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/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(¤t->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 */