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#include <linux/set_memory.h>
28
29#include <trace/syscall.h>
30
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
38static int ftrace_poke_late = 0;
39
40void 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
52void 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(lock: &text_mutex);
63}
64
65static const char *ftrace_nop_replace(void)
66{
67 return x86_nops[5];
68}
69
70static const char *ftrace_call_replace(unsigned long ip, unsigned long addr)
71{
72 /*
73 * No need to translate into a callthunk. The trampoline does
74 * the depth accounting itself.
75 */
76 return text_gen_insn(CALL_INSN_OPCODE, addr: (void *)ip, dest: (void *)addr);
77}
78
79static int ftrace_verify_code(unsigned long ip, const char *old_code)
80{
81 char cur_code[MCOUNT_INSN_SIZE];
82
83 /*
84 * Note:
85 * We are paranoid about modifying text, as if a bug was to happen, it
86 * could cause us to read or write to someplace that could cause harm.
87 * Carefully read and modify the code with probe_kernel_*(), and make
88 * sure what we read is what we expected it to be before modifying it.
89 */
90 /* read the text we want to modify */
91 if (copy_from_kernel_nofault(dst: cur_code, src: (void *)ip, MCOUNT_INSN_SIZE)) {
92 WARN_ON(1);
93 return -EFAULT;
94 }
95
96 /* Make sure it is what we expect it to be */
97 if (memcmp(p: cur_code, q: old_code, MCOUNT_INSN_SIZE) != 0) {
98 ftrace_expected = old_code;
99 WARN_ON(1);
100 return -EINVAL;
101 }
102
103 return 0;
104}
105
106/*
107 * Marked __ref because it calls text_poke_early() which is .init.text. That is
108 * ok because that call will happen early, during boot, when .init sections are
109 * still present.
110 */
111static int __ref
112ftrace_modify_code_direct(unsigned long ip, const char *old_code,
113 const char *new_code)
114{
115 int ret = ftrace_verify_code(ip, old_code);
116 if (ret)
117 return ret;
118
119 /* replace the text with the new text */
120 if (ftrace_poke_late)
121 text_poke_queue(addr: (void *)ip, opcode: new_code, MCOUNT_INSN_SIZE, NULL);
122 else
123 text_poke_early(addr: (void *)ip, opcode: new_code, MCOUNT_INSN_SIZE);
124 return 0;
125}
126
127int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr)
128{
129 unsigned long ip = rec->ip;
130 const char *new, *old;
131
132 old = ftrace_call_replace(ip, addr);
133 new = ftrace_nop_replace();
134
135 /*
136 * On boot up, and when modules are loaded, the MCOUNT_ADDR
137 * is converted to a nop, and will never become MCOUNT_ADDR
138 * again. This code is either running before SMP (on boot up)
139 * or before the code will ever be executed (module load).
140 * We do not want to use the breakpoint version in this case,
141 * just modify the code directly.
142 */
143 if (addr == MCOUNT_ADDR)
144 return ftrace_modify_code_direct(ip, old_code: old, new_code: new);
145
146 /*
147 * x86 overrides ftrace_replace_code -- this function will never be used
148 * in this case.
149 */
150 WARN_ONCE(1, "invalid use of ftrace_make_nop");
151 return -EINVAL;
152}
153
154int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
155{
156 unsigned long ip = rec->ip;
157 const char *new, *old;
158
159 old = ftrace_nop_replace();
160 new = ftrace_call_replace(ip, addr);
161
162 /* Should only be called when module is loaded */
163 return ftrace_modify_code_direct(ip: rec->ip, old_code: old, new_code: new);
164}
165
166/*
167 * Should never be called:
168 * As it is only called by __ftrace_replace_code() which is called by
169 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
170 * which is called to turn mcount into nops or nops into function calls
171 * but not to convert a function from not using regs to one that uses
172 * regs, which ftrace_modify_call() is for.
173 */
174int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
175 unsigned long addr)
176{
177 WARN_ON(1);
178 return -EINVAL;
179}
180
181int ftrace_update_ftrace_func(ftrace_func_t func)
182{
183 unsigned long ip;
184 const char *new;
185
186 ip = (unsigned long)(&ftrace_call);
187 new = ftrace_call_replace(ip, addr: (unsigned long)func);
188 text_poke_bp(addr: (void *)ip, opcode: new, MCOUNT_INSN_SIZE, NULL);
189
190 ip = (unsigned long)(&ftrace_regs_call);
191 new = ftrace_call_replace(ip, addr: (unsigned long)func);
192 text_poke_bp(addr: (void *)ip, opcode: new, MCOUNT_INSN_SIZE, NULL);
193
194 return 0;
195}
196
197void ftrace_replace_code(int enable)
198{
199 struct ftrace_rec_iter *iter;
200 struct dyn_ftrace *rec;
201 const char *new, *old;
202 int ret;
203
204 for_ftrace_rec_iter(iter) {
205 rec = ftrace_rec_iter_record(iter);
206
207 switch (ftrace_test_record(rec, enable)) {
208 case FTRACE_UPDATE_IGNORE:
209 default:
210 continue;
211
212 case FTRACE_UPDATE_MAKE_CALL:
213 old = ftrace_nop_replace();
214 break;
215
216 case FTRACE_UPDATE_MODIFY_CALL:
217 case FTRACE_UPDATE_MAKE_NOP:
218 old = ftrace_call_replace(ip: rec->ip, addr: ftrace_get_addr_curr(rec));
219 break;
220 }
221
222 ret = ftrace_verify_code(ip: rec->ip, old_code: old);
223 if (ret) {
224 ftrace_expected = old;
225 ftrace_bug(err: ret, rec);
226 ftrace_expected = NULL;
227 return;
228 }
229 }
230
231 for_ftrace_rec_iter(iter) {
232 rec = ftrace_rec_iter_record(iter);
233
234 switch (ftrace_test_record(rec, enable)) {
235 case FTRACE_UPDATE_IGNORE:
236 default:
237 continue;
238
239 case FTRACE_UPDATE_MAKE_CALL:
240 case FTRACE_UPDATE_MODIFY_CALL:
241 new = ftrace_call_replace(ip: rec->ip, addr: ftrace_get_addr_new(rec));
242 break;
243
244 case FTRACE_UPDATE_MAKE_NOP:
245 new = ftrace_nop_replace();
246 break;
247 }
248
249 text_poke_queue(addr: (void *)rec->ip, opcode: new, MCOUNT_INSN_SIZE, NULL);
250 ftrace_update_record(rec, enable);
251 }
252 text_poke_finish();
253}
254
255void arch_ftrace_update_code(int command)
256{
257 ftrace_modify_all_code(command);
258}
259
260/* Currently only x86_64 supports dynamic trampolines */
261#ifdef CONFIG_X86_64
262
263#ifdef CONFIG_MODULES
264#include <linux/moduleloader.h>
265/* Module allocation simplifies allocating memory for code */
266static inline void *alloc_tramp(unsigned long size)
267{
268 return module_alloc(size);
269}
270static inline void tramp_free(void *tramp)
271{
272 module_memfree(module_region: tramp);
273}
274#else
275/* Trampolines can only be created if modules are supported */
276static inline void *alloc_tramp(unsigned long size)
277{
278 return NULL;
279}
280static inline void tramp_free(void *tramp) { }
281#endif
282
283/* Defined as markers to the end of the ftrace default trampolines */
284extern void ftrace_regs_caller_end(void);
285extern void ftrace_caller_end(void);
286extern void ftrace_caller_op_ptr(void);
287extern void ftrace_regs_caller_op_ptr(void);
288extern void ftrace_regs_caller_jmp(void);
289
290/* movq function_trace_op(%rip), %rdx */
291/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
292#define OP_REF_SIZE 7
293
294/*
295 * The ftrace_ops is passed to the function callback. Since the
296 * trampoline only services a single ftrace_ops, we can pass in
297 * that ops directly.
298 *
299 * The ftrace_op_code_union is used to create a pointer to the
300 * ftrace_ops that will be passed to the callback function.
301 */
302union ftrace_op_code_union {
303 char code[OP_REF_SIZE];
304 struct {
305 char op[3];
306 int offset;
307 } __attribute__((packed));
308};
309
310#define RET_SIZE \
311 (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) ? 5 : 1 + IS_ENABLED(CONFIG_MITIGATION_SLS))
312
313static unsigned long
314create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
315{
316 unsigned long start_offset;
317 unsigned long end_offset;
318 unsigned long op_offset;
319 unsigned long call_offset;
320 unsigned long jmp_offset;
321 unsigned long offset;
322 unsigned long npages;
323 unsigned long size;
324 unsigned long *ptr;
325 void *trampoline;
326 void *ip, *dest;
327 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
328 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
329 unsigned const char retq[] = { RET_INSN_OPCODE, INT3_INSN_OPCODE };
330 union ftrace_op_code_union op_ptr;
331 int ret;
332
333 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
334 start_offset = (unsigned long)ftrace_regs_caller;
335 end_offset = (unsigned long)ftrace_regs_caller_end;
336 op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
337 call_offset = (unsigned long)ftrace_regs_call;
338 jmp_offset = (unsigned long)ftrace_regs_caller_jmp;
339 } else {
340 start_offset = (unsigned long)ftrace_caller;
341 end_offset = (unsigned long)ftrace_caller_end;
342 op_offset = (unsigned long)ftrace_caller_op_ptr;
343 call_offset = (unsigned long)ftrace_call;
344 jmp_offset = 0;
345 }
346
347 size = end_offset - start_offset;
348
349 /*
350 * Allocate enough size to store the ftrace_caller code,
351 * the iret , as well as the address of the ftrace_ops this
352 * trampoline is used for.
353 */
354 trampoline = alloc_tramp(size: size + RET_SIZE + sizeof(void *));
355 if (!trampoline)
356 return 0;
357
358 *tramp_size = size + RET_SIZE + sizeof(void *);
359 npages = DIV_ROUND_UP(*tramp_size, PAGE_SIZE);
360
361 /* Copy ftrace_caller onto the trampoline memory */
362 ret = copy_from_kernel_nofault(dst: trampoline, src: (void *)start_offset, size);
363 if (WARN_ON(ret < 0))
364 goto fail;
365
366 ip = trampoline + size;
367 if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
368 __text_gen_insn(buf: ip, JMP32_INSN_OPCODE, addr: ip, dest: x86_return_thunk, JMP32_INSN_SIZE);
369 else
370 memcpy(ip, retq, sizeof(retq));
371
372 /* No need to test direct calls on created trampolines */
373 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
374 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
375 ip = trampoline + (jmp_offset - start_offset);
376 if (WARN_ON(*(char *)ip != 0x75))
377 goto fail;
378 ret = copy_from_kernel_nofault(dst: ip, src: x86_nops[2], size: 2);
379 if (ret < 0)
380 goto fail;
381 }
382
383 /*
384 * The address of the ftrace_ops that is used for this trampoline
385 * is stored at the end of the trampoline. This will be used to
386 * load the third parameter for the callback. Basically, that
387 * location at the end of the trampoline takes the place of
388 * the global function_trace_op variable.
389 */
390
391 ptr = (unsigned long *)(trampoline + size + RET_SIZE);
392 *ptr = (unsigned long)ops;
393
394 op_offset -= start_offset;
395 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
396
397 /* Are we pointing to the reference? */
398 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0))
399 goto fail;
400
401 /* Load the contents of ptr into the callback parameter */
402 offset = (unsigned long)ptr;
403 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
404
405 op_ptr.offset = offset;
406
407 /* put in the new offset to the ftrace_ops */
408 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
409
410 /* put in the call to the function */
411 mutex_lock(&text_mutex);
412 call_offset -= start_offset;
413 /*
414 * No need to translate into a callthunk. The trampoline does
415 * the depth accounting before the call already.
416 */
417 dest = ftrace_ops_get_func(ops);
418 memcpy(trampoline + call_offset,
419 text_gen_insn(CALL_INSN_OPCODE, trampoline + call_offset, dest),
420 CALL_INSN_SIZE);
421 mutex_unlock(lock: &text_mutex);
422
423 /* ALLOC_TRAMP flags lets us know we created it */
424 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
425
426 set_memory_rox(addr: (unsigned long)trampoline, numpages: npages);
427 return (unsigned long)trampoline;
428fail:
429 tramp_free(tramp: trampoline);
430 return 0;
431}
432
433void set_ftrace_ops_ro(void)
434{
435 struct ftrace_ops *ops;
436 unsigned long start_offset;
437 unsigned long end_offset;
438 unsigned long npages;
439 unsigned long size;
440
441 do_for_each_ftrace_op(ops, ftrace_ops_list) {
442 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
443 continue;
444
445 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
446 start_offset = (unsigned long)ftrace_regs_caller;
447 end_offset = (unsigned long)ftrace_regs_caller_end;
448 } else {
449 start_offset = (unsigned long)ftrace_caller;
450 end_offset = (unsigned long)ftrace_caller_end;
451 }
452 size = end_offset - start_offset;
453 size = size + RET_SIZE + sizeof(void *);
454 npages = DIV_ROUND_UP(size, PAGE_SIZE);
455 set_memory_ro(addr: (unsigned long)ops->trampoline, numpages: npages);
456 } while_for_each_ftrace_op(ops);
457}
458
459static unsigned long calc_trampoline_call_offset(bool save_regs)
460{
461 unsigned long start_offset;
462 unsigned long call_offset;
463
464 if (save_regs) {
465 start_offset = (unsigned long)ftrace_regs_caller;
466 call_offset = (unsigned long)ftrace_regs_call;
467 } else {
468 start_offset = (unsigned long)ftrace_caller;
469 call_offset = (unsigned long)ftrace_call;
470 }
471
472 return call_offset - start_offset;
473}
474
475void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
476{
477 ftrace_func_t func;
478 unsigned long offset;
479 unsigned long ip;
480 unsigned int size;
481 const char *new;
482
483 if (!ops->trampoline) {
484 ops->trampoline = create_trampoline(ops, tramp_size: &size);
485 if (!ops->trampoline)
486 return;
487 ops->trampoline_size = size;
488 return;
489 }
490
491 /*
492 * The ftrace_ops caller may set up its own trampoline.
493 * In such a case, this code must not modify it.
494 */
495 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
496 return;
497
498 offset = calc_trampoline_call_offset(save_regs: ops->flags & FTRACE_OPS_FL_SAVE_REGS);
499 ip = ops->trampoline + offset;
500 func = ftrace_ops_get_func(ops);
501
502 mutex_lock(&text_mutex);
503 /* Do a safe modify in case the trampoline is executing */
504 new = ftrace_call_replace(ip, addr: (unsigned long)func);
505 text_poke_bp(addr: (void *)ip, opcode: new, MCOUNT_INSN_SIZE, NULL);
506 mutex_unlock(lock: &text_mutex);
507}
508
509/* Return the address of the function the trampoline calls */
510static void *addr_from_call(void *ptr)
511{
512 union text_poke_insn call;
513 int ret;
514
515 ret = copy_from_kernel_nofault(dst: &call, src: ptr, CALL_INSN_SIZE);
516 if (WARN_ON_ONCE(ret < 0))
517 return NULL;
518
519 /* Make sure this is a call */
520 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) {
521 pr_warn("Expected E8, got %x\n", call.opcode);
522 return NULL;
523 }
524
525 return ptr + CALL_INSN_SIZE + call.disp;
526}
527
528/*
529 * If the ops->trampoline was not allocated, then it probably
530 * has a static trampoline func, or is the ftrace caller itself.
531 */
532static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
533{
534 unsigned long offset;
535 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
536 void *ptr;
537
538 if (ops && ops->trampoline) {
539#if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \
540 defined(CONFIG_FUNCTION_GRAPH_TRACER)
541 /*
542 * We only know about function graph tracer setting as static
543 * trampoline.
544 */
545 if (ops->trampoline == FTRACE_GRAPH_ADDR)
546 return (void *)prepare_ftrace_return;
547#endif
548 return NULL;
549 }
550
551 offset = calc_trampoline_call_offset(save_regs);
552
553 if (save_regs)
554 ptr = (void *)FTRACE_REGS_ADDR + offset;
555 else
556 ptr = (void *)FTRACE_ADDR + offset;
557
558 return addr_from_call(ptr);
559}
560
561void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
562{
563 unsigned long offset;
564
565 /* If we didn't allocate this trampoline, consider it static */
566 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
567 return static_tramp_func(ops, rec);
568
569 offset = calc_trampoline_call_offset(save_regs: ops->flags & FTRACE_OPS_FL_SAVE_REGS);
570 return addr_from_call(ptr: (void *)ops->trampoline + offset);
571}
572
573void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
574{
575 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
576 return;
577
578 tramp_free(tramp: (void *)ops->trampoline);
579 ops->trampoline = 0;
580}
581
582#endif /* CONFIG_X86_64 */
583#endif /* CONFIG_DYNAMIC_FTRACE */
584
585#ifdef CONFIG_FUNCTION_GRAPH_TRACER
586
587#if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS)
588extern void ftrace_graph_call(void);
589static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
590{
591 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
592}
593
594static int ftrace_mod_jmp(unsigned long ip, void *func)
595{
596 const char *new;
597
598 new = ftrace_jmp_replace(ip, (unsigned long)func);
599 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
600 return 0;
601}
602
603int ftrace_enable_ftrace_graph_caller(void)
604{
605 unsigned long ip = (unsigned long)(&ftrace_graph_call);
606
607 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
608}
609
610int ftrace_disable_ftrace_graph_caller(void)
611{
612 unsigned long ip = (unsigned long)(&ftrace_graph_call);
613
614 return ftrace_mod_jmp(ip, &ftrace_stub);
615}
616#endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */
617
618/*
619 * Hook the return address and push it in the stack of return addrs
620 * in current thread info.
621 */
622void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
623 unsigned long frame_pointer)
624{
625 unsigned long return_hooker = (unsigned long)&return_to_handler;
626 int bit;
627
628 /*
629 * When resuming from suspend-to-ram, this function can be indirectly
630 * called from early CPU startup code while the CPU is in real mode,
631 * which would fail miserably. Make sure the stack pointer is a
632 * virtual address.
633 *
634 * This check isn't as accurate as virt_addr_valid(), but it should be
635 * good enough for this purpose, and it's fast.
636 */
637 if (unlikely((long)__builtin_frame_address(0) >= 0))
638 return;
639
640 if (unlikely(ftrace_graph_is_dead()))
641 return;
642
643 if (unlikely(atomic_read(&current->tracing_graph_pause)))
644 return;
645
646 bit = ftrace_test_recursion_trylock(ip, parent_ip: *parent);
647 if (bit < 0)
648 return;
649
650 if (!function_graph_enter(ret: *parent, func: ip, frame_pointer, retp: parent))
651 *parent = return_hooker;
652
653 ftrace_test_recursion_unlock(bit);
654}
655
656#ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
657void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
658 struct ftrace_ops *op, struct ftrace_regs *fregs)
659{
660 struct pt_regs *regs = &fregs->regs;
661 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs);
662
663 prepare_ftrace_return(ip, parent: (unsigned long *)stack, frame_pointer: 0);
664}
665#endif
666
667#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
668

source code of linux/arch/x86/kernel/ftrace.c