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 | ||
395a59d0 | 21 | #include <asm/ftrace.h> |
caf4b323 | 22 | #include <linux/ftrace.h> |
732f3ca7 | 23 | #include <asm/nops.h> |
caf4b323 | 24 | #include <asm/nmi.h> |
3d083395 | 25 | |
3d083395 | 26 | |
caf4b323 | 27 | #ifdef CONFIG_DYNAMIC_FTRACE |
3d083395 | 28 | |
3d083395 | 29 | union ftrace_code_union { |
395a59d0 | 30 | char code[MCOUNT_INSN_SIZE]; |
3d083395 SR |
31 | struct { |
32 | char e8; | |
33 | int offset; | |
34 | } __attribute__((packed)); | |
35 | }; | |
36 | ||
15adc048 | 37 | static int ftrace_calc_offset(long ip, long addr) |
3c1720f0 SR |
38 | { |
39 | return (int)(addr - ip); | |
40 | } | |
3d083395 | 41 | |
31e88909 | 42 | static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) |
3c1720f0 SR |
43 | { |
44 | static union ftrace_code_union calc; | |
3d083395 | 45 | |
3c1720f0 | 46 | calc.e8 = 0xe8; |
395a59d0 | 47 | calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr); |
3c1720f0 SR |
48 | |
49 | /* | |
50 | * No locking needed, this must be called via kstop_machine | |
51 | * which in essence is like running on a uniprocessor machine. | |
52 | */ | |
53 | return calc.code; | |
3d083395 SR |
54 | } |
55 | ||
17666f02 SR |
56 | /* |
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. | |
64 | * | |
65 | * Two buffers are added: An IP buffer and a "code" buffer. | |
66 | * | |
a26a2a27 | 67 | * 1) Put the instruction pointer into the IP buffer |
17666f02 SR |
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. | |
71 | * 4) Write the code | |
72 | * 5) clear the flag. | |
73 | * 6) Wait for any running NMIs to finish. | |
74 | * | |
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. | |
78 | * | |
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. | |
83 | */ | |
84 | ||
4e6ea144 | 85 | static atomic_t nmi_running = ATOMIC_INIT(0); |
a26a2a27 SR |
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 */ | |
17666f02 | 90 | |
a26a2a27 SR |
91 | static unsigned nmi_wait_count; |
92 | static atomic_t nmi_update_count = ATOMIC_INIT(0); | |
b807c3d0 SR |
93 | |
94 | int ftrace_arch_read_dyn_info(char *buf, int size) | |
95 | { | |
96 | int r; | |
97 | ||
98 | r = snprintf(buf, size, "%u %u", | |
99 | nmi_wait_count, | |
100 | atomic_read(&nmi_update_count)); | |
101 | return r; | |
102 | } | |
103 | ||
17666f02 SR |
104 | static void ftrace_mod_code(void) |
105 | { | |
106 | /* | |
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. | |
111 | */ | |
112 | mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode, | |
113 | MCOUNT_INSN_SIZE); | |
17666f02 SR |
114 | } |
115 | ||
78d904b4 | 116 | void arch_ftrace_nmi_enter(void) |
17666f02 | 117 | { |
4e6ea144 SR |
118 | atomic_inc(&nmi_running); |
119 | /* Must have nmi_running seen before reading write flag */ | |
17666f02 | 120 | smp_mb(); |
b807c3d0 | 121 | if (mod_code_write) { |
17666f02 | 122 | ftrace_mod_code(); |
b807c3d0 SR |
123 | atomic_inc(&nmi_update_count); |
124 | } | |
17666f02 SR |
125 | } |
126 | ||
78d904b4 | 127 | void arch_ftrace_nmi_exit(void) |
17666f02 | 128 | { |
4e6ea144 | 129 | /* Finish all executions before clearing nmi_running */ |
17666f02 | 130 | smp_wmb(); |
4e6ea144 | 131 | atomic_dec(&nmi_running); |
17666f02 SR |
132 | } |
133 | ||
134 | static void wait_for_nmi(void) | |
135 | { | |
4e6ea144 | 136 | if (!atomic_read(&nmi_running)) |
89025282 | 137 | return; |
b807c3d0 | 138 | |
89025282 | 139 | do { |
17666f02 | 140 | cpu_relax(); |
4e6ea144 | 141 | } while (atomic_read(&nmi_running)); |
b807c3d0 | 142 | |
89025282 | 143 | nmi_wait_count++; |
17666f02 SR |
144 | } |
145 | ||
146 | static int | |
147 | do_ftrace_mod_code(unsigned long ip, void *new_code) | |
148 | { | |
149 | mod_code_ip = (void *)ip; | |
150 | mod_code_newcode = new_code; | |
151 | ||
152 | /* The buffers need to be visible before we let NMIs write them */ | |
153 | smp_wmb(); | |
154 | ||
155 | mod_code_write = 1; | |
156 | ||
157 | /* Make sure write bit is visible before we wait on NMIs */ | |
158 | smp_mb(); | |
159 | ||
160 | wait_for_nmi(); | |
161 | ||
162 | /* Make sure all running NMIs have finished before we write the code */ | |
163 | smp_mb(); | |
164 | ||
165 | ftrace_mod_code(); | |
166 | ||
167 | /* Make sure the write happens before clearing the bit */ | |
168 | smp_wmb(); | |
169 | ||
170 | mod_code_write = 0; | |
171 | ||
172 | /* make sure NMIs see the cleared bit */ | |
173 | smp_mb(); | |
174 | ||
175 | wait_for_nmi(); | |
176 | ||
177 | return mod_code_status; | |
178 | } | |
179 | ||
180 | ||
caf4b323 FW |
181 | |
182 | ||
183 | static unsigned char ftrace_nop[MCOUNT_INSN_SIZE]; | |
184 | ||
31e88909 | 185 | static unsigned char *ftrace_nop_replace(void) |
caf4b323 FW |
186 | { |
187 | return ftrace_nop; | |
188 | } | |
189 | ||
31e88909 | 190 | static int |
3d083395 SR |
191 | ftrace_modify_code(unsigned long ip, unsigned char *old_code, |
192 | unsigned char *new_code) | |
193 | { | |
6f93fc07 | 194 | unsigned char replaced[MCOUNT_INSN_SIZE]; |
3d083395 SR |
195 | |
196 | /* | |
197 | * Note: Due to modules and __init, code can | |
198 | * disappear and change, we need to protect against faulting | |
76aefee5 | 199 | * as well as code changing. We do this by using the |
ab9a0918 | 200 | * probe_kernel_* functions. |
3d083395 SR |
201 | * |
202 | * No real locking needed, this code is run through | |
6f93fc07 | 203 | * kstop_machine, or before SMP starts. |
3d083395 | 204 | */ |
76aefee5 SR |
205 | |
206 | /* read the text we want to modify */ | |
ab9a0918 | 207 | if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) |
593eb8a2 | 208 | return -EFAULT; |
6f93fc07 | 209 | |
76aefee5 | 210 | /* Make sure it is what we expect it to be */ |
6f93fc07 | 211 | if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) |
593eb8a2 | 212 | return -EINVAL; |
3d083395 | 213 | |
76aefee5 | 214 | /* replace the text with the new text */ |
17666f02 | 215 | if (do_ftrace_mod_code(ip, new_code)) |
593eb8a2 | 216 | return -EPERM; |
6f93fc07 SR |
217 | |
218 | sync_core(); | |
3d083395 | 219 | |
6f93fc07 | 220 | return 0; |
3d083395 SR |
221 | } |
222 | ||
31e88909 SR |
223 | int ftrace_make_nop(struct module *mod, |
224 | struct dyn_ftrace *rec, unsigned long addr) | |
225 | { | |
226 | unsigned char *new, *old; | |
227 | unsigned long ip = rec->ip; | |
228 | ||
229 | old = ftrace_call_replace(ip, addr); | |
230 | new = ftrace_nop_replace(); | |
231 | ||
232 | return ftrace_modify_code(rec->ip, old, new); | |
233 | } | |
234 | ||
235 | int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) | |
236 | { | |
237 | unsigned char *new, *old; | |
238 | unsigned long ip = rec->ip; | |
239 | ||
240 | old = ftrace_nop_replace(); | |
241 | new = ftrace_call_replace(ip, addr); | |
242 | ||
243 | return ftrace_modify_code(rec->ip, old, new); | |
244 | } | |
245 | ||
15adc048 | 246 | int ftrace_update_ftrace_func(ftrace_func_t func) |
d61f82d0 SR |
247 | { |
248 | unsigned long ip = (unsigned long)(&ftrace_call); | |
395a59d0 | 249 | unsigned char old[MCOUNT_INSN_SIZE], *new; |
d61f82d0 SR |
250 | int ret; |
251 | ||
395a59d0 | 252 | memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE); |
d61f82d0 SR |
253 | new = ftrace_call_replace(ip, (unsigned long)func); |
254 | ret = ftrace_modify_code(ip, old, new); | |
255 | ||
256 | return ret; | |
257 | } | |
258 | ||
d61f82d0 | 259 | int __init ftrace_dyn_arch_init(void *data) |
3d083395 | 260 | { |
732f3ca7 SR |
261 | extern const unsigned char ftrace_test_p6nop[]; |
262 | extern const unsigned char ftrace_test_nop5[]; | |
263 | extern const unsigned char ftrace_test_jmp[]; | |
264 | int faulted = 0; | |
d61f82d0 | 265 | |
732f3ca7 SR |
266 | /* |
267 | * There is no good nop for all x86 archs. | |
268 | * We will default to using the P6_NOP5, but first we | |
269 | * will test to make sure that the nop will actually | |
270 | * work on this CPU. If it faults, we will then | |
271 | * go to a lesser efficient 5 byte nop. If that fails | |
272 | * we then just use a jmp as our nop. This isn't the most | |
273 | * efficient nop, but we can not use a multi part nop | |
274 | * since we would then risk being preempted in the middle | |
275 | * of that nop, and if we enabled tracing then, it might | |
276 | * cause a system crash. | |
277 | * | |
278 | * TODO: check the cpuid to determine the best nop. | |
279 | */ | |
280 | asm volatile ( | |
732f3ca7 SR |
281 | "ftrace_test_jmp:" |
282 | "jmp ftrace_test_p6nop\n" | |
8b27386a AK |
283 | "nop\n" |
284 | "nop\n" | |
285 | "nop\n" /* 2 byte jmp + 3 bytes */ | |
732f3ca7 SR |
286 | "ftrace_test_p6nop:" |
287 | P6_NOP5 | |
288 | "jmp 1f\n" | |
289 | "ftrace_test_nop5:" | |
290 | ".byte 0x66,0x66,0x66,0x66,0x90\n" | |
732f3ca7 SR |
291 | "1:" |
292 | ".section .fixup, \"ax\"\n" | |
293 | "2: movl $1, %0\n" | |
294 | " jmp ftrace_test_nop5\n" | |
295 | "3: movl $2, %0\n" | |
296 | " jmp 1b\n" | |
297 | ".previous\n" | |
298 | _ASM_EXTABLE(ftrace_test_p6nop, 2b) | |
299 | _ASM_EXTABLE(ftrace_test_nop5, 3b) | |
300 | : "=r"(faulted) : "0" (faulted)); | |
301 | ||
302 | switch (faulted) { | |
303 | case 0: | |
304 | pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n"); | |
8115f3f0 | 305 | memcpy(ftrace_nop, ftrace_test_p6nop, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
306 | break; |
307 | case 1: | |
308 | pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n"); | |
8115f3f0 | 309 | memcpy(ftrace_nop, ftrace_test_nop5, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
310 | break; |
311 | case 2: | |
8b27386a | 312 | pr_info("ftrace: converting mcount calls to jmp . + 5\n"); |
8115f3f0 | 313 | memcpy(ftrace_nop, ftrace_test_jmp, MCOUNT_INSN_SIZE); |
732f3ca7 SR |
314 | break; |
315 | } | |
316 | ||
317 | /* The return code is retured via data */ | |
318 | *(unsigned long *)data = 0; | |
dfa60aba | 319 | |
3d083395 SR |
320 | return 0; |
321 | } | |
caf4b323 | 322 | #endif |
e7d3737e | 323 | |
fb52607a | 324 | #ifdef CONFIG_FUNCTION_GRAPH_TRACER |
e7d3737e | 325 | |
5a45cfe1 SR |
326 | #ifdef CONFIG_DYNAMIC_FTRACE |
327 | extern void ftrace_graph_call(void); | |
328 | ||
329 | static int ftrace_mod_jmp(unsigned long ip, | |
330 | int old_offset, int new_offset) | |
331 | { | |
332 | unsigned char code[MCOUNT_INSN_SIZE]; | |
333 | ||
334 | if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE)) | |
335 | return -EFAULT; | |
336 | ||
337 | if (code[0] != 0xe9 || old_offset != *(int *)(&code[1])) | |
338 | return -EINVAL; | |
339 | ||
340 | *(int *)(&code[1]) = new_offset; | |
341 | ||
342 | if (do_ftrace_mod_code(ip, &code)) | |
343 | return -EPERM; | |
344 | ||
345 | return 0; | |
346 | } | |
347 | ||
348 | int ftrace_enable_ftrace_graph_caller(void) | |
349 | { | |
350 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
351 | int old_offset, new_offset; | |
352 | ||
353 | old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
354 | new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
355 | ||
356 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
357 | } | |
358 | ||
359 | int ftrace_disable_ftrace_graph_caller(void) | |
360 | { | |
361 | unsigned long ip = (unsigned long)(&ftrace_graph_call); | |
362 | int old_offset, new_offset; | |
363 | ||
364 | old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE); | |
365 | new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE); | |
366 | ||
367 | return ftrace_mod_jmp(ip, old_offset, new_offset); | |
368 | } | |
369 | ||
e7d3737e FW |
370 | #endif /* !CONFIG_DYNAMIC_FTRACE */ |
371 | ||
372 | /* Add a function return address to the trace stack on thread info.*/ | |
373 | static int push_return_trace(unsigned long ret, unsigned long long time, | |
287b6e68 | 374 | unsigned long func, int *depth) |
e7d3737e FW |
375 | { |
376 | int index; | |
f201ae23 FW |
377 | |
378 | if (!current->ret_stack) | |
379 | return -EBUSY; | |
e7d3737e FW |
380 | |
381 | /* The return trace stack is full */ | |
f201ae23 FW |
382 | if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) { |
383 | atomic_inc(¤t->trace_overrun); | |
e7d3737e | 384 | return -EBUSY; |
0231022c | 385 | } |
e7d3737e | 386 | |
f201ae23 | 387 | index = ++current->curr_ret_stack; |
e7d3737e | 388 | barrier(); |
f201ae23 FW |
389 | current->ret_stack[index].ret = ret; |
390 | current->ret_stack[index].func = func; | |
391 | current->ret_stack[index].calltime = time; | |
287b6e68 | 392 | *depth = index; |
e7d3737e FW |
393 | |
394 | return 0; | |
395 | } | |
396 | ||
397 | /* Retrieve a function return address to the trace stack on thread info.*/ | |
287b6e68 | 398 | static void pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret) |
e7d3737e FW |
399 | { |
400 | int index; | |
401 | ||
f201ae23 | 402 | index = current->curr_ret_stack; |
62679efe SR |
403 | |
404 | if (unlikely(index < 0)) { | |
405 | ftrace_graph_stop(); | |
406 | WARN_ON(1); | |
407 | /* Might as well panic, otherwise we have no where to go */ | |
408 | *ret = (unsigned long)panic; | |
409 | return; | |
410 | } | |
411 | ||
f201ae23 | 412 | *ret = current->ret_stack[index].ret; |
287b6e68 FW |
413 | trace->func = current->ret_stack[index].func; |
414 | trace->calltime = current->ret_stack[index].calltime; | |
415 | trace->overrun = atomic_read(¤t->trace_overrun); | |
416 | trace->depth = index; | |
e49dc19c | 417 | barrier(); |
f201ae23 | 418 | current->curr_ret_stack--; |
62679efe | 419 | |
e7d3737e FW |
420 | } |
421 | ||
422 | /* | |
423 | * Send the trace to the ring-buffer. | |
424 | * @return the original return address. | |
425 | */ | |
426 | unsigned long ftrace_return_to_handler(void) | |
427 | { | |
fb52607a | 428 | struct ftrace_graph_ret trace; |
287b6e68 FW |
429 | unsigned long ret; |
430 | ||
431 | pop_return_trace(&trace, &ret); | |
e7d3737e | 432 | trace.rettime = cpu_clock(raw_smp_processor_id()); |
287b6e68 | 433 | ftrace_graph_return(&trace); |
e7d3737e | 434 | |
62679efe SR |
435 | if (unlikely(!ret)) { |
436 | ftrace_graph_stop(); | |
437 | WARN_ON(1); | |
438 | /* Might as well panic. What else to do? */ | |
439 | ret = (unsigned long)panic; | |
440 | } | |
441 | ||
287b6e68 | 442 | return ret; |
e7d3737e FW |
443 | } |
444 | ||
445 | /* | |
446 | * Hook the return address and push it in the stack of return addrs | |
447 | * in current thread info. | |
448 | */ | |
449 | void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr) | |
450 | { | |
451 | unsigned long old; | |
452 | unsigned long long calltime; | |
453 | int faulted; | |
287b6e68 | 454 | struct ftrace_graph_ent trace; |
e7d3737e FW |
455 | unsigned long return_hooker = (unsigned long) |
456 | &return_to_handler; | |
457 | ||
458 | /* Nmi's are currently unsupported */ | |
9a5fd902 | 459 | if (unlikely(in_nmi())) |
380c4b14 FW |
460 | return; |
461 | ||
462 | if (unlikely(atomic_read(¤t->tracing_graph_pause))) | |
e7d3737e FW |
463 | return; |
464 | ||
465 | /* | |
466 | * Protect against fault, even if it shouldn't | |
467 | * happen. This tool is too much intrusive to | |
468 | * ignore such a protection. | |
469 | */ | |
470 | asm volatile( | |
347fdd9d SR |
471 | "1: " _ASM_MOV " (%[parent_old]), %[old]\n" |
472 | "2: " _ASM_MOV " %[return_hooker], (%[parent_replaced])\n" | |
e7d3737e FW |
473 | " movl $0, %[faulted]\n" |
474 | ||
475 | ".section .fixup, \"ax\"\n" | |
476 | "3: movl $1, %[faulted]\n" | |
477 | ".previous\n" | |
478 | ||
347fdd9d SR |
479 | _ASM_EXTABLE(1b, 3b) |
480 | _ASM_EXTABLE(2b, 3b) | |
e7d3737e FW |
481 | |
482 | : [parent_replaced] "=r" (parent), [old] "=r" (old), | |
483 | [faulted] "=r" (faulted) | |
484 | : [parent_old] "0" (parent), [return_hooker] "r" (return_hooker) | |
485 | : "memory" | |
486 | ); | |
487 | ||
14a866c5 SR |
488 | if (unlikely(faulted)) { |
489 | ftrace_graph_stop(); | |
490 | WARN_ON(1); | |
e7d3737e FW |
491 | return; |
492 | } | |
493 | ||
14a866c5 SR |
494 | if (unlikely(!__kernel_text_address(old))) { |
495 | ftrace_graph_stop(); | |
e7d3737e | 496 | *parent = old; |
14a866c5 | 497 | WARN_ON(1); |
e7d3737e FW |
498 | return; |
499 | } | |
500 | ||
501 | calltime = cpu_clock(raw_smp_processor_id()); | |
502 | ||
287b6e68 FW |
503 | if (push_return_trace(old, calltime, |
504 | self_addr, &trace.depth) == -EBUSY) { | |
e7d3737e | 505 | *parent = old; |
287b6e68 FW |
506 | return; |
507 | } | |
508 | ||
509 | trace.func = self_addr; | |
287b6e68 | 510 | |
e49dc19c SR |
511 | /* Only trace if the calling function expects to */ |
512 | if (!ftrace_graph_entry(&trace)) { | |
513 | current->curr_ret_stack--; | |
514 | *parent = old; | |
515 | } | |
e7d3737e | 516 | } |
fb52607a | 517 | #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ |