (unsigned long)ptr; 395 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE; 396 397 op_ptr.offset = offset; 398 399 /* put in the new offset to the ftrace_ops */ 400 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE); 401 402 /* put in the call to the function */ 403 mutex_lock(&text_mutex); 404 call_offset -= start_offset; 405 memcpy(trampoline + call_offset, 406 text_gen_insn(CALL_INSN_OPCODE, 407 trampoline + call_offset, 408 ftrace_ops_get_func(ops)), CALL_INSN_SIZE); 409 mutex_unlock(&text_mutex); 410 411 /* ALLOC_TRAMP flags lets us know we created it */ 412 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP; 413 414 set_vm_flush_reset_perms(trampoline); 415 416 if (likely(system_state != SYSTEM_BOOTING)) 417 set_memory_ro((unsigned long)trampoline, npages); 418 set_memory_x((unsigned long)trampoline, npages); 419 return (unsigned long)trampoline; 420 fail: 421 tramp_free(trampoline); 422 return 0; 423 } 424 425 void set_ftrace_ops_ro(void) 426 { 427 struct ftrace_ops *ops; 428 unsigned long start_offset; 429 unsigned long end_offset; 430 unsigned long npages; 431 unsigned long size; 432 433 do_for_each_ftrace_op(ops, ftrace_ops_list) { 434 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 435 continue; 436 437 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { 438 start_offset = (unsigned long)ftrace_regs_caller; 439 end_offset = (unsigned long)ftrace_regs_caller_end; 440 } else { 441 start_offset = (unsigned long)ftrace_caller; 442 end_offset = (unsigned long)ftrace_caller_end; 443 } 444 size = end_offset - start_offset; 445 size = size + RET_SIZE + sizeof(void *); 446 npages = DIV_ROUND_UP(size, PAGE_SIZE); 447 set_memory_ro((unsigned long)ops->trampoline, npages); 448 } while_for_each_ftrace_op(ops); 449 } 450 451 static unsigned long calc_trampoline_call_offset(bool save_regs) 452 { 453 unsigned long start_offset; 454 unsigned long call_offset; 455 456 if (save_regs) { 457 start_offset = (unsigned long)ftrace_regs_caller; 458 call_offset = (unsigned long)ftrace_regs_call; 459 } else { 460 start_offset = (unsigned long)ftrace_caller; 461 call_offset = (unsigned long)ftrace_call; 462 } 463 464 return call_offset - start_offset; 465 } 466 467 void arch_ftrace_update_trampoline(struct ftrace_ops *ops) 468 { 469 ftrace_func_t func; 470 unsigned long offset; 471 unsigned long ip; 472 unsigned int size; 473 const char *new; 474 475 if (!ops->trampoline) { 476 ops->trampoline = create_trampoline(ops, &size); 477 if (!ops->trampoline) 478 return; 479 ops->trampoline_size = size; 480 return; 481 } 482 483 /* 484 * The ftrace_ops caller may set up its own trampoline. 485 * In such a case, this code must not modify it. 486 */ 487 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 488 return; 489 490 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); 491 ip = ops->trampoline + offset; 492 func = ftrace_ops_get_func(ops); 493 494 mutex_lock(&text_mutex); 495 /* Do a safe modify in case the trampoline is executing */ 496 new = ftrace_call_replace(ip, (unsigned long)func); 497 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 498 mutex_unlock(&text_mutex); 499 } 500 501 /* Return the address of the function the trampoline calls */ 502 static void *addr_from_call(void *ptr) 503 { 504 union text_poke_insn call; 505 int ret; 506 507 ret = copy_from_kernel_nofault(&call, ptr, CALL_INSN_SIZE); 508 if (WARN_ON_ONCE(ret < 0)) 509 return NULL; 510 511 /* Make sure this is a call */ 512 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) { 513 pr_warn("Expected E8, got %x\n", call.opcode); 514 return NULL; 515 } 516 517 return ptr + CALL_INSN_SIZE + call.disp; 518 } 519 520 void prepare_ftrace_return(unsigned long ip, unsigned long *parent, 521 unsigned long frame_pointer); 522 523 /* 524 * If the ops->trampoline was not allocated, then it probably 525 * has a static trampoline func, or is the ftrace caller itself. 526 */ 527 static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) 528 { 529 unsigned long offset; 530 bool save_regs = rec->flags & FTRACE_FL_REGS_EN; 531 void *ptr; 532 533 if (ops && ops->trampoline) { 534 #if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \ 535 defined(CONFIG_FUNCTION_GRAPH_TRACER) 536 /* 537 * We only know about function graph tracer setting as static 538 * trampoline. 539 */ 540 if (ops->trampoline == FTRACE_GRAPH_ADDR) 541 return (void *)prepare_ftrace_return; 542 #endif 543 return NULL; 544 } 545 546 offset = calc_trampoline_call_offset(save_regs); 547 548 if (save_regs) 549 ptr = (void *)FTRACE_REGS_ADDR + offset; 550 else 551 ptr = (void *)FTRACE_ADDR + offset; 552 553 return addr_from_call(ptr); 554 } 555 556 void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) 557 { 558 unsigned long offset; 559 560 /* If we didn't allocate this trampoline, consider it static */ 561 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 562 return static_tramp_func(ops, rec); 563 564 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); 565 return addr_from_call((void *)ops->trampoline + offset); 566 } 567 568 void arch_ftrace_trampoline_free(struct ftrace_ops *ops) 569 { 570 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) 571 return; 572 573 tramp_free((void *)ops->trampoline); 574 ops->trampoline = 0; 575 } 576 577 #endif /* CONFIG_X86_64 */ 578 #endif /* CONFIG_DYNAMIC_FTRACE */ 579 580 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 581 582 #if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) 583 extern void ftrace_graph_call(void); 584 static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr) 585 { 586 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr); 587 } 588 589 static int ftrace_mod_jmp(unsigned long ip, void *func) 590 { 591 const char *new; 592 593 new = ftrace_jmp_replace(ip, (unsigned long)func); 594 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL); 595 return 0; 596 } 597 598 int ftrace_enable_ftrace_graph_caller(void) 599 { 600 unsigned long ip = (unsigned long)(&ftrace_graph_call); 601 602 return ftrace_mod_jmp(ip, &ftrace_graph_caller); 603 } 604 605 int ftrace_disable_ftrace_graph_caller(void) 606 { 607 unsigned long ip = (unsigned long)(&ftrace_graph_call); 608 609 return ftrace_mod_jmp(ip, &ftrace_stub); 610 } 611 #endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */ 612 613 /* 614 * Hook the return address and push it in the stack of return addrs 615 * in current thread info. 616 */ 617 void prepare_ftrace_return(unsigned long ip, unsigned long *parent, 618 unsigned long frame_pointer) 619 { 620 unsigned long return_hooker = (unsigned long)&return_to_handler; 621 int bit; 622 623 /* 624 * When resuming from suspend-to-ram, this function can be indirectly 625 * called from early CPU startup code while the CPU is in real mode, 626 * which would fail miserably. Make sure the stack pointer is a 627 * virtual address. 628 * 629 * This check isn't as accurate as virt_addr_valid(), but it should be 630 * good enough for this purpose, and it's fast. 631 */ 632 if (unlikely((long)__builtin_frame_address(0) >= 0)) 633 return; 634 635 if (unlikely(ftrace_graph_is_dead())) 636 return; 637 638 if (unlikely(atomic_read(¤t->tracing_graph_pause))) 639 return; 640 641 bit = ftrace_test_recursion_trylock(ip, *parent); 642 if (bit < 0) 643 return; 644 645 if (!function_graph_enter(*parent, ip, frame_pointer, parent)) 646 *parent = return_hooker; 647 648 ftrace_test_recursion_unlock(bit); 649 } 650 651 #ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS 652 void ftrace_graph_func(unsigned long ip, unsigned long parent_ip, 653 struct ftrace_ops *op, struct ftrace_regs *fregs) 654 { 655 struct pt_regs *regs = &fregs->regs; 656 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs); 657 658 prepare_ftrace_return(ip, (unsigned long *)stack, 0); 659 } 660 #endif 661 662 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ 663