Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v6.13.7
  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(&current->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 */
v4.6
 
   1/*
   2 * Dynamic function tracing support.
   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/slab.h>
  21#include <linux/init.h>
  22#include <linux/list.h>
  23#include <linux/module.h>
 
 
 
 
  24
  25#include <trace/syscall.h>
  26
  27#include <asm/cacheflush.h>
  28#include <asm/kprobes.h>
  29#include <asm/ftrace.h>
  30#include <asm/nops.h>
 
  31
  32#ifdef CONFIG_DYNAMIC_FTRACE
  33
  34int ftrace_arch_code_modify_prepare(void)
  35{
  36	set_kernel_text_rw();
  37	set_all_modules_text_rw();
  38	return 0;
  39}
  40
  41int ftrace_arch_code_modify_post_process(void)
 
  42{
  43	set_all_modules_text_ro();
  44	set_kernel_text_ro();
  45	return 0;
  46}
  47
  48union ftrace_code_union {
  49	char code[MCOUNT_INSN_SIZE];
  50	struct {
  51		unsigned char e8;
  52		int offset;
  53	} __attribute__((packed));
  54};
  55
  56static int ftrace_calc_offset(long ip, long addr)
  57{
  58	return (int)(addr - ip);
  59}
  60
  61static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
 
  62{
  63	static union ftrace_code_union calc;
  64
  65	calc.e8		= 0xe8;
  66	calc.offset	= ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
  67
  68	/*
  69	 * No locking needed, this must be called via kstop_machine
  70	 * which in essence is like running on a uniprocessor machine.
 
  71	 */
  72	return calc.code;
 
 
  73}
  74
  75static inline int
  76within(unsigned long addr, unsigned long start, unsigned long end)
  77{
  78	return addr >= start && addr < end;
  79}
  80
  81static unsigned long text_ip_addr(unsigned long ip)
  82{
  83	/*
  84	 * On x86_64, kernel text mappings are mapped read-only, so we use
  85	 * the kernel identity mapping instead of the kernel text mapping
  86	 * 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_symbol(ip));
  93
  94	return ip;
  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	ftrace_expected = old_code;
 109
 110	/*
 111	 * Note:
 112	 * We are paranoid about modifying text, as if a bug was to happen, it
 113	 * could cause us to read or write to someplace that could cause harm.
 114	 * Carefully read and modify the code with probe_kernel_*(), and make
 115	 * sure what we read is what we expected it to be before modifying it.
 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	ip = text_ip_addr(ip);
 
 
 
 
 
 
 
 
 
 
 
 127
 128	/* replace the text with the new text */
 129	if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE))
 130		return -EPERM;
 131
 132	sync_core();
 133
 
 
 134	return 0;
 135}
 136
 137int ftrace_make_nop(struct module *mod,
 138		    struct dyn_ftrace *rec, unsigned long addr)
 139{
 140	unsigned const char *new, *old;
 141	unsigned long ip = rec->ip;
 
 142
 143	old = ftrace_call_replace(ip, addr);
 144	new = ftrace_nop_replace();
 145
 146	/*
 147	 * On boot up, and when modules are loaded, the MCOUNT_ADDR
 148	 * is converted to a nop, and will never become MCOUNT_ADDR
 149	 * again. This code is either running before SMP (on boot up)
 150	 * or before the code will ever be executed (module load).
 151	 * We do not want to use the breakpoint version in this case,
 152	 * just modify the code directly.
 153	 */
 154	if (addr == MCOUNT_ADDR)
 155		return ftrace_modify_code_direct(rec->ip, old, new);
 156
 157	ftrace_expected = NULL;
 158
 159	/* Normal cases use add_brk_on_nop */
 
 160	WARN_ONCE(1, "invalid use of ftrace_make_nop");
 161	return -EINVAL;
 162}
 163
 164int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 165{
 166	unsigned const char *new, *old;
 167	unsigned long ip = rec->ip;
 
 168
 169	old = ftrace_nop_replace();
 170	new = ftrace_call_replace(ip, addr);
 171
 172	/* Should only be called when module is loaded */
 173	return ftrace_modify_code_direct(rec->ip, old, new);
 174}
 175
 176/*
 177 * The modifying_ftrace_code is used to tell the breakpoint
 178 * handler to call ftrace_int3_handler(). If it fails to
 179 * call this handler for a breakpoint added by ftrace, then
 180 * the kernel may crash.
 181 *
 182 * As atomic_writes on x86 do not need a barrier, we do not
 183 * need to add smp_mb()s for this to work. It is also considered
 184 * that we can not read the modifying_ftrace_code before
 185 * executing the breakpoint. That would be quite remarkable if
 186 * it could do that. Here's the flow that is required:
 187 *
 188 *   CPU-0                          CPU-1
 189 *
 190 * atomic_inc(mfc);
 191 * write int3s
 192 *				<trap-int3> // implicit (r)mb
 193 *				if (atomic_read(mfc))
 194 *					call ftrace_int3_handler()
 195 *
 196 * Then when we are finished:
 197 *
 198 * atomic_dec(mfc);
 199 *
 200 * If we hit a breakpoint that was not set by ftrace, it does not
 201 * matter if ftrace_int3_handler() is called or not. It will
 202 * simply be ignored. But it is crucial that a ftrace nop/caller
 203 * breakpoint is handled. No other user should ever place a
 204 * breakpoint on an ftrace nop/caller location. It must only
 205 * be done by this code.
 206 */
 207atomic_t modifying_ftrace_code __read_mostly;
 208
 209static int
 210ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
 211		   unsigned const char *new_code);
 212
 213/*
 214 * Should never be called:
 215 *  As it is only called by __ftrace_replace_code() which is called by
 216 *  ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
 217 *  which is called to turn mcount into nops or nops into function calls
 218 *  but not to convert a function from not using regs to one that uses
 219 *  regs, which ftrace_modify_call() is for.
 220 */
 221int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
 222				 unsigned long addr)
 223{
 224	WARN_ON(1);
 225	ftrace_expected = NULL;
 226	return -EINVAL;
 227}
 228
 229static unsigned long ftrace_update_func;
 230
 231static int update_ftrace_func(unsigned long ip, void *new)
 232{
 233	unsigned char old[MCOUNT_INSN_SIZE];
 234	int ret;
 235
 236	memcpy(old, (void *)ip, MCOUNT_INSN_SIZE);
 237
 238	ftrace_update_func = ip;
 239	/* Make sure the breakpoints see the ftrace_update_func update */
 240	smp_wmb();
 241
 242	/* See comment above by declaration of modifying_ftrace_code */
 243	atomic_inc(&modifying_ftrace_code);
 244
 245	ret = ftrace_modify_code(ip, old, new);
 246
 247	atomic_dec(&modifying_ftrace_code);
 248
 249	return ret;
 250}
 251
 252int ftrace_update_ftrace_func(ftrace_func_t func)
 253{
 254	unsigned long ip = (unsigned long)(&ftrace_call);
 255	unsigned char *new;
 256	int ret;
 257
 
 258	new = ftrace_call_replace(ip, (unsigned long)func);
 259	ret = update_ftrace_func(ip, new);
 260
 261	/* Also update the regs callback function */
 262	if (!ret) {
 263		ip = (unsigned long)(&ftrace_regs_call);
 264		new = ftrace_call_replace(ip, (unsigned long)func);
 265		ret = update_ftrace_func(ip, new);
 266	}
 267
 268	return ret;
 269}
 270
 271static int is_ftrace_caller(unsigned long ip)
 272{
 273	if (ip == ftrace_update_func)
 274		return 1;
 275
 276	return 0;
 277}
 278
 279/*
 280 * A breakpoint was added to the code address we are about to
 281 * modify, and this is the handle that will just skip over it.
 282 * We are either changing a nop into a trace call, or a trace
 283 * call to a nop. While the change is taking place, we treat
 284 * it just like it was a nop.
 285 */
 286int ftrace_int3_handler(struct pt_regs *regs)
 287{
 288	unsigned long ip;
 289
 290	if (WARN_ON_ONCE(!regs))
 291		return 0;
 292
 293	ip = regs->ip - 1;
 294	if (!ftrace_location(ip) && !is_ftrace_caller(ip))
 295		return 0;
 296
 297	regs->ip += MCOUNT_INSN_SIZE - 1;
 298
 299	return 1;
 300}
 301
 302static int ftrace_write(unsigned long ip, const char *val, int size)
 303{
 304	ip = text_ip_addr(ip);
 305
 306	if (probe_kernel_write((void *)ip, val, size))
 307		return -EPERM;
 308
 309	return 0;
 310}
 311
 312static int add_break(unsigned long ip, const char *old)
 313{
 314	unsigned char replaced[MCOUNT_INSN_SIZE];
 315	unsigned char brk = BREAKPOINT_INSTRUCTION;
 316
 317	if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
 318		return -EFAULT;
 319
 320	ftrace_expected = old;
 321
 322	/* Make sure it is what we expect it to be */
 323	if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
 324		return -EINVAL;
 325
 326	return ftrace_write(ip, &brk, 1);
 327}
 328
 329static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
 330{
 331	unsigned const char *old;
 332	unsigned long ip = rec->ip;
 333
 334	old = ftrace_call_replace(ip, addr);
 335
 336	return add_break(rec->ip, old);
 337}
 338
 339
 340static int add_brk_on_nop(struct dyn_ftrace *rec)
 341{
 342	unsigned const char *old;
 343
 344	old = ftrace_nop_replace();
 345
 346	return add_break(rec->ip, old);
 347}
 348
 349static int add_breakpoints(struct dyn_ftrace *rec, int enable)
 350{
 351	unsigned long ftrace_addr;
 352	int ret;
 353
 354	ftrace_addr = ftrace_get_addr_curr(rec);
 355
 356	ret = ftrace_test_record(rec, enable);
 357
 358	switch (ret) {
 359	case FTRACE_UPDATE_IGNORE:
 360		return 0;
 361
 362	case FTRACE_UPDATE_MAKE_CALL:
 363		/* converting nop to call */
 364		return add_brk_on_nop(rec);
 365
 366	case FTRACE_UPDATE_MODIFY_CALL:
 367	case FTRACE_UPDATE_MAKE_NOP:
 368		/* converting a call to a nop */
 369		return add_brk_on_call(rec, ftrace_addr);
 370	}
 371	return 0;
 372}
 373
 374/*
 375 * On error, we need to remove breakpoints. This needs to
 376 * be done caefully. If the address does not currently have a
 377 * breakpoint, we know we are done. Otherwise, we look at the
 378 * remaining 4 bytes of the instruction. If it matches a nop
 379 * we replace the breakpoint with the nop. Otherwise we replace
 380 * it with the call instruction.
 381 */
 382static int remove_breakpoint(struct dyn_ftrace *rec)
 383{
 384	unsigned char ins[MCOUNT_INSN_SIZE];
 385	unsigned char brk = BREAKPOINT_INSTRUCTION;
 386	const unsigned char *nop;
 387	unsigned long ftrace_addr;
 388	unsigned long ip = rec->ip;
 389
 390	/* If we fail the read, just give up */
 391	if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
 392		return -EFAULT;
 393
 394	/* If this does not have a breakpoint, we are done */
 395	if (ins[0] != brk)
 396		return 0;
 397
 398	nop = ftrace_nop_replace();
 399
 400	/*
 401	 * If the last 4 bytes of the instruction do not match
 402	 * a nop, then we assume that this is a call to ftrace_addr.
 403	 */
 404	if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
 405		/*
 406		 * For extra paranoidism, we check if the breakpoint is on
 407		 * a call that would actually jump to the ftrace_addr.
 408		 * If not, don't touch the breakpoint, we make just create
 409		 * a disaster.
 410		 */
 411		ftrace_addr = ftrace_get_addr_new(rec);
 412		nop = ftrace_call_replace(ip, ftrace_addr);
 413
 414		if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
 415			goto update;
 416
 417		/* Check both ftrace_addr and ftrace_old_addr */
 418		ftrace_addr = ftrace_get_addr_curr(rec);
 419		nop = ftrace_call_replace(ip, ftrace_addr);
 420
 421		ftrace_expected = nop;
 422
 423		if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
 424			return -EINVAL;
 425	}
 426
 427 update:
 428	return ftrace_write(ip, nop, 1);
 429}
 430
 431static int add_update_code(unsigned long ip, unsigned const char *new)
 432{
 433	/* skip breakpoint */
 434	ip++;
 435	new++;
 436	return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1);
 437}
 438
 439static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
 440{
 441	unsigned long ip = rec->ip;
 442	unsigned const char *new;
 443
 444	new = ftrace_call_replace(ip, addr);
 445	return add_update_code(ip, new);
 446}
 447
 448static int add_update_nop(struct dyn_ftrace *rec)
 449{
 450	unsigned long ip = rec->ip;
 451	unsigned const char *new;
 452
 453	new = ftrace_nop_replace();
 454	return add_update_code(ip, new);
 455}
 456
 457static int add_update(struct dyn_ftrace *rec, int enable)
 458{
 459	unsigned long ftrace_addr;
 460	int ret;
 461
 462	ret = ftrace_test_record(rec, enable);
 463
 464	ftrace_addr  = ftrace_get_addr_new(rec);
 465
 466	switch (ret) {
 467	case FTRACE_UPDATE_IGNORE:
 468		return 0;
 469
 470	case FTRACE_UPDATE_MODIFY_CALL:
 471	case FTRACE_UPDATE_MAKE_CALL:
 472		/* converting nop to call */
 473		return add_update_call(rec, ftrace_addr);
 474
 475	case FTRACE_UPDATE_MAKE_NOP:
 476		/* converting a call to a nop */
 477		return add_update_nop(rec);
 478	}
 479
 480	return 0;
 481}
 482
 483static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
 484{
 485	unsigned long ip = rec->ip;
 486	unsigned const char *new;
 487
 488	new = ftrace_call_replace(ip, addr);
 489
 490	return ftrace_write(ip, new, 1);
 491}
 492
 493static int finish_update_nop(struct dyn_ftrace *rec)
 494{
 495	unsigned long ip = rec->ip;
 496	unsigned const char *new;
 497
 498	new = ftrace_nop_replace();
 499
 500	return ftrace_write(ip, new, 1);
 501}
 502
 503static int finish_update(struct dyn_ftrace *rec, int enable)
 504{
 505	unsigned long ftrace_addr;
 506	int ret;
 507
 508	ret = ftrace_update_record(rec, enable);
 509
 510	ftrace_addr = ftrace_get_addr_new(rec);
 511
 512	switch (ret) {
 513	case FTRACE_UPDATE_IGNORE:
 514		return 0;
 515
 516	case FTRACE_UPDATE_MODIFY_CALL:
 517	case FTRACE_UPDATE_MAKE_CALL:
 518		/* converting nop to call */
 519		return finish_update_call(rec, ftrace_addr);
 520
 521	case FTRACE_UPDATE_MAKE_NOP:
 522		/* converting a call to a nop */
 523		return finish_update_nop(rec);
 524	}
 525
 526	return 0;
 527}
 528
 529static void do_sync_core(void *data)
 530{
 531	sync_core();
 532}
 533
 534static void run_sync(void)
 535{
 536	int enable_irqs = irqs_disabled();
 537
 538	/* We may be called with interrupts disbled (on bootup). */
 539	if (enable_irqs)
 540		local_irq_enable();
 541	on_each_cpu(do_sync_core, NULL, 1);
 542	if (enable_irqs)
 543		local_irq_disable();
 544}
 545
 546void ftrace_replace_code(int enable)
 547{
 548	struct ftrace_rec_iter *iter;
 549	struct dyn_ftrace *rec;
 550	const char *report = "adding breakpoints";
 551	int count = 0;
 552	int ret;
 553
 554	for_ftrace_rec_iter(iter) {
 555		rec = ftrace_rec_iter_record(iter);
 556
 557		ret = add_breakpoints(rec, enable);
 558		if (ret)
 559			goto remove_breakpoints;
 560		count++;
 561	}
 562
 563	run_sync();
 564
 565	report = "updating code";
 566	count = 0;
 567
 568	for_ftrace_rec_iter(iter) {
 569		rec = ftrace_rec_iter_record(iter);
 570
 571		ret = add_update(rec, enable);
 572		if (ret)
 573			goto remove_breakpoints;
 574		count++;
 
 
 
 
 575	}
 576
 577	run_sync();
 578
 579	report = "removing breakpoints";
 580	count = 0;
 581
 582	for_ftrace_rec_iter(iter) {
 583		rec = ftrace_rec_iter_record(iter);
 584
 585		ret = finish_update(rec, enable);
 586		if (ret)
 587			goto remove_breakpoints;
 588		count++;
 589	}
 590
 591	run_sync();
 592
 593	return;
 
 
 
 
 
 594
 595 remove_breakpoints:
 596	pr_warn("Failed on %s (%d):\n", report, count);
 597	ftrace_bug(ret, rec);
 598	for_ftrace_rec_iter(iter) {
 599		rec = ftrace_rec_iter_record(iter);
 600		/*
 601		 * Breakpoints are handled only when this function is in
 602		 * progress. The system could not work with them.
 603		 */
 604		if (remove_breakpoint(rec))
 605			BUG();
 606	}
 607	run_sync();
 608}
 609
 610static int
 611ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
 612		   unsigned const char *new_code)
 613{
 614	int ret;
 615
 616	ret = add_break(ip, old_code);
 617	if (ret)
 618		goto out;
 619
 620	run_sync();
 621
 622	ret = add_update_code(ip, new_code);
 623	if (ret)
 624		goto fail_update;
 625
 626	run_sync();
 627
 628	ret = ftrace_write(ip, new_code, 1);
 629	/*
 630	 * The breakpoint is handled only when this function is in progress.
 631	 * The system could not work if we could not remove it.
 632	 */
 633	BUG_ON(ret);
 634 out:
 635	run_sync();
 636	return ret;
 637
 638 fail_update:
 639	/* Also here the system could not work with the breakpoint */
 640	if (ftrace_write(ip, old_code, 1))
 641		BUG();
 642	goto out;
 643}
 644
 645void arch_ftrace_update_code(int command)
 646{
 647	/* See comment above by declaration of modifying_ftrace_code */
 648	atomic_inc(&modifying_ftrace_code);
 649
 650	ftrace_modify_all_code(command);
 651
 652	atomic_dec(&modifying_ftrace_code);
 653}
 654
 655int __init ftrace_dyn_arch_init(void)
 656{
 657	return 0;
 658}
 659
 660#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
 661static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
 662{
 663	static union ftrace_code_union calc;
 664
 665	/* Jmp not a call (ignore the .e8) */
 666	calc.e8		= 0xe9;
 667	calc.offset	= ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
 668
 669	/*
 670	 * ftrace external locks synchronize the access to the static variable.
 671	 */
 672	return calc.code;
 673}
 674#endif
 675
 676/* Currently only x86_64 supports dynamic trampolines */
 677#ifdef CONFIG_X86_64
 678
 679#ifdef CONFIG_MODULES
 680#include <linux/moduleloader.h>
 681/* Module allocation simplifies allocating memory for code */
 682static inline void *alloc_tramp(unsigned long size)
 683{
 684	return module_alloc(size);
 685}
 686static inline void tramp_free(void *tramp)
 687{
 688	module_memfree(tramp);
 689}
 690#else
 691/* Trampolines can only be created if modules are supported */
 692static inline void *alloc_tramp(unsigned long size)
 693{
 694	return NULL;
 695}
 696static inline void tramp_free(void *tramp) { }
 697#endif
 698
 699/* Defined as markers to the end of the ftrace default trampolines */
 700extern void ftrace_regs_caller_end(void);
 701extern void ftrace_epilogue(void);
 702extern void ftrace_caller_op_ptr(void);
 703extern void ftrace_regs_caller_op_ptr(void);
 
 704
 705/* movq function_trace_op(%rip), %rdx */
 706/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
 707#define OP_REF_SIZE	7
 708
 709/*
 710 * The ftrace_ops is passed to the function callback. Since the
 711 * trampoline only services a single ftrace_ops, we can pass in
 712 * that ops directly.
 713 *
 714 * The ftrace_op_code_union is used to create a pointer to the
 715 * ftrace_ops that will be passed to the callback function.
 716 */
 717union ftrace_op_code_union {
 718	char code[OP_REF_SIZE];
 719	struct {
 720		char op[3];
 721		int offset;
 722	} __attribute__((packed));
 723};
 724
 
 
 
 725static unsigned long
 726create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
 727{
 728	unsigned const char *jmp;
 729	unsigned long start_offset;
 730	unsigned long end_offset;
 731	unsigned long op_offset;
 
 
 732	unsigned long offset;
 
 733	unsigned long size;
 734	unsigned long ip;
 735	unsigned long *ptr;
 736	void *trampoline;
 
 737	/* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
 738	unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
 
 739	union ftrace_op_code_union op_ptr;
 740	int ret;
 741
 742	if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
 743		start_offset = (unsigned long)ftrace_regs_caller;
 744		end_offset = (unsigned long)ftrace_regs_caller_end;
 745		op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
 
 
 746	} else {
 747		start_offset = (unsigned long)ftrace_caller;
 748		end_offset = (unsigned long)ftrace_epilogue;
 749		op_offset = (unsigned long)ftrace_caller_op_ptr;
 
 
 750	}
 751
 752	size = end_offset - start_offset;
 753
 754	/*
 755	 * Allocate enough size to store the ftrace_caller code,
 756	 * the jmp to ftrace_epilogue, as well as the address of
 757	 * the ftrace_ops this trampoline is used for.
 758	 */
 759	trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
 760	if (!trampoline)
 761		return 0;
 762
 763	*tramp_size = size + MCOUNT_INSN_SIZE + sizeof(void *);
 
 764
 765	/* Copy ftrace_caller onto the trampoline memory */
 766	ret = probe_kernel_read(trampoline, (void *)start_offset, size);
 767	if (WARN_ON(ret < 0)) {
 768		tramp_free(trampoline);
 769		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 770	}
 771
 772	ip = (unsigned long)trampoline + size;
 773
 774	/* The trampoline ends with a jmp to ftrace_epilogue */
 775	jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_epilogue);
 776	memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
 777
 778	/*
 779	 * The address of the ftrace_ops that is used for this trampoline
 780	 * is stored at the end of the trampoline. This will be used to
 781	 * load the third parameter for the callback. Basically, that
 782	 * location at the end of the trampoline takes the place of
 783	 * the global function_trace_op variable.
 784	 */
 785
 786	ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
 787	*ptr = (unsigned long)ops;
 788
 789	op_offset -= start_offset;
 790	memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
 791
 792	/* Are we pointing to the reference? */
 793	if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
 794		tramp_free(trampoline);
 795		return 0;
 796	}
 797
 798	/* Load the contents of ptr into the callback parameter */
 799	offset = (unsigned long)ptr;
 800	offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
 801
 802	op_ptr.offset = offset;
 803
 804	/* put in the new offset to the ftrace_ops */
 805	memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 806
 807	/* ALLOC_TRAMP flags lets us know we created it */
 808	ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
 809
 
 810	return (unsigned long)trampoline;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 811}
 812
 813static unsigned long calc_trampoline_call_offset(bool save_regs)
 814{
 815	unsigned long start_offset;
 816	unsigned long call_offset;
 817
 818	if (save_regs) {
 819		start_offset = (unsigned long)ftrace_regs_caller;
 820		call_offset = (unsigned long)ftrace_regs_call;
 821	} else {
 822		start_offset = (unsigned long)ftrace_caller;
 823		call_offset = (unsigned long)ftrace_call;
 824	}
 825
 826	return call_offset - start_offset;
 827}
 828
 829void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
 830{
 831	ftrace_func_t func;
 832	unsigned char *new;
 833	unsigned long offset;
 834	unsigned long ip;
 835	unsigned int size;
 836	int ret;
 837
 838	if (ops->trampoline) {
 839		/*
 840		 * The ftrace_ops caller may set up its own trampoline.
 841		 * In such a case, this code must not modify it.
 842		 */
 843		if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
 844			return;
 845	} else {
 846		ops->trampoline = create_trampoline(ops, &size);
 847		if (!ops->trampoline)
 848			return;
 849		ops->trampoline_size = size;
 
 850	}
 851
 
 
 
 
 
 
 
 852	offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
 853	ip = ops->trampoline + offset;
 854
 855	func = ftrace_ops_get_func(ops);
 856
 
 857	/* Do a safe modify in case the trampoline is executing */
 858	new = ftrace_call_replace(ip, (unsigned long)func);
 859	ret = update_ftrace_func(ip, new);
 860
 861	/* The update should never fail */
 862	WARN_ON(ret);
 863}
 864
 865/* Return the address of the function the trampoline calls */
 866static void *addr_from_call(void *ptr)
 867{
 868	union ftrace_code_union calc;
 869	int ret;
 870
 871	ret = probe_kernel_read(&calc, ptr, MCOUNT_INSN_SIZE);
 872	if (WARN_ON_ONCE(ret < 0))
 873		return NULL;
 874
 875	/* Make sure this is a call */
 876	if (WARN_ON_ONCE(calc.e8 != 0xe8)) {
 877		pr_warn("Expected e8, got %x\n", calc.e8);
 878		return NULL;
 879	}
 880
 881	return ptr + MCOUNT_INSN_SIZE + calc.offset;
 882}
 883
 884void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
 885			   unsigned long frame_pointer);
 886
 887/*
 888 * If the ops->trampoline was not allocated, then it probably
 889 * has a static trampoline func, or is the ftrace caller itself.
 890 */
 891static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
 892{
 893	unsigned long offset;
 894	bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
 895	void *ptr;
 896
 897	if (ops && ops->trampoline) {
 898#ifdef CONFIG_FUNCTION_GRAPH_TRACER
 
 899		/*
 900		 * We only know about function graph tracer setting as static
 901		 * trampoline.
 902		 */
 903		if (ops->trampoline == FTRACE_GRAPH_ADDR)
 904			return (void *)prepare_ftrace_return;
 905#endif
 906		return NULL;
 907	}
 908
 909	offset = calc_trampoline_call_offset(save_regs);
 910
 911	if (save_regs)
 912		ptr = (void *)FTRACE_REGS_ADDR + offset;
 913	else
 914		ptr = (void *)FTRACE_ADDR + offset;
 915
 916	return addr_from_call(ptr);
 917}
 918
 919void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
 920{
 921	unsigned long offset;
 922
 923	/* If we didn't allocate this trampoline, consider it static */
 924	if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
 925		return static_tramp_func(ops, rec);
 926
 927	offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
 928	return addr_from_call((void *)ops->trampoline + offset);
 929}
 930
 931void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
 932{
 933	if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
 934		return;
 935
 936	tramp_free((void *)ops->trampoline);
 937	ops->trampoline = 0;
 938}
 939
 940#endif /* CONFIG_X86_64 */
 941#endif /* CONFIG_DYNAMIC_FTRACE */
 942
 943#ifdef CONFIG_FUNCTION_GRAPH_TRACER
 944
 945#ifdef CONFIG_DYNAMIC_FTRACE
 946extern void ftrace_graph_call(void);
 
 
 
 
 947
 948static int ftrace_mod_jmp(unsigned long ip, void *func)
 949{
 950	unsigned char *new;
 951
 952	new = ftrace_jmp_replace(ip, (unsigned long)func);
 953
 954	return update_ftrace_func(ip, new);
 955}
 956
 957int ftrace_enable_ftrace_graph_caller(void)
 958{
 959	unsigned long ip = (unsigned long)(&ftrace_graph_call);
 960
 961	return ftrace_mod_jmp(ip, &ftrace_graph_caller);
 962}
 963
 964int ftrace_disable_ftrace_graph_caller(void)
 965{
 966	unsigned long ip = (unsigned long)(&ftrace_graph_call);
 967
 968	return ftrace_mod_jmp(ip, &ftrace_stub);
 969}
 970
 971#endif /* !CONFIG_DYNAMIC_FTRACE */
 972
 973/*
 974 * Hook the return address and push it in the stack of return addrs
 975 * in current thread info.
 976 */
 977void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
 978			   unsigned long frame_pointer)
 979{
 980	unsigned long old;
 981	int faulted;
 982	struct ftrace_graph_ent trace;
 983	unsigned long return_hooker = (unsigned long)
 984				&return_to_handler;
 
 
 
 
 
 
 
 
 
 985
 986	if (unlikely(ftrace_graph_is_dead()))
 987		return;
 988
 989	if (unlikely(atomic_read(&current->tracing_graph_pause)))
 990		return;
 991
 992	/*
 993	 * Protect against fault, even if it shouldn't
 994	 * happen. This tool is too much intrusive to
 995	 * ignore such a protection.
 996	 */
 997	asm volatile(
 998		"1: " _ASM_MOV " (%[parent]), %[old]\n"
 999		"2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
1000		"   movl $0, %[faulted]\n"
1001		"3:\n"
1002
1003		".section .fixup, \"ax\"\n"
1004		"4: movl $1, %[faulted]\n"
1005		"   jmp 3b\n"
1006		".previous\n"
1007
1008		_ASM_EXTABLE(1b, 4b)
1009		_ASM_EXTABLE(2b, 4b)
1010
1011		: [old] "=&r" (old), [faulted] "=r" (faulted)
1012		: [parent] "r" (parent), [return_hooker] "r" (return_hooker)
1013		: "memory"
1014	);
1015
1016	if (unlikely(faulted)) {
1017		ftrace_graph_stop();
1018		WARN_ON(1);
1019		return;
1020	}
1021
1022	trace.func = self_addr;
1023	trace.depth = current->curr_ret_stack + 1;
1024
1025	/* Only trace if the calling function expects to */
1026	if (!ftrace_graph_entry(&trace)) {
1027		*parent = old;
1028		return;
1029	}
 
1030
1031	if (ftrace_push_return_trace(old, self_addr, &trace.depth,
1032		    frame_pointer) == -EBUSY) {
1033		*parent = old;
1034		return;
1035	}
1036}
 
 
1037#endif /* CONFIG_FUNCTION_GRAPH_TRACER */