xref: /linux/kernel/trace/trace_functions_graph.c (revision a5c4300389bb33ade2515c082709217f0614cf15)
1 /*
2  *
3  * Function graph tracer.
4  * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
5  * Mostly borrowed from function tracer which
6  * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
7  *
8  */
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/ftrace.h>
12 #include <linux/slab.h>
13 #include <linux/fs.h>
14 
15 #include "trace.h"
16 #include "trace_output.h"
17 
18 struct fgraph_cpu_data {
19 	pid_t		last_pid;
20 	int		depth;
21 	int		ignore;
22 	unsigned long	enter_funcs[FTRACE_RETFUNC_DEPTH];
23 };
24 
25 struct fgraph_data {
26 	struct fgraph_cpu_data		*cpu_data;
27 
28 	/* Place to preserve last processed entry. */
29 	struct ftrace_graph_ent_entry	ent;
30 	struct ftrace_graph_ret_entry	ret;
31 	int				failed;
32 	int				cpu;
33 };
34 
35 #define TRACE_GRAPH_INDENT	2
36 
37 /* Flag options */
38 #define TRACE_GRAPH_PRINT_OVERRUN	0x1
39 #define TRACE_GRAPH_PRINT_CPU		0x2
40 #define TRACE_GRAPH_PRINT_OVERHEAD	0x4
41 #define TRACE_GRAPH_PRINT_PROC		0x8
42 #define TRACE_GRAPH_PRINT_DURATION	0x10
43 #define TRACE_GRAPH_PRINT_ABS_TIME	0x20
44 
45 static struct tracer_opt trace_opts[] = {
46 	/* Display overruns? (for self-debug purpose) */
47 	{ TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
48 	/* Display CPU ? */
49 	{ TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
50 	/* Display Overhead ? */
51 	{ TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
52 	/* Display proc name/pid */
53 	{ TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
54 	/* Display duration of execution */
55 	{ TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
56 	/* Display absolute time of an entry */
57 	{ TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
58 	{ } /* Empty entry */
59 };
60 
61 static struct tracer_flags tracer_flags = {
62 	/* Don't display overruns and proc by default */
63 	.val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
64 	       TRACE_GRAPH_PRINT_DURATION,
65 	.opts = trace_opts
66 };
67 
68 static struct trace_array *graph_array;
69 
70 
71 /* Add a function return address to the trace stack on thread info.*/
72 int
73 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
74 			 unsigned long frame_pointer)
75 {
76 	unsigned long long calltime;
77 	int index;
78 
79 	if (!current->ret_stack)
80 		return -EBUSY;
81 
82 	/*
83 	 * We must make sure the ret_stack is tested before we read
84 	 * anything else.
85 	 */
86 	smp_rmb();
87 
88 	/* The return trace stack is full */
89 	if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
90 		atomic_inc(&current->trace_overrun);
91 		return -EBUSY;
92 	}
93 
94 	calltime = trace_clock_local();
95 
96 	index = ++current->curr_ret_stack;
97 	barrier();
98 	current->ret_stack[index].ret = ret;
99 	current->ret_stack[index].func = func;
100 	current->ret_stack[index].calltime = calltime;
101 	current->ret_stack[index].subtime = 0;
102 	current->ret_stack[index].fp = frame_pointer;
103 	*depth = index;
104 
105 	return 0;
106 }
107 
108 /* Retrieve a function return address to the trace stack on thread info.*/
109 static void
110 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
111 			unsigned long frame_pointer)
112 {
113 	int index;
114 
115 	index = current->curr_ret_stack;
116 
117 	if (unlikely(index < 0)) {
118 		ftrace_graph_stop();
119 		WARN_ON(1);
120 		/* Might as well panic, otherwise we have no where to go */
121 		*ret = (unsigned long)panic;
122 		return;
123 	}
124 
125 #ifdef CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST
126 	/*
127 	 * The arch may choose to record the frame pointer used
128 	 * and check it here to make sure that it is what we expect it
129 	 * to be. If gcc does not set the place holder of the return
130 	 * address in the frame pointer, and does a copy instead, then
131 	 * the function graph trace will fail. This test detects this
132 	 * case.
133 	 *
134 	 * Currently, x86_32 with optimize for size (-Os) makes the latest
135 	 * gcc do the above.
136 	 */
137 	if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
138 		ftrace_graph_stop();
139 		WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
140 		     "  from func %ps return to %lx\n",
141 		     current->ret_stack[index].fp,
142 		     frame_pointer,
143 		     (void *)current->ret_stack[index].func,
144 		     current->ret_stack[index].ret);
145 		*ret = (unsigned long)panic;
146 		return;
147 	}
148 #endif
149 
150 	*ret = current->ret_stack[index].ret;
151 	trace->func = current->ret_stack[index].func;
152 	trace->calltime = current->ret_stack[index].calltime;
153 	trace->overrun = atomic_read(&current->trace_overrun);
154 	trace->depth = index;
155 }
156 
157 /*
158  * Send the trace to the ring-buffer.
159  * @return the original return address.
160  */
161 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
162 {
163 	struct ftrace_graph_ret trace;
164 	unsigned long ret;
165 
166 	ftrace_pop_return_trace(&trace, &ret, frame_pointer);
167 	trace.rettime = trace_clock_local();
168 	ftrace_graph_return(&trace);
169 	barrier();
170 	current->curr_ret_stack--;
171 
172 	if (unlikely(!ret)) {
173 		ftrace_graph_stop();
174 		WARN_ON(1);
175 		/* Might as well panic. What else to do? */
176 		ret = (unsigned long)panic;
177 	}
178 
179 	return ret;
180 }
181 
182 int __trace_graph_entry(struct trace_array *tr,
183 				struct ftrace_graph_ent *trace,
184 				unsigned long flags,
185 				int pc)
186 {
187 	struct ftrace_event_call *call = &event_funcgraph_entry;
188 	struct ring_buffer_event *event;
189 	struct ring_buffer *buffer = tr->buffer;
190 	struct ftrace_graph_ent_entry *entry;
191 
192 	if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
193 		return 0;
194 
195 	event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
196 					  sizeof(*entry), flags, pc);
197 	if (!event)
198 		return 0;
199 	entry	= ring_buffer_event_data(event);
200 	entry->graph_ent			= *trace;
201 	if (!filter_current_check_discard(buffer, call, entry, event))
202 		ring_buffer_unlock_commit(buffer, event);
203 
204 	return 1;
205 }
206 
207 int trace_graph_entry(struct ftrace_graph_ent *trace)
208 {
209 	struct trace_array *tr = graph_array;
210 	struct trace_array_cpu *data;
211 	unsigned long flags;
212 	long disabled;
213 	int ret;
214 	int cpu;
215 	int pc;
216 
217 	if (!ftrace_trace_task(current))
218 		return 0;
219 
220 	/* trace it when it is-nested-in or is a function enabled. */
221 	if (!(trace->depth || ftrace_graph_addr(trace->func)))
222 		return 0;
223 
224 	local_irq_save(flags);
225 	cpu = raw_smp_processor_id();
226 	data = tr->data[cpu];
227 	disabled = atomic_inc_return(&data->disabled);
228 	if (likely(disabled == 1)) {
229 		pc = preempt_count();
230 		ret = __trace_graph_entry(tr, trace, flags, pc);
231 	} else {
232 		ret = 0;
233 	}
234 
235 	atomic_dec(&data->disabled);
236 	local_irq_restore(flags);
237 
238 	return ret;
239 }
240 
241 int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
242 {
243 	if (tracing_thresh)
244 		return 1;
245 	else
246 		return trace_graph_entry(trace);
247 }
248 
249 void __trace_graph_return(struct trace_array *tr,
250 				struct ftrace_graph_ret *trace,
251 				unsigned long flags,
252 				int pc)
253 {
254 	struct ftrace_event_call *call = &event_funcgraph_exit;
255 	struct ring_buffer_event *event;
256 	struct ring_buffer *buffer = tr->buffer;
257 	struct ftrace_graph_ret_entry *entry;
258 
259 	if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
260 		return;
261 
262 	event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
263 					  sizeof(*entry), flags, pc);
264 	if (!event)
265 		return;
266 	entry	= ring_buffer_event_data(event);
267 	entry->ret				= *trace;
268 	if (!filter_current_check_discard(buffer, call, entry, event))
269 		ring_buffer_unlock_commit(buffer, event);
270 }
271 
272 void trace_graph_return(struct ftrace_graph_ret *trace)
273 {
274 	struct trace_array *tr = graph_array;
275 	struct trace_array_cpu *data;
276 	unsigned long flags;
277 	long disabled;
278 	int cpu;
279 	int pc;
280 
281 	local_irq_save(flags);
282 	cpu = raw_smp_processor_id();
283 	data = tr->data[cpu];
284 	disabled = atomic_inc_return(&data->disabled);
285 	if (likely(disabled == 1)) {
286 		pc = preempt_count();
287 		__trace_graph_return(tr, trace, flags, pc);
288 	}
289 	atomic_dec(&data->disabled);
290 	local_irq_restore(flags);
291 }
292 
293 void set_graph_array(struct trace_array *tr)
294 {
295 	graph_array = tr;
296 
297 	/* Make graph_array visible before we start tracing */
298 
299 	smp_mb();
300 }
301 
302 void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
303 {
304 	if (tracing_thresh &&
305 	    (trace->rettime - trace->calltime < tracing_thresh))
306 		return;
307 	else
308 		trace_graph_return(trace);
309 }
310 
311 static int graph_trace_init(struct trace_array *tr)
312 {
313 	int ret;
314 
315 	set_graph_array(tr);
316 	if (tracing_thresh)
317 		ret = register_ftrace_graph(&trace_graph_thresh_return,
318 					    &trace_graph_thresh_entry);
319 	else
320 		ret = register_ftrace_graph(&trace_graph_return,
321 					    &trace_graph_entry);
322 	if (ret)
323 		return ret;
324 	tracing_start_cmdline_record();
325 
326 	return 0;
327 }
328 
329 static void graph_trace_reset(struct trace_array *tr)
330 {
331 	tracing_stop_cmdline_record();
332 	unregister_ftrace_graph();
333 }
334 
335 static int max_bytes_for_cpu;
336 
337 static enum print_line_t
338 print_graph_cpu(struct trace_seq *s, int cpu)
339 {
340 	int ret;
341 
342 	/*
343 	 * Start with a space character - to make it stand out
344 	 * to the right a bit when trace output is pasted into
345 	 * email:
346 	 */
347 	ret = trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
348 	if (!ret)
349 		return TRACE_TYPE_PARTIAL_LINE;
350 
351 	return TRACE_TYPE_HANDLED;
352 }
353 
354 #define TRACE_GRAPH_PROCINFO_LENGTH	14
355 
356 static enum print_line_t
357 print_graph_proc(struct trace_seq *s, pid_t pid)
358 {
359 	char comm[TASK_COMM_LEN];
360 	/* sign + log10(MAX_INT) + '\0' */
361 	char pid_str[11];
362 	int spaces = 0;
363 	int ret;
364 	int len;
365 	int i;
366 
367 	trace_find_cmdline(pid, comm);
368 	comm[7] = '\0';
369 	sprintf(pid_str, "%d", pid);
370 
371 	/* 1 stands for the "-" character */
372 	len = strlen(comm) + strlen(pid_str) + 1;
373 
374 	if (len < TRACE_GRAPH_PROCINFO_LENGTH)
375 		spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
376 
377 	/* First spaces to align center */
378 	for (i = 0; i < spaces / 2; i++) {
379 		ret = trace_seq_printf(s, " ");
380 		if (!ret)
381 			return TRACE_TYPE_PARTIAL_LINE;
382 	}
383 
384 	ret = trace_seq_printf(s, "%s-%s", comm, pid_str);
385 	if (!ret)
386 		return TRACE_TYPE_PARTIAL_LINE;
387 
388 	/* Last spaces to align center */
389 	for (i = 0; i < spaces - (spaces / 2); i++) {
390 		ret = trace_seq_printf(s, " ");
391 		if (!ret)
392 			return TRACE_TYPE_PARTIAL_LINE;
393 	}
394 	return TRACE_TYPE_HANDLED;
395 }
396 
397 
398 static enum print_line_t
399 print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
400 {
401 	if (!trace_seq_putc(s, ' '))
402 		return 0;
403 
404 	return trace_print_lat_fmt(s, entry);
405 }
406 
407 /* If the pid changed since the last trace, output this event */
408 static enum print_line_t
409 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
410 {
411 	pid_t prev_pid;
412 	pid_t *last_pid;
413 	int ret;
414 
415 	if (!data)
416 		return TRACE_TYPE_HANDLED;
417 
418 	last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
419 
420 	if (*last_pid == pid)
421 		return TRACE_TYPE_HANDLED;
422 
423 	prev_pid = *last_pid;
424 	*last_pid = pid;
425 
426 	if (prev_pid == -1)
427 		return TRACE_TYPE_HANDLED;
428 /*
429  * Context-switch trace line:
430 
431  ------------------------------------------
432  | 1)  migration/0--1  =>  sshd-1755
433  ------------------------------------------
434 
435  */
436 	ret = trace_seq_printf(s,
437 		" ------------------------------------------\n");
438 	if (!ret)
439 		return TRACE_TYPE_PARTIAL_LINE;
440 
441 	ret = print_graph_cpu(s, cpu);
442 	if (ret == TRACE_TYPE_PARTIAL_LINE)
443 		return TRACE_TYPE_PARTIAL_LINE;
444 
445 	ret = print_graph_proc(s, prev_pid);
446 	if (ret == TRACE_TYPE_PARTIAL_LINE)
447 		return TRACE_TYPE_PARTIAL_LINE;
448 
449 	ret = trace_seq_printf(s, " => ");
450 	if (!ret)
451 		return TRACE_TYPE_PARTIAL_LINE;
452 
453 	ret = print_graph_proc(s, pid);
454 	if (ret == TRACE_TYPE_PARTIAL_LINE)
455 		return TRACE_TYPE_PARTIAL_LINE;
456 
457 	ret = trace_seq_printf(s,
458 		"\n ------------------------------------------\n\n");
459 	if (!ret)
460 		return TRACE_TYPE_PARTIAL_LINE;
461 
462 	return TRACE_TYPE_HANDLED;
463 }
464 
465 static struct ftrace_graph_ret_entry *
466 get_return_for_leaf(struct trace_iterator *iter,
467 		struct ftrace_graph_ent_entry *curr)
468 {
469 	struct fgraph_data *data = iter->private;
470 	struct ring_buffer_iter *ring_iter = NULL;
471 	struct ring_buffer_event *event;
472 	struct ftrace_graph_ret_entry *next;
473 
474 	/*
475 	 * If the previous output failed to write to the seq buffer,
476 	 * then we just reuse the data from before.
477 	 */
478 	if (data && data->failed) {
479 		curr = &data->ent;
480 		next = &data->ret;
481 	} else {
482 
483 		ring_iter = iter->buffer_iter[iter->cpu];
484 
485 		/* First peek to compare current entry and the next one */
486 		if (ring_iter)
487 			event = ring_buffer_iter_peek(ring_iter, NULL);
488 		else {
489 			/*
490 			 * We need to consume the current entry to see
491 			 * the next one.
492 			 */
493 			ring_buffer_consume(iter->tr->buffer, iter->cpu,
494 					    NULL, NULL);
495 			event = ring_buffer_peek(iter->tr->buffer, iter->cpu,
496 						 NULL, NULL);
497 		}
498 
499 		if (!event)
500 			return NULL;
501 
502 		next = ring_buffer_event_data(event);
503 
504 		if (data) {
505 			/*
506 			 * Save current and next entries for later reference
507 			 * if the output fails.
508 			 */
509 			data->ent = *curr;
510 			data->ret = *next;
511 		}
512 	}
513 
514 	if (next->ent.type != TRACE_GRAPH_RET)
515 		return NULL;
516 
517 	if (curr->ent.pid != next->ent.pid ||
518 			curr->graph_ent.func != next->ret.func)
519 		return NULL;
520 
521 	/* this is a leaf, now advance the iterator */
522 	if (ring_iter)
523 		ring_buffer_read(ring_iter, NULL);
524 
525 	return next;
526 }
527 
528 /* Signal a overhead of time execution to the output */
529 static int
530 print_graph_overhead(unsigned long long duration, struct trace_seq *s,
531 		     u32 flags)
532 {
533 	/* If duration disappear, we don't need anything */
534 	if (!(flags & TRACE_GRAPH_PRINT_DURATION))
535 		return 1;
536 
537 	/* Non nested entry or return */
538 	if (duration == -1)
539 		return trace_seq_printf(s, "  ");
540 
541 	if (flags & TRACE_GRAPH_PRINT_OVERHEAD) {
542 		/* Duration exceeded 100 msecs */
543 		if (duration > 100000ULL)
544 			return trace_seq_printf(s, "! ");
545 
546 		/* Duration exceeded 10 msecs */
547 		if (duration > 10000ULL)
548 			return trace_seq_printf(s, "+ ");
549 	}
550 
551 	return trace_seq_printf(s, "  ");
552 }
553 
554 static int print_graph_abs_time(u64 t, struct trace_seq *s)
555 {
556 	unsigned long usecs_rem;
557 
558 	usecs_rem = do_div(t, NSEC_PER_SEC);
559 	usecs_rem /= 1000;
560 
561 	return trace_seq_printf(s, "%5lu.%06lu |  ",
562 			(unsigned long)t, usecs_rem);
563 }
564 
565 static enum print_line_t
566 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
567 		enum trace_type type, int cpu, pid_t pid, u32 flags)
568 {
569 	int ret;
570 	struct trace_seq *s = &iter->seq;
571 
572 	if (addr < (unsigned long)__irqentry_text_start ||
573 		addr >= (unsigned long)__irqentry_text_end)
574 		return TRACE_TYPE_UNHANDLED;
575 
576 	/* Absolute time */
577 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
578 		ret = print_graph_abs_time(iter->ts, s);
579 		if (!ret)
580 			return TRACE_TYPE_PARTIAL_LINE;
581 	}
582 
583 	/* Cpu */
584 	if (flags & TRACE_GRAPH_PRINT_CPU) {
585 		ret = print_graph_cpu(s, cpu);
586 		if (ret == TRACE_TYPE_PARTIAL_LINE)
587 			return TRACE_TYPE_PARTIAL_LINE;
588 	}
589 
590 	/* Proc */
591 	if (flags & TRACE_GRAPH_PRINT_PROC) {
592 		ret = print_graph_proc(s, pid);
593 		if (ret == TRACE_TYPE_PARTIAL_LINE)
594 			return TRACE_TYPE_PARTIAL_LINE;
595 		ret = trace_seq_printf(s, " | ");
596 		if (!ret)
597 			return TRACE_TYPE_PARTIAL_LINE;
598 	}
599 
600 	/* No overhead */
601 	ret = print_graph_overhead(-1, s, flags);
602 	if (!ret)
603 		return TRACE_TYPE_PARTIAL_LINE;
604 
605 	if (type == TRACE_GRAPH_ENT)
606 		ret = trace_seq_printf(s, "==========>");
607 	else
608 		ret = trace_seq_printf(s, "<==========");
609 
610 	if (!ret)
611 		return TRACE_TYPE_PARTIAL_LINE;
612 
613 	/* Don't close the duration column if haven't one */
614 	if (flags & TRACE_GRAPH_PRINT_DURATION)
615 		trace_seq_printf(s, " |");
616 	ret = trace_seq_printf(s, "\n");
617 
618 	if (!ret)
619 		return TRACE_TYPE_PARTIAL_LINE;
620 	return TRACE_TYPE_HANDLED;
621 }
622 
623 enum print_line_t
624 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
625 {
626 	unsigned long nsecs_rem = do_div(duration, 1000);
627 	/* log10(ULONG_MAX) + '\0' */
628 	char msecs_str[21];
629 	char nsecs_str[5];
630 	int ret, len;
631 	int i;
632 
633 	sprintf(msecs_str, "%lu", (unsigned long) duration);
634 
635 	/* Print msecs */
636 	ret = trace_seq_printf(s, "%s", msecs_str);
637 	if (!ret)
638 		return TRACE_TYPE_PARTIAL_LINE;
639 
640 	len = strlen(msecs_str);
641 
642 	/* Print nsecs (we don't want to exceed 7 numbers) */
643 	if (len < 7) {
644 		snprintf(nsecs_str, 8 - len, "%03lu", nsecs_rem);
645 		ret = trace_seq_printf(s, ".%s", nsecs_str);
646 		if (!ret)
647 			return TRACE_TYPE_PARTIAL_LINE;
648 		len += strlen(nsecs_str);
649 	}
650 
651 	ret = trace_seq_printf(s, " us ");
652 	if (!ret)
653 		return TRACE_TYPE_PARTIAL_LINE;
654 
655 	/* Print remaining spaces to fit the row's width */
656 	for (i = len; i < 7; i++) {
657 		ret = trace_seq_printf(s, " ");
658 		if (!ret)
659 			return TRACE_TYPE_PARTIAL_LINE;
660 	}
661 	return TRACE_TYPE_HANDLED;
662 }
663 
664 static enum print_line_t
665 print_graph_duration(unsigned long long duration, struct trace_seq *s)
666 {
667 	int ret;
668 
669 	ret = trace_print_graph_duration(duration, s);
670 	if (ret != TRACE_TYPE_HANDLED)
671 		return ret;
672 
673 	ret = trace_seq_printf(s, "|  ");
674 	if (!ret)
675 		return TRACE_TYPE_PARTIAL_LINE;
676 
677 	return TRACE_TYPE_HANDLED;
678 }
679 
680 /* Case of a leaf function on its call entry */
681 static enum print_line_t
682 print_graph_entry_leaf(struct trace_iterator *iter,
683 		struct ftrace_graph_ent_entry *entry,
684 		struct ftrace_graph_ret_entry *ret_entry,
685 		struct trace_seq *s, u32 flags)
686 {
687 	struct fgraph_data *data = iter->private;
688 	struct ftrace_graph_ret *graph_ret;
689 	struct ftrace_graph_ent *call;
690 	unsigned long long duration;
691 	int ret;
692 	int i;
693 
694 	graph_ret = &ret_entry->ret;
695 	call = &entry->graph_ent;
696 	duration = graph_ret->rettime - graph_ret->calltime;
697 
698 	if (data) {
699 		struct fgraph_cpu_data *cpu_data;
700 		int cpu = iter->cpu;
701 
702 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
703 
704 		/*
705 		 * Comments display at + 1 to depth. Since
706 		 * this is a leaf function, keep the comments
707 		 * equal to this depth.
708 		 */
709 		cpu_data->depth = call->depth - 1;
710 
711 		/* No need to keep this function around for this depth */
712 		if (call->depth < FTRACE_RETFUNC_DEPTH)
713 			cpu_data->enter_funcs[call->depth] = 0;
714 	}
715 
716 	/* Overhead */
717 	ret = print_graph_overhead(duration, s, flags);
718 	if (!ret)
719 		return TRACE_TYPE_PARTIAL_LINE;
720 
721 	/* Duration */
722 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
723 		ret = print_graph_duration(duration, s);
724 		if (ret == TRACE_TYPE_PARTIAL_LINE)
725 			return TRACE_TYPE_PARTIAL_LINE;
726 	}
727 
728 	/* Function */
729 	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
730 		ret = trace_seq_printf(s, " ");
731 		if (!ret)
732 			return TRACE_TYPE_PARTIAL_LINE;
733 	}
734 
735 	ret = trace_seq_printf(s, "%ps();\n", (void *)call->func);
736 	if (!ret)
737 		return TRACE_TYPE_PARTIAL_LINE;
738 
739 	return TRACE_TYPE_HANDLED;
740 }
741 
742 static enum print_line_t
743 print_graph_entry_nested(struct trace_iterator *iter,
744 			 struct ftrace_graph_ent_entry *entry,
745 			 struct trace_seq *s, int cpu, u32 flags)
746 {
747 	struct ftrace_graph_ent *call = &entry->graph_ent;
748 	struct fgraph_data *data = iter->private;
749 	int ret;
750 	int i;
751 
752 	if (data) {
753 		struct fgraph_cpu_data *cpu_data;
754 		int cpu = iter->cpu;
755 
756 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
757 		cpu_data->depth = call->depth;
758 
759 		/* Save this function pointer to see if the exit matches */
760 		if (call->depth < FTRACE_RETFUNC_DEPTH)
761 			cpu_data->enter_funcs[call->depth] = call->func;
762 	}
763 
764 	/* No overhead */
765 	ret = print_graph_overhead(-1, s, flags);
766 	if (!ret)
767 		return TRACE_TYPE_PARTIAL_LINE;
768 
769 	/* No time */
770 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
771 		ret = trace_seq_printf(s, "            |  ");
772 		if (!ret)
773 			return TRACE_TYPE_PARTIAL_LINE;
774 	}
775 
776 	/* Function */
777 	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
778 		ret = trace_seq_printf(s, " ");
779 		if (!ret)
780 			return TRACE_TYPE_PARTIAL_LINE;
781 	}
782 
783 	ret = trace_seq_printf(s, "%ps() {\n", (void *)call->func);
784 	if (!ret)
785 		return TRACE_TYPE_PARTIAL_LINE;
786 
787 	/*
788 	 * we already consumed the current entry to check the next one
789 	 * and see if this is a leaf.
790 	 */
791 	return TRACE_TYPE_NO_CONSUME;
792 }
793 
794 static enum print_line_t
795 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
796 		     int type, unsigned long addr, u32 flags)
797 {
798 	struct fgraph_data *data = iter->private;
799 	struct trace_entry *ent = iter->ent;
800 	int cpu = iter->cpu;
801 	int ret;
802 
803 	/* Pid */
804 	if (verif_pid(s, ent->pid, cpu, data) == TRACE_TYPE_PARTIAL_LINE)
805 		return TRACE_TYPE_PARTIAL_LINE;
806 
807 	if (type) {
808 		/* Interrupt */
809 		ret = print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
810 		if (ret == TRACE_TYPE_PARTIAL_LINE)
811 			return TRACE_TYPE_PARTIAL_LINE;
812 	}
813 
814 	/* Absolute time */
815 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
816 		ret = print_graph_abs_time(iter->ts, s);
817 		if (!ret)
818 			return TRACE_TYPE_PARTIAL_LINE;
819 	}
820 
821 	/* Cpu */
822 	if (flags & TRACE_GRAPH_PRINT_CPU) {
823 		ret = print_graph_cpu(s, cpu);
824 		if (ret == TRACE_TYPE_PARTIAL_LINE)
825 			return TRACE_TYPE_PARTIAL_LINE;
826 	}
827 
828 	/* Proc */
829 	if (flags & TRACE_GRAPH_PRINT_PROC) {
830 		ret = print_graph_proc(s, ent->pid);
831 		if (ret == TRACE_TYPE_PARTIAL_LINE)
832 			return TRACE_TYPE_PARTIAL_LINE;
833 
834 		ret = trace_seq_printf(s, " | ");
835 		if (!ret)
836 			return TRACE_TYPE_PARTIAL_LINE;
837 	}
838 
839 	/* Latency format */
840 	if (trace_flags & TRACE_ITER_LATENCY_FMT) {
841 		ret = print_graph_lat_fmt(s, ent);
842 		if (ret == TRACE_TYPE_PARTIAL_LINE)
843 			return TRACE_TYPE_PARTIAL_LINE;
844 	}
845 
846 	return 0;
847 }
848 
849 static enum print_line_t
850 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
851 			struct trace_iterator *iter, u32 flags)
852 {
853 	struct fgraph_data *data = iter->private;
854 	struct ftrace_graph_ent *call = &field->graph_ent;
855 	struct ftrace_graph_ret_entry *leaf_ret;
856 	static enum print_line_t ret;
857 	int cpu = iter->cpu;
858 
859 	if (print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags))
860 		return TRACE_TYPE_PARTIAL_LINE;
861 
862 	leaf_ret = get_return_for_leaf(iter, field);
863 	if (leaf_ret)
864 		ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
865 	else
866 		ret = print_graph_entry_nested(iter, field, s, cpu, flags);
867 
868 	if (data) {
869 		/*
870 		 * If we failed to write our output, then we need to make
871 		 * note of it. Because we already consumed our entry.
872 		 */
873 		if (s->full) {
874 			data->failed = 1;
875 			data->cpu = cpu;
876 		} else
877 			data->failed = 0;
878 	}
879 
880 	return ret;
881 }
882 
883 static enum print_line_t
884 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
885 		   struct trace_entry *ent, struct trace_iterator *iter,
886 		   u32 flags)
887 {
888 	unsigned long long duration = trace->rettime - trace->calltime;
889 	struct fgraph_data *data = iter->private;
890 	pid_t pid = ent->pid;
891 	int cpu = iter->cpu;
892 	int func_match = 1;
893 	int ret;
894 	int i;
895 
896 	if (data) {
897 		struct fgraph_cpu_data *cpu_data;
898 		int cpu = iter->cpu;
899 
900 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
901 
902 		/*
903 		 * Comments display at + 1 to depth. This is the
904 		 * return from a function, we now want the comments
905 		 * to display at the same level of the bracket.
906 		 */
907 		cpu_data->depth = trace->depth - 1;
908 
909 		if (trace->depth < FTRACE_RETFUNC_DEPTH) {
910 			if (cpu_data->enter_funcs[trace->depth] != trace->func)
911 				func_match = 0;
912 			cpu_data->enter_funcs[trace->depth] = 0;
913 		}
914 	}
915 
916 	if (print_graph_prologue(iter, s, 0, 0, flags))
917 		return TRACE_TYPE_PARTIAL_LINE;
918 
919 	/* Overhead */
920 	ret = print_graph_overhead(duration, s, flags);
921 	if (!ret)
922 		return TRACE_TYPE_PARTIAL_LINE;
923 
924 	/* Duration */
925 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
926 		ret = print_graph_duration(duration, s);
927 		if (ret == TRACE_TYPE_PARTIAL_LINE)
928 			return TRACE_TYPE_PARTIAL_LINE;
929 	}
930 
931 	/* Closing brace */
932 	for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++) {
933 		ret = trace_seq_printf(s, " ");
934 		if (!ret)
935 			return TRACE_TYPE_PARTIAL_LINE;
936 	}
937 
938 	/*
939 	 * If the return function does not have a matching entry,
940 	 * then the entry was lost. Instead of just printing
941 	 * the '}' and letting the user guess what function this
942 	 * belongs to, write out the function name.
943 	 */
944 	if (func_match) {
945 		ret = trace_seq_printf(s, "}\n");
946 		if (!ret)
947 			return TRACE_TYPE_PARTIAL_LINE;
948 	} else {
949 		ret = trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
950 		if (!ret)
951 			return TRACE_TYPE_PARTIAL_LINE;
952 	}
953 
954 	/* Overrun */
955 	if (flags & TRACE_GRAPH_PRINT_OVERRUN) {
956 		ret = trace_seq_printf(s, " (Overruns: %lu)\n",
957 					trace->overrun);
958 		if (!ret)
959 			return TRACE_TYPE_PARTIAL_LINE;
960 	}
961 
962 	ret = print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
963 			      cpu, pid, flags);
964 	if (ret == TRACE_TYPE_PARTIAL_LINE)
965 		return TRACE_TYPE_PARTIAL_LINE;
966 
967 	return TRACE_TYPE_HANDLED;
968 }
969 
970 static enum print_line_t
971 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
972 		    struct trace_iterator *iter, u32 flags)
973 {
974 	unsigned long sym_flags = (trace_flags & TRACE_ITER_SYM_MASK);
975 	struct fgraph_data *data = iter->private;
976 	struct trace_event *event;
977 	int depth = 0;
978 	int ret;
979 	int i;
980 
981 	if (data)
982 		depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
983 
984 	if (print_graph_prologue(iter, s, 0, 0, flags))
985 		return TRACE_TYPE_PARTIAL_LINE;
986 
987 	/* No overhead */
988 	ret = print_graph_overhead(-1, s, flags);
989 	if (!ret)
990 		return TRACE_TYPE_PARTIAL_LINE;
991 
992 	/* No time */
993 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
994 		ret = trace_seq_printf(s, "            |  ");
995 		if (!ret)
996 			return TRACE_TYPE_PARTIAL_LINE;
997 	}
998 
999 	/* Indentation */
1000 	if (depth > 0)
1001 		for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++) {
1002 			ret = trace_seq_printf(s, " ");
1003 			if (!ret)
1004 				return TRACE_TYPE_PARTIAL_LINE;
1005 		}
1006 
1007 	/* The comment */
1008 	ret = trace_seq_printf(s, "/* ");
1009 	if (!ret)
1010 		return TRACE_TYPE_PARTIAL_LINE;
1011 
1012 	switch (iter->ent->type) {
1013 	case TRACE_BPRINT:
1014 		ret = trace_print_bprintk_msg_only(iter);
1015 		if (ret != TRACE_TYPE_HANDLED)
1016 			return ret;
1017 		break;
1018 	case TRACE_PRINT:
1019 		ret = trace_print_printk_msg_only(iter);
1020 		if (ret != TRACE_TYPE_HANDLED)
1021 			return ret;
1022 		break;
1023 	default:
1024 		event = ftrace_find_event(ent->type);
1025 		if (!event)
1026 			return TRACE_TYPE_UNHANDLED;
1027 
1028 		ret = event->trace(iter, sym_flags);
1029 		if (ret != TRACE_TYPE_HANDLED)
1030 			return ret;
1031 	}
1032 
1033 	/* Strip ending newline */
1034 	if (s->buffer[s->len - 1] == '\n') {
1035 		s->buffer[s->len - 1] = '\0';
1036 		s->len--;
1037 	}
1038 
1039 	ret = trace_seq_printf(s, " */\n");
1040 	if (!ret)
1041 		return TRACE_TYPE_PARTIAL_LINE;
1042 
1043 	return TRACE_TYPE_HANDLED;
1044 }
1045 
1046 
1047 enum print_line_t
1048 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1049 {
1050 	struct ftrace_graph_ent_entry *field;
1051 	struct fgraph_data *data = iter->private;
1052 	struct trace_entry *entry = iter->ent;
1053 	struct trace_seq *s = &iter->seq;
1054 	int cpu = iter->cpu;
1055 	int ret;
1056 
1057 	if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1058 		per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1059 		return TRACE_TYPE_HANDLED;
1060 	}
1061 
1062 	/*
1063 	 * If the last output failed, there's a possibility we need
1064 	 * to print out the missing entry which would never go out.
1065 	 */
1066 	if (data && data->failed) {
1067 		field = &data->ent;
1068 		iter->cpu = data->cpu;
1069 		ret = print_graph_entry(field, s, iter, flags);
1070 		if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1071 			per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1072 			ret = TRACE_TYPE_NO_CONSUME;
1073 		}
1074 		iter->cpu = cpu;
1075 		return ret;
1076 	}
1077 
1078 	switch (entry->type) {
1079 	case TRACE_GRAPH_ENT: {
1080 		/*
1081 		 * print_graph_entry() may consume the current event,
1082 		 * thus @field may become invalid, so we need to save it.
1083 		 * sizeof(struct ftrace_graph_ent_entry) is very small,
1084 		 * it can be safely saved at the stack.
1085 		 */
1086 		struct ftrace_graph_ent_entry saved;
1087 		trace_assign_type(field, entry);
1088 		saved = *field;
1089 		return print_graph_entry(&saved, s, iter, flags);
1090 	}
1091 	case TRACE_GRAPH_RET: {
1092 		struct ftrace_graph_ret_entry *field;
1093 		trace_assign_type(field, entry);
1094 		return print_graph_return(&field->ret, s, entry, iter, flags);
1095 	}
1096 	case TRACE_STACK:
1097 	case TRACE_FN:
1098 		/* dont trace stack and functions as comments */
1099 		return TRACE_TYPE_UNHANDLED;
1100 
1101 	default:
1102 		return print_graph_comment(s, entry, iter, flags);
1103 	}
1104 
1105 	return TRACE_TYPE_HANDLED;
1106 }
1107 
1108 static enum print_line_t
1109 print_graph_function(struct trace_iterator *iter)
1110 {
1111 	return print_graph_function_flags(iter, tracer_flags.val);
1112 }
1113 
1114 static enum print_line_t
1115 print_graph_function_event(struct trace_iterator *iter, int flags)
1116 {
1117 	return print_graph_function(iter);
1118 }
1119 
1120 static void print_lat_header(struct seq_file *s, u32 flags)
1121 {
1122 	static const char spaces[] = "                "	/* 16 spaces */
1123 		"    "					/* 4 spaces */
1124 		"                 ";			/* 17 spaces */
1125 	int size = 0;
1126 
1127 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1128 		size += 16;
1129 	if (flags & TRACE_GRAPH_PRINT_CPU)
1130 		size += 4;
1131 	if (flags & TRACE_GRAPH_PRINT_PROC)
1132 		size += 17;
1133 
1134 	seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1135 	seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1136 	seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1137 	seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1138 	seq_printf(s, "#%.*s||| / _-=> lock-depth      \n", size, spaces);
1139 	seq_printf(s, "#%.*s|||| /                     \n", size, spaces);
1140 }
1141 
1142 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1143 {
1144 	int lat = trace_flags & TRACE_ITER_LATENCY_FMT;
1145 
1146 	if (lat)
1147 		print_lat_header(s, flags);
1148 
1149 	/* 1st line */
1150 	seq_printf(s, "#");
1151 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1152 		seq_printf(s, "     TIME       ");
1153 	if (flags & TRACE_GRAPH_PRINT_CPU)
1154 		seq_printf(s, " CPU");
1155 	if (flags & TRACE_GRAPH_PRINT_PROC)
1156 		seq_printf(s, "  TASK/PID       ");
1157 	if (lat)
1158 		seq_printf(s, "|||||");
1159 	if (flags & TRACE_GRAPH_PRINT_DURATION)
1160 		seq_printf(s, "  DURATION   ");
1161 	seq_printf(s, "               FUNCTION CALLS\n");
1162 
1163 	/* 2nd line */
1164 	seq_printf(s, "#");
1165 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1166 		seq_printf(s, "      |         ");
1167 	if (flags & TRACE_GRAPH_PRINT_CPU)
1168 		seq_printf(s, " |  ");
1169 	if (flags & TRACE_GRAPH_PRINT_PROC)
1170 		seq_printf(s, "   |    |        ");
1171 	if (lat)
1172 		seq_printf(s, "|||||");
1173 	if (flags & TRACE_GRAPH_PRINT_DURATION)
1174 		seq_printf(s, "   |   |      ");
1175 	seq_printf(s, "               |   |   |   |\n");
1176 }
1177 
1178 void print_graph_headers(struct seq_file *s)
1179 {
1180 	print_graph_headers_flags(s, tracer_flags.val);
1181 }
1182 
1183 void graph_trace_open(struct trace_iterator *iter)
1184 {
1185 	/* pid and depth on the last trace processed */
1186 	struct fgraph_data *data;
1187 	int cpu;
1188 
1189 	iter->private = NULL;
1190 
1191 	data = kzalloc(sizeof(*data), GFP_KERNEL);
1192 	if (!data)
1193 		goto out_err;
1194 
1195 	data->cpu_data = alloc_percpu(struct fgraph_cpu_data);
1196 	if (!data->cpu_data)
1197 		goto out_err_free;
1198 
1199 	for_each_possible_cpu(cpu) {
1200 		pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1201 		int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1202 		int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1203 		*pid = -1;
1204 		*depth = 0;
1205 		*ignore = 0;
1206 	}
1207 
1208 	iter->private = data;
1209 
1210 	return;
1211 
1212  out_err_free:
1213 	kfree(data);
1214  out_err:
1215 	pr_warning("function graph tracer: not enough memory\n");
1216 }
1217 
1218 void graph_trace_close(struct trace_iterator *iter)
1219 {
1220 	struct fgraph_data *data = iter->private;
1221 
1222 	if (data) {
1223 		free_percpu(data->cpu_data);
1224 		kfree(data);
1225 	}
1226 }
1227 
1228 static struct trace_event graph_trace_entry_event = {
1229 	.type		= TRACE_GRAPH_ENT,
1230 	.trace		= print_graph_function_event,
1231 };
1232 
1233 static struct trace_event graph_trace_ret_event = {
1234 	.type		= TRACE_GRAPH_RET,
1235 	.trace		= print_graph_function_event,
1236 };
1237 
1238 static struct tracer graph_trace __read_mostly = {
1239 	.name		= "function_graph",
1240 	.open		= graph_trace_open,
1241 	.pipe_open	= graph_trace_open,
1242 	.close		= graph_trace_close,
1243 	.pipe_close	= graph_trace_close,
1244 	.wait_pipe	= poll_wait_pipe,
1245 	.init		= graph_trace_init,
1246 	.reset		= graph_trace_reset,
1247 	.print_line	= print_graph_function,
1248 	.print_header	= print_graph_headers,
1249 	.flags		= &tracer_flags,
1250 #ifdef CONFIG_FTRACE_SELFTEST
1251 	.selftest	= trace_selftest_startup_function_graph,
1252 #endif
1253 };
1254 
1255 static __init int init_graph_trace(void)
1256 {
1257 	max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1258 
1259 	if (!register_ftrace_event(&graph_trace_entry_event)) {
1260 		pr_warning("Warning: could not register graph trace events\n");
1261 		return 1;
1262 	}
1263 
1264 	if (!register_ftrace_event(&graph_trace_ret_event)) {
1265 		pr_warning("Warning: could not register graph trace events\n");
1266 		return 1;
1267 	}
1268 
1269 	return register_tracer(&graph_trace);
1270 }
1271 
1272 device_initcall(init_graph_trace);
1273