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 #include <linux/spinlock.h>
13 #include <linux/hardirq.h>
14 #include <linux/uaccess.h>
15 #include <linux/ftrace.h>
16 #include <linux/percpu.h>
17 #include <linux/sched.h>
18 #include <linux/init.h>
19 #include <linux/list.h>
21 #include <asm/cacheflush.h>
22 #include <asm/ftrace.h>
23 #include <linux/ftrace.h>
28 #ifdef CONFIG_DYNAMIC_FTRACE
30 int ftrace_arch_code_modify_prepare(void)
36 int ftrace_arch_code_modify_post_process(void)
42 union ftrace_code_union
{
43 char code
[MCOUNT_INSN_SIZE
];
47 } __attribute__((packed
));
50 static int ftrace_calc_offset(long ip
, long addr
)
52 return (int)(addr
- ip
);
55 static unsigned char *ftrace_call_replace(unsigned long ip
, unsigned long addr
)
57 static union ftrace_code_union calc
;
60 calc
.offset
= ftrace_calc_offset(ip
+ MCOUNT_INSN_SIZE
, addr
);
63 * No locking needed, this must be called via kstop_machine
64 * which in essence is like running on a uniprocessor machine.
70 * Modifying code must take extra care. On an SMP machine, if
71 * the code being modified is also being executed on another CPU
72 * that CPU will have undefined results and possibly take a GPF.
73 * We use kstop_machine to stop other CPUS from exectuing code.
74 * But this does not stop NMIs from happening. We still need
75 * to protect against that. We separate out the modification of
76 * the code to take care of this.
78 * Two buffers are added: An IP buffer and a "code" buffer.
80 * 1) Put the instruction pointer into the IP buffer
81 * and the new code into the "code" buffer.
82 * 2) Wait for any running NMIs to finish and set a flag that says
83 * we are modifying code, it is done in an atomic operation.
86 * 5) Wait for any running NMIs to finish.
88 * If an NMI is executed, the first thing it does is to call
89 * "ftrace_nmi_enter". This will check if the flag is set to write
90 * and if it is, it will write what is in the IP and "code" buffers.
92 * The trick is, it does not matter if everyone is writing the same
93 * content to the code location. Also, if a CPU is executing code
94 * it is OK to write to that code location if the contents being written
95 * are the same as what exists.
98 #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */
99 static atomic_t nmi_running
= ATOMIC_INIT(0);
100 static int mod_code_status
; /* holds return value of text write */
101 static void *mod_code_ip
; /* holds the IP to write to */
102 static void *mod_code_newcode
; /* holds the text to write to the IP */
104 static unsigned nmi_wait_count
;
105 static atomic_t nmi_update_count
= ATOMIC_INIT(0);
107 int ftrace_arch_read_dyn_info(char *buf
, int size
)
111 r
= snprintf(buf
, size
, "%u %u",
113 atomic_read(&nmi_update_count
));
117 static void clear_mod_flag(void)
119 int old
= atomic_read(&nmi_running
);
122 int new = old
& ~MOD_CODE_WRITE_FLAG
;
127 old
= atomic_cmpxchg(&nmi_running
, old
, new);
131 static void ftrace_mod_code(void)
134 * Yes, more than one CPU process can be writing to mod_code_status.
135 * (and the code itself)
136 * But if one were to fail, then they all should, and if one were
137 * to succeed, then they all should.
139 mod_code_status
= probe_kernel_write(mod_code_ip
, mod_code_newcode
,
142 /* if we fail, then kill any new writers */
147 void ftrace_nmi_enter(void)
149 if (atomic_inc_return(&nmi_running
) & MOD_CODE_WRITE_FLAG
) {
152 atomic_inc(&nmi_update_count
);
154 /* Must have previous changes seen before executions */
158 void ftrace_nmi_exit(void)
160 /* Finish all executions before clearing nmi_running */
162 atomic_dec(&nmi_running
);
165 static void wait_for_nmi_and_set_mod_flag(void)
167 if (!atomic_cmpxchg(&nmi_running
, 0, MOD_CODE_WRITE_FLAG
))
172 } while (atomic_cmpxchg(&nmi_running
, 0, MOD_CODE_WRITE_FLAG
));
177 static void wait_for_nmi(void)
179 if (!atomic_read(&nmi_running
))
184 } while (atomic_read(&nmi_running
));
190 do_ftrace_mod_code(unsigned long ip
, void *new_code
)
192 mod_code_ip
= (void *)ip
;
193 mod_code_newcode
= new_code
;
195 /* The buffers need to be visible before we let NMIs write them */
198 wait_for_nmi_and_set_mod_flag();
200 /* Make sure all running NMIs have finished before we write the code */
205 /* Make sure the write happens before clearing the bit */
211 return mod_code_status
;
217 static unsigned char ftrace_nop
[MCOUNT_INSN_SIZE
];
219 static unsigned char *ftrace_nop_replace(void)
225 ftrace_modify_code(unsigned long ip
, unsigned char *old_code
,
226 unsigned char *new_code
)
228 unsigned char replaced
[MCOUNT_INSN_SIZE
];
231 * Note: Due to modules and __init, code can
232 * disappear and change, we need to protect against faulting
233 * as well as code changing. We do this by using the
234 * probe_kernel_* functions.
236 * No real locking needed, this code is run through
237 * kstop_machine, or before SMP starts.
240 /* read the text we want to modify */
241 if (probe_kernel_read(replaced
, (void *)ip
, MCOUNT_INSN_SIZE
))
244 /* Make sure it is what we expect it to be */
245 if (memcmp(replaced
, old_code
, MCOUNT_INSN_SIZE
) != 0)
248 /* replace the text with the new text */
249 if (do_ftrace_mod_code(ip
, new_code
))
257 int ftrace_make_nop(struct module
*mod
,
258 struct dyn_ftrace
*rec
, unsigned long addr
)
260 unsigned char *new, *old
;
261 unsigned long ip
= rec
->ip
;
263 old
= ftrace_call_replace(ip
, addr
);
264 new = ftrace_nop_replace();
266 return ftrace_modify_code(rec
->ip
, old
, new);
269 int ftrace_make_call(struct dyn_ftrace
*rec
, unsigned long addr
)
271 unsigned char *new, *old
;
272 unsigned long ip
= rec
->ip
;
274 old
= ftrace_nop_replace();
275 new = ftrace_call_replace(ip
, addr
);
277 return ftrace_modify_code(rec
->ip
, old
, new);
280 int ftrace_update_ftrace_func(ftrace_func_t func
)
282 unsigned long ip
= (unsigned long)(&ftrace_call
);
283 unsigned char old
[MCOUNT_INSN_SIZE
], *new;
286 memcpy(old
, &ftrace_call
, MCOUNT_INSN_SIZE
);
287 new = ftrace_call_replace(ip
, (unsigned long)func
);
288 ret
= ftrace_modify_code(ip
, old
, new);
293 int __init
ftrace_dyn_arch_init(void *data
)
295 extern const unsigned char ftrace_test_p6nop
[];
296 extern const unsigned char ftrace_test_nop5
[];
297 extern const unsigned char ftrace_test_jmp
[];
301 * There is no good nop for all x86 archs.
302 * We will default to using the P6_NOP5, but first we
303 * will test to make sure that the nop will actually
304 * work on this CPU. If it faults, we will then
305 * go to a lesser efficient 5 byte nop. If that fails
306 * we then just use a jmp as our nop. This isn't the most
307 * efficient nop, but we can not use a multi part nop
308 * since we would then risk being preempted in the middle
309 * of that nop, and if we enabled tracing then, it might
310 * cause a system crash.
312 * TODO: check the cpuid to determine the best nop.
316 "jmp ftrace_test_p6nop\n"
319 "nop\n" /* 2 byte jmp + 3 bytes */
324 ".byte 0x66,0x66,0x66,0x66,0x90\n"
326 ".section .fixup, \"ax\"\n"
328 " jmp ftrace_test_nop5\n"
332 _ASM_EXTABLE(ftrace_test_p6nop
, 2b
)
333 _ASM_EXTABLE(ftrace_test_nop5
, 3b
)
334 : "=r"(faulted
) : "0" (faulted
));
338 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
339 memcpy(ftrace_nop
, ftrace_test_p6nop
, MCOUNT_INSN_SIZE
);
342 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
343 memcpy(ftrace_nop
, ftrace_test_nop5
, MCOUNT_INSN_SIZE
);
346 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
347 memcpy(ftrace_nop
, ftrace_test_jmp
, MCOUNT_INSN_SIZE
);
351 /* The return code is retured via data */
352 *(unsigned long *)data
= 0;
358 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
360 #ifdef CONFIG_DYNAMIC_FTRACE
361 extern void ftrace_graph_call(void);
363 static int ftrace_mod_jmp(unsigned long ip
,
364 int old_offset
, int new_offset
)
366 unsigned char code
[MCOUNT_INSN_SIZE
];
368 if (probe_kernel_read(code
, (void *)ip
, MCOUNT_INSN_SIZE
))
371 if (code
[0] != 0xe9 || old_offset
!= *(int *)(&code
[1]))
374 *(int *)(&code
[1]) = new_offset
;
376 if (do_ftrace_mod_code(ip
, &code
))
382 int ftrace_enable_ftrace_graph_caller(void)
384 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
385 int old_offset
, new_offset
;
387 old_offset
= (unsigned long)(&ftrace_stub
) - (ip
+ MCOUNT_INSN_SIZE
);
388 new_offset
= (unsigned long)(&ftrace_graph_caller
) - (ip
+ MCOUNT_INSN_SIZE
);
390 return ftrace_mod_jmp(ip
, old_offset
, new_offset
);
393 int ftrace_disable_ftrace_graph_caller(void)
395 unsigned long ip
= (unsigned long)(&ftrace_graph_call
);
396 int old_offset
, new_offset
;
398 old_offset
= (unsigned long)(&ftrace_graph_caller
) - (ip
+ MCOUNT_INSN_SIZE
);
399 new_offset
= (unsigned long)(&ftrace_stub
) - (ip
+ MCOUNT_INSN_SIZE
);
401 return ftrace_mod_jmp(ip
, old_offset
, new_offset
);
404 #endif /* !CONFIG_DYNAMIC_FTRACE */
407 * Hook the return address and push it in the stack of return addrs
408 * in current thread info.
410 void prepare_ftrace_return(unsigned long *parent
, unsigned long self_addr
)
414 struct ftrace_graph_ent trace
;
415 unsigned long return_hooker
= (unsigned long)
418 /* Nmi's are currently unsupported */
419 if (unlikely(in_nmi()))
422 if (unlikely(atomic_read(¤t
->tracing_graph_pause
)))
426 * Protect against fault, even if it shouldn't
427 * happen. This tool is too much intrusive to
428 * ignore such a protection.
431 "1: " _ASM_MOV
" (%[parent]), %[old]\n"
432 "2: " _ASM_MOV
" %[return_hooker], (%[parent])\n"
433 " movl $0, %[faulted]\n"
436 ".section .fixup, \"ax\"\n"
437 "4: movl $1, %[faulted]\n"
444 : [old
] "=r" (old
), [faulted
] "=r" (faulted
)
445 : [parent
] "r" (parent
), [return_hooker
] "r" (return_hooker
)
449 if (unlikely(faulted
)) {
455 if (ftrace_push_return_trace(old
, self_addr
, &trace
.depth
) == -EBUSY
) {
460 trace
.func
= self_addr
;
462 /* Only trace if the calling function expects to */
463 if (!ftrace_graph_entry(&trace
)) {
464 current
->curr_ret_stack
--;
468 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
470 #ifdef CONFIG_FTRACE_SYSCALLS
472 extern unsigned long __start_syscalls_metadata
[];
473 extern unsigned long __stop_syscalls_metadata
[];
474 extern unsigned long *sys_call_table
;
476 static struct syscall_metadata
**syscalls_metadata
;
478 static struct syscall_metadata
*find_syscall_meta(unsigned long *syscall
)
480 struct syscall_metadata
*start
;
481 struct syscall_metadata
*stop
;
482 char str
[KSYM_SYMBOL_LEN
];
485 start
= (struct syscall_metadata
*)__start_syscalls_metadata
;
486 stop
= (struct syscall_metadata
*)__stop_syscalls_metadata
;
487 kallsyms_lookup((unsigned long) syscall
, NULL
, NULL
, NULL
, str
);
489 for ( ; start
< stop
; start
++) {
490 if (start
->name
&& !strcmp(start
->name
, str
))
496 struct syscall_metadata
*syscall_nr_to_meta(int nr
)
498 if (!syscalls_metadata
|| nr
>= FTRACE_SYSCALL_MAX
|| nr
< 0)
501 return syscalls_metadata
[nr
];
504 void arch_init_ftrace_syscalls(void)
507 struct syscall_metadata
*meta
;
508 unsigned long **psys_syscall_table
= &sys_call_table
;
509 static atomic_t refs
;
511 if (atomic_inc_return(&refs
) != 1)
514 syscalls_metadata
= kzalloc(sizeof(*syscalls_metadata
) *
515 FTRACE_SYSCALL_MAX
, GFP_KERNEL
);
516 if (!syscalls_metadata
) {
521 for (i
= 0; i
< FTRACE_SYSCALL_MAX
; i
++) {
522 meta
= find_syscall_meta(psys_syscall_table
[i
]);
523 syscalls_metadata
[i
] = meta
;
527 /* Paranoid: avoid overflow */