Commit | Line | Data |
---|---|---|
3d083395 SR |
1 | /* |
2 | * Code for replacing ftrace calls with jumps. | |
3 | * | |
4 | * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> | |
5 | * | |
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. | |
10 | */ | |
11 | ||
3bb258bf JP |
12 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
13 | ||
3d083395 SR |
14 | #include <linux/spinlock.h> |
15 | #include <linux/hardirq.h> | |
6f93fc07 | 16 | #include <linux/uaccess.h> |
3d083395 SR |
17 | #include <linux/ftrace.h> |
18 | #include <linux/percpu.h> | |
19b3e967 | 19 | #include <linux/sched.h> |
3d083395 SR |
20 | #include <linux/init.h> |
21 | #include <linux/list.h> | |
22 | ||
47788c58 FW |
23 | #include <trace/syscall.h> |
24 | ||
16239630 | 25 | #include <asm/cacheflush.h> |
395a59d0 | 26 | #include <asm/ftrace.h> |
732f3ca7 | 27 | #include <asm/nops.h> |
caf4b323 | 28 | #include <asm/nmi.h> |
3d083395 | 29 | |
3d083395 | 30 | |
caf4b323 | 31 | #ifdef CONFIG_DYNAMIC_FTRACE |
3d083395 | 32 | |
0c54dd34 SR |
33 | /* |
34 | * modifying_code is set to notify NMIs that they need to use | |
35 | * memory barriers when entering or exiting. But we don't want | |
36 | * to burden NMIs with unnecessary memory barriers when code | |
37 | * modification is not being done (which is most of the time). | |
38 | * | |
39 | * A mutex is already held when ftrace_arch_code_modify_prepare | |
40 | * and post_process are called. No locks need to be taken here. | |
41 | * | |
42 | * Stop machine will make sure currently running NMIs are done | |
43 | * and new NMIs will see the updated variable before we need | |
44 | * to worry about NMIs doing memory barriers. | |
45 | */ | |
46 | static int modifying_code __read_mostly; | |
47 | static DEFINE_PER_CPU(int, save_modifying_code); | |
48 | ||
16239630 SR |
49 | int ftrace_arch_code_modify_prepare(void) |
50 | { | |
51 | set_kernel_text_rw(); | |
0c54dd34 | 52 | modifying_code = 1; |
16239630 SR |
53 | return 0; |
54 | } | |
55 | ||
56 | int ftrace_arch_code_modify_post_process(void) | |
57 | { | |
0c54dd34 | 58 | modifying_code = 0; |
16239630 SR |
59 | set_kernel_text_ro(); |
60 | return 0; | |
61 | } | |
62 | ||
3d083395 | 63 | union ftrace_code_union { |
395a59d0 | 64 | char code[MCOUNT_INSN_SIZE]; |
3d083395 SR |
65 | struct { |
66 | char e8; | |
67 | int offset; | |
68 | } __attribute__((packed)); | |
69 | }; | |
70 | ||
15adc048 | 71 | static int ftrace_calc_offset(long ip, long addr) |
3c1720f0 SR |
72 | { |
73 | return (int)(addr - ip); | |
74 | } | |
3d083395 | 75 | |
31e88909 | 76 | static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) |
3c1720f0 SR |
77 | { |
78 | static union ftrace_code_union calc; | |
3d083395 | 79 | |
3c1720f0 | 80 | calc.e8 = 0xe8; |
395a59d0 | 81 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
3c1720f0 SR |
82 | |
83 | /* | |
84 | * No locking needed, this must be called via kstop_machine | |
85 | * which in essence is like running on a uniprocessor machine. | |
86 | */ | |
87 | return calc.code; | |
3d083395 SR |
88 | } |
89 | ||
17666f02 SR |
90 | /* |
91 | * Modifying code must take extra care. On an SMP machine, if | |
92 | * the code being modified is also being executed on another CPU | |
93 | * that CPU will have undefined results and possibly take a GPF. | |
94 | * We use kstop_machine to stop other CPUS from exectuing code. | |
95 | * But this does not stop NMIs from happening. We still need | |
96 | * to protect against that. We separate out the modification of | |
97 | * the code to take care of this. | |
98 | * | |
99 | * Two buffers are added: An IP buffer and a "code" buffer. | |
100 | * | |
a26a2a27 | 101 | * 1) Put the instruction pointer into the IP buffer |
17666f02 | 102 | * and the new code into the "code" buffer. |
e9d9df44 LJ |
103 | * 2) Wait for any running NMIs to finish and set a flag that says |
104 | * we are modifying code, it is done in an atomic operation. | |
105 | * 3) Write the code | |
106 | * 4) clear the flag. | |
107 | * 5) Wait for any running NMIs to finish. | |
17666f02 SR |
108 | * |
109 | * If an NMI is executed, the first thing it does is to call | |
110 | * "ftrace_nmi_enter". This will check if the flag is set to write | |
111 | * and if it is, it will write what is in the IP and "code" buffers. | |
112 | * | |
113 | * The trick is, it does not matter if everyone is writing the same | |
114 | * content to the code location. Also, if a CPU is executing code | |
115 | * it is OK to write to that code location if the contents being written | |
116 | * are the same as what exists. | |
117 | */ | |
118 | ||
e9d9df44 | 119 | #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */ |
4e6ea144 | 120 | static atomic_t nmi_running = ATOMIC_INIT(0); |
a26a2a27 | 121 | static int mod_code_status; /* holds return value of text write */ |
a26a2a27 SR |
122 | static void *mod_code_ip; /* holds the IP to write to */ |
123 | static void *mod_code_newcode; /* holds the text to write to the IP */ | |
17666f02 | 124 | |
a26a2a27 SR |
125 | static unsigned nmi_wait_count; |
126 | static atomic_t nmi_update_count = ATOMIC_INIT(0); | |
b807c3d0 SR |
127 | |
128 | int ftrace_arch_read_dyn_info(char *buf, int size) | |
129 | { | |
130 | int r; | |
131 | ||
132 | r = snprintf(buf, size, "%u %u", | |
133 | nmi_wait_count, | |
134 | atomic_read(&nmi_update_count)); | |
135 | return r; | |
136 | } | |
137 | ||
e9d9df44 LJ |
138 | static void clear_mod_flag(void) |
139 | { | |
140 | int old = atomic_read(&nmi_running); | |
141 | ||
142 | for (;;) { | |
143 | int new = old & ~MOD_CODE_WRITE_FLAG; | |
144 | ||
145 | if (old == new) | |
146 | break; | |
147 | ||
148 | old = atomic_cmpxchg(&nmi_running, old, new); | |
149 | } | |
150 | } | |
151 | ||
17666f02 SR |
152 | static void ftrace_mod_code(void) |
153 | { | |
154 | /* | |
155 | * Yes, more than one CPU process can be writing to mod_code_status. | |
156 | * (and the code itself) | |
157 | * But if one were to fail, then they all should, and if one were | |
158 | * to succeed, then they all should. | |
159 | */ | |
160 | mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode, | |
161 | MCOUNT_INSN_SIZE); | |
90c7ac49 SR |
162 | |
163 | /* if we fail, then kill any new writers */ | |
164 | if (mod_code_status) | |
e9d9df44 | 165 | clear_mod_flag(); |
17666f02 SR |
166 | } |
167 | ||
a81bd80a | 168 | void ftrace_nmi_enter(void) |
17666f02 | 169 | { |
0c54dd34 SR |
170 | __get_cpu_var(save_modifying_code) = modifying_code; |
171 | ||
172 | if (!__get_cpu_var(save_modifying_code)) | |
173 | return; | |
174 | ||
e9d9df44 LJ |
175 | if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) { |
176 | smp_rmb(); | |
17666f02 | 177 | ftrace_mod_code(); |
b807c3d0 SR |
178 | atomic_inc(&nmi_update_count); |
179 | } | |
e9d9df44 LJ |
180 | /* Must have previous changes seen before executions */ |
181 | smp_mb(); | |
17666f02 SR |
182 | } |
183 | ||
a81bd80a | 184 | void ftrace_nmi_exit(void) |
17666f02 | 185 | { |
0c54dd34 SR |
186 | if (!__get_cpu_var(save_modifying_code)) |
187 | return; | |
188 | ||
4e6ea144 | 189 | /* Finish all executions before clearing nmi_running */ |
e9d9df44 | 190 | smp_mb(); |
4e6ea144 | 191 | atomic_dec(&nmi_running); |
17666f02 SR |
192 | } |
193 | ||
e9d9df44 LJ |
194 | static void wait_for_nmi_and_set_mod_flag(void) |
195 | { | |
196 | if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)) | |
197 | return; | |
198 | ||
199 | do { | |
200 | cpu_relax(); | |
201 | } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)); | |
202 | ||
203 | nmi_wait_count++; | |
204 | } | |
205 | ||
17666f02 SR |
206 | static void wait_for_nmi(void) |
207 | { | |
4e6ea144 | 208 | if (!atomic_read(&nmi_running)) |
89025282 | 209 | return; |
b807c3d0 | 210 | |
89025282 | 211 | do { |
17666f02 | 212 | cpu_relax(); |
4e6ea144 | 213 | } while (atomic_read(&nmi_running)); |
b807c3d0 | 214 | |
89025282 | 215 | nmi_wait_count++; |
17666f02 SR |
216 | } |
217 | ||
55ca3cc1 SS |
218 | static inline int |
219 | within(unsigned long addr, unsigned long start, unsigned long end) | |
220 | { | |
221 | return addr >= start && addr < end; | |
222 | } | |
223 | ||
17666f02 SR |
224 | static int |
225 | do_ftrace_mod_code(unsigned long ip, void *new_code) | |
226 | { | |
55ca3cc1 SS |
227 | /* |
228 | * On x86_64, kernel text mappings are mapped read-only with | |
229 | * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead | |
230 | * of the kernel text mapping to modify the kernel text. | |
231 | * | |
232 | * For 32bit kernels, these mappings are same and we can use | |
233 | * kernel identity mapping to modify code. | |
234 | */ | |
235 | if (within(ip, (unsigned long)_text, (unsigned long)_etext)) | |
236 | ip = (unsigned long)__va(__pa(ip)); | |
237 | ||
17666f02 SR |
238 | mod_code_ip = (void *)ip; |
239 | mod_code_newcode = new_code; | |
240 | ||
241 | /* The buffers need to be visible before we let NMIs write them */ | |
17666f02 SR |
242 | smp_mb(); |
243 | ||
e9d9df44 | 244 | wait_for_nmi_and_set_mod_flag(); |
17666f02 SR |
245 | |
246 | /* Make sure all running NMIs have finished before we write the code */ | |
247 | smp_mb(); | |
248 | ||
249 | ftrace_mod_code(); | |
250 | ||
251 | /* Make sure the write happens before clearing the bit */ | |
17666f02 SR |
252 | smp_mb(); |
253 | ||
e9d9df44 | 254 | clear_mod_flag(); |
17666f02 SR |
255 | wait_for_nmi(); |
256 | ||
257 | return mod_code_status; | |
258 | } | |
259 | ||
31e88909 | 260 | static unsigned char *ftrace_nop_replace(void) |
caf4b323 | 261 | { |
f49aa448 | 262 | return ideal_nop5; |
caf4b323 FW |
263 | } |
264 | ||
31e88909 | 265 | static int |
3d083395 SR |
266 | ftrace_modify_code(unsigned long ip, unsigned char *old_code, |
267 | unsigned char *new_code) | |
268 | { | |
6f93fc07 | 269 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
3d083395 SR |
270 | |
271 | /* | |
272 | * Note: Due to modules and __init, code can | |
273 | * disappear and change, we need to protect against faulting | |
76aefee5 | 274 | * as well as code changing. We do this by using the |
ab9a0918 | 275 | * probe_kernel_* functions. |
3d083395 SR |
276 | * |
277 | * No real locking needed, this code is run through | |
6f93fc07 | 278 | * kstop_machine, or before SMP starts. |
3d083395 | 279 | */ |
76aefee5 SR |
280 | |
281 | /* read the text we want to modify */ | |
ab9a0918 | 282 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
593eb8a2 | 283 | return -EFAULT; |
6f93fc07 | 284 | |
76aefee5 | 285 | /* Make sure it is what we expect it to be */ |
6f93fc07 | 286 | if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) |
593eb8a2 | 287 | return -EINVAL; |
3d083395 | 288 | |
76aefee5 | 289 | /* replace the text with the new text */ |
17666f02 | 290 | if (do_ftrace_mod_code(ip, new_code)) |
593eb8a2 | 291 | return -EPERM; |
6f93fc07 SR |
292 | |
293 | sync_core(); | |
3d083395 | 294 | |
6f93fc07 | 295 | return 0; |
3d083395 SR |
296 | } |
297 | ||
31e88909 SR |
298 | int ftrace_make_nop(struct module *mod, |
299 | struct dyn_ftrace *rec, unsigned long addr) | |
300 | { | |
301 | unsigned char *new, *old; | |
302 | unsigned long ip = rec->ip; | |
303 | ||
304 | old = ftrace_call_replace(ip, addr); | |
305 | new = ftrace_nop_replace(); | |
306 | ||
307 | return ftrace_modify_code(rec->ip, old, new); | |
308 | } | |
309 | ||
310 | int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) | |
311 | { | |
312 | unsigned char *new, *old; | |
313 | unsigned long ip = rec->ip; | |
314 | ||
315 | old = ftrace_nop_replace(); | |
316 | new = ftrace_call_replace(ip, addr); | |
317 | ||
318 | return ftrace_modify_code(rec->ip, old, new); | |
319 | } | |
320 | ||
15adc048 | 321 | int ftrace_update_ftrace_func(ftrace_func_t func) |
d61f82d0 SR |
322 | { |
323 | unsigned long ip = (unsigned long)(&ftrace_call); | |
395a59d0 | 324 | unsigned char old[MCOUNT_INSN_SIZE], *new; |
d61f82d0 SR |
325 | int ret; |
326 | ||
395a59d0 | 327 | memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE); |
d61f82d0 SR |
328 | new = ftrace_call_replace(ip, (unsigned long)func); |
329 | ret = ftrace_modify_code(ip, old, new); | |
330 | ||
331 | return ret; | |
332 | } | |
333 | ||
d61f82d0 | 334 | int __init ftrace_dyn_arch_init(void *data) |
3d083395 | 335 | { |
732f3ca7 SR |
336 | /* The return code is retured via data */ |
337 | *(unsigned long *)data = 0; | |
dfa60aba | 338 | |
3d083395 SR |
339 | return 0; |
340 | } | |
caf4b323 | 341 | #endif |
e7d3737e | 342 | |
fb52607a | 343 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER |
e7d3737e | 344 | |
5a45cfe1 SR |
345 | #ifdef CONFIG_DYNAMIC_FTRACE |
346 | extern void ftrace_graph_call(void); | |
347 | ||
348 | static int ftrace_mod_jmp(unsigned long ip, | |
349 | int old_offset, int new_offset) | |
350 | { | |
351 | unsigned char code[MCOUNT_INSN_SIZE]; | |
352 | ||
353 | if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE)) | |
354 | return -EFAULT; | |
355 | ||
356 | if (code[0] != 0xe9 || old_offset != *(int *)(&code[1])) | |
357 | return -EINVAL; | |
358 | ||
359 | *(int *)(&code[1]) = new_offset; | |
360 | ||
361 | if (do_ftrace_mod_code(ip, &code)) | |
362 | return -EPERM; | |
363 | ||
364 | return 0; | |
365 | } | |
366 | ||
367 | int ftrace_enable_ftrace_graph_caller(void) | |
368 | { | |
369 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
370 | int old_offset, new_offset; | |
371 | ||
372 | old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
373 | new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
374 | ||
375 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
376 | } | |
377 | ||
378 | int ftrace_disable_ftrace_graph_caller(void) | |
379 | { | |
380 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
381 | int old_offset, new_offset; | |
382 | ||
383 | old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
384 | new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
385 | ||
386 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
387 | } | |
388 | ||
e7d3737e FW |
389 | #endif /* !CONFIG_DYNAMIC_FTRACE */ |
390 | ||
e7d3737e FW |
391 | /* |
392 | * Hook the return address and push it in the stack of return addrs | |
393 | * in current thread info. | |
394 | */ | |
71e308a2 SR |
395 | void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr, |
396 | unsigned long frame_pointer) | |
e7d3737e FW |
397 | { |
398 | unsigned long old; | |
e7d3737e | 399 | int faulted; |
287b6e68 | 400 | struct ftrace_graph_ent trace; |
e7d3737e FW |
401 | unsigned long return_hooker = (unsigned long) |
402 | &return_to_handler; | |
403 | ||
380c4b14 | 404 | if (unlikely(atomic_read(¤t->tracing_graph_pause))) |
e7d3737e FW |
405 | return; |
406 | ||
407 | /* | |
408 | * Protect against fault, even if it shouldn't | |
409 | * happen. This tool is too much intrusive to | |
410 | * ignore such a protection. | |
411 | */ | |
412 | asm volatile( | |
96665788 SR |
413 | "1: " _ASM_MOV " (%[parent]), %[old]\n" |
414 | "2: " _ASM_MOV " %[return_hooker], (%[parent])\n" | |
e7d3737e | 415 | " movl $0, %[faulted]\n" |
e3944bfa | 416 | "3:\n" |
e7d3737e FW |
417 | |
418 | ".section .fixup, \"ax\"\n" | |
e3944bfa SR |
419 | "4: movl $1, %[faulted]\n" |
420 | " jmp 3b\n" | |
e7d3737e FW |
421 | ".previous\n" |
422 | ||
e3944bfa SR |
423 | _ASM_EXTABLE(1b, 4b) |
424 | _ASM_EXTABLE(2b, 4b) | |
e7d3737e | 425 | |
aa512a27 | 426 | : [old] "=&r" (old), [faulted] "=r" (faulted) |
96665788 | 427 | : [parent] "r" (parent), [return_hooker] "r" (return_hooker) |
e7d3737e FW |
428 | : "memory" |
429 | ); | |
430 | ||
14a866c5 SR |
431 | if (unlikely(faulted)) { |
432 | ftrace_graph_stop(); | |
433 | WARN_ON(1); | |
e7d3737e FW |
434 | return; |
435 | } | |
436 | ||
71e308a2 SR |
437 | if (ftrace_push_return_trace(old, self_addr, &trace.depth, |
438 | frame_pointer) == -EBUSY) { | |
e7d3737e | 439 | *parent = old; |
287b6e68 FW |
440 | return; |
441 | } | |
442 | ||
443 | trace.func = self_addr; | |
287b6e68 | 444 | |
e49dc19c SR |
445 | /* Only trace if the calling function expects to */ |
446 | if (!ftrace_graph_entry(&trace)) { | |
447 | current->curr_ret_stack--; | |
448 | *parent = old; | |
449 | } | |
e7d3737e | 450 | } |
fb52607a | 451 | #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ |