]>
Commit | Line | Data |
---|---|---|
3d083395 | 1 | /* |
9d2099ab | 2 | * Dynamic function tracing support. |
3d083395 SR |
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> |
f3bea491 | 20 | #include <linux/slab.h> |
3d083395 SR |
21 | #include <linux/init.h> |
22 | #include <linux/list.h> | |
84e1c6bb | 23 | #include <linux/module.h> |
3d083395 | 24 | |
47788c58 FW |
25 | #include <trace/syscall.h> |
26 | ||
d1163651 | 27 | #include <asm/set_memory.h> |
59a094c9 | 28 | #include <asm/kprobes.h> |
395a59d0 | 29 | #include <asm/ftrace.h> |
732f3ca7 | 30 | #include <asm/nops.h> |
3d083395 | 31 | |
caf4b323 | 32 | #ifdef CONFIG_DYNAMIC_FTRACE |
3d083395 | 33 | |
16239630 SR |
34 | int ftrace_arch_code_modify_prepare(void) |
35 | { | |
36 | set_kernel_text_rw(); | |
84e1c6bb | 37 | set_all_modules_text_rw(); |
16239630 SR |
38 | return 0; |
39 | } | |
40 | ||
41 | int ftrace_arch_code_modify_post_process(void) | |
42 | { | |
84e1c6bb | 43 | set_all_modules_text_ro(); |
16239630 SR |
44 | set_kernel_text_ro(); |
45 | return 0; | |
46 | } | |
47 | ||
3d083395 | 48 | union ftrace_code_union { |
395a59d0 | 49 | char code[MCOUNT_INSN_SIZE]; |
3d083395 | 50 | struct { |
15d5b02c | 51 | unsigned char e8; |
3d083395 SR |
52 | int offset; |
53 | } __attribute__((packed)); | |
54 | }; | |
55 | ||
15adc048 | 56 | static int ftrace_calc_offset(long ip, long addr) |
3c1720f0 SR |
57 | { |
58 | return (int)(addr - ip); | |
59 | } | |
3d083395 | 60 | |
31e88909 | 61 | static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) |
3c1720f0 SR |
62 | { |
63 | static union ftrace_code_union calc; | |
3d083395 | 64 | |
3c1720f0 | 65 | calc.e8 = 0xe8; |
395a59d0 | 66 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
3c1720f0 SR |
67 | |
68 | /* | |
69 | * No locking needed, this must be called via kstop_machine | |
70 | * which in essence is like running on a uniprocessor machine. | |
71 | */ | |
72 | return calc.code; | |
3d083395 SR |
73 | } |
74 | ||
55ca3cc1 SS |
75 | static inline int |
76 | within(unsigned long addr, unsigned long start, unsigned long end) | |
77 | { | |
78 | return addr >= start && addr < end; | |
79 | } | |
80 | ||
87fbb2ac | 81 | static unsigned long text_ip_addr(unsigned long ip) |
17666f02 | 82 | { |
55ca3cc1 | 83 | /* |
9ccaf77c KC |
84 | * On x86_64, kernel text mappings are mapped read-only, so we use |
85 | * the kernel identity mapping instead of the kernel text mapping | |
86 | * to modify the kernel text. | |
55ca3cc1 SS |
87 | * |
88 | * For 32bit kernels, these mappings are same and we can use | |
89 | * kernel identity mapping to modify code. | |
90 | */ | |
91 | if (within(ip, (unsigned long)_text, (unsigned long)_etext)) | |
217f155e | 92 | ip = (unsigned long)__va(__pa_symbol(ip)); |
55ca3cc1 | 93 | |
87fbb2ac | 94 | return ip; |
17666f02 SR |
95 | } |
96 | ||
dc326fca | 97 | static const unsigned char *ftrace_nop_replace(void) |
caf4b323 | 98 | { |
dc326fca | 99 | return ideal_nops[NOP_ATOMIC5]; |
caf4b323 FW |
100 | } |
101 | ||
31e88909 | 102 | static int |
8a4d0a68 | 103 | ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code, |
0d098a7d | 104 | unsigned const char *new_code) |
3d083395 | 105 | { |
6f93fc07 | 106 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
3d083395 | 107 | |
b05086c7 SRRH |
108 | ftrace_expected = old_code; |
109 | ||
3d083395 | 110 | /* |
c5d641f9 LB |
111 | * Note: |
112 | * We are paranoid about modifying text, as if a bug was to happen, it | |
113 | * could cause us to read or write to someplace that could cause harm. | |
114 | * Carefully read and modify the code with probe_kernel_*(), and make | |
115 | * sure what we read is what we expected it to be before modifying it. | |
3d083395 | 116 | */ |
76aefee5 SR |
117 | |
118 | /* read the text we want to modify */ | |
ab9a0918 | 119 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
593eb8a2 | 120 | return -EFAULT; |
6f93fc07 | 121 | |
76aefee5 | 122 | /* Make sure it is what we expect it to be */ |
6f93fc07 | 123 | if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) |
593eb8a2 | 124 | return -EINVAL; |
3d083395 | 125 | |
87fbb2ac SRRH |
126 | ip = text_ip_addr(ip); |
127 | ||
76aefee5 | 128 | /* replace the text with the new text */ |
87fbb2ac | 129 | if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE)) |
593eb8a2 | 130 | return -EPERM; |
6f93fc07 SR |
131 | |
132 | sync_core(); | |
3d083395 | 133 | |
6f93fc07 | 134 | return 0; |
3d083395 SR |
135 | } |
136 | ||
31e88909 SR |
137 | int ftrace_make_nop(struct module *mod, |
138 | struct dyn_ftrace *rec, unsigned long addr) | |
139 | { | |
0d098a7d | 140 | unsigned const char *new, *old; |
31e88909 SR |
141 | unsigned long ip = rec->ip; |
142 | ||
143 | old = ftrace_call_replace(ip, addr); | |
144 | new = ftrace_nop_replace(); | |
145 | ||
8a4d0a68 SR |
146 | /* |
147 | * On boot up, and when modules are loaded, the MCOUNT_ADDR | |
148 | * is converted to a nop, and will never become MCOUNT_ADDR | |
149 | * again. This code is either running before SMP (on boot up) | |
150 | * or before the code will ever be executed (module load). | |
151 | * We do not want to use the breakpoint version in this case, | |
152 | * just modify the code directly. | |
153 | */ | |
154 | if (addr == MCOUNT_ADDR) | |
155 | return ftrace_modify_code_direct(rec->ip, old, new); | |
156 | ||
b05086c7 SRRH |
157 | ftrace_expected = NULL; |
158 | ||
8a4d0a68 SR |
159 | /* Normal cases use add_brk_on_nop */ |
160 | WARN_ONCE(1, "invalid use of ftrace_make_nop"); | |
161 | return -EINVAL; | |
31e88909 SR |
162 | } |
163 | ||
164 | int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) | |
165 | { | |
0d098a7d | 166 | unsigned const char *new, *old; |
31e88909 SR |
167 | unsigned long ip = rec->ip; |
168 | ||
169 | old = ftrace_nop_replace(); | |
170 | new = ftrace_call_replace(ip, addr); | |
171 | ||
8a4d0a68 SR |
172 | /* Should only be called when module is loaded */ |
173 | return ftrace_modify_code_direct(rec->ip, old, new); | |
d61f82d0 SR |
174 | } |
175 | ||
a192cd04 SR |
176 | /* |
177 | * The modifying_ftrace_code is used to tell the breakpoint | |
178 | * handler to call ftrace_int3_handler(). If it fails to | |
179 | * call this handler for a breakpoint added by ftrace, then | |
180 | * the kernel may crash. | |
181 | * | |
182 | * As atomic_writes on x86 do not need a barrier, we do not | |
183 | * need to add smp_mb()s for this to work. It is also considered | |
184 | * that we can not read the modifying_ftrace_code before | |
185 | * executing the breakpoint. That would be quite remarkable if | |
186 | * it could do that. Here's the flow that is required: | |
187 | * | |
188 | * CPU-0 CPU-1 | |
189 | * | |
190 | * atomic_inc(mfc); | |
191 | * write int3s | |
192 | * <trap-int3> // implicit (r)mb | |
193 | * if (atomic_read(mfc)) | |
194 | * call ftrace_int3_handler() | |
195 | * | |
196 | * Then when we are finished: | |
197 | * | |
198 | * atomic_dec(mfc); | |
199 | * | |
200 | * If we hit a breakpoint that was not set by ftrace, it does not | |
201 | * matter if ftrace_int3_handler() is called or not. It will | |
202 | * simply be ignored. But it is crucial that a ftrace nop/caller | |
203 | * breakpoint is handled. No other user should ever place a | |
204 | * breakpoint on an ftrace nop/caller location. It must only | |
205 | * be done by this code. | |
206 | */ | |
207 | atomic_t modifying_ftrace_code __read_mostly; | |
08d636b6 | 208 | |
8a4d0a68 SR |
209 | static int |
210 | ftrace_modify_code(unsigned long ip, unsigned const char *old_code, | |
211 | unsigned const char *new_code); | |
212 | ||
08f6fba5 SR |
213 | /* |
214 | * Should never be called: | |
215 | * As it is only called by __ftrace_replace_code() which is called by | |
216 | * ftrace_replace_code() that x86 overrides, and by ftrace_update_code() | |
217 | * which is called to turn mcount into nops or nops into function calls | |
218 | * but not to convert a function from not using regs to one that uses | |
219 | * regs, which ftrace_modify_call() is for. | |
220 | */ | |
221 | int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr, | |
222 | unsigned long addr) | |
223 | { | |
224 | WARN_ON(1); | |
b05086c7 | 225 | ftrace_expected = NULL; |
08f6fba5 SR |
226 | return -EINVAL; |
227 | } | |
08f6fba5 | 228 | |
87fbb2ac SRRH |
229 | static unsigned long ftrace_update_func; |
230 | ||
231 | static int update_ftrace_func(unsigned long ip, void *new) | |
8a4d0a68 | 232 | { |
87fbb2ac | 233 | unsigned char old[MCOUNT_INSN_SIZE]; |
8a4d0a68 SR |
234 | int ret; |
235 | ||
87fbb2ac SRRH |
236 | memcpy(old, (void *)ip, MCOUNT_INSN_SIZE); |
237 | ||
238 | ftrace_update_func = ip; | |
239 | /* Make sure the breakpoints see the ftrace_update_func update */ | |
240 | smp_wmb(); | |
8a4d0a68 SR |
241 | |
242 | /* See comment above by declaration of modifying_ftrace_code */ | |
243 | atomic_inc(&modifying_ftrace_code); | |
244 | ||
245 | ret = ftrace_modify_code(ip, old, new); | |
246 | ||
87fbb2ac SRRH |
247 | atomic_dec(&modifying_ftrace_code); |
248 | ||
249 | return ret; | |
250 | } | |
251 | ||
252 | int ftrace_update_ftrace_func(ftrace_func_t func) | |
253 | { | |
254 | unsigned long ip = (unsigned long)(&ftrace_call); | |
255 | unsigned char *new; | |
256 | int ret; | |
257 | ||
258 | new = ftrace_call_replace(ip, (unsigned long)func); | |
259 | ret = update_ftrace_func(ip, new); | |
260 | ||
08f6fba5 SR |
261 | /* Also update the regs callback function */ |
262 | if (!ret) { | |
263 | ip = (unsigned long)(&ftrace_regs_call); | |
08f6fba5 | 264 | new = ftrace_call_replace(ip, (unsigned long)func); |
87fbb2ac | 265 | ret = update_ftrace_func(ip, new); |
08f6fba5 | 266 | } |
08f6fba5 | 267 | |
8a4d0a68 SR |
268 | return ret; |
269 | } | |
270 | ||
ab4ead02 KH |
271 | static int is_ftrace_caller(unsigned long ip) |
272 | { | |
87fbb2ac | 273 | if (ip == ftrace_update_func) |
ab4ead02 KH |
274 | return 1; |
275 | ||
276 | return 0; | |
277 | } | |
278 | ||
08d636b6 SR |
279 | /* |
280 | * A breakpoint was added to the code address we are about to | |
281 | * modify, and this is the handle that will just skip over it. | |
282 | * We are either changing a nop into a trace call, or a trace | |
283 | * call to a nop. While the change is taking place, we treat | |
284 | * it just like it was a nop. | |
285 | */ | |
286 | int ftrace_int3_handler(struct pt_regs *regs) | |
287 | { | |
ab4ead02 KH |
288 | unsigned long ip; |
289 | ||
08d636b6 SR |
290 | if (WARN_ON_ONCE(!regs)) |
291 | return 0; | |
292 | ||
ab4ead02 KH |
293 | ip = regs->ip - 1; |
294 | if (!ftrace_location(ip) && !is_ftrace_caller(ip)) | |
08d636b6 SR |
295 | return 0; |
296 | ||
297 | regs->ip += MCOUNT_INSN_SIZE - 1; | |
298 | ||
299 | return 1; | |
300 | } | |
301 | ||
302 | static int ftrace_write(unsigned long ip, const char *val, int size) | |
303 | { | |
964f7b6b | 304 | ip = text_ip_addr(ip); |
08d636b6 | 305 | |
92550405 SRRH |
306 | if (probe_kernel_write((void *)ip, val, size)) |
307 | return -EPERM; | |
308 | ||
309 | return 0; | |
08d636b6 SR |
310 | } |
311 | ||
312 | static int add_break(unsigned long ip, const char *old) | |
313 | { | |
314 | unsigned char replaced[MCOUNT_INSN_SIZE]; | |
315 | unsigned char brk = BREAKPOINT_INSTRUCTION; | |
316 | ||
317 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) | |
318 | return -EFAULT; | |
319 | ||
b05086c7 SRRH |
320 | ftrace_expected = old; |
321 | ||
08d636b6 SR |
322 | /* Make sure it is what we expect it to be */ |
323 | if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0) | |
324 | return -EINVAL; | |
325 | ||
92550405 | 326 | return ftrace_write(ip, &brk, 1); |
08d636b6 SR |
327 | } |
328 | ||
329 | static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr) | |
330 | { | |
331 | unsigned const char *old; | |
332 | unsigned long ip = rec->ip; | |
333 | ||
334 | old = ftrace_call_replace(ip, addr); | |
335 | ||
336 | return add_break(rec->ip, old); | |
337 | } | |
338 | ||
339 | ||
340 | static int add_brk_on_nop(struct dyn_ftrace *rec) | |
341 | { | |
342 | unsigned const char *old; | |
343 | ||
344 | old = ftrace_nop_replace(); | |
345 | ||
346 | return add_break(rec->ip, old); | |
347 | } | |
348 | ||
349 | static int add_breakpoints(struct dyn_ftrace *rec, int enable) | |
350 | { | |
351 | unsigned long ftrace_addr; | |
352 | int ret; | |
353 | ||
7413af1f | 354 | ftrace_addr = ftrace_get_addr_curr(rec); |
08d636b6 | 355 | |
94792ea0 | 356 | ret = ftrace_test_record(rec, enable); |
08d636b6 SR |
357 | |
358 | switch (ret) { | |
359 | case FTRACE_UPDATE_IGNORE: | |
360 | return 0; | |
361 | ||
362 | case FTRACE_UPDATE_MAKE_CALL: | |
363 | /* converting nop to call */ | |
364 | return add_brk_on_nop(rec); | |
365 | ||
08f6fba5 | 366 | case FTRACE_UPDATE_MODIFY_CALL: |
08d636b6 SR |
367 | case FTRACE_UPDATE_MAKE_NOP: |
368 | /* converting a call to a nop */ | |
369 | return add_brk_on_call(rec, ftrace_addr); | |
370 | } | |
371 | return 0; | |
372 | } | |
373 | ||
374 | /* | |
375 | * On error, we need to remove breakpoints. This needs to | |
376 | * be done caefully. If the address does not currently have a | |
377 | * breakpoint, we know we are done. Otherwise, we look at the | |
378 | * remaining 4 bytes of the instruction. If it matches a nop | |
379 | * we replace the breakpoint with the nop. Otherwise we replace | |
380 | * it with the call instruction. | |
381 | */ | |
382 | static int remove_breakpoint(struct dyn_ftrace *rec) | |
383 | { | |
384 | unsigned char ins[MCOUNT_INSN_SIZE]; | |
385 | unsigned char brk = BREAKPOINT_INSTRUCTION; | |
386 | const unsigned char *nop; | |
387 | unsigned long ftrace_addr; | |
388 | unsigned long ip = rec->ip; | |
389 | ||
390 | /* If we fail the read, just give up */ | |
391 | if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE)) | |
392 | return -EFAULT; | |
393 | ||
394 | /* If this does not have a breakpoint, we are done */ | |
395 | if (ins[0] != brk) | |
7f11f5ec | 396 | return 0; |
08d636b6 SR |
397 | |
398 | nop = ftrace_nop_replace(); | |
399 | ||
400 | /* | |
401 | * If the last 4 bytes of the instruction do not match | |
402 | * a nop, then we assume that this is a call to ftrace_addr. | |
403 | */ | |
404 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) { | |
405 | /* | |
406 | * For extra paranoidism, we check if the breakpoint is on | |
407 | * a call that would actually jump to the ftrace_addr. | |
408 | * If not, don't touch the breakpoint, we make just create | |
409 | * a disaster. | |
410 | */ | |
7413af1f | 411 | ftrace_addr = ftrace_get_addr_new(rec); |
08f6fba5 SR |
412 | nop = ftrace_call_replace(ip, ftrace_addr); |
413 | ||
414 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0) | |
415 | goto update; | |
416 | ||
417 | /* Check both ftrace_addr and ftrace_old_addr */ | |
7413af1f | 418 | ftrace_addr = ftrace_get_addr_curr(rec); |
08d636b6 SR |
419 | nop = ftrace_call_replace(ip, ftrace_addr); |
420 | ||
b05086c7 SRRH |
421 | ftrace_expected = nop; |
422 | ||
08d636b6 SR |
423 | if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) |
424 | return -EINVAL; | |
425 | } | |
426 | ||
08f6fba5 | 427 | update: |
c932c6b7 | 428 | return ftrace_write(ip, nop, 1); |
08d636b6 SR |
429 | } |
430 | ||
431 | static int add_update_code(unsigned long ip, unsigned const char *new) | |
432 | { | |
433 | /* skip breakpoint */ | |
434 | ip++; | |
435 | new++; | |
92550405 | 436 | return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1); |
08d636b6 SR |
437 | } |
438 | ||
439 | static int add_update_call(struct dyn_ftrace *rec, unsigned long addr) | |
440 | { | |
441 | unsigned long ip = rec->ip; | |
442 | unsigned const char *new; | |
443 | ||
444 | new = ftrace_call_replace(ip, addr); | |
445 | return add_update_code(ip, new); | |
446 | } | |
447 | ||
448 | static int add_update_nop(struct dyn_ftrace *rec) | |
449 | { | |
450 | unsigned long ip = rec->ip; | |
451 | unsigned const char *new; | |
452 | ||
453 | new = ftrace_nop_replace(); | |
454 | return add_update_code(ip, new); | |
455 | } | |
456 | ||
457 | static int add_update(struct dyn_ftrace *rec, int enable) | |
458 | { | |
459 | unsigned long ftrace_addr; | |
460 | int ret; | |
461 | ||
462 | ret = ftrace_test_record(rec, enable); | |
463 | ||
7413af1f | 464 | ftrace_addr = ftrace_get_addr_new(rec); |
08d636b6 SR |
465 | |
466 | switch (ret) { | |
467 | case FTRACE_UPDATE_IGNORE: | |
468 | return 0; | |
469 | ||
08f6fba5 | 470 | case FTRACE_UPDATE_MODIFY_CALL: |
08d636b6 SR |
471 | case FTRACE_UPDATE_MAKE_CALL: |
472 | /* converting nop to call */ | |
473 | return add_update_call(rec, ftrace_addr); | |
474 | ||
475 | case FTRACE_UPDATE_MAKE_NOP: | |
476 | /* converting a call to a nop */ | |
477 | return add_update_nop(rec); | |
478 | } | |
479 | ||
480 | return 0; | |
481 | } | |
482 | ||
483 | static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr) | |
484 | { | |
485 | unsigned long ip = rec->ip; | |
486 | unsigned const char *new; | |
487 | ||
488 | new = ftrace_call_replace(ip, addr); | |
489 | ||
92550405 | 490 | return ftrace_write(ip, new, 1); |
08d636b6 SR |
491 | } |
492 | ||
493 | static int finish_update_nop(struct dyn_ftrace *rec) | |
494 | { | |
495 | unsigned long ip = rec->ip; | |
496 | unsigned const char *new; | |
497 | ||
498 | new = ftrace_nop_replace(); | |
499 | ||
92550405 | 500 | return ftrace_write(ip, new, 1); |
08d636b6 SR |
501 | } |
502 | ||
503 | static int finish_update(struct dyn_ftrace *rec, int enable) | |
504 | { | |
505 | unsigned long ftrace_addr; | |
506 | int ret; | |
507 | ||
508 | ret = ftrace_update_record(rec, enable); | |
509 | ||
7413af1f | 510 | ftrace_addr = ftrace_get_addr_new(rec); |
08d636b6 SR |
511 | |
512 | switch (ret) { | |
513 | case FTRACE_UPDATE_IGNORE: | |
514 | return 0; | |
515 | ||
08f6fba5 | 516 | case FTRACE_UPDATE_MODIFY_CALL: |
08d636b6 SR |
517 | case FTRACE_UPDATE_MAKE_CALL: |
518 | /* converting nop to call */ | |
519 | return finish_update_call(rec, ftrace_addr); | |
520 | ||
521 | case FTRACE_UPDATE_MAKE_NOP: | |
522 | /* converting a call to a nop */ | |
523 | return finish_update_nop(rec); | |
524 | } | |
525 | ||
526 | return 0; | |
527 | } | |
528 | ||
529 | static void do_sync_core(void *data) | |
530 | { | |
531 | sync_core(); | |
532 | } | |
533 | ||
534 | static void run_sync(void) | |
535 | { | |
2b87965a SRV |
536 | int enable_irqs; |
537 | ||
538 | /* No need to sync if there's only one CPU */ | |
539 | if (num_online_cpus() == 1) | |
540 | return; | |
541 | ||
542 | enable_irqs = irqs_disabled(); | |
08d636b6 | 543 | |
8a1115ff | 544 | /* We may be called with interrupts disabled (on bootup). */ |
08d636b6 SR |
545 | if (enable_irqs) |
546 | local_irq_enable(); | |
547 | on_each_cpu(do_sync_core, NULL, 1); | |
548 | if (enable_irqs) | |
549 | local_irq_disable(); | |
550 | } | |
551 | ||
e4f5d544 | 552 | void ftrace_replace_code(int enable) |
08d636b6 SR |
553 | { |
554 | struct ftrace_rec_iter *iter; | |
555 | struct dyn_ftrace *rec; | |
556 | const char *report = "adding breakpoints"; | |
557 | int count = 0; | |
558 | int ret; | |
559 | ||
560 | for_ftrace_rec_iter(iter) { | |
561 | rec = ftrace_rec_iter_record(iter); | |
562 | ||
563 | ret = add_breakpoints(rec, enable); | |
564 | if (ret) | |
565 | goto remove_breakpoints; | |
566 | count++; | |
567 | } | |
568 | ||
569 | run_sync(); | |
570 | ||
571 | report = "updating code"; | |
883a1e86 | 572 | count = 0; |
08d636b6 SR |
573 | |
574 | for_ftrace_rec_iter(iter) { | |
575 | rec = ftrace_rec_iter_record(iter); | |
576 | ||
577 | ret = add_update(rec, enable); | |
578 | if (ret) | |
579 | goto remove_breakpoints; | |
883a1e86 | 580 | count++; |
08d636b6 SR |
581 | } |
582 | ||
583 | run_sync(); | |
584 | ||
585 | report = "removing breakpoints"; | |
883a1e86 | 586 | count = 0; |
08d636b6 SR |
587 | |
588 | for_ftrace_rec_iter(iter) { | |
589 | rec = ftrace_rec_iter_record(iter); | |
590 | ||
591 | ret = finish_update(rec, enable); | |
592 | if (ret) | |
593 | goto remove_breakpoints; | |
883a1e86 | 594 | count++; |
08d636b6 SR |
595 | } |
596 | ||
597 | run_sync(); | |
598 | ||
599 | return; | |
600 | ||
601 | remove_breakpoints: | |
74bb8c45 | 602 | pr_warn("Failed on %s (%d):\n", report, count); |
4fd3279b | 603 | ftrace_bug(ret, rec); |
08d636b6 SR |
604 | for_ftrace_rec_iter(iter) { |
605 | rec = ftrace_rec_iter_record(iter); | |
7f11f5ec PM |
606 | /* |
607 | * Breakpoints are handled only when this function is in | |
608 | * progress. The system could not work with them. | |
609 | */ | |
610 | if (remove_breakpoint(rec)) | |
611 | BUG(); | |
08d636b6 | 612 | } |
c932c6b7 | 613 | run_sync(); |
08d636b6 SR |
614 | } |
615 | ||
8a4d0a68 SR |
616 | static int |
617 | ftrace_modify_code(unsigned long ip, unsigned const char *old_code, | |
618 | unsigned const char *new_code) | |
619 | { | |
620 | int ret; | |
621 | ||
622 | ret = add_break(ip, old_code); | |
623 | if (ret) | |
624 | goto out; | |
625 | ||
626 | run_sync(); | |
627 | ||
628 | ret = add_update_code(ip, new_code); | |
629 | if (ret) | |
630 | goto fail_update; | |
631 | ||
632 | run_sync(); | |
633 | ||
634 | ret = ftrace_write(ip, new_code, 1); | |
7f11f5ec PM |
635 | /* |
636 | * The breakpoint is handled only when this function is in progress. | |
637 | * The system could not work if we could not remove it. | |
638 | */ | |
639 | BUG_ON(ret); | |
8a4d0a68 | 640 | out: |
12729f14 | 641 | run_sync(); |
8a4d0a68 SR |
642 | return ret; |
643 | ||
644 | fail_update: | |
7f11f5ec PM |
645 | /* Also here the system could not work with the breakpoint */ |
646 | if (ftrace_write(ip, old_code, 1)) | |
647 | BUG(); | |
8a4d0a68 SR |
648 | goto out; |
649 | } | |
650 | ||
08d636b6 SR |
651 | void arch_ftrace_update_code(int command) |
652 | { | |
a192cd04 SR |
653 | /* See comment above by declaration of modifying_ftrace_code */ |
654 | atomic_inc(&modifying_ftrace_code); | |
08d636b6 | 655 | |
e4f5d544 | 656 | ftrace_modify_all_code(command); |
08d636b6 | 657 | |
a192cd04 | 658 | atomic_dec(&modifying_ftrace_code); |
08d636b6 SR |
659 | } |
660 | ||
3a36cb11 | 661 | int __init ftrace_dyn_arch_init(void) |
3d083395 | 662 | { |
3d083395 SR |
663 | return 0; |
664 | } | |
5a45cfe1 | 665 | |
f3bea491 | 666 | #if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER) |
87fbb2ac | 667 | static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr) |
5a45cfe1 | 668 | { |
87fbb2ac | 669 | static union ftrace_code_union calc; |
5a45cfe1 | 670 | |
87fbb2ac SRRH |
671 | /* Jmp not a call (ignore the .e8) */ |
672 | calc.e8 = 0xe9; | |
673 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); | |
5a45cfe1 | 674 | |
87fbb2ac SRRH |
675 | /* |
676 | * ftrace external locks synchronize the access to the static variable. | |
677 | */ | |
678 | return calc.code; | |
679 | } | |
f3bea491 SRRH |
680 | #endif |
681 | ||
682 | /* Currently only x86_64 supports dynamic trampolines */ | |
683 | #ifdef CONFIG_X86_64 | |
684 | ||
685 | #ifdef CONFIG_MODULES | |
686 | #include <linux/moduleloader.h> | |
687 | /* Module allocation simplifies allocating memory for code */ | |
688 | static inline void *alloc_tramp(unsigned long size) | |
689 | { | |
690 | return module_alloc(size); | |
691 | } | |
6ee98ffe | 692 | static inline void tramp_free(void *tramp, int size) |
f3bea491 | 693 | { |
6ee98ffe TG |
694 | int npages = PAGE_ALIGN(size) >> PAGE_SHIFT; |
695 | ||
696 | set_memory_nx((unsigned long)tramp, npages); | |
697 | set_memory_rw((unsigned long)tramp, npages); | |
be1f221c | 698 | module_memfree(tramp); |
f3bea491 SRRH |
699 | } |
700 | #else | |
701 | /* Trampolines can only be created if modules are supported */ | |
702 | static inline void *alloc_tramp(unsigned long size) | |
703 | { | |
704 | return NULL; | |
705 | } | |
6ee98ffe | 706 | static inline void tramp_free(void *tramp, int size) { } |
f3bea491 SRRH |
707 | #endif |
708 | ||
709 | /* Defined as markers to the end of the ftrace default trampolines */ | |
f3bea491 | 710 | extern void ftrace_regs_caller_end(void); |
f1b92bb6 | 711 | extern void ftrace_epilogue(void); |
f3bea491 SRRH |
712 | extern void ftrace_caller_op_ptr(void); |
713 | extern void ftrace_regs_caller_op_ptr(void); | |
714 | ||
715 | /* movq function_trace_op(%rip), %rdx */ | |
716 | /* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */ | |
717 | #define OP_REF_SIZE 7 | |
718 | ||
719 | /* | |
720 | * The ftrace_ops is passed to the function callback. Since the | |
721 | * trampoline only services a single ftrace_ops, we can pass in | |
722 | * that ops directly. | |
723 | * | |
724 | * The ftrace_op_code_union is used to create a pointer to the | |
725 | * ftrace_ops that will be passed to the callback function. | |
726 | */ | |
727 | union ftrace_op_code_union { | |
728 | char code[OP_REF_SIZE]; | |
729 | struct { | |
730 | char op[3]; | |
731 | int offset; | |
732 | } __attribute__((packed)); | |
733 | }; | |
734 | ||
aec0be2d SRRH |
735 | static unsigned long |
736 | create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size) | |
f3bea491 SRRH |
737 | { |
738 | unsigned const char *jmp; | |
739 | unsigned long start_offset; | |
740 | unsigned long end_offset; | |
741 | unsigned long op_offset; | |
742 | unsigned long offset; | |
743 | unsigned long size; | |
744 | unsigned long ip; | |
745 | unsigned long *ptr; | |
746 | void *trampoline; | |
747 | /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */ | |
748 | unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 }; | |
749 | union ftrace_op_code_union op_ptr; | |
750 | int ret; | |
751 | ||
752 | if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) { | |
753 | start_offset = (unsigned long)ftrace_regs_caller; | |
754 | end_offset = (unsigned long)ftrace_regs_caller_end; | |
755 | op_offset = (unsigned long)ftrace_regs_caller_op_ptr; | |
756 | } else { | |
757 | start_offset = (unsigned long)ftrace_caller; | |
f1b92bb6 | 758 | end_offset = (unsigned long)ftrace_epilogue; |
f3bea491 SRRH |
759 | op_offset = (unsigned long)ftrace_caller_op_ptr; |
760 | } | |
761 | ||
762 | size = end_offset - start_offset; | |
763 | ||
764 | /* | |
765 | * Allocate enough size to store the ftrace_caller code, | |
f1b92bb6 | 766 | * the jmp to ftrace_epilogue, as well as the address of |
f3bea491 SRRH |
767 | * the ftrace_ops this trampoline is used for. |
768 | */ | |
769 | trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *)); | |
770 | if (!trampoline) | |
771 | return 0; | |
772 | ||
aec0be2d SRRH |
773 | *tramp_size = size + MCOUNT_INSN_SIZE + sizeof(void *); |
774 | ||
f3bea491 SRRH |
775 | /* Copy ftrace_caller onto the trampoline memory */ |
776 | ret = probe_kernel_read(trampoline, (void *)start_offset, size); | |
777 | if (WARN_ON(ret < 0)) { | |
6ee98ffe | 778 | tramp_free(trampoline, *tramp_size); |
f3bea491 SRRH |
779 | return 0; |
780 | } | |
781 | ||
782 | ip = (unsigned long)trampoline + size; | |
783 | ||
f1b92bb6 BP |
784 | /* The trampoline ends with a jmp to ftrace_epilogue */ |
785 | jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_epilogue); | |
f3bea491 SRRH |
786 | memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE); |
787 | ||
788 | /* | |
789 | * The address of the ftrace_ops that is used for this trampoline | |
790 | * is stored at the end of the trampoline. This will be used to | |
791 | * load the third parameter for the callback. Basically, that | |
792 | * location at the end of the trampoline takes the place of | |
793 | * the global function_trace_op variable. | |
794 | */ | |
795 | ||
796 | ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE); | |
797 | *ptr = (unsigned long)ops; | |
798 | ||
799 | op_offset -= start_offset; | |
800 | memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE); | |
801 | ||
802 | /* Are we pointing to the reference? */ | |
803 | if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) { | |
6ee98ffe | 804 | tramp_free(trampoline, *tramp_size); |
f3bea491 SRRH |
805 | return 0; |
806 | } | |
807 | ||
808 | /* Load the contents of ptr into the callback parameter */ | |
809 | offset = (unsigned long)ptr; | |
810 | offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE; | |
811 | ||
812 | op_ptr.offset = offset; | |
813 | ||
814 | /* put in the new offset to the ftrace_ops */ | |
815 | memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE); | |
816 | ||
817 | /* ALLOC_TRAMP flags lets us know we created it */ | |
818 | ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP; | |
819 | ||
820 | return (unsigned long)trampoline; | |
821 | } | |
822 | ||
15d5b02c SRRH |
823 | static unsigned long calc_trampoline_call_offset(bool save_regs) |
824 | { | |
825 | unsigned long start_offset; | |
826 | unsigned long call_offset; | |
827 | ||
828 | if (save_regs) { | |
829 | start_offset = (unsigned long)ftrace_regs_caller; | |
830 | call_offset = (unsigned long)ftrace_regs_call; | |
831 | } else { | |
832 | start_offset = (unsigned long)ftrace_caller; | |
833 | call_offset = (unsigned long)ftrace_call; | |
834 | } | |
835 | ||
836 | return call_offset - start_offset; | |
837 | } | |
838 | ||
f3bea491 SRRH |
839 | void arch_ftrace_update_trampoline(struct ftrace_ops *ops) |
840 | { | |
841 | ftrace_func_t func; | |
842 | unsigned char *new; | |
f3bea491 SRRH |
843 | unsigned long offset; |
844 | unsigned long ip; | |
aec0be2d | 845 | unsigned int size; |
6ee98ffe | 846 | int ret, npages; |
f3bea491 SRRH |
847 | |
848 | if (ops->trampoline) { | |
849 | /* | |
850 | * The ftrace_ops caller may set up its own trampoline. | |
851 | * In such a case, this code must not modify it. | |
852 | */ | |
853 | if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) | |
854 | return; | |
6ee98ffe TG |
855 | npages = PAGE_ALIGN(ops->trampoline_size) >> PAGE_SHIFT; |
856 | set_memory_rw(ops->trampoline, npages); | |
f3bea491 | 857 | } else { |
aec0be2d | 858 | ops->trampoline = create_trampoline(ops, &size); |
f3bea491 SRRH |
859 | if (!ops->trampoline) |
860 | return; | |
aec0be2d | 861 | ops->trampoline_size = size; |
6ee98ffe | 862 | npages = PAGE_ALIGN(size) >> PAGE_SHIFT; |
f3bea491 SRRH |
863 | } |
864 | ||
15d5b02c | 865 | offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); |
f3bea491 SRRH |
866 | ip = ops->trampoline + offset; |
867 | ||
868 | func = ftrace_ops_get_func(ops); | |
869 | ||
870 | /* Do a safe modify in case the trampoline is executing */ | |
871 | new = ftrace_call_replace(ip, (unsigned long)func); | |
872 | ret = update_ftrace_func(ip, new); | |
6ee98ffe | 873 | set_memory_ro(ops->trampoline, npages); |
f3bea491 SRRH |
874 | |
875 | /* The update should never fail */ | |
876 | WARN_ON(ret); | |
877 | } | |
15d5b02c SRRH |
878 | |
879 | /* Return the address of the function the trampoline calls */ | |
880 | static void *addr_from_call(void *ptr) | |
881 | { | |
882 | union ftrace_code_union calc; | |
883 | int ret; | |
884 | ||
885 | ret = probe_kernel_read(&calc, ptr, MCOUNT_INSN_SIZE); | |
886 | if (WARN_ON_ONCE(ret < 0)) | |
887 | return NULL; | |
888 | ||
889 | /* Make sure this is a call */ | |
890 | if (WARN_ON_ONCE(calc.e8 != 0xe8)) { | |
891 | pr_warn("Expected e8, got %x\n", calc.e8); | |
892 | return NULL; | |
893 | } | |
894 | ||
895 | return ptr + MCOUNT_INSN_SIZE + calc.offset; | |
896 | } | |
897 | ||
6a06bdbf | 898 | void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent, |
15d5b02c SRRH |
899 | unsigned long frame_pointer); |
900 | ||
901 | /* | |
902 | * If the ops->trampoline was not allocated, then it probably | |
903 | * has a static trampoline func, or is the ftrace caller itself. | |
904 | */ | |
905 | static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) | |
906 | { | |
907 | unsigned long offset; | |
908 | bool save_regs = rec->flags & FTRACE_FL_REGS_EN; | |
909 | void *ptr; | |
910 | ||
911 | if (ops && ops->trampoline) { | |
912 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER | |
913 | /* | |
914 | * We only know about function graph tracer setting as static | |
915 | * trampoline. | |
916 | */ | |
917 | if (ops->trampoline == FTRACE_GRAPH_ADDR) | |
918 | return (void *)prepare_ftrace_return; | |
919 | #endif | |
920 | return NULL; | |
921 | } | |
922 | ||
923 | offset = calc_trampoline_call_offset(save_regs); | |
924 | ||
925 | if (save_regs) | |
926 | ptr = (void *)FTRACE_REGS_ADDR + offset; | |
927 | else | |
928 | ptr = (void *)FTRACE_ADDR + offset; | |
929 | ||
930 | return addr_from_call(ptr); | |
931 | } | |
932 | ||
933 | void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec) | |
934 | { | |
935 | unsigned long offset; | |
936 | ||
937 | /* If we didn't allocate this trampoline, consider it static */ | |
938 | if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) | |
939 | return static_tramp_func(ops, rec); | |
940 | ||
941 | offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS); | |
942 | return addr_from_call((void *)ops->trampoline + offset); | |
943 | } | |
944 | ||
12cce594 SRRH |
945 | void arch_ftrace_trampoline_free(struct ftrace_ops *ops) |
946 | { | |
947 | if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) | |
948 | return; | |
949 | ||
6ee98ffe | 950 | tramp_free((void *)ops->trampoline, ops->trampoline_size); |
12cce594 SRRH |
951 | ops->trampoline = 0; |
952 | } | |
15d5b02c | 953 | |
f3bea491 SRRH |
954 | #endif /* CONFIG_X86_64 */ |
955 | #endif /* CONFIG_DYNAMIC_FTRACE */ | |
956 | ||
957 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER | |
958 | ||
959 | #ifdef CONFIG_DYNAMIC_FTRACE | |
960 | extern void ftrace_graph_call(void); | |
5a45cfe1 | 961 | |
87fbb2ac SRRH |
962 | static int ftrace_mod_jmp(unsigned long ip, void *func) |
963 | { | |
964 | unsigned char *new; | |
5a45cfe1 | 965 | |
87fbb2ac | 966 | new = ftrace_jmp_replace(ip, (unsigned long)func); |
5a45cfe1 | 967 | |
87fbb2ac | 968 | return update_ftrace_func(ip, new); |
5a45cfe1 SR |
969 | } |
970 | ||
971 | int ftrace_enable_ftrace_graph_caller(void) | |
972 | { | |
973 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
5a45cfe1 | 974 | |
87fbb2ac | 975 | return ftrace_mod_jmp(ip, &ftrace_graph_caller); |
5a45cfe1 SR |
976 | } |
977 | ||
978 | int ftrace_disable_ftrace_graph_caller(void) | |
979 | { | |
980 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
5a45cfe1 | 981 | |
87fbb2ac | 982 | return ftrace_mod_jmp(ip, &ftrace_stub); |
5a45cfe1 SR |
983 | } |
984 | ||
e7d3737e FW |
985 | #endif /* !CONFIG_DYNAMIC_FTRACE */ |
986 | ||
e7d3737e FW |
987 | /* |
988 | * Hook the return address and push it in the stack of return addrs | |
989 | * in current thread info. | |
990 | */ | |
6a06bdbf | 991 | void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent, |
71e308a2 | 992 | unsigned long frame_pointer) |
e7d3737e FW |
993 | { |
994 | unsigned long old; | |
e7d3737e | 995 | int faulted; |
287b6e68 | 996 | struct ftrace_graph_ent trace; |
e7d3737e FW |
997 | unsigned long return_hooker = (unsigned long) |
998 | &return_to_handler; | |
999 | ||
34a477e5 JP |
1000 | /* |
1001 | * When resuming from suspend-to-ram, this function can be indirectly | |
1002 | * called from early CPU startup code while the CPU is in real mode, | |
1003 | * which would fail miserably. Make sure the stack pointer is a | |
1004 | * virtual address. | |
1005 | * | |
1006 | * This check isn't as accurate as virt_addr_valid(), but it should be | |
1007 | * good enough for this purpose, and it's fast. | |
1008 | */ | |
1009 | if (unlikely((long)__builtin_frame_address(0) >= 0)) | |
1010 | return; | |
1011 | ||
84b2bc7f SRRH |
1012 | if (unlikely(ftrace_graph_is_dead())) |
1013 | return; | |
1014 | ||
380c4b14 | 1015 | if (unlikely(atomic_read(¤t->tracing_graph_pause))) |
e7d3737e FW |
1016 | return; |
1017 | ||
1018 | /* | |
1019 | * Protect against fault, even if it shouldn't | |
1020 | * happen. This tool is too much intrusive to | |
1021 | * ignore such a protection. | |
1022 | */ | |
1023 | asm volatile( | |
96665788 SR |
1024 | "1: " _ASM_MOV " (%[parent]), %[old]\n" |
1025 | "2: " _ASM_MOV " %[return_hooker], (%[parent])\n" | |
e7d3737e | 1026 | " movl $0, %[faulted]\n" |
e3944bfa | 1027 | "3:\n" |
e7d3737e FW |
1028 | |
1029 | ".section .fixup, \"ax\"\n" | |
e3944bfa SR |
1030 | "4: movl $1, %[faulted]\n" |
1031 | " jmp 3b\n" | |
e7d3737e FW |
1032 | ".previous\n" |
1033 | ||
e3944bfa SR |
1034 | _ASM_EXTABLE(1b, 4b) |
1035 | _ASM_EXTABLE(2b, 4b) | |
e7d3737e | 1036 | |
aa512a27 | 1037 | : [old] "=&r" (old), [faulted] "=r" (faulted) |
96665788 | 1038 | : [parent] "r" (parent), [return_hooker] "r" (return_hooker) |
e7d3737e FW |
1039 | : "memory" |
1040 | ); | |
1041 | ||
14a866c5 SR |
1042 | if (unlikely(faulted)) { |
1043 | ftrace_graph_stop(); | |
1044 | WARN_ON(1); | |
e7d3737e FW |
1045 | return; |
1046 | } | |
1047 | ||
287b6e68 | 1048 | trace.func = self_addr; |
722b3c74 | 1049 | trace.depth = current->curr_ret_stack + 1; |
287b6e68 | 1050 | |
e49dc19c SR |
1051 | /* Only trace if the calling function expects to */ |
1052 | if (!ftrace_graph_entry(&trace)) { | |
e49dc19c | 1053 | *parent = old; |
722b3c74 SR |
1054 | return; |
1055 | } | |
1056 | ||
1057 | if (ftrace_push_return_trace(old, self_addr, &trace.depth, | |
471bd10f | 1058 | frame_pointer, parent) == -EBUSY) { |
722b3c74 SR |
1059 | *parent = old; |
1060 | return; | |
e49dc19c | 1061 | } |
e7d3737e | 1062 | } |
fb52607a | 1063 | #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ |