Linux Audio

Check our new training course

Loading...
v3.5.6
 
  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 int
 81do_ftrace_mod_code(unsigned long ip, const void *new_code)
 82{
 83	/*
 84	 * On x86_64, kernel text mappings are mapped read-only with
 85	 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
 86	 * of the kernel text mapping to modify the kernel text.
 87	 *
 88	 * For 32bit kernels, these mappings are same and we can use
 89	 * kernel identity mapping to modify code.
 90	 */
 91	if (within(ip, (unsigned long)_text, (unsigned long)_etext))
 92		ip = (unsigned long)__va(__pa(ip));
 93
 94	return probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE);
 95}
 96
 97static const unsigned char *ftrace_nop_replace(void)
 98{
 99	return ideal_nops[NOP_ATOMIC5];
100}
101
102static int
103ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code,
104		   unsigned const char *new_code)
105{
106	unsigned char replaced[MCOUNT_INSN_SIZE];
107
108	/*
109	 * Note: Due to modules and __init, code can
110	 *  disappear and change, we need to protect against faulting
111	 *  as well as code changing. We do this by using the
112	 *  probe_kernel_* functions.
113	 *
114	 * No real locking needed, this code is run through
115	 * kstop_machine, or before SMP starts.
116	 */
117
118	/* read the text we want to modify */
119	if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
 
120		return -EFAULT;
 
121
122	/* Make sure it is what we expect it to be */
123	if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
 
124		return -EINVAL;
 
125
126	/* replace the text with the new text */
127	if (do_ftrace_mod_code(ip, new_code))
128		return -EPERM;
129
130	sync_core();
 
 
 
 
 
 
 
 
 
 
 
131
 
 
 
 
 
132	return 0;
133}
134
135int ftrace_make_nop(struct module *mod,
136		    struct dyn_ftrace *rec, unsigned long addr)
137{
138	unsigned const char *new, *old;
139	unsigned long ip = rec->ip;
 
140
141	old = ftrace_call_replace(ip, addr);
142	new = ftrace_nop_replace();
143
144	/*
145	 * On boot up, and when modules are loaded, the MCOUNT_ADDR
146	 * is converted to a nop, and will never become MCOUNT_ADDR
147	 * again. This code is either running before SMP (on boot up)
148	 * or before the code will ever be executed (module load).
149	 * We do not want to use the breakpoint version in this case,
150	 * just modify the code directly.
151	 */
152	if (addr == MCOUNT_ADDR)
153		return ftrace_modify_code_direct(rec->ip, old, new);
154
155	/* Normal cases use add_brk_on_nop */
 
 
 
156	WARN_ONCE(1, "invalid use of ftrace_make_nop");
157	return -EINVAL;
158}
159
160int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
161{
162	unsigned const char *new, *old;
163	unsigned long ip = rec->ip;
 
164
165	old = ftrace_nop_replace();
166	new = ftrace_call_replace(ip, addr);
167
168	/* Should only be called when module is loaded */
169	return ftrace_modify_code_direct(rec->ip, old, new);
170}
171
172/*
173 * The modifying_ftrace_code is used to tell the breakpoint
174 * handler to call ftrace_int3_handler(). If it fails to
175 * call this handler for a breakpoint added by ftrace, then
176 * the kernel may crash.
177 *
178 * As atomic_writes on x86 do not need a barrier, we do not
179 * need to add smp_mb()s for this to work. It is also considered
180 * that we can not read the modifying_ftrace_code before
181 * executing the breakpoint. That would be quite remarkable if
182 * it could do that. Here's the flow that is required:
183 *
184 *   CPU-0                          CPU-1
185 *
186 * atomic_inc(mfc);
187 * write int3s
188 *				<trap-int3> // implicit (r)mb
189 *				if (atomic_read(mfc))
190 *					call ftrace_int3_handler()
191 *
192 * Then when we are finished:
193 *
194 * atomic_dec(mfc);
195 *
196 * If we hit a breakpoint that was not set by ftrace, it does not
197 * matter if ftrace_int3_handler() is called or not. It will
198 * simply be ignored. But it is crucial that a ftrace nop/caller
199 * breakpoint is handled. No other user should ever place a
200 * breakpoint on an ftrace nop/caller location. It must only
201 * be done by this code.
202 */
203atomic_t modifying_ftrace_code __read_mostly;
204
205static int
206ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
207		   unsigned const char *new_code);
 
208
209int ftrace_update_ftrace_func(ftrace_func_t func)
210{
211	unsigned long ip = (unsigned long)(&ftrace_call);
212	unsigned char old[MCOUNT_INSN_SIZE], *new;
213	int ret;
214
215	memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
216	new = ftrace_call_replace(ip, (unsigned long)func);
 
217
218	/* See comment above by declaration of modifying_ftrace_code */
219	atomic_inc(&modifying_ftrace_code);
220
221	ret = ftrace_modify_code(ip, old, new);
222
223	atomic_dec(&modifying_ftrace_code);
224
225	return ret;
226}
227
228/*
229 * A breakpoint was added to the code address we are about to
230 * modify, and this is the handle that will just skip over it.
231 * We are either changing a nop into a trace call, or a trace
232 * call to a nop. While the change is taking place, we treat
233 * it just like it was a nop.
234 */
235int ftrace_int3_handler(struct pt_regs *regs)
236{
237	if (WARN_ON_ONCE(!regs))
238		return 0;
 
 
239
240	if (!ftrace_location(regs->ip - 1))
241		return 0;
242
243	regs->ip += MCOUNT_INSN_SIZE - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
244
245	return 1;
246}
247
248static int ftrace_write(unsigned long ip, const char *val, int size)
249{
250	/*
251	 * On x86_64, kernel text mappings are mapped read-only with
252	 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
253	 * of the kernel text mapping to modify the kernel text.
254	 *
255	 * For 32bit kernels, these mappings are same and we can use
256	 * kernel identity mapping to modify code.
257	 */
258	if (within(ip, (unsigned long)_text, (unsigned long)_etext))
259		ip = (unsigned long)__va(__pa(ip));
 
 
260
261	return probe_kernel_write((void *)ip, val, size);
 
 
 
262}
263
264static int add_break(unsigned long ip, const char *old)
265{
266	unsigned char replaced[MCOUNT_INSN_SIZE];
267	unsigned char brk = BREAKPOINT_INSTRUCTION;
268
269	if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
270		return -EFAULT;
271
272	/* Make sure it is what we expect it to be */
273	if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
274		return -EINVAL;
275
276	if (ftrace_write(ip, &brk, 1))
277		return -EPERM;
278
279	return 0;
280}
281
282static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
283{
284	unsigned const char *old;
285	unsigned long ip = rec->ip;
286
287	old = ftrace_call_replace(ip, addr);
288
289	return add_break(rec->ip, old);
290}
291
 
 
292
293static int add_brk_on_nop(struct dyn_ftrace *rec)
 
 
 
294{
295	unsigned const char *old;
296
297	old = ftrace_nop_replace();
298
299	return add_break(rec->ip, old);
300}
301
302static int add_breakpoints(struct dyn_ftrace *rec, int enable)
303{
304	unsigned long ftrace_addr;
305	int ret;
306
307	ret = ftrace_test_record(rec, enable);
308
309	ftrace_addr = (unsigned long)FTRACE_ADDR;
310
311	switch (ret) {
312	case FTRACE_UPDATE_IGNORE:
313		return 0;
314
315	case FTRACE_UPDATE_MAKE_CALL:
316		/* converting nop to call */
317		return add_brk_on_nop(rec);
318
319	case FTRACE_UPDATE_MAKE_NOP:
320		/* converting a call to a nop */
321		return add_brk_on_call(rec, ftrace_addr);
322	}
323	return 0;
324}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
325
326/*
327 * On error, we need to remove breakpoints. This needs to
328 * be done caefully. If the address does not currently have a
329 * breakpoint, we know we are done. Otherwise, we look at the
330 * remaining 4 bytes of the instruction. If it matches a nop
331 * we replace the breakpoint with the nop. Otherwise we replace
332 * it with the call instruction.
333 */
334static int remove_breakpoint(struct dyn_ftrace *rec)
335{
336	unsigned char ins[MCOUNT_INSN_SIZE];
337	unsigned char brk = BREAKPOINT_INSTRUCTION;
338	const unsigned char *nop;
339	unsigned long ftrace_addr;
340	unsigned long ip = rec->ip;
341
342	/* If we fail the read, just give up */
343	if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
344		return -EFAULT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
345
346	/* If this does not have a breakpoint, we are done */
347	if (ins[0] != brk)
348		return -1;
 
 
 
 
 
 
 
 
 
 
349
350	nop = ftrace_nop_replace();
351
352	/*
353	 * If the last 4 bytes of the instruction do not match
354	 * a nop, then we assume that this is a call to ftrace_addr.
 
355	 */
356	if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
357		/*
358		 * For extra paranoidism, we check if the breakpoint is on
359		 * a call that would actually jump to the ftrace_addr.
360		 * If not, don't touch the breakpoint, we make just create
361		 * a disaster.
362		 */
363		ftrace_addr = (unsigned long)FTRACE_ADDR;
364		nop = ftrace_call_replace(ip, ftrace_addr);
365
366		if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
367			return -EINVAL;
368	}
369
370	return probe_kernel_write((void *)ip, &nop[0], 1);
371}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
372
373static int add_update_code(unsigned long ip, unsigned const char *new)
374{
375	/* skip breakpoint */
376	ip++;
377	new++;
378	if (ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1))
379		return -EPERM;
380	return 0;
381}
382
383static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
384{
385	unsigned long ip = rec->ip;
386	unsigned const char *new;
387
388	new = ftrace_call_replace(ip, addr);
389	return add_update_code(ip, new);
390}
391
392static int add_update_nop(struct dyn_ftrace *rec)
393{
394	unsigned long ip = rec->ip;
395	unsigned const char *new;
396
397	new = ftrace_nop_replace();
398	return add_update_code(ip, new);
399}
400
401static int add_update(struct dyn_ftrace *rec, int enable)
402{
403	unsigned long ftrace_addr;
404	int ret;
405
406	ret = ftrace_test_record(rec, enable);
 
407
408	ftrace_addr = (unsigned long)FTRACE_ADDR;
 
 
 
 
 
 
 
409
410	switch (ret) {
411	case FTRACE_UPDATE_IGNORE:
412		return 0;
413
414	case FTRACE_UPDATE_MAKE_CALL:
415		/* converting nop to call */
416		return add_update_call(rec, ftrace_addr);
417
418	case FTRACE_UPDATE_MAKE_NOP:
419		/* converting a call to a nop */
420		return add_update_nop(rec);
421	}
422
 
 
 
 
 
 
423	return 0;
424}
425
426static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
427{
428	unsigned long ip = rec->ip;
429	unsigned const char *new;
 
 
 
430
431	new = ftrace_call_replace(ip, addr);
432
433	if (ftrace_write(ip, new, 1))
434		return -EPERM;
435
436	return 0;
 
 
 
 
 
 
 
 
 
 
 
437}
438
439static int finish_update_nop(struct dyn_ftrace *rec)
440{
441	unsigned long ip = rec->ip;
442	unsigned const char *new;
443
444	new = ftrace_nop_replace();
 
 
 
 
 
 
445
446	if (ftrace_write(ip, new, 1))
447		return -EPERM;
448	return 0;
449}
450
451static int finish_update(struct dyn_ftrace *rec, int enable)
452{
453	unsigned long ftrace_addr;
454	int ret;
 
 
 
455
456	ret = ftrace_update_record(rec, enable);
457
458	ftrace_addr = (unsigned long)FTRACE_ADDR;
459
460	switch (ret) {
461	case FTRACE_UPDATE_IGNORE:
462		return 0;
463
464	case FTRACE_UPDATE_MAKE_CALL:
465		/* converting nop to call */
466		return finish_update_call(rec, ftrace_addr);
467
468	case FTRACE_UPDATE_MAKE_NOP:
469		/* converting a call to a nop */
470		return finish_update_nop(rec);
471	}
472
473	return 0;
474}
475
476static void do_sync_core(void *data)
477{
478	sync_core();
479}
480
481static void run_sync(void)
482{
483	int enable_irqs = irqs_disabled();
484
485	/* We may be called with interrupts disbled (on bootup). */
486	if (enable_irqs)
487		local_irq_enable();
488	on_each_cpu(do_sync_core, NULL, 1);
489	if (enable_irqs)
490		local_irq_disable();
491}
492
493void ftrace_replace_code(int enable)
 
494{
495	struct ftrace_rec_iter *iter;
496	struct dyn_ftrace *rec;
497	const char *report = "adding breakpoints";
498	int count = 0;
499	int ret;
500
501	for_ftrace_rec_iter(iter) {
502		rec = ftrace_rec_iter_record(iter);
 
503
504		ret = add_breakpoints(rec, enable);
505		if (ret)
506			goto remove_breakpoints;
507		count++;
508	}
509
510	run_sync();
511
512	report = "updating code";
513
514	for_ftrace_rec_iter(iter) {
515		rec = ftrace_rec_iter_record(iter);
516
517		ret = add_update(rec, enable);
518		if (ret)
519			goto remove_breakpoints;
520	}
521
522	run_sync();
523
524	report = "removing breakpoints";
 
525
526	for_ftrace_rec_iter(iter) {
527		rec = ftrace_rec_iter_record(iter);
 
 
 
 
 
 
 
528
529		ret = finish_update(rec, enable);
530		if (ret)
531			goto remove_breakpoints;
 
 
 
 
 
 
 
532	}
533
534	run_sync();
535
536	return;
 
 
 
537
538 remove_breakpoints:
539	ftrace_bug(ret, rec ? rec->ip : 0);
540	printk(KERN_WARNING "Failed on %s (%d):\n", report, count);
541	for_ftrace_rec_iter(iter) {
542		rec = ftrace_rec_iter_record(iter);
543		remove_breakpoint(rec);
544	}
545}
546
547static int
548ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
549		   unsigned const char *new_code)
550{
551	int ret;
552
553	ret = add_break(ip, old_code);
554	if (ret)
555		goto out;
556
557	run_sync();
 
 
558
559	ret = add_update_code(ip, new_code);
560	if (ret)
561		goto fail_update;
562
563	run_sync();
564
565	ret = ftrace_write(ip, new_code, 1);
566	if (ret) {
567		ret = -EPERM;
568		goto out;
569	}
570	run_sync();
571 out:
572	return ret;
573
574 fail_update:
575	probe_kernel_write((void *)ip, &old_code[0], 1);
576	goto out;
577}
578
579void arch_ftrace_update_code(int command)
580{
581	/* See comment above by declaration of modifying_ftrace_code */
582	atomic_inc(&modifying_ftrace_code);
583
584	ftrace_modify_all_code(command);
585
586	atomic_dec(&modifying_ftrace_code);
 
587}
588
589int __init ftrace_dyn_arch_init(void *data)
590{
591	/* The return code is retured via data */
592	*(unsigned long *)data = 0;
593
594	return 0;
595}
596#endif
597
598#ifdef CONFIG_FUNCTION_GRAPH_TRACER
599
600#ifdef CONFIG_DYNAMIC_FTRACE
601extern void ftrace_graph_call(void);
602
603static int ftrace_mod_jmp(unsigned long ip,
604			  int old_offset, int new_offset)
605{
606	unsigned char code[MCOUNT_INSN_SIZE];
607
608	if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
609		return -EFAULT;
610
611	if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
612		return -EINVAL;
613
614	*(int *)(&code[1]) = new_offset;
615
616	if (do_ftrace_mod_code(ip, &code))
617		return -EPERM;
 
618
 
 
619	return 0;
620}
621
622int ftrace_enable_ftrace_graph_caller(void)
623{
624	unsigned long ip = (unsigned long)(&ftrace_graph_call);
625	int old_offset, new_offset;
626
627	old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
628	new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
629
630	return ftrace_mod_jmp(ip, old_offset, new_offset);
631}
632
633int ftrace_disable_ftrace_graph_caller(void)
634{
635	unsigned long ip = (unsigned long)(&ftrace_graph_call);
636	int old_offset, new_offset;
637
638	old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
639	new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
640
641	return ftrace_mod_jmp(ip, old_offset, new_offset);
642}
643
644#endif /* !CONFIG_DYNAMIC_FTRACE */
645
646/*
647 * Hook the return address and push it in the stack of return addrs
648 * in current thread info.
649 */
650void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
651			   unsigned long frame_pointer)
652{
 
653	unsigned long old;
654	int faulted;
655	struct ftrace_graph_ent trace;
656	unsigned long return_hooker = (unsigned long)
657				&return_to_handler;
 
 
 
 
 
 
 
 
 
 
 
 
658
659	if (unlikely(atomic_read(&current->tracing_graph_pause)))
660		return;
661
662	/*
663	 * Protect against fault, even if it shouldn't
664	 * happen. This tool is too much intrusive to
665	 * ignore such a protection.
666	 */
667	asm volatile(
668		"1: " _ASM_MOV " (%[parent]), %[old]\n"
669		"2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
670		"   movl $0, %[faulted]\n"
671		"3:\n"
672
673		".section .fixup, \"ax\"\n"
674		"4: movl $1, %[faulted]\n"
675		"   jmp 3b\n"
676		".previous\n"
677
678		_ASM_EXTABLE(1b, 4b)
679		_ASM_EXTABLE(2b, 4b)
680
681		: [old] "=&r" (old), [faulted] "=r" (faulted)
682		: [parent] "r" (parent), [return_hooker] "r" (return_hooker)
683		: "memory"
684	);
685
686	if (unlikely(faulted)) {
687		ftrace_graph_stop();
688		WARN_ON(1);
689		return;
690	}
691
692	trace.func = self_addr;
693	trace.depth = current->curr_ret_stack + 1;
694
695	/* Only trace if the calling function expects to */
696	if (!ftrace_graph_entry(&trace)) {
697		*parent = old;
698		return;
699	}
700
701	if (ftrace_push_return_trace(old, self_addr, &trace.depth,
702		    frame_pointer) == -EBUSY) {
703		*parent = old;
704		return;
705	}
706}
707#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
v5.9
  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
 38static int ftrace_poke_late = 0;
 39
 40int 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
 53int 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
 67static const char *ftrace_nop_replace(void)
 
 68{
 69	return ideal_nops[NOP_ATOMIC5];
 
 
 
 
 
 
 
 
 
 
 
 70}
 71
 72static 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
 77static 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 */
108static int __ref
109ftrace_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
124int 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
151int 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 */
171int 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
178int 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
194void 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
250void arch_ftrace_update_code(int command)
251{
252	ftrace_modify_all_code(command);
 
 
 
 
 
 
 
 
 
 
 
 
 
253}
254
255int __init ftrace_dyn_arch_init(void)
256{
257	return 0;
 
 
 
 
 
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(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_regs_caller_ret(void);
286extern void ftrace_caller_end(void);
287extern void ftrace_caller_op_ptr(void);
288extern void ftrace_regs_caller_op_ptr(void);
289extern void ftrace_regs_caller_jmp(void);
290
291/* movq function_trace_op(%rip), %rdx */
292/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
293#define OP_REF_SIZE	7
294
295/*
296 * The ftrace_ops is passed to the function callback. Since the
297 * trampoline only services a single ftrace_ops, we can pass in
298 * that ops directly.
299 *
300 * The ftrace_op_code_union is used to create a pointer to the
301 * ftrace_ops that will be passed to the callback function.
302 */
303union ftrace_op_code_union {
304	char code[OP_REF_SIZE];
305	struct {
306		char op[3];
307		int offset;
308	} __attribute__((packed));
309};
310
311#define RET_SIZE		1
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 retq;
325	unsigned long *ptr;
326	void *trampoline;
327	void *ip;
328	/* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
329	unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
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 + 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(trampoline, (void *)start_offset, size);
363	if (WARN_ON(ret < 0))
364		goto fail;
365
366	ip = trampoline + size;
367
368	/* The trampoline ends with ret(q) */
369	retq = (unsigned long)ftrace_stub;
370	ret = copy_from_kernel_nofault(ip, (void *)retq, RET_SIZE);
371	if (WARN_ON(ret < 0))
372		goto fail;
373
374	/* No need to test direct calls on created trampolines */
375	if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
376		/* NOP the jnz 1f; but make sure it's a 2 byte jnz */
377		ip = trampoline + (jmp_offset - start_offset);
378		if (WARN_ON(*(char *)ip != 0x75))
379			goto fail;
380		ret = copy_from_kernel_nofault(ip, ideal_nops[2], 2);
381		if (ret < 0)
382			goto fail;
383	}
384
385	/*
386	 * The address of the ftrace_ops that is used for this trampoline
387	 * is stored at the end of the trampoline. This will be used to
388	 * load the third parameter for the callback. Basically, that
389	 * location at the end of the trampoline takes the place of
390	 * the global function_trace_op variable.
391	 */
 
 
392
393	ptr = (unsigned long *)(trampoline + size + RET_SIZE);
394	*ptr = (unsigned long)ops;
 
 
395
396	op_offset -= start_offset;
397	memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
 
398
399	/* Are we pointing to the reference? */
400	if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0))
401		goto fail;
 
402
403	/* Load the contents of ptr into the callback parameter */
404	offset = (unsigned long)ptr;
405	offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
406
407	op_ptr.offset = offset;
 
 
 
408
409	/* put in the new offset to the ftrace_ops */
410	memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
411
412	/* put in the call to the function */
413	mutex_lock(&text_mutex);
414	call_offset -= start_offset;
415	memcpy(trampoline + call_offset,
416	       text_gen_insn(CALL_INSN_OPCODE,
417			     trampoline + call_offset,
418			     ftrace_ops_get_func(ops)), CALL_INSN_SIZE);
419	mutex_unlock(&text_mutex);
420
421	/* ALLOC_TRAMP flags lets us know we created it */
422	ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
 
423
424	set_vm_flush_reset_perms(trampoline);
 
 
 
 
 
 
 
425
426	if (likely(system_state != SYSTEM_BOOTING))
427		set_memory_ro((unsigned long)trampoline, npages);
428	set_memory_x((unsigned long)trampoline, npages);
429	return (unsigned long)trampoline;
430fail:
431	tramp_free(trampoline);
432	return 0;
433}
434
435void set_ftrace_ops_ro(void)
436{
437	struct ftrace_ops *ops;
438	unsigned long start_offset;
439	unsigned long end_offset;
440	unsigned long npages;
441	unsigned long size;
442
443	do_for_each_ftrace_op(ops, ftrace_ops_list) {
444		if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
445			continue;
 
446
447		if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
448			start_offset = (unsigned long)ftrace_regs_caller;
449			end_offset = (unsigned long)ftrace_regs_caller_end;
450		} else {
451			start_offset = (unsigned long)ftrace_caller;
452			end_offset = (unsigned long)ftrace_caller_end;
453		}
454		size = end_offset - start_offset;
455		size = size + RET_SIZE + sizeof(void *);
456		npages = DIV_ROUND_UP(size, PAGE_SIZE);
457		set_memory_ro((unsigned long)ops->trampoline, npages);
458	} while_for_each_ftrace_op(ops);
459}
460
461static unsigned long calc_trampoline_call_offset(bool save_regs)
462{
463	unsigned long start_offset;
464	unsigned long call_offset;
465
466	if (save_regs) {
467		start_offset = (unsigned long)ftrace_regs_caller;
468		call_offset = (unsigned long)ftrace_regs_call;
469	} else {
470		start_offset = (unsigned long)ftrace_caller;
471		call_offset = (unsigned long)ftrace_call;
472	}
473
474	return call_offset - start_offset;
 
 
475}
476
477void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
478{
479	ftrace_func_t func;
480	unsigned long offset;
481	unsigned long ip;
482	unsigned int size;
483	const char *new;
484
485	if (!ops->trampoline) {
486		ops->trampoline = create_trampoline(ops, &size);
487		if (!ops->trampoline)
488			return;
489		ops->trampoline_size = size;
490		return;
 
 
 
 
 
 
 
 
 
491	}
492
493	/*
494	 * The ftrace_ops caller may set up its own trampoline.
495	 * In such a case, this code must not modify it.
496	 */
497	if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
498		return;
 
499
500	offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
501	ip = ops->trampoline + offset;
502	func = ftrace_ops_get_func(ops);
503
504	mutex_lock(&text_mutex);
505	/* Do a safe modify in case the trampoline is executing */
506	new = ftrace_call_replace(ip, (unsigned long)func);
507	text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
508	mutex_unlock(&text_mutex);
 
509}
510
511/* Return the address of the function the trampoline calls */
512static void *addr_from_call(void *ptr)
513{
514	union text_poke_insn call;
 
 
 
515	int ret;
516
517	ret = copy_from_kernel_nofault(&call, ptr, CALL_INSN_SIZE);
518	if (WARN_ON_ONCE(ret < 0))
519		return NULL;
520
521	/* Make sure this is a call */
522	if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) {
523		pr_warn("Expected E8, got %x\n", call.opcode);
524		return NULL;
525	}
526
527	return ptr + CALL_INSN_SIZE + call.disp;
528}
 
 
 
 
 
 
 
 
 
 
 
529
530void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
531			   unsigned long frame_pointer);
532
533/*
534 * If the ops->trampoline was not allocated, then it probably
535 * has a static trampoline func, or is the ftrace caller itself.
536 */
537static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
538{
539	unsigned long offset;
540	bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
541	void *ptr;
542
543	if (ops && ops->trampoline) {
544#ifdef CONFIG_FUNCTION_GRAPH_TRACER
545		/*
546		 * We only know about function graph tracer setting as static
547		 * trampoline.
548		 */
549		if (ops->trampoline == FTRACE_GRAPH_ADDR)
550			return (void *)prepare_ftrace_return;
551#endif
552		return NULL;
553	}
554
555	offset = calc_trampoline_call_offset(save_regs);
556
557	if (save_regs)
558		ptr = (void *)FTRACE_REGS_ADDR + offset;
559	else
560		ptr = (void *)FTRACE_ADDR + offset;
561
562	return addr_from_call(ptr);
 
 
 
 
 
 
563}
564
565void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
 
 
566{
567	unsigned long offset;
 
 
 
 
568
569	/* If we didn't allocate this trampoline, consider it static */
570	if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
571		return static_tramp_func(ops, rec);
572
573	offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
574	return addr_from_call((void *)ops->trampoline + offset);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
575}
576
577void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
578{
579	if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
580		return;
 
 
581
582	tramp_free((void *)ops->trampoline);
583	ops->trampoline = 0;
584}
585
586#endif /* CONFIG_X86_64 */
587#endif /* CONFIG_DYNAMIC_FTRACE */
 
 
 
 
 
 
588
589#ifdef CONFIG_FUNCTION_GRAPH_TRACER
590
591#ifdef CONFIG_DYNAMIC_FTRACE
592extern void ftrace_graph_call(void);
593
594static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
 
595{
596	return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
597}
 
 
 
 
 
 
 
598
599static int ftrace_mod_jmp(unsigned long ip, void *func)
600{
601	const char *new;
602
603	new = ftrace_jmp_replace(ip, (unsigned long)func);
604	text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
605	return 0;
606}
607
608int ftrace_enable_ftrace_graph_caller(void)
609{
610	unsigned long ip = (unsigned long)(&ftrace_graph_call);
 
 
 
 
611
612	return ftrace_mod_jmp(ip, &ftrace_graph_caller);
613}
614
615int ftrace_disable_ftrace_graph_caller(void)
616{
617	unsigned long ip = (unsigned long)(&ftrace_graph_call);
 
 
 
 
618
619	return ftrace_mod_jmp(ip, &ftrace_stub);
620}
621
622#endif /* !CONFIG_DYNAMIC_FTRACE */
623
624/*
625 * Hook the return address and push it in the stack of return addrs
626 * in current thread info.
627 */
628void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
629			   unsigned long frame_pointer)
630{
631	unsigned long return_hooker = (unsigned long)&return_to_handler;
632	unsigned long old;
633	int faulted;
634
635	/*
636	 * When resuming from suspend-to-ram, this function can be indirectly
637	 * called from early CPU startup code while the CPU is in real mode,
638	 * which would fail miserably.  Make sure the stack pointer is a
639	 * virtual address.
640	 *
641	 * This check isn't as accurate as virt_addr_valid(), but it should be
642	 * good enough for this purpose, and it's fast.
643	 */
644	if (unlikely((long)__builtin_frame_address(0) >= 0))
645		return;
646
647	if (unlikely(ftrace_graph_is_dead()))
648		return;
649
650	if (unlikely(atomic_read(&current->tracing_graph_pause)))
651		return;
652
653	/*
654	 * Protect against fault, even if it shouldn't
655	 * happen. This tool is too much intrusive to
656	 * ignore such a protection.
657	 */
658	asm volatile(
659		"1: " _ASM_MOV " (%[parent]), %[old]\n"
660		"2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
661		"   movl $0, %[faulted]\n"
662		"3:\n"
663
664		".section .fixup, \"ax\"\n"
665		"4: movl $1, %[faulted]\n"
666		"   jmp 3b\n"
667		".previous\n"
668
669		_ASM_EXTABLE(1b, 4b)
670		_ASM_EXTABLE(2b, 4b)
671
672		: [old] "=&r" (old), [faulted] "=r" (faulted)
673		: [parent] "r" (parent), [return_hooker] "r" (return_hooker)
674		: "memory"
675	);
676
677	if (unlikely(faulted)) {
678		ftrace_graph_stop();
679		WARN_ON(1);
680		return;
681	}
682
683	if (function_graph_enter(old, self_addr, frame_pointer, parent))
 
 
 
 
684		*parent = old;
 
 
 
 
 
 
 
 
685}
686#endif /* CONFIG_FUNCTION_GRAPH_TRACER */