Linux Audio

Check our new training course

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 */
v3.1
 
  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/ftrace.h>
 28#include <asm/nops.h>
 29#include <asm/nmi.h>
 30
 31
 32#ifdef CONFIG_DYNAMIC_FTRACE
 33
 34/*
 35 * modifying_code is set to notify NMIs that they need to use
 36 * memory barriers when entering or exiting. But we don't want
 37 * to burden NMIs with unnecessary memory barriers when code
 38 * modification is not being done (which is most of the time).
 39 *
 40 * A mutex is already held when ftrace_arch_code_modify_prepare
 41 * and post_process are called. No locks need to be taken here.
 42 *
 43 * Stop machine will make sure currently running NMIs are done
 44 * and new NMIs will see the updated variable before we need
 45 * to worry about NMIs doing memory barriers.
 46 */
 47static int modifying_code __read_mostly;
 48static DEFINE_PER_CPU(int, save_modifying_code);
 49
 50int ftrace_arch_code_modify_prepare(void)
 
 51{
 52	set_kernel_text_rw();
 53	set_all_modules_text_rw();
 54	modifying_code = 1;
 55	return 0;
 
 
 
 56}
 57
 58int ftrace_arch_code_modify_post_process(void)
 
 59{
 60	modifying_code = 0;
 61	set_all_modules_text_ro();
 62	set_kernel_text_ro();
 63	return 0;
 
 
 
 
 64}
 65
 66union ftrace_code_union {
 67	char code[MCOUNT_INSN_SIZE];
 68	struct {
 69		char e8;
 70		int offset;
 71	} __attribute__((packed));
 72};
 73
 74static int ftrace_calc_offset(long ip, long addr)
 75{
 76	return (int)(addr - ip);
 
 
 
 
 77}
 78
 79static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
 80{
 81	static union ftrace_code_union calc;
 82
 83	calc.e8		= 0xe8;
 84	calc.offset	= ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
 85
 86	/*
 87	 * No locking needed, this must be called via kstop_machine
 88	 * which in essence is like running on a uniprocessor machine.
 
 
 
 89	 */
 90	return calc.code;
 
 
 
 
 
 
 
 
 
 
 
 
 
 91}
 92
 93/*
 94 * Modifying code must take extra care. On an SMP machine, if
 95 * the code being modified is also being executed on another CPU
 96 * that CPU will have undefined results and possibly take a GPF.
 97 * We use kstop_machine to stop other CPUS from exectuing code.
 98 * But this does not stop NMIs from happening. We still need
 99 * to protect against that. We separate out the modification of
100 * the code to take care of this.
101 *
102 * Two buffers are added: An IP buffer and a "code" buffer.
103 *
104 * 1) Put the instruction pointer into the IP buffer
105 *    and the new code into the "code" buffer.
106 * 2) Wait for any running NMIs to finish and set a flag that says
107 *    we are modifying code, it is done in an atomic operation.
108 * 3) Write the code
109 * 4) clear the flag.
110 * 5) Wait for any running NMIs to finish.
111 *
112 * If an NMI is executed, the first thing it does is to call
113 * "ftrace_nmi_enter". This will check if the flag is set to write
114 * and if it is, it will write what is in the IP and "code" buffers.
115 *
116 * The trick is, it does not matter if everyone is writing the same
117 * content to the code location. Also, if a CPU is executing code
118 * it is OK to write to that code location if the contents being written
119 * are the same as what exists.
120 */
 
 
 
 
 
 
 
121
122#define MOD_CODE_WRITE_FLAG (1 << 31)	/* set when NMI should do the write */
123static atomic_t nmi_running = ATOMIC_INIT(0);
124static int mod_code_status;		/* holds return value of text write */
125static void *mod_code_ip;		/* holds the IP to write to */
126static const void *mod_code_newcode;	/* holds the text to write to the IP */
 
 
 
 
 
127
128static unsigned nmi_wait_count;
129static atomic_t nmi_update_count = ATOMIC_INIT(0);
 
 
130
131int ftrace_arch_read_dyn_info(char *buf, int size)
132{
133	int r;
 
 
 
 
 
 
 
 
 
 
134
135	r = snprintf(buf, size, "%u %u",
136		     nmi_wait_count,
137		     atomic_read(&nmi_update_count));
138	return r;
 
 
139}
140
141static void clear_mod_flag(void)
142{
143	int old = atomic_read(&nmi_running);
 
144
145	for (;;) {
146		int new = old & ~MOD_CODE_WRITE_FLAG;
147
148		if (old == new)
149			break;
 
150
151		old = atomic_cmpxchg(&nmi_running, old, new);
152	}
 
 
 
 
 
 
 
 
 
 
 
153}
154
155static void ftrace_mod_code(void)
156{
157	/*
158	 * Yes, more than one CPU process can be writing to mod_code_status.
159	 *    (and the code itself)
160	 * But if one were to fail, then they all should, and if one were
161	 * to succeed, then they all should.
162	 */
163	mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
164					     MCOUNT_INSN_SIZE);
 
 
165
166	/* if we fail, then kill any new writers */
167	if (mod_code_status)
168		clear_mod_flag();
169}
170
171void ftrace_nmi_enter(void)
172{
173	__this_cpu_write(save_modifying_code, modifying_code);
 
 
 
 
 
 
 
 
 
 
 
174
175	if (!__this_cpu_read(save_modifying_code))
176		return;
 
 
 
 
 
 
 
177
178	if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) {
179		smp_rmb();
180		ftrace_mod_code();
181		atomic_inc(&nmi_update_count);
 
 
 
182	}
183	/* Must have previous changes seen before executions */
184	smp_mb();
185}
186
187void ftrace_nmi_exit(void)
188{
189	if (!__this_cpu_read(save_modifying_code))
190		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
191
192	/* Finish all executions before clearing nmi_running */
193	smp_mb();
194	atomic_dec(&nmi_running);
 
195}
196
197static void wait_for_nmi_and_set_mod_flag(void)
198{
199	if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG))
200		return;
201
202	do {
203		cpu_relax();
204	} while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG));
205
206	nmi_wait_count++;
 
 
 
 
 
 
207}
208
209static void wait_for_nmi(void)
210{
211	if (!atomic_read(&nmi_running))
212		return;
 
 
 
 
 
 
213
214	do {
215		cpu_relax();
216	} while (atomic_read(&nmi_running));
 
 
 
 
 
 
 
 
 
 
 
 
217
218	nmi_wait_count++;
219}
220
221static inline int
222within(unsigned long addr, unsigned long start, unsigned long end)
223{
224	return addr >= start && addr < end;
225}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
226
227static int
228do_ftrace_mod_code(unsigned long ip, const void *new_code)
229{
230	/*
231	 * On x86_64, kernel text mappings are mapped read-only with
232	 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
233	 * of the kernel text mapping to modify the kernel text.
234	 *
235	 * For 32bit kernels, these mappings are same and we can use
236	 * kernel identity mapping to modify code.
237	 */
238	if (within(ip, (unsigned long)_text, (unsigned long)_etext))
239		ip = (unsigned long)__va(__pa(ip));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
240
241	mod_code_ip = (void *)ip;
242	mod_code_newcode = new_code;
 
 
 
 
 
243
244	/* The buffers need to be visible before we let NMIs write them */
245	smp_mb();
246
247	wait_for_nmi_and_set_mod_flag();
 
248
249	/* Make sure all running NMIs have finished before we write the code */
250	smp_mb();
 
251
252	ftrace_mod_code();
 
 
253
254	/* Make sure the write happens before clearing the bit */
255	smp_mb();
256
257	clear_mod_flag();
258	wait_for_nmi();
259
260	return mod_code_status;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
261}
262
263static const unsigned char *ftrace_nop_replace(void)
264{
265	return ideal_nops[NOP_ATOMIC5];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
266}
267
268static int
269ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
270		   unsigned const char *new_code)
271{
272	unsigned char replaced[MCOUNT_INSN_SIZE];
 
 
 
 
 
 
 
 
 
 
 
 
273
274	/*
275	 * Note: Due to modules and __init, code can
276	 *  disappear and change, we need to protect against faulting
277	 *  as well as code changing. We do this by using the
278	 *  probe_kernel_* functions.
279	 *
280	 * No real locking needed, this code is run through
281	 * kstop_machine, or before SMP starts.
282	 */
 
 
283
284	/* read the text we want to modify */
285	if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
286		return -EFAULT;
287
288	/* Make sure it is what we expect it to be */
289	if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
290		return -EINVAL;
291
292	/* replace the text with the new text */
293	if (do_ftrace_mod_code(ip, new_code))
294		return -EPERM;
295
296	sync_core();
297
298	return 0;
299}
300
301int ftrace_make_nop(struct module *mod,
302		    struct dyn_ftrace *rec, unsigned long addr)
303{
304	unsigned const char *new, *old;
305	unsigned long ip = rec->ip;
306
307	old = ftrace_call_replace(ip, addr);
308	new = ftrace_nop_replace();
 
 
 
 
 
 
 
309
310	return ftrace_modify_code(rec->ip, old, new);
311}
312
313int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 
 
 
 
314{
315	unsigned const char *new, *old;
316	unsigned long ip = rec->ip;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
317
318	old = ftrace_nop_replace();
319	new = ftrace_call_replace(ip, addr);
 
 
320
321	return ftrace_modify_code(rec->ip, old, new);
322}
323
324int ftrace_update_ftrace_func(ftrace_func_t func)
325{
326	unsigned long ip = (unsigned long)(&ftrace_call);
327	unsigned char old[MCOUNT_INSN_SIZE], *new;
328	int ret;
329
330	memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
331	new = ftrace_call_replace(ip, (unsigned long)func);
332	ret = ftrace_modify_code(ip, old, new);
333
334	return ret;
 
335}
336
337int __init ftrace_dyn_arch_init(void *data)
338{
339	/* The return code is retured via data */
340	*(unsigned long *)data = 0;
341
342	return 0;
 
343}
344#endif
 
 
345
346#ifdef CONFIG_FUNCTION_GRAPH_TRACER
347
348#ifdef CONFIG_DYNAMIC_FTRACE
349extern void ftrace_graph_call(void);
 
 
 
 
350
351static int ftrace_mod_jmp(unsigned long ip,
352			  int old_offset, int new_offset)
353{
354	unsigned char code[MCOUNT_INSN_SIZE];
355
356	if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
357		return -EFAULT;
358
359	if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
360		return -EINVAL;
361
362	*(int *)(&code[1]) = new_offset;
363
364	if (do_ftrace_mod_code(ip, &code))
365		return -EPERM;
366
 
 
367	return 0;
368}
369
370int ftrace_enable_ftrace_graph_caller(void)
371{
372	unsigned long ip = (unsigned long)(&ftrace_graph_call);
373	int old_offset, new_offset;
374
375	old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
376	new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
377
378	return ftrace_mod_jmp(ip, old_offset, new_offset);
379}
380
381int ftrace_disable_ftrace_graph_caller(void)
382{
383	unsigned long ip = (unsigned long)(&ftrace_graph_call);
384	int old_offset, new_offset;
385
386	old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
387	new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
388
389	return ftrace_mod_jmp(ip, old_offset, new_offset);
390}
391
392#endif /* !CONFIG_DYNAMIC_FTRACE */
393
394/*
395 * Hook the return address and push it in the stack of return addrs
396 * in current thread info.
397 */
398void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
399			   unsigned long frame_pointer)
400{
401	unsigned long old;
402	int faulted;
403	struct ftrace_graph_ent trace;
404	unsigned long return_hooker = (unsigned long)
405				&return_to_handler;
 
 
 
 
 
 
 
 
 
 
 
 
406
407	if (unlikely(atomic_read(&current->tracing_graph_pause)))
408		return;
409
410	/*
411	 * Protect against fault, even if it shouldn't
412	 * happen. This tool is too much intrusive to
413	 * ignore such a protection.
414	 */
415	asm volatile(
416		"1: " _ASM_MOV " (%[parent]), %[old]\n"
417		"2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
418		"   movl $0, %[faulted]\n"
419		"3:\n"
420
421		".section .fixup, \"ax\"\n"
422		"4: movl $1, %[faulted]\n"
423		"   jmp 3b\n"
424		".previous\n"
425
426		_ASM_EXTABLE(1b, 4b)
427		_ASM_EXTABLE(2b, 4b)
428
429		: [old] "=&r" (old), [faulted] "=r" (faulted)
430		: [parent] "r" (parent), [return_hooker] "r" (return_hooker)
431		: "memory"
432	);
433
434	if (unlikely(faulted)) {
435		ftrace_graph_stop();
436		WARN_ON(1);
437		return;
438	}
439
440	trace.func = self_addr;
441	trace.depth = current->curr_ret_stack + 1;
442
443	/* Only trace if the calling function expects to */
444	if (!ftrace_graph_entry(&trace)) {
445		*parent = old;
446		return;
447	}
 
448
449	if (ftrace_push_return_trace(old, self_addr, &trace.depth,
450		    frame_pointer) == -EBUSY) {
451		*parent = old;
452		return;
453	}
454}
 
 
455#endif /* CONFIG_FUNCTION_GRAPH_TRACER */