Loading...
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/interrupt.h>
12#include <linux/slab.h>
13#include <linux/fs.h>
14
15#include "trace.h"
16#include "trace_output.h"
17
18static bool kill_ftrace_graph;
19
20/**
21 * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
22 *
23 * ftrace_graph_stop() is called when a severe error is detected in
24 * the function graph tracing. This function is called by the critical
25 * paths of function graph to keep those paths from doing any more harm.
26 */
27bool ftrace_graph_is_dead(void)
28{
29 return kill_ftrace_graph;
30}
31
32/**
33 * ftrace_graph_stop - set to permanently disable function graph tracincg
34 *
35 * In case of an error int function graph tracing, this is called
36 * to try to keep function graph tracing from causing any more harm.
37 * Usually this is pretty severe and this is called to try to at least
38 * get a warning out to the user.
39 */
40void ftrace_graph_stop(void)
41{
42 kill_ftrace_graph = true;
43}
44
45/* When set, irq functions will be ignored */
46static int ftrace_graph_skip_irqs;
47
48struct fgraph_cpu_data {
49 pid_t last_pid;
50 int depth;
51 int depth_irq;
52 int ignore;
53 unsigned long enter_funcs[FTRACE_RETFUNC_DEPTH];
54};
55
56struct fgraph_data {
57 struct fgraph_cpu_data __percpu *cpu_data;
58
59 /* Place to preserve last processed entry. */
60 struct ftrace_graph_ent_entry ent;
61 struct ftrace_graph_ret_entry ret;
62 int failed;
63 int cpu;
64};
65
66#define TRACE_GRAPH_INDENT 2
67
68static unsigned int max_depth;
69
70static struct tracer_opt trace_opts[] = {
71 /* Display overruns? (for self-debug purpose) */
72 { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
73 /* Display CPU ? */
74 { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
75 /* Display Overhead ? */
76 { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
77 /* Display proc name/pid */
78 { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
79 /* Display duration of execution */
80 { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
81 /* Display absolute time of an entry */
82 { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
83 /* Display interrupts */
84 { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
85 /* Display function name after trailing } */
86 { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
87 /* Include sleep time (scheduled out) between entry and return */
88 { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
89 /* Include time within nested functions */
90 { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
91 { } /* Empty entry */
92};
93
94static struct tracer_flags tracer_flags = {
95 /* Don't display overruns, proc, or tail by default */
96 .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
97 TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
98 TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
99 .opts = trace_opts
100};
101
102static struct trace_array *graph_array;
103
104/*
105 * DURATION column is being also used to display IRQ signs,
106 * following values are used by print_graph_irq and others
107 * to fill in space into DURATION column.
108 */
109enum {
110 FLAGS_FILL_FULL = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
111 FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
112 FLAGS_FILL_END = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
113};
114
115static void
116print_graph_duration(struct trace_array *tr, unsigned long long duration,
117 struct trace_seq *s, u32 flags);
118
119/* Add a function return address to the trace stack on thread info.*/
120int
121ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
122 unsigned long frame_pointer)
123{
124 unsigned long long calltime;
125 int index;
126
127 if (unlikely(ftrace_graph_is_dead()))
128 return -EBUSY;
129
130 if (!current->ret_stack)
131 return -EBUSY;
132
133 /*
134 * We must make sure the ret_stack is tested before we read
135 * anything else.
136 */
137 smp_rmb();
138
139 /* The return trace stack is full */
140 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
141 atomic_inc(¤t->trace_overrun);
142 return -EBUSY;
143 }
144
145 /*
146 * The curr_ret_stack is an index to ftrace return stack of
147 * current task. Its value should be in [0, FTRACE_RETFUNC_
148 * DEPTH) when the function graph tracer is used. To support
149 * filtering out specific functions, it makes the index
150 * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
151 * so when it sees a negative index the ftrace will ignore
152 * the record. And the index gets recovered when returning
153 * from the filtered function by adding the FTRACE_NOTRACE_
154 * DEPTH and then it'll continue to record functions normally.
155 *
156 * The curr_ret_stack is initialized to -1 and get increased
157 * in this function. So it can be less than -1 only if it was
158 * filtered out via ftrace_graph_notrace_addr() which can be
159 * set from set_graph_notrace file in tracefs by user.
160 */
161 if (current->curr_ret_stack < -1)
162 return -EBUSY;
163
164 calltime = trace_clock_local();
165
166 index = ++current->curr_ret_stack;
167 if (ftrace_graph_notrace_addr(func))
168 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
169 barrier();
170 current->ret_stack[index].ret = ret;
171 current->ret_stack[index].func = func;
172 current->ret_stack[index].calltime = calltime;
173 current->ret_stack[index].subtime = 0;
174 current->ret_stack[index].fp = frame_pointer;
175 *depth = current->curr_ret_stack;
176
177 return 0;
178}
179
180/* Retrieve a function return address to the trace stack on thread info.*/
181static void
182ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
183 unsigned long frame_pointer)
184{
185 int index;
186
187 index = current->curr_ret_stack;
188
189 /*
190 * A negative index here means that it's just returned from a
191 * notrace'd function. Recover index to get an original
192 * return address. See ftrace_push_return_trace().
193 *
194 * TODO: Need to check whether the stack gets corrupted.
195 */
196 if (index < 0)
197 index += FTRACE_NOTRACE_DEPTH;
198
199 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
200 ftrace_graph_stop();
201 WARN_ON(1);
202 /* Might as well panic, otherwise we have no where to go */
203 *ret = (unsigned long)panic;
204 return;
205 }
206
207#if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
208 /*
209 * The arch may choose to record the frame pointer used
210 * and check it here to make sure that it is what we expect it
211 * to be. If gcc does not set the place holder of the return
212 * address in the frame pointer, and does a copy instead, then
213 * the function graph trace will fail. This test detects this
214 * case.
215 *
216 * Currently, x86_32 with optimize for size (-Os) makes the latest
217 * gcc do the above.
218 *
219 * Note, -mfentry does not use frame pointers, and this test
220 * is not needed if CC_USING_FENTRY is set.
221 */
222 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
223 ftrace_graph_stop();
224 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
225 " from func %ps return to %lx\n",
226 current->ret_stack[index].fp,
227 frame_pointer,
228 (void *)current->ret_stack[index].func,
229 current->ret_stack[index].ret);
230 *ret = (unsigned long)panic;
231 return;
232 }
233#endif
234
235 *ret = current->ret_stack[index].ret;
236 trace->func = current->ret_stack[index].func;
237 trace->calltime = current->ret_stack[index].calltime;
238 trace->overrun = atomic_read(¤t->trace_overrun);
239 trace->depth = index;
240}
241
242/*
243 * Send the trace to the ring-buffer.
244 * @return the original return address.
245 */
246unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
247{
248 struct ftrace_graph_ret trace;
249 unsigned long ret;
250
251 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
252 trace.rettime = trace_clock_local();
253 barrier();
254 current->curr_ret_stack--;
255 /*
256 * The curr_ret_stack can be less than -1 only if it was
257 * filtered out and it's about to return from the function.
258 * Recover the index and continue to trace normal functions.
259 */
260 if (current->curr_ret_stack < -1) {
261 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
262 return ret;
263 }
264
265 /*
266 * The trace should run after decrementing the ret counter
267 * in case an interrupt were to come in. We don't want to
268 * lose the interrupt if max_depth is set.
269 */
270 ftrace_graph_return(&trace);
271
272 if (unlikely(!ret)) {
273 ftrace_graph_stop();
274 WARN_ON(1);
275 /* Might as well panic. What else to do? */
276 ret = (unsigned long)panic;
277 }
278
279 return ret;
280}
281
282int __trace_graph_entry(struct trace_array *tr,
283 struct ftrace_graph_ent *trace,
284 unsigned long flags,
285 int pc)
286{
287 struct trace_event_call *call = &event_funcgraph_entry;
288 struct ring_buffer_event *event;
289 struct ring_buffer *buffer = tr->trace_buffer.buffer;
290 struct ftrace_graph_ent_entry *entry;
291
292 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
293 sizeof(*entry), flags, pc);
294 if (!event)
295 return 0;
296 entry = ring_buffer_event_data(event);
297 entry->graph_ent = *trace;
298 if (!call_filter_check_discard(call, entry, buffer, event))
299 __buffer_unlock_commit(buffer, event);
300
301 return 1;
302}
303
304static inline int ftrace_graph_ignore_irqs(void)
305{
306 if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
307 return 0;
308
309 return in_irq();
310}
311
312int trace_graph_entry(struct ftrace_graph_ent *trace)
313{
314 struct trace_array *tr = graph_array;
315 struct trace_array_cpu *data;
316 unsigned long flags;
317 long disabled;
318 int ret;
319 int cpu;
320 int pc;
321
322 if (!ftrace_trace_task(current))
323 return 0;
324
325 /* trace it when it is-nested-in or is a function enabled. */
326 if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
327 ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
328 (max_depth && trace->depth >= max_depth))
329 return 0;
330
331 /*
332 * Do not trace a function if it's filtered by set_graph_notrace.
333 * Make the index of ret stack negative to indicate that it should
334 * ignore further functions. But it needs its own ret stack entry
335 * to recover the original index in order to continue tracing after
336 * returning from the function.
337 */
338 if (ftrace_graph_notrace_addr(trace->func))
339 return 1;
340
341 local_irq_save(flags);
342 cpu = raw_smp_processor_id();
343 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
344 disabled = atomic_inc_return(&data->disabled);
345 if (likely(disabled == 1)) {
346 pc = preempt_count();
347 ret = __trace_graph_entry(tr, trace, flags, pc);
348 } else {
349 ret = 0;
350 }
351
352 atomic_dec(&data->disabled);
353 local_irq_restore(flags);
354
355 return ret;
356}
357
358static int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
359{
360 if (tracing_thresh)
361 return 1;
362 else
363 return trace_graph_entry(trace);
364}
365
366static void
367__trace_graph_function(struct trace_array *tr,
368 unsigned long ip, unsigned long flags, int pc)
369{
370 u64 time = trace_clock_local();
371 struct ftrace_graph_ent ent = {
372 .func = ip,
373 .depth = 0,
374 };
375 struct ftrace_graph_ret ret = {
376 .func = ip,
377 .depth = 0,
378 .calltime = time,
379 .rettime = time,
380 };
381
382 __trace_graph_entry(tr, &ent, flags, pc);
383 __trace_graph_return(tr, &ret, flags, pc);
384}
385
386void
387trace_graph_function(struct trace_array *tr,
388 unsigned long ip, unsigned long parent_ip,
389 unsigned long flags, int pc)
390{
391 __trace_graph_function(tr, ip, flags, pc);
392}
393
394void __trace_graph_return(struct trace_array *tr,
395 struct ftrace_graph_ret *trace,
396 unsigned long flags,
397 int pc)
398{
399 struct trace_event_call *call = &event_funcgraph_exit;
400 struct ring_buffer_event *event;
401 struct ring_buffer *buffer = tr->trace_buffer.buffer;
402 struct ftrace_graph_ret_entry *entry;
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_array *tr = iter->tr;
657 struct trace_seq *s = &iter->seq;
658 struct trace_entry *ent = iter->ent;
659
660 if (addr < (unsigned long)__irqentry_text_start ||
661 addr >= (unsigned long)__irqentry_text_end)
662 return;
663
664 if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
665 /* Absolute time */
666 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
667 print_graph_abs_time(iter->ts, s);
668
669 /* Cpu */
670 if (flags & TRACE_GRAPH_PRINT_CPU)
671 print_graph_cpu(s, cpu);
672
673 /* Proc */
674 if (flags & TRACE_GRAPH_PRINT_PROC) {
675 print_graph_proc(s, pid);
676 trace_seq_puts(s, " | ");
677 }
678
679 /* Latency format */
680 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
681 print_graph_lat_fmt(s, ent);
682 }
683
684 /* No overhead */
685 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
686
687 if (type == TRACE_GRAPH_ENT)
688 trace_seq_puts(s, "==========>");
689 else
690 trace_seq_puts(s, "<==========");
691
692 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
693 trace_seq_putc(s, '\n');
694}
695
696void
697trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
698{
699 unsigned long nsecs_rem = do_div(duration, 1000);
700 /* log10(ULONG_MAX) + '\0' */
701 char usecs_str[21];
702 char nsecs_str[5];
703 int len;
704 int i;
705
706 sprintf(usecs_str, "%lu", (unsigned long) duration);
707
708 /* Print msecs */
709 trace_seq_printf(s, "%s", usecs_str);
710
711 len = strlen(usecs_str);
712
713 /* Print nsecs (we don't want to exceed 7 numbers) */
714 if (len < 7) {
715 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
716
717 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
718 trace_seq_printf(s, ".%s", nsecs_str);
719 len += strlen(nsecs_str) + 1;
720 }
721
722 trace_seq_puts(s, " us ");
723
724 /* Print remaining spaces to fit the row's width */
725 for (i = len; i < 8; i++)
726 trace_seq_putc(s, ' ');
727}
728
729static void
730print_graph_duration(struct trace_array *tr, unsigned long long duration,
731 struct trace_seq *s, u32 flags)
732{
733 if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
734 !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
735 return;
736
737 /* No real adata, just filling the column with spaces */
738 switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
739 case FLAGS_FILL_FULL:
740 trace_seq_puts(s, " | ");
741 return;
742 case FLAGS_FILL_START:
743 trace_seq_puts(s, " ");
744 return;
745 case FLAGS_FILL_END:
746 trace_seq_puts(s, " |");
747 return;
748 }
749
750 /* Signal a overhead of time execution to the output */
751 if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
752 trace_seq_printf(s, "%c ", trace_find_mark(duration));
753 else
754 trace_seq_puts(s, " ");
755
756 trace_print_graph_duration(duration, s);
757 trace_seq_puts(s, "| ");
758}
759
760/* Case of a leaf function on its call entry */
761static enum print_line_t
762print_graph_entry_leaf(struct trace_iterator *iter,
763 struct ftrace_graph_ent_entry *entry,
764 struct ftrace_graph_ret_entry *ret_entry,
765 struct trace_seq *s, u32 flags)
766{
767 struct fgraph_data *data = iter->private;
768 struct trace_array *tr = iter->tr;
769 struct ftrace_graph_ret *graph_ret;
770 struct ftrace_graph_ent *call;
771 unsigned long long duration;
772 int i;
773
774 graph_ret = &ret_entry->ret;
775 call = &entry->graph_ent;
776 duration = graph_ret->rettime - graph_ret->calltime;
777
778 if (data) {
779 struct fgraph_cpu_data *cpu_data;
780 int cpu = iter->cpu;
781
782 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
783
784 /*
785 * Comments display at + 1 to depth. Since
786 * this is a leaf function, keep the comments
787 * equal to this depth.
788 */
789 cpu_data->depth = call->depth - 1;
790
791 /* No need to keep this function around for this depth */
792 if (call->depth < FTRACE_RETFUNC_DEPTH)
793 cpu_data->enter_funcs[call->depth] = 0;
794 }
795
796 /* Overhead and duration */
797 print_graph_duration(tr, duration, s, flags);
798
799 /* Function */
800 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
801 trace_seq_putc(s, ' ');
802
803 trace_seq_printf(s, "%ps();\n", (void *)call->func);
804
805 return trace_handle_return(s);
806}
807
808static enum print_line_t
809print_graph_entry_nested(struct trace_iterator *iter,
810 struct ftrace_graph_ent_entry *entry,
811 struct trace_seq *s, int cpu, u32 flags)
812{
813 struct ftrace_graph_ent *call = &entry->graph_ent;
814 struct fgraph_data *data = iter->private;
815 struct trace_array *tr = iter->tr;
816 int i;
817
818 if (data) {
819 struct fgraph_cpu_data *cpu_data;
820 int cpu = iter->cpu;
821
822 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
823 cpu_data->depth = call->depth;
824
825 /* Save this function pointer to see if the exit matches */
826 if (call->depth < FTRACE_RETFUNC_DEPTH)
827 cpu_data->enter_funcs[call->depth] = call->func;
828 }
829
830 /* No time */
831 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
832
833 /* Function */
834 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
835 trace_seq_putc(s, ' ');
836
837 trace_seq_printf(s, "%ps() {\n", (void *)call->func);
838
839 if (trace_seq_has_overflowed(s))
840 return TRACE_TYPE_PARTIAL_LINE;
841
842 /*
843 * we already consumed the current entry to check the next one
844 * and see if this is a leaf.
845 */
846 return TRACE_TYPE_NO_CONSUME;
847}
848
849static void
850print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
851 int type, unsigned long addr, u32 flags)
852{
853 struct fgraph_data *data = iter->private;
854 struct trace_entry *ent = iter->ent;
855 struct trace_array *tr = iter->tr;
856 int cpu = iter->cpu;
857
858 /* Pid */
859 verif_pid(s, ent->pid, cpu, data);
860
861 if (type)
862 /* Interrupt */
863 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
864
865 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
866 return;
867
868 /* Absolute time */
869 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
870 print_graph_abs_time(iter->ts, s);
871
872 /* Cpu */
873 if (flags & TRACE_GRAPH_PRINT_CPU)
874 print_graph_cpu(s, cpu);
875
876 /* Proc */
877 if (flags & TRACE_GRAPH_PRINT_PROC) {
878 print_graph_proc(s, ent->pid);
879 trace_seq_puts(s, " | ");
880 }
881
882 /* Latency format */
883 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
884 print_graph_lat_fmt(s, ent);
885
886 return;
887}
888
889/*
890 * Entry check for irq code
891 *
892 * returns 1 if
893 * - we are inside irq code
894 * - we just entered irq code
895 *
896 * retunns 0 if
897 * - funcgraph-interrupts option is set
898 * - we are not inside irq code
899 */
900static int
901check_irq_entry(struct trace_iterator *iter, u32 flags,
902 unsigned long addr, int depth)
903{
904 int cpu = iter->cpu;
905 int *depth_irq;
906 struct fgraph_data *data = iter->private;
907
908 /*
909 * If we are either displaying irqs, or we got called as
910 * a graph event and private data does not exist,
911 * then we bypass the irq check.
912 */
913 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
914 (!data))
915 return 0;
916
917 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
918
919 /*
920 * We are inside the irq code
921 */
922 if (*depth_irq >= 0)
923 return 1;
924
925 if ((addr < (unsigned long)__irqentry_text_start) ||
926 (addr >= (unsigned long)__irqentry_text_end))
927 return 0;
928
929 /*
930 * We are entering irq code.
931 */
932 *depth_irq = depth;
933 return 1;
934}
935
936/*
937 * Return check for irq code
938 *
939 * returns 1 if
940 * - we are inside irq code
941 * - we just left irq code
942 *
943 * returns 0 if
944 * - funcgraph-interrupts option is set
945 * - we are not inside irq code
946 */
947static int
948check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
949{
950 int cpu = iter->cpu;
951 int *depth_irq;
952 struct fgraph_data *data = iter->private;
953
954 /*
955 * If we are either displaying irqs, or we got called as
956 * a graph event and private data does not exist,
957 * then we bypass the irq check.
958 */
959 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
960 (!data))
961 return 0;
962
963 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
964
965 /*
966 * We are not inside the irq code.
967 */
968 if (*depth_irq == -1)
969 return 0;
970
971 /*
972 * We are inside the irq code, and this is returning entry.
973 * Let's not trace it and clear the entry depth, since
974 * we are out of irq code.
975 *
976 * This condition ensures that we 'leave the irq code' once
977 * we are out of the entry depth. Thus protecting us from
978 * the RETURN entry loss.
979 */
980 if (*depth_irq >= depth) {
981 *depth_irq = -1;
982 return 1;
983 }
984
985 /*
986 * We are inside the irq code, and this is not the entry.
987 */
988 return 1;
989}
990
991static enum print_line_t
992print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
993 struct trace_iterator *iter, u32 flags)
994{
995 struct fgraph_data *data = iter->private;
996 struct ftrace_graph_ent *call = &field->graph_ent;
997 struct ftrace_graph_ret_entry *leaf_ret;
998 static enum print_line_t ret;
999 int cpu = iter->cpu;
1000
1001 if (check_irq_entry(iter, flags, call->func, call->depth))
1002 return TRACE_TYPE_HANDLED;
1003
1004 print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1005
1006 leaf_ret = get_return_for_leaf(iter, field);
1007 if (leaf_ret)
1008 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1009 else
1010 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1011
1012 if (data) {
1013 /*
1014 * If we failed to write our output, then we need to make
1015 * note of it. Because we already consumed our entry.
1016 */
1017 if (s->full) {
1018 data->failed = 1;
1019 data->cpu = cpu;
1020 } else
1021 data->failed = 0;
1022 }
1023
1024 return ret;
1025}
1026
1027static enum print_line_t
1028print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1029 struct trace_entry *ent, struct trace_iterator *iter,
1030 u32 flags)
1031{
1032 unsigned long long duration = trace->rettime - trace->calltime;
1033 struct fgraph_data *data = iter->private;
1034 struct trace_array *tr = iter->tr;
1035 pid_t pid = ent->pid;
1036 int cpu = iter->cpu;
1037 int func_match = 1;
1038 int i;
1039
1040 if (check_irq_return(iter, flags, trace->depth))
1041 return TRACE_TYPE_HANDLED;
1042
1043 if (data) {
1044 struct fgraph_cpu_data *cpu_data;
1045 int cpu = iter->cpu;
1046
1047 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1048
1049 /*
1050 * Comments display at + 1 to depth. This is the
1051 * return from a function, we now want the comments
1052 * to display at the same level of the bracket.
1053 */
1054 cpu_data->depth = trace->depth - 1;
1055
1056 if (trace->depth < FTRACE_RETFUNC_DEPTH) {
1057 if (cpu_data->enter_funcs[trace->depth] != trace->func)
1058 func_match = 0;
1059 cpu_data->enter_funcs[trace->depth] = 0;
1060 }
1061 }
1062
1063 print_graph_prologue(iter, s, 0, 0, flags);
1064
1065 /* Overhead and duration */
1066 print_graph_duration(tr, duration, s, flags);
1067
1068 /* Closing brace */
1069 for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1070 trace_seq_putc(s, ' ');
1071
1072 /*
1073 * If the return function does not have a matching entry,
1074 * then the entry was lost. Instead of just printing
1075 * the '}' and letting the user guess what function this
1076 * belongs to, write out the function name. Always do
1077 * that if the funcgraph-tail option is enabled.
1078 */
1079 if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1080 trace_seq_puts(s, "}\n");
1081 else
1082 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1083
1084 /* Overrun */
1085 if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1086 trace_seq_printf(s, " (Overruns: %lu)\n",
1087 trace->overrun);
1088
1089 print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1090 cpu, pid, flags);
1091
1092 return trace_handle_return(s);
1093}
1094
1095static enum print_line_t
1096print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1097 struct trace_iterator *iter, u32 flags)
1098{
1099 struct trace_array *tr = iter->tr;
1100 unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1101 struct fgraph_data *data = iter->private;
1102 struct trace_event *event;
1103 int depth = 0;
1104 int ret;
1105 int i;
1106
1107 if (data)
1108 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1109
1110 print_graph_prologue(iter, s, 0, 0, flags);
1111
1112 /* No time */
1113 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1114
1115 /* Indentation */
1116 if (depth > 0)
1117 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1118 trace_seq_putc(s, ' ');
1119
1120 /* The comment */
1121 trace_seq_puts(s, "/* ");
1122
1123 switch (iter->ent->type) {
1124 case TRACE_BPRINT:
1125 ret = trace_print_bprintk_msg_only(iter);
1126 if (ret != TRACE_TYPE_HANDLED)
1127 return ret;
1128 break;
1129 case TRACE_PRINT:
1130 ret = trace_print_printk_msg_only(iter);
1131 if (ret != TRACE_TYPE_HANDLED)
1132 return ret;
1133 break;
1134 default:
1135 event = ftrace_find_event(ent->type);
1136 if (!event)
1137 return TRACE_TYPE_UNHANDLED;
1138
1139 ret = event->funcs->trace(iter, sym_flags, event);
1140 if (ret != TRACE_TYPE_HANDLED)
1141 return ret;
1142 }
1143
1144 if (trace_seq_has_overflowed(s))
1145 goto out;
1146
1147 /* Strip ending newline */
1148 if (s->buffer[s->seq.len - 1] == '\n') {
1149 s->buffer[s->seq.len - 1] = '\0';
1150 s->seq.len--;
1151 }
1152
1153 trace_seq_puts(s, " */\n");
1154 out:
1155 return trace_handle_return(s);
1156}
1157
1158
1159enum print_line_t
1160print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1161{
1162 struct ftrace_graph_ent_entry *field;
1163 struct fgraph_data *data = iter->private;
1164 struct trace_entry *entry = iter->ent;
1165 struct trace_seq *s = &iter->seq;
1166 int cpu = iter->cpu;
1167 int ret;
1168
1169 if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1170 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1171 return TRACE_TYPE_HANDLED;
1172 }
1173
1174 /*
1175 * If the last output failed, there's a possibility we need
1176 * to print out the missing entry which would never go out.
1177 */
1178 if (data && data->failed) {
1179 field = &data->ent;
1180 iter->cpu = data->cpu;
1181 ret = print_graph_entry(field, s, iter, flags);
1182 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1183 per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1184 ret = TRACE_TYPE_NO_CONSUME;
1185 }
1186 iter->cpu = cpu;
1187 return ret;
1188 }
1189
1190 switch (entry->type) {
1191 case TRACE_GRAPH_ENT: {
1192 /*
1193 * print_graph_entry() may consume the current event,
1194 * thus @field may become invalid, so we need to save it.
1195 * sizeof(struct ftrace_graph_ent_entry) is very small,
1196 * it can be safely saved at the stack.
1197 */
1198 struct ftrace_graph_ent_entry saved;
1199 trace_assign_type(field, entry);
1200 saved = *field;
1201 return print_graph_entry(&saved, s, iter, flags);
1202 }
1203 case TRACE_GRAPH_RET: {
1204 struct ftrace_graph_ret_entry *field;
1205 trace_assign_type(field, entry);
1206 return print_graph_return(&field->ret, s, entry, iter, flags);
1207 }
1208 case TRACE_STACK:
1209 case TRACE_FN:
1210 /* dont trace stack and functions as comments */
1211 return TRACE_TYPE_UNHANDLED;
1212
1213 default:
1214 return print_graph_comment(s, entry, iter, flags);
1215 }
1216
1217 return TRACE_TYPE_HANDLED;
1218}
1219
1220static enum print_line_t
1221print_graph_function(struct trace_iterator *iter)
1222{
1223 return print_graph_function_flags(iter, tracer_flags.val);
1224}
1225
1226static enum print_line_t
1227print_graph_function_event(struct trace_iterator *iter, int flags,
1228 struct trace_event *event)
1229{
1230 return print_graph_function(iter);
1231}
1232
1233static void print_lat_header(struct seq_file *s, u32 flags)
1234{
1235 static const char spaces[] = " " /* 16 spaces */
1236 " " /* 4 spaces */
1237 " "; /* 17 spaces */
1238 int size = 0;
1239
1240 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1241 size += 16;
1242 if (flags & TRACE_GRAPH_PRINT_CPU)
1243 size += 4;
1244 if (flags & TRACE_GRAPH_PRINT_PROC)
1245 size += 17;
1246
1247 seq_printf(s, "#%.*s _-----=> irqs-off \n", size, spaces);
1248 seq_printf(s, "#%.*s / _----=> need-resched \n", size, spaces);
1249 seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1250 seq_printf(s, "#%.*s|| / _--=> preempt-depth \n", size, spaces);
1251 seq_printf(s, "#%.*s||| / \n", size, spaces);
1252}
1253
1254static void __print_graph_headers_flags(struct trace_array *tr,
1255 struct seq_file *s, u32 flags)
1256{
1257 int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1258
1259 if (lat)
1260 print_lat_header(s, flags);
1261
1262 /* 1st line */
1263 seq_putc(s, '#');
1264 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1265 seq_puts(s, " TIME ");
1266 if (flags & TRACE_GRAPH_PRINT_CPU)
1267 seq_puts(s, " CPU");
1268 if (flags & TRACE_GRAPH_PRINT_PROC)
1269 seq_puts(s, " TASK/PID ");
1270 if (lat)
1271 seq_puts(s, "||||");
1272 if (flags & TRACE_GRAPH_PRINT_DURATION)
1273 seq_puts(s, " DURATION ");
1274 seq_puts(s, " FUNCTION CALLS\n");
1275
1276 /* 2nd line */
1277 seq_putc(s, '#');
1278 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1279 seq_puts(s, " | ");
1280 if (flags & TRACE_GRAPH_PRINT_CPU)
1281 seq_puts(s, " | ");
1282 if (flags & TRACE_GRAPH_PRINT_PROC)
1283 seq_puts(s, " | | ");
1284 if (lat)
1285 seq_puts(s, "||||");
1286 if (flags & TRACE_GRAPH_PRINT_DURATION)
1287 seq_puts(s, " | | ");
1288 seq_puts(s, " | | | |\n");
1289}
1290
1291static void print_graph_headers(struct seq_file *s)
1292{
1293 print_graph_headers_flags(s, tracer_flags.val);
1294}
1295
1296void print_graph_headers_flags(struct seq_file *s, u32 flags)
1297{
1298 struct trace_iterator *iter = s->private;
1299 struct trace_array *tr = iter->tr;
1300
1301 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1302 return;
1303
1304 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1305 /* print nothing if the buffers are empty */
1306 if (trace_empty(iter))
1307 return;
1308
1309 print_trace_header(s, iter);
1310 }
1311
1312 __print_graph_headers_flags(tr, s, flags);
1313}
1314
1315void graph_trace_open(struct trace_iterator *iter)
1316{
1317 /* pid and depth on the last trace processed */
1318 struct fgraph_data *data;
1319 gfp_t gfpflags;
1320 int cpu;
1321
1322 iter->private = NULL;
1323
1324 /* We can be called in atomic context via ftrace_dump() */
1325 gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1326
1327 data = kzalloc(sizeof(*data), gfpflags);
1328 if (!data)
1329 goto out_err;
1330
1331 data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1332 if (!data->cpu_data)
1333 goto out_err_free;
1334
1335 for_each_possible_cpu(cpu) {
1336 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1337 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1338 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1339 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1340
1341 *pid = -1;
1342 *depth = 0;
1343 *ignore = 0;
1344 *depth_irq = -1;
1345 }
1346
1347 iter->private = data;
1348
1349 return;
1350
1351 out_err_free:
1352 kfree(data);
1353 out_err:
1354 pr_warn("function graph tracer: not enough memory\n");
1355}
1356
1357void graph_trace_close(struct trace_iterator *iter)
1358{
1359 struct fgraph_data *data = iter->private;
1360
1361 if (data) {
1362 free_percpu(data->cpu_data);
1363 kfree(data);
1364 }
1365}
1366
1367static int
1368func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1369{
1370 if (bit == TRACE_GRAPH_PRINT_IRQS)
1371 ftrace_graph_skip_irqs = !set;
1372
1373 if (bit == TRACE_GRAPH_SLEEP_TIME)
1374 ftrace_graph_sleep_time_control(set);
1375
1376 if (bit == TRACE_GRAPH_GRAPH_TIME)
1377 ftrace_graph_graph_time_control(set);
1378
1379 return 0;
1380}
1381
1382static struct trace_event_functions graph_functions = {
1383 .trace = print_graph_function_event,
1384};
1385
1386static struct trace_event graph_trace_entry_event = {
1387 .type = TRACE_GRAPH_ENT,
1388 .funcs = &graph_functions,
1389};
1390
1391static struct trace_event graph_trace_ret_event = {
1392 .type = TRACE_GRAPH_RET,
1393 .funcs = &graph_functions
1394};
1395
1396static struct tracer graph_trace __tracer_data = {
1397 .name = "function_graph",
1398 .update_thresh = graph_trace_update_thresh,
1399 .open = graph_trace_open,
1400 .pipe_open = graph_trace_open,
1401 .close = graph_trace_close,
1402 .pipe_close = graph_trace_close,
1403 .init = graph_trace_init,
1404 .reset = graph_trace_reset,
1405 .print_line = print_graph_function,
1406 .print_header = print_graph_headers,
1407 .flags = &tracer_flags,
1408 .set_flag = func_graph_set_flag,
1409#ifdef CONFIG_FTRACE_SELFTEST
1410 .selftest = trace_selftest_startup_function_graph,
1411#endif
1412};
1413
1414
1415static ssize_t
1416graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1417 loff_t *ppos)
1418{
1419 unsigned long val;
1420 int ret;
1421
1422 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1423 if (ret)
1424 return ret;
1425
1426 max_depth = val;
1427
1428 *ppos += cnt;
1429
1430 return cnt;
1431}
1432
1433static ssize_t
1434graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1435 loff_t *ppos)
1436{
1437 char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1438 int n;
1439
1440 n = sprintf(buf, "%d\n", max_depth);
1441
1442 return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1443}
1444
1445static const struct file_operations graph_depth_fops = {
1446 .open = tracing_open_generic,
1447 .write = graph_depth_write,
1448 .read = graph_depth_read,
1449 .llseek = generic_file_llseek,
1450};
1451
1452static __init int init_graph_tracefs(void)
1453{
1454 struct dentry *d_tracer;
1455
1456 d_tracer = tracing_init_dentry();
1457 if (IS_ERR(d_tracer))
1458 return 0;
1459
1460 trace_create_file("max_graph_depth", 0644, d_tracer,
1461 NULL, &graph_depth_fops);
1462
1463 return 0;
1464}
1465fs_initcall(init_graph_tracefs);
1466
1467static __init int init_graph_trace(void)
1468{
1469 max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1470
1471 if (!register_trace_event(&graph_trace_entry_event)) {
1472 pr_warn("Warning: could not register graph trace events\n");
1473 return 1;
1474 }
1475
1476 if (!register_trace_event(&graph_trace_ret_event)) {
1477 pr_warn("Warning: could not register graph trace events\n");
1478 return 1;
1479 }
1480
1481 return register_tracer(&graph_trace);
1482}
1483
1484core_initcall(init_graph_trace);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 *
4 * Function graph tracer.
5 * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
6 * Mostly borrowed from function tracer which
7 * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
8 *
9 */
10#include <linux/uaccess.h>
11#include <linux/ftrace.h>
12#include <linux/interrupt.h>
13#include <linux/slab.h>
14#include <linux/fs.h>
15
16#include "trace.h"
17#include "trace_output.h"
18
19/* When set, irq functions will be ignored */
20static int ftrace_graph_skip_irqs;
21
22struct fgraph_cpu_data {
23 pid_t last_pid;
24 int depth;
25 int depth_irq;
26 int ignore;
27 unsigned long enter_funcs[FTRACE_RETFUNC_DEPTH];
28};
29
30struct fgraph_data {
31 struct fgraph_cpu_data __percpu *cpu_data;
32
33 /* Place to preserve last processed entry. */
34 struct ftrace_graph_ent_entry ent;
35 struct ftrace_graph_ret_entry ret;
36 int failed;
37 int cpu;
38};
39
40#define TRACE_GRAPH_INDENT 2
41
42unsigned int fgraph_max_depth;
43
44static struct tracer_opt trace_opts[] = {
45 /* Display overruns? (for self-debug purpose) */
46 { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
47 /* Display CPU ? */
48 { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
49 /* Display Overhead ? */
50 { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
51 /* Display proc name/pid */
52 { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
53 /* Display duration of execution */
54 { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
55 /* Display absolute time of an entry */
56 { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
57 /* Display interrupts */
58 { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
59 /* Display function name after trailing } */
60 { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
61 /* Include sleep time (scheduled out) between entry and return */
62 { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
63
64#ifdef CONFIG_FUNCTION_PROFILER
65 /* Include time within nested functions */
66 { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
67#endif
68
69 { } /* Empty entry */
70};
71
72static struct tracer_flags tracer_flags = {
73 /* Don't display overruns, proc, or tail by default */
74 .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
75 TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
76 TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
77 .opts = trace_opts
78};
79
80static struct trace_array *graph_array;
81
82/*
83 * DURATION column is being also used to display IRQ signs,
84 * following values are used by print_graph_irq and others
85 * to fill in space into DURATION column.
86 */
87enum {
88 FLAGS_FILL_FULL = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
89 FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
90 FLAGS_FILL_END = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
91};
92
93static void
94print_graph_duration(struct trace_array *tr, unsigned long long duration,
95 struct trace_seq *s, u32 flags);
96
97int __trace_graph_entry(struct trace_array *tr,
98 struct ftrace_graph_ent *trace,
99 unsigned long flags,
100 int pc)
101{
102 struct trace_event_call *call = &event_funcgraph_entry;
103 struct ring_buffer_event *event;
104 struct ring_buffer *buffer = tr->trace_buffer.buffer;
105 struct ftrace_graph_ent_entry *entry;
106
107 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
108 sizeof(*entry), flags, pc);
109 if (!event)
110 return 0;
111 entry = ring_buffer_event_data(event);
112 entry->graph_ent = *trace;
113 if (!call_filter_check_discard(call, entry, buffer, event))
114 trace_buffer_unlock_commit_nostack(buffer, event);
115
116 return 1;
117}
118
119static inline int ftrace_graph_ignore_irqs(void)
120{
121 if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
122 return 0;
123
124 return in_irq();
125}
126
127int trace_graph_entry(struct ftrace_graph_ent *trace)
128{
129 struct trace_array *tr = graph_array;
130 struct trace_array_cpu *data;
131 unsigned long flags;
132 long disabled;
133 int ret;
134 int cpu;
135 int pc;
136
137 if (trace_recursion_test(TRACE_GRAPH_NOTRACE_BIT))
138 return 0;
139
140 /*
141 * Do not trace a function if it's filtered by set_graph_notrace.
142 * Make the index of ret stack negative to indicate that it should
143 * ignore further functions. But it needs its own ret stack entry
144 * to recover the original index in order to continue tracing after
145 * returning from the function.
146 */
147 if (ftrace_graph_notrace_addr(trace->func)) {
148 trace_recursion_set(TRACE_GRAPH_NOTRACE_BIT);
149 /*
150 * Need to return 1 to have the return called
151 * that will clear the NOTRACE bit.
152 */
153 return 1;
154 }
155
156 if (!ftrace_trace_task(tr))
157 return 0;
158
159 if (ftrace_graph_ignore_func(trace))
160 return 0;
161
162 if (ftrace_graph_ignore_irqs())
163 return 0;
164
165 /*
166 * Stop here if tracing_threshold is set. We only write function return
167 * events to the ring buffer.
168 */
169 if (tracing_thresh)
170 return 1;
171
172 local_irq_save(flags);
173 cpu = raw_smp_processor_id();
174 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
175 disabled = atomic_inc_return(&data->disabled);
176 if (likely(disabled == 1)) {
177 pc = preempt_count();
178 ret = __trace_graph_entry(tr, trace, flags, pc);
179 } else {
180 ret = 0;
181 }
182
183 atomic_dec(&data->disabled);
184 local_irq_restore(flags);
185
186 return ret;
187}
188
189static void
190__trace_graph_function(struct trace_array *tr,
191 unsigned long ip, unsigned long flags, int pc)
192{
193 u64 time = trace_clock_local();
194 struct ftrace_graph_ent ent = {
195 .func = ip,
196 .depth = 0,
197 };
198 struct ftrace_graph_ret ret = {
199 .func = ip,
200 .depth = 0,
201 .calltime = time,
202 .rettime = time,
203 };
204
205 __trace_graph_entry(tr, &ent, flags, pc);
206 __trace_graph_return(tr, &ret, flags, pc);
207}
208
209void
210trace_graph_function(struct trace_array *tr,
211 unsigned long ip, unsigned long parent_ip,
212 unsigned long flags, int pc)
213{
214 __trace_graph_function(tr, ip, flags, pc);
215}
216
217void __trace_graph_return(struct trace_array *tr,
218 struct ftrace_graph_ret *trace,
219 unsigned long flags,
220 int pc)
221{
222 struct trace_event_call *call = &event_funcgraph_exit;
223 struct ring_buffer_event *event;
224 struct ring_buffer *buffer = tr->trace_buffer.buffer;
225 struct ftrace_graph_ret_entry *entry;
226
227 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
228 sizeof(*entry), flags, pc);
229 if (!event)
230 return;
231 entry = ring_buffer_event_data(event);
232 entry->ret = *trace;
233 if (!call_filter_check_discard(call, entry, buffer, event))
234 trace_buffer_unlock_commit_nostack(buffer, event);
235}
236
237void trace_graph_return(struct ftrace_graph_ret *trace)
238{
239 struct trace_array *tr = graph_array;
240 struct trace_array_cpu *data;
241 unsigned long flags;
242 long disabled;
243 int cpu;
244 int pc;
245
246 ftrace_graph_addr_finish(trace);
247
248 if (trace_recursion_test(TRACE_GRAPH_NOTRACE_BIT)) {
249 trace_recursion_clear(TRACE_GRAPH_NOTRACE_BIT);
250 return;
251 }
252
253 local_irq_save(flags);
254 cpu = raw_smp_processor_id();
255 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
256 disabled = atomic_inc_return(&data->disabled);
257 if (likely(disabled == 1)) {
258 pc = preempt_count();
259 __trace_graph_return(tr, trace, flags, pc);
260 }
261 atomic_dec(&data->disabled);
262 local_irq_restore(flags);
263}
264
265void set_graph_array(struct trace_array *tr)
266{
267 graph_array = tr;
268
269 /* Make graph_array visible before we start tracing */
270
271 smp_mb();
272}
273
274static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
275{
276 ftrace_graph_addr_finish(trace);
277
278 if (trace_recursion_test(TRACE_GRAPH_NOTRACE_BIT)) {
279 trace_recursion_clear(TRACE_GRAPH_NOTRACE_BIT);
280 return;
281 }
282
283 if (tracing_thresh &&
284 (trace->rettime - trace->calltime < tracing_thresh))
285 return;
286 else
287 trace_graph_return(trace);
288}
289
290static struct fgraph_ops funcgraph_thresh_ops = {
291 .entryfunc = &trace_graph_entry,
292 .retfunc = &trace_graph_thresh_return,
293};
294
295static struct fgraph_ops funcgraph_ops = {
296 .entryfunc = &trace_graph_entry,
297 .retfunc = &trace_graph_return,
298};
299
300static int graph_trace_init(struct trace_array *tr)
301{
302 int ret;
303
304 set_graph_array(tr);
305 if (tracing_thresh)
306 ret = register_ftrace_graph(&funcgraph_thresh_ops);
307 else
308 ret = register_ftrace_graph(&funcgraph_ops);
309 if (ret)
310 return ret;
311 tracing_start_cmdline_record();
312
313 return 0;
314}
315
316static void graph_trace_reset(struct trace_array *tr)
317{
318 tracing_stop_cmdline_record();
319 if (tracing_thresh)
320 unregister_ftrace_graph(&funcgraph_thresh_ops);
321 else
322 unregister_ftrace_graph(&funcgraph_ops);
323}
324
325static int graph_trace_update_thresh(struct trace_array *tr)
326{
327 graph_trace_reset(tr);
328 return graph_trace_init(tr);
329}
330
331static int max_bytes_for_cpu;
332
333static void print_graph_cpu(struct trace_seq *s, int cpu)
334{
335 /*
336 * Start with a space character - to make it stand out
337 * to the right a bit when trace output is pasted into
338 * email:
339 */
340 trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
341}
342
343#define TRACE_GRAPH_PROCINFO_LENGTH 14
344
345static void print_graph_proc(struct trace_seq *s, pid_t pid)
346{
347 char comm[TASK_COMM_LEN];
348 /* sign + log10(MAX_INT) + '\0' */
349 char pid_str[11];
350 int spaces = 0;
351 int len;
352 int i;
353
354 trace_find_cmdline(pid, comm);
355 comm[7] = '\0';
356 sprintf(pid_str, "%d", pid);
357
358 /* 1 stands for the "-" character */
359 len = strlen(comm) + strlen(pid_str) + 1;
360
361 if (len < TRACE_GRAPH_PROCINFO_LENGTH)
362 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
363
364 /* First spaces to align center */
365 for (i = 0; i < spaces / 2; i++)
366 trace_seq_putc(s, ' ');
367
368 trace_seq_printf(s, "%s-%s", comm, pid_str);
369
370 /* Last spaces to align center */
371 for (i = 0; i < spaces - (spaces / 2); i++)
372 trace_seq_putc(s, ' ');
373}
374
375
376static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
377{
378 trace_seq_putc(s, ' ');
379 trace_print_lat_fmt(s, entry);
380 trace_seq_puts(s, " | ");
381}
382
383/* If the pid changed since the last trace, output this event */
384static void
385verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
386{
387 pid_t prev_pid;
388 pid_t *last_pid;
389
390 if (!data)
391 return;
392
393 last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
394
395 if (*last_pid == pid)
396 return;
397
398 prev_pid = *last_pid;
399 *last_pid = pid;
400
401 if (prev_pid == -1)
402 return;
403/*
404 * Context-switch trace line:
405
406 ------------------------------------------
407 | 1) migration/0--1 => sshd-1755
408 ------------------------------------------
409
410 */
411 trace_seq_puts(s, " ------------------------------------------\n");
412 print_graph_cpu(s, cpu);
413 print_graph_proc(s, prev_pid);
414 trace_seq_puts(s, " => ");
415 print_graph_proc(s, pid);
416 trace_seq_puts(s, "\n ------------------------------------------\n\n");
417}
418
419static struct ftrace_graph_ret_entry *
420get_return_for_leaf(struct trace_iterator *iter,
421 struct ftrace_graph_ent_entry *curr)
422{
423 struct fgraph_data *data = iter->private;
424 struct ring_buffer_iter *ring_iter = NULL;
425 struct ring_buffer_event *event;
426 struct ftrace_graph_ret_entry *next;
427
428 /*
429 * If the previous output failed to write to the seq buffer,
430 * then we just reuse the data from before.
431 */
432 if (data && data->failed) {
433 curr = &data->ent;
434 next = &data->ret;
435 } else {
436
437 ring_iter = trace_buffer_iter(iter, iter->cpu);
438
439 /* First peek to compare current entry and the next one */
440 if (ring_iter)
441 event = ring_buffer_iter_peek(ring_iter, NULL);
442 else {
443 /*
444 * We need to consume the current entry to see
445 * the next one.
446 */
447 ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
448 NULL, NULL);
449 event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
450 NULL, NULL);
451 }
452
453 if (!event)
454 return NULL;
455
456 next = ring_buffer_event_data(event);
457
458 if (data) {
459 /*
460 * Save current and next entries for later reference
461 * if the output fails.
462 */
463 data->ent = *curr;
464 /*
465 * If the next event is not a return type, then
466 * we only care about what type it is. Otherwise we can
467 * safely copy the entire event.
468 */
469 if (next->ent.type == TRACE_GRAPH_RET)
470 data->ret = *next;
471 else
472 data->ret.ent.type = next->ent.type;
473 }
474 }
475
476 if (next->ent.type != TRACE_GRAPH_RET)
477 return NULL;
478
479 if (curr->ent.pid != next->ent.pid ||
480 curr->graph_ent.func != next->ret.func)
481 return NULL;
482
483 /* this is a leaf, now advance the iterator */
484 if (ring_iter)
485 ring_buffer_read(ring_iter, NULL);
486
487 return next;
488}
489
490static void print_graph_abs_time(u64 t, struct trace_seq *s)
491{
492 unsigned long usecs_rem;
493
494 usecs_rem = do_div(t, NSEC_PER_SEC);
495 usecs_rem /= 1000;
496
497 trace_seq_printf(s, "%5lu.%06lu | ",
498 (unsigned long)t, usecs_rem);
499}
500
501static void
502print_graph_rel_time(struct trace_iterator *iter, struct trace_seq *s)
503{
504 unsigned long long usecs;
505
506 usecs = iter->ts - iter->trace_buffer->time_start;
507 do_div(usecs, NSEC_PER_USEC);
508
509 trace_seq_printf(s, "%9llu us | ", usecs);
510}
511
512static void
513print_graph_irq(struct trace_iterator *iter, unsigned long addr,
514 enum trace_type type, int cpu, pid_t pid, u32 flags)
515{
516 struct trace_array *tr = iter->tr;
517 struct trace_seq *s = &iter->seq;
518 struct trace_entry *ent = iter->ent;
519
520 if (addr < (unsigned long)__irqentry_text_start ||
521 addr >= (unsigned long)__irqentry_text_end)
522 return;
523
524 if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
525 /* Absolute time */
526 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
527 print_graph_abs_time(iter->ts, s);
528
529 /* Relative time */
530 if (flags & TRACE_GRAPH_PRINT_REL_TIME)
531 print_graph_rel_time(iter, s);
532
533 /* Cpu */
534 if (flags & TRACE_GRAPH_PRINT_CPU)
535 print_graph_cpu(s, cpu);
536
537 /* Proc */
538 if (flags & TRACE_GRAPH_PRINT_PROC) {
539 print_graph_proc(s, pid);
540 trace_seq_puts(s, " | ");
541 }
542
543 /* Latency format */
544 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
545 print_graph_lat_fmt(s, ent);
546 }
547
548 /* No overhead */
549 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
550
551 if (type == TRACE_GRAPH_ENT)
552 trace_seq_puts(s, "==========>");
553 else
554 trace_seq_puts(s, "<==========");
555
556 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
557 trace_seq_putc(s, '\n');
558}
559
560void
561trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
562{
563 unsigned long nsecs_rem = do_div(duration, 1000);
564 /* log10(ULONG_MAX) + '\0' */
565 char usecs_str[21];
566 char nsecs_str[5];
567 int len;
568 int i;
569
570 sprintf(usecs_str, "%lu", (unsigned long) duration);
571
572 /* Print msecs */
573 trace_seq_printf(s, "%s", usecs_str);
574
575 len = strlen(usecs_str);
576
577 /* Print nsecs (we don't want to exceed 7 numbers) */
578 if (len < 7) {
579 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
580
581 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
582 trace_seq_printf(s, ".%s", nsecs_str);
583 len += strlen(nsecs_str) + 1;
584 }
585
586 trace_seq_puts(s, " us ");
587
588 /* Print remaining spaces to fit the row's width */
589 for (i = len; i < 8; i++)
590 trace_seq_putc(s, ' ');
591}
592
593static void
594print_graph_duration(struct trace_array *tr, unsigned long long duration,
595 struct trace_seq *s, u32 flags)
596{
597 if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
598 !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
599 return;
600
601 /* No real adata, just filling the column with spaces */
602 switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
603 case FLAGS_FILL_FULL:
604 trace_seq_puts(s, " | ");
605 return;
606 case FLAGS_FILL_START:
607 trace_seq_puts(s, " ");
608 return;
609 case FLAGS_FILL_END:
610 trace_seq_puts(s, " |");
611 return;
612 }
613
614 /* Signal a overhead of time execution to the output */
615 if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
616 trace_seq_printf(s, "%c ", trace_find_mark(duration));
617 else
618 trace_seq_puts(s, " ");
619
620 trace_print_graph_duration(duration, s);
621 trace_seq_puts(s, "| ");
622}
623
624/* Case of a leaf function on its call entry */
625static enum print_line_t
626print_graph_entry_leaf(struct trace_iterator *iter,
627 struct ftrace_graph_ent_entry *entry,
628 struct ftrace_graph_ret_entry *ret_entry,
629 struct trace_seq *s, u32 flags)
630{
631 struct fgraph_data *data = iter->private;
632 struct trace_array *tr = iter->tr;
633 struct ftrace_graph_ret *graph_ret;
634 struct ftrace_graph_ent *call;
635 unsigned long long duration;
636 int cpu = iter->cpu;
637 int i;
638
639 graph_ret = &ret_entry->ret;
640 call = &entry->graph_ent;
641 duration = graph_ret->rettime - graph_ret->calltime;
642
643 if (data) {
644 struct fgraph_cpu_data *cpu_data;
645
646 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
647
648 /*
649 * Comments display at + 1 to depth. Since
650 * this is a leaf function, keep the comments
651 * equal to this depth.
652 */
653 cpu_data->depth = call->depth - 1;
654
655 /* No need to keep this function around for this depth */
656 if (call->depth < FTRACE_RETFUNC_DEPTH &&
657 !WARN_ON_ONCE(call->depth < 0))
658 cpu_data->enter_funcs[call->depth] = 0;
659 }
660
661 /* Overhead and duration */
662 print_graph_duration(tr, duration, s, flags);
663
664 /* Function */
665 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
666 trace_seq_putc(s, ' ');
667
668 trace_seq_printf(s, "%ps();\n", (void *)call->func);
669
670 print_graph_irq(iter, graph_ret->func, TRACE_GRAPH_RET,
671 cpu, iter->ent->pid, flags);
672
673 return trace_handle_return(s);
674}
675
676static enum print_line_t
677print_graph_entry_nested(struct trace_iterator *iter,
678 struct ftrace_graph_ent_entry *entry,
679 struct trace_seq *s, int cpu, u32 flags)
680{
681 struct ftrace_graph_ent *call = &entry->graph_ent;
682 struct fgraph_data *data = iter->private;
683 struct trace_array *tr = iter->tr;
684 int i;
685
686 if (data) {
687 struct fgraph_cpu_data *cpu_data;
688 int cpu = iter->cpu;
689
690 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
691 cpu_data->depth = call->depth;
692
693 /* Save this function pointer to see if the exit matches */
694 if (call->depth < FTRACE_RETFUNC_DEPTH &&
695 !WARN_ON_ONCE(call->depth < 0))
696 cpu_data->enter_funcs[call->depth] = call->func;
697 }
698
699 /* No time */
700 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
701
702 /* Function */
703 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
704 trace_seq_putc(s, ' ');
705
706 trace_seq_printf(s, "%ps() {\n", (void *)call->func);
707
708 if (trace_seq_has_overflowed(s))
709 return TRACE_TYPE_PARTIAL_LINE;
710
711 /*
712 * we already consumed the current entry to check the next one
713 * and see if this is a leaf.
714 */
715 return TRACE_TYPE_NO_CONSUME;
716}
717
718static void
719print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
720 int type, unsigned long addr, u32 flags)
721{
722 struct fgraph_data *data = iter->private;
723 struct trace_entry *ent = iter->ent;
724 struct trace_array *tr = iter->tr;
725 int cpu = iter->cpu;
726
727 /* Pid */
728 verif_pid(s, ent->pid, cpu, data);
729
730 if (type)
731 /* Interrupt */
732 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
733
734 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
735 return;
736
737 /* Absolute time */
738 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
739 print_graph_abs_time(iter->ts, s);
740
741 /* Relative time */
742 if (flags & TRACE_GRAPH_PRINT_REL_TIME)
743 print_graph_rel_time(iter, s);
744
745 /* Cpu */
746 if (flags & TRACE_GRAPH_PRINT_CPU)
747 print_graph_cpu(s, cpu);
748
749 /* Proc */
750 if (flags & TRACE_GRAPH_PRINT_PROC) {
751 print_graph_proc(s, ent->pid);
752 trace_seq_puts(s, " | ");
753 }
754
755 /* Latency format */
756 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
757 print_graph_lat_fmt(s, ent);
758
759 return;
760}
761
762/*
763 * Entry check for irq code
764 *
765 * returns 1 if
766 * - we are inside irq code
767 * - we just entered irq code
768 *
769 * retunns 0 if
770 * - funcgraph-interrupts option is set
771 * - we are not inside irq code
772 */
773static int
774check_irq_entry(struct trace_iterator *iter, u32 flags,
775 unsigned long addr, int depth)
776{
777 int cpu = iter->cpu;
778 int *depth_irq;
779 struct fgraph_data *data = iter->private;
780
781 /*
782 * If we are either displaying irqs, or we got called as
783 * a graph event and private data does not exist,
784 * then we bypass the irq check.
785 */
786 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
787 (!data))
788 return 0;
789
790 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
791
792 /*
793 * We are inside the irq code
794 */
795 if (*depth_irq >= 0)
796 return 1;
797
798 if ((addr < (unsigned long)__irqentry_text_start) ||
799 (addr >= (unsigned long)__irqentry_text_end))
800 return 0;
801
802 /*
803 * We are entering irq code.
804 */
805 *depth_irq = depth;
806 return 1;
807}
808
809/*
810 * Return check for irq code
811 *
812 * returns 1 if
813 * - we are inside irq code
814 * - we just left irq code
815 *
816 * returns 0 if
817 * - funcgraph-interrupts option is set
818 * - we are not inside irq code
819 */
820static int
821check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
822{
823 int cpu = iter->cpu;
824 int *depth_irq;
825 struct fgraph_data *data = iter->private;
826
827 /*
828 * If we are either displaying irqs, or we got called as
829 * a graph event and private data does not exist,
830 * then we bypass the irq check.
831 */
832 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
833 (!data))
834 return 0;
835
836 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
837
838 /*
839 * We are not inside the irq code.
840 */
841 if (*depth_irq == -1)
842 return 0;
843
844 /*
845 * We are inside the irq code, and this is returning entry.
846 * Let's not trace it and clear the entry depth, since
847 * we are out of irq code.
848 *
849 * This condition ensures that we 'leave the irq code' once
850 * we are out of the entry depth. Thus protecting us from
851 * the RETURN entry loss.
852 */
853 if (*depth_irq >= depth) {
854 *depth_irq = -1;
855 return 1;
856 }
857
858 /*
859 * We are inside the irq code, and this is not the entry.
860 */
861 return 1;
862}
863
864static enum print_line_t
865print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
866 struct trace_iterator *iter, u32 flags)
867{
868 struct fgraph_data *data = iter->private;
869 struct ftrace_graph_ent *call = &field->graph_ent;
870 struct ftrace_graph_ret_entry *leaf_ret;
871 static enum print_line_t ret;
872 int cpu = iter->cpu;
873
874 if (check_irq_entry(iter, flags, call->func, call->depth))
875 return TRACE_TYPE_HANDLED;
876
877 print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
878
879 leaf_ret = get_return_for_leaf(iter, field);
880 if (leaf_ret)
881 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
882 else
883 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
884
885 if (data) {
886 /*
887 * If we failed to write our output, then we need to make
888 * note of it. Because we already consumed our entry.
889 */
890 if (s->full) {
891 data->failed = 1;
892 data->cpu = cpu;
893 } else
894 data->failed = 0;
895 }
896
897 return ret;
898}
899
900static enum print_line_t
901print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
902 struct trace_entry *ent, struct trace_iterator *iter,
903 u32 flags)
904{
905 unsigned long long duration = trace->rettime - trace->calltime;
906 struct fgraph_data *data = iter->private;
907 struct trace_array *tr = iter->tr;
908 pid_t pid = ent->pid;
909 int cpu = iter->cpu;
910 int func_match = 1;
911 int i;
912
913 if (check_irq_return(iter, flags, trace->depth))
914 return TRACE_TYPE_HANDLED;
915
916 if (data) {
917 struct fgraph_cpu_data *cpu_data;
918 int cpu = iter->cpu;
919
920 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
921
922 /*
923 * Comments display at + 1 to depth. This is the
924 * return from a function, we now want the comments
925 * to display at the same level of the bracket.
926 */
927 cpu_data->depth = trace->depth - 1;
928
929 if (trace->depth < FTRACE_RETFUNC_DEPTH &&
930 !WARN_ON_ONCE(trace->depth < 0)) {
931 if (cpu_data->enter_funcs[trace->depth] != trace->func)
932 func_match = 0;
933 cpu_data->enter_funcs[trace->depth] = 0;
934 }
935 }
936
937 print_graph_prologue(iter, s, 0, 0, flags);
938
939 /* Overhead and duration */
940 print_graph_duration(tr, duration, s, flags);
941
942 /* Closing brace */
943 for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
944 trace_seq_putc(s, ' ');
945
946 /*
947 * If the return function does not have a matching entry,
948 * then the entry was lost. Instead of just printing
949 * the '}' and letting the user guess what function this
950 * belongs to, write out the function name. Always do
951 * that if the funcgraph-tail option is enabled.
952 */
953 if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
954 trace_seq_puts(s, "}\n");
955 else
956 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
957
958 /* Overrun */
959 if (flags & TRACE_GRAPH_PRINT_OVERRUN)
960 trace_seq_printf(s, " (Overruns: %lu)\n",
961 trace->overrun);
962
963 print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
964 cpu, pid, flags);
965
966 return trace_handle_return(s);
967}
968
969static enum print_line_t
970print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
971 struct trace_iterator *iter, u32 flags)
972{
973 struct trace_array *tr = iter->tr;
974 unsigned long sym_flags = (tr->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 print_graph_prologue(iter, s, 0, 0, flags);
985
986 /* No time */
987 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
988
989 /* Indentation */
990 if (depth > 0)
991 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
992 trace_seq_putc(s, ' ');
993
994 /* The comment */
995 trace_seq_puts(s, "/* ");
996
997 switch (iter->ent->type) {
998 case TRACE_BPUTS:
999 ret = trace_print_bputs_msg_only(iter);
1000 if (ret != TRACE_TYPE_HANDLED)
1001 return ret;
1002 break;
1003 case TRACE_BPRINT:
1004 ret = trace_print_bprintk_msg_only(iter);
1005 if (ret != TRACE_TYPE_HANDLED)
1006 return ret;
1007 break;
1008 case TRACE_PRINT:
1009 ret = trace_print_printk_msg_only(iter);
1010 if (ret != TRACE_TYPE_HANDLED)
1011 return ret;
1012 break;
1013 default:
1014 event = ftrace_find_event(ent->type);
1015 if (!event)
1016 return TRACE_TYPE_UNHANDLED;
1017
1018 ret = event->funcs->trace(iter, sym_flags, event);
1019 if (ret != TRACE_TYPE_HANDLED)
1020 return ret;
1021 }
1022
1023 if (trace_seq_has_overflowed(s))
1024 goto out;
1025
1026 /* Strip ending newline */
1027 if (s->buffer[s->seq.len - 1] == '\n') {
1028 s->buffer[s->seq.len - 1] = '\0';
1029 s->seq.len--;
1030 }
1031
1032 trace_seq_puts(s, " */\n");
1033 out:
1034 return trace_handle_return(s);
1035}
1036
1037
1038enum print_line_t
1039print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1040{
1041 struct ftrace_graph_ent_entry *field;
1042 struct fgraph_data *data = iter->private;
1043 struct trace_entry *entry = iter->ent;
1044 struct trace_seq *s = &iter->seq;
1045 int cpu = iter->cpu;
1046 int ret;
1047
1048 if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1049 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1050 return TRACE_TYPE_HANDLED;
1051 }
1052
1053 /*
1054 * If the last output failed, there's a possibility we need
1055 * to print out the missing entry which would never go out.
1056 */
1057 if (data && data->failed) {
1058 field = &data->ent;
1059 iter->cpu = data->cpu;
1060 ret = print_graph_entry(field, s, iter, flags);
1061 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1062 per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1063 ret = TRACE_TYPE_NO_CONSUME;
1064 }
1065 iter->cpu = cpu;
1066 return ret;
1067 }
1068
1069 switch (entry->type) {
1070 case TRACE_GRAPH_ENT: {
1071 /*
1072 * print_graph_entry() may consume the current event,
1073 * thus @field may become invalid, so we need to save it.
1074 * sizeof(struct ftrace_graph_ent_entry) is very small,
1075 * it can be safely saved at the stack.
1076 */
1077 struct ftrace_graph_ent_entry saved;
1078 trace_assign_type(field, entry);
1079 saved = *field;
1080 return print_graph_entry(&saved, s, iter, flags);
1081 }
1082 case TRACE_GRAPH_RET: {
1083 struct ftrace_graph_ret_entry *field;
1084 trace_assign_type(field, entry);
1085 return print_graph_return(&field->ret, s, entry, iter, flags);
1086 }
1087 case TRACE_STACK:
1088 case TRACE_FN:
1089 /* dont trace stack and functions as comments */
1090 return TRACE_TYPE_UNHANDLED;
1091
1092 default:
1093 return print_graph_comment(s, entry, iter, flags);
1094 }
1095
1096 return TRACE_TYPE_HANDLED;
1097}
1098
1099static enum print_line_t
1100print_graph_function(struct trace_iterator *iter)
1101{
1102 return print_graph_function_flags(iter, tracer_flags.val);
1103}
1104
1105static enum print_line_t
1106print_graph_function_event(struct trace_iterator *iter, int flags,
1107 struct trace_event *event)
1108{
1109 return print_graph_function(iter);
1110}
1111
1112static void print_lat_header(struct seq_file *s, u32 flags)
1113{
1114 static const char spaces[] = " " /* 16 spaces */
1115 " " /* 4 spaces */
1116 " "; /* 17 spaces */
1117 int size = 0;
1118
1119 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1120 size += 16;
1121 if (flags & TRACE_GRAPH_PRINT_REL_TIME)
1122 size += 16;
1123 if (flags & TRACE_GRAPH_PRINT_CPU)
1124 size += 4;
1125 if (flags & TRACE_GRAPH_PRINT_PROC)
1126 size += 17;
1127
1128 seq_printf(s, "#%.*s _-----=> irqs-off \n", size, spaces);
1129 seq_printf(s, "#%.*s / _----=> need-resched \n", size, spaces);
1130 seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1131 seq_printf(s, "#%.*s|| / _--=> preempt-depth \n", size, spaces);
1132 seq_printf(s, "#%.*s||| / \n", size, spaces);
1133}
1134
1135static void __print_graph_headers_flags(struct trace_array *tr,
1136 struct seq_file *s, u32 flags)
1137{
1138 int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1139
1140 if (lat)
1141 print_lat_header(s, flags);
1142
1143 /* 1st line */
1144 seq_putc(s, '#');
1145 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1146 seq_puts(s, " TIME ");
1147 if (flags & TRACE_GRAPH_PRINT_REL_TIME)
1148 seq_puts(s, " REL TIME ");
1149 if (flags & TRACE_GRAPH_PRINT_CPU)
1150 seq_puts(s, " CPU");
1151 if (flags & TRACE_GRAPH_PRINT_PROC)
1152 seq_puts(s, " TASK/PID ");
1153 if (lat)
1154 seq_puts(s, "|||| ");
1155 if (flags & TRACE_GRAPH_PRINT_DURATION)
1156 seq_puts(s, " DURATION ");
1157 seq_puts(s, " FUNCTION CALLS\n");
1158
1159 /* 2nd line */
1160 seq_putc(s, '#');
1161 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1162 seq_puts(s, " | ");
1163 if (flags & TRACE_GRAPH_PRINT_REL_TIME)
1164 seq_puts(s, " | ");
1165 if (flags & TRACE_GRAPH_PRINT_CPU)
1166 seq_puts(s, " | ");
1167 if (flags & TRACE_GRAPH_PRINT_PROC)
1168 seq_puts(s, " | | ");
1169 if (lat)
1170 seq_puts(s, "|||| ");
1171 if (flags & TRACE_GRAPH_PRINT_DURATION)
1172 seq_puts(s, " | | ");
1173 seq_puts(s, " | | | |\n");
1174}
1175
1176static void print_graph_headers(struct seq_file *s)
1177{
1178 print_graph_headers_flags(s, tracer_flags.val);
1179}
1180
1181void print_graph_headers_flags(struct seq_file *s, u32 flags)
1182{
1183 struct trace_iterator *iter = s->private;
1184 struct trace_array *tr = iter->tr;
1185
1186 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1187 return;
1188
1189 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1190 /* print nothing if the buffers are empty */
1191 if (trace_empty(iter))
1192 return;
1193
1194 print_trace_header(s, iter);
1195 }
1196
1197 __print_graph_headers_flags(tr, s, flags);
1198}
1199
1200void graph_trace_open(struct trace_iterator *iter)
1201{
1202 /* pid and depth on the last trace processed */
1203 struct fgraph_data *data;
1204 gfp_t gfpflags;
1205 int cpu;
1206
1207 iter->private = NULL;
1208
1209 /* We can be called in atomic context via ftrace_dump() */
1210 gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1211
1212 data = kzalloc(sizeof(*data), gfpflags);
1213 if (!data)
1214 goto out_err;
1215
1216 data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1217 if (!data->cpu_data)
1218 goto out_err_free;
1219
1220 for_each_possible_cpu(cpu) {
1221 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1222 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1223 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1224 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1225
1226 *pid = -1;
1227 *depth = 0;
1228 *ignore = 0;
1229 *depth_irq = -1;
1230 }
1231
1232 iter->private = data;
1233
1234 return;
1235
1236 out_err_free:
1237 kfree(data);
1238 out_err:
1239 pr_warn("function graph tracer: not enough memory\n");
1240}
1241
1242void graph_trace_close(struct trace_iterator *iter)
1243{
1244 struct fgraph_data *data = iter->private;
1245
1246 if (data) {
1247 free_percpu(data->cpu_data);
1248 kfree(data);
1249 }
1250}
1251
1252static int
1253func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1254{
1255 if (bit == TRACE_GRAPH_PRINT_IRQS)
1256 ftrace_graph_skip_irqs = !set;
1257
1258 if (bit == TRACE_GRAPH_SLEEP_TIME)
1259 ftrace_graph_sleep_time_control(set);
1260
1261 if (bit == TRACE_GRAPH_GRAPH_TIME)
1262 ftrace_graph_graph_time_control(set);
1263
1264 return 0;
1265}
1266
1267static struct trace_event_functions graph_functions = {
1268 .trace = print_graph_function_event,
1269};
1270
1271static struct trace_event graph_trace_entry_event = {
1272 .type = TRACE_GRAPH_ENT,
1273 .funcs = &graph_functions,
1274};
1275
1276static struct trace_event graph_trace_ret_event = {
1277 .type = TRACE_GRAPH_RET,
1278 .funcs = &graph_functions
1279};
1280
1281static struct tracer graph_trace __tracer_data = {
1282 .name = "function_graph",
1283 .update_thresh = graph_trace_update_thresh,
1284 .open = graph_trace_open,
1285 .pipe_open = graph_trace_open,
1286 .close = graph_trace_close,
1287 .pipe_close = graph_trace_close,
1288 .init = graph_trace_init,
1289 .reset = graph_trace_reset,
1290 .print_line = print_graph_function,
1291 .print_header = print_graph_headers,
1292 .flags = &tracer_flags,
1293 .set_flag = func_graph_set_flag,
1294#ifdef CONFIG_FTRACE_SELFTEST
1295 .selftest = trace_selftest_startup_function_graph,
1296#endif
1297};
1298
1299
1300static ssize_t
1301graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1302 loff_t *ppos)
1303{
1304 unsigned long val;
1305 int ret;
1306
1307 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1308 if (ret)
1309 return ret;
1310
1311 fgraph_max_depth = val;
1312
1313 *ppos += cnt;
1314
1315 return cnt;
1316}
1317
1318static ssize_t
1319graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1320 loff_t *ppos)
1321{
1322 char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1323 int n;
1324
1325 n = sprintf(buf, "%d\n", fgraph_max_depth);
1326
1327 return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1328}
1329
1330static const struct file_operations graph_depth_fops = {
1331 .open = tracing_open_generic,
1332 .write = graph_depth_write,
1333 .read = graph_depth_read,
1334 .llseek = generic_file_llseek,
1335};
1336
1337static __init int init_graph_tracefs(void)
1338{
1339 struct dentry *d_tracer;
1340
1341 d_tracer = tracing_init_dentry();
1342 if (IS_ERR(d_tracer))
1343 return 0;
1344
1345 trace_create_file("max_graph_depth", 0644, d_tracer,
1346 NULL, &graph_depth_fops);
1347
1348 return 0;
1349}
1350fs_initcall(init_graph_tracefs);
1351
1352static __init int init_graph_trace(void)
1353{
1354 max_bytes_for_cpu = snprintf(NULL, 0, "%u", nr_cpu_ids - 1);
1355
1356 if (!register_trace_event(&graph_trace_entry_event)) {
1357 pr_warn("Warning: could not register graph trace events\n");
1358 return 1;
1359 }
1360
1361 if (!register_trace_event(&graph_trace_ret_event)) {
1362 pr_warn("Warning: could not register graph trace events\n");
1363 return 1;
1364 }
1365
1366 return register_tracer(&graph_trace);
1367}
1368
1369core_initcall(init_graph_trace);