1 // SPDX-License-Identifier: GPL-2.0
3 * Dynamic function tracing support.
5 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
7 * Thanks goes to Ingo Molnar, for suggesting the idea.
8 * Mathieu Desnoyers, for suggesting postponing the modifications.
9 * Arjan van de Ven, for keeping me straight, and explaining to me
10 * the dangers of modifying code on the run.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/spinlock.h>
16 #include <linux/hardirq.h>
17 #include <linux/uaccess.h>
18 #include <linux/ftrace.h>
19 #include <linux/percpu.h>
20 #include <linux/sched.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/list.h>
24 #include <linux/module.h>
25 #include <linux/memory.h>
26 #include <linux/vmalloc.h>
28 #include <trace/syscall.h>
30 #include <asm/set_memory.h>
31 #include <asm/kprobes.h>
32 #include <asm/ftrace.h>
34 #include <asm/text-patching.h>
36 #ifdef CONFIG_DYNAMIC_FTRACE
38 static int ftrace_poke_late
= 0;
40 int ftrace_arch_code_modify_prepare(void)
41 __acquires(&text_mutex
)
44 * Need to grab text_mutex to prevent a race from module loading
45 * and live kernel patching from changing the text permissions while
46 * ftrace has it set to "read/write".
48 mutex_lock(&text_mutex
);
53 int ftrace_arch_code_modify_post_process(void)
54 __releases(&text_mutex
)
57 * ftrace_make_{call,nop}() may be called during
58 * module load, and we need to finish the text_poke_queue()
63 mutex_unlock(&text_mutex
);
67 static const char *ftrace_nop_replace(void)
72 static const char *ftrace_call_replace(unsigned long ip
, unsigned long addr
)
74 return text_gen_insn(CALL_INSN_OPCODE
, (void *)ip
, (void *)addr
);
77 static int ftrace_verify_code(unsigned long ip
, const char *old_code
)
79 char cur_code
[MCOUNT_INSN_SIZE
];
83 * We are paranoid about modifying text, as if a bug was to happen, it
84 * could cause us to read or write to someplace that could cause harm.
85 * Carefully read and modify the code with probe_kernel_*(), and make
86 * sure what we read is what we expected it to be before modifying it.
88 /* read the text we want to modify */
89 if (copy_from_kernel_nofault(cur_code
, (void *)ip
, MCOUNT_INSN_SIZE
)) {
94 /* Make sure it is what we expect it to be */
95 if (memcmp(cur_code
, old_code
, MCOUNT_INSN_SIZE
) != 0) {
104 * Marked __ref because it calls text_poke_early() which is .init.text. That is
105 * ok because that call will happen early, during boot, when .init sections are
109 ftrace_modify_code_direct(unsigned long ip
, const char *old_code
,
110 const char *new_code
)
112 int ret
= ftrace_verify_code(ip
, old_code
);
116 /* replace the text with the new text */
117 if (ftrace_poke_late
)
118 text_poke_queue((void *)ip
, new_code
, MCOUNT_INSN_SIZE
, NULL
);
120 text_poke_early((void *)ip
, new_code
, MCOUNT_INSN_SIZE
);
124 int ftrace_make_nop(struct module
*mod
, struct dyn_ftrace
*rec
, unsigned long addr
)
126 unsigned long ip
= rec
->ip
;
127 const char *new, *old
;
129 old
= ftrace_call_replace(ip
, addr
);
130 new = ftrace_nop_replace();
133 * On boot up, and when modules are loaded, the MCOUNT_ADDR
134 * is converted to a nop, and will never become MCOUNT_ADDR
135 * again. This code is either running before SMP (on boot up)
136 * or before the code will ever be executed (module load).
137 * We do not want to use the breakpoint version in this case,
138 * just modify the code directly.
140 if (addr
== MCOUNT_ADDR
)
141 return ftrace_modify_code_direct(ip
, old
, new);
144 * x86 overrides ftrace_replace_code -- this function will never be used
147 WARN_ONCE(1, "invalid use of ftrace_make_nop");
151 int ftrace_make_call(struct dyn_ftrace
*rec
, unsigned long addr
)
153 unsigned long ip
= rec
->ip
;
154 const char *new, *old
;
156 old
= ftrace_nop_replace();
157 new = ftrace_call_replace(ip
, addr
);
159 /* Should only be called when module is loaded */
160 return ftrace_modify_code_direct(rec
->ip
, old
, new);
164 * Should never be called:
165 * As it is only called by __ftrace_replace_code() which is called by
166 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
167 * which is called to turn mcount into nops or nops into function calls
168 * but not to convert a function from not using regs to one that uses
169 * regs, which ftrace_modify_call() is for.
171 int ftrace_modify_call(struct dyn_ftrace
*rec
, unsigned long old_addr
,
178 int ftrace_update_ftrace_func(ftrace_func_t func
)
183 ip
= (unsigned long)(&ftrace_call
);
184 new = ftrace_call_replace(ip
, (unsigned long)func
);
185 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
187 ip
= (unsigned long)(&ftrace_regs_call
);
188 new = ftrace_call_replace(ip
, (unsigned long)func
);
189 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
194 void ftrace_replace_code(int enable
)
196 struct ftrace_rec_iter
*iter
;
197 struct dyn_ftrace
*rec
;
198 const char *new, *old
;
201 for_ftrace_rec_iter(iter
) {
202 rec
= ftrace_rec_iter_record(iter
);
204 switch (ftrace_test_record(rec
, enable
)) {
205 case FTRACE_UPDATE_IGNORE
:
209 case FTRACE_UPDATE_MAKE_CALL
:
210 old
= ftrace_nop_replace();
213 case FTRACE_UPDATE_MODIFY_CALL
:
214 case FTRACE_UPDATE_MAKE_NOP
:
215 old
= ftrace_call_replace(rec
->ip
, ftrace_get_addr_curr(rec
));
219 ret
= ftrace_verify_code(rec
->ip
, old
);
221 ftrace_bug(ret
, rec
);
226 for_ftrace_rec_iter(iter
) {
227 rec
= ftrace_rec_iter_record(iter
);
229 switch (ftrace_test_record(rec
, enable
)) {
230 case FTRACE_UPDATE_IGNORE
:
234 case FTRACE_UPDATE_MAKE_CALL
:
235 case FTRACE_UPDATE_MODIFY_CALL
:
236 new = ftrace_call_replace(rec
->ip
, ftrace_get_addr_new(rec
));
239 case FTRACE_UPDATE_MAKE_NOP
:
240 new = ftrace_nop_replace();
244 text_poke_queue((void *)rec
->ip
, new, MCOUNT_INSN_SIZE
, NULL
);
245 ftrace_update_record(rec
, enable
);
250 void arch_ftrace_update_code(int command
)
252 ftrace_modify_all_code(command
);
255 int __init
ftrace_dyn_arch_init(void)
260 /* Currently only x86_64 supports dynamic trampolines */
263 #ifdef CONFIG_MODULES
264 #include <linux/moduleloader.h>
265 /* Module allocation simplifies allocating memory for code */
266 static inline void *alloc_tramp(unsigned long size
)
268 return module_alloc(size
);
270 static inline void tramp_free(void *tramp
)
272 module_memfree(tramp
);
275 /* Trampolines can only be created if modules are supported */
276 static inline void *alloc_tramp(unsigned long size
)
280 static inline void tramp_free(void *tramp
) { }
283 /* Defined as markers to the end of the ftrace default trampolines */
284 extern void ftrace_regs_caller_end(void);
285 extern void ftrace_regs_caller_ret(void);
286 extern void ftrace_caller_end(void);
287 extern void ftrace_caller_op_ptr(void);
288 extern void ftrace_regs_caller_op_ptr(void);
289 extern void ftrace_regs_caller_jmp(void);
291 /* movq function_trace_op(%rip), %rdx */
292 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
293 #define OP_REF_SIZE 7
296 * The ftrace_ops is passed to the function callback. Since the
297 * trampoline only services a single ftrace_ops, we can pass in
300 * The ftrace_op_code_union is used to create a pointer to the
301 * ftrace_ops that will be passed to the callback function.
303 union ftrace_op_code_union
{
304 char code
[OP_REF_SIZE
];
308 } __attribute__((packed
));
314 create_trampoline(struct ftrace_ops
*ops
, unsigned int *tramp_size
)
316 unsigned long start_offset
;
317 unsigned long end_offset
;
318 unsigned long op_offset
;
319 unsigned long call_offset
;
320 unsigned long jmp_offset
;
321 unsigned long offset
;
322 unsigned long npages
;
328 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
329 unsigned const char op_ref
[] = { 0x48, 0x8b, 0x15 };
330 union ftrace_op_code_union op_ptr
;
333 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
334 start_offset
= (unsigned long)ftrace_regs_caller
;
335 end_offset
= (unsigned long)ftrace_regs_caller_end
;
336 op_offset
= (unsigned long)ftrace_regs_caller_op_ptr
;
337 call_offset
= (unsigned long)ftrace_regs_call
;
338 jmp_offset
= (unsigned long)ftrace_regs_caller_jmp
;
340 start_offset
= (unsigned long)ftrace_caller
;
341 end_offset
= (unsigned long)ftrace_caller_end
;
342 op_offset
= (unsigned long)ftrace_caller_op_ptr
;
343 call_offset
= (unsigned long)ftrace_call
;
347 size
= end_offset
- start_offset
;
350 * Allocate enough size to store the ftrace_caller code,
351 * the iret , as well as the address of the ftrace_ops this
352 * trampoline is used for.
354 trampoline
= alloc_tramp(size
+ RET_SIZE
+ sizeof(void *));
358 *tramp_size
= size
+ RET_SIZE
+ sizeof(void *);
359 npages
= DIV_ROUND_UP(*tramp_size
, PAGE_SIZE
);
361 /* Copy ftrace_caller onto the trampoline memory */
362 ret
= copy_from_kernel_nofault(trampoline
, (void *)start_offset
, size
);
363 if (WARN_ON(ret
< 0))
366 ip
= trampoline
+ size
;
368 /* The trampoline ends with ret(q) */
369 retq
= (unsigned long)ftrace_stub
;
370 ret
= copy_from_kernel_nofault(ip
, (void *)retq
, RET_SIZE
);
371 if (WARN_ON(ret
< 0))
374 /* No need to test direct calls on created trampolines */
375 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
376 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
377 ip
= trampoline
+ (jmp_offset
- start_offset
);
378 if (WARN_ON(*(char *)ip
!= 0x75))
380 ret
= copy_from_kernel_nofault(ip
, x86_nops
[2], 2);
386 * The address of the ftrace_ops that is used for this trampoline
387 * is stored at the end of the trampoline. This will be used to
388 * load the third parameter for the callback. Basically, that
389 * location at the end of the trampoline takes the place of
390 * the global function_trace_op variable.
393 ptr
= (unsigned long *)(trampoline
+ size
+ RET_SIZE
);
394 *ptr
= (unsigned long)ops
;
396 op_offset
-= start_offset
;
397 memcpy(&op_ptr
, trampoline
+ op_offset
, OP_REF_SIZE
);
399 /* Are we pointing to the reference? */
400 if (WARN_ON(memcmp(op_ptr
.op
, op_ref
, 3) != 0))
403 /* Load the contents of ptr into the callback parameter */
404 offset
= (unsigned long)ptr
;
405 offset
-= (unsigned long)trampoline
+ op_offset
+ OP_REF_SIZE
;
407 op_ptr
.offset
= offset
;
409 /* put in the new offset to the ftrace_ops */
410 memcpy(trampoline
+ op_offset
, &op_ptr
, OP_REF_SIZE
);
412 /* put in the call to the function */
413 mutex_lock(&text_mutex
);
414 call_offset
-= start_offset
;
415 memcpy(trampoline
+ call_offset
,
416 text_gen_insn(CALL_INSN_OPCODE
,
417 trampoline
+ call_offset
,
418 ftrace_ops_get_func(ops
)), CALL_INSN_SIZE
);
419 mutex_unlock(&text_mutex
);
421 /* ALLOC_TRAMP flags lets us know we created it */
422 ops
->flags
|= FTRACE_OPS_FL_ALLOC_TRAMP
;
424 set_vm_flush_reset_perms(trampoline
);
426 if (likely(system_state
!= SYSTEM_BOOTING
))
427 set_memory_ro((unsigned long)trampoline
, npages
);
428 set_memory_x((unsigned long)trampoline
, npages
);
429 return (unsigned long)trampoline
;
431 tramp_free(trampoline
);
435 void set_ftrace_ops_ro(void)
437 struct ftrace_ops
*ops
;
438 unsigned long start_offset
;
439 unsigned long end_offset
;
440 unsigned long npages
;
443 do_for_each_ftrace_op(ops
, ftrace_ops_list
) {
444 if (!(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
447 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
448 start_offset
= (unsigned long)ftrace_regs_caller
;
449 end_offset
= (unsigned long)ftrace_regs_caller_end
;
451 start_offset
= (unsigned long)ftrace_caller
;
452 end_offset
= (unsigned long)ftrace_caller_end
;
454 size
= end_offset
- start_offset
;
455 size
= size
+ RET_SIZE
+ sizeof(void *);
456 npages
= DIV_ROUND_UP(size
, PAGE_SIZE
);
457 set_memory_ro((unsigned long)ops
->trampoline
, npages
);
458 } while_for_each_ftrace_op(ops
);
461 static unsigned long calc_trampoline_call_offset(bool save_regs
)
463 unsigned long start_offset
;
464 unsigned long call_offset
;
467 start_offset
= (unsigned long)ftrace_regs_caller
;
468 call_offset
= (unsigned long)ftrace_regs_call
;
470 start_offset
= (unsigned long)ftrace_caller
;
471 call_offset
= (unsigned long)ftrace_call
;
474 return call_offset
- start_offset
;
477 void arch_ftrace_update_trampoline(struct ftrace_ops
*ops
)
480 unsigned long offset
;
485 if (!ops
->trampoline
) {
486 ops
->trampoline
= create_trampoline(ops
, &size
);
487 if (!ops
->trampoline
)
489 ops
->trampoline_size
= size
;
494 * The ftrace_ops caller may set up its own trampoline.
495 * In such a case, this code must not modify it.
497 if (!(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
500 offset
= calc_trampoline_call_offset(ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
);
501 ip
= ops
->trampoline
+ offset
;
502 func
= ftrace_ops_get_func(ops
);
504 mutex_lock(&text_mutex
);
505 /* Do a safe modify in case the trampoline is executing */
506 new = ftrace_call_replace(ip
, (unsigned long)func
);
507 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
508 mutex_unlock(&text_mutex
);
511 /* Return the address of the function the trampoline calls */
512 static void *addr_from_call(void *ptr
)
514 union text_poke_insn call
;
517 ret
= copy_from_kernel_nofault(&call
, ptr
, CALL_INSN_SIZE
);
518 if (WARN_ON_ONCE(ret
< 0))
521 /* Make sure this is a call */
522 if (WARN_ON_ONCE(call
.opcode
!= CALL_INSN_OPCODE
)) {
523 pr_warn("Expected E8, got %x\n", call
.opcode
);
527 return ptr
+ CALL_INSN_SIZE
+ call
.disp
;
530 void prepare_ftrace_return(unsigned long self_addr
, unsigned long *parent
,
531 unsigned long frame_pointer
);
534 * If the ops->trampoline was not allocated, then it probably
535 * has a static trampoline func, or is the ftrace caller itself.
537 static void *static_tramp_func(struct ftrace_ops
*ops
, struct dyn_ftrace
*rec
)
539 unsigned long offset
;
540 bool save_regs
= rec
->flags
& FTRACE_FL_REGS_EN
;
543 if (ops
&& ops
->trampoline
) {
544 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
546 * We only know about function graph tracer setting as static
549 if (ops
->trampoline
== FTRACE_GRAPH_ADDR
)
550 return (void *)prepare_ftrace_return
;
555 offset
= calc_trampoline_call_offset(save_regs
);
558 ptr
= (void *)FTRACE_REGS_ADDR
+ offset
;
560 ptr
= (void *)FTRACE_ADDR
+ offset
;
562 return addr_from_call(ptr
);
565 void *arch_ftrace_trampoline_func(struct ftrace_ops
*ops
, struct dyn_ftrace
*rec
)
567 unsigned long offset
;
569 /* If we didn't allocate this trampoline, consider it static */
570 if (!ops
|| !(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
571 return static_tramp_func(ops
, rec
);
573 offset
= calc_trampoline_call_offset(ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
);
574 return addr_from_call((void *)ops
->trampoline
+ offset
);
577 void arch_ftrace_trampoline_free(struct ftrace_ops
*ops
)
579 if (!ops
|| !(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
582 tramp_free((void *)ops
->trampoline
);
586 #endif /* CONFIG_X86_64 */
587 #endif /* CONFIG_DYNAMIC_FTRACE */
589 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
591 #ifdef CONFIG_DYNAMIC_FTRACE
592 extern void ftrace_graph_call(void);
594 static const char *ftrace_jmp_replace(unsigned long ip
, unsigned long addr
)
596 return text_gen_insn(JMP32_INSN_OPCODE
, (void *)ip
, (void *)addr
);
599 static int ftrace_mod_jmp(unsigned long ip
, void *func
)
603 new = ftrace_jmp_replace(ip
, (unsigned long)func
);
604 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
608 int ftrace_enable_ftrace_graph_caller(void)
610 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
612 return ftrace_mod_jmp(ip
, &ftrace_graph_caller
);
615 int ftrace_disable_ftrace_graph_caller(void)
617 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
619 return ftrace_mod_jmp(ip
, &ftrace_stub
);
622 #endif /* !CONFIG_DYNAMIC_FTRACE */
625 * Hook the return address and push it in the stack of return addrs
626 * in current thread info.
628 void prepare_ftrace_return(unsigned long self_addr
, unsigned long *parent
,
629 unsigned long frame_pointer
)
631 unsigned long return_hooker
= (unsigned long)&return_to_handler
;
636 * When resuming from suspend-to-ram, this function can be indirectly
637 * called from early CPU startup code while the CPU is in real mode,
638 * which would fail miserably. Make sure the stack pointer is a
641 * This check isn't as accurate as virt_addr_valid(), but it should be
642 * good enough for this purpose, and it's fast.
644 if (unlikely((long)__builtin_frame_address(0) >= 0))
647 if (unlikely(ftrace_graph_is_dead()))
650 if (unlikely(atomic_read(¤t
->tracing_graph_pause
)))
654 * Protect against fault, even if it shouldn't
655 * happen. This tool is too much intrusive to
656 * ignore such a protection.
659 "1: " _ASM_MOV
" (%[parent]), %[old]\n"
660 "2: " _ASM_MOV
" %[return_hooker], (%[parent])\n"
661 " movl $0, %[faulted]\n"
664 ".section .fixup, \"ax\"\n"
665 "4: movl $1, %[faulted]\n"
672 : [old
] "=&r" (old
), [faulted
] "=r" (faulted
)
673 : [parent
] "r" (parent
), [return_hooker
] "r" (return_hooker
)
677 if (unlikely(faulted
)) {
683 if (function_graph_enter(old
, self_addr
, frame_pointer
, parent
))
686 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */