xref: /openbmc/linux/arch/x86/kernel/ftrace.c (revision 275876e2)
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 
33 int ftrace_arch_code_modify_prepare(void)
34 {
35 	set_kernel_text_rw();
36 	set_all_modules_text_rw();
37 	return 0;
38 }
39 
40 int ftrace_arch_code_modify_post_process(void)
41 {
42 	set_all_modules_text_ro();
43 	set_kernel_text_ro();
44 	return 0;
45 }
46 
47 union ftrace_code_union {
48 	char code[MCOUNT_INSN_SIZE];
49 	struct {
50 		char e8;
51 		int offset;
52 	} __attribute__((packed));
53 };
54 
55 static int ftrace_calc_offset(long ip, long addr)
56 {
57 	return (int)(addr - ip);
58 }
59 
60 static 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 
74 static inline int
75 within(unsigned long addr, unsigned long start, unsigned long end)
76 {
77 	return addr >= start && addr < end;
78 }
79 
80 static 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 
96 static const unsigned char *ftrace_nop_replace(void)
97 {
98 	return ideal_nops[NOP_ATOMIC5];
99 }
100 
101 static int
102 ftrace_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 
136 int 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 
161 int 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  */
204 atomic_t modifying_ftrace_code __read_mostly;
205 
206 static int
207 ftrace_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  */
218 int 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 
225 static unsigned long ftrace_update_func;
226 
227 static 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 
248 int 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 
267 static 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  */
282 int 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 
298 static int ftrace_write(unsigned long ip, const char *val, int size)
299 {
300 	ip = text_ip_addr(ip);
301 
302 	if (probe_kernel_write((void *)ip, val, size))
303 		return -EPERM;
304 
305 	return 0;
306 }
307 
308 static int add_break(unsigned long ip, const char *old)
309 {
310 	unsigned char replaced[MCOUNT_INSN_SIZE];
311 	unsigned char brk = BREAKPOINT_INSTRUCTION;
312 
313 	if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
314 		return -EFAULT;
315 
316 	/* Make sure it is what we expect it to be */
317 	if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
318 		return -EINVAL;
319 
320 	return ftrace_write(ip, &brk, 1);
321 }
322 
323 static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
324 {
325 	unsigned const char *old;
326 	unsigned long ip = rec->ip;
327 
328 	old = ftrace_call_replace(ip, addr);
329 
330 	return add_break(rec->ip, old);
331 }
332 
333 
334 static int add_brk_on_nop(struct dyn_ftrace *rec)
335 {
336 	unsigned const char *old;
337 
338 	old = ftrace_nop_replace();
339 
340 	return add_break(rec->ip, old);
341 }
342 
343 static int add_breakpoints(struct dyn_ftrace *rec, int enable)
344 {
345 	unsigned long ftrace_addr;
346 	int ret;
347 
348 	ftrace_addr = ftrace_get_addr_curr(rec);
349 
350 	ret = ftrace_test_record(rec, enable);
351 
352 	switch (ret) {
353 	case FTRACE_UPDATE_IGNORE:
354 		return 0;
355 
356 	case FTRACE_UPDATE_MAKE_CALL:
357 		/* converting nop to call */
358 		return add_brk_on_nop(rec);
359 
360 	case FTRACE_UPDATE_MODIFY_CALL:
361 	case FTRACE_UPDATE_MAKE_NOP:
362 		/* converting a call to a nop */
363 		return add_brk_on_call(rec, ftrace_addr);
364 	}
365 	return 0;
366 }
367 
368 /*
369  * On error, we need to remove breakpoints. This needs to
370  * be done caefully. If the address does not currently have a
371  * breakpoint, we know we are done. Otherwise, we look at the
372  * remaining 4 bytes of the instruction. If it matches a nop
373  * we replace the breakpoint with the nop. Otherwise we replace
374  * it with the call instruction.
375  */
376 static int remove_breakpoint(struct dyn_ftrace *rec)
377 {
378 	unsigned char ins[MCOUNT_INSN_SIZE];
379 	unsigned char brk = BREAKPOINT_INSTRUCTION;
380 	const unsigned char *nop;
381 	unsigned long ftrace_addr;
382 	unsigned long ip = rec->ip;
383 
384 	/* If we fail the read, just give up */
385 	if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
386 		return -EFAULT;
387 
388 	/* If this does not have a breakpoint, we are done */
389 	if (ins[0] != brk)
390 		return 0;
391 
392 	nop = ftrace_nop_replace();
393 
394 	/*
395 	 * If the last 4 bytes of the instruction do not match
396 	 * a nop, then we assume that this is a call to ftrace_addr.
397 	 */
398 	if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
399 		/*
400 		 * For extra paranoidism, we check if the breakpoint is on
401 		 * a call that would actually jump to the ftrace_addr.
402 		 * If not, don't touch the breakpoint, we make just create
403 		 * a disaster.
404 		 */
405 		ftrace_addr = ftrace_get_addr_new(rec);
406 		nop = ftrace_call_replace(ip, ftrace_addr);
407 
408 		if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
409 			goto update;
410 
411 		/* Check both ftrace_addr and ftrace_old_addr */
412 		ftrace_addr = ftrace_get_addr_curr(rec);
413 		nop = ftrace_call_replace(ip, ftrace_addr);
414 
415 		if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
416 			return -EINVAL;
417 	}
418 
419  update:
420 	return ftrace_write(ip, nop, 1);
421 }
422 
423 static int add_update_code(unsigned long ip, unsigned const char *new)
424 {
425 	/* skip breakpoint */
426 	ip++;
427 	new++;
428 	return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1);
429 }
430 
431 static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
432 {
433 	unsigned long ip = rec->ip;
434 	unsigned const char *new;
435 
436 	new = ftrace_call_replace(ip, addr);
437 	return add_update_code(ip, new);
438 }
439 
440 static int add_update_nop(struct dyn_ftrace *rec)
441 {
442 	unsigned long ip = rec->ip;
443 	unsigned const char *new;
444 
445 	new = ftrace_nop_replace();
446 	return add_update_code(ip, new);
447 }
448 
449 static int add_update(struct dyn_ftrace *rec, int enable)
450 {
451 	unsigned long ftrace_addr;
452 	int ret;
453 
454 	ret = ftrace_test_record(rec, enable);
455 
456 	ftrace_addr  = ftrace_get_addr_new(rec);
457 
458 	switch (ret) {
459 	case FTRACE_UPDATE_IGNORE:
460 		return 0;
461 
462 	case FTRACE_UPDATE_MODIFY_CALL:
463 	case FTRACE_UPDATE_MAKE_CALL:
464 		/* converting nop to call */
465 		return add_update_call(rec, ftrace_addr);
466 
467 	case FTRACE_UPDATE_MAKE_NOP:
468 		/* converting a call to a nop */
469 		return add_update_nop(rec);
470 	}
471 
472 	return 0;
473 }
474 
475 static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
476 {
477 	unsigned long ip = rec->ip;
478 	unsigned const char *new;
479 
480 	new = ftrace_call_replace(ip, addr);
481 
482 	return ftrace_write(ip, new, 1);
483 }
484 
485 static int finish_update_nop(struct dyn_ftrace *rec)
486 {
487 	unsigned long ip = rec->ip;
488 	unsigned const char *new;
489 
490 	new = ftrace_nop_replace();
491 
492 	return ftrace_write(ip, new, 1);
493 }
494 
495 static int finish_update(struct dyn_ftrace *rec, int enable)
496 {
497 	unsigned long ftrace_addr;
498 	int ret;
499 
500 	ret = ftrace_update_record(rec, enable);
501 
502 	ftrace_addr = ftrace_get_addr_new(rec);
503 
504 	switch (ret) {
505 	case FTRACE_UPDATE_IGNORE:
506 		return 0;
507 
508 	case FTRACE_UPDATE_MODIFY_CALL:
509 	case FTRACE_UPDATE_MAKE_CALL:
510 		/* converting nop to call */
511 		return finish_update_call(rec, ftrace_addr);
512 
513 	case FTRACE_UPDATE_MAKE_NOP:
514 		/* converting a call to a nop */
515 		return finish_update_nop(rec);
516 	}
517 
518 	return 0;
519 }
520 
521 static void do_sync_core(void *data)
522 {
523 	sync_core();
524 }
525 
526 static void run_sync(void)
527 {
528 	int enable_irqs = irqs_disabled();
529 
530 	/* We may be called with interrupts disbled (on bootup). */
531 	if (enable_irqs)
532 		local_irq_enable();
533 	on_each_cpu(do_sync_core, NULL, 1);
534 	if (enable_irqs)
535 		local_irq_disable();
536 }
537 
538 void ftrace_replace_code(int enable)
539 {
540 	struct ftrace_rec_iter *iter;
541 	struct dyn_ftrace *rec;
542 	const char *report = "adding breakpoints";
543 	int count = 0;
544 	int ret;
545 
546 	for_ftrace_rec_iter(iter) {
547 		rec = ftrace_rec_iter_record(iter);
548 
549 		ret = add_breakpoints(rec, enable);
550 		if (ret)
551 			goto remove_breakpoints;
552 		count++;
553 	}
554 
555 	run_sync();
556 
557 	report = "updating code";
558 
559 	for_ftrace_rec_iter(iter) {
560 		rec = ftrace_rec_iter_record(iter);
561 
562 		ret = add_update(rec, enable);
563 		if (ret)
564 			goto remove_breakpoints;
565 	}
566 
567 	run_sync();
568 
569 	report = "removing breakpoints";
570 
571 	for_ftrace_rec_iter(iter) {
572 		rec = ftrace_rec_iter_record(iter);
573 
574 		ret = finish_update(rec, enable);
575 		if (ret)
576 			goto remove_breakpoints;
577 	}
578 
579 	run_sync();
580 
581 	return;
582 
583  remove_breakpoints:
584 	pr_warn("Failed on %s (%d):\n", report, count);
585 	ftrace_bug(ret, rec ? rec->ip : 0);
586 	for_ftrace_rec_iter(iter) {
587 		rec = ftrace_rec_iter_record(iter);
588 		/*
589 		 * Breakpoints are handled only when this function is in
590 		 * progress. The system could not work with them.
591 		 */
592 		if (remove_breakpoint(rec))
593 			BUG();
594 	}
595 	run_sync();
596 }
597 
598 static int
599 ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
600 		   unsigned const char *new_code)
601 {
602 	int ret;
603 
604 	ret = add_break(ip, old_code);
605 	if (ret)
606 		goto out;
607 
608 	run_sync();
609 
610 	ret = add_update_code(ip, new_code);
611 	if (ret)
612 		goto fail_update;
613 
614 	run_sync();
615 
616 	ret = ftrace_write(ip, new_code, 1);
617 	/*
618 	 * The breakpoint is handled only when this function is in progress.
619 	 * The system could not work if we could not remove it.
620 	 */
621 	BUG_ON(ret);
622  out:
623 	run_sync();
624 	return ret;
625 
626  fail_update:
627 	/* Also here the system could not work with the breakpoint */
628 	if (ftrace_write(ip, old_code, 1))
629 		BUG();
630 	goto out;
631 }
632 
633 void arch_ftrace_update_code(int command)
634 {
635 	/* See comment above by declaration of modifying_ftrace_code */
636 	atomic_inc(&modifying_ftrace_code);
637 
638 	ftrace_modify_all_code(command);
639 
640 	atomic_dec(&modifying_ftrace_code);
641 }
642 
643 int __init ftrace_dyn_arch_init(void)
644 {
645 	return 0;
646 }
647 #endif
648 
649 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
650 
651 #ifdef CONFIG_DYNAMIC_FTRACE
652 extern void ftrace_graph_call(void);
653 
654 static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
655 {
656 	static union ftrace_code_union calc;
657 
658 	/* Jmp not a call (ignore the .e8) */
659 	calc.e8		= 0xe9;
660 	calc.offset	= ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
661 
662 	/*
663 	 * ftrace external locks synchronize the access to the static variable.
664 	 */
665 	return calc.code;
666 }
667 
668 static int ftrace_mod_jmp(unsigned long ip, void *func)
669 {
670 	unsigned char *new;
671 
672 	new = ftrace_jmp_replace(ip, (unsigned long)func);
673 
674 	return update_ftrace_func(ip, new);
675 }
676 
677 int ftrace_enable_ftrace_graph_caller(void)
678 {
679 	unsigned long ip = (unsigned long)(&ftrace_graph_call);
680 
681 	return ftrace_mod_jmp(ip, &ftrace_graph_caller);
682 }
683 
684 int ftrace_disable_ftrace_graph_caller(void)
685 {
686 	unsigned long ip = (unsigned long)(&ftrace_graph_call);
687 
688 	return ftrace_mod_jmp(ip, &ftrace_stub);
689 }
690 
691 #endif /* !CONFIG_DYNAMIC_FTRACE */
692 
693 /*
694  * Hook the return address and push it in the stack of return addrs
695  * in current thread info.
696  */
697 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
698 			   unsigned long frame_pointer)
699 {
700 	unsigned long old;
701 	int faulted;
702 	struct ftrace_graph_ent trace;
703 	unsigned long return_hooker = (unsigned long)
704 				&return_to_handler;
705 
706 	if (unlikely(ftrace_graph_is_dead()))
707 		return;
708 
709 	if (unlikely(atomic_read(&current->tracing_graph_pause)))
710 		return;
711 
712 	/*
713 	 * Protect against fault, even if it shouldn't
714 	 * happen. This tool is too much intrusive to
715 	 * ignore such a protection.
716 	 */
717 	asm volatile(
718 		"1: " _ASM_MOV " (%[parent]), %[old]\n"
719 		"2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
720 		"   movl $0, %[faulted]\n"
721 		"3:\n"
722 
723 		".section .fixup, \"ax\"\n"
724 		"4: movl $1, %[faulted]\n"
725 		"   jmp 3b\n"
726 		".previous\n"
727 
728 		_ASM_EXTABLE(1b, 4b)
729 		_ASM_EXTABLE(2b, 4b)
730 
731 		: [old] "=&r" (old), [faulted] "=r" (faulted)
732 		: [parent] "r" (parent), [return_hooker] "r" (return_hooker)
733 		: "memory"
734 	);
735 
736 	if (unlikely(faulted)) {
737 		ftrace_graph_stop();
738 		WARN_ON(1);
739 		return;
740 	}
741 
742 	trace.func = self_addr;
743 	trace.depth = current->curr_ret_stack + 1;
744 
745 	/* Only trace if the calling function expects to */
746 	if (!ftrace_graph_entry(&trace)) {
747 		*parent = old;
748 		return;
749 	}
750 
751 	if (ftrace_push_return_trace(old, self_addr, &trace.depth,
752 		    frame_pointer) == -EBUSY) {
753 		*parent = old;
754 		return;
755 	}
756 }
757 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
758