2 * Code for replacing ftrace calls with jumps.
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
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.
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
24 #include <trace/syscall.h>
26 #include <asm/cacheflush.h>
27 #include <asm/kprobes.h>
28 #include <asm/ftrace.h>
31 #ifdef CONFIG_DYNAMIC_FTRACE
33 int ftrace_arch_code_modify_prepare(void)
36 set_all_modules_text_rw();
40 int ftrace_arch_code_modify_post_process(void)
42 set_all_modules_text_ro();
47 union ftrace_code_union {
48 char code[MCOUNT_INSN_SIZE];
52 } __attribute__((packed));
55 static int ftrace_calc_offset(long ip, long addr)
57 return (int)(addr - ip);
60 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
62 static union ftrace_code_union calc;
65 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
68 * No locking needed, this must be called via kstop_machine
69 * which in essence is like running on a uniprocessor machine.
75 within(unsigned long addr, unsigned long start, unsigned long end)
77 return addr >= start && addr < end;
81 do_ftrace_mod_code(unsigned long ip, const void *new_code)
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.
88 * For 32bit kernels, these mappings are same and we can use
89 * kernel identity mapping to modify code.
91 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
92 ip = (unsigned long)__va(__pa(ip));
94 return probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE);
97 static const unsigned char *ftrace_nop_replace(void)
99 return ideal_nops[NOP_ATOMIC5];
103 ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code,
104 unsigned const char *new_code)
106 unsigned char replaced[MCOUNT_INSN_SIZE];
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.
114 * No real locking needed, this code is run through
115 * kstop_machine, or before SMP starts.
118 /* read the text we want to modify */
119 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
122 /* Make sure it is what we expect it to be */
123 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
126 /* replace the text with the new text */
127 if (do_ftrace_mod_code(ip, new_code))
135 int ftrace_make_nop(struct module *mod,
136 struct dyn_ftrace *rec, unsigned long addr)
138 unsigned const char *new, *old;
139 unsigned long ip = rec->ip;
141 old = ftrace_call_replace(ip, addr);
142 new = ftrace_nop_replace();
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.
152 if (addr == MCOUNT_ADDR)
153 return ftrace_modify_code_direct(rec->ip, old, new);
155 /* Normal cases use add_brk_on_nop */
156 WARN_ONCE(1, "invalid use of ftrace_make_nop");
160 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
162 unsigned const char *new, *old;
163 unsigned long ip = rec->ip;
165 old = ftrace_nop_replace();
166 new = ftrace_call_replace(ip, addr);
168 /* Should only be called when module is loaded */
169 return ftrace_modify_code_direct(rec->ip, old, new);
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.
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:
188 * <trap-int3> // implicit (r)mb
189 * if (atomic_read(mfc))
190 * call ftrace_int3_handler()
192 * Then when we are finished:
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.
203 atomic_t modifying_ftrace_code __read_mostly;
206 ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
207 unsigned const char *new_code);
209 #ifdef ARCH_SUPPORTS_FTRACE_SAVE_REGS
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.
218 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
226 int ftrace_update_ftrace_func(ftrace_func_t func)
228 unsigned long ip = (unsigned long)(&ftrace_call);
229 unsigned char old[MCOUNT_INSN_SIZE], *new;
232 memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
233 new = ftrace_call_replace(ip, (unsigned long)func);
235 /* See comment above by declaration of modifying_ftrace_code */
236 atomic_inc(&modifying_ftrace_code);
238 ret = ftrace_modify_code(ip, old, new);
240 #ifdef ARCH_SUPPORTS_FTRACE_SAVE_REGS
241 /* Also update the regs callback function */
243 ip = (unsigned long)(&ftrace_regs_call);
244 memcpy(old, &ftrace_regs_call, MCOUNT_INSN_SIZE);
245 new = ftrace_call_replace(ip, (unsigned long)func);
246 ret = ftrace_modify_code(ip, old, new);
250 atomic_dec(&modifying_ftrace_code);
256 * A breakpoint was added to the code address we are about to
257 * modify, and this is the handle that will just skip over it.
258 * We are either changing a nop into a trace call, or a trace
259 * call to a nop. While the change is taking place, we treat
260 * it just like it was a nop.
262 int ftrace_int3_handler(struct pt_regs *regs)
264 if (WARN_ON_ONCE(!regs))
267 if (!ftrace_location(regs->ip - 1))
270 regs->ip += MCOUNT_INSN_SIZE - 1;
275 static int ftrace_write(unsigned long ip, const char *val, int size)
278 * On x86_64, kernel text mappings are mapped read-only with
279 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
280 * of the kernel text mapping to modify the kernel text.
282 * For 32bit kernels, these mappings are same and we can use
283 * kernel identity mapping to modify code.
285 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
286 ip = (unsigned long)__va(__pa(ip));
288 return probe_kernel_write((void *)ip, val, size);
291 static int add_break(unsigned long ip, const char *old)
293 unsigned char replaced[MCOUNT_INSN_SIZE];
294 unsigned char brk = BREAKPOINT_INSTRUCTION;
296 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
299 /* Make sure it is what we expect it to be */
300 if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
303 if (ftrace_write(ip, &brk, 1))
309 static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
311 unsigned const char *old;
312 unsigned long ip = rec->ip;
314 old = ftrace_call_replace(ip, addr);
316 return add_break(rec->ip, old);
320 static int add_brk_on_nop(struct dyn_ftrace *rec)
322 unsigned const char *old;
324 old = ftrace_nop_replace();
326 return add_break(rec->ip, old);
330 * If the record has the FTRACE_FL_REGS set, that means that it
331 * wants to convert to a callback that saves all regs. If FTRACE_FL_REGS
332 * is not not set, then it wants to convert to the normal callback.
334 static unsigned long get_ftrace_addr(struct dyn_ftrace *rec)
336 if (rec->flags & FTRACE_FL_REGS)
337 return (unsigned long)FTRACE_REGS_ADDR;
339 return (unsigned long)FTRACE_ADDR;
343 * The FTRACE_FL_REGS_EN is set when the record already points to
344 * a function that saves all the regs. Basically the '_EN' version
345 * represents the current state of the function.
347 static unsigned long get_ftrace_old_addr(struct dyn_ftrace *rec)
349 if (rec->flags & FTRACE_FL_REGS_EN)
350 return (unsigned long)FTRACE_REGS_ADDR;
352 return (unsigned long)FTRACE_ADDR;
355 static int add_breakpoints(struct dyn_ftrace *rec, int enable)
357 unsigned long ftrace_addr;
360 ret = ftrace_test_record(rec, enable);
362 ftrace_addr = get_ftrace_addr(rec);
365 case FTRACE_UPDATE_IGNORE:
368 case FTRACE_UPDATE_MAKE_CALL:
369 /* converting nop to call */
370 return add_brk_on_nop(rec);
372 case FTRACE_UPDATE_MODIFY_CALL_REGS:
373 case FTRACE_UPDATE_MODIFY_CALL:
374 ftrace_addr = get_ftrace_old_addr(rec);
376 case FTRACE_UPDATE_MAKE_NOP:
377 /* converting a call to a nop */
378 return add_brk_on_call(rec, ftrace_addr);
384 * On error, we need to remove breakpoints. This needs to
385 * be done caefully. If the address does not currently have a
386 * breakpoint, we know we are done. Otherwise, we look at the
387 * remaining 4 bytes of the instruction. If it matches a nop
388 * we replace the breakpoint with the nop. Otherwise we replace
389 * it with the call instruction.
391 static int remove_breakpoint(struct dyn_ftrace *rec)
393 unsigned char ins[MCOUNT_INSN_SIZE];
394 unsigned char brk = BREAKPOINT_INSTRUCTION;
395 const unsigned char *nop;
396 unsigned long ftrace_addr;
397 unsigned long ip = rec->ip;
399 /* If we fail the read, just give up */
400 if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
403 /* If this does not have a breakpoint, we are done */
407 nop = ftrace_nop_replace();
410 * If the last 4 bytes of the instruction do not match
411 * a nop, then we assume that this is a call to ftrace_addr.
413 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
415 * For extra paranoidism, we check if the breakpoint is on
416 * a call that would actually jump to the ftrace_addr.
417 * If not, don't touch the breakpoint, we make just create
420 ftrace_addr = get_ftrace_addr(rec);
421 nop = ftrace_call_replace(ip, ftrace_addr);
423 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
426 /* Check both ftrace_addr and ftrace_old_addr */
427 ftrace_addr = get_ftrace_old_addr(rec);
428 nop = ftrace_call_replace(ip, ftrace_addr);
430 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
435 return probe_kernel_write((void *)ip, &nop[0], 1);
438 static int add_update_code(unsigned long ip, unsigned const char *new)
440 /* skip breakpoint */
443 if (ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1))
448 static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
450 unsigned long ip = rec->ip;
451 unsigned const char *new;
453 new = ftrace_call_replace(ip, addr);
454 return add_update_code(ip, new);
457 static int add_update_nop(struct dyn_ftrace *rec)
459 unsigned long ip = rec->ip;
460 unsigned const char *new;
462 new = ftrace_nop_replace();
463 return add_update_code(ip, new);
466 static int add_update(struct dyn_ftrace *rec, int enable)
468 unsigned long ftrace_addr;
471 ret = ftrace_test_record(rec, enable);
473 ftrace_addr = get_ftrace_addr(rec);
476 case FTRACE_UPDATE_IGNORE:
479 case FTRACE_UPDATE_MODIFY_CALL_REGS:
480 case FTRACE_UPDATE_MODIFY_CALL:
481 case FTRACE_UPDATE_MAKE_CALL:
482 /* converting nop to call */
483 return add_update_call(rec, ftrace_addr);
485 case FTRACE_UPDATE_MAKE_NOP:
486 /* converting a call to a nop */
487 return add_update_nop(rec);
493 static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
495 unsigned long ip = rec->ip;
496 unsigned const char *new;
498 new = ftrace_call_replace(ip, addr);
500 if (ftrace_write(ip, new, 1))
506 static int finish_update_nop(struct dyn_ftrace *rec)
508 unsigned long ip = rec->ip;
509 unsigned const char *new;
511 new = ftrace_nop_replace();
513 if (ftrace_write(ip, new, 1))
518 static int finish_update(struct dyn_ftrace *rec, int enable)
520 unsigned long ftrace_addr;
523 ret = ftrace_update_record(rec, enable);
525 ftrace_addr = get_ftrace_addr(rec);
528 case FTRACE_UPDATE_IGNORE:
531 case FTRACE_UPDATE_MODIFY_CALL_REGS:
532 case FTRACE_UPDATE_MODIFY_CALL:
533 case FTRACE_UPDATE_MAKE_CALL:
534 /* converting nop to call */
535 return finish_update_call(rec, ftrace_addr);
537 case FTRACE_UPDATE_MAKE_NOP:
538 /* converting a call to a nop */
539 return finish_update_nop(rec);
545 static void do_sync_core(void *data)
550 static void run_sync(void)
552 int enable_irqs = irqs_disabled();
554 /* We may be called with interrupts disbled (on bootup). */
557 on_each_cpu(do_sync_core, NULL, 1);
562 void ftrace_replace_code(int enable)
564 struct ftrace_rec_iter *iter;
565 struct dyn_ftrace *rec;
566 const char *report = "adding breakpoints";
570 for_ftrace_rec_iter(iter) {
571 rec = ftrace_rec_iter_record(iter);
573 ret = add_breakpoints(rec, enable);
575 goto remove_breakpoints;
581 report = "updating code";
583 for_ftrace_rec_iter(iter) {
584 rec = ftrace_rec_iter_record(iter);
586 ret = add_update(rec, enable);
588 goto remove_breakpoints;
593 report = "removing breakpoints";
595 for_ftrace_rec_iter(iter) {
596 rec = ftrace_rec_iter_record(iter);
598 ret = finish_update(rec, enable);
600 goto remove_breakpoints;
608 ftrace_bug(ret, rec ? rec->ip : 0);
609 printk(KERN_WARNING "Failed on %s (%d):\n", report, count);
610 for_ftrace_rec_iter(iter) {
611 rec = ftrace_rec_iter_record(iter);
612 remove_breakpoint(rec);
617 ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
618 unsigned const char *new_code)
622 ret = add_break(ip, old_code);
628 ret = add_update_code(ip, new_code);
634 ret = ftrace_write(ip, new_code, 1);
644 probe_kernel_write((void *)ip, &old_code[0], 1);
648 void arch_ftrace_update_code(int command)
650 /* See comment above by declaration of modifying_ftrace_code */
651 atomic_inc(&modifying_ftrace_code);
653 ftrace_modify_all_code(command);
655 atomic_dec(&modifying_ftrace_code);
658 int __init ftrace_dyn_arch_init(void *data)
660 /* The return code is retured via data */
661 *(unsigned long *)data = 0;
667 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
669 #ifdef CONFIG_DYNAMIC_FTRACE
670 extern void ftrace_graph_call(void);
672 static int ftrace_mod_jmp(unsigned long ip,
673 int old_offset, int new_offset)
675 unsigned char code[MCOUNT_INSN_SIZE];
677 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
680 if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
683 *(int *)(&code[1]) = new_offset;
685 if (do_ftrace_mod_code(ip, &code))
691 int ftrace_enable_ftrace_graph_caller(void)
693 unsigned long ip = (unsigned long)(&ftrace_graph_call);
694 int old_offset, new_offset;
696 old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
697 new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
699 return ftrace_mod_jmp(ip, old_offset, new_offset);
702 int ftrace_disable_ftrace_graph_caller(void)
704 unsigned long ip = (unsigned long)(&ftrace_graph_call);
705 int old_offset, new_offset;
707 old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
708 new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
710 return ftrace_mod_jmp(ip, old_offset, new_offset);
713 #endif /* !CONFIG_DYNAMIC_FTRACE */
716 * Hook the return address and push it in the stack of return addrs
717 * in current thread info.
719 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
720 unsigned long frame_pointer)
724 struct ftrace_graph_ent trace;
725 unsigned long return_hooker = (unsigned long)
728 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
732 * Protect against fault, even if it shouldn't
733 * happen. This tool is too much intrusive to
734 * ignore such a protection.
737 "1: " _ASM_MOV " (%[parent]), %[old]\n"
738 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
739 " movl $0, %[faulted]\n"
742 ".section .fixup, \"ax\"\n"
743 "4: movl $1, %[faulted]\n"
750 : [old] "=&r" (old), [faulted] "=r" (faulted)
751 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
755 if (unlikely(faulted)) {
761 trace.func = self_addr;
762 trace.depth = current->curr_ret_stack + 1;
764 /* Only trace if the calling function expects to */
765 if (!ftrace_graph_entry(&trace)) {
770 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
771 frame_pointer) == -EBUSY) {
776 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */