Loading...
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(¤t->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 */
1/*
2 * Code for replacing ftrace calls with jumps.
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 *
6 * Thanks goes to Ingo Molnar, for suggesting the idea.
7 * Mathieu Desnoyers, for suggesting postponing the modifications.
8 * Arjan van de Ven, for keeping me straight, and explaining to me
9 * the dangers of modifying code on the run.
10 */
11
12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13
14#include <linux/spinlock.h>
15#include <linux/hardirq.h>
16#include <linux/uaccess.h>
17#include <linux/ftrace.h>
18#include <linux/percpu.h>
19#include <linux/sched.h>
20#include <linux/init.h>
21#include <linux/list.h>
22#include <linux/module.h>
23
24#include <trace/syscall.h>
25
26#include <asm/cacheflush.h>
27#include <asm/kprobes.h>
28#include <asm/ftrace.h>
29#include <asm/nops.h>
30
31#ifdef CONFIG_DYNAMIC_FTRACE
32
33int ftrace_arch_code_modify_prepare(void)
34{
35 set_kernel_text_rw();
36 set_all_modules_text_rw();
37 return 0;
38}
39
40int ftrace_arch_code_modify_post_process(void)
41{
42 set_all_modules_text_ro();
43 set_kernel_text_ro();
44 return 0;
45}
46
47union ftrace_code_union {
48 char code[MCOUNT_INSN_SIZE];
49 struct {
50 char e8;
51 int offset;
52 } __attribute__((packed));
53};
54
55static int ftrace_calc_offset(long ip, long addr)
56{
57 return (int)(addr - ip);
58}
59
60static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
61{
62 static union ftrace_code_union calc;
63
64 calc.e8 = 0xe8;
65 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
66
67 /*
68 * No locking needed, this must be called via kstop_machine
69 * which in essence is like running on a uniprocessor machine.
70 */
71 return calc.code;
72}
73
74static inline int
75within(unsigned long addr, unsigned long start, unsigned long end)
76{
77 return addr >= start && addr < end;
78}
79
80static unsigned long text_ip_addr(unsigned long ip)
81{
82 /*
83 * On x86_64, kernel text mappings are mapped read-only with
84 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
85 * of the kernel text mapping to modify the kernel text.
86 *
87 * For 32bit kernels, these mappings are same and we can use
88 * kernel identity mapping to modify code.
89 */
90 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
91 ip = (unsigned long)__va(__pa_symbol(ip));
92
93 return ip;
94}
95
96static const unsigned char *ftrace_nop_replace(void)
97{
98 return ideal_nops[NOP_ATOMIC5];
99}
100
101static int
102ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code,
103 unsigned const char *new_code)
104{
105 unsigned char replaced[MCOUNT_INSN_SIZE];
106
107 /*
108 * Note: Due to modules and __init, code can
109 * disappear and change, we need to protect against faulting
110 * as well as code changing. We do this by using the
111 * probe_kernel_* functions.
112 *
113 * No real locking needed, this code is run through
114 * kstop_machine, or before SMP starts.
115 */
116
117 /* read the text we want to modify */
118 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
119 return -EFAULT;
120
121 /* Make sure it is what we expect it to be */
122 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
123 return -EINVAL;
124
125 ip = text_ip_addr(ip);
126
127 /* replace the text with the new text */
128 if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE))
129 return -EPERM;
130
131 sync_core();
132
133 return 0;
134}
135
136int ftrace_make_nop(struct module *mod,
137 struct dyn_ftrace *rec, unsigned long addr)
138{
139 unsigned const char *new, *old;
140 unsigned long ip = rec->ip;
141
142 old = ftrace_call_replace(ip, addr);
143 new = ftrace_nop_replace();
144
145 /*
146 * On boot up, and when modules are loaded, the MCOUNT_ADDR
147 * is converted to a nop, and will never become MCOUNT_ADDR
148 * again. This code is either running before SMP (on boot up)
149 * or before the code will ever be executed (module load).
150 * We do not want to use the breakpoint version in this case,
151 * just modify the code directly.
152 */
153 if (addr == MCOUNT_ADDR)
154 return ftrace_modify_code_direct(rec->ip, old, new);
155
156 /* Normal cases use add_brk_on_nop */
157 WARN_ONCE(1, "invalid use of ftrace_make_nop");
158 return -EINVAL;
159}
160
161int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
162{
163 unsigned const char *new, *old;
164 unsigned long ip = rec->ip;
165
166 old = ftrace_nop_replace();
167 new = ftrace_call_replace(ip, addr);
168
169 /* Should only be called when module is loaded */
170 return ftrace_modify_code_direct(rec->ip, old, new);
171}
172
173/*
174 * The modifying_ftrace_code is used to tell the breakpoint
175 * handler to call ftrace_int3_handler(). If it fails to
176 * call this handler for a breakpoint added by ftrace, then
177 * the kernel may crash.
178 *
179 * As atomic_writes on x86 do not need a barrier, we do not
180 * need to add smp_mb()s for this to work. It is also considered
181 * that we can not read the modifying_ftrace_code before
182 * executing the breakpoint. That would be quite remarkable if
183 * it could do that. Here's the flow that is required:
184 *
185 * CPU-0 CPU-1
186 *
187 * atomic_inc(mfc);
188 * write int3s
189 * <trap-int3> // implicit (r)mb
190 * if (atomic_read(mfc))
191 * call ftrace_int3_handler()
192 *
193 * Then when we are finished:
194 *
195 * atomic_dec(mfc);
196 *
197 * If we hit a breakpoint that was not set by ftrace, it does not
198 * matter if ftrace_int3_handler() is called or not. It will
199 * simply be ignored. But it is crucial that a ftrace nop/caller
200 * breakpoint is handled. No other user should ever place a
201 * breakpoint on an ftrace nop/caller location. It must only
202 * be done by this code.
203 */
204atomic_t modifying_ftrace_code __read_mostly;
205
206static int
207ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
208 unsigned const char *new_code);
209
210/*
211 * Should never be called:
212 * As it is only called by __ftrace_replace_code() which is called by
213 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
214 * which is called to turn mcount into nops or nops into function calls
215 * but not to convert a function from not using regs to one that uses
216 * regs, which ftrace_modify_call() is for.
217 */
218int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
219 unsigned long addr)
220{
221 WARN_ON(1);
222 return -EINVAL;
223}
224
225static unsigned long ftrace_update_func;
226
227static int update_ftrace_func(unsigned long ip, void *new)
228{
229 unsigned char old[MCOUNT_INSN_SIZE];
230 int ret;
231
232 memcpy(old, (void *)ip, MCOUNT_INSN_SIZE);
233
234 ftrace_update_func = ip;
235 /* Make sure the breakpoints see the ftrace_update_func update */
236 smp_wmb();
237
238 /* See comment above by declaration of modifying_ftrace_code */
239 atomic_inc(&modifying_ftrace_code);
240
241 ret = ftrace_modify_code(ip, old, new);
242
243 atomic_dec(&modifying_ftrace_code);
244
245 return ret;
246}
247
248int ftrace_update_ftrace_func(ftrace_func_t func)
249{
250 unsigned long ip = (unsigned long)(&ftrace_call);
251 unsigned char *new;
252 int ret;
253
254 new = ftrace_call_replace(ip, (unsigned long)func);
255 ret = update_ftrace_func(ip, new);
256
257 /* Also update the regs callback function */
258 if (!ret) {
259 ip = (unsigned long)(&ftrace_regs_call);
260 new = ftrace_call_replace(ip, (unsigned long)func);
261 ret = update_ftrace_func(ip, new);
262 }
263
264 return ret;
265}
266
267static int is_ftrace_caller(unsigned long ip)
268{
269 if (ip == ftrace_update_func)
270 return 1;
271
272 return 0;
273}
274
275/*
276 * A breakpoint was added to the code address we are about to
277 * modify, and this is the handle that will just skip over it.
278 * We are either changing a nop into a trace call, or a trace
279 * call to a nop. While the change is taking place, we treat
280 * it just like it was a nop.
281 */
282int ftrace_int3_handler(struct pt_regs *regs)
283{
284 unsigned long ip;
285
286 if (WARN_ON_ONCE(!regs))
287 return 0;
288
289 ip = regs->ip - 1;
290 if (!ftrace_location(ip) && !is_ftrace_caller(ip))
291 return 0;
292
293 regs->ip += MCOUNT_INSN_SIZE - 1;
294
295 return 1;
296}
297
298static int ftrace_write(unsigned long ip, const char *val, int size)
299{
300 /*
301 * On x86_64, kernel text mappings are mapped read-only with
302 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
303 * of the kernel text mapping to modify the kernel text.
304 *
305 * For 32bit kernels, these mappings are same and we can use
306 * kernel identity mapping to modify code.
307 */
308 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
309 ip = (unsigned long)__va(__pa_symbol(ip));
310
311 if (probe_kernel_write((void *)ip, val, size))
312 return -EPERM;
313
314 return 0;
315}
316
317static int add_break(unsigned long ip, const char *old)
318{
319 unsigned char replaced[MCOUNT_INSN_SIZE];
320 unsigned char brk = BREAKPOINT_INSTRUCTION;
321
322 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
323 return -EFAULT;
324
325 /* Make sure it is what we expect it to be */
326 if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
327 return -EINVAL;
328
329 return ftrace_write(ip, &brk, 1);
330}
331
332static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
333{
334 unsigned const char *old;
335 unsigned long ip = rec->ip;
336
337 old = ftrace_call_replace(ip, addr);
338
339 return add_break(rec->ip, old);
340}
341
342
343static int add_brk_on_nop(struct dyn_ftrace *rec)
344{
345 unsigned const char *old;
346
347 old = ftrace_nop_replace();
348
349 return add_break(rec->ip, old);
350}
351
352/*
353 * If the record has the FTRACE_FL_REGS set, that means that it
354 * wants to convert to a callback that saves all regs. If FTRACE_FL_REGS
355 * is not not set, then it wants to convert to the normal callback.
356 */
357static unsigned long get_ftrace_addr(struct dyn_ftrace *rec)
358{
359 if (rec->flags & FTRACE_FL_REGS)
360 return (unsigned long)FTRACE_REGS_ADDR;
361 else
362 return (unsigned long)FTRACE_ADDR;
363}
364
365/*
366 * The FTRACE_FL_REGS_EN is set when the record already points to
367 * a function that saves all the regs. Basically the '_EN' version
368 * represents the current state of the function.
369 */
370static unsigned long get_ftrace_old_addr(struct dyn_ftrace *rec)
371{
372 if (rec->flags & FTRACE_FL_REGS_EN)
373 return (unsigned long)FTRACE_REGS_ADDR;
374 else
375 return (unsigned long)FTRACE_ADDR;
376}
377
378static int add_breakpoints(struct dyn_ftrace *rec, int enable)
379{
380 unsigned long ftrace_addr;
381 int ret;
382
383 ret = ftrace_test_record(rec, enable);
384
385 ftrace_addr = get_ftrace_addr(rec);
386
387 switch (ret) {
388 case FTRACE_UPDATE_IGNORE:
389 return 0;
390
391 case FTRACE_UPDATE_MAKE_CALL:
392 /* converting nop to call */
393 return add_brk_on_nop(rec);
394
395 case FTRACE_UPDATE_MODIFY_CALL_REGS:
396 case FTRACE_UPDATE_MODIFY_CALL:
397 ftrace_addr = get_ftrace_old_addr(rec);
398 /* fall through */
399 case FTRACE_UPDATE_MAKE_NOP:
400 /* converting a call to a nop */
401 return add_brk_on_call(rec, ftrace_addr);
402 }
403 return 0;
404}
405
406/*
407 * On error, we need to remove breakpoints. This needs to
408 * be done caefully. If the address does not currently have a
409 * breakpoint, we know we are done. Otherwise, we look at the
410 * remaining 4 bytes of the instruction. If it matches a nop
411 * we replace the breakpoint with the nop. Otherwise we replace
412 * it with the call instruction.
413 */
414static int remove_breakpoint(struct dyn_ftrace *rec)
415{
416 unsigned char ins[MCOUNT_INSN_SIZE];
417 unsigned char brk = BREAKPOINT_INSTRUCTION;
418 const unsigned char *nop;
419 unsigned long ftrace_addr;
420 unsigned long ip = rec->ip;
421
422 /* If we fail the read, just give up */
423 if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
424 return -EFAULT;
425
426 /* If this does not have a breakpoint, we are done */
427 if (ins[0] != brk)
428 return 0;
429
430 nop = ftrace_nop_replace();
431
432 /*
433 * If the last 4 bytes of the instruction do not match
434 * a nop, then we assume that this is a call to ftrace_addr.
435 */
436 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
437 /*
438 * For extra paranoidism, we check if the breakpoint is on
439 * a call that would actually jump to the ftrace_addr.
440 * If not, don't touch the breakpoint, we make just create
441 * a disaster.
442 */
443 ftrace_addr = get_ftrace_addr(rec);
444 nop = ftrace_call_replace(ip, ftrace_addr);
445
446 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
447 goto update;
448
449 /* Check both ftrace_addr and ftrace_old_addr */
450 ftrace_addr = get_ftrace_old_addr(rec);
451 nop = ftrace_call_replace(ip, ftrace_addr);
452
453 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
454 return -EINVAL;
455 }
456
457 update:
458 return ftrace_write(ip, nop, 1);
459}
460
461static int add_update_code(unsigned long ip, unsigned const char *new)
462{
463 /* skip breakpoint */
464 ip++;
465 new++;
466 return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1);
467}
468
469static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
470{
471 unsigned long ip = rec->ip;
472 unsigned const char *new;
473
474 new = ftrace_call_replace(ip, addr);
475 return add_update_code(ip, new);
476}
477
478static int add_update_nop(struct dyn_ftrace *rec)
479{
480 unsigned long ip = rec->ip;
481 unsigned const char *new;
482
483 new = ftrace_nop_replace();
484 return add_update_code(ip, new);
485}
486
487static int add_update(struct dyn_ftrace *rec, int enable)
488{
489 unsigned long ftrace_addr;
490 int ret;
491
492 ret = ftrace_test_record(rec, enable);
493
494 ftrace_addr = get_ftrace_addr(rec);
495
496 switch (ret) {
497 case FTRACE_UPDATE_IGNORE:
498 return 0;
499
500 case FTRACE_UPDATE_MODIFY_CALL_REGS:
501 case FTRACE_UPDATE_MODIFY_CALL:
502 case FTRACE_UPDATE_MAKE_CALL:
503 /* converting nop to call */
504 return add_update_call(rec, ftrace_addr);
505
506 case FTRACE_UPDATE_MAKE_NOP:
507 /* converting a call to a nop */
508 return add_update_nop(rec);
509 }
510
511 return 0;
512}
513
514static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
515{
516 unsigned long ip = rec->ip;
517 unsigned const char *new;
518
519 new = ftrace_call_replace(ip, addr);
520
521 return ftrace_write(ip, new, 1);
522}
523
524static int finish_update_nop(struct dyn_ftrace *rec)
525{
526 unsigned long ip = rec->ip;
527 unsigned const char *new;
528
529 new = ftrace_nop_replace();
530
531 return ftrace_write(ip, new, 1);
532}
533
534static int finish_update(struct dyn_ftrace *rec, int enable)
535{
536 unsigned long ftrace_addr;
537 int ret;
538
539 ret = ftrace_update_record(rec, enable);
540
541 ftrace_addr = get_ftrace_addr(rec);
542
543 switch (ret) {
544 case FTRACE_UPDATE_IGNORE:
545 return 0;
546
547 case FTRACE_UPDATE_MODIFY_CALL_REGS:
548 case FTRACE_UPDATE_MODIFY_CALL:
549 case FTRACE_UPDATE_MAKE_CALL:
550 /* converting nop to call */
551 return finish_update_call(rec, ftrace_addr);
552
553 case FTRACE_UPDATE_MAKE_NOP:
554 /* converting a call to a nop */
555 return finish_update_nop(rec);
556 }
557
558 return 0;
559}
560
561static void do_sync_core(void *data)
562{
563 sync_core();
564}
565
566static void run_sync(void)
567{
568 int enable_irqs = irqs_disabled();
569
570 /* We may be called with interrupts disbled (on bootup). */
571 if (enable_irqs)
572 local_irq_enable();
573 on_each_cpu(do_sync_core, NULL, 1);
574 if (enable_irqs)
575 local_irq_disable();
576}
577
578void ftrace_replace_code(int enable)
579{
580 struct ftrace_rec_iter *iter;
581 struct dyn_ftrace *rec;
582 const char *report = "adding breakpoints";
583 int count = 0;
584 int ret;
585
586 for_ftrace_rec_iter(iter) {
587 rec = ftrace_rec_iter_record(iter);
588
589 ret = add_breakpoints(rec, enable);
590 if (ret)
591 goto remove_breakpoints;
592 count++;
593 }
594
595 run_sync();
596
597 report = "updating code";
598
599 for_ftrace_rec_iter(iter) {
600 rec = ftrace_rec_iter_record(iter);
601
602 ret = add_update(rec, enable);
603 if (ret)
604 goto remove_breakpoints;
605 }
606
607 run_sync();
608
609 report = "removing breakpoints";
610
611 for_ftrace_rec_iter(iter) {
612 rec = ftrace_rec_iter_record(iter);
613
614 ret = finish_update(rec, enable);
615 if (ret)
616 goto remove_breakpoints;
617 }
618
619 run_sync();
620
621 return;
622
623 remove_breakpoints:
624 ftrace_bug(ret, rec ? rec->ip : 0);
625 printk(KERN_WARNING "Failed on %s (%d):\n", report, count);
626 for_ftrace_rec_iter(iter) {
627 rec = ftrace_rec_iter_record(iter);
628 /*
629 * Breakpoints are handled only when this function is in
630 * progress. The system could not work with them.
631 */
632 if (remove_breakpoint(rec))
633 BUG();
634 }
635 run_sync();
636}
637
638static int
639ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
640 unsigned const char *new_code)
641{
642 int ret;
643
644 ret = add_break(ip, old_code);
645 if (ret)
646 goto out;
647
648 run_sync();
649
650 ret = add_update_code(ip, new_code);
651 if (ret)
652 goto fail_update;
653
654 run_sync();
655
656 ret = ftrace_write(ip, new_code, 1);
657 /*
658 * The breakpoint is handled only when this function is in progress.
659 * The system could not work if we could not remove it.
660 */
661 BUG_ON(ret);
662 out:
663 run_sync();
664 return ret;
665
666 fail_update:
667 /* Also here the system could not work with the breakpoint */
668 if (ftrace_write(ip, old_code, 1))
669 BUG();
670 goto out;
671}
672
673void arch_ftrace_update_code(int command)
674{
675 /* See comment above by declaration of modifying_ftrace_code */
676 atomic_inc(&modifying_ftrace_code);
677
678 ftrace_modify_all_code(command);
679
680 atomic_dec(&modifying_ftrace_code);
681}
682
683int __init ftrace_dyn_arch_init(void)
684{
685 return 0;
686}
687#endif
688
689#ifdef CONFIG_FUNCTION_GRAPH_TRACER
690
691#ifdef CONFIG_DYNAMIC_FTRACE
692extern void ftrace_graph_call(void);
693
694static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
695{
696 static union ftrace_code_union calc;
697
698 /* Jmp not a call (ignore the .e8) */
699 calc.e8 = 0xe9;
700 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
701
702 /*
703 * ftrace external locks synchronize the access to the static variable.
704 */
705 return calc.code;
706}
707
708static int ftrace_mod_jmp(unsigned long ip, void *func)
709{
710 unsigned char *new;
711
712 new = ftrace_jmp_replace(ip, (unsigned long)func);
713
714 return update_ftrace_func(ip, new);
715}
716
717int ftrace_enable_ftrace_graph_caller(void)
718{
719 unsigned long ip = (unsigned long)(&ftrace_graph_call);
720
721 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
722}
723
724int ftrace_disable_ftrace_graph_caller(void)
725{
726 unsigned long ip = (unsigned long)(&ftrace_graph_call);
727
728 return ftrace_mod_jmp(ip, &ftrace_stub);
729}
730
731#endif /* !CONFIG_DYNAMIC_FTRACE */
732
733/*
734 * Hook the return address and push it in the stack of return addrs
735 * in current thread info.
736 */
737void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
738 unsigned long frame_pointer)
739{
740 unsigned long old;
741 int faulted;
742 struct ftrace_graph_ent trace;
743 unsigned long return_hooker = (unsigned long)
744 &return_to_handler;
745
746 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
747 return;
748
749 /*
750 * Protect against fault, even if it shouldn't
751 * happen. This tool is too much intrusive to
752 * ignore such a protection.
753 */
754 asm volatile(
755 "1: " _ASM_MOV " (%[parent]), %[old]\n"
756 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
757 " movl $0, %[faulted]\n"
758 "3:\n"
759
760 ".section .fixup, \"ax\"\n"
761 "4: movl $1, %[faulted]\n"
762 " jmp 3b\n"
763 ".previous\n"
764
765 _ASM_EXTABLE(1b, 4b)
766 _ASM_EXTABLE(2b, 4b)
767
768 : [old] "=&r" (old), [faulted] "=r" (faulted)
769 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
770 : "memory"
771 );
772
773 if (unlikely(faulted)) {
774 ftrace_graph_stop();
775 WARN_ON(1);
776 return;
777 }
778
779 trace.func = self_addr;
780 trace.depth = current->curr_ret_stack + 1;
781
782 /* Only trace if the calling function expects to */
783 if (!ftrace_graph_entry(&trace)) {
784 *parent = old;
785 return;
786 }
787
788 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
789 frame_pointer) == -EBUSY) {
790 *parent = old;
791 return;
792 }
793}
794#endif /* CONFIG_FUNCTION_GRAPH_TRACER */