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 void 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);
52 void ftrace_arch_code_modify_post_process(void)
53 __releases(&text_mutex)
56 * ftrace_make_{call,nop}() may be called during
57 * module load, and we need to finish the text_poke_queue()
62 mutex_unlock(&text_mutex);
65 static const char *ftrace_nop_replace(void)
70 static const char *ftrace_call_replace(unsigned long ip, unsigned long addr)
72 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
75 static int ftrace_verify_code(unsigned long ip, const char *old_code)
77 char cur_code[MCOUNT_INSN_SIZE];
81 * We are paranoid about modifying text, as if a bug was to happen, it
82 * could cause us to read or write to someplace that could cause harm.
83 * Carefully read and modify the code with probe_kernel_*(), and make
84 * sure what we read is what we expected it to be before modifying it.
86 /* read the text we want to modify */
87 if (copy_from_kernel_nofault(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) {
92 /* Make sure it is what we expect it to be */
93 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) {
94 ftrace_expected = old_code;
103 * Marked __ref because it calls text_poke_early() which is .init.text. That is
104 * ok because that call will happen early, during boot, when .init sections are
108 ftrace_modify_code_direct(unsigned long ip, const char *old_code,
109 const char *new_code)
111 int ret = ftrace_verify_code(ip, old_code);
115 /* replace the text with the new text */
116 if (ftrace_poke_late)
117 text_poke_queue((void *)ip, new_code, MCOUNT_INSN_SIZE, NULL);
119 text_poke_early((void *)ip, new_code, MCOUNT_INSN_SIZE);
123 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr)
125 unsigned long ip = rec->ip;
126 const char *new, *old;
128 old = ftrace_call_replace(ip, addr);
129 new = ftrace_nop_replace();
132 * On boot up, and when modules are loaded, the MCOUNT_ADDR
133 * is converted to a nop, and will never become MCOUNT_ADDR
134 * again. This code is either running before SMP (on boot up)
135 * or before the code will ever be executed (module load).
136 * We do not want to use the breakpoint version in this case,
137 * just modify the code directly.
139 if (addr == MCOUNT_ADDR)
140 return ftrace_modify_code_direct(ip, old, new);
143 * x86 overrides ftrace_replace_code -- this function will never be used
146 WARN_ONCE(1, "invalid use of ftrace_make_nop");
150 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
152 unsigned long ip = rec->ip;
153 const char *new, *old;
155 old = ftrace_nop_replace();
156 new = ftrace_call_replace(ip, addr);
158 /* Should only be called when module is loaded */
159 return ftrace_modify_code_direct(rec->ip, old, new);
163 * Should never be called:
164 * As it is only called by __ftrace_replace_code() which is called by
165 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
166 * which is called to turn mcount into nops or nops into function calls
167 * but not to convert a function from not using regs to one that uses
168 * regs, which ftrace_modify_call() is for.
170 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
177 int ftrace_update_ftrace_func(ftrace_func_t func)
182 ip = (unsigned long)(&ftrace_call);
183 new = ftrace_call_replace(ip, (unsigned long)func);
184 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
186 ip = (unsigned long)(&ftrace_regs_call);
187 new = ftrace_call_replace(ip, (unsigned long)func);
188 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
193 void ftrace_replace_code(int enable)
195 struct ftrace_rec_iter *iter;
196 struct dyn_ftrace *rec;
197 const char *new, *old;
200 for_ftrace_rec_iter(iter) {
201 rec = ftrace_rec_iter_record(iter);
203 switch (ftrace_test_record(rec, enable)) {
204 case FTRACE_UPDATE_IGNORE:
208 case FTRACE_UPDATE_MAKE_CALL:
209 old = ftrace_nop_replace();
212 case FTRACE_UPDATE_MODIFY_CALL:
213 case FTRACE_UPDATE_MAKE_NOP:
214 old = ftrace_call_replace(rec->ip, ftrace_get_addr_curr(rec));
218 ret = ftrace_verify_code(rec->ip, old);
220 ftrace_bug(ret, rec);
225 for_ftrace_rec_iter(iter) {
226 rec = ftrace_rec_iter_record(iter);
228 switch (ftrace_test_record(rec, enable)) {
229 case FTRACE_UPDATE_IGNORE:
233 case FTRACE_UPDATE_MAKE_CALL:
234 case FTRACE_UPDATE_MODIFY_CALL:
235 new = ftrace_call_replace(rec->ip, ftrace_get_addr_new(rec));
238 case FTRACE_UPDATE_MAKE_NOP:
239 new = ftrace_nop_replace();
243 text_poke_queue((void *)rec->ip, new, MCOUNT_INSN_SIZE, NULL);
244 ftrace_update_record(rec, enable);
249 void arch_ftrace_update_code(int command)
251 ftrace_modify_all_code(command);
254 /* Currently only x86_64 supports dynamic trampolines */
257 #ifdef CONFIG_MODULES
258 #include <linux/moduleloader.h>
259 /* Module allocation simplifies allocating memory for code */
260 static inline void *alloc_tramp(unsigned long size)
262 return module_alloc(size);
264 static inline void tramp_free(void *tramp)
266 module_memfree(tramp);
269 /* Trampolines can only be created if modules are supported */
270 static inline void *alloc_tramp(unsigned long size)
274 static inline void tramp_free(void *tramp) { }
277 /* Defined as markers to the end of the ftrace default trampolines */
278 extern void ftrace_regs_caller_end(void);
279 extern void ftrace_regs_caller_ret(void);
280 extern void ftrace_caller_end(void);
281 extern void ftrace_caller_op_ptr(void);
282 extern void ftrace_regs_caller_op_ptr(void);
283 extern void ftrace_regs_caller_jmp(void);
285 /* movq function_trace_op(%rip), %rdx */
286 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
287 #define OP_REF_SIZE 7
290 * The ftrace_ops is passed to the function callback. Since the
291 * trampoline only services a single ftrace_ops, we can pass in
294 * The ftrace_op_code_union is used to create a pointer to the
295 * ftrace_ops that will be passed to the callback function.
297 union ftrace_op_code_union {
298 char code[OP_REF_SIZE];
302 } __attribute__((packed));
305 #define RET_SIZE (IS_ENABLED(CONFIG_RETPOLINE) ? 5 : 1 + IS_ENABLED(CONFIG_SLS))
308 create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
310 unsigned long start_offset;
311 unsigned long end_offset;
312 unsigned long op_offset;
313 unsigned long call_offset;
314 unsigned long jmp_offset;
315 unsigned long offset;
316 unsigned long npages;
321 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
322 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
323 unsigned const char retq[] = { RET_INSN_OPCODE, INT3_INSN_OPCODE };
324 union ftrace_op_code_union op_ptr;
327 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
328 start_offset = (unsigned long)ftrace_regs_caller;
329 end_offset = (unsigned long)ftrace_regs_caller_end;
330 op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
331 call_offset = (unsigned long)ftrace_regs_call;
332 jmp_offset = (unsigned long)ftrace_regs_caller_jmp;
334 start_offset = (unsigned long)ftrace_caller;
335 end_offset = (unsigned long)ftrace_caller_end;
336 op_offset = (unsigned long)ftrace_caller_op_ptr;
337 call_offset = (unsigned long)ftrace_call;
341 size = end_offset - start_offset;
344 * Allocate enough size to store the ftrace_caller code,
345 * the iret , as well as the address of the ftrace_ops this
346 * trampoline is used for.
348 trampoline = alloc_tramp(size + RET_SIZE + sizeof(void *));
352 *tramp_size = size + RET_SIZE + sizeof(void *);
353 npages = DIV_ROUND_UP(*tramp_size, PAGE_SIZE);
355 /* Copy ftrace_caller onto the trampoline memory */
356 ret = copy_from_kernel_nofault(trampoline, (void *)start_offset, size);
357 if (WARN_ON(ret < 0))
360 ip = trampoline + size;
361 if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
362 __text_gen_insn(ip, JMP32_INSN_OPCODE, ip, &__x86_return_thunk, JMP32_INSN_SIZE);
364 memcpy(ip, retq, sizeof(retq));
366 /* No need to test direct calls on created trampolines */
367 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
368 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
369 ip = trampoline + (jmp_offset - start_offset);
370 if (WARN_ON(*(char *)ip != 0x75))
372 ret = copy_from_kernel_nofault(ip, x86_nops[2], 2);
378 * The address of the ftrace_ops that is used for this trampoline
379 * is stored at the end of the trampoline. This will be used to
380 * load the third parameter for the callback. Basically, that
381 * location at the end of the trampoline takes the place of
382 * the global function_trace_op variable.
385 ptr = (unsigned long *)(trampoline + size + RET_SIZE);
386 *ptr = (unsigned long)ops;
388 op_offset -= start_offset;
389 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
391 /* Are we pointing to the reference? */
392 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0))
395 /* Load the contents of ptr into the callback parameter */
396 offset = (unsigned long)ptr;
397 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
399 op_ptr.offset = offset;
401 /* put in the new offset to the ftrace_ops */
402 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
404 /* put in the call to the function */
405 mutex_lock(&text_mutex);
406 call_offset -= start_offset;
407 memcpy(trampoline + call_offset,
408 text_gen_insn(CALL_INSN_OPCODE,
409 trampoline + call_offset,
410 ftrace_ops_get_func(ops)), CALL_INSN_SIZE);
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_vm_flush_reset_perms(trampoline);
418 if (likely(system_state != SYSTEM_BOOTING))
419 set_memory_ro((unsigned long)trampoline, npages);
420 set_memory_x((unsigned long)trampoline, npages);
421 return (unsigned long)trampoline;
423 tramp_free(trampoline);
427 void set_ftrace_ops_ro(void)
429 struct ftrace_ops *ops;
430 unsigned long start_offset;
431 unsigned long end_offset;
432 unsigned long npages;
435 do_for_each_ftrace_op(ops, ftrace_ops_list) {
436 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
439 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
440 start_offset = (unsigned long)ftrace_regs_caller;
441 end_offset = (unsigned long)ftrace_regs_caller_end;
443 start_offset = (unsigned long)ftrace_caller;
444 end_offset = (unsigned long)ftrace_caller_end;
446 size = end_offset - start_offset;
447 size = size + RET_SIZE + sizeof(void *);
448 npages = DIV_ROUND_UP(size, PAGE_SIZE);
449 set_memory_ro((unsigned long)ops->trampoline, npages);
450 } while_for_each_ftrace_op(ops);
453 static unsigned long calc_trampoline_call_offset(bool save_regs)
455 unsigned long start_offset;
456 unsigned long call_offset;
459 start_offset = (unsigned long)ftrace_regs_caller;
460 call_offset = (unsigned long)ftrace_regs_call;
462 start_offset = (unsigned long)ftrace_caller;
463 call_offset = (unsigned long)ftrace_call;
466 return call_offset - start_offset;
469 void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
472 unsigned long offset;
477 if (!ops->trampoline) {
478 ops->trampoline = create_trampoline(ops, &size);
479 if (!ops->trampoline)
481 ops->trampoline_size = size;
486 * The ftrace_ops caller may set up its own trampoline.
487 * In such a case, this code must not modify it.
489 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
492 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
493 ip = ops->trampoline + offset;
494 func = ftrace_ops_get_func(ops);
496 mutex_lock(&text_mutex);
497 /* Do a safe modify in case the trampoline is executing */
498 new = ftrace_call_replace(ip, (unsigned long)func);
499 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
500 mutex_unlock(&text_mutex);
503 /* Return the address of the function the trampoline calls */
504 static void *addr_from_call(void *ptr)
506 union text_poke_insn call;
509 ret = copy_from_kernel_nofault(&call, ptr, CALL_INSN_SIZE);
510 if (WARN_ON_ONCE(ret < 0))
513 /* Make sure this is a call */
514 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) {
515 pr_warn("Expected E8, got %x\n", call.opcode);
519 return ptr + CALL_INSN_SIZE + call.disp;
522 void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
523 unsigned long frame_pointer);
526 * If the ops->trampoline was not allocated, then it probably
527 * has a static trampoline func, or is the ftrace caller itself.
529 static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
531 unsigned long offset;
532 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
535 if (ops && ops->trampoline) {
536 #if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \
537 defined(CONFIG_FUNCTION_GRAPH_TRACER)
539 * We only know about function graph tracer setting as static
542 if (ops->trampoline == FTRACE_GRAPH_ADDR)
543 return (void *)prepare_ftrace_return;
548 offset = calc_trampoline_call_offset(save_regs);
551 ptr = (void *)FTRACE_REGS_ADDR + offset;
553 ptr = (void *)FTRACE_ADDR + offset;
555 return addr_from_call(ptr);
558 void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
560 unsigned long offset;
562 /* If we didn't allocate this trampoline, consider it static */
563 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
564 return static_tramp_func(ops, rec);
566 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
567 return addr_from_call((void *)ops->trampoline + offset);
570 void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
572 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
575 tramp_free((void *)ops->trampoline);
579 #endif /* CONFIG_X86_64 */
580 #endif /* CONFIG_DYNAMIC_FTRACE */
582 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
584 #if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS)
585 extern void ftrace_graph_call(void);
586 static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
588 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
591 static int ftrace_mod_jmp(unsigned long ip, void *func)
595 new = ftrace_jmp_replace(ip, (unsigned long)func);
596 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
600 int ftrace_enable_ftrace_graph_caller(void)
602 unsigned long ip = (unsigned long)(&ftrace_graph_call);
604 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
607 int ftrace_disable_ftrace_graph_caller(void)
609 unsigned long ip = (unsigned long)(&ftrace_graph_call);
611 return ftrace_mod_jmp(ip, &ftrace_stub);
613 #endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */
616 * Hook the return address and push it in the stack of return addrs
617 * in current thread info.
619 void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
620 unsigned long frame_pointer)
622 unsigned long return_hooker = (unsigned long)&return_to_handler;
626 * When resuming from suspend-to-ram, this function can be indirectly
627 * called from early CPU startup code while the CPU is in real mode,
628 * which would fail miserably. Make sure the stack pointer is a
631 * This check isn't as accurate as virt_addr_valid(), but it should be
632 * good enough for this purpose, and it's fast.
634 if (unlikely((long)__builtin_frame_address(0) >= 0))
637 if (unlikely(ftrace_graph_is_dead()))
640 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
643 bit = ftrace_test_recursion_trylock(ip, *parent);
647 if (!function_graph_enter(*parent, ip, frame_pointer, parent))
648 *parent = return_hooker;
650 ftrace_test_recursion_unlock(bit);
653 #ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
654 void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
655 struct ftrace_ops *op, struct ftrace_regs *fregs)
657 struct pt_regs *regs = &fregs->regs;
658 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs);
660 prepare_ftrace_return(ip, (unsigned long *)stack, 0);
664 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */