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 int ftrace_update_ftrace_func(ftrace_func_t func)
211 unsigned long ip = (unsigned long)(&ftrace_call);
212 unsigned char old[MCOUNT_INSN_SIZE], *new;
215 memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
216 new = ftrace_call_replace(ip, (unsigned long)func);
218 /* See comment above by declaration of modifying_ftrace_code */
219 atomic_inc(&modifying_ftrace_code);
221 ret = ftrace_modify_code(ip, old, new);
223 atomic_dec(&modifying_ftrace_code);
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.
235 int ftrace_int3_handler(struct pt_regs *regs)
237 if (WARN_ON_ONCE(!regs))
240 if (!ftrace_location(regs->ip - 1))
243 regs->ip += MCOUNT_INSN_SIZE - 1;
248 static int ftrace_write(unsigned long ip, const char *val, int size)
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.
255 * For 32bit kernels, these mappings are same and we can use
256 * kernel identity mapping to modify code.
258 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
259 ip = (unsigned long)__va(__pa(ip));
261 return probe_kernel_write((void *)ip, val, size);
264 static int add_break(unsigned long ip, const char *old)
266 unsigned char replaced[MCOUNT_INSN_SIZE];
267 unsigned char brk = BREAKPOINT_INSTRUCTION;
269 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
272 /* Make sure it is what we expect it to be */
273 if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
276 if (ftrace_write(ip, &brk, 1))
282 static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
284 unsigned const char *old;
285 unsigned long ip = rec->ip;
287 old = ftrace_call_replace(ip, addr);
289 return add_break(rec->ip, old);
293 static int add_brk_on_nop(struct dyn_ftrace *rec)
295 unsigned const char *old;
297 old = ftrace_nop_replace();
299 return add_break(rec->ip, old);
302 static int add_breakpoints(struct dyn_ftrace *rec, int enable)
304 unsigned long ftrace_addr;
307 ret = ftrace_test_record(rec, enable);
309 ftrace_addr = (unsigned long)FTRACE_ADDR;
312 case FTRACE_UPDATE_IGNORE:
315 case FTRACE_UPDATE_MAKE_CALL:
316 /* converting nop to call */
317 return add_brk_on_nop(rec);
319 case FTRACE_UPDATE_MAKE_NOP:
320 /* converting a call to a nop */
321 return add_brk_on_call(rec, ftrace_addr);
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.
334 static int remove_breakpoint(struct dyn_ftrace *rec)
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;
342 /* If we fail the read, just give up */
343 if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
346 /* If this does not have a breakpoint, we are done */
350 nop = ftrace_nop_replace();
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.
356 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
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
363 ftrace_addr = (unsigned long)FTRACE_ADDR;
364 nop = ftrace_call_replace(ip, ftrace_addr);
366 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
370 return probe_kernel_write((void *)ip, &nop[0], 1);
373 static int add_update_code(unsigned long ip, unsigned const char *new)
375 /* skip breakpoint */
378 if (ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1))
383 static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
385 unsigned long ip = rec->ip;
386 unsigned const char *new;
388 new = ftrace_call_replace(ip, addr);
389 return add_update_code(ip, new);
392 static int add_update_nop(struct dyn_ftrace *rec)
394 unsigned long ip = rec->ip;
395 unsigned const char *new;
397 new = ftrace_nop_replace();
398 return add_update_code(ip, new);
401 static int add_update(struct dyn_ftrace *rec, int enable)
403 unsigned long ftrace_addr;
406 ret = ftrace_test_record(rec, enable);
408 ftrace_addr = (unsigned long)FTRACE_ADDR;
411 case FTRACE_UPDATE_IGNORE:
414 case FTRACE_UPDATE_MAKE_CALL:
415 /* converting nop to call */
416 return add_update_call(rec, ftrace_addr);
418 case FTRACE_UPDATE_MAKE_NOP:
419 /* converting a call to a nop */
420 return add_update_nop(rec);
426 static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
428 unsigned long ip = rec->ip;
429 unsigned const char *new;
431 new = ftrace_call_replace(ip, addr);
433 if (ftrace_write(ip, new, 1))
439 static int finish_update_nop(struct dyn_ftrace *rec)
441 unsigned long ip = rec->ip;
442 unsigned const char *new;
444 new = ftrace_nop_replace();
446 if (ftrace_write(ip, new, 1))
451 static int finish_update(struct dyn_ftrace *rec, int enable)
453 unsigned long ftrace_addr;
456 ret = ftrace_update_record(rec, enable);
458 ftrace_addr = (unsigned long)FTRACE_ADDR;
461 case FTRACE_UPDATE_IGNORE:
464 case FTRACE_UPDATE_MAKE_CALL:
465 /* converting nop to call */
466 return finish_update_call(rec, ftrace_addr);
468 case FTRACE_UPDATE_MAKE_NOP:
469 /* converting a call to a nop */
470 return finish_update_nop(rec);
476 static void do_sync_core(void *data)
481 static void run_sync(void)
483 int enable_irqs = irqs_disabled();
485 /* We may be called with interrupts disbled (on bootup). */
488 on_each_cpu(do_sync_core, NULL, 1);
493 void ftrace_replace_code(int enable)
495 struct ftrace_rec_iter *iter;
496 struct dyn_ftrace *rec;
497 const char *report = "adding breakpoints";
501 for_ftrace_rec_iter(iter) {
502 rec = ftrace_rec_iter_record(iter);
504 ret = add_breakpoints(rec, enable);
506 goto remove_breakpoints;
512 report = "updating code";
514 for_ftrace_rec_iter(iter) {
515 rec = ftrace_rec_iter_record(iter);
517 ret = add_update(rec, enable);
519 goto remove_breakpoints;
524 report = "removing breakpoints";
526 for_ftrace_rec_iter(iter) {
527 rec = ftrace_rec_iter_record(iter);
529 ret = finish_update(rec, enable);
531 goto 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);
548 ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
549 unsigned const char *new_code)
553 ret = add_break(ip, old_code);
559 ret = add_update_code(ip, new_code);
565 ret = ftrace_write(ip, new_code, 1);
575 probe_kernel_write((void *)ip, &old_code[0], 1);
579 void arch_ftrace_update_code(int command)
581 /* See comment above by declaration of modifying_ftrace_code */
582 atomic_inc(&modifying_ftrace_code);
584 ftrace_modify_all_code(command);
586 atomic_dec(&modifying_ftrace_code);
589 int __init ftrace_dyn_arch_init(void *data)
591 /* The return code is retured via data */
592 *(unsigned long *)data = 0;
598 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
600 #ifdef CONFIG_DYNAMIC_FTRACE
601 extern void ftrace_graph_call(void);
603 static int ftrace_mod_jmp(unsigned long ip,
604 int old_offset, int new_offset)
606 unsigned char code[MCOUNT_INSN_SIZE];
608 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
611 if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
614 *(int *)(&code[1]) = new_offset;
616 if (do_ftrace_mod_code(ip, &code))
622 int ftrace_enable_ftrace_graph_caller(void)
624 unsigned long ip = (unsigned long)(&ftrace_graph_call);
625 int old_offset, new_offset;
627 old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
628 new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
630 return ftrace_mod_jmp(ip, old_offset, new_offset);
633 int ftrace_disable_ftrace_graph_caller(void)
635 unsigned long ip = (unsigned long)(&ftrace_graph_call);
636 int old_offset, new_offset;
638 old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
639 new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
641 return ftrace_mod_jmp(ip, old_offset, new_offset);
644 #endif /* !CONFIG_DYNAMIC_FTRACE */
647 * Hook the return address and push it in the stack of return addrs
648 * in current thread info.
650 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
651 unsigned long frame_pointer)
655 struct ftrace_graph_ent trace;
656 unsigned long return_hooker = (unsigned long)
659 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
663 * Protect against fault, even if it shouldn't
664 * happen. This tool is too much intrusive to
665 * ignore such a protection.
668 "1: " _ASM_MOV " (%[parent]), %[old]\n"
669 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
670 " movl $0, %[faulted]\n"
673 ".section .fixup, \"ax\"\n"
674 "4: movl $1, %[faulted]\n"
681 : [old] "=&r" (old), [faulted] "=r" (faulted)
682 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
686 if (unlikely(faulted)) {
692 trace.func = self_addr;
693 trace.depth = current->curr_ret_stack + 1;
695 /* Only trace if the calling function expects to */
696 if (!ftrace_graph_entry(&trace)) {
701 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
702 frame_pointer) == -EBUSY) {
707 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */