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/uaccess.h>
10#include <linux/ftrace.h>
11#include <linux/slab.h>
12#include <linux/fs.h>
13
14#include "trace.h"
15#include "trace_output.h"
16
17static bool kill_ftrace_graph;
18
19/**
20 * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
21 *
22 * ftrace_graph_stop() is called when a severe error is detected in
23 * the function graph tracing. This function is called by the critical
24 * paths of function graph to keep those paths from doing any more harm.
25 */
26bool ftrace_graph_is_dead(void)
27{
28	return kill_ftrace_graph;
29}
30
31/**
32 * ftrace_graph_stop - set to permanently disable function graph tracincg
33 *
34 * In case of an error int function graph tracing, this is called
35 * to try to keep function graph tracing from causing any more harm.
36 * Usually this is pretty severe and this is called to try to at least
37 * get a warning out to the user.
38 */
39void ftrace_graph_stop(void)
40{
41	kill_ftrace_graph = true;
42}
43
44/* When set, irq functions will be ignored */
45static int ftrace_graph_skip_irqs;
46
47struct fgraph_cpu_data {
48	pid_t		last_pid;
49	int		depth;
50	int		depth_irq;
51	int		ignore;
52	unsigned long	enter_funcs[FTRACE_RETFUNC_DEPTH];
53};
54
55struct fgraph_data {
56	struct fgraph_cpu_data __percpu *cpu_data;
57
58	/* Place to preserve last processed entry. */
59	struct ftrace_graph_ent_entry	ent;
60	struct ftrace_graph_ret_entry	ret;
61	int				failed;
62	int				cpu;
63};
64
65#define TRACE_GRAPH_INDENT	2
66
67static unsigned int max_depth;
68
69static struct tracer_opt trace_opts[] = {
70	/* Display overruns? (for self-debug purpose) */
71	{ TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
72	/* Display CPU ? */
73	{ TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
74	/* Display Overhead ? */
75	{ TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
76	/* Display proc name/pid */
77	{ TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
78	/* Display duration of execution */
79	{ TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
80	/* Display absolute time of an entry */
81	{ TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
82	/* Display interrupts */
83	{ TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
84	/* Display function name after trailing } */
85	{ TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
86	{ } /* Empty entry */
87};
88
89static struct tracer_flags tracer_flags = {
90	/* Don't display overruns, proc, or tail by default */
91	.val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
92	       TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS,
93	.opts = trace_opts
94};
95
96static struct trace_array *graph_array;
97
98/*
99 * DURATION column is being also used to display IRQ signs,
100 * following values are used by print_graph_irq and others
101 * to fill in space into DURATION column.
102 */
103enum {
104	FLAGS_FILL_FULL  = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
105	FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
106	FLAGS_FILL_END   = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
107};
108
109static void
110print_graph_duration(unsigned long long duration, struct trace_seq *s,
111		     u32 flags);
112
113/* Add a function return address to the trace stack on thread info.*/
114int
115ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
116			 unsigned long frame_pointer)
117{
118	unsigned long long calltime;
119	int index;
120
121	if (unlikely(ftrace_graph_is_dead()))
122		return -EBUSY;
123
124	if (!current->ret_stack)
125		return -EBUSY;
126
127	/*
128	 * We must make sure the ret_stack is tested before we read
129	 * anything else.
130	 */
131	smp_rmb();
132
133	/* The return trace stack is full */
134	if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
135		atomic_inc(&current->trace_overrun);
136		return -EBUSY;
137	}
138
139	/*
140	 * The curr_ret_stack is an index to ftrace return stack of
141	 * current task.  Its value should be in [0, FTRACE_RETFUNC_
142	 * DEPTH) when the function graph tracer is used.  To support
143	 * filtering out specific functions, it makes the index
144	 * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
145	 * so when it sees a negative index the ftrace will ignore
146	 * the record.  And the index gets recovered when returning
147	 * from the filtered function by adding the FTRACE_NOTRACE_
148	 * DEPTH and then it'll continue to record functions normally.
149	 *
150	 * The curr_ret_stack is initialized to -1 and get increased
151	 * in this function.  So it can be less than -1 only if it was
152	 * filtered out via ftrace_graph_notrace_addr() which can be
153	 * set from set_graph_notrace file in tracefs by user.
154	 */
155	if (current->curr_ret_stack < -1)
156		return -EBUSY;
157
158	calltime = trace_clock_local();
159
160	index = ++current->curr_ret_stack;
161	if (ftrace_graph_notrace_addr(func))
162		current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
163	barrier();
164	current->ret_stack[index].ret = ret;
165	current->ret_stack[index].func = func;
166	current->ret_stack[index].calltime = calltime;
167	current->ret_stack[index].subtime = 0;
168	current->ret_stack[index].fp = frame_pointer;
169	*depth = current->curr_ret_stack;
170
171	return 0;
172}
173
174/* Retrieve a function return address to the trace stack on thread info.*/
175static void
176ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
177			unsigned long frame_pointer)
178{
179	int index;
180
181	index = current->curr_ret_stack;
182
183	/*
184	 * A negative index here means that it's just returned from a
185	 * notrace'd function.  Recover index to get an original
186	 * return address.  See ftrace_push_return_trace().
187	 *
188	 * TODO: Need to check whether the stack gets corrupted.
189	 */
190	if (index < 0)
191		index += FTRACE_NOTRACE_DEPTH;
192
193	if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
194		ftrace_graph_stop();
195		WARN_ON(1);
196		/* Might as well panic, otherwise we have no where to go */
197		*ret = (unsigned long)panic;
198		return;
199	}
200
201#if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
202	/*
203	 * The arch may choose to record the frame pointer used
204	 * and check it here to make sure that it is what we expect it
205	 * to be. If gcc does not set the place holder of the return
206	 * address in the frame pointer, and does a copy instead, then
207	 * the function graph trace will fail. This test detects this
208	 * case.
209	 *
210	 * Currently, x86_32 with optimize for size (-Os) makes the latest
211	 * gcc do the above.
212	 *
213	 * Note, -mfentry does not use frame pointers, and this test
214	 *  is not needed if CC_USING_FENTRY is set.
215	 */
216	if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
217		ftrace_graph_stop();
218		WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
219		     "  from func %ps return to %lx\n",
220		     current->ret_stack[index].fp,
221		     frame_pointer,
222		     (void *)current->ret_stack[index].func,
223		     current->ret_stack[index].ret);
224		*ret = (unsigned long)panic;
225		return;
226	}
227#endif
228
229	*ret = current->ret_stack[index].ret;
230	trace->func = current->ret_stack[index].func;
231	trace->calltime = current->ret_stack[index].calltime;
232	trace->overrun = atomic_read(&current->trace_overrun);
233	trace->depth = index;
234}
235
236/*
237 * Send the trace to the ring-buffer.
238 * @return the original return address.
239 */
240unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
241{
242	struct ftrace_graph_ret trace;
243	unsigned long ret;
244
245	ftrace_pop_return_trace(&trace, &ret, frame_pointer);
246	trace.rettime = trace_clock_local();
247	barrier();
248	current->curr_ret_stack--;
249	/*
250	 * The curr_ret_stack can be less than -1 only if it was
251	 * filtered out and it's about to return from the function.
252	 * Recover the index and continue to trace normal functions.
253	 */
254	if (current->curr_ret_stack < -1) {
255		current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
256		return ret;
257	}
258
259	/*
260	 * The trace should run after decrementing the ret counter
261	 * in case an interrupt were to come in. We don't want to
262	 * lose the interrupt if max_depth is set.
263	 */
264	ftrace_graph_return(&trace);
265
266	if (unlikely(!ret)) {
267		ftrace_graph_stop();
268		WARN_ON(1);
269		/* Might as well panic. What else to do? */
270		ret = (unsigned long)panic;
271	}
272
273	return ret;
274}
275
276int __trace_graph_entry(struct trace_array *tr,
277				struct ftrace_graph_ent *trace,
278				unsigned long flags,
279				int pc)
280{
281	struct ftrace_event_call *call = &event_funcgraph_entry;
282	struct ring_buffer_event *event;
283	struct ring_buffer *buffer = tr->trace_buffer.buffer;
284	struct ftrace_graph_ent_entry *entry;
285
286	if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
287		return 0;
288
289	event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
290					  sizeof(*entry), flags, pc);
291	if (!event)
292		return 0;
293	entry	= ring_buffer_event_data(event);
294	entry->graph_ent			= *trace;
295	if (!call_filter_check_discard(call, entry, buffer, event))
296		__buffer_unlock_commit(buffer, event);
297
298	return 1;
299}
300
301static inline int ftrace_graph_ignore_irqs(void)
302{
303	if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
304		return 0;
305
306	return in_irq();
307}
308
309int trace_graph_entry(struct ftrace_graph_ent *trace)
310{
311	struct trace_array *tr = graph_array;
312	struct trace_array_cpu *data;
313	unsigned long flags;
314	long disabled;
315	int ret;
316	int cpu;
317	int pc;
318
319	if (!ftrace_trace_task(current))
320		return 0;
321
322	/* trace it when it is-nested-in or is a function enabled. */
323	if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
324	     ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
325	    (max_depth && trace->depth >= max_depth))
326		return 0;
327
328	/*
329	 * Do not trace a function if it's filtered by set_graph_notrace.
330	 * Make the index of ret stack negative to indicate that it should
331	 * ignore further functions.  But it needs its own ret stack entry
332	 * to recover the original index in order to continue tracing after
333	 * returning from the function.
334	 */
335	if (ftrace_graph_notrace_addr(trace->func))
336		return 1;
337
338	local_irq_save(flags);
339	cpu = raw_smp_processor_id();
340	data = per_cpu_ptr(tr->trace_buffer.data, cpu);
341	disabled = atomic_inc_return(&data->disabled);
342	if (likely(disabled == 1)) {
343		pc = preempt_count();
344		ret = __trace_graph_entry(tr, trace, flags, pc);
345	} else {
346		ret = 0;
347	}
348
349	atomic_dec(&data->disabled);
350	local_irq_restore(flags);
351
352	return ret;
353}
354
355static int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
356{
357	if (tracing_thresh)
358		return 1;
359	else
360		return trace_graph_entry(trace);
361}
362
363static void
364__trace_graph_function(struct trace_array *tr,
365		unsigned long ip, unsigned long flags, int pc)
366{
367	u64 time = trace_clock_local();
368	struct ftrace_graph_ent ent = {
369		.func  = ip,
370		.depth = 0,
371	};
372	struct ftrace_graph_ret ret = {
373		.func     = ip,
374		.depth    = 0,
375		.calltime = time,
376		.rettime  = time,
377	};
378
379	__trace_graph_entry(tr, &ent, flags, pc);
380	__trace_graph_return(tr, &ret, flags, pc);
381}
382
383void
384trace_graph_function(struct trace_array *tr,
385		unsigned long ip, unsigned long parent_ip,
386		unsigned long flags, int pc)
387{
388	__trace_graph_function(tr, ip, flags, pc);
389}
390
391void __trace_graph_return(struct trace_array *tr,
392				struct ftrace_graph_ret *trace,
393				unsigned long flags,
394				int pc)
395{
396	struct ftrace_event_call *call = &event_funcgraph_exit;
397	struct ring_buffer_event *event;
398	struct ring_buffer *buffer = tr->trace_buffer.buffer;
399	struct ftrace_graph_ret_entry *entry;
400
401	if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
402		return;
403
404	event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
405					  sizeof(*entry), flags, pc);
406	if (!event)
407		return;
408	entry	= ring_buffer_event_data(event);
409	entry->ret				= *trace;
410	if (!call_filter_check_discard(call, entry, buffer, event))
411		__buffer_unlock_commit(buffer, event);
412}
413
414void trace_graph_return(struct ftrace_graph_ret *trace)
415{
416	struct trace_array *tr = graph_array;
417	struct trace_array_cpu *data;
418	unsigned long flags;
419	long disabled;
420	int cpu;
421	int pc;
422
423	local_irq_save(flags);
424	cpu = raw_smp_processor_id();
425	data = per_cpu_ptr(tr->trace_buffer.data, cpu);
426	disabled = atomic_inc_return(&data->disabled);
427	if (likely(disabled == 1)) {
428		pc = preempt_count();
429		__trace_graph_return(tr, trace, flags, pc);
430	}
431	atomic_dec(&data->disabled);
432	local_irq_restore(flags);
433}
434
435void set_graph_array(struct trace_array *tr)
436{
437	graph_array = tr;
438
439	/* Make graph_array visible before we start tracing */
440
441	smp_mb();
442}
443
444static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
445{
446	if (tracing_thresh &&
447	    (trace->rettime - trace->calltime < tracing_thresh))
448		return;
449	else
450		trace_graph_return(trace);
451}
452
453static int graph_trace_init(struct trace_array *tr)
454{
455	int ret;
456
457	set_graph_array(tr);
458	if (tracing_thresh)
459		ret = register_ftrace_graph(&trace_graph_thresh_return,
460					    &trace_graph_thresh_entry);
461	else
462		ret = register_ftrace_graph(&trace_graph_return,
463					    &trace_graph_entry);
464	if (ret)
465		return ret;
466	tracing_start_cmdline_record();
467
468	return 0;
469}
470
471static void graph_trace_reset(struct trace_array *tr)
472{
473	tracing_stop_cmdline_record();
474	unregister_ftrace_graph();
475}
476
477static int graph_trace_update_thresh(struct trace_array *tr)
478{
479	graph_trace_reset(tr);
480	return graph_trace_init(tr);
481}
482
483static int max_bytes_for_cpu;
484
485static void print_graph_cpu(struct trace_seq *s, int cpu)
486{
487	/*
488	 * Start with a space character - to make it stand out
489	 * to the right a bit when trace output is pasted into
490	 * email:
491	 */
492	trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
493}
494
495#define TRACE_GRAPH_PROCINFO_LENGTH	14
496
497static void print_graph_proc(struct trace_seq *s, pid_t pid)
498{
499	char comm[TASK_COMM_LEN];
500	/* sign + log10(MAX_INT) + '\0' */
501	char pid_str[11];
502	int spaces = 0;
503	int len;
504	int i;
505
506	trace_find_cmdline(pid, comm);
507	comm[7] = '\0';
508	sprintf(pid_str, "%d", pid);
509
510	/* 1 stands for the "-" character */
511	len = strlen(comm) + strlen(pid_str) + 1;
512
513	if (len < TRACE_GRAPH_PROCINFO_LENGTH)
514		spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
515
516	/* First spaces to align center */
517	for (i = 0; i < spaces / 2; i++)
518		trace_seq_putc(s, ' ');
519
520	trace_seq_printf(s, "%s-%s", comm, pid_str);
521
522	/* Last spaces to align center */
523	for (i = 0; i < spaces - (spaces / 2); i++)
524		trace_seq_putc(s, ' ');
525}
526
527
528static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
529{
530	trace_seq_putc(s, ' ');
531	trace_print_lat_fmt(s, entry);
532}
533
534/* If the pid changed since the last trace, output this event */
535static void
536verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
537{
538	pid_t prev_pid;
539	pid_t *last_pid;
540
541	if (!data)
542		return;
543
544	last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
545
546	if (*last_pid == pid)
547		return;
548
549	prev_pid = *last_pid;
550	*last_pid = pid;
551
552	if (prev_pid == -1)
553		return;
554/*
555 * Context-switch trace line:
556
557 ------------------------------------------
558 | 1)  migration/0--1  =>  sshd-1755
559 ------------------------------------------
560
561 */
562	trace_seq_puts(s, " ------------------------------------------\n");
563	print_graph_cpu(s, cpu);
564	print_graph_proc(s, prev_pid);
565	trace_seq_puts(s, " => ");
566	print_graph_proc(s, pid);
567	trace_seq_puts(s, "\n ------------------------------------------\n\n");
568}
569
570static struct ftrace_graph_ret_entry *
571get_return_for_leaf(struct trace_iterator *iter,
572		struct ftrace_graph_ent_entry *curr)
573{
574	struct fgraph_data *data = iter->private;
575	struct ring_buffer_iter *ring_iter = NULL;
576	struct ring_buffer_event *event;
577	struct ftrace_graph_ret_entry *next;
578
579	/*
580	 * If the previous output failed to write to the seq buffer,
581	 * then we just reuse the data from before.
582	 */
583	if (data && data->failed) {
584		curr = &data->ent;
585		next = &data->ret;
586	} else {
587
588		ring_iter = trace_buffer_iter(iter, iter->cpu);
589
590		/* First peek to compare current entry and the next one */
591		if (ring_iter)
592			event = ring_buffer_iter_peek(ring_iter, NULL);
593		else {
594			/*
595			 * We need to consume the current entry to see
596			 * the next one.
597			 */
598			ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
599					    NULL, NULL);
600			event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
601						 NULL, NULL);
602		}
603
604		if (!event)
605			return NULL;
606
607		next = ring_buffer_event_data(event);
608
609		if (data) {
610			/*
611			 * Save current and next entries for later reference
612			 * if the output fails.
613			 */
614			data->ent = *curr;
615			/*
616			 * If the next event is not a return type, then
617			 * we only care about what type it is. Otherwise we can
618			 * safely copy the entire event.
619			 */
620			if (next->ent.type == TRACE_GRAPH_RET)
621				data->ret = *next;
622			else
623				data->ret.ent.type = next->ent.type;
624		}
625	}
626
627	if (next->ent.type != TRACE_GRAPH_RET)
628		return NULL;
629
630	if (curr->ent.pid != next->ent.pid ||
631			curr->graph_ent.func != next->ret.func)
632		return NULL;
633
634	/* this is a leaf, now advance the iterator */
635	if (ring_iter)
636		ring_buffer_read(ring_iter, NULL);
637
638	return next;
639}
640
641static void print_graph_abs_time(u64 t, struct trace_seq *s)
642{
643	unsigned long usecs_rem;
644
645	usecs_rem = do_div(t, NSEC_PER_SEC);
646	usecs_rem /= 1000;
647
648	trace_seq_printf(s, "%5lu.%06lu |  ",
649			 (unsigned long)t, usecs_rem);
650}
651
652static void
653print_graph_irq(struct trace_iterator *iter, unsigned long addr,
654		enum trace_type type, int cpu, pid_t pid, u32 flags)
655{
656	struct trace_seq *s = &iter->seq;
657	struct trace_entry *ent = iter->ent;
658
659	if (addr < (unsigned long)__irqentry_text_start ||
660		addr >= (unsigned long)__irqentry_text_end)
661		return;
662
663	if (trace_flags & TRACE_ITER_CONTEXT_INFO) {
664		/* Absolute time */
665		if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
666			print_graph_abs_time(iter->ts, s);
667
668		/* Cpu */
669		if (flags & TRACE_GRAPH_PRINT_CPU)
670			print_graph_cpu(s, cpu);
671
672		/* Proc */
673		if (flags & TRACE_GRAPH_PRINT_PROC) {
674			print_graph_proc(s, pid);
675			trace_seq_puts(s, " | ");
676		}
677
678		/* Latency format */
679		if (trace_flags & TRACE_ITER_LATENCY_FMT)
680			print_graph_lat_fmt(s, ent);
681	}
682
683	/* No overhead */
684	print_graph_duration(0, s, flags | FLAGS_FILL_START);
685
686	if (type == TRACE_GRAPH_ENT)
687		trace_seq_puts(s, "==========>");
688	else
689		trace_seq_puts(s, "<==========");
690
691	print_graph_duration(0, s, flags | FLAGS_FILL_END);
692	trace_seq_putc(s, '\n');
693}
694
695void
696trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
697{
698	unsigned long nsecs_rem = do_div(duration, 1000);
699	/* log10(ULONG_MAX) + '\0' */
700	char usecs_str[21];
701	char nsecs_str[5];
702	int len;
703	int i;
704
705	sprintf(usecs_str, "%lu", (unsigned long) duration);
706
707	/* Print msecs */
708	trace_seq_printf(s, "%s", usecs_str);
709
710	len = strlen(usecs_str);
711
712	/* Print nsecs (we don't want to exceed 7 numbers) */
713	if (len < 7) {
714		size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
715
716		snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
717		trace_seq_printf(s, ".%s", nsecs_str);
718		len += strlen(nsecs_str);
719	}
720
721	trace_seq_puts(s, " us ");
722
723	/* Print remaining spaces to fit the row's width */
724	for (i = len; i < 7; i++)
725		trace_seq_putc(s, ' ');
726}
727
728static void
729print_graph_duration(unsigned long long duration, struct trace_seq *s,
730		     u32 flags)
731{
732	if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
733	    !(trace_flags & TRACE_ITER_CONTEXT_INFO))
734		return;
735
736	/* No real adata, just filling the column with spaces */
737	switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
738	case FLAGS_FILL_FULL:
739		trace_seq_puts(s, "              |  ");
740		return;
741	case FLAGS_FILL_START:
742		trace_seq_puts(s, "  ");
743		return;
744	case FLAGS_FILL_END:
745		trace_seq_puts(s, " |");
746		return;
747	}
748
749	/* Signal a overhead of time execution to the output */
750	if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
751		trace_seq_printf(s, "%c ", trace_find_mark(duration));
752	else
753		trace_seq_puts(s, "  ");
754
755	trace_print_graph_duration(duration, s);
756	trace_seq_puts(s, "|  ");
757}
758
759/* Case of a leaf function on its call entry */
760static enum print_line_t
761print_graph_entry_leaf(struct trace_iterator *iter,
762		struct ftrace_graph_ent_entry *entry,
763		struct ftrace_graph_ret_entry *ret_entry,
764		struct trace_seq *s, u32 flags)
765{
766	struct fgraph_data *data = iter->private;
767	struct ftrace_graph_ret *graph_ret;
768	struct ftrace_graph_ent *call;
769	unsigned long long duration;
770	int i;
771
772	graph_ret = &ret_entry->ret;
773	call = &entry->graph_ent;
774	duration = graph_ret->rettime - graph_ret->calltime;
775
776	if (data) {
777		struct fgraph_cpu_data *cpu_data;
778		int cpu = iter->cpu;
779
780		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
781
782		/*
783		 * Comments display at + 1 to depth. Since
784		 * this is a leaf function, keep the comments
785		 * equal to this depth.
786		 */
787		cpu_data->depth = call->depth - 1;
788
789		/* No need to keep this function around for this depth */
790		if (call->depth < FTRACE_RETFUNC_DEPTH)
791			cpu_data->enter_funcs[call->depth] = 0;
792	}
793
794	/* Overhead and duration */
795	print_graph_duration(duration, s, flags);
796
797	/* Function */
798	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
799		trace_seq_putc(s, ' ');
800
801	trace_seq_printf(s, "%ps();\n", (void *)call->func);
802
803	return trace_handle_return(s);
804}
805
806static enum print_line_t
807print_graph_entry_nested(struct trace_iterator *iter,
808			 struct ftrace_graph_ent_entry *entry,
809			 struct trace_seq *s, int cpu, u32 flags)
810{
811	struct ftrace_graph_ent *call = &entry->graph_ent;
812	struct fgraph_data *data = iter->private;
813	int i;
814
815	if (data) {
816		struct fgraph_cpu_data *cpu_data;
817		int cpu = iter->cpu;
818
819		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
820		cpu_data->depth = call->depth;
821
822		/* Save this function pointer to see if the exit matches */
823		if (call->depth < FTRACE_RETFUNC_DEPTH)
824			cpu_data->enter_funcs[call->depth] = call->func;
825	}
826
827	/* No time */
828	print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
829
830	/* Function */
831	for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
832		trace_seq_putc(s, ' ');
833
834	trace_seq_printf(s, "%ps() {\n", (void *)call->func);
835
836	if (trace_seq_has_overflowed(s))
837		return TRACE_TYPE_PARTIAL_LINE;
838
839	/*
840	 * we already consumed the current entry to check the next one
841	 * and see if this is a leaf.
842	 */
843	return TRACE_TYPE_NO_CONSUME;
844}
845
846static void
847print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
848		     int type, unsigned long addr, u32 flags)
849{
850	struct fgraph_data *data = iter->private;
851	struct trace_entry *ent = iter->ent;
852	int cpu = iter->cpu;
853
854	/* Pid */
855	verif_pid(s, ent->pid, cpu, data);
856
857	if (type)
858		/* Interrupt */
859		print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
860
861	if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
862		return;
863
864	/* Absolute time */
865	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
866		print_graph_abs_time(iter->ts, s);
867
868	/* Cpu */
869	if (flags & TRACE_GRAPH_PRINT_CPU)
870		print_graph_cpu(s, cpu);
871
872	/* Proc */
873	if (flags & TRACE_GRAPH_PRINT_PROC) {
874		print_graph_proc(s, ent->pid);
875		trace_seq_puts(s, " | ");
876	}
877
878	/* Latency format */
879	if (trace_flags & TRACE_ITER_LATENCY_FMT)
880		print_graph_lat_fmt(s, ent);
881
882	return;
883}
884
885/*
886 * Entry check for irq code
887 *
888 * returns 1 if
889 *  - we are inside irq code
890 *  - we just entered irq code
891 *
892 * retunns 0 if
893 *  - funcgraph-interrupts option is set
894 *  - we are not inside irq code
895 */
896static int
897check_irq_entry(struct trace_iterator *iter, u32 flags,
898		unsigned long addr, int depth)
899{
900	int cpu = iter->cpu;
901	int *depth_irq;
902	struct fgraph_data *data = iter->private;
903
904	/*
905	 * If we are either displaying irqs, or we got called as
906	 * a graph event and private data does not exist,
907	 * then we bypass the irq check.
908	 */
909	if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
910	    (!data))
911		return 0;
912
913	depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
914
915	/*
916	 * We are inside the irq code
917	 */
918	if (*depth_irq >= 0)
919		return 1;
920
921	if ((addr < (unsigned long)__irqentry_text_start) ||
922	    (addr >= (unsigned long)__irqentry_text_end))
923		return 0;
924
925	/*
926	 * We are entering irq code.
927	 */
928	*depth_irq = depth;
929	return 1;
930}
931
932/*
933 * Return check for irq code
934 *
935 * returns 1 if
936 *  - we are inside irq code
937 *  - we just left irq code
938 *
939 * returns 0 if
940 *  - funcgraph-interrupts option is set
941 *  - we are not inside irq code
942 */
943static int
944check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
945{
946	int cpu = iter->cpu;
947	int *depth_irq;
948	struct fgraph_data *data = iter->private;
949
950	/*
951	 * If we are either displaying irqs, or we got called as
952	 * a graph event and private data does not exist,
953	 * then we bypass the irq check.
954	 */
955	if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
956	    (!data))
957		return 0;
958
959	depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
960
961	/*
962	 * We are not inside the irq code.
963	 */
964	if (*depth_irq == -1)
965		return 0;
966
967	/*
968	 * We are inside the irq code, and this is returning entry.
969	 * Let's not trace it and clear the entry depth, since
970	 * we are out of irq code.
971	 *
972	 * This condition ensures that we 'leave the irq code' once
973	 * we are out of the entry depth. Thus protecting us from
974	 * the RETURN entry loss.
975	 */
976	if (*depth_irq >= depth) {
977		*depth_irq = -1;
978		return 1;
979	}
980
981	/*
982	 * We are inside the irq code, and this is not the entry.
983	 */
984	return 1;
985}
986
987static enum print_line_t
988print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
989			struct trace_iterator *iter, u32 flags)
990{
991	struct fgraph_data *data = iter->private;
992	struct ftrace_graph_ent *call = &field->graph_ent;
993	struct ftrace_graph_ret_entry *leaf_ret;
994	static enum print_line_t ret;
995	int cpu = iter->cpu;
996
997	if (check_irq_entry(iter, flags, call->func, call->depth))
998		return TRACE_TYPE_HANDLED;
999
1000	print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1001
1002	leaf_ret = get_return_for_leaf(iter, field);
1003	if (leaf_ret)
1004		ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1005	else
1006		ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1007
1008	if (data) {
1009		/*
1010		 * If we failed to write our output, then we need to make
1011		 * note of it. Because we already consumed our entry.
1012		 */
1013		if (s->full) {
1014			data->failed = 1;
1015			data->cpu = cpu;
1016		} else
1017			data->failed = 0;
1018	}
1019
1020	return ret;
1021}
1022
1023static enum print_line_t
1024print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1025		   struct trace_entry *ent, struct trace_iterator *iter,
1026		   u32 flags)
1027{
1028	unsigned long long duration = trace->rettime - trace->calltime;
1029	struct fgraph_data *data = iter->private;
1030	pid_t pid = ent->pid;
1031	int cpu = iter->cpu;
1032	int func_match = 1;
1033	int i;
1034
1035	if (check_irq_return(iter, flags, trace->depth))
1036		return TRACE_TYPE_HANDLED;
1037
1038	if (data) {
1039		struct fgraph_cpu_data *cpu_data;
1040		int cpu = iter->cpu;
1041
1042		cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1043
1044		/*
1045		 * Comments display at + 1 to depth. This is the
1046		 * return from a function, we now want the comments
1047		 * to display at the same level of the bracket.
1048		 */
1049		cpu_data->depth = trace->depth - 1;
1050
1051		if (trace->depth < FTRACE_RETFUNC_DEPTH) {
1052			if (cpu_data->enter_funcs[trace->depth] != trace->func)
1053				func_match = 0;
1054			cpu_data->enter_funcs[trace->depth] = 0;
1055		}
1056	}
1057
1058	print_graph_prologue(iter, s, 0, 0, flags);
1059
1060	/* Overhead and duration */
1061	print_graph_duration(duration, s, flags);
1062
1063	/* Closing brace */
1064	for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1065		trace_seq_putc(s, ' ');
1066
1067	/*
1068	 * If the return function does not have a matching entry,
1069	 * then the entry was lost. Instead of just printing
1070	 * the '}' and letting the user guess what function this
1071	 * belongs to, write out the function name. Always do
1072	 * that if the funcgraph-tail option is enabled.
1073	 */
1074	if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1075		trace_seq_puts(s, "}\n");
1076	else
1077		trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1078
1079	/* Overrun */
1080	if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1081		trace_seq_printf(s, " (Overruns: %lu)\n",
1082				 trace->overrun);
1083
1084	print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1085			cpu, pid, flags);
1086
1087	return trace_handle_return(s);
1088}
1089
1090static enum print_line_t
1091print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1092		    struct trace_iterator *iter, u32 flags)
1093{
1094	unsigned long sym_flags = (trace_flags & TRACE_ITER_SYM_MASK);
1095	struct fgraph_data *data = iter->private;
1096	struct trace_event *event;
1097	int depth = 0;
1098	int ret;
1099	int i;
1100
1101	if (data)
1102		depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1103
1104	print_graph_prologue(iter, s, 0, 0, flags);
1105
1106	/* No time */
1107	print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
1108
1109	/* Indentation */
1110	if (depth > 0)
1111		for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1112			trace_seq_putc(s, ' ');
1113
1114	/* The comment */
1115	trace_seq_puts(s, "/* ");
1116
1117	switch (iter->ent->type) {
1118	case TRACE_BPRINT:
1119		ret = trace_print_bprintk_msg_only(iter);
1120		if (ret != TRACE_TYPE_HANDLED)
1121			return ret;
1122		break;
1123	case TRACE_PRINT:
1124		ret = trace_print_printk_msg_only(iter);
1125		if (ret != TRACE_TYPE_HANDLED)
1126			return ret;
1127		break;
1128	default:
1129		event = ftrace_find_event(ent->type);
1130		if (!event)
1131			return TRACE_TYPE_UNHANDLED;
1132
1133		ret = event->funcs->trace(iter, sym_flags, event);
1134		if (ret != TRACE_TYPE_HANDLED)
1135			return ret;
1136	}
1137
1138	if (trace_seq_has_overflowed(s))
1139		goto out;
1140
1141	/* Strip ending newline */
1142	if (s->buffer[s->seq.len - 1] == '\n') {
1143		s->buffer[s->seq.len - 1] = '\0';
1144		s->seq.len--;
1145	}
1146
1147	trace_seq_puts(s, " */\n");
1148 out:
1149	return trace_handle_return(s);
1150}
1151
1152
1153enum print_line_t
1154print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1155{
1156	struct ftrace_graph_ent_entry *field;
1157	struct fgraph_data *data = iter->private;
1158	struct trace_entry *entry = iter->ent;
1159	struct trace_seq *s = &iter->seq;
1160	int cpu = iter->cpu;
1161	int ret;
1162
1163	if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1164		per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1165		return TRACE_TYPE_HANDLED;
1166	}
1167
1168	/*
1169	 * If the last output failed, there's a possibility we need
1170	 * to print out the missing entry which would never go out.
1171	 */
1172	if (data && data->failed) {
1173		field = &data->ent;
1174		iter->cpu = data->cpu;
1175		ret = print_graph_entry(field, s, iter, flags);
1176		if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1177			per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1178			ret = TRACE_TYPE_NO_CONSUME;
1179		}
1180		iter->cpu = cpu;
1181		return ret;
1182	}
1183
1184	switch (entry->type) {
1185	case TRACE_GRAPH_ENT: {
1186		/*
1187		 * print_graph_entry() may consume the current event,
1188		 * thus @field may become invalid, so we need to save it.
1189		 * sizeof(struct ftrace_graph_ent_entry) is very small,
1190		 * it can be safely saved at the stack.
1191		 */
1192		struct ftrace_graph_ent_entry saved;
1193		trace_assign_type(field, entry);
1194		saved = *field;
1195		return print_graph_entry(&saved, s, iter, flags);
1196	}
1197	case TRACE_GRAPH_RET: {
1198		struct ftrace_graph_ret_entry *field;
1199		trace_assign_type(field, entry);
1200		return print_graph_return(&field->ret, s, entry, iter, flags);
1201	}
1202	case TRACE_STACK:
1203	case TRACE_FN:
1204		/* dont trace stack and functions as comments */
1205		return TRACE_TYPE_UNHANDLED;
1206
1207	default:
1208		return print_graph_comment(s, entry, iter, flags);
1209	}
1210
1211	return TRACE_TYPE_HANDLED;
1212}
1213
1214static enum print_line_t
1215print_graph_function(struct trace_iterator *iter)
1216{
1217	return print_graph_function_flags(iter, tracer_flags.val);
1218}
1219
1220static enum print_line_t
1221print_graph_function_event(struct trace_iterator *iter, int flags,
1222			   struct trace_event *event)
1223{
1224	return print_graph_function(iter);
1225}
1226
1227static void print_lat_header(struct seq_file *s, u32 flags)
1228{
1229	static const char spaces[] = "                "	/* 16 spaces */
1230		"    "					/* 4 spaces */
1231		"                 ";			/* 17 spaces */
1232	int size = 0;
1233
1234	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1235		size += 16;
1236	if (flags & TRACE_GRAPH_PRINT_CPU)
1237		size += 4;
1238	if (flags & TRACE_GRAPH_PRINT_PROC)
1239		size += 17;
1240
1241	seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1242	seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1243	seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1244	seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1245	seq_printf(s, "#%.*s||| /                      \n", size, spaces);
1246}
1247
1248static void __print_graph_headers_flags(struct seq_file *s, u32 flags)
1249{
1250	int lat = trace_flags & TRACE_ITER_LATENCY_FMT;
1251
1252	if (lat)
1253		print_lat_header(s, flags);
1254
1255	/* 1st line */
1256	seq_putc(s, '#');
1257	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1258		seq_puts(s, "     TIME       ");
1259	if (flags & TRACE_GRAPH_PRINT_CPU)
1260		seq_puts(s, " CPU");
1261	if (flags & TRACE_GRAPH_PRINT_PROC)
1262		seq_puts(s, "  TASK/PID       ");
1263	if (lat)
1264		seq_puts(s, "||||");
1265	if (flags & TRACE_GRAPH_PRINT_DURATION)
1266		seq_puts(s, "  DURATION   ");
1267	seq_puts(s, "               FUNCTION CALLS\n");
1268
1269	/* 2nd line */
1270	seq_putc(s, '#');
1271	if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1272		seq_puts(s, "      |         ");
1273	if (flags & TRACE_GRAPH_PRINT_CPU)
1274		seq_puts(s, " |  ");
1275	if (flags & TRACE_GRAPH_PRINT_PROC)
1276		seq_puts(s, "   |    |        ");
1277	if (lat)
1278		seq_puts(s, "||||");
1279	if (flags & TRACE_GRAPH_PRINT_DURATION)
1280		seq_puts(s, "   |   |      ");
1281	seq_puts(s, "               |   |   |   |\n");
1282}
1283
1284static void print_graph_headers(struct seq_file *s)
1285{
1286	print_graph_headers_flags(s, tracer_flags.val);
1287}
1288
1289void print_graph_headers_flags(struct seq_file *s, u32 flags)
1290{
1291	struct trace_iterator *iter = s->private;
1292
1293	if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
1294		return;
1295
1296	if (trace_flags & TRACE_ITER_LATENCY_FMT) {
1297		/* print nothing if the buffers are empty */
1298		if (trace_empty(iter))
1299			return;
1300
1301		print_trace_header(s, iter);
1302	}
1303
1304	__print_graph_headers_flags(s, flags);
1305}
1306
1307void graph_trace_open(struct trace_iterator *iter)
1308{
1309	/* pid and depth on the last trace processed */
1310	struct fgraph_data *data;
1311	gfp_t gfpflags;
1312	int cpu;
1313
1314	iter->private = NULL;
1315
1316	/* We can be called in atomic context via ftrace_dump() */
1317	gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1318
1319	data = kzalloc(sizeof(*data), gfpflags);
1320	if (!data)
1321		goto out_err;
1322
1323	data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1324	if (!data->cpu_data)
1325		goto out_err_free;
1326
1327	for_each_possible_cpu(cpu) {
1328		pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1329		int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1330		int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1331		int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1332
1333		*pid = -1;
1334		*depth = 0;
1335		*ignore = 0;
1336		*depth_irq = -1;
1337	}
1338
1339	iter->private = data;
1340
1341	return;
1342
1343 out_err_free:
1344	kfree(data);
1345 out_err:
1346	pr_warning("function graph tracer: not enough memory\n");
1347}
1348
1349void graph_trace_close(struct trace_iterator *iter)
1350{
1351	struct fgraph_data *data = iter->private;
1352
1353	if (data) {
1354		free_percpu(data->cpu_data);
1355		kfree(data);
1356	}
1357}
1358
1359static int
1360func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1361{
1362	if (bit == TRACE_GRAPH_PRINT_IRQS)
1363		ftrace_graph_skip_irqs = !set;
1364
1365	return 0;
1366}
1367
1368static struct trace_event_functions graph_functions = {
1369	.trace		= print_graph_function_event,
1370};
1371
1372static struct trace_event graph_trace_entry_event = {
1373	.type		= TRACE_GRAPH_ENT,
1374	.funcs		= &graph_functions,
1375};
1376
1377static struct trace_event graph_trace_ret_event = {
1378	.type		= TRACE_GRAPH_RET,
1379	.funcs		= &graph_functions
1380};
1381
1382static struct tracer graph_trace __tracer_data = {
1383	.name		= "function_graph",
1384	.update_thresh	= graph_trace_update_thresh,
1385	.open		= graph_trace_open,
1386	.pipe_open	= graph_trace_open,
1387	.close		= graph_trace_close,
1388	.pipe_close	= graph_trace_close,
1389	.init		= graph_trace_init,
1390	.reset		= graph_trace_reset,
1391	.print_line	= print_graph_function,
1392	.print_header	= print_graph_headers,
1393	.flags		= &tracer_flags,
1394	.set_flag	= func_graph_set_flag,
1395#ifdef CONFIG_FTRACE_SELFTEST
1396	.selftest	= trace_selftest_startup_function_graph,
1397#endif
1398};
1399
1400
1401static ssize_t
1402graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1403		  loff_t *ppos)
1404{
1405	unsigned long val;
1406	int ret;
1407
1408	ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1409	if (ret)
1410		return ret;
1411
1412	max_depth = val;
1413
1414	*ppos += cnt;
1415
1416	return cnt;
1417}
1418
1419static ssize_t
1420graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1421		 loff_t *ppos)
1422{
1423	char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1424	int n;
1425
1426	n = sprintf(buf, "%d\n", max_depth);
1427
1428	return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1429}
1430
1431static const struct file_operations graph_depth_fops = {
1432	.open		= tracing_open_generic,
1433	.write		= graph_depth_write,
1434	.read		= graph_depth_read,
1435	.llseek		= generic_file_llseek,
1436};
1437
1438static __init int init_graph_tracefs(void)
1439{
1440	struct dentry *d_tracer;
1441
1442	d_tracer = tracing_init_dentry();
1443	if (IS_ERR(d_tracer))
1444		return 0;
1445
1446	trace_create_file("max_graph_depth", 0644, d_tracer,
1447			  NULL, &graph_depth_fops);
1448
1449	return 0;
1450}
1451fs_initcall(init_graph_tracefs);
1452
1453static __init int init_graph_trace(void)
1454{
1455	max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1456
1457	if (!register_ftrace_event(&graph_trace_entry_event)) {
1458		pr_warning("Warning: could not register graph trace events\n");
1459		return 1;
1460	}
1461
1462	if (!register_ftrace_event(&graph_trace_ret_event)) {
1463		pr_warning("Warning: could not register graph trace events\n");
1464		return 1;
1465	}
1466
1467	return register_tracer(&graph_trace);
1468}
1469
1470core_initcall(init_graph_trace);
1471