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>
27 #include <trace/syscall.h>
29 #include <asm/set_memory.h>
30 #include <asm/kprobes.h>
31 #include <asm/ftrace.h>
33 #include <asm/text-patching.h>
35 #ifdef CONFIG_DYNAMIC_FTRACE
37 static int ftrace_poke_late = 0;
39 int ftrace_arch_code_modify_prepare(void)
40 __acquires(&text_mutex)
43 * Need to grab text_mutex to prevent a race from module loading
44 * and live kernel patching from changing the text permissions while
45 * ftrace has it set to "read/write".
47 mutex_lock(&text_mutex);
52 int 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);
66 static const char *ftrace_nop_replace(void)
68 return ideal_nops[NOP_ATOMIC5];
71 static const char *ftrace_call_replace(unsigned long ip, unsigned long addr)
73 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
76 static int ftrace_verify_code(unsigned long ip, const char *old_code)
78 char cur_code[MCOUNT_INSN_SIZE];
82 * We are paranoid about modifying text, as if a bug was to happen, it
83 * could cause us to read or write to someplace that could cause harm.
84 * Carefully read and modify the code with probe_kernel_*(), and make
85 * sure what we read is what we expected it to be before modifying it.
87 /* read the text we want to modify */
88 if (probe_kernel_read(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) {
93 /* Make sure it is what we expect it to be */
94 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) {
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 int __init ftrace_dyn_arch_init(void)
259 /* Currently only x86_64 supports dynamic trampolines */
262 #ifdef CONFIG_MODULES
263 #include <linux/moduleloader.h>
264 /* Module allocation simplifies allocating memory for code */
265 static inline void *alloc_tramp(unsigned long size)
267 return module_alloc(size);
269 static inline void tramp_free(void *tramp)
271 module_memfree(tramp);
274 /* Trampolines can only be created if modules are supported */
275 static inline void *alloc_tramp(unsigned long size)
279 static inline void tramp_free(void *tramp) { }
282 /* Defined as markers to the end of the ftrace default trampolines */
283 extern void ftrace_regs_caller_end(void);
284 extern void ftrace_epilogue(void);
285 extern void ftrace_caller_op_ptr(void);
286 extern void ftrace_regs_caller_op_ptr(void);
288 /* movq function_trace_op(%rip), %rdx */
289 /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
290 #define OP_REF_SIZE 7
293 * The ftrace_ops is passed to the function callback. Since the
294 * trampoline only services a single ftrace_ops, we can pass in
297 * The ftrace_op_code_union is used to create a pointer to the
298 * ftrace_ops that will be passed to the callback function.
300 union ftrace_op_code_union {
301 char code[OP_REF_SIZE];
305 } __attribute__((packed));
311 create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
313 unsigned long start_offset;
314 unsigned long end_offset;
315 unsigned long op_offset;
316 unsigned long call_offset;
317 unsigned long offset;
318 unsigned long npages;
324 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
325 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
326 union ftrace_op_code_union op_ptr;
329 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
330 start_offset = (unsigned long)ftrace_regs_caller;
331 end_offset = (unsigned long)ftrace_regs_caller_end;
332 op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
333 call_offset = (unsigned long)ftrace_regs_call;
335 start_offset = (unsigned long)ftrace_caller;
336 end_offset = (unsigned long)ftrace_epilogue;
337 op_offset = (unsigned long)ftrace_caller_op_ptr;
338 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 = probe_kernel_read(trampoline, (void *)start_offset, size);
357 if (WARN_ON(ret < 0))
360 ip = trampoline + size;
362 /* The trampoline ends with ret(q) */
363 retq = (unsigned long)ftrace_stub;
364 ret = probe_kernel_read(ip, (void *)retq, RET_SIZE);
365 if (WARN_ON(ret < 0))
369 * The address of the ftrace_ops that is used for this trampoline
370 * is stored at the end of the trampoline. This will be used to
371 * load the third parameter for the callback. Basically, that
372 * location at the end of the trampoline takes the place of
373 * the global function_trace_op variable.
376 ptr = (unsigned long *)(trampoline + size + RET_SIZE);
377 *ptr = (unsigned long)ops;
379 op_offset -= start_offset;
380 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
382 /* Are we pointing to the reference? */
383 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0))
386 /* Load the contents of ptr into the callback parameter */
387 offset = (unsigned long)ptr;
388 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
390 op_ptr.offset = offset;
392 /* put in the new offset to the ftrace_ops */
393 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
395 /* put in the call to the function */
396 mutex_lock(&text_mutex);
397 call_offset -= start_offset;
398 memcpy(trampoline + call_offset,
399 text_gen_insn(CALL_INSN_OPCODE,
400 trampoline + call_offset,
401 ftrace_ops_get_func(ops)), CALL_INSN_SIZE);
402 mutex_unlock(&text_mutex);
404 /* ALLOC_TRAMP flags lets us know we created it */
405 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
407 set_vm_flush_reset_perms(trampoline);
409 set_memory_ro((unsigned long)trampoline, npages);
410 set_memory_x((unsigned long)trampoline, npages);
411 return (unsigned long)trampoline;
413 tramp_free(trampoline);
417 static unsigned long calc_trampoline_call_offset(bool save_regs)
419 unsigned long start_offset;
420 unsigned long call_offset;
423 start_offset = (unsigned long)ftrace_regs_caller;
424 call_offset = (unsigned long)ftrace_regs_call;
426 start_offset = (unsigned long)ftrace_caller;
427 call_offset = (unsigned long)ftrace_call;
430 return call_offset - start_offset;
433 void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
436 unsigned long offset;
441 if (!ops->trampoline) {
442 ops->trampoline = create_trampoline(ops, &size);
443 if (!ops->trampoline)
445 ops->trampoline_size = size;
450 * The ftrace_ops caller may set up its own trampoline.
451 * In such a case, this code must not modify it.
453 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
456 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
457 ip = ops->trampoline + offset;
458 func = ftrace_ops_get_func(ops);
460 mutex_lock(&text_mutex);
461 /* Do a safe modify in case the trampoline is executing */
462 new = ftrace_call_replace(ip, (unsigned long)func);
463 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
464 mutex_unlock(&text_mutex);
467 /* Return the address of the function the trampoline calls */
468 static void *addr_from_call(void *ptr)
470 union text_poke_insn call;
473 ret = probe_kernel_read(&call, ptr, CALL_INSN_SIZE);
474 if (WARN_ON_ONCE(ret < 0))
477 /* Make sure this is a call */
478 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) {
479 pr_warn("Expected E8, got %x\n", call.opcode);
483 return ptr + CALL_INSN_SIZE + call.disp;
486 void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
487 unsigned long frame_pointer);
490 * If the ops->trampoline was not allocated, then it probably
491 * has a static trampoline func, or is the ftrace caller itself.
493 static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
495 unsigned long offset;
496 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
499 if (ops && ops->trampoline) {
500 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
502 * We only know about function graph tracer setting as static
505 if (ops->trampoline == FTRACE_GRAPH_ADDR)
506 return (void *)prepare_ftrace_return;
511 offset = calc_trampoline_call_offset(save_regs);
514 ptr = (void *)FTRACE_REGS_ADDR + offset;
516 ptr = (void *)FTRACE_ADDR + offset;
518 return addr_from_call(ptr);
521 void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
523 unsigned long offset;
525 /* If we didn't allocate this trampoline, consider it static */
526 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
527 return static_tramp_func(ops, rec);
529 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
530 return addr_from_call((void *)ops->trampoline + offset);
533 void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
535 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
538 tramp_free((void *)ops->trampoline);
542 #endif /* CONFIG_X86_64 */
543 #endif /* CONFIG_DYNAMIC_FTRACE */
545 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
547 #ifdef CONFIG_DYNAMIC_FTRACE
548 extern void ftrace_graph_call(void);
550 static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
552 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
555 static int ftrace_mod_jmp(unsigned long ip, void *func)
559 new = ftrace_jmp_replace(ip, (unsigned long)func);
560 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
564 int ftrace_enable_ftrace_graph_caller(void)
566 unsigned long ip = (unsigned long)(&ftrace_graph_call);
568 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
571 int ftrace_disable_ftrace_graph_caller(void)
573 unsigned long ip = (unsigned long)(&ftrace_graph_call);
575 return ftrace_mod_jmp(ip, &ftrace_stub);
578 #endif /* !CONFIG_DYNAMIC_FTRACE */
581 * Hook the return address and push it in the stack of return addrs
582 * in current thread info.
584 void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
585 unsigned long frame_pointer)
587 unsigned long return_hooker = (unsigned long)&return_to_handler;
592 * When resuming from suspend-to-ram, this function can be indirectly
593 * called from early CPU startup code while the CPU is in real mode,
594 * which would fail miserably. Make sure the stack pointer is a
597 * This check isn't as accurate as virt_addr_valid(), but it should be
598 * good enough for this purpose, and it's fast.
600 if (unlikely((long)__builtin_frame_address(0) >= 0))
603 if (unlikely(ftrace_graph_is_dead()))
606 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
610 * If the return location is actually pointing directly to
611 * the start of a direct trampoline (if we trace the trampoline
612 * it will still be offset by MCOUNT_INSN_SIZE), then the
613 * return address is actually off by one word, and we
614 * need to adjust for that.
616 if (ftrace_direct_func_count) {
617 if (ftrace_find_direct_func(self_addr + MCOUNT_INSN_SIZE)) {
624 * Protect against fault, even if it shouldn't
625 * happen. This tool is too much intrusive to
626 * ignore such a protection.
629 "1: " _ASM_MOV " (%[parent]), %[old]\n"
630 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
631 " movl $0, %[faulted]\n"
634 ".section .fixup, \"ax\"\n"
635 "4: movl $1, %[faulted]\n"
642 : [old] "=&r" (old), [faulted] "=r" (faulted)
643 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
647 if (unlikely(faulted)) {
653 if (function_graph_enter(old, self_addr, frame_pointer, parent))
656 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */