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>
27 #include <linux/set_memory.h>
28 #include <linux/execmem.h>
30 #include <trace/syscall.h>
32 #include <asm/kprobes.h>
33 #include <asm/ftrace.h>
35 #include <asm/text-patching.h>
37 #ifdef CONFIG_DYNAMIC_FTRACE
39 static int ftrace_poke_late
= 0;
41 void ftrace_arch_code_modify_prepare(void)
42 __acquires(&text_mutex
)
45 * Need to grab text_mutex to prevent a race from module loading
46 * and live kernel patching from changing the text permissions while
47 * ftrace has it set to "read/write".
49 mutex_lock(&text_mutex
);
53 void 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
);
66 static const char *ftrace_nop_replace(void)
71 static const char *ftrace_call_replace(unsigned long ip
, unsigned long addr
)
74 * No need to translate into a callthunk. The trampoline does
75 * the depth accounting itself.
77 return text_gen_insn(CALL_INSN_OPCODE
, (void *)ip
, (void *)addr
);
80 static int ftrace_verify_code(unsigned long ip
, const char *old_code
)
82 char cur_code
[MCOUNT_INSN_SIZE
];
86 * We are paranoid about modifying text, as if a bug was to happen, it
87 * could cause us to read or write to someplace that could cause harm.
88 * Carefully read and modify the code with probe_kernel_*(), and make
89 * sure what we read is what we expected it to be before modifying it.
91 /* read the text we want to modify */
92 if (copy_from_kernel_nofault(cur_code
, (void *)ip
, MCOUNT_INSN_SIZE
)) {
97 /* Make sure it is what we expect it to be */
98 if (memcmp(cur_code
, old_code
, MCOUNT_INSN_SIZE
) != 0) {
99 ftrace_expected
= old_code
;
108 * Marked __ref because it calls text_poke_early() which is .init.text. That is
109 * ok because that call will happen early, during boot, when .init sections are
113 ftrace_modify_code_direct(unsigned long ip
, const char *old_code
,
114 const char *new_code
)
116 int ret
= ftrace_verify_code(ip
, old_code
);
120 /* replace the text with the new text */
121 if (ftrace_poke_late
)
122 text_poke_queue((void *)ip
, new_code
, MCOUNT_INSN_SIZE
, NULL
);
124 text_poke_early((void *)ip
, new_code
, MCOUNT_INSN_SIZE
);
128 int ftrace_make_nop(struct module
*mod
, struct dyn_ftrace
*rec
, unsigned long addr
)
130 unsigned long ip
= rec
->ip
;
131 const char *new, *old
;
133 old
= ftrace_call_replace(ip
, addr
);
134 new = ftrace_nop_replace();
137 * On boot up, and when modules are loaded, the MCOUNT_ADDR
138 * is converted to a nop, and will never become MCOUNT_ADDR
139 * again. This code is either running before SMP (on boot up)
140 * or before the code will ever be executed (module load).
141 * We do not want to use the breakpoint version in this case,
142 * just modify the code directly.
144 if (addr
== MCOUNT_ADDR
)
145 return ftrace_modify_code_direct(ip
, old
, new);
148 * x86 overrides ftrace_replace_code -- this function will never be used
151 WARN_ONCE(1, "invalid use of ftrace_make_nop");
155 int ftrace_make_call(struct dyn_ftrace
*rec
, unsigned long addr
)
157 unsigned long ip
= rec
->ip
;
158 const char *new, *old
;
160 old
= ftrace_nop_replace();
161 new = ftrace_call_replace(ip
, addr
);
163 /* Should only be called when module is loaded */
164 return ftrace_modify_code_direct(rec
->ip
, old
, new);
168 * Should never be called:
169 * As it is only called by __ftrace_replace_code() which is called by
170 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
171 * which is called to turn mcount into nops or nops into function calls
172 * but not to convert a function from not using regs to one that uses
173 * regs, which ftrace_modify_call() is for.
175 int ftrace_modify_call(struct dyn_ftrace
*rec
, unsigned long old_addr
,
182 int ftrace_update_ftrace_func(ftrace_func_t func
)
187 ip
= (unsigned long)(&ftrace_call
);
188 new = ftrace_call_replace(ip
, (unsigned long)func
);
189 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
191 ip
= (unsigned long)(&ftrace_regs_call
);
192 new = ftrace_call_replace(ip
, (unsigned long)func
);
193 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
198 void ftrace_replace_code(int enable
)
200 struct ftrace_rec_iter
*iter
;
201 struct dyn_ftrace
*rec
;
202 const char *new, *old
;
205 for_ftrace_rec_iter(iter
) {
206 rec
= ftrace_rec_iter_record(iter
);
208 switch (ftrace_test_record(rec
, enable
)) {
209 case FTRACE_UPDATE_IGNORE
:
213 case FTRACE_UPDATE_MAKE_CALL
:
214 old
= ftrace_nop_replace();
217 case FTRACE_UPDATE_MODIFY_CALL
:
218 case FTRACE_UPDATE_MAKE_NOP
:
219 old
= ftrace_call_replace(rec
->ip
, ftrace_get_addr_curr(rec
));
223 ret
= ftrace_verify_code(rec
->ip
, old
);
225 ftrace_expected
= old
;
226 ftrace_bug(ret
, rec
);
227 ftrace_expected
= NULL
;
232 for_ftrace_rec_iter(iter
) {
233 rec
= ftrace_rec_iter_record(iter
);
235 switch (ftrace_test_record(rec
, enable
)) {
236 case FTRACE_UPDATE_IGNORE
:
240 case FTRACE_UPDATE_MAKE_CALL
:
241 case FTRACE_UPDATE_MODIFY_CALL
:
242 new = ftrace_call_replace(rec
->ip
, ftrace_get_addr_new(rec
));
245 case FTRACE_UPDATE_MAKE_NOP
:
246 new = ftrace_nop_replace();
250 text_poke_queue((void *)rec
->ip
, new, MCOUNT_INSN_SIZE
, NULL
);
251 ftrace_update_record(rec
, enable
);
256 void arch_ftrace_update_code(int command
)
258 ftrace_modify_all_code(command
);
261 /* Currently only x86_64 supports dynamic trampolines */
264 static inline void *alloc_tramp(unsigned long size
)
266 return execmem_alloc(EXECMEM_FTRACE
, size
);
268 static inline void tramp_free(void *tramp
)
273 /* Defined as markers to the end of the ftrace default trampolines */
274 extern void ftrace_regs_caller_end(void);
275 extern void ftrace_caller_end(void);
276 extern void ftrace_caller_op_ptr(void);
277 extern void ftrace_regs_caller_op_ptr(void);
278 extern void ftrace_regs_caller_jmp(void);
280 /* movq function_trace_op(%rip), %rdx */
281 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
282 #define OP_REF_SIZE 7
285 * The ftrace_ops is passed to the function callback. Since the
286 * trampoline only services a single ftrace_ops, we can pass in
289 * The ftrace_op_code_union is used to create a pointer to the
290 * ftrace_ops that will be passed to the callback function.
292 union ftrace_op_code_union
{
293 char code
[OP_REF_SIZE
];
297 } __attribute__((packed
));
301 (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) ? 5 : 1 + IS_ENABLED(CONFIG_MITIGATION_SLS))
304 create_trampoline(struct ftrace_ops
*ops
, unsigned int *tramp_size
)
306 unsigned long start_offset
;
307 unsigned long end_offset
;
308 unsigned long op_offset
;
309 unsigned long call_offset
;
310 unsigned long jmp_offset
;
311 unsigned long offset
;
312 unsigned long npages
;
317 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
318 unsigned const char op_ref
[] = { 0x48, 0x8b, 0x15 };
319 unsigned const char retq
[] = { RET_INSN_OPCODE
, INT3_INSN_OPCODE
};
320 union ftrace_op_code_union op_ptr
;
323 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
324 start_offset
= (unsigned long)ftrace_regs_caller
;
325 end_offset
= (unsigned long)ftrace_regs_caller_end
;
326 op_offset
= (unsigned long)ftrace_regs_caller_op_ptr
;
327 call_offset
= (unsigned long)ftrace_regs_call
;
328 jmp_offset
= (unsigned long)ftrace_regs_caller_jmp
;
330 start_offset
= (unsigned long)ftrace_caller
;
331 end_offset
= (unsigned long)ftrace_caller_end
;
332 op_offset
= (unsigned long)ftrace_caller_op_ptr
;
333 call_offset
= (unsigned long)ftrace_call
;
337 size
= end_offset
- start_offset
;
340 * Allocate enough size to store the ftrace_caller code,
341 * the iret , as well as the address of the ftrace_ops this
342 * trampoline is used for.
344 trampoline
= alloc_tramp(size
+ RET_SIZE
+ sizeof(void *));
348 *tramp_size
= size
+ RET_SIZE
+ sizeof(void *);
349 npages
= DIV_ROUND_UP(*tramp_size
, PAGE_SIZE
);
351 /* Copy ftrace_caller onto the trampoline memory */
352 ret
= copy_from_kernel_nofault(trampoline
, (void *)start_offset
, size
);
353 if (WARN_ON(ret
< 0))
356 ip
= trampoline
+ size
;
357 if (cpu_feature_enabled(X86_FEATURE_RETHUNK
))
358 __text_gen_insn(ip
, JMP32_INSN_OPCODE
, ip
, x86_return_thunk
, JMP32_INSN_SIZE
);
360 memcpy(ip
, retq
, sizeof(retq
));
362 /* No need to test direct calls on created trampolines */
363 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
364 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
365 ip
= trampoline
+ (jmp_offset
- start_offset
);
366 if (WARN_ON(*(char *)ip
!= 0x75))
368 ret
= copy_from_kernel_nofault(ip
, x86_nops
[2], 2);
374 * The address of the ftrace_ops that is used for this trampoline
375 * is stored at the end of the trampoline. This will be used to
376 * load the third parameter for the callback. Basically, that
377 * location at the end of the trampoline takes the place of
378 * the global function_trace_op variable.
381 ptr
= (unsigned long *)(trampoline
+ size
+ RET_SIZE
);
382 *ptr
= (unsigned long)ops
;
384 op_offset
-= start_offset
;
385 memcpy(&op_ptr
, trampoline
+ op_offset
, OP_REF_SIZE
);
387 /* Are we pointing to the reference? */
388 if (WARN_ON(memcmp(op_ptr
.op
, op_ref
, 3) != 0))
391 /* Load the contents of ptr into the callback parameter */
392 offset
= (unsigned long)ptr
;
393 offset
-= (unsigned long)trampoline
+ op_offset
+ OP_REF_SIZE
;
395 op_ptr
.offset
= offset
;
397 /* put in the new offset to the ftrace_ops */
398 memcpy(trampoline
+ op_offset
, &op_ptr
, OP_REF_SIZE
);
400 /* put in the call to the function */
401 mutex_lock(&text_mutex
);
402 call_offset
-= start_offset
;
404 * No need to translate into a callthunk. The trampoline does
405 * the depth accounting before the call already.
407 dest
= ftrace_ops_get_func(ops
);
408 memcpy(trampoline
+ call_offset
,
409 text_gen_insn(CALL_INSN_OPCODE
, trampoline
+ call_offset
, dest
),
411 mutex_unlock(&text_mutex
);
413 /* ALLOC_TRAMP flags lets us know we created it */
414 ops
->flags
|= FTRACE_OPS_FL_ALLOC_TRAMP
;
416 set_memory_rox((unsigned long)trampoline
, npages
);
417 return (unsigned long)trampoline
;
419 tramp_free(trampoline
);
423 void set_ftrace_ops_ro(void)
425 struct ftrace_ops
*ops
;
426 unsigned long start_offset
;
427 unsigned long end_offset
;
428 unsigned long npages
;
431 do_for_each_ftrace_op(ops
, ftrace_ops_list
) {
432 if (!(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
435 if (ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
) {
436 start_offset
= (unsigned long)ftrace_regs_caller
;
437 end_offset
= (unsigned long)ftrace_regs_caller_end
;
439 start_offset
= (unsigned long)ftrace_caller
;
440 end_offset
= (unsigned long)ftrace_caller_end
;
442 size
= end_offset
- start_offset
;
443 size
= size
+ RET_SIZE
+ sizeof(void *);
444 npages
= DIV_ROUND_UP(size
, PAGE_SIZE
);
445 set_memory_ro((unsigned long)ops
->trampoline
, npages
);
446 } while_for_each_ftrace_op(ops
);
449 static unsigned long calc_trampoline_call_offset(bool save_regs
)
451 unsigned long start_offset
;
452 unsigned long call_offset
;
455 start_offset
= (unsigned long)ftrace_regs_caller
;
456 call_offset
= (unsigned long)ftrace_regs_call
;
458 start_offset
= (unsigned long)ftrace_caller
;
459 call_offset
= (unsigned long)ftrace_call
;
462 return call_offset
- start_offset
;
465 void arch_ftrace_update_trampoline(struct ftrace_ops
*ops
)
468 unsigned long offset
;
473 if (!ops
->trampoline
) {
474 ops
->trampoline
= create_trampoline(ops
, &size
);
475 if (!ops
->trampoline
)
477 ops
->trampoline_size
= size
;
482 * The ftrace_ops caller may set up its own trampoline.
483 * In such a case, this code must not modify it.
485 if (!(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
488 offset
= calc_trampoline_call_offset(ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
);
489 ip
= ops
->trampoline
+ offset
;
490 func
= ftrace_ops_get_func(ops
);
492 mutex_lock(&text_mutex
);
493 /* Do a safe modify in case the trampoline is executing */
494 new = ftrace_call_replace(ip
, (unsigned long)func
);
495 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
496 mutex_unlock(&text_mutex
);
499 /* Return the address of the function the trampoline calls */
500 static void *addr_from_call(void *ptr
)
502 union text_poke_insn call
;
505 ret
= copy_from_kernel_nofault(&call
, ptr
, CALL_INSN_SIZE
);
506 if (WARN_ON_ONCE(ret
< 0))
509 /* Make sure this is a call */
510 if (WARN_ON_ONCE(call
.opcode
!= CALL_INSN_OPCODE
)) {
511 pr_warn("Expected E8, got %x\n", call
.opcode
);
515 return ptr
+ CALL_INSN_SIZE
+ call
.disp
;
519 * If the ops->trampoline was not allocated, then it probably
520 * has a static trampoline func, or is the ftrace caller itself.
522 static void *static_tramp_func(struct ftrace_ops
*ops
, struct dyn_ftrace
*rec
)
524 unsigned long offset
;
525 bool save_regs
= rec
->flags
& FTRACE_FL_REGS_EN
;
528 if (ops
&& ops
->trampoline
) {
529 #if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \
530 defined(CONFIG_FUNCTION_GRAPH_TRACER)
532 * We only know about function graph tracer setting as static
535 if (ops
->trampoline
== FTRACE_GRAPH_ADDR
)
536 return (void *)prepare_ftrace_return
;
541 offset
= calc_trampoline_call_offset(save_regs
);
544 ptr
= (void *)FTRACE_REGS_ADDR
+ offset
;
546 ptr
= (void *)FTRACE_ADDR
+ offset
;
548 return addr_from_call(ptr
);
551 void *arch_ftrace_trampoline_func(struct ftrace_ops
*ops
, struct dyn_ftrace
*rec
)
553 unsigned long offset
;
555 /* If we didn't allocate this trampoline, consider it static */
556 if (!ops
|| !(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
557 return static_tramp_func(ops
, rec
);
559 offset
= calc_trampoline_call_offset(ops
->flags
& FTRACE_OPS_FL_SAVE_REGS
);
560 return addr_from_call((void *)ops
->trampoline
+ offset
);
563 void arch_ftrace_trampoline_free(struct ftrace_ops
*ops
)
565 if (!ops
|| !(ops
->flags
& FTRACE_OPS_FL_ALLOC_TRAMP
))
568 tramp_free((void *)ops
->trampoline
);
572 #endif /* CONFIG_X86_64 */
573 #endif /* CONFIG_DYNAMIC_FTRACE */
575 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
577 #if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS)
578 extern void ftrace_graph_call(void);
579 static const char *ftrace_jmp_replace(unsigned long ip
, unsigned long addr
)
581 return text_gen_insn(JMP32_INSN_OPCODE
, (void *)ip
, (void *)addr
);
584 static int ftrace_mod_jmp(unsigned long ip
, void *func
)
588 new = ftrace_jmp_replace(ip
, (unsigned long)func
);
589 text_poke_bp((void *)ip
, new, MCOUNT_INSN_SIZE
, NULL
);
593 int ftrace_enable_ftrace_graph_caller(void)
595 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
597 return ftrace_mod_jmp(ip
, &ftrace_graph_caller
);
600 int ftrace_disable_ftrace_graph_caller(void)
602 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
604 return ftrace_mod_jmp(ip
, &ftrace_stub
);
606 #endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */
609 * Hook the return address and push it in the stack of return addrs
610 * in current thread info.
612 void prepare_ftrace_return(unsigned long ip
, unsigned long *parent
,
613 unsigned long frame_pointer
)
615 unsigned long return_hooker
= (unsigned long)&return_to_handler
;
619 * When resuming from suspend-to-ram, this function can be indirectly
620 * called from early CPU startup code while the CPU is in real mode,
621 * which would fail miserably. Make sure the stack pointer is a
624 * This check isn't as accurate as virt_addr_valid(), but it should be
625 * good enough for this purpose, and it's fast.
627 if (unlikely((long)__builtin_frame_address(0) >= 0))
630 if (unlikely(ftrace_graph_is_dead()))
633 if (unlikely(atomic_read(¤t
->tracing_graph_pause
)))
636 bit
= ftrace_test_recursion_trylock(ip
, *parent
);
640 if (!function_graph_enter(*parent
, ip
, frame_pointer
, parent
))
641 *parent
= return_hooker
;
643 ftrace_test_recursion_unlock(bit
);
646 #ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
647 void ftrace_graph_func(unsigned long ip
, unsigned long parent_ip
,
648 struct ftrace_ops
*op
, struct ftrace_regs
*fregs
)
650 struct pt_regs
*regs
= &fregs
->regs
;
651 unsigned long *stack
= (unsigned long *)kernel_stack_pointer(regs
);
653 prepare_ftrace_return(ip
, (unsigned long *)stack
, 0);
657 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */