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 | ||
12 | #include <linux/spinlock.h> | |
13 | #include <linux/hardirq.h> | |
6f93fc07 | 14 | #include <linux/uaccess.h> |
3d083395 SR |
15 | #include <linux/ftrace.h> |
16 | #include <linux/percpu.h> | |
19b3e967 | 17 | #include <linux/sched.h> |
3d083395 SR |
18 | #include <linux/init.h> |
19 | #include <linux/list.h> | |
20 | ||
47788c58 FW |
21 | #include <trace/syscall.h> |
22 | ||
16239630 | 23 | #include <asm/cacheflush.h> |
395a59d0 | 24 | #include <asm/ftrace.h> |
732f3ca7 | 25 | #include <asm/nops.h> |
caf4b323 | 26 | #include <asm/nmi.h> |
3d083395 | 27 | |
3d083395 | 28 | |
caf4b323 | 29 | #ifdef CONFIG_DYNAMIC_FTRACE |
3d083395 | 30 | |
16239630 SR |
31 | int ftrace_arch_code_modify_prepare(void) |
32 | { | |
33 | set_kernel_text_rw(); | |
34 | return 0; | |
35 | } | |
36 | ||
37 | int ftrace_arch_code_modify_post_process(void) | |
38 | { | |
39 | set_kernel_text_ro(); | |
40 | return 0; | |
41 | } | |
42 | ||
3d083395 | 43 | union ftrace_code_union { |
395a59d0 | 44 | char code[MCOUNT_INSN_SIZE]; |
3d083395 SR |
45 | struct { |
46 | char e8; | |
47 | int offset; | |
48 | } __attribute__((packed)); | |
49 | }; | |
50 | ||
15adc048 | 51 | static int ftrace_calc_offset(long ip, long addr) |
3c1720f0 SR |
52 | { |
53 | return (int)(addr - ip); | |
54 | } | |
3d083395 | 55 | |
31e88909 | 56 | static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) |
3c1720f0 SR |
57 | { |
58 | static union ftrace_code_union calc; | |
3d083395 | 59 | |
3c1720f0 | 60 | calc.e8 = 0xe8; |
395a59d0 | 61 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
3c1720f0 SR |
62 | |
63 | /* | |
64 | * No locking needed, this must be called via kstop_machine | |
65 | * which in essence is like running on a uniprocessor machine. | |
66 | */ | |
67 | return calc.code; | |
3d083395 SR |
68 | } |
69 | ||
17666f02 SR |
70 | /* |
71 | * Modifying code must take extra care. On an SMP machine, if | |
72 | * the code being modified is also being executed on another CPU | |
73 | * that CPU will have undefined results and possibly take a GPF. | |
74 | * We use kstop_machine to stop other CPUS from exectuing code. | |
75 | * But this does not stop NMIs from happening. We still need | |
76 | * to protect against that. We separate out the modification of | |
77 | * the code to take care of this. | |
78 | * | |
79 | * Two buffers are added: An IP buffer and a "code" buffer. | |
80 | * | |
a26a2a27 | 81 | * 1) Put the instruction pointer into the IP buffer |
17666f02 | 82 | * and the new code into the "code" buffer. |
e9d9df44 LJ |
83 | * 2) Wait for any running NMIs to finish and set a flag that says |
84 | * we are modifying code, it is done in an atomic operation. | |
85 | * 3) Write the code | |
86 | * 4) clear the flag. | |
87 | * 5) Wait for any running NMIs to finish. | |
17666f02 SR |
88 | * |
89 | * If an NMI is executed, the first thing it does is to call | |
90 | * "ftrace_nmi_enter". This will check if the flag is set to write | |
91 | * and if it is, it will write what is in the IP and "code" buffers. | |
92 | * | |
93 | * The trick is, it does not matter if everyone is writing the same | |
94 | * content to the code location. Also, if a CPU is executing code | |
95 | * it is OK to write to that code location if the contents being written | |
96 | * are the same as what exists. | |
97 | */ | |
98 | ||
e9d9df44 | 99 | #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */ |
4e6ea144 | 100 | static atomic_t nmi_running = ATOMIC_INIT(0); |
a26a2a27 | 101 | static int mod_code_status; /* holds return value of text write */ |
a26a2a27 SR |
102 | static void *mod_code_ip; /* holds the IP to write to */ |
103 | static void *mod_code_newcode; /* holds the text to write to the IP */ | |
17666f02 | 104 | |
a26a2a27 SR |
105 | static unsigned nmi_wait_count; |
106 | static atomic_t nmi_update_count = ATOMIC_INIT(0); | |
b807c3d0 SR |
107 | |
108 | int ftrace_arch_read_dyn_info(char *buf, int size) | |
109 | { | |
110 | int r; | |
111 | ||
112 | r = snprintf(buf, size, "%u %u", | |
113 | nmi_wait_count, | |
114 | atomic_read(&nmi_update_count)); | |
115 | return r; | |
116 | } | |
117 | ||
e9d9df44 LJ |
118 | static void clear_mod_flag(void) |
119 | { | |
120 | int old = atomic_read(&nmi_running); | |
121 | ||
122 | for (;;) { | |
123 | int new = old & ~MOD_CODE_WRITE_FLAG; | |
124 | ||
125 | if (old == new) | |
126 | break; | |
127 | ||
128 | old = atomic_cmpxchg(&nmi_running, old, new); | |
129 | } | |
130 | } | |
131 | ||
17666f02 SR |
132 | static void ftrace_mod_code(void) |
133 | { | |
134 | /* | |
135 | * Yes, more than one CPU process can be writing to mod_code_status. | |
136 | * (and the code itself) | |
137 | * But if one were to fail, then they all should, and if one were | |
138 | * to succeed, then they all should. | |
139 | */ | |
140 | mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode, | |
141 | MCOUNT_INSN_SIZE); | |
90c7ac49 SR |
142 | |
143 | /* if we fail, then kill any new writers */ | |
144 | if (mod_code_status) | |
e9d9df44 | 145 | clear_mod_flag(); |
17666f02 SR |
146 | } |
147 | ||
a81bd80a | 148 | void ftrace_nmi_enter(void) |
17666f02 | 149 | { |
e9d9df44 LJ |
150 | if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) { |
151 | smp_rmb(); | |
17666f02 | 152 | ftrace_mod_code(); |
b807c3d0 SR |
153 | atomic_inc(&nmi_update_count); |
154 | } | |
e9d9df44 LJ |
155 | /* Must have previous changes seen before executions */ |
156 | smp_mb(); | |
17666f02 SR |
157 | } |
158 | ||
a81bd80a | 159 | void ftrace_nmi_exit(void) |
17666f02 | 160 | { |
4e6ea144 | 161 | /* Finish all executions before clearing nmi_running */ |
e9d9df44 | 162 | smp_mb(); |
4e6ea144 | 163 | atomic_dec(&nmi_running); |
17666f02 SR |
164 | } |
165 | ||
e9d9df44 LJ |
166 | static void wait_for_nmi_and_set_mod_flag(void) |
167 | { | |
168 | if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)) | |
169 | return; | |
170 | ||
171 | do { | |
172 | cpu_relax(); | |
173 | } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)); | |
174 | ||
175 | nmi_wait_count++; | |
176 | } | |
177 | ||
17666f02 SR |
178 | static void wait_for_nmi(void) |
179 | { | |
4e6ea144 | 180 | if (!atomic_read(&nmi_running)) |
89025282 | 181 | return; |
b807c3d0 | 182 | |
89025282 | 183 | do { |
17666f02 | 184 | cpu_relax(); |
4e6ea144 | 185 | } while (atomic_read(&nmi_running)); |
b807c3d0 | 186 | |
89025282 | 187 | nmi_wait_count++; |
17666f02 SR |
188 | } |
189 | ||
190 | static int | |
191 | do_ftrace_mod_code(unsigned long ip, void *new_code) | |
192 | { | |
193 | mod_code_ip = (void *)ip; | |
194 | mod_code_newcode = new_code; | |
195 | ||
196 | /* The buffers need to be visible before we let NMIs write them */ | |
17666f02 SR |
197 | smp_mb(); |
198 | ||
e9d9df44 | 199 | wait_for_nmi_and_set_mod_flag(); |
17666f02 SR |
200 | |
201 | /* Make sure all running NMIs have finished before we write the code */ | |
202 | smp_mb(); | |
203 | ||
204 | ftrace_mod_code(); | |
205 | ||
206 | /* Make sure the write happens before clearing the bit */ | |
17666f02 SR |
207 | smp_mb(); |
208 | ||
e9d9df44 | 209 | clear_mod_flag(); |
17666f02 SR |
210 | wait_for_nmi(); |
211 | ||
212 | return mod_code_status; | |
213 | } | |
214 | ||
215 | ||
caf4b323 FW |
216 | |
217 | ||
218 | static unsigned char ftrace_nop[MCOUNT_INSN_SIZE]; | |
219 | ||
31e88909 | 220 | static unsigned char *ftrace_nop_replace(void) |
caf4b323 FW |
221 | { |
222 | return ftrace_nop; | |
223 | } | |
224 | ||
31e88909 | 225 | static int |
3d083395 SR |
226 | ftrace_modify_code(unsigned long ip, unsigned char *old_code, |
227 | unsigned char *new_code) | |
228 | { | |
6f93fc07 | 229 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
3d083395 SR |
230 | |
231 | /* | |
232 | * Note: Due to modules and __init, code can | |
233 | * disappear and change, we need to protect against faulting | |
76aefee5 | 234 | * as well as code changing. We do this by using the |
ab9a0918 | 235 | * probe_kernel_* functions. |
3d083395 SR |
236 | * |
237 | * No real locking needed, this code is run through | |
6f93fc07 | 238 | * kstop_machine, or before SMP starts. |
3d083395 | 239 | */ |
76aefee5 SR |
240 | |
241 | /* read the text we want to modify */ | |
ab9a0918 | 242 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
593eb8a2 | 243 | return -EFAULT; |
6f93fc07 | 244 | |
76aefee5 | 245 | /* Make sure it is what we expect it to be */ |
6f93fc07 | 246 | if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) |
593eb8a2 | 247 | return -EINVAL; |
3d083395 | 248 | |
76aefee5 | 249 | /* replace the text with the new text */ |
17666f02 | 250 | if (do_ftrace_mod_code(ip, new_code)) |
593eb8a2 | 251 | return -EPERM; |
6f93fc07 SR |
252 | |
253 | sync_core(); | |
3d083395 | 254 | |
6f93fc07 | 255 | return 0; |
3d083395 SR |
256 | } |
257 | ||
31e88909 SR |
258 | int ftrace_make_nop(struct module *mod, |
259 | struct dyn_ftrace *rec, unsigned long addr) | |
260 | { | |
261 | unsigned char *new, *old; | |
262 | unsigned long ip = rec->ip; | |
263 | ||
264 | old = ftrace_call_replace(ip, addr); | |
265 | new = ftrace_nop_replace(); | |
266 | ||
267 | return ftrace_modify_code(rec->ip, old, new); | |
268 | } | |
269 | ||
270 | int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) | |
271 | { | |
272 | unsigned char *new, *old; | |
273 | unsigned long ip = rec->ip; | |
274 | ||
275 | old = ftrace_nop_replace(); | |
276 | new = ftrace_call_replace(ip, addr); | |
277 | ||
278 | return ftrace_modify_code(rec->ip, old, new); | |
279 | } | |
280 | ||
15adc048 | 281 | int ftrace_update_ftrace_func(ftrace_func_t func) |
d61f82d0 SR |
282 | { |
283 | unsigned long ip = (unsigned long)(&ftrace_call); | |
395a59d0 | 284 | unsigned char old[MCOUNT_INSN_SIZE], *new; |
d61f82d0 SR |
285 | int ret; |
286 | ||
395a59d0 | 287 | memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE); |
d61f82d0 SR |
288 | new = ftrace_call_replace(ip, (unsigned long)func); |
289 | ret = ftrace_modify_code(ip, old, new); | |
290 | ||
291 | return ret; | |
292 | } | |
293 | ||
d61f82d0 | 294 | int __init ftrace_dyn_arch_init(void *data) |
3d083395 | 295 | { |
732f3ca7 SR |
296 | extern const unsigned char ftrace_test_p6nop[]; |
297 | extern const unsigned char ftrace_test_nop5[]; | |
298 | extern const unsigned char ftrace_test_jmp[]; | |
299 | int faulted = 0; | |
d61f82d0 | 300 | |
732f3ca7 SR |
301 | /* |
302 | * There is no good nop for all x86 archs. | |
303 | * We will default to using the P6_NOP5, but first we | |
304 | * will test to make sure that the nop will actually | |
305 | * work on this CPU. If it faults, we will then | |
306 | * go to a lesser efficient 5 byte nop. If that fails | |
307 | * we then just use a jmp as our nop. This isn't the most | |
308 | * efficient nop, but we can not use a multi part nop | |
309 | * since we would then risk being preempted in the middle | |
310 | * of that nop, and if we enabled tracing then, it might | |
311 | * cause a system crash. | |
312 | * | |
313 | * TODO: check the cpuid to determine the best nop. | |
314 | */ | |
315 | asm volatile ( | |
732f3ca7 SR |
316 | "ftrace_test_jmp:" |
317 | "jmp ftrace_test_p6nop\n" | |
8b27386a AK |
318 | "nop\n" |
319 | "nop\n" | |
320 | "nop\n" /* 2 byte jmp + 3 bytes */ | |
732f3ca7 SR |
321 | "ftrace_test_p6nop:" |
322 | P6_NOP5 | |
323 | "jmp 1f\n" | |
324 | "ftrace_test_nop5:" | |
325 | ".byte 0x66,0x66,0x66,0x66,0x90\n" | |
732f3ca7 SR |
326 | "1:" |
327 | ".section .fixup, \"ax\"\n" | |
328 | "2: movl $1, %0\n" | |
329 | " jmp ftrace_test_nop5\n" | |
330 | "3: movl $2, %0\n" | |
331 | " jmp 1b\n" | |
332 | ".previous\n" | |
333 | _ASM_EXTABLE(ftrace_test_p6nop, 2b) | |
334 | _ASM_EXTABLE(ftrace_test_nop5, 3b) | |
335 | : "=r"(faulted) : "0" (faulted)); | |
336 | ||
337 | switch (faulted) { | |
338 | case 0: | |
339 | pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n"); | |
8115f3f0 | 340 | memcpy(ftrace_nop, ftrace_test_p6nop, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
341 | break; |
342 | case 1: | |
343 | pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n"); | |
8115f3f0 | 344 | memcpy(ftrace_nop, ftrace_test_nop5, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
345 | break; |
346 | case 2: | |
8b27386a | 347 | pr_info("ftrace: converting mcount calls to jmp . + 5\n"); |
8115f3f0 | 348 | memcpy(ftrace_nop, ftrace_test_jmp, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
349 | break; |
350 | } | |
351 | ||
352 | /* The return code is retured via data */ | |
353 | *(unsigned long *)data = 0; | |
dfa60aba | 354 | |
3d083395 SR |
355 | return 0; |
356 | } | |
caf4b323 | 357 | #endif |
e7d3737e | 358 | |
fb52607a | 359 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER |
e7d3737e | 360 | |
5a45cfe1 SR |
361 | #ifdef CONFIG_DYNAMIC_FTRACE |
362 | extern void ftrace_graph_call(void); | |
363 | ||
364 | static int ftrace_mod_jmp(unsigned long ip, | |
365 | int old_offset, int new_offset) | |
366 | { | |
367 | unsigned char code[MCOUNT_INSN_SIZE]; | |
368 | ||
369 | if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE)) | |
370 | return -EFAULT; | |
371 | ||
372 | if (code[0] != 0xe9 || old_offset != *(int *)(&code[1])) | |
373 | return -EINVAL; | |
374 | ||
375 | *(int *)(&code[1]) = new_offset; | |
376 | ||
377 | if (do_ftrace_mod_code(ip, &code)) | |
378 | return -EPERM; | |
379 | ||
380 | return 0; | |
381 | } | |
382 | ||
383 | int ftrace_enable_ftrace_graph_caller(void) | |
384 | { | |
385 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
386 | int old_offset, new_offset; | |
387 | ||
388 | old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
389 | new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
390 | ||
391 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
392 | } | |
393 | ||
394 | int ftrace_disable_ftrace_graph_caller(void) | |
395 | { | |
396 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
397 | int old_offset, new_offset; | |
398 | ||
399 | old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
400 | new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
401 | ||
402 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
403 | } | |
404 | ||
e7d3737e FW |
405 | #endif /* !CONFIG_DYNAMIC_FTRACE */ |
406 | ||
e7d3737e FW |
407 | /* |
408 | * Hook the return address and push it in the stack of return addrs | |
409 | * in current thread info. | |
410 | */ | |
71e308a2 SR |
411 | void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr, |
412 | unsigned long frame_pointer) | |
e7d3737e FW |
413 | { |
414 | unsigned long old; | |
e7d3737e | 415 | int faulted; |
287b6e68 | 416 | struct ftrace_graph_ent trace; |
e7d3737e FW |
417 | unsigned long return_hooker = (unsigned long) |
418 | &return_to_handler; | |
419 | ||
380c4b14 | 420 | if (unlikely(atomic_read(¤t->tracing_graph_pause))) |
e7d3737e FW |
421 | return; |
422 | ||
423 | /* | |
424 | * Protect against fault, even if it shouldn't | |
425 | * happen. This tool is too much intrusive to | |
426 | * ignore such a protection. | |
427 | */ | |
428 | asm volatile( | |
96665788 SR |
429 | "1: " _ASM_MOV " (%[parent]), %[old]\n" |
430 | "2: " _ASM_MOV " %[return_hooker], (%[parent])\n" | |
e7d3737e | 431 | " movl $0, %[faulted]\n" |
e3944bfa | 432 | "3:\n" |
e7d3737e FW |
433 | |
434 | ".section .fixup, \"ax\"\n" | |
e3944bfa SR |
435 | "4: movl $1, %[faulted]\n" |
436 | " jmp 3b\n" | |
e7d3737e FW |
437 | ".previous\n" |
438 | ||
e3944bfa SR |
439 | _ASM_EXTABLE(1b, 4b) |
440 | _ASM_EXTABLE(2b, 4b) | |
e7d3737e | 441 | |
aa512a27 | 442 | : [old] "=&r" (old), [faulted] "=r" (faulted) |
96665788 | 443 | : [parent] "r" (parent), [return_hooker] "r" (return_hooker) |
e7d3737e FW |
444 | : "memory" |
445 | ); | |
446 | ||
14a866c5 SR |
447 | if (unlikely(faulted)) { |
448 | ftrace_graph_stop(); | |
449 | WARN_ON(1); | |
e7d3737e FW |
450 | return; |
451 | } | |
452 | ||
71e308a2 SR |
453 | if (ftrace_push_return_trace(old, self_addr, &trace.depth, |
454 | frame_pointer) == -EBUSY) { | |
e7d3737e | 455 | *parent = old; |
287b6e68 FW |
456 | return; |
457 | } | |
458 | ||
459 | trace.func = self_addr; | |
287b6e68 | 460 | |
e49dc19c SR |
461 | /* Only trace if the calling function expects to */ |
462 | if (!ftrace_graph_entry(&trace)) { | |
463 | current->curr_ret_stack--; | |
464 | *parent = old; | |
465 | } | |
e7d3737e | 466 | } |
fb52607a | 467 | #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ |
f58ba100 FW |
468 | |
469 | #ifdef CONFIG_FTRACE_SYSCALLS | |
470 | ||
471 | extern unsigned long __start_syscalls_metadata[]; | |
472 | extern unsigned long __stop_syscalls_metadata[]; | |
473 | extern unsigned long *sys_call_table; | |
474 | ||
475 | static struct syscall_metadata **syscalls_metadata; | |
476 | ||
477 | static struct syscall_metadata *find_syscall_meta(unsigned long *syscall) | |
478 | { | |
479 | struct syscall_metadata *start; | |
480 | struct syscall_metadata *stop; | |
481 | char str[KSYM_SYMBOL_LEN]; | |
482 | ||
483 | ||
484 | start = (struct syscall_metadata *)__start_syscalls_metadata; | |
485 | stop = (struct syscall_metadata *)__stop_syscalls_metadata; | |
486 | kallsyms_lookup((unsigned long) syscall, NULL, NULL, NULL, str); | |
487 | ||
488 | for ( ; start < stop; start++) { | |
489 | if (start->name && !strcmp(start->name, str)) | |
490 | return start; | |
491 | } | |
492 | return NULL; | |
493 | } | |
494 | ||
495 | struct syscall_metadata *syscall_nr_to_meta(int nr) | |
496 | { | |
497 | if (!syscalls_metadata || nr >= FTRACE_SYSCALL_MAX || nr < 0) | |
498 | return NULL; | |
499 | ||
500 | return syscalls_metadata[nr]; | |
501 | } | |
502 | ||
503 | void arch_init_ftrace_syscalls(void) | |
504 | { | |
505 | int i; | |
506 | struct syscall_metadata *meta; | |
507 | unsigned long **psys_syscall_table = &sys_call_table; | |
508 | static atomic_t refs; | |
509 | ||
510 | if (atomic_inc_return(&refs) != 1) | |
511 | goto end; | |
512 | ||
513 | syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) * | |
514 | FTRACE_SYSCALL_MAX, GFP_KERNEL); | |
515 | if (!syscalls_metadata) { | |
516 | WARN_ON(1); | |
517 | return; | |
518 | } | |
519 | ||
520 | for (i = 0; i < FTRACE_SYSCALL_MAX; i++) { | |
521 | meta = find_syscall_meta(psys_syscall_table[i]); | |
522 | syscalls_metadata[i] = meta; | |
523 | } | |
524 | return; | |
525 | ||
526 | /* Paranoid: avoid overflow */ | |
527 | end: | |
528 | atomic_dec(&refs); | |
529 | } | |
530 | #endif |