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/ftrace.h>
22 #include <linux/ftrace.h>
27 #ifdef CONFIG_DYNAMIC_FTRACE
29 union ftrace_code_union
{
30 char code
[MCOUNT_INSN_SIZE
];
34 } __attribute__((packed
));
37 static int ftrace_calc_offset(long ip
, long addr
)
39 return (int)(addr
- ip
);
42 static unsigned char *ftrace_call_replace(unsigned long ip
, unsigned long addr
)
44 static union ftrace_code_union calc
;
47 calc
.offset
= ftrace_calc_offset(ip
+ MCOUNT_INSN_SIZE
, addr
);
50 * No locking needed, this must be called via kstop_machine
51 * which in essence is like running on a uniprocessor machine.
57 * Modifying code must take extra care. On an SMP machine, if
58 * the code being modified is also being executed on another CPU
59 * that CPU will have undefined results and possibly take a GPF.
60 * We use kstop_machine to stop other CPUS from exectuing code.
61 * But this does not stop NMIs from happening. We still need
62 * to protect against that. We separate out the modification of
63 * the code to take care of this.
65 * Two buffers are added: An IP buffer and a "code" buffer.
67 * 1) Put the instruction pointer into the IP buffer
68 * and the new code into the "code" buffer.
69 * 2) Set a flag that says we are modifying code
70 * 3) Wait for any running NMIs to finish.
73 * 6) Wait for any running NMIs to finish.
75 * If an NMI is executed, the first thing it does is to call
76 * "ftrace_nmi_enter". This will check if the flag is set to write
77 * and if it is, it will write what is in the IP and "code" buffers.
79 * The trick is, it does not matter if everyone is writing the same
80 * content to the code location. Also, if a CPU is executing code
81 * it is OK to write to that code location if the contents being written
82 * are the same as what exists.
85 static atomic_t in_nmi
= ATOMIC_INIT(0);
86 static int mod_code_status
; /* holds return value of text write */
87 static int mod_code_write
; /* set when NMI should do the write */
88 static void *mod_code_ip
; /* holds the IP to write to */
89 static void *mod_code_newcode
; /* holds the text to write to the IP */
91 static unsigned nmi_wait_count
;
92 static atomic_t nmi_update_count
= ATOMIC_INIT(0);
94 int ftrace_arch_read_dyn_info(char *buf
, int size
)
98 r
= snprintf(buf
, size
, "%u %u",
100 atomic_read(&nmi_update_count
));
104 static void ftrace_mod_code(void)
107 * Yes, more than one CPU process can be writing to mod_code_status.
108 * (and the code itself)
109 * But if one were to fail, then they all should, and if one were
110 * to succeed, then they all should.
112 mod_code_status
= probe_kernel_write(mod_code_ip
, mod_code_newcode
,
117 void ftrace_nmi_enter(void)
120 /* Must have in_nmi seen before reading write flag */
122 if (mod_code_write
) {
124 atomic_inc(&nmi_update_count
);
128 void ftrace_nmi_exit(void)
130 /* Finish all executions before clearing in_nmi */
135 static void wait_for_nmi(void)
139 while (atomic_read(&in_nmi
)) {
149 do_ftrace_mod_code(unsigned long ip
, void *new_code
)
151 mod_code_ip
= (void *)ip
;
152 mod_code_newcode
= new_code
;
154 /* The buffers need to be visible before we let NMIs write them */
159 /* Make sure write bit is visible before we wait on NMIs */
164 /* Make sure all running NMIs have finished before we write the code */
169 /* Make sure the write happens before clearing the bit */
174 /* make sure NMIs see the cleared bit */
179 return mod_code_status
;
185 static unsigned char ftrace_nop
[MCOUNT_INSN_SIZE
];
187 static unsigned char *ftrace_nop_replace(void)
193 ftrace_modify_code(unsigned long ip
, unsigned char *old_code
,
194 unsigned char *new_code
)
196 unsigned char replaced
[MCOUNT_INSN_SIZE
];
199 * Note: Due to modules and __init, code can
200 * disappear and change, we need to protect against faulting
201 * as well as code changing. We do this by using the
202 * probe_kernel_* functions.
204 * No real locking needed, this code is run through
205 * kstop_machine, or before SMP starts.
208 /* read the text we want to modify */
209 if (probe_kernel_read(replaced
, (void *)ip
, MCOUNT_INSN_SIZE
))
212 /* Make sure it is what we expect it to be */
213 if (memcmp(replaced
, old_code
, MCOUNT_INSN_SIZE
) != 0)
216 /* replace the text with the new text */
217 if (do_ftrace_mod_code(ip
, new_code
))
225 int ftrace_make_nop(struct module
*mod
,
226 struct dyn_ftrace
*rec
, unsigned long addr
)
228 unsigned char *new, *old
;
229 unsigned long ip
= rec
->ip
;
231 old
= ftrace_call_replace(ip
, addr
);
232 new = ftrace_nop_replace();
234 return ftrace_modify_code(rec
->ip
, old
, new);
237 int ftrace_make_call(struct dyn_ftrace
*rec
, unsigned long addr
)
239 unsigned char *new, *old
;
240 unsigned long ip
= rec
->ip
;
242 old
= ftrace_nop_replace();
243 new = ftrace_call_replace(ip
, addr
);
245 return ftrace_modify_code(rec
->ip
, old
, new);
248 int ftrace_update_ftrace_func(ftrace_func_t func
)
250 unsigned long ip
= (unsigned long)(&ftrace_call
);
251 unsigned char old
[MCOUNT_INSN_SIZE
], *new;
254 memcpy(old
, &ftrace_call
, MCOUNT_INSN_SIZE
);
255 new = ftrace_call_replace(ip
, (unsigned long)func
);
256 ret
= ftrace_modify_code(ip
, old
, new);
261 int __init
ftrace_dyn_arch_init(void *data
)
263 extern const unsigned char ftrace_test_p6nop
[];
264 extern const unsigned char ftrace_test_nop5
[];
265 extern const unsigned char ftrace_test_jmp
[];
269 * There is no good nop for all x86 archs.
270 * We will default to using the P6_NOP5, but first we
271 * will test to make sure that the nop will actually
272 * work on this CPU. If it faults, we will then
273 * go to a lesser efficient 5 byte nop. If that fails
274 * we then just use a jmp as our nop. This isn't the most
275 * efficient nop, but we can not use a multi part nop
276 * since we would then risk being preempted in the middle
277 * of that nop, and if we enabled tracing then, it might
278 * cause a system crash.
280 * TODO: check the cpuid to determine the best nop.
284 "jmp ftrace_test_p6nop\n"
287 "nop\n" /* 2 byte jmp + 3 bytes */
292 ".byte 0x66,0x66,0x66,0x66,0x90\n"
294 ".section .fixup, \"ax\"\n"
296 " jmp ftrace_test_nop5\n"
300 _ASM_EXTABLE(ftrace_test_p6nop
, 2b
)
301 _ASM_EXTABLE(ftrace_test_nop5
, 3b
)
302 : "=r"(faulted
) : "0" (faulted
));
306 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
307 memcpy(ftrace_nop
, ftrace_test_p6nop
, MCOUNT_INSN_SIZE
);
310 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
311 memcpy(ftrace_nop
, ftrace_test_nop5
, MCOUNT_INSN_SIZE
);
314 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
315 memcpy(ftrace_nop
, ftrace_test_jmp
, MCOUNT_INSN_SIZE
);
319 /* The return code is retured via data */
320 *(unsigned long *)data
= 0;
326 #ifdef CONFIG_FUNCTION_RET_TRACER
328 #ifndef CONFIG_DYNAMIC_FTRACE
331 * These functions are picked from those used on
332 * this page for dynamic ftrace. They have been
333 * simplified to ignore all traces in NMI context.
335 static atomic_t in_nmi
;
337 void ftrace_nmi_enter(void)
342 void ftrace_nmi_exit(void)
346 #endif /* !CONFIG_DYNAMIC_FTRACE */
348 /* Add a function return address to the trace stack on thread info.*/
349 static int push_return_trace(unsigned long ret
, unsigned long long time
,
354 if (!current
->ret_stack
)
357 /* The return trace stack is full */
358 if (current
->curr_ret_stack
== FTRACE_RETFUNC_DEPTH
- 1) {
359 atomic_inc(¤t
->trace_overrun
);
363 index
= ++current
->curr_ret_stack
;
365 current
->ret_stack
[index
].ret
= ret
;
366 current
->ret_stack
[index
].func
= func
;
367 current
->ret_stack
[index
].calltime
= time
;
372 /* Retrieve a function return address to the trace stack on thread info.*/
373 static void pop_return_trace(unsigned long *ret
, unsigned long long *time
,
374 unsigned long *func
, unsigned long *overrun
)
378 index
= current
->curr_ret_stack
;
379 *ret
= current
->ret_stack
[index
].ret
;
380 *func
= current
->ret_stack
[index
].func
;
381 *time
= current
->ret_stack
[index
].calltime
;
382 *overrun
= atomic_read(¤t
->trace_overrun
);
383 current
->curr_ret_stack
--;
387 * Send the trace to the ring-buffer.
388 * @return the original return address.
390 unsigned long ftrace_return_to_handler(void)
392 struct ftrace_retfunc trace
;
393 pop_return_trace(&trace
.ret
, &trace
.calltime
, &trace
.func
,
395 trace
.rettime
= cpu_clock(raw_smp_processor_id());
396 ftrace_function_return(&trace
);
402 * Hook the return address and push it in the stack of return addrs
403 * in current thread info.
405 void prepare_ftrace_return(unsigned long *parent
, unsigned long self_addr
)
408 unsigned long long calltime
;
410 unsigned long return_hooker
= (unsigned long)
413 /* Nmi's are currently unsupported */
414 if (atomic_read(&in_nmi
))
418 * Protect against fault, even if it shouldn't
419 * happen. This tool is too much intrusive to
420 * ignore such a protection.
423 "1: movl (%[parent_old]), %[old]\n"
424 "2: movl %[return_hooker], (%[parent_replaced])\n"
425 " movl $0, %[faulted]\n"
427 ".section .fixup, \"ax\"\n"
428 "3: movl $1, %[faulted]\n"
431 ".section __ex_table, \"a\"\n"
436 : [parent_replaced
] "=r" (parent
), [old
] "=r" (old
),
437 [faulted
] "=r" (faulted
)
438 : [parent_old
] "0" (parent
), [return_hooker
] "r" (return_hooker
)
442 if (WARN_ON(faulted
)) {
443 unregister_ftrace_return();
447 if (WARN_ON(!__kernel_text_address(old
))) {
448 unregister_ftrace_return();
453 calltime
= cpu_clock(raw_smp_processor_id());
455 if (push_return_trace(old
, calltime
, self_addr
) == -EBUSY
)
459 #endif /* CONFIG_FUNCTION_RET_TRACER */