Loading...
1/*
2 * trace task wakeup timings
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6 *
7 * Based on code from the latency_tracer, that is:
8 *
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 Nadia Yvette Chambers
11 */
12#include <linux/module.h>
13#include <linux/kallsyms.h>
14#include <linux/uaccess.h>
15#include <linux/ftrace.h>
16#include <linux/sched/rt.h>
17#include <linux/sched/deadline.h>
18#include <trace/events/sched.h>
19#include "trace.h"
20
21static struct trace_array *wakeup_trace;
22static int __read_mostly tracer_enabled;
23
24static struct task_struct *wakeup_task;
25static int wakeup_cpu;
26static int wakeup_current_cpu;
27static unsigned wakeup_prio = -1;
28static int wakeup_rt;
29static int wakeup_dl;
30static int tracing_dl = 0;
31
32static arch_spinlock_t wakeup_lock =
33 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
34
35static void wakeup_reset(struct trace_array *tr);
36static void __wakeup_reset(struct trace_array *tr);
37
38static int save_flags;
39
40#ifdef CONFIG_FUNCTION_GRAPH_TRACER
41static int wakeup_display_graph(struct trace_array *tr, int set);
42# define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
43#else
44static inline int wakeup_display_graph(struct trace_array *tr, int set)
45{
46 return 0;
47}
48# define is_graph(tr) false
49#endif
50
51
52#ifdef CONFIG_FUNCTION_TRACER
53
54static int wakeup_graph_entry(struct ftrace_graph_ent *trace);
55static void wakeup_graph_return(struct ftrace_graph_ret *trace);
56
57static bool function_enabled;
58
59/*
60 * Prologue for the wakeup function tracers.
61 *
62 * Returns 1 if it is OK to continue, and preemption
63 * is disabled and data->disabled is incremented.
64 * 0 if the trace is to be ignored, and preemption
65 * is not disabled and data->disabled is
66 * kept the same.
67 *
68 * Note, this function is also used outside this ifdef but
69 * inside the #ifdef of the function graph tracer below.
70 * This is OK, since the function graph tracer is
71 * dependent on the function tracer.
72 */
73static int
74func_prolog_preempt_disable(struct trace_array *tr,
75 struct trace_array_cpu **data,
76 int *pc)
77{
78 long disabled;
79 int cpu;
80
81 if (likely(!wakeup_task))
82 return 0;
83
84 *pc = preempt_count();
85 preempt_disable_notrace();
86
87 cpu = raw_smp_processor_id();
88 if (cpu != wakeup_current_cpu)
89 goto out_enable;
90
91 *data = per_cpu_ptr(tr->trace_buffer.data, cpu);
92 disabled = atomic_inc_return(&(*data)->disabled);
93 if (unlikely(disabled != 1))
94 goto out;
95
96 return 1;
97
98out:
99 atomic_dec(&(*data)->disabled);
100
101out_enable:
102 preempt_enable_notrace();
103 return 0;
104}
105
106/*
107 * wakeup uses its own tracer function to keep the overhead down:
108 */
109static void
110wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
111 struct ftrace_ops *op, struct pt_regs *pt_regs)
112{
113 struct trace_array *tr = wakeup_trace;
114 struct trace_array_cpu *data;
115 unsigned long flags;
116 int pc;
117
118 if (!func_prolog_preempt_disable(tr, &data, &pc))
119 return;
120
121 local_irq_save(flags);
122 trace_function(tr, ip, parent_ip, flags, pc);
123 local_irq_restore(flags);
124
125 atomic_dec(&data->disabled);
126 preempt_enable_notrace();
127}
128
129static int register_wakeup_function(struct trace_array *tr, int graph, int set)
130{
131 int ret;
132
133 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
134 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
135 return 0;
136
137 if (graph)
138 ret = register_ftrace_graph(&wakeup_graph_return,
139 &wakeup_graph_entry);
140 else
141 ret = register_ftrace_function(tr->ops);
142
143 if (!ret)
144 function_enabled = true;
145
146 return ret;
147}
148
149static void unregister_wakeup_function(struct trace_array *tr, int graph)
150{
151 if (!function_enabled)
152 return;
153
154 if (graph)
155 unregister_ftrace_graph();
156 else
157 unregister_ftrace_function(tr->ops);
158
159 function_enabled = false;
160}
161
162static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
163{
164 if (!(mask & TRACE_ITER_FUNCTION))
165 return 0;
166
167 if (set)
168 register_wakeup_function(tr, is_graph(tr), 1);
169 else
170 unregister_wakeup_function(tr, is_graph(tr));
171 return 1;
172}
173#else
174static int register_wakeup_function(struct trace_array *tr, int graph, int set)
175{
176 return 0;
177}
178static void unregister_wakeup_function(struct trace_array *tr, int graph) { }
179static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
180{
181 return 0;
182}
183#endif /* CONFIG_FUNCTION_TRACER */
184
185static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
186{
187 struct tracer *tracer = tr->current_trace;
188
189 if (wakeup_function_set(tr, mask, set))
190 return 0;
191
192#ifdef CONFIG_FUNCTION_GRAPH_TRACER
193 if (mask & TRACE_ITER_DISPLAY_GRAPH)
194 return wakeup_display_graph(tr, set);
195#endif
196
197 return trace_keep_overwrite(tracer, mask, set);
198}
199
200static int start_func_tracer(struct trace_array *tr, int graph)
201{
202 int ret;
203
204 ret = register_wakeup_function(tr, graph, 0);
205
206 if (!ret && tracing_is_enabled())
207 tracer_enabled = 1;
208 else
209 tracer_enabled = 0;
210
211 return ret;
212}
213
214static void stop_func_tracer(struct trace_array *tr, int graph)
215{
216 tracer_enabled = 0;
217
218 unregister_wakeup_function(tr, graph);
219}
220
221#ifdef CONFIG_FUNCTION_GRAPH_TRACER
222static int wakeup_display_graph(struct trace_array *tr, int set)
223{
224 if (!(is_graph(tr) ^ set))
225 return 0;
226
227 stop_func_tracer(tr, !set);
228
229 wakeup_reset(wakeup_trace);
230 tr->max_latency = 0;
231
232 return start_func_tracer(tr, set);
233}
234
235static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
236{
237 struct trace_array *tr = wakeup_trace;
238 struct trace_array_cpu *data;
239 unsigned long flags;
240 int pc, ret = 0;
241
242 if (ftrace_graph_ignore_func(trace))
243 return 0;
244 /*
245 * Do not trace a function if it's filtered by set_graph_notrace.
246 * Make the index of ret stack negative to indicate that it should
247 * ignore further functions. But it needs its own ret stack entry
248 * to recover the original index in order to continue tracing after
249 * returning from the function.
250 */
251 if (ftrace_graph_notrace_addr(trace->func))
252 return 1;
253
254 if (!func_prolog_preempt_disable(tr, &data, &pc))
255 return 0;
256
257 local_save_flags(flags);
258 ret = __trace_graph_entry(tr, trace, flags, pc);
259 atomic_dec(&data->disabled);
260 preempt_enable_notrace();
261
262 return ret;
263}
264
265static void wakeup_graph_return(struct ftrace_graph_ret *trace)
266{
267 struct trace_array *tr = wakeup_trace;
268 struct trace_array_cpu *data;
269 unsigned long flags;
270 int pc;
271
272 if (!func_prolog_preempt_disable(tr, &data, &pc))
273 return;
274
275 local_save_flags(flags);
276 __trace_graph_return(tr, trace, flags, pc);
277 atomic_dec(&data->disabled);
278
279 preempt_enable_notrace();
280 return;
281}
282
283static void wakeup_trace_open(struct trace_iterator *iter)
284{
285 if (is_graph(iter->tr))
286 graph_trace_open(iter);
287}
288
289static void wakeup_trace_close(struct trace_iterator *iter)
290{
291 if (iter->private)
292 graph_trace_close(iter);
293}
294
295#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
296 TRACE_GRAPH_PRINT_ABS_TIME | \
297 TRACE_GRAPH_PRINT_DURATION)
298
299static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
300{
301 /*
302 * In graph mode call the graph tracer output function,
303 * otherwise go with the TRACE_FN event handler
304 */
305 if (is_graph(iter->tr))
306 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
307
308 return TRACE_TYPE_UNHANDLED;
309}
310
311static void wakeup_print_header(struct seq_file *s)
312{
313 if (is_graph(wakeup_trace))
314 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
315 else
316 trace_default_header(s);
317}
318
319static void
320__trace_function(struct trace_array *tr,
321 unsigned long ip, unsigned long parent_ip,
322 unsigned long flags, int pc)
323{
324 if (is_graph(tr))
325 trace_graph_function(tr, ip, parent_ip, flags, pc);
326 else
327 trace_function(tr, ip, parent_ip, flags, pc);
328}
329#else
330#define __trace_function trace_function
331
332static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
333{
334 return TRACE_TYPE_UNHANDLED;
335}
336
337static void wakeup_trace_open(struct trace_iterator *iter) { }
338static void wakeup_trace_close(struct trace_iterator *iter) { }
339
340#ifdef CONFIG_FUNCTION_TRACER
341static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
342{
343 return -1;
344}
345static void wakeup_graph_return(struct ftrace_graph_ret *trace) { }
346static void wakeup_print_header(struct seq_file *s)
347{
348 trace_default_header(s);
349}
350#else
351static void wakeup_print_header(struct seq_file *s)
352{
353 trace_latency_header(s);
354}
355#endif /* CONFIG_FUNCTION_TRACER */
356#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
357
358/*
359 * Should this new latency be reported/recorded?
360 */
361static bool report_latency(struct trace_array *tr, u64 delta)
362{
363 if (tracing_thresh) {
364 if (delta < tracing_thresh)
365 return false;
366 } else {
367 if (delta <= tr->max_latency)
368 return false;
369 }
370 return true;
371}
372
373static void
374probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
375{
376 if (task != wakeup_task)
377 return;
378
379 wakeup_current_cpu = cpu;
380}
381
382static void
383tracing_sched_switch_trace(struct trace_array *tr,
384 struct task_struct *prev,
385 struct task_struct *next,
386 unsigned long flags, int pc)
387{
388 struct trace_event_call *call = &event_context_switch;
389 struct ring_buffer *buffer = tr->trace_buffer.buffer;
390 struct ring_buffer_event *event;
391 struct ctx_switch_entry *entry;
392
393 event = trace_buffer_lock_reserve(buffer, TRACE_CTX,
394 sizeof(*entry), flags, pc);
395 if (!event)
396 return;
397 entry = ring_buffer_event_data(event);
398 entry->prev_pid = prev->pid;
399 entry->prev_prio = prev->prio;
400 entry->prev_state = prev->state;
401 entry->next_pid = next->pid;
402 entry->next_prio = next->prio;
403 entry->next_state = next->state;
404 entry->next_cpu = task_cpu(next);
405
406 if (!call_filter_check_discard(call, entry, buffer, event))
407 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
408}
409
410static void
411tracing_sched_wakeup_trace(struct trace_array *tr,
412 struct task_struct *wakee,
413 struct task_struct *curr,
414 unsigned long flags, int pc)
415{
416 struct trace_event_call *call = &event_wakeup;
417 struct ring_buffer_event *event;
418 struct ctx_switch_entry *entry;
419 struct ring_buffer *buffer = tr->trace_buffer.buffer;
420
421 event = trace_buffer_lock_reserve(buffer, TRACE_WAKE,
422 sizeof(*entry), flags, pc);
423 if (!event)
424 return;
425 entry = ring_buffer_event_data(event);
426 entry->prev_pid = curr->pid;
427 entry->prev_prio = curr->prio;
428 entry->prev_state = curr->state;
429 entry->next_pid = wakee->pid;
430 entry->next_prio = wakee->prio;
431 entry->next_state = wakee->state;
432 entry->next_cpu = task_cpu(wakee);
433
434 if (!call_filter_check_discard(call, entry, buffer, event))
435 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
436}
437
438static void notrace
439probe_wakeup_sched_switch(void *ignore, bool preempt,
440 struct task_struct *prev, struct task_struct *next)
441{
442 struct trace_array_cpu *data;
443 u64 T0, T1, delta;
444 unsigned long flags;
445 long disabled;
446 int cpu;
447 int pc;
448
449 tracing_record_cmdline(prev);
450
451 if (unlikely(!tracer_enabled))
452 return;
453
454 /*
455 * When we start a new trace, we set wakeup_task to NULL
456 * and then set tracer_enabled = 1. We want to make sure
457 * that another CPU does not see the tracer_enabled = 1
458 * and the wakeup_task with an older task, that might
459 * actually be the same as next.
460 */
461 smp_rmb();
462
463 if (next != wakeup_task)
464 return;
465
466 pc = preempt_count();
467
468 /* disable local data, not wakeup_cpu data */
469 cpu = raw_smp_processor_id();
470 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
471 if (likely(disabled != 1))
472 goto out;
473
474 local_irq_save(flags);
475 arch_spin_lock(&wakeup_lock);
476
477 /* We could race with grabbing wakeup_lock */
478 if (unlikely(!tracer_enabled || next != wakeup_task))
479 goto out_unlock;
480
481 /* The task we are waiting for is waking up */
482 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
483
484 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
485 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
486
487 T0 = data->preempt_timestamp;
488 T1 = ftrace_now(cpu);
489 delta = T1-T0;
490
491 if (!report_latency(wakeup_trace, delta))
492 goto out_unlock;
493
494 if (likely(!is_tracing_stopped())) {
495 wakeup_trace->max_latency = delta;
496 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
497 }
498
499out_unlock:
500 __wakeup_reset(wakeup_trace);
501 arch_spin_unlock(&wakeup_lock);
502 local_irq_restore(flags);
503out:
504 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
505}
506
507static void __wakeup_reset(struct trace_array *tr)
508{
509 wakeup_cpu = -1;
510 wakeup_prio = -1;
511 tracing_dl = 0;
512
513 if (wakeup_task)
514 put_task_struct(wakeup_task);
515
516 wakeup_task = NULL;
517}
518
519static void wakeup_reset(struct trace_array *tr)
520{
521 unsigned long flags;
522
523 tracing_reset_online_cpus(&tr->trace_buffer);
524
525 local_irq_save(flags);
526 arch_spin_lock(&wakeup_lock);
527 __wakeup_reset(tr);
528 arch_spin_unlock(&wakeup_lock);
529 local_irq_restore(flags);
530}
531
532static void
533probe_wakeup(void *ignore, struct task_struct *p)
534{
535 struct trace_array_cpu *data;
536 int cpu = smp_processor_id();
537 unsigned long flags;
538 long disabled;
539 int pc;
540
541 if (likely(!tracer_enabled))
542 return;
543
544 tracing_record_cmdline(p);
545 tracing_record_cmdline(current);
546
547 /*
548 * Semantic is like this:
549 * - wakeup tracer handles all tasks in the system, independently
550 * from their scheduling class;
551 * - wakeup_rt tracer handles tasks belonging to sched_dl and
552 * sched_rt class;
553 * - wakeup_dl handles tasks belonging to sched_dl class only.
554 */
555 if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
556 (wakeup_rt && !dl_task(p) && !rt_task(p)) ||
557 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
558 return;
559
560 pc = preempt_count();
561 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
562 if (unlikely(disabled != 1))
563 goto out;
564
565 /* interrupts should be off from try_to_wake_up */
566 arch_spin_lock(&wakeup_lock);
567
568 /* check for races. */
569 if (!tracer_enabled || tracing_dl ||
570 (!dl_task(p) && p->prio >= wakeup_prio))
571 goto out_locked;
572
573 /* reset the trace */
574 __wakeup_reset(wakeup_trace);
575
576 wakeup_cpu = task_cpu(p);
577 wakeup_current_cpu = wakeup_cpu;
578 wakeup_prio = p->prio;
579
580 /*
581 * Once you start tracing a -deadline task, don't bother tracing
582 * another task until the first one wakes up.
583 */
584 if (dl_task(p))
585 tracing_dl = 1;
586 else
587 tracing_dl = 0;
588
589 wakeup_task = p;
590 get_task_struct(wakeup_task);
591
592 local_save_flags(flags);
593
594 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
595 data->preempt_timestamp = ftrace_now(cpu);
596 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
597
598 /*
599 * We must be careful in using CALLER_ADDR2. But since wake_up
600 * is not called by an assembly function (where as schedule is)
601 * it should be safe to use it here.
602 */
603 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
604
605out_locked:
606 arch_spin_unlock(&wakeup_lock);
607out:
608 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
609}
610
611static void start_wakeup_tracer(struct trace_array *tr)
612{
613 int ret;
614
615 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
616 if (ret) {
617 pr_info("wakeup trace: Couldn't activate tracepoint"
618 " probe to kernel_sched_wakeup\n");
619 return;
620 }
621
622 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
623 if (ret) {
624 pr_info("wakeup trace: Couldn't activate tracepoint"
625 " probe to kernel_sched_wakeup_new\n");
626 goto fail_deprobe;
627 }
628
629 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
630 if (ret) {
631 pr_info("sched trace: Couldn't activate tracepoint"
632 " probe to kernel_sched_switch\n");
633 goto fail_deprobe_wake_new;
634 }
635
636 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
637 if (ret) {
638 pr_info("wakeup trace: Couldn't activate tracepoint"
639 " probe to kernel_sched_migrate_task\n");
640 return;
641 }
642
643 wakeup_reset(tr);
644
645 /*
646 * Don't let the tracer_enabled = 1 show up before
647 * the wakeup_task is reset. This may be overkill since
648 * wakeup_reset does a spin_unlock after setting the
649 * wakeup_task to NULL, but I want to be safe.
650 * This is a slow path anyway.
651 */
652 smp_wmb();
653
654 if (start_func_tracer(tr, is_graph(tr)))
655 printk(KERN_ERR "failed to start wakeup tracer\n");
656
657 return;
658fail_deprobe_wake_new:
659 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
660fail_deprobe:
661 unregister_trace_sched_wakeup(probe_wakeup, NULL);
662}
663
664static void stop_wakeup_tracer(struct trace_array *tr)
665{
666 tracer_enabled = 0;
667 stop_func_tracer(tr, is_graph(tr));
668 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
669 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
670 unregister_trace_sched_wakeup(probe_wakeup, NULL);
671 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
672}
673
674static bool wakeup_busy;
675
676static int __wakeup_tracer_init(struct trace_array *tr)
677{
678 save_flags = tr->trace_flags;
679
680 /* non overwrite screws up the latency tracers */
681 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
682 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
683
684 tr->max_latency = 0;
685 wakeup_trace = tr;
686 ftrace_init_array_ops(tr, wakeup_tracer_call);
687 start_wakeup_tracer(tr);
688
689 wakeup_busy = true;
690 return 0;
691}
692
693static int wakeup_tracer_init(struct trace_array *tr)
694{
695 if (wakeup_busy)
696 return -EBUSY;
697
698 wakeup_dl = 0;
699 wakeup_rt = 0;
700 return __wakeup_tracer_init(tr);
701}
702
703static int wakeup_rt_tracer_init(struct trace_array *tr)
704{
705 if (wakeup_busy)
706 return -EBUSY;
707
708 wakeup_dl = 0;
709 wakeup_rt = 1;
710 return __wakeup_tracer_init(tr);
711}
712
713static int wakeup_dl_tracer_init(struct trace_array *tr)
714{
715 if (wakeup_busy)
716 return -EBUSY;
717
718 wakeup_dl = 1;
719 wakeup_rt = 0;
720 return __wakeup_tracer_init(tr);
721}
722
723static void wakeup_tracer_reset(struct trace_array *tr)
724{
725 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
726 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
727
728 stop_wakeup_tracer(tr);
729 /* make sure we put back any tasks we are tracing */
730 wakeup_reset(tr);
731
732 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
733 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
734 ftrace_reset_array_ops(tr);
735 wakeup_busy = false;
736}
737
738static void wakeup_tracer_start(struct trace_array *tr)
739{
740 wakeup_reset(tr);
741 tracer_enabled = 1;
742}
743
744static void wakeup_tracer_stop(struct trace_array *tr)
745{
746 tracer_enabled = 0;
747}
748
749static struct tracer wakeup_tracer __read_mostly =
750{
751 .name = "wakeup",
752 .init = wakeup_tracer_init,
753 .reset = wakeup_tracer_reset,
754 .start = wakeup_tracer_start,
755 .stop = wakeup_tracer_stop,
756 .print_max = true,
757 .print_header = wakeup_print_header,
758 .print_line = wakeup_print_line,
759 .flag_changed = wakeup_flag_changed,
760#ifdef CONFIG_FTRACE_SELFTEST
761 .selftest = trace_selftest_startup_wakeup,
762#endif
763 .open = wakeup_trace_open,
764 .close = wakeup_trace_close,
765 .allow_instances = true,
766 .use_max_tr = true,
767};
768
769static struct tracer wakeup_rt_tracer __read_mostly =
770{
771 .name = "wakeup_rt",
772 .init = wakeup_rt_tracer_init,
773 .reset = wakeup_tracer_reset,
774 .start = wakeup_tracer_start,
775 .stop = wakeup_tracer_stop,
776 .print_max = true,
777 .print_header = wakeup_print_header,
778 .print_line = wakeup_print_line,
779 .flag_changed = wakeup_flag_changed,
780#ifdef CONFIG_FTRACE_SELFTEST
781 .selftest = trace_selftest_startup_wakeup,
782#endif
783 .open = wakeup_trace_open,
784 .close = wakeup_trace_close,
785 .allow_instances = true,
786 .use_max_tr = true,
787};
788
789static struct tracer wakeup_dl_tracer __read_mostly =
790{
791 .name = "wakeup_dl",
792 .init = wakeup_dl_tracer_init,
793 .reset = wakeup_tracer_reset,
794 .start = wakeup_tracer_start,
795 .stop = wakeup_tracer_stop,
796 .print_max = true,
797 .print_header = wakeup_print_header,
798 .print_line = wakeup_print_line,
799 .flag_changed = wakeup_flag_changed,
800#ifdef CONFIG_FTRACE_SELFTEST
801 .selftest = trace_selftest_startup_wakeup,
802#endif
803 .open = wakeup_trace_open,
804 .close = wakeup_trace_close,
805 .allow_instances = true,
806 .use_max_tr = true,
807};
808
809__init static int init_wakeup_tracer(void)
810{
811 int ret;
812
813 ret = register_tracer(&wakeup_tracer);
814 if (ret)
815 return ret;
816
817 ret = register_tracer(&wakeup_rt_tracer);
818 if (ret)
819 return ret;
820
821 ret = register_tracer(&wakeup_dl_tracer);
822 if (ret)
823 return ret;
824
825 return 0;
826}
827core_initcall(init_wakeup_tracer);
1/*
2 * trace task wakeup timings
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6 *
7 * Based on code from the latency_tracer, that is:
8 *
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 William Lee Irwin III
11 */
12#include <linux/module.h>
13#include <linux/fs.h>
14#include <linux/debugfs.h>
15#include <linux/kallsyms.h>
16#include <linux/uaccess.h>
17#include <linux/ftrace.h>
18#include <trace/events/sched.h>
19
20#include "trace.h"
21
22static struct trace_array *wakeup_trace;
23static int __read_mostly tracer_enabled;
24
25static struct task_struct *wakeup_task;
26static int wakeup_cpu;
27static int wakeup_current_cpu;
28static unsigned wakeup_prio = -1;
29static int wakeup_rt;
30
31static arch_spinlock_t wakeup_lock =
32 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
33
34static void wakeup_reset(struct trace_array *tr);
35static void __wakeup_reset(struct trace_array *tr);
36static int wakeup_graph_entry(struct ftrace_graph_ent *trace);
37static void wakeup_graph_return(struct ftrace_graph_ret *trace);
38
39static int save_lat_flag;
40
41#define TRACE_DISPLAY_GRAPH 1
42
43static struct tracer_opt trace_opts[] = {
44#ifdef CONFIG_FUNCTION_GRAPH_TRACER
45 /* display latency trace as call graph */
46 { TRACER_OPT(display-graph, TRACE_DISPLAY_GRAPH) },
47#endif
48 { } /* Empty entry */
49};
50
51static struct tracer_flags tracer_flags = {
52 .val = 0,
53 .opts = trace_opts,
54};
55
56#define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH)
57
58#ifdef CONFIG_FUNCTION_TRACER
59
60/*
61 * Prologue for the wakeup function tracers.
62 *
63 * Returns 1 if it is OK to continue, and preemption
64 * is disabled and data->disabled is incremented.
65 * 0 if the trace is to be ignored, and preemption
66 * is not disabled and data->disabled is
67 * kept the same.
68 *
69 * Note, this function is also used outside this ifdef but
70 * inside the #ifdef of the function graph tracer below.
71 * This is OK, since the function graph tracer is
72 * dependent on the function tracer.
73 */
74static int
75func_prolog_preempt_disable(struct trace_array *tr,
76 struct trace_array_cpu **data,
77 int *pc)
78{
79 long disabled;
80 int cpu;
81
82 if (likely(!wakeup_task))
83 return 0;
84
85 *pc = preempt_count();
86 preempt_disable_notrace();
87
88 cpu = raw_smp_processor_id();
89 if (cpu != wakeup_current_cpu)
90 goto out_enable;
91
92 *data = tr->data[cpu];
93 disabled = atomic_inc_return(&(*data)->disabled);
94 if (unlikely(disabled != 1))
95 goto out;
96
97 return 1;
98
99out:
100 atomic_dec(&(*data)->disabled);
101
102out_enable:
103 preempt_enable_notrace();
104 return 0;
105}
106
107/*
108 * wakeup uses its own tracer function to keep the overhead down:
109 */
110static void
111wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
112{
113 struct trace_array *tr = wakeup_trace;
114 struct trace_array_cpu *data;
115 unsigned long flags;
116 int pc;
117
118 if (!func_prolog_preempt_disable(tr, &data, &pc))
119 return;
120
121 local_irq_save(flags);
122 trace_function(tr, ip, parent_ip, flags, pc);
123 local_irq_restore(flags);
124
125 atomic_dec(&data->disabled);
126 preempt_enable_notrace();
127}
128
129static struct ftrace_ops trace_ops __read_mostly =
130{
131 .func = wakeup_tracer_call,
132 .flags = FTRACE_OPS_FL_GLOBAL,
133};
134#endif /* CONFIG_FUNCTION_TRACER */
135
136static int start_func_tracer(int graph)
137{
138 int ret;
139
140 if (!graph)
141 ret = register_ftrace_function(&trace_ops);
142 else
143 ret = register_ftrace_graph(&wakeup_graph_return,
144 &wakeup_graph_entry);
145
146 if (!ret && tracing_is_enabled())
147 tracer_enabled = 1;
148 else
149 tracer_enabled = 0;
150
151 return ret;
152}
153
154static void stop_func_tracer(int graph)
155{
156 tracer_enabled = 0;
157
158 if (!graph)
159 unregister_ftrace_function(&trace_ops);
160 else
161 unregister_ftrace_graph();
162}
163
164#ifdef CONFIG_FUNCTION_GRAPH_TRACER
165static int wakeup_set_flag(u32 old_flags, u32 bit, int set)
166{
167
168 if (!(bit & TRACE_DISPLAY_GRAPH))
169 return -EINVAL;
170
171 if (!(is_graph() ^ set))
172 return 0;
173
174 stop_func_tracer(!set);
175
176 wakeup_reset(wakeup_trace);
177 tracing_max_latency = 0;
178
179 return start_func_tracer(set);
180}
181
182static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
183{
184 struct trace_array *tr = wakeup_trace;
185 struct trace_array_cpu *data;
186 unsigned long flags;
187 int pc, ret = 0;
188
189 if (!func_prolog_preempt_disable(tr, &data, &pc))
190 return 0;
191
192 local_save_flags(flags);
193 ret = __trace_graph_entry(tr, trace, flags, pc);
194 atomic_dec(&data->disabled);
195 preempt_enable_notrace();
196
197 return ret;
198}
199
200static void wakeup_graph_return(struct ftrace_graph_ret *trace)
201{
202 struct trace_array *tr = wakeup_trace;
203 struct trace_array_cpu *data;
204 unsigned long flags;
205 int pc;
206
207 if (!func_prolog_preempt_disable(tr, &data, &pc))
208 return;
209
210 local_save_flags(flags);
211 __trace_graph_return(tr, trace, flags, pc);
212 atomic_dec(&data->disabled);
213
214 preempt_enable_notrace();
215 return;
216}
217
218static void wakeup_trace_open(struct trace_iterator *iter)
219{
220 if (is_graph())
221 graph_trace_open(iter);
222}
223
224static void wakeup_trace_close(struct trace_iterator *iter)
225{
226 if (iter->private)
227 graph_trace_close(iter);
228}
229
230#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
231 TRACE_GRAPH_PRINT_ABS_TIME | \
232 TRACE_GRAPH_PRINT_DURATION)
233
234static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
235{
236 /*
237 * In graph mode call the graph tracer output function,
238 * otherwise go with the TRACE_FN event handler
239 */
240 if (is_graph())
241 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
242
243 return TRACE_TYPE_UNHANDLED;
244}
245
246static void wakeup_print_header(struct seq_file *s)
247{
248 if (is_graph())
249 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
250 else
251 trace_default_header(s);
252}
253
254static void
255__trace_function(struct trace_array *tr,
256 unsigned long ip, unsigned long parent_ip,
257 unsigned long flags, int pc)
258{
259 if (is_graph())
260 trace_graph_function(tr, ip, parent_ip, flags, pc);
261 else
262 trace_function(tr, ip, parent_ip, flags, pc);
263}
264#else
265#define __trace_function trace_function
266
267static int wakeup_set_flag(u32 old_flags, u32 bit, int set)
268{
269 return -EINVAL;
270}
271
272static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
273{
274 return -1;
275}
276
277static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
278{
279 return TRACE_TYPE_UNHANDLED;
280}
281
282static void wakeup_graph_return(struct ftrace_graph_ret *trace) { }
283static void wakeup_print_header(struct seq_file *s) { }
284static void wakeup_trace_open(struct trace_iterator *iter) { }
285static void wakeup_trace_close(struct trace_iterator *iter) { }
286#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
287
288/*
289 * Should this new latency be reported/recorded?
290 */
291static int report_latency(cycle_t delta)
292{
293 if (tracing_thresh) {
294 if (delta < tracing_thresh)
295 return 0;
296 } else {
297 if (delta <= tracing_max_latency)
298 return 0;
299 }
300 return 1;
301}
302
303static void
304probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
305{
306 if (task != wakeup_task)
307 return;
308
309 wakeup_current_cpu = cpu;
310}
311
312static void notrace
313probe_wakeup_sched_switch(void *ignore,
314 struct task_struct *prev, struct task_struct *next)
315{
316 struct trace_array_cpu *data;
317 cycle_t T0, T1, delta;
318 unsigned long flags;
319 long disabled;
320 int cpu;
321 int pc;
322
323 tracing_record_cmdline(prev);
324
325 if (unlikely(!tracer_enabled))
326 return;
327
328 /*
329 * When we start a new trace, we set wakeup_task to NULL
330 * and then set tracer_enabled = 1. We want to make sure
331 * that another CPU does not see the tracer_enabled = 1
332 * and the wakeup_task with an older task, that might
333 * actually be the same as next.
334 */
335 smp_rmb();
336
337 if (next != wakeup_task)
338 return;
339
340 pc = preempt_count();
341
342 /* disable local data, not wakeup_cpu data */
343 cpu = raw_smp_processor_id();
344 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
345 if (likely(disabled != 1))
346 goto out;
347
348 local_irq_save(flags);
349 arch_spin_lock(&wakeup_lock);
350
351 /* We could race with grabbing wakeup_lock */
352 if (unlikely(!tracer_enabled || next != wakeup_task))
353 goto out_unlock;
354
355 /* The task we are waiting for is waking up */
356 data = wakeup_trace->data[wakeup_cpu];
357
358 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
359 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
360
361 T0 = data->preempt_timestamp;
362 T1 = ftrace_now(cpu);
363 delta = T1-T0;
364
365 if (!report_latency(delta))
366 goto out_unlock;
367
368 if (likely(!is_tracing_stopped())) {
369 tracing_max_latency = delta;
370 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
371 }
372
373out_unlock:
374 __wakeup_reset(wakeup_trace);
375 arch_spin_unlock(&wakeup_lock);
376 local_irq_restore(flags);
377out:
378 atomic_dec(&wakeup_trace->data[cpu]->disabled);
379}
380
381static void __wakeup_reset(struct trace_array *tr)
382{
383 wakeup_cpu = -1;
384 wakeup_prio = -1;
385
386 if (wakeup_task)
387 put_task_struct(wakeup_task);
388
389 wakeup_task = NULL;
390}
391
392static void wakeup_reset(struct trace_array *tr)
393{
394 unsigned long flags;
395
396 tracing_reset_online_cpus(tr);
397
398 local_irq_save(flags);
399 arch_spin_lock(&wakeup_lock);
400 __wakeup_reset(tr);
401 arch_spin_unlock(&wakeup_lock);
402 local_irq_restore(flags);
403}
404
405static void
406probe_wakeup(void *ignore, struct task_struct *p, int success)
407{
408 struct trace_array_cpu *data;
409 int cpu = smp_processor_id();
410 unsigned long flags;
411 long disabled;
412 int pc;
413
414 if (likely(!tracer_enabled))
415 return;
416
417 tracing_record_cmdline(p);
418 tracing_record_cmdline(current);
419
420 if ((wakeup_rt && !rt_task(p)) ||
421 p->prio >= wakeup_prio ||
422 p->prio >= current->prio)
423 return;
424
425 pc = preempt_count();
426 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
427 if (unlikely(disabled != 1))
428 goto out;
429
430 /* interrupts should be off from try_to_wake_up */
431 arch_spin_lock(&wakeup_lock);
432
433 /* check for races. */
434 if (!tracer_enabled || p->prio >= wakeup_prio)
435 goto out_locked;
436
437 /* reset the trace */
438 __wakeup_reset(wakeup_trace);
439
440 wakeup_cpu = task_cpu(p);
441 wakeup_current_cpu = wakeup_cpu;
442 wakeup_prio = p->prio;
443
444 wakeup_task = p;
445 get_task_struct(wakeup_task);
446
447 local_save_flags(flags);
448
449 data = wakeup_trace->data[wakeup_cpu];
450 data->preempt_timestamp = ftrace_now(cpu);
451 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
452
453 /*
454 * We must be careful in using CALLER_ADDR2. But since wake_up
455 * is not called by an assembly function (where as schedule is)
456 * it should be safe to use it here.
457 */
458 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
459
460out_locked:
461 arch_spin_unlock(&wakeup_lock);
462out:
463 atomic_dec(&wakeup_trace->data[cpu]->disabled);
464}
465
466static void start_wakeup_tracer(struct trace_array *tr)
467{
468 int ret;
469
470 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
471 if (ret) {
472 pr_info("wakeup trace: Couldn't activate tracepoint"
473 " probe to kernel_sched_wakeup\n");
474 return;
475 }
476
477 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
478 if (ret) {
479 pr_info("wakeup trace: Couldn't activate tracepoint"
480 " probe to kernel_sched_wakeup_new\n");
481 goto fail_deprobe;
482 }
483
484 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
485 if (ret) {
486 pr_info("sched trace: Couldn't activate tracepoint"
487 " probe to kernel_sched_switch\n");
488 goto fail_deprobe_wake_new;
489 }
490
491 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
492 if (ret) {
493 pr_info("wakeup trace: Couldn't activate tracepoint"
494 " probe to kernel_sched_migrate_task\n");
495 return;
496 }
497
498 wakeup_reset(tr);
499
500 /*
501 * Don't let the tracer_enabled = 1 show up before
502 * the wakeup_task is reset. This may be overkill since
503 * wakeup_reset does a spin_unlock after setting the
504 * wakeup_task to NULL, but I want to be safe.
505 * This is a slow path anyway.
506 */
507 smp_wmb();
508
509 if (start_func_tracer(is_graph()))
510 printk(KERN_ERR "failed to start wakeup tracer\n");
511
512 return;
513fail_deprobe_wake_new:
514 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
515fail_deprobe:
516 unregister_trace_sched_wakeup(probe_wakeup, NULL);
517}
518
519static void stop_wakeup_tracer(struct trace_array *tr)
520{
521 tracer_enabled = 0;
522 stop_func_tracer(is_graph());
523 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
524 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
525 unregister_trace_sched_wakeup(probe_wakeup, NULL);
526 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
527}
528
529static int __wakeup_tracer_init(struct trace_array *tr)
530{
531 save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
532 trace_flags |= TRACE_ITER_LATENCY_FMT;
533
534 tracing_max_latency = 0;
535 wakeup_trace = tr;
536 start_wakeup_tracer(tr);
537 return 0;
538}
539
540static int wakeup_tracer_init(struct trace_array *tr)
541{
542 wakeup_rt = 0;
543 return __wakeup_tracer_init(tr);
544}
545
546static int wakeup_rt_tracer_init(struct trace_array *tr)
547{
548 wakeup_rt = 1;
549 return __wakeup_tracer_init(tr);
550}
551
552static void wakeup_tracer_reset(struct trace_array *tr)
553{
554 stop_wakeup_tracer(tr);
555 /* make sure we put back any tasks we are tracing */
556 wakeup_reset(tr);
557
558 if (!save_lat_flag)
559 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
560}
561
562static void wakeup_tracer_start(struct trace_array *tr)
563{
564 wakeup_reset(tr);
565 tracer_enabled = 1;
566}
567
568static void wakeup_tracer_stop(struct trace_array *tr)
569{
570 tracer_enabled = 0;
571}
572
573static struct tracer wakeup_tracer __read_mostly =
574{
575 .name = "wakeup",
576 .init = wakeup_tracer_init,
577 .reset = wakeup_tracer_reset,
578 .start = wakeup_tracer_start,
579 .stop = wakeup_tracer_stop,
580 .print_max = 1,
581 .print_header = wakeup_print_header,
582 .print_line = wakeup_print_line,
583 .flags = &tracer_flags,
584 .set_flag = wakeup_set_flag,
585#ifdef CONFIG_FTRACE_SELFTEST
586 .selftest = trace_selftest_startup_wakeup,
587#endif
588 .open = wakeup_trace_open,
589 .close = wakeup_trace_close,
590 .use_max_tr = 1,
591};
592
593static struct tracer wakeup_rt_tracer __read_mostly =
594{
595 .name = "wakeup_rt",
596 .init = wakeup_rt_tracer_init,
597 .reset = wakeup_tracer_reset,
598 .start = wakeup_tracer_start,
599 .stop = wakeup_tracer_stop,
600 .wait_pipe = poll_wait_pipe,
601 .print_max = 1,
602 .print_header = wakeup_print_header,
603 .print_line = wakeup_print_line,
604 .flags = &tracer_flags,
605 .set_flag = wakeup_set_flag,
606#ifdef CONFIG_FTRACE_SELFTEST
607 .selftest = trace_selftest_startup_wakeup,
608#endif
609 .open = wakeup_trace_open,
610 .close = wakeup_trace_close,
611 .use_max_tr = 1,
612};
613
614__init static int init_wakeup_tracer(void)
615{
616 int ret;
617
618 ret = register_tracer(&wakeup_tracer);
619 if (ret)
620 return ret;
621
622 ret = register_tracer(&wakeup_rt_tracer);
623 if (ret)
624 return ret;
625
626 return 0;
627}
628device_initcall(init_wakeup_tracer);