Loading...
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#include <linux/execmem.h>
29
30#include <trace/syscall.h>
31
32#include <asm/kprobes.h>
33#include <asm/ftrace.h>
34#include <asm/nops.h>
35#include <asm/text-patching.h>
36
37#ifdef CONFIG_DYNAMIC_FTRACE
38
39static int ftrace_poke_late = 0;
40
41void ftrace_arch_code_modify_prepare(void)
42 __acquires(&text_mutex)
43{
44 /*
45 * Need to grab text_mutex to prevent a race from module loading
46 * and live kernel patching from changing the text permissions while
47 * ftrace has it set to "read/write".
48 */
49 mutex_lock(&text_mutex);
50 ftrace_poke_late = 1;
51}
52
53void 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}
65
66static const char *ftrace_nop_replace(void)
67{
68 return x86_nops[5];
69}
70
71static const char *ftrace_call_replace(unsigned long ip, unsigned long addr)
72{
73 /*
74 * No need to translate into a callthunk. The trampoline does
75 * the depth accounting itself.
76 */
77 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
78}
79
80static int ftrace_verify_code(unsigned long ip, const char *old_code)
81{
82 char cur_code[MCOUNT_INSN_SIZE];
83
84 /*
85 * Note:
86 * We are paranoid about modifying text, as if a bug was to happen, it
87 * could cause us to read or write to someplace that could cause harm.
88 * Carefully read and modify the code with probe_kernel_*(), and make
89 * sure what we read is what we expected it to be before modifying it.
90 */
91 /* read the text we want to modify */
92 if (copy_from_kernel_nofault(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) {
93 WARN_ON(1);
94 return -EFAULT;
95 }
96
97 /* Make sure it is what we expect it to be */
98 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) {
99 ftrace_expected = old_code;
100 WARN_ON(1);
101 return -EINVAL;
102 }
103
104 return 0;
105}
106
107/*
108 * Marked __ref because it calls text_poke_early() which is .init.text. That is
109 * ok because that call will happen early, during boot, when .init sections are
110 * still present.
111 */
112static int __ref
113ftrace_modify_code_direct(unsigned long ip, const char *old_code,
114 const char *new_code)
115{
116 int ret = ftrace_verify_code(ip, old_code);
117 if (ret)
118 return ret;
119
120 /* replace the text with the new text */
121 if (ftrace_poke_late) {
122 text_poke_queue((void *)ip, new_code, MCOUNT_INSN_SIZE, NULL);
123 } else {
124 mutex_lock(&text_mutex);
125 text_poke((void *)ip, new_code, MCOUNT_INSN_SIZE);
126 mutex_unlock(&text_mutex);
127 }
128 return 0;
129}
130
131int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr)
132{
133 unsigned long ip = rec->ip;
134 const char *new, *old;
135
136 old = ftrace_call_replace(ip, addr);
137 new = ftrace_nop_replace();
138
139 /*
140 * On boot up, and when modules are loaded, the MCOUNT_ADDR
141 * is converted to a nop, and will never become MCOUNT_ADDR
142 * again. This code is either running before SMP (on boot up)
143 * or before the code will ever be executed (module load).
144 * We do not want to use the breakpoint version in this case,
145 * just modify the code directly.
146 */
147 if (addr == MCOUNT_ADDR)
148 return ftrace_modify_code_direct(ip, old, new);
149
150 /*
151 * x86 overrides ftrace_replace_code -- this function will never be used
152 * in this case.
153 */
154 WARN_ONCE(1, "invalid use of ftrace_make_nop");
155 return -EINVAL;
156}
157
158int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
159{
160 unsigned long ip = rec->ip;
161 const char *new, *old;
162
163 old = ftrace_nop_replace();
164 new = ftrace_call_replace(ip, addr);
165
166 /* Should only be called when module is loaded */
167 return ftrace_modify_code_direct(rec->ip, old, new);
168}
169
170/*
171 * Should never be called:
172 * As it is only called by __ftrace_replace_code() which is called by
173 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
174 * which is called to turn mcount into nops or nops into function calls
175 * but not to convert a function from not using regs to one that uses
176 * regs, which ftrace_modify_call() is for.
177 */
178int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
179 unsigned long addr)
180{
181 WARN_ON(1);
182 return -EINVAL;
183}
184
185int ftrace_update_ftrace_func(ftrace_func_t func)
186{
187 unsigned long ip;
188 const char *new;
189
190 ip = (unsigned long)(&ftrace_call);
191 new = ftrace_call_replace(ip, (unsigned long)func);
192 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
193
194 ip = (unsigned long)(&ftrace_regs_call);
195 new = ftrace_call_replace(ip, (unsigned long)func);
196 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
197
198 return 0;
199}
200
201void ftrace_replace_code(int enable)
202{
203 struct ftrace_rec_iter *iter;
204 struct dyn_ftrace *rec;
205 const char *new, *old;
206 int ret;
207
208 for_ftrace_rec_iter(iter) {
209 rec = ftrace_rec_iter_record(iter);
210
211 switch (ftrace_test_record(rec, enable)) {
212 case FTRACE_UPDATE_IGNORE:
213 default:
214 continue;
215
216 case FTRACE_UPDATE_MAKE_CALL:
217 old = ftrace_nop_replace();
218 break;
219
220 case FTRACE_UPDATE_MODIFY_CALL:
221 case FTRACE_UPDATE_MAKE_NOP:
222 old = ftrace_call_replace(rec->ip, ftrace_get_addr_curr(rec));
223 break;
224 }
225
226 ret = ftrace_verify_code(rec->ip, old);
227 if (ret) {
228 ftrace_expected = old;
229 ftrace_bug(ret, rec);
230 ftrace_expected = NULL;
231 return;
232 }
233 }
234
235 for_ftrace_rec_iter(iter) {
236 rec = ftrace_rec_iter_record(iter);
237
238 switch (ftrace_test_record(rec, enable)) {
239 case FTRACE_UPDATE_IGNORE:
240 default:
241 continue;
242
243 case FTRACE_UPDATE_MAKE_CALL:
244 case FTRACE_UPDATE_MODIFY_CALL:
245 new = ftrace_call_replace(rec->ip, ftrace_get_addr_new(rec));
246 break;
247
248 case FTRACE_UPDATE_MAKE_NOP:
249 new = ftrace_nop_replace();
250 break;
251 }
252
253 text_poke_queue((void *)rec->ip, new, MCOUNT_INSN_SIZE, NULL);
254 ftrace_update_record(rec, enable);
255 }
256 text_poke_finish();
257}
258
259void arch_ftrace_update_code(int command)
260{
261 ftrace_modify_all_code(command);
262}
263
264/* Currently only x86_64 supports dynamic trampolines */
265#ifdef CONFIG_X86_64
266
267static inline void *alloc_tramp(unsigned long size)
268{
269 return execmem_alloc(EXECMEM_FTRACE, size);
270}
271static inline void tramp_free(void *tramp)
272{
273 execmem_free(tramp);
274}
275
276/* Defined as markers to the end of the ftrace default trampolines */
277extern void ftrace_regs_caller_end(void);
278extern void ftrace_caller_end(void);
279extern void ftrace_caller_op_ptr(void);
280extern void ftrace_regs_caller_op_ptr(void);
281extern void ftrace_regs_caller_jmp(void);
282
283/* movq function_trace_op(%rip), %rdx */
284/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
285#define OP_REF_SIZE 7
286
287/*
288 * The ftrace_ops is passed to the function callback. Since the
289 * trampoline only services a single ftrace_ops, we can pass in
290 * that ops directly.
291 *
292 * The ftrace_op_code_union is used to create a pointer to the
293 * ftrace_ops that will be passed to the callback function.
294 */
295union ftrace_op_code_union {
296 char code[OP_REF_SIZE];
297 struct {
298 char op[3];
299 int offset;
300 } __attribute__((packed));
301};
302
303#define RET_SIZE \
304 (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) ? 5 : 1 + IS_ENABLED(CONFIG_MITIGATION_SLS))
305
306static unsigned long
307create_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, *dest;
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 void *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 = text_poke_copy(trampoline, (void *)start_offset, size);
356 if (WARN_ON(!ret))
357 goto fail;
358
359 ip = trampoline + size;
360 if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
361 __text_gen_insn(ip, JMP32_INSN_OPCODE, ip, x86_return_thunk, JMP32_INSN_SIZE);
362 else
363 text_poke_copy(ip, retq, sizeof(retq));
364
365 /* No need to test direct calls on created trampolines */
366 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
367 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
368 ip = trampoline + (jmp_offset - start_offset);
369 if (WARN_ON(*(char *)ip != 0x75))
370 goto fail;
371 if (!text_poke_copy(ip, x86_nops[2], 2))
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 text_poke_copy(ptr, &ops, sizeof(unsigned long));
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 text_poke_copy(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 /*
406 * No need to translate into a callthunk. The trampoline does
407 * the depth accounting before the call already.
408 */
409 dest = ftrace_ops_get_func(ops);
410 text_poke_copy_locked(trampoline + call_offset,
411 text_gen_insn(CALL_INSN_OPCODE, trampoline + call_offset, dest),
412 CALL_INSN_SIZE, false);
413 mutex_unlock(&text_mutex);
414
415 /* ALLOC_TRAMP flags lets us know we created it */
416 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
417
418 set_memory_rox((unsigned long)trampoline, npages);
419 return (unsigned long)trampoline;
420fail:
421 tramp_free(trampoline);
422 return 0;
423}
424
425void 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
451static 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
467void 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 */
502static 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/*
521 * If the ops->trampoline was not allocated, then it probably
522 * has a static trampoline func, or is the ftrace caller itself.
523 */
524static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
525{
526 unsigned long offset;
527 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
528 void *ptr;
529
530 if (ops && ops->trampoline) {
531#if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \
532 defined(CONFIG_FUNCTION_GRAPH_TRACER)
533 /*
534 * We only know about function graph tracer setting as static
535 * trampoline.
536 */
537 if (ops->trampoline == FTRACE_GRAPH_ADDR)
538 return (void *)prepare_ftrace_return;
539#endif
540 return NULL;
541 }
542
543 offset = calc_trampoline_call_offset(save_regs);
544
545 if (save_regs)
546 ptr = (void *)FTRACE_REGS_ADDR + offset;
547 else
548 ptr = (void *)FTRACE_ADDR + offset;
549
550 return addr_from_call(ptr);
551}
552
553void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
554{
555 unsigned long offset;
556
557 /* If we didn't allocate this trampoline, consider it static */
558 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
559 return static_tramp_func(ops, rec);
560
561 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
562 return addr_from_call((void *)ops->trampoline + offset);
563}
564
565void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
566{
567 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
568 return;
569
570 tramp_free((void *)ops->trampoline);
571 ops->trampoline = 0;
572}
573
574#endif /* CONFIG_X86_64 */
575#endif /* CONFIG_DYNAMIC_FTRACE */
576
577#ifdef CONFIG_FUNCTION_GRAPH_TRACER
578
579#if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS)
580extern void ftrace_graph_call(void);
581static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
582{
583 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
584}
585
586static int ftrace_mod_jmp(unsigned long ip, void *func)
587{
588 const char *new;
589
590 new = ftrace_jmp_replace(ip, (unsigned long)func);
591 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
592 return 0;
593}
594
595int ftrace_enable_ftrace_graph_caller(void)
596{
597 unsigned long ip = (unsigned long)(&ftrace_graph_call);
598
599 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
600}
601
602int ftrace_disable_ftrace_graph_caller(void)
603{
604 unsigned long ip = (unsigned long)(&ftrace_graph_call);
605
606 return ftrace_mod_jmp(ip, &ftrace_stub);
607}
608#endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */
609
610/*
611 * Hook the return address and push it in the stack of return addrs
612 * in current thread info.
613 */
614void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
615 unsigned long frame_pointer)
616{
617 unsigned long return_hooker = (unsigned long)&return_to_handler;
618 int bit;
619
620 /*
621 * When resuming from suspend-to-ram, this function can be indirectly
622 * called from early CPU startup code while the CPU is in real mode,
623 * which would fail miserably. Make sure the stack pointer is a
624 * virtual address.
625 *
626 * This check isn't as accurate as virt_addr_valid(), but it should be
627 * good enough for this purpose, and it's fast.
628 */
629 if (unlikely((long)__builtin_frame_address(0) >= 0))
630 return;
631
632 if (unlikely(ftrace_graph_is_dead()))
633 return;
634
635 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
636 return;
637
638 bit = ftrace_test_recursion_trylock(ip, *parent);
639 if (bit < 0)
640 return;
641
642 if (!function_graph_enter(*parent, ip, frame_pointer, parent))
643 *parent = return_hooker;
644
645 ftrace_test_recursion_unlock(bit);
646}
647
648#ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
649void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
650 struct ftrace_ops *op, struct ftrace_regs *fregs)
651{
652 struct pt_regs *regs = &arch_ftrace_regs(fregs)->regs;
653 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs);
654
655 prepare_ftrace_return(ip, (unsigned long *)stack, 0);
656}
657#endif
658
659#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
1/*
2 * Code for replacing ftrace calls with jumps.
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/init.h>
21#include <linux/list.h>
22#include <linux/module.h>
23
24#include <trace/syscall.h>
25
26#include <asm/cacheflush.h>
27#include <asm/kprobes.h>
28#include <asm/ftrace.h>
29#include <asm/nops.h>
30
31#ifdef CONFIG_DYNAMIC_FTRACE
32
33int ftrace_arch_code_modify_prepare(void)
34{
35 set_kernel_text_rw();
36 set_all_modules_text_rw();
37 return 0;
38}
39
40int ftrace_arch_code_modify_post_process(void)
41{
42 set_all_modules_text_ro();
43 set_kernel_text_ro();
44 return 0;
45}
46
47union ftrace_code_union {
48 char code[MCOUNT_INSN_SIZE];
49 struct {
50 char e8;
51 int offset;
52 } __attribute__((packed));
53};
54
55static int ftrace_calc_offset(long ip, long addr)
56{
57 return (int)(addr - ip);
58}
59
60static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
61{
62 static union ftrace_code_union calc;
63
64 calc.e8 = 0xe8;
65 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
66
67 /*
68 * No locking needed, this must be called via kstop_machine
69 * which in essence is like running on a uniprocessor machine.
70 */
71 return calc.code;
72}
73
74static inline int
75within(unsigned long addr, unsigned long start, unsigned long end)
76{
77 return addr >= start && addr < end;
78}
79
80static unsigned long text_ip_addr(unsigned long ip)
81{
82 /*
83 * On x86_64, kernel text mappings are mapped read-only with
84 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
85 * of the kernel text mapping to modify the kernel text.
86 *
87 * For 32bit kernels, these mappings are same and we can use
88 * kernel identity mapping to modify code.
89 */
90 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
91 ip = (unsigned long)__va(__pa_symbol(ip));
92
93 return ip;
94}
95
96static const unsigned char *ftrace_nop_replace(void)
97{
98 return ideal_nops[NOP_ATOMIC5];
99}
100
101static int
102ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code,
103 unsigned const char *new_code)
104{
105 unsigned char replaced[MCOUNT_INSN_SIZE];
106
107 /*
108 * Note: Due to modules and __init, code can
109 * disappear and change, we need to protect against faulting
110 * as well as code changing. We do this by using the
111 * probe_kernel_* functions.
112 *
113 * No real locking needed, this code is run through
114 * kstop_machine, or before SMP starts.
115 */
116
117 /* read the text we want to modify */
118 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
119 return -EFAULT;
120
121 /* Make sure it is what we expect it to be */
122 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
123 return -EINVAL;
124
125 ip = text_ip_addr(ip);
126
127 /* replace the text with the new text */
128 if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE))
129 return -EPERM;
130
131 sync_core();
132
133 return 0;
134}
135
136int ftrace_make_nop(struct module *mod,
137 struct dyn_ftrace *rec, unsigned long addr)
138{
139 unsigned const char *new, *old;
140 unsigned long ip = rec->ip;
141
142 old = ftrace_call_replace(ip, addr);
143 new = ftrace_nop_replace();
144
145 /*
146 * On boot up, and when modules are loaded, the MCOUNT_ADDR
147 * is converted to a nop, and will never become MCOUNT_ADDR
148 * again. This code is either running before SMP (on boot up)
149 * or before the code will ever be executed (module load).
150 * We do not want to use the breakpoint version in this case,
151 * just modify the code directly.
152 */
153 if (addr == MCOUNT_ADDR)
154 return ftrace_modify_code_direct(rec->ip, old, new);
155
156 /* Normal cases use add_brk_on_nop */
157 WARN_ONCE(1, "invalid use of ftrace_make_nop");
158 return -EINVAL;
159}
160
161int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
162{
163 unsigned const char *new, *old;
164 unsigned long ip = rec->ip;
165
166 old = ftrace_nop_replace();
167 new = ftrace_call_replace(ip, addr);
168
169 /* Should only be called when module is loaded */
170 return ftrace_modify_code_direct(rec->ip, old, new);
171}
172
173/*
174 * The modifying_ftrace_code is used to tell the breakpoint
175 * handler to call ftrace_int3_handler(). If it fails to
176 * call this handler for a breakpoint added by ftrace, then
177 * the kernel may crash.
178 *
179 * As atomic_writes on x86 do not need a barrier, we do not
180 * need to add smp_mb()s for this to work. It is also considered
181 * that we can not read the modifying_ftrace_code before
182 * executing the breakpoint. That would be quite remarkable if
183 * it could do that. Here's the flow that is required:
184 *
185 * CPU-0 CPU-1
186 *
187 * atomic_inc(mfc);
188 * write int3s
189 * <trap-int3> // implicit (r)mb
190 * if (atomic_read(mfc))
191 * call ftrace_int3_handler()
192 *
193 * Then when we are finished:
194 *
195 * atomic_dec(mfc);
196 *
197 * If we hit a breakpoint that was not set by ftrace, it does not
198 * matter if ftrace_int3_handler() is called or not. It will
199 * simply be ignored. But it is crucial that a ftrace nop/caller
200 * breakpoint is handled. No other user should ever place a
201 * breakpoint on an ftrace nop/caller location. It must only
202 * be done by this code.
203 */
204atomic_t modifying_ftrace_code __read_mostly;
205
206static int
207ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
208 unsigned const char *new_code);
209
210/*
211 * Should never be called:
212 * As it is only called by __ftrace_replace_code() which is called by
213 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
214 * which is called to turn mcount into nops or nops into function calls
215 * but not to convert a function from not using regs to one that uses
216 * regs, which ftrace_modify_call() is for.
217 */
218int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
219 unsigned long addr)
220{
221 WARN_ON(1);
222 return -EINVAL;
223}
224
225static unsigned long ftrace_update_func;
226
227static int update_ftrace_func(unsigned long ip, void *new)
228{
229 unsigned char old[MCOUNT_INSN_SIZE];
230 int ret;
231
232 memcpy(old, (void *)ip, MCOUNT_INSN_SIZE);
233
234 ftrace_update_func = ip;
235 /* Make sure the breakpoints see the ftrace_update_func update */
236 smp_wmb();
237
238 /* See comment above by declaration of modifying_ftrace_code */
239 atomic_inc(&modifying_ftrace_code);
240
241 ret = ftrace_modify_code(ip, old, new);
242
243 atomic_dec(&modifying_ftrace_code);
244
245 return ret;
246}
247
248int ftrace_update_ftrace_func(ftrace_func_t func)
249{
250 unsigned long ip = (unsigned long)(&ftrace_call);
251 unsigned char *new;
252 int ret;
253
254 new = ftrace_call_replace(ip, (unsigned long)func);
255 ret = update_ftrace_func(ip, new);
256
257 /* Also update the regs callback function */
258 if (!ret) {
259 ip = (unsigned long)(&ftrace_regs_call);
260 new = ftrace_call_replace(ip, (unsigned long)func);
261 ret = update_ftrace_func(ip, new);
262 }
263
264 return ret;
265}
266
267static int is_ftrace_caller(unsigned long ip)
268{
269 if (ip == ftrace_update_func)
270 return 1;
271
272 return 0;
273}
274
275/*
276 * A breakpoint was added to the code address we are about to
277 * modify, and this is the handle that will just skip over it.
278 * We are either changing a nop into a trace call, or a trace
279 * call to a nop. While the change is taking place, we treat
280 * it just like it was a nop.
281 */
282int ftrace_int3_handler(struct pt_regs *regs)
283{
284 unsigned long ip;
285
286 if (WARN_ON_ONCE(!regs))
287 return 0;
288
289 ip = regs->ip - 1;
290 if (!ftrace_location(ip) && !is_ftrace_caller(ip))
291 return 0;
292
293 regs->ip += MCOUNT_INSN_SIZE - 1;
294
295 return 1;
296}
297
298static int ftrace_write(unsigned long ip, const char *val, int size)
299{
300 /*
301 * On x86_64, kernel text mappings are mapped read-only with
302 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
303 * of the kernel text mapping to modify the kernel text.
304 *
305 * For 32bit kernels, these mappings are same and we can use
306 * kernel identity mapping to modify code.
307 */
308 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
309 ip = (unsigned long)__va(__pa_symbol(ip));
310
311 if (probe_kernel_write((void *)ip, val, size))
312 return -EPERM;
313
314 return 0;
315}
316
317static int add_break(unsigned long ip, const char *old)
318{
319 unsigned char replaced[MCOUNT_INSN_SIZE];
320 unsigned char brk = BREAKPOINT_INSTRUCTION;
321
322 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
323 return -EFAULT;
324
325 /* Make sure it is what we expect it to be */
326 if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
327 return -EINVAL;
328
329 return ftrace_write(ip, &brk, 1);
330}
331
332static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
333{
334 unsigned const char *old;
335 unsigned long ip = rec->ip;
336
337 old = ftrace_call_replace(ip, addr);
338
339 return add_break(rec->ip, old);
340}
341
342
343static int add_brk_on_nop(struct dyn_ftrace *rec)
344{
345 unsigned const char *old;
346
347 old = ftrace_nop_replace();
348
349 return add_break(rec->ip, old);
350}
351
352/*
353 * If the record has the FTRACE_FL_REGS set, that means that it
354 * wants to convert to a callback that saves all regs. If FTRACE_FL_REGS
355 * is not not set, then it wants to convert to the normal callback.
356 */
357static unsigned long get_ftrace_addr(struct dyn_ftrace *rec)
358{
359 if (rec->flags & FTRACE_FL_REGS)
360 return (unsigned long)FTRACE_REGS_ADDR;
361 else
362 return (unsigned long)FTRACE_ADDR;
363}
364
365/*
366 * The FTRACE_FL_REGS_EN is set when the record already points to
367 * a function that saves all the regs. Basically the '_EN' version
368 * represents the current state of the function.
369 */
370static unsigned long get_ftrace_old_addr(struct dyn_ftrace *rec)
371{
372 if (rec->flags & FTRACE_FL_REGS_EN)
373 return (unsigned long)FTRACE_REGS_ADDR;
374 else
375 return (unsigned long)FTRACE_ADDR;
376}
377
378static int add_breakpoints(struct dyn_ftrace *rec, int enable)
379{
380 unsigned long ftrace_addr;
381 int ret;
382
383 ret = ftrace_test_record(rec, enable);
384
385 ftrace_addr = get_ftrace_addr(rec);
386
387 switch (ret) {
388 case FTRACE_UPDATE_IGNORE:
389 return 0;
390
391 case FTRACE_UPDATE_MAKE_CALL:
392 /* converting nop to call */
393 return add_brk_on_nop(rec);
394
395 case FTRACE_UPDATE_MODIFY_CALL_REGS:
396 case FTRACE_UPDATE_MODIFY_CALL:
397 ftrace_addr = get_ftrace_old_addr(rec);
398 /* fall through */
399 case FTRACE_UPDATE_MAKE_NOP:
400 /* converting a call to a nop */
401 return add_brk_on_call(rec, ftrace_addr);
402 }
403 return 0;
404}
405
406/*
407 * On error, we need to remove breakpoints. This needs to
408 * be done caefully. If the address does not currently have a
409 * breakpoint, we know we are done. Otherwise, we look at the
410 * remaining 4 bytes of the instruction. If it matches a nop
411 * we replace the breakpoint with the nop. Otherwise we replace
412 * it with the call instruction.
413 */
414static int remove_breakpoint(struct dyn_ftrace *rec)
415{
416 unsigned char ins[MCOUNT_INSN_SIZE];
417 unsigned char brk = BREAKPOINT_INSTRUCTION;
418 const unsigned char *nop;
419 unsigned long ftrace_addr;
420 unsigned long ip = rec->ip;
421
422 /* If we fail the read, just give up */
423 if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
424 return -EFAULT;
425
426 /* If this does not have a breakpoint, we are done */
427 if (ins[0] != brk)
428 return 0;
429
430 nop = ftrace_nop_replace();
431
432 /*
433 * If the last 4 bytes of the instruction do not match
434 * a nop, then we assume that this is a call to ftrace_addr.
435 */
436 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
437 /*
438 * For extra paranoidism, we check if the breakpoint is on
439 * a call that would actually jump to the ftrace_addr.
440 * If not, don't touch the breakpoint, we make just create
441 * a disaster.
442 */
443 ftrace_addr = get_ftrace_addr(rec);
444 nop = ftrace_call_replace(ip, ftrace_addr);
445
446 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
447 goto update;
448
449 /* Check both ftrace_addr and ftrace_old_addr */
450 ftrace_addr = get_ftrace_old_addr(rec);
451 nop = ftrace_call_replace(ip, ftrace_addr);
452
453 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
454 return -EINVAL;
455 }
456
457 update:
458 return ftrace_write(ip, nop, 1);
459}
460
461static int add_update_code(unsigned long ip, unsigned const char *new)
462{
463 /* skip breakpoint */
464 ip++;
465 new++;
466 return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1);
467}
468
469static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
470{
471 unsigned long ip = rec->ip;
472 unsigned const char *new;
473
474 new = ftrace_call_replace(ip, addr);
475 return add_update_code(ip, new);
476}
477
478static int add_update_nop(struct dyn_ftrace *rec)
479{
480 unsigned long ip = rec->ip;
481 unsigned const char *new;
482
483 new = ftrace_nop_replace();
484 return add_update_code(ip, new);
485}
486
487static int add_update(struct dyn_ftrace *rec, int enable)
488{
489 unsigned long ftrace_addr;
490 int ret;
491
492 ret = ftrace_test_record(rec, enable);
493
494 ftrace_addr = get_ftrace_addr(rec);
495
496 switch (ret) {
497 case FTRACE_UPDATE_IGNORE:
498 return 0;
499
500 case FTRACE_UPDATE_MODIFY_CALL_REGS:
501 case FTRACE_UPDATE_MODIFY_CALL:
502 case FTRACE_UPDATE_MAKE_CALL:
503 /* converting nop to call */
504 return add_update_call(rec, ftrace_addr);
505
506 case FTRACE_UPDATE_MAKE_NOP:
507 /* converting a call to a nop */
508 return add_update_nop(rec);
509 }
510
511 return 0;
512}
513
514static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
515{
516 unsigned long ip = rec->ip;
517 unsigned const char *new;
518
519 new = ftrace_call_replace(ip, addr);
520
521 return ftrace_write(ip, new, 1);
522}
523
524static int finish_update_nop(struct dyn_ftrace *rec)
525{
526 unsigned long ip = rec->ip;
527 unsigned const char *new;
528
529 new = ftrace_nop_replace();
530
531 return ftrace_write(ip, new, 1);
532}
533
534static int finish_update(struct dyn_ftrace *rec, int enable)
535{
536 unsigned long ftrace_addr;
537 int ret;
538
539 ret = ftrace_update_record(rec, enable);
540
541 ftrace_addr = get_ftrace_addr(rec);
542
543 switch (ret) {
544 case FTRACE_UPDATE_IGNORE:
545 return 0;
546
547 case FTRACE_UPDATE_MODIFY_CALL_REGS:
548 case FTRACE_UPDATE_MODIFY_CALL:
549 case FTRACE_UPDATE_MAKE_CALL:
550 /* converting nop to call */
551 return finish_update_call(rec, ftrace_addr);
552
553 case FTRACE_UPDATE_MAKE_NOP:
554 /* converting a call to a nop */
555 return finish_update_nop(rec);
556 }
557
558 return 0;
559}
560
561static void do_sync_core(void *data)
562{
563 sync_core();
564}
565
566static void run_sync(void)
567{
568 int enable_irqs = irqs_disabled();
569
570 /* We may be called with interrupts disbled (on bootup). */
571 if (enable_irqs)
572 local_irq_enable();
573 on_each_cpu(do_sync_core, NULL, 1);
574 if (enable_irqs)
575 local_irq_disable();
576}
577
578void ftrace_replace_code(int enable)
579{
580 struct ftrace_rec_iter *iter;
581 struct dyn_ftrace *rec;
582 const char *report = "adding breakpoints";
583 int count = 0;
584 int ret;
585
586 for_ftrace_rec_iter(iter) {
587 rec = ftrace_rec_iter_record(iter);
588
589 ret = add_breakpoints(rec, enable);
590 if (ret)
591 goto remove_breakpoints;
592 count++;
593 }
594
595 run_sync();
596
597 report = "updating code";
598
599 for_ftrace_rec_iter(iter) {
600 rec = ftrace_rec_iter_record(iter);
601
602 ret = add_update(rec, enable);
603 if (ret)
604 goto remove_breakpoints;
605 }
606
607 run_sync();
608
609 report = "removing breakpoints";
610
611 for_ftrace_rec_iter(iter) {
612 rec = ftrace_rec_iter_record(iter);
613
614 ret = finish_update(rec, enable);
615 if (ret)
616 goto remove_breakpoints;
617 }
618
619 run_sync();
620
621 return;
622
623 remove_breakpoints:
624 ftrace_bug(ret, rec ? rec->ip : 0);
625 printk(KERN_WARNING "Failed on %s (%d):\n", report, count);
626 for_ftrace_rec_iter(iter) {
627 rec = ftrace_rec_iter_record(iter);
628 /*
629 * Breakpoints are handled only when this function is in
630 * progress. The system could not work with them.
631 */
632 if (remove_breakpoint(rec))
633 BUG();
634 }
635 run_sync();
636}
637
638static int
639ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
640 unsigned const char *new_code)
641{
642 int ret;
643
644 ret = add_break(ip, old_code);
645 if (ret)
646 goto out;
647
648 run_sync();
649
650 ret = add_update_code(ip, new_code);
651 if (ret)
652 goto fail_update;
653
654 run_sync();
655
656 ret = ftrace_write(ip, new_code, 1);
657 /*
658 * The breakpoint is handled only when this function is in progress.
659 * The system could not work if we could not remove it.
660 */
661 BUG_ON(ret);
662 out:
663 run_sync();
664 return ret;
665
666 fail_update:
667 /* Also here the system could not work with the breakpoint */
668 if (ftrace_write(ip, old_code, 1))
669 BUG();
670 goto out;
671}
672
673void arch_ftrace_update_code(int command)
674{
675 /* See comment above by declaration of modifying_ftrace_code */
676 atomic_inc(&modifying_ftrace_code);
677
678 ftrace_modify_all_code(command);
679
680 atomic_dec(&modifying_ftrace_code);
681}
682
683int __init ftrace_dyn_arch_init(void)
684{
685 return 0;
686}
687#endif
688
689#ifdef CONFIG_FUNCTION_GRAPH_TRACER
690
691#ifdef CONFIG_DYNAMIC_FTRACE
692extern void ftrace_graph_call(void);
693
694static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
695{
696 static union ftrace_code_union calc;
697
698 /* Jmp not a call (ignore the .e8) */
699 calc.e8 = 0xe9;
700 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
701
702 /*
703 * ftrace external locks synchronize the access to the static variable.
704 */
705 return calc.code;
706}
707
708static int ftrace_mod_jmp(unsigned long ip, void *func)
709{
710 unsigned char *new;
711
712 new = ftrace_jmp_replace(ip, (unsigned long)func);
713
714 return update_ftrace_func(ip, new);
715}
716
717int ftrace_enable_ftrace_graph_caller(void)
718{
719 unsigned long ip = (unsigned long)(&ftrace_graph_call);
720
721 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
722}
723
724int ftrace_disable_ftrace_graph_caller(void)
725{
726 unsigned long ip = (unsigned long)(&ftrace_graph_call);
727
728 return ftrace_mod_jmp(ip, &ftrace_stub);
729}
730
731#endif /* !CONFIG_DYNAMIC_FTRACE */
732
733/*
734 * Hook the return address and push it in the stack of return addrs
735 * in current thread info.
736 */
737void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
738 unsigned long frame_pointer)
739{
740 unsigned long old;
741 int faulted;
742 struct ftrace_graph_ent trace;
743 unsigned long return_hooker = (unsigned long)
744 &return_to_handler;
745
746 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
747 return;
748
749 /*
750 * Protect against fault, even if it shouldn't
751 * happen. This tool is too much intrusive to
752 * ignore such a protection.
753 */
754 asm volatile(
755 "1: " _ASM_MOV " (%[parent]), %[old]\n"
756 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
757 " movl $0, %[faulted]\n"
758 "3:\n"
759
760 ".section .fixup, \"ax\"\n"
761 "4: movl $1, %[faulted]\n"
762 " jmp 3b\n"
763 ".previous\n"
764
765 _ASM_EXTABLE(1b, 4b)
766 _ASM_EXTABLE(2b, 4b)
767
768 : [old] "=&r" (old), [faulted] "=r" (faulted)
769 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
770 : "memory"
771 );
772
773 if (unlikely(faulted)) {
774 ftrace_graph_stop();
775 WARN_ON(1);
776 return;
777 }
778
779 trace.func = self_addr;
780 trace.depth = current->curr_ret_stack + 1;
781
782 /* Only trace if the calling function expects to */
783 if (!ftrace_graph_entry(&trace)) {
784 *parent = old;
785 return;
786 }
787
788 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
789 frame_pointer) == -EBUSY) {
790 *parent = old;
791 return;
792 }
793}
794#endif /* CONFIG_FUNCTION_GRAPH_TRACER */