1 /*
2  * ring buffer based function tracer
3  *
4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6  *
7  * Based on code from the latency_tracer, that is:
8  *
9  *  Copyright (C) 2004-2006 Ingo Molnar
10  *  Copyright (C) 2004 Nadia Yvette Chambers
11  */
12 #include <linux/ring_buffer.h>
13 #include <linux/debugfs.h>
14 #include <linux/uaccess.h>
15 #include <linux/ftrace.h>
16 #include <linux/slab.h>
17 #include <linux/fs.h>
18 
19 #include "trace.h"
20 
21 static void tracing_start_function_trace(struct trace_array *tr);
22 static void tracing_stop_function_trace(struct trace_array *tr);
23 static void
24 function_trace_call(unsigned long ip, unsigned long parent_ip,
25 		    struct ftrace_ops *op, struct pt_regs *pt_regs);
26 static void
27 function_stack_trace_call(unsigned long ip, unsigned long parent_ip,
28 			  struct ftrace_ops *op, struct pt_regs *pt_regs);
29 static struct tracer_flags func_flags;
30 
31 /* Our option */
32 enum {
33 	TRACE_FUNC_OPT_STACK	= 0x1,
34 };
35 
36 static int allocate_ftrace_ops(struct trace_array *tr)
37 {
38 	struct ftrace_ops *ops;
39 
40 	ops = kzalloc(sizeof(*ops), GFP_KERNEL);
41 	if (!ops)
42 		return -ENOMEM;
43 
44 	/* Currently only the non stack verision is supported */
45 	ops->func = function_trace_call;
46 	ops->flags = FTRACE_OPS_FL_RECURSION_SAFE;
47 
48 	tr->ops = ops;
49 	ops->private = tr;
50 	return 0;
51 }
52 
53 
54 int ftrace_create_function_files(struct trace_array *tr,
55 				 struct dentry *parent)
56 {
57 	int ret;
58 
59 	/*
60 	 * The top level array uses the "global_ops", and the files are
61 	 * created on boot up.
62 	 */
63 	if (tr->flags & TRACE_ARRAY_FL_GLOBAL)
64 		return 0;
65 
66 	ret = allocate_ftrace_ops(tr);
67 	if (ret)
68 		return ret;
69 
70 	ftrace_create_filter_files(tr->ops, parent);
71 
72 	return 0;
73 }
74 
75 void ftrace_destroy_function_files(struct trace_array *tr)
76 {
77 	ftrace_destroy_filter_files(tr->ops);
78 	kfree(tr->ops);
79 	tr->ops = NULL;
80 }
81 
82 static int function_trace_init(struct trace_array *tr)
83 {
84 	ftrace_func_t func;
85 
86 	/*
87 	 * Instance trace_arrays get their ops allocated
88 	 * at instance creation. Unless it failed
89 	 * the allocation.
90 	 */
91 	if (!tr->ops)
92 		return -ENOMEM;
93 
94 	/* Currently only the global instance can do stack tracing */
95 	if (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
96 	    func_flags.val & TRACE_FUNC_OPT_STACK)
97 		func = function_stack_trace_call;
98 	else
99 		func = function_trace_call;
100 
101 	ftrace_init_array_ops(tr, func);
102 
103 	tr->trace_buffer.cpu = get_cpu();
104 	put_cpu();
105 
106 	tracing_start_cmdline_record();
107 	tracing_start_function_trace(tr);
108 	return 0;
109 }
110 
111 static void function_trace_reset(struct trace_array *tr)
112 {
113 	tracing_stop_function_trace(tr);
114 	tracing_stop_cmdline_record();
115 	ftrace_reset_array_ops(tr);
116 }
117 
118 static void function_trace_start(struct trace_array *tr)
119 {
120 	tracing_reset_online_cpus(&tr->trace_buffer);
121 }
122 
123 static void
124 function_trace_call(unsigned long ip, unsigned long parent_ip,
125 		    struct ftrace_ops *op, struct pt_regs *pt_regs)
126 {
127 	struct trace_array *tr = op->private;
128 	struct trace_array_cpu *data;
129 	unsigned long flags;
130 	int bit;
131 	int cpu;
132 	int pc;
133 
134 	if (unlikely(!tr->function_enabled))
135 		return;
136 
137 	pc = preempt_count();
138 	preempt_disable_notrace();
139 
140 	bit = trace_test_and_set_recursion(TRACE_FTRACE_START, TRACE_FTRACE_MAX);
141 	if (bit < 0)
142 		goto out;
143 
144 	cpu = smp_processor_id();
145 	data = per_cpu_ptr(tr->trace_buffer.data, cpu);
146 	if (!atomic_read(&data->disabled)) {
147 		local_save_flags(flags);
148 		trace_function(tr, ip, parent_ip, flags, pc);
149 	}
150 	trace_clear_recursion(bit);
151 
152  out:
153 	preempt_enable_notrace();
154 }
155 
156 static void
157 function_stack_trace_call(unsigned long ip, unsigned long parent_ip,
158 			  struct ftrace_ops *op, struct pt_regs *pt_regs)
159 {
160 	struct trace_array *tr = op->private;
161 	struct trace_array_cpu *data;
162 	unsigned long flags;
163 	long disabled;
164 	int cpu;
165 	int pc;
166 
167 	if (unlikely(!tr->function_enabled))
168 		return;
169 
170 	/*
171 	 * Need to use raw, since this must be called before the
172 	 * recursive protection is performed.
173 	 */
174 	local_irq_save(flags);
175 	cpu = raw_smp_processor_id();
176 	data = per_cpu_ptr(tr->trace_buffer.data, cpu);
177 	disabled = atomic_inc_return(&data->disabled);
178 
179 	if (likely(disabled == 1)) {
180 		pc = preempt_count();
181 		trace_function(tr, ip, parent_ip, flags, pc);
182 		/*
183 		 * skip over 5 funcs:
184 		 *    __ftrace_trace_stack,
185 		 *    __trace_stack,
186 		 *    function_stack_trace_call
187 		 *    ftrace_list_func
188 		 *    ftrace_call
189 		 */
190 		__trace_stack(tr, flags, 5, pc);
191 	}
192 
193 	atomic_dec(&data->disabled);
194 	local_irq_restore(flags);
195 }
196 
197 static struct tracer_opt func_opts[] = {
198 #ifdef CONFIG_STACKTRACE
199 	{ TRACER_OPT(func_stack_trace, TRACE_FUNC_OPT_STACK) },
200 #endif
201 	{ } /* Always set a last empty entry */
202 };
203 
204 static struct tracer_flags func_flags = {
205 	.val = 0, /* By default: all flags disabled */
206 	.opts = func_opts
207 };
208 
209 static void tracing_start_function_trace(struct trace_array *tr)
210 {
211 	tr->function_enabled = 0;
212 	register_ftrace_function(tr->ops);
213 	tr->function_enabled = 1;
214 }
215 
216 static void tracing_stop_function_trace(struct trace_array *tr)
217 {
218 	tr->function_enabled = 0;
219 	unregister_ftrace_function(tr->ops);
220 }
221 
222 static int
223 func_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
224 {
225 	switch (bit) {
226 	case TRACE_FUNC_OPT_STACK:
227 		/* do nothing if already set */
228 		if (!!set == !!(func_flags.val & TRACE_FUNC_OPT_STACK))
229 			break;
230 
231 		unregister_ftrace_function(tr->ops);
232 
233 		if (set) {
234 			tr->ops->func = function_stack_trace_call;
235 			register_ftrace_function(tr->ops);
236 		} else {
237 			tr->ops->func = function_trace_call;
238 			register_ftrace_function(tr->ops);
239 		}
240 
241 		break;
242 	default:
243 		return -EINVAL;
244 	}
245 
246 	return 0;
247 }
248 
249 static struct tracer function_trace __tracer_data =
250 {
251 	.name		= "function",
252 	.init		= function_trace_init,
253 	.reset		= function_trace_reset,
254 	.start		= function_trace_start,
255 	.flags		= &func_flags,
256 	.set_flag	= func_set_flag,
257 	.allow_instances = true,
258 #ifdef CONFIG_FTRACE_SELFTEST
259 	.selftest	= trace_selftest_startup_function,
260 #endif
261 };
262 
263 #ifdef CONFIG_DYNAMIC_FTRACE
264 static void update_traceon_count(void **data, bool on)
265 {
266 	long *count = (long *)data;
267 	long old_count = *count;
268 
269 	/*
270 	 * Tracing gets disabled (or enabled) once per count.
271 	 * This function can be called at the same time on multiple CPUs.
272 	 * It is fine if both disable (or enable) tracing, as disabling
273 	 * (or enabling) the second time doesn't do anything as the
274 	 * state of the tracer is already disabled (or enabled).
275 	 * What needs to be synchronized in this case is that the count
276 	 * only gets decremented once, even if the tracer is disabled
277 	 * (or enabled) twice, as the second one is really a nop.
278 	 *
279 	 * The memory barriers guarantee that we only decrement the
280 	 * counter once. First the count is read to a local variable
281 	 * and a read barrier is used to make sure that it is loaded
282 	 * before checking if the tracer is in the state we want.
283 	 * If the tracer is not in the state we want, then the count
284 	 * is guaranteed to be the old count.
285 	 *
286 	 * Next the tracer is set to the state we want (disabled or enabled)
287 	 * then a write memory barrier is used to make sure that
288 	 * the new state is visible before changing the counter by
289 	 * one minus the old counter. This guarantees that another CPU
290 	 * executing this code will see the new state before seeing
291 	 * the new counter value, and would not do anything if the new
292 	 * counter is seen.
293 	 *
294 	 * Note, there is no synchronization between this and a user
295 	 * setting the tracing_on file. But we currently don't care
296 	 * about that.
297 	 */
298 	if (!old_count)
299 		return;
300 
301 	/* Make sure we see count before checking tracing state */
302 	smp_rmb();
303 
304 	if (on == !!tracing_is_on())
305 		return;
306 
307 	if (on)
308 		tracing_on();
309 	else
310 		tracing_off();
311 
312 	/* unlimited? */
313 	if (old_count == -1)
314 		return;
315 
316 	/* Make sure tracing state is visible before updating count */
317 	smp_wmb();
318 
319 	*count = old_count - 1;
320 }
321 
322 static void
323 ftrace_traceon_count(unsigned long ip, unsigned long parent_ip, void **data)
324 {
325 	update_traceon_count(data, 1);
326 }
327 
328 static void
329 ftrace_traceoff_count(unsigned long ip, unsigned long parent_ip, void **data)
330 {
331 	update_traceon_count(data, 0);
332 }
333 
334 static void
335 ftrace_traceon(unsigned long ip, unsigned long parent_ip, void **data)
336 {
337 	if (tracing_is_on())
338 		return;
339 
340 	tracing_on();
341 }
342 
343 static void
344 ftrace_traceoff(unsigned long ip, unsigned long parent_ip, void **data)
345 {
346 	if (!tracing_is_on())
347 		return;
348 
349 	tracing_off();
350 }
351 
352 /*
353  * Skip 4:
354  *   ftrace_stacktrace()
355  *   function_trace_probe_call()
356  *   ftrace_ops_list_func()
357  *   ftrace_call()
358  */
359 #define STACK_SKIP 4
360 
361 static void
362 ftrace_stacktrace(unsigned long ip, unsigned long parent_ip, void **data)
363 {
364 	trace_dump_stack(STACK_SKIP);
365 }
366 
367 static void
368 ftrace_stacktrace_count(unsigned long ip, unsigned long parent_ip, void **data)
369 {
370 	long *count = (long *)data;
371 	long old_count;
372 	long new_count;
373 
374 	/*
375 	 * Stack traces should only execute the number of times the
376 	 * user specified in the counter.
377 	 */
378 	do {
379 
380 		if (!tracing_is_on())
381 			return;
382 
383 		old_count = *count;
384 
385 		if (!old_count)
386 			return;
387 
388 		/* unlimited? */
389 		if (old_count == -1) {
390 			trace_dump_stack(STACK_SKIP);
391 			return;
392 		}
393 
394 		new_count = old_count - 1;
395 		new_count = cmpxchg(count, old_count, new_count);
396 		if (new_count == old_count)
397 			trace_dump_stack(STACK_SKIP);
398 
399 	} while (new_count != old_count);
400 }
401 
402 static int update_count(void **data)
403 {
404 	unsigned long *count = (long *)data;
405 
406 	if (!*count)
407 		return 0;
408 
409 	if (*count != -1)
410 		(*count)--;
411 
412 	return 1;
413 }
414 
415 static void
416 ftrace_dump_probe(unsigned long ip, unsigned long parent_ip, void **data)
417 {
418 	if (update_count(data))
419 		ftrace_dump(DUMP_ALL);
420 }
421 
422 /* Only dump the current CPU buffer. */
423 static void
424 ftrace_cpudump_probe(unsigned long ip, unsigned long parent_ip, void **data)
425 {
426 	if (update_count(data))
427 		ftrace_dump(DUMP_ORIG);
428 }
429 
430 static int
431 ftrace_probe_print(const char *name, struct seq_file *m,
432 		   unsigned long ip, void *data)
433 {
434 	long count = (long)data;
435 
436 	seq_printf(m, "%ps:%s", (void *)ip, name);
437 
438 	if (count == -1)
439 		seq_puts(m, ":unlimited\n");
440 	else
441 		seq_printf(m, ":count=%ld\n", count);
442 
443 	return 0;
444 }
445 
446 static int
447 ftrace_traceon_print(struct seq_file *m, unsigned long ip,
448 			 struct ftrace_probe_ops *ops, void *data)
449 {
450 	return ftrace_probe_print("traceon", m, ip, data);
451 }
452 
453 static int
454 ftrace_traceoff_print(struct seq_file *m, unsigned long ip,
455 			 struct ftrace_probe_ops *ops, void *data)
456 {
457 	return ftrace_probe_print("traceoff", m, ip, data);
458 }
459 
460 static int
461 ftrace_stacktrace_print(struct seq_file *m, unsigned long ip,
462 			struct ftrace_probe_ops *ops, void *data)
463 {
464 	return ftrace_probe_print("stacktrace", m, ip, data);
465 }
466 
467 static int
468 ftrace_dump_print(struct seq_file *m, unsigned long ip,
469 			struct ftrace_probe_ops *ops, void *data)
470 {
471 	return ftrace_probe_print("dump", m, ip, data);
472 }
473 
474 static int
475 ftrace_cpudump_print(struct seq_file *m, unsigned long ip,
476 			struct ftrace_probe_ops *ops, void *data)
477 {
478 	return ftrace_probe_print("cpudump", m, ip, data);
479 }
480 
481 static struct ftrace_probe_ops traceon_count_probe_ops = {
482 	.func			= ftrace_traceon_count,
483 	.print			= ftrace_traceon_print,
484 };
485 
486 static struct ftrace_probe_ops traceoff_count_probe_ops = {
487 	.func			= ftrace_traceoff_count,
488 	.print			= ftrace_traceoff_print,
489 };
490 
491 static struct ftrace_probe_ops stacktrace_count_probe_ops = {
492 	.func			= ftrace_stacktrace_count,
493 	.print			= ftrace_stacktrace_print,
494 };
495 
496 static struct ftrace_probe_ops dump_probe_ops = {
497 	.func			= ftrace_dump_probe,
498 	.print			= ftrace_dump_print,
499 };
500 
501 static struct ftrace_probe_ops cpudump_probe_ops = {
502 	.func			= ftrace_cpudump_probe,
503 	.print			= ftrace_cpudump_print,
504 };
505 
506 static struct ftrace_probe_ops traceon_probe_ops = {
507 	.func			= ftrace_traceon,
508 	.print			= ftrace_traceon_print,
509 };
510 
511 static struct ftrace_probe_ops traceoff_probe_ops = {
512 	.func			= ftrace_traceoff,
513 	.print			= ftrace_traceoff_print,
514 };
515 
516 static struct ftrace_probe_ops stacktrace_probe_ops = {
517 	.func			= ftrace_stacktrace,
518 	.print			= ftrace_stacktrace_print,
519 };
520 
521 static int
522 ftrace_trace_probe_callback(struct ftrace_probe_ops *ops,
523 			    struct ftrace_hash *hash, char *glob,
524 			    char *cmd, char *param, int enable)
525 {
526 	void *count = (void *)-1;
527 	char *number;
528 	int ret;
529 
530 	/* hash funcs only work with set_ftrace_filter */
531 	if (!enable)
532 		return -EINVAL;
533 
534 	if (glob[0] == '!') {
535 		unregister_ftrace_function_probe_func(glob+1, ops);
536 		return 0;
537 	}
538 
539 	if (!param)
540 		goto out_reg;
541 
542 	number = strsep(&param, ":");
543 
544 	if (!strlen(number))
545 		goto out_reg;
546 
547 	/*
548 	 * We use the callback data field (which is a pointer)
549 	 * as our counter.
550 	 */
551 	ret = kstrtoul(number, 0, (unsigned long *)&count);
552 	if (ret)
553 		return ret;
554 
555  out_reg:
556 	ret = register_ftrace_function_probe(glob, ops, count);
557 
558 	return ret < 0 ? ret : 0;
559 }
560 
561 static int
562 ftrace_trace_onoff_callback(struct ftrace_hash *hash,
563 			    char *glob, char *cmd, char *param, int enable)
564 {
565 	struct ftrace_probe_ops *ops;
566 
567 	/* we register both traceon and traceoff to this callback */
568 	if (strcmp(cmd, "traceon") == 0)
569 		ops = param ? &traceon_count_probe_ops : &traceon_probe_ops;
570 	else
571 		ops = param ? &traceoff_count_probe_ops : &traceoff_probe_ops;
572 
573 	return ftrace_trace_probe_callback(ops, hash, glob, cmd,
574 					   param, enable);
575 }
576 
577 static int
578 ftrace_stacktrace_callback(struct ftrace_hash *hash,
579 			   char *glob, char *cmd, char *param, int enable)
580 {
581 	struct ftrace_probe_ops *ops;
582 
583 	ops = param ? &stacktrace_count_probe_ops : &stacktrace_probe_ops;
584 
585 	return ftrace_trace_probe_callback(ops, hash, glob, cmd,
586 					   param, enable);
587 }
588 
589 static int
590 ftrace_dump_callback(struct ftrace_hash *hash,
591 			   char *glob, char *cmd, char *param, int enable)
592 {
593 	struct ftrace_probe_ops *ops;
594 
595 	ops = &dump_probe_ops;
596 
597 	/* Only dump once. */
598 	return ftrace_trace_probe_callback(ops, hash, glob, cmd,
599 					   "1", enable);
600 }
601 
602 static int
603 ftrace_cpudump_callback(struct ftrace_hash *hash,
604 			   char *glob, char *cmd, char *param, int enable)
605 {
606 	struct ftrace_probe_ops *ops;
607 
608 	ops = &cpudump_probe_ops;
609 
610 	/* Only dump once. */
611 	return ftrace_trace_probe_callback(ops, hash, glob, cmd,
612 					   "1", enable);
613 }
614 
615 static struct ftrace_func_command ftrace_traceon_cmd = {
616 	.name			= "traceon",
617 	.func			= ftrace_trace_onoff_callback,
618 };
619 
620 static struct ftrace_func_command ftrace_traceoff_cmd = {
621 	.name			= "traceoff",
622 	.func			= ftrace_trace_onoff_callback,
623 };
624 
625 static struct ftrace_func_command ftrace_stacktrace_cmd = {
626 	.name			= "stacktrace",
627 	.func			= ftrace_stacktrace_callback,
628 };
629 
630 static struct ftrace_func_command ftrace_dump_cmd = {
631 	.name			= "dump",
632 	.func			= ftrace_dump_callback,
633 };
634 
635 static struct ftrace_func_command ftrace_cpudump_cmd = {
636 	.name			= "cpudump",
637 	.func			= ftrace_cpudump_callback,
638 };
639 
640 static int __init init_func_cmd_traceon(void)
641 {
642 	int ret;
643 
644 	ret = register_ftrace_command(&ftrace_traceoff_cmd);
645 	if (ret)
646 		return ret;
647 
648 	ret = register_ftrace_command(&ftrace_traceon_cmd);
649 	if (ret)
650 		goto out_free_traceoff;
651 
652 	ret = register_ftrace_command(&ftrace_stacktrace_cmd);
653 	if (ret)
654 		goto out_free_traceon;
655 
656 	ret = register_ftrace_command(&ftrace_dump_cmd);
657 	if (ret)
658 		goto out_free_stacktrace;
659 
660 	ret = register_ftrace_command(&ftrace_cpudump_cmd);
661 	if (ret)
662 		goto out_free_dump;
663 
664 	return 0;
665 
666  out_free_dump:
667 	unregister_ftrace_command(&ftrace_dump_cmd);
668  out_free_stacktrace:
669 	unregister_ftrace_command(&ftrace_stacktrace_cmd);
670  out_free_traceon:
671 	unregister_ftrace_command(&ftrace_traceon_cmd);
672  out_free_traceoff:
673 	unregister_ftrace_command(&ftrace_traceoff_cmd);
674 
675 	return ret;
676 }
677 #else
678 static inline int init_func_cmd_traceon(void)
679 {
680 	return 0;
681 }
682 #endif /* CONFIG_DYNAMIC_FTRACE */
683 
684 static __init int init_function_trace(void)
685 {
686 	init_func_cmd_traceon();
687 	return register_tracer(&function_trace);
688 }
689 core_initcall(init_function_trace);
690