xref: /linux/kernel/trace/trace_functions_graph.c (revision a234ca0faa65dcd5cc473915bd925130ebb7b74b)
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 			/*
511 			 * If the next event is not a return type, then
512 			 * we only care about what type it is. Otherwise we can
513 			 * safely copy the entire event.
514 			 */
515 			if (next->ent.type == TRACE_GRAPH_RET)
516 				data->ret = *next;
517 			else
518 				data->ret.ent.type = next->ent.type;
519 		}
520 	}
521 
522 	if (next->ent.type != TRACE_GRAPH_RET)
523 		return NULL;
524 
525 	if (curr->ent.pid != next->ent.pid ||
526 			curr->graph_ent.func != next->ret.func)
527 		return NULL;
528 
529 	/* this is a leaf, now advance the iterator */
530 	if (ring_iter)
531 		ring_buffer_read(ring_iter, NULL);
532 
533 	return next;
534 }
535 
536 /* Signal a overhead of time execution to the output */
537 static int
538 print_graph_overhead(unsigned long long duration, struct trace_seq *s,
539 		     u32 flags)
540 {
541 	/* If duration disappear, we don't need anything */
542 	if (!(flags & TRACE_GRAPH_PRINT_DURATION))
543 		return 1;
544 
545 	/* Non nested entry or return */
546 	if (duration == -1)
547 		return trace_seq_printf(s, "  ");
548 
549 	if (flags & TRACE_GRAPH_PRINT_OVERHEAD) {
550 		/* Duration exceeded 100 msecs */
551 		if (duration > 100000ULL)
552 			return trace_seq_printf(s, "! ");
553 
554 		/* Duration exceeded 10 msecs */
555 		if (duration > 10000ULL)
556 			return trace_seq_printf(s, "+ ");
557 	}
558 
559 	return trace_seq_printf(s, "  ");
560 }
561 
562 static int print_graph_abs_time(u64 t, struct trace_seq *s)
563 {
564 	unsigned long usecs_rem;
565 
566 	usecs_rem = do_div(t, NSEC_PER_SEC);
567 	usecs_rem /= 1000;
568 
569 	return trace_seq_printf(s, "%5lu.%06lu |  ",
570 			(unsigned long)t, usecs_rem);
571 }
572 
573 static enum print_line_t
574 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
575 		enum trace_type type, int cpu, pid_t pid, u32 flags)
576 {
577 	int ret;
578 	struct trace_seq *s = &iter->seq;
579 
580 	if (addr < (unsigned long)__irqentry_text_start ||
581 		addr >= (unsigned long)__irqentry_text_end)
582 		return TRACE_TYPE_UNHANDLED;
583 
584 	/* Absolute time */
585 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
586 		ret = print_graph_abs_time(iter->ts, s);
587 		if (!ret)
588 			return TRACE_TYPE_PARTIAL_LINE;
589 	}
590 
591 	/* Cpu */
592 	if (flags & TRACE_GRAPH_PRINT_CPU) {
593 		ret = print_graph_cpu(s, cpu);
594 		if (ret == TRACE_TYPE_PARTIAL_LINE)
595 			return TRACE_TYPE_PARTIAL_LINE;
596 	}
597 
598 	/* Proc */
599 	if (flags & TRACE_GRAPH_PRINT_PROC) {
600 		ret = print_graph_proc(s, pid);
601 		if (ret == TRACE_TYPE_PARTIAL_LINE)
602 			return TRACE_TYPE_PARTIAL_LINE;
603 		ret = trace_seq_printf(s, " | ");
604 		if (!ret)
605 			return TRACE_TYPE_PARTIAL_LINE;
606 	}
607 
608 	/* No overhead */
609 	ret = print_graph_overhead(-1, s, flags);
610 	if (!ret)
611 		return TRACE_TYPE_PARTIAL_LINE;
612 
613 	if (type == TRACE_GRAPH_ENT)
614 		ret = trace_seq_printf(s, "==========>");
615 	else
616 		ret = trace_seq_printf(s, "<==========");
617 
618 	if (!ret)
619 		return TRACE_TYPE_PARTIAL_LINE;
620 
621 	/* Don't close the duration column if haven't one */
622 	if (flags & TRACE_GRAPH_PRINT_DURATION)
623 		trace_seq_printf(s, " |");
624 	ret = trace_seq_printf(s, "\n");
625 
626 	if (!ret)
627 		return TRACE_TYPE_PARTIAL_LINE;
628 	return TRACE_TYPE_HANDLED;
629 }
630 
631 enum print_line_t
632 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
633 {
634 	unsigned long nsecs_rem = do_div(duration, 1000);
635 	/* log10(ULONG_MAX) + '\0' */
636 	char msecs_str[21];
637 	char nsecs_str[5];
638 	int ret, len;
639 	int i;
640 
641 	sprintf(msecs_str, "%lu", (unsigned long) duration);
642 
643 	/* Print msecs */
644 	ret = trace_seq_printf(s, "%s", msecs_str);
645 	if (!ret)
646 		return TRACE_TYPE_PARTIAL_LINE;
647 
648 	len = strlen(msecs_str);
649 
650 	/* Print nsecs (we don't want to exceed 7 numbers) */
651 	if (len < 7) {
652 		snprintf(nsecs_str, min(sizeof(nsecs_str), 8UL - len), "%03lu",
653 			 nsecs_rem);
654 		ret = trace_seq_printf(s, ".%s", nsecs_str);
655 		if (!ret)
656 			return TRACE_TYPE_PARTIAL_LINE;
657 		len += strlen(nsecs_str);
658 	}
659 
660 	ret = trace_seq_printf(s, " us ");
661 	if (!ret)
662 		return TRACE_TYPE_PARTIAL_LINE;
663 
664 	/* Print remaining spaces to fit the row's width */
665 	for (i = len; i < 7; i++) {
666 		ret = trace_seq_printf(s, " ");
667 		if (!ret)
668 			return TRACE_TYPE_PARTIAL_LINE;
669 	}
670 	return TRACE_TYPE_HANDLED;
671 }
672 
673 static enum print_line_t
674 print_graph_duration(unsigned long long duration, struct trace_seq *s)
675 {
676 	int ret;
677 
678 	ret = trace_print_graph_duration(duration, s);
679 	if (ret != TRACE_TYPE_HANDLED)
680 		return ret;
681 
682 	ret = trace_seq_printf(s, "|  ");
683 	if (!ret)
684 		return TRACE_TYPE_PARTIAL_LINE;
685 
686 	return TRACE_TYPE_HANDLED;
687 }
688 
689 /* Case of a leaf function on its call entry */
690 static enum print_line_t
691 print_graph_entry_leaf(struct trace_iterator *iter,
692 		struct ftrace_graph_ent_entry *entry,
693 		struct ftrace_graph_ret_entry *ret_entry,
694 		struct trace_seq *s, u32 flags)
695 {
696 	struct fgraph_data *data = iter->private;
697 	struct ftrace_graph_ret *graph_ret;
698 	struct ftrace_graph_ent *call;
699 	unsigned long long duration;
700 	int ret;
701 	int i;
702 
703 	graph_ret = &ret_entry->ret;
704 	call = &entry->graph_ent;
705 	duration = graph_ret->rettime - graph_ret->calltime;
706 
707 	if (data) {
708 		struct fgraph_cpu_data *cpu_data;
709 		int cpu = iter->cpu;
710 
711 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
712 
713 		/*
714 		 * Comments display at + 1 to depth. Since
715 		 * this is a leaf function, keep the comments
716 		 * equal to this depth.
717 		 */
718 		cpu_data->depth = call->depth - 1;
719 
720 		/* No need to keep this function around for this depth */
721 		if (call->depth < FTRACE_RETFUNC_DEPTH)
722 			cpu_data->enter_funcs[call->depth] = 0;
723 	}
724 
725 	/* Overhead */
726 	ret = print_graph_overhead(duration, s, flags);
727 	if (!ret)
728 		return TRACE_TYPE_PARTIAL_LINE;
729 
730 	/* Duration */
731 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
732 		ret = print_graph_duration(duration, s);
733 		if (ret == TRACE_TYPE_PARTIAL_LINE)
734 			return TRACE_TYPE_PARTIAL_LINE;
735 	}
736 
737 	/* Function */
738 	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
739 		ret = trace_seq_printf(s, " ");
740 		if (!ret)
741 			return TRACE_TYPE_PARTIAL_LINE;
742 	}
743 
744 	ret = trace_seq_printf(s, "%ps();\n", (void *)call->func);
745 	if (!ret)
746 		return TRACE_TYPE_PARTIAL_LINE;
747 
748 	return TRACE_TYPE_HANDLED;
749 }
750 
751 static enum print_line_t
752 print_graph_entry_nested(struct trace_iterator *iter,
753 			 struct ftrace_graph_ent_entry *entry,
754 			 struct trace_seq *s, int cpu, u32 flags)
755 {
756 	struct ftrace_graph_ent *call = &entry->graph_ent;
757 	struct fgraph_data *data = iter->private;
758 	int ret;
759 	int i;
760 
761 	if (data) {
762 		struct fgraph_cpu_data *cpu_data;
763 		int cpu = iter->cpu;
764 
765 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
766 		cpu_data->depth = call->depth;
767 
768 		/* Save this function pointer to see if the exit matches */
769 		if (call->depth < FTRACE_RETFUNC_DEPTH)
770 			cpu_data->enter_funcs[call->depth] = call->func;
771 	}
772 
773 	/* No overhead */
774 	ret = print_graph_overhead(-1, s, flags);
775 	if (!ret)
776 		return TRACE_TYPE_PARTIAL_LINE;
777 
778 	/* No time */
779 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
780 		ret = trace_seq_printf(s, "            |  ");
781 		if (!ret)
782 			return TRACE_TYPE_PARTIAL_LINE;
783 	}
784 
785 	/* Function */
786 	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
787 		ret = trace_seq_printf(s, " ");
788 		if (!ret)
789 			return TRACE_TYPE_PARTIAL_LINE;
790 	}
791 
792 	ret = trace_seq_printf(s, "%ps() {\n", (void *)call->func);
793 	if (!ret)
794 		return TRACE_TYPE_PARTIAL_LINE;
795 
796 	/*
797 	 * we already consumed the current entry to check the next one
798 	 * and see if this is a leaf.
799 	 */
800 	return TRACE_TYPE_NO_CONSUME;
801 }
802 
803 static enum print_line_t
804 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
805 		     int type, unsigned long addr, u32 flags)
806 {
807 	struct fgraph_data *data = iter->private;
808 	struct trace_entry *ent = iter->ent;
809 	int cpu = iter->cpu;
810 	int ret;
811 
812 	/* Pid */
813 	if (verif_pid(s, ent->pid, cpu, data) == TRACE_TYPE_PARTIAL_LINE)
814 		return TRACE_TYPE_PARTIAL_LINE;
815 
816 	if (type) {
817 		/* Interrupt */
818 		ret = print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
819 		if (ret == TRACE_TYPE_PARTIAL_LINE)
820 			return TRACE_TYPE_PARTIAL_LINE;
821 	}
822 
823 	/* Absolute time */
824 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
825 		ret = print_graph_abs_time(iter->ts, s);
826 		if (!ret)
827 			return TRACE_TYPE_PARTIAL_LINE;
828 	}
829 
830 	/* Cpu */
831 	if (flags & TRACE_GRAPH_PRINT_CPU) {
832 		ret = print_graph_cpu(s, cpu);
833 		if (ret == TRACE_TYPE_PARTIAL_LINE)
834 			return TRACE_TYPE_PARTIAL_LINE;
835 	}
836 
837 	/* Proc */
838 	if (flags & TRACE_GRAPH_PRINT_PROC) {
839 		ret = print_graph_proc(s, ent->pid);
840 		if (ret == TRACE_TYPE_PARTIAL_LINE)
841 			return TRACE_TYPE_PARTIAL_LINE;
842 
843 		ret = trace_seq_printf(s, " | ");
844 		if (!ret)
845 			return TRACE_TYPE_PARTIAL_LINE;
846 	}
847 
848 	/* Latency format */
849 	if (trace_flags & TRACE_ITER_LATENCY_FMT) {
850 		ret = print_graph_lat_fmt(s, ent);
851 		if (ret == TRACE_TYPE_PARTIAL_LINE)
852 			return TRACE_TYPE_PARTIAL_LINE;
853 	}
854 
855 	return 0;
856 }
857 
858 static enum print_line_t
859 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
860 			struct trace_iterator *iter, u32 flags)
861 {
862 	struct fgraph_data *data = iter->private;
863 	struct ftrace_graph_ent *call = &field->graph_ent;
864 	struct ftrace_graph_ret_entry *leaf_ret;
865 	static enum print_line_t ret;
866 	int cpu = iter->cpu;
867 
868 	if (print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags))
869 		return TRACE_TYPE_PARTIAL_LINE;
870 
871 	leaf_ret = get_return_for_leaf(iter, field);
872 	if (leaf_ret)
873 		ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
874 	else
875 		ret = print_graph_entry_nested(iter, field, s, cpu, flags);
876 
877 	if (data) {
878 		/*
879 		 * If we failed to write our output, then we need to make
880 		 * note of it. Because we already consumed our entry.
881 		 */
882 		if (s->full) {
883 			data->failed = 1;
884 			data->cpu = cpu;
885 		} else
886 			data->failed = 0;
887 	}
888 
889 	return ret;
890 }
891 
892 static enum print_line_t
893 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
894 		   struct trace_entry *ent, struct trace_iterator *iter,
895 		   u32 flags)
896 {
897 	unsigned long long duration = trace->rettime - trace->calltime;
898 	struct fgraph_data *data = iter->private;
899 	pid_t pid = ent->pid;
900 	int cpu = iter->cpu;
901 	int func_match = 1;
902 	int ret;
903 	int i;
904 
905 	if (data) {
906 		struct fgraph_cpu_data *cpu_data;
907 		int cpu = iter->cpu;
908 
909 		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
910 
911 		/*
912 		 * Comments display at + 1 to depth. This is the
913 		 * return from a function, we now want the comments
914 		 * to display at the same level of the bracket.
915 		 */
916 		cpu_data->depth = trace->depth - 1;
917 
918 		if (trace->depth < FTRACE_RETFUNC_DEPTH) {
919 			if (cpu_data->enter_funcs[trace->depth] != trace->func)
920 				func_match = 0;
921 			cpu_data->enter_funcs[trace->depth] = 0;
922 		}
923 	}
924 
925 	if (print_graph_prologue(iter, s, 0, 0, flags))
926 		return TRACE_TYPE_PARTIAL_LINE;
927 
928 	/* Overhead */
929 	ret = print_graph_overhead(duration, s, flags);
930 	if (!ret)
931 		return TRACE_TYPE_PARTIAL_LINE;
932 
933 	/* Duration */
934 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
935 		ret = print_graph_duration(duration, s);
936 		if (ret == TRACE_TYPE_PARTIAL_LINE)
937 			return TRACE_TYPE_PARTIAL_LINE;
938 	}
939 
940 	/* Closing brace */
941 	for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++) {
942 		ret = trace_seq_printf(s, " ");
943 		if (!ret)
944 			return TRACE_TYPE_PARTIAL_LINE;
945 	}
946 
947 	/*
948 	 * If the return function does not have a matching entry,
949 	 * then the entry was lost. Instead of just printing
950 	 * the '}' and letting the user guess what function this
951 	 * belongs to, write out the function name.
952 	 */
953 	if (func_match) {
954 		ret = trace_seq_printf(s, "}\n");
955 		if (!ret)
956 			return TRACE_TYPE_PARTIAL_LINE;
957 	} else {
958 		ret = trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
959 		if (!ret)
960 			return TRACE_TYPE_PARTIAL_LINE;
961 	}
962 
963 	/* Overrun */
964 	if (flags & TRACE_GRAPH_PRINT_OVERRUN) {
965 		ret = trace_seq_printf(s, " (Overruns: %lu)\n",
966 					trace->overrun);
967 		if (!ret)
968 			return TRACE_TYPE_PARTIAL_LINE;
969 	}
970 
971 	ret = print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
972 			      cpu, pid, flags);
973 	if (ret == TRACE_TYPE_PARTIAL_LINE)
974 		return TRACE_TYPE_PARTIAL_LINE;
975 
976 	return TRACE_TYPE_HANDLED;
977 }
978 
979 static enum print_line_t
980 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
981 		    struct trace_iterator *iter, u32 flags)
982 {
983 	unsigned long sym_flags = (trace_flags & TRACE_ITER_SYM_MASK);
984 	struct fgraph_data *data = iter->private;
985 	struct trace_event *event;
986 	int depth = 0;
987 	int ret;
988 	int i;
989 
990 	if (data)
991 		depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
992 
993 	if (print_graph_prologue(iter, s, 0, 0, flags))
994 		return TRACE_TYPE_PARTIAL_LINE;
995 
996 	/* No overhead */
997 	ret = print_graph_overhead(-1, s, flags);
998 	if (!ret)
999 		return TRACE_TYPE_PARTIAL_LINE;
1000 
1001 	/* No time */
1002 	if (flags & TRACE_GRAPH_PRINT_DURATION) {
1003 		ret = trace_seq_printf(s, "            |  ");
1004 		if (!ret)
1005 			return TRACE_TYPE_PARTIAL_LINE;
1006 	}
1007 
1008 	/* Indentation */
1009 	if (depth > 0)
1010 		for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++) {
1011 			ret = trace_seq_printf(s, " ");
1012 			if (!ret)
1013 				return TRACE_TYPE_PARTIAL_LINE;
1014 		}
1015 
1016 	/* The comment */
1017 	ret = trace_seq_printf(s, "/* ");
1018 	if (!ret)
1019 		return TRACE_TYPE_PARTIAL_LINE;
1020 
1021 	switch (iter->ent->type) {
1022 	case TRACE_BPRINT:
1023 		ret = trace_print_bprintk_msg_only(iter);
1024 		if (ret != TRACE_TYPE_HANDLED)
1025 			return ret;
1026 		break;
1027 	case TRACE_PRINT:
1028 		ret = trace_print_printk_msg_only(iter);
1029 		if (ret != TRACE_TYPE_HANDLED)
1030 			return ret;
1031 		break;
1032 	default:
1033 		event = ftrace_find_event(ent->type);
1034 		if (!event)
1035 			return TRACE_TYPE_UNHANDLED;
1036 
1037 		ret = event->funcs->trace(iter, sym_flags, event);
1038 		if (ret != TRACE_TYPE_HANDLED)
1039 			return ret;
1040 	}
1041 
1042 	/* Strip ending newline */
1043 	if (s->buffer[s->len - 1] == '\n') {
1044 		s->buffer[s->len - 1] = '\0';
1045 		s->len--;
1046 	}
1047 
1048 	ret = trace_seq_printf(s, " */\n");
1049 	if (!ret)
1050 		return TRACE_TYPE_PARTIAL_LINE;
1051 
1052 	return TRACE_TYPE_HANDLED;
1053 }
1054 
1055 
1056 enum print_line_t
1057 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1058 {
1059 	struct ftrace_graph_ent_entry *field;
1060 	struct fgraph_data *data = iter->private;
1061 	struct trace_entry *entry = iter->ent;
1062 	struct trace_seq *s = &iter->seq;
1063 	int cpu = iter->cpu;
1064 	int ret;
1065 
1066 	if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1067 		per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1068 		return TRACE_TYPE_HANDLED;
1069 	}
1070 
1071 	/*
1072 	 * If the last output failed, there's a possibility we need
1073 	 * to print out the missing entry which would never go out.
1074 	 */
1075 	if (data && data->failed) {
1076 		field = &data->ent;
1077 		iter->cpu = data->cpu;
1078 		ret = print_graph_entry(field, s, iter, flags);
1079 		if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1080 			per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1081 			ret = TRACE_TYPE_NO_CONSUME;
1082 		}
1083 		iter->cpu = cpu;
1084 		return ret;
1085 	}
1086 
1087 	switch (entry->type) {
1088 	case TRACE_GRAPH_ENT: {
1089 		/*
1090 		 * print_graph_entry() may consume the current event,
1091 		 * thus @field may become invalid, so we need to save it.
1092 		 * sizeof(struct ftrace_graph_ent_entry) is very small,
1093 		 * it can be safely saved at the stack.
1094 		 */
1095 		struct ftrace_graph_ent_entry saved;
1096 		trace_assign_type(field, entry);
1097 		saved = *field;
1098 		return print_graph_entry(&saved, s, iter, flags);
1099 	}
1100 	case TRACE_GRAPH_RET: {
1101 		struct ftrace_graph_ret_entry *field;
1102 		trace_assign_type(field, entry);
1103 		return print_graph_return(&field->ret, s, entry, iter, flags);
1104 	}
1105 	case TRACE_STACK:
1106 	case TRACE_FN:
1107 		/* dont trace stack and functions as comments */
1108 		return TRACE_TYPE_UNHANDLED;
1109 
1110 	default:
1111 		return print_graph_comment(s, entry, iter, flags);
1112 	}
1113 
1114 	return TRACE_TYPE_HANDLED;
1115 }
1116 
1117 static enum print_line_t
1118 print_graph_function(struct trace_iterator *iter)
1119 {
1120 	return print_graph_function_flags(iter, tracer_flags.val);
1121 }
1122 
1123 static enum print_line_t
1124 print_graph_function_event(struct trace_iterator *iter, int flags,
1125 			   struct trace_event *event)
1126 {
1127 	return print_graph_function(iter);
1128 }
1129 
1130 static void print_lat_header(struct seq_file *s, u32 flags)
1131 {
1132 	static const char spaces[] = "                "	/* 16 spaces */
1133 		"    "					/* 4 spaces */
1134 		"                 ";			/* 17 spaces */
1135 	int size = 0;
1136 
1137 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1138 		size += 16;
1139 	if (flags & TRACE_GRAPH_PRINT_CPU)
1140 		size += 4;
1141 	if (flags & TRACE_GRAPH_PRINT_PROC)
1142 		size += 17;
1143 
1144 	seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1145 	seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1146 	seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1147 	seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1148 	seq_printf(s, "#%.*s||| / _-=> lock-depth      \n", size, spaces);
1149 	seq_printf(s, "#%.*s|||| /                     \n", size, spaces);
1150 }
1151 
1152 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1153 {
1154 	int lat = trace_flags & TRACE_ITER_LATENCY_FMT;
1155 
1156 	if (lat)
1157 		print_lat_header(s, flags);
1158 
1159 	/* 1st line */
1160 	seq_printf(s, "#");
1161 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1162 		seq_printf(s, "     TIME       ");
1163 	if (flags & TRACE_GRAPH_PRINT_CPU)
1164 		seq_printf(s, " CPU");
1165 	if (flags & TRACE_GRAPH_PRINT_PROC)
1166 		seq_printf(s, "  TASK/PID       ");
1167 	if (lat)
1168 		seq_printf(s, "|||||");
1169 	if (flags & TRACE_GRAPH_PRINT_DURATION)
1170 		seq_printf(s, "  DURATION   ");
1171 	seq_printf(s, "               FUNCTION CALLS\n");
1172 
1173 	/* 2nd line */
1174 	seq_printf(s, "#");
1175 	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1176 		seq_printf(s, "      |         ");
1177 	if (flags & TRACE_GRAPH_PRINT_CPU)
1178 		seq_printf(s, " |  ");
1179 	if (flags & TRACE_GRAPH_PRINT_PROC)
1180 		seq_printf(s, "   |    |        ");
1181 	if (lat)
1182 		seq_printf(s, "|||||");
1183 	if (flags & TRACE_GRAPH_PRINT_DURATION)
1184 		seq_printf(s, "   |   |      ");
1185 	seq_printf(s, "               |   |   |   |\n");
1186 }
1187 
1188 void print_graph_headers(struct seq_file *s)
1189 {
1190 	print_graph_headers_flags(s, tracer_flags.val);
1191 }
1192 
1193 void graph_trace_open(struct trace_iterator *iter)
1194 {
1195 	/* pid and depth on the last trace processed */
1196 	struct fgraph_data *data;
1197 	int cpu;
1198 
1199 	iter->private = NULL;
1200 
1201 	data = kzalloc(sizeof(*data), GFP_KERNEL);
1202 	if (!data)
1203 		goto out_err;
1204 
1205 	data->cpu_data = alloc_percpu(struct fgraph_cpu_data);
1206 	if (!data->cpu_data)
1207 		goto out_err_free;
1208 
1209 	for_each_possible_cpu(cpu) {
1210 		pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1211 		int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1212 		int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1213 		*pid = -1;
1214 		*depth = 0;
1215 		*ignore = 0;
1216 	}
1217 
1218 	iter->private = data;
1219 
1220 	return;
1221 
1222  out_err_free:
1223 	kfree(data);
1224  out_err:
1225 	pr_warning("function graph tracer: not enough memory\n");
1226 }
1227 
1228 void graph_trace_close(struct trace_iterator *iter)
1229 {
1230 	struct fgraph_data *data = iter->private;
1231 
1232 	if (data) {
1233 		free_percpu(data->cpu_data);
1234 		kfree(data);
1235 	}
1236 }
1237 
1238 static struct trace_event_functions graph_functions = {
1239 	.trace		= print_graph_function_event,
1240 };
1241 
1242 static struct trace_event graph_trace_entry_event = {
1243 	.type		= TRACE_GRAPH_ENT,
1244 	.funcs		= &graph_functions,
1245 };
1246 
1247 static struct trace_event graph_trace_ret_event = {
1248 	.type		= TRACE_GRAPH_RET,
1249 	.funcs		= &graph_functions
1250 };
1251 
1252 static struct tracer graph_trace __read_mostly = {
1253 	.name		= "function_graph",
1254 	.open		= graph_trace_open,
1255 	.pipe_open	= graph_trace_open,
1256 	.close		= graph_trace_close,
1257 	.pipe_close	= graph_trace_close,
1258 	.wait_pipe	= poll_wait_pipe,
1259 	.init		= graph_trace_init,
1260 	.reset		= graph_trace_reset,
1261 	.print_line	= print_graph_function,
1262 	.print_header	= print_graph_headers,
1263 	.flags		= &tracer_flags,
1264 #ifdef CONFIG_FTRACE_SELFTEST
1265 	.selftest	= trace_selftest_startup_function_graph,
1266 #endif
1267 };
1268 
1269 static __init int init_graph_trace(void)
1270 {
1271 	max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1272 
1273 	if (!register_ftrace_event(&graph_trace_entry_event)) {
1274 		pr_warning("Warning: could not register graph trace events\n");
1275 		return 1;
1276 	}
1277 
1278 	if (!register_ftrace_event(&graph_trace_ret_event)) {
1279 		pr_warning("Warning: could not register graph trace events\n");
1280 		return 1;
1281 	}
1282 
1283 	return register_tracer(&graph_trace);
1284 }
1285 
1286 device_initcall(init_graph_trace);
1287