Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * trace task wakeup timings
4 *
5 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
6 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
7 *
8 * Based on code from the latency_tracer, that is:
9 *
10 * Copyright (C) 2004-2006 Ingo Molnar
11 * Copyright (C) 2004 Nadia Yvette Chambers
12 */
13#include <linux/module.h>
14#include <linux/kallsyms.h>
15#include <linux/uaccess.h>
16#include <linux/ftrace.h>
17#include <linux/sched/rt.h>
18#include <linux/sched/deadline.h>
19#include <trace/events/sched.h>
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;
30static int wakeup_dl;
31static int tracing_dl = 0;
32
33static arch_spinlock_t wakeup_lock =
34 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
35
36static void wakeup_reset(struct trace_array *tr);
37static void __wakeup_reset(struct trace_array *tr);
38static int start_func_tracer(struct trace_array *tr, int graph);
39static void stop_func_tracer(struct trace_array *tr, int graph);
40
41static int save_flags;
42
43#ifdef CONFIG_FUNCTION_GRAPH_TRACER
44# define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
45#else
46# define is_graph(tr) false
47#endif
48
49#ifdef CONFIG_FUNCTION_TRACER
50
51static bool function_enabled;
52
53/*
54 * Prologue for the wakeup function tracers.
55 *
56 * Returns 1 if it is OK to continue, and preemption
57 * is disabled and data->disabled is incremented.
58 * 0 if the trace is to be ignored, and preemption
59 * is not disabled and data->disabled is
60 * kept the same.
61 *
62 * Note, this function is also used outside this ifdef but
63 * inside the #ifdef of the function graph tracer below.
64 * This is OK, since the function graph tracer is
65 * dependent on the function tracer.
66 */
67static int
68func_prolog_preempt_disable(struct trace_array *tr,
69 struct trace_array_cpu **data,
70 int *pc)
71{
72 long disabled;
73 int cpu;
74
75 if (likely(!wakeup_task))
76 return 0;
77
78 *pc = preempt_count();
79 preempt_disable_notrace();
80
81 cpu = raw_smp_processor_id();
82 if (cpu != wakeup_current_cpu)
83 goto out_enable;
84
85 *data = per_cpu_ptr(tr->trace_buffer.data, cpu);
86 disabled = atomic_inc_return(&(*data)->disabled);
87 if (unlikely(disabled != 1))
88 goto out;
89
90 return 1;
91
92out:
93 atomic_dec(&(*data)->disabled);
94
95out_enable:
96 preempt_enable_notrace();
97 return 0;
98}
99
100#ifdef CONFIG_FUNCTION_GRAPH_TRACER
101
102static int wakeup_display_graph(struct trace_array *tr, int set)
103{
104 if (!(is_graph(tr) ^ set))
105 return 0;
106
107 stop_func_tracer(tr, !set);
108
109 wakeup_reset(wakeup_trace);
110 tr->max_latency = 0;
111
112 return start_func_tracer(tr, set);
113}
114
115static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
116{
117 struct trace_array *tr = wakeup_trace;
118 struct trace_array_cpu *data;
119 unsigned long flags;
120 int pc, ret = 0;
121
122 if (ftrace_graph_ignore_func(trace))
123 return 0;
124 /*
125 * Do not trace a function if it's filtered by set_graph_notrace.
126 * Make the index of ret stack negative to indicate that it should
127 * ignore further functions. But it needs its own ret stack entry
128 * to recover the original index in order to continue tracing after
129 * returning from the function.
130 */
131 if (ftrace_graph_notrace_addr(trace->func))
132 return 1;
133
134 if (!func_prolog_preempt_disable(tr, &data, &pc))
135 return 0;
136
137 local_save_flags(flags);
138 ret = __trace_graph_entry(tr, trace, flags, pc);
139 atomic_dec(&data->disabled);
140 preempt_enable_notrace();
141
142 return ret;
143}
144
145static void wakeup_graph_return(struct ftrace_graph_ret *trace)
146{
147 struct trace_array *tr = wakeup_trace;
148 struct trace_array_cpu *data;
149 unsigned long flags;
150 int pc;
151
152 ftrace_graph_addr_finish(trace);
153
154 if (!func_prolog_preempt_disable(tr, &data, &pc))
155 return;
156
157 local_save_flags(flags);
158 __trace_graph_return(tr, trace, flags, pc);
159 atomic_dec(&data->disabled);
160
161 preempt_enable_notrace();
162 return;
163}
164
165static struct fgraph_ops fgraph_wakeup_ops = {
166 .entryfunc = &wakeup_graph_entry,
167 .retfunc = &wakeup_graph_return,
168};
169
170static void wakeup_trace_open(struct trace_iterator *iter)
171{
172 if (is_graph(iter->tr))
173 graph_trace_open(iter);
174}
175
176static void wakeup_trace_close(struct trace_iterator *iter)
177{
178 if (iter->private)
179 graph_trace_close(iter);
180}
181
182#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
183 TRACE_GRAPH_PRINT_CPU | \
184 TRACE_GRAPH_PRINT_REL_TIME | \
185 TRACE_GRAPH_PRINT_DURATION | \
186 TRACE_GRAPH_PRINT_OVERHEAD | \
187 TRACE_GRAPH_PRINT_IRQS)
188
189static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
190{
191 /*
192 * In graph mode call the graph tracer output function,
193 * otherwise go with the TRACE_FN event handler
194 */
195 if (is_graph(iter->tr))
196 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
197
198 return TRACE_TYPE_UNHANDLED;
199}
200
201static void wakeup_print_header(struct seq_file *s)
202{
203 if (is_graph(wakeup_trace))
204 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
205 else
206 trace_default_header(s);
207}
208#endif /* else CONFIG_FUNCTION_GRAPH_TRACER */
209
210/*
211 * wakeup uses its own tracer function to keep the overhead down:
212 */
213static void
214wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
215 struct ftrace_ops *op, struct pt_regs *pt_regs)
216{
217 struct trace_array *tr = wakeup_trace;
218 struct trace_array_cpu *data;
219 unsigned long flags;
220 int pc;
221
222 if (!func_prolog_preempt_disable(tr, &data, &pc))
223 return;
224
225 local_irq_save(flags);
226 trace_function(tr, ip, parent_ip, flags, pc);
227 local_irq_restore(flags);
228
229 atomic_dec(&data->disabled);
230 preempt_enable_notrace();
231}
232
233static int register_wakeup_function(struct trace_array *tr, int graph, int set)
234{
235 int ret;
236
237 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
238 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
239 return 0;
240
241 if (graph)
242 ret = register_ftrace_graph(&fgraph_wakeup_ops);
243 else
244 ret = register_ftrace_function(tr->ops);
245
246 if (!ret)
247 function_enabled = true;
248
249 return ret;
250}
251
252static void unregister_wakeup_function(struct trace_array *tr, int graph)
253{
254 if (!function_enabled)
255 return;
256
257 if (graph)
258 unregister_ftrace_graph(&fgraph_wakeup_ops);
259 else
260 unregister_ftrace_function(tr->ops);
261
262 function_enabled = false;
263}
264
265static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
266{
267 if (!(mask & TRACE_ITER_FUNCTION))
268 return 0;
269
270 if (set)
271 register_wakeup_function(tr, is_graph(tr), 1);
272 else
273 unregister_wakeup_function(tr, is_graph(tr));
274 return 1;
275}
276#else /* CONFIG_FUNCTION_TRACER */
277static int register_wakeup_function(struct trace_array *tr, int graph, int set)
278{
279 return 0;
280}
281static void unregister_wakeup_function(struct trace_array *tr, int graph) { }
282static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
283{
284 return 0;
285}
286#endif /* else CONFIG_FUNCTION_TRACER */
287
288#ifndef CONFIG_FUNCTION_GRAPH_TRACER
289static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
290{
291 return TRACE_TYPE_UNHANDLED;
292}
293
294static void wakeup_trace_open(struct trace_iterator *iter) { }
295static void wakeup_trace_close(struct trace_iterator *iter) { }
296
297static void wakeup_print_header(struct seq_file *s)
298{
299 trace_default_header(s);
300}
301#endif /* !CONFIG_FUNCTION_GRAPH_TRACER */
302
303static void
304__trace_function(struct trace_array *tr,
305 unsigned long ip, unsigned long parent_ip,
306 unsigned long flags, int pc)
307{
308 if (is_graph(tr))
309 trace_graph_function(tr, ip, parent_ip, flags, pc);
310 else
311 trace_function(tr, ip, parent_ip, flags, pc);
312}
313
314static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
315{
316 struct tracer *tracer = tr->current_trace;
317
318 if (wakeup_function_set(tr, mask, set))
319 return 0;
320
321#ifdef CONFIG_FUNCTION_GRAPH_TRACER
322 if (mask & TRACE_ITER_DISPLAY_GRAPH)
323 return wakeup_display_graph(tr, set);
324#endif
325
326 return trace_keep_overwrite(tracer, mask, set);
327}
328
329static int start_func_tracer(struct trace_array *tr, int graph)
330{
331 int ret;
332
333 ret = register_wakeup_function(tr, graph, 0);
334
335 if (!ret && tracing_is_enabled())
336 tracer_enabled = 1;
337 else
338 tracer_enabled = 0;
339
340 return ret;
341}
342
343static void stop_func_tracer(struct trace_array *tr, int graph)
344{
345 tracer_enabled = 0;
346
347 unregister_wakeup_function(tr, graph);
348}
349
350/*
351 * Should this new latency be reported/recorded?
352 */
353static bool report_latency(struct trace_array *tr, u64 delta)
354{
355 if (tracing_thresh) {
356 if (delta < tracing_thresh)
357 return false;
358 } else {
359 if (delta <= tr->max_latency)
360 return false;
361 }
362 return true;
363}
364
365static void
366probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
367{
368 if (task != wakeup_task)
369 return;
370
371 wakeup_current_cpu = cpu;
372}
373
374static void
375tracing_sched_switch_trace(struct trace_array *tr,
376 struct task_struct *prev,
377 struct task_struct *next,
378 unsigned long flags, int pc)
379{
380 struct trace_event_call *call = &event_context_switch;
381 struct ring_buffer *buffer = tr->trace_buffer.buffer;
382 struct ring_buffer_event *event;
383 struct ctx_switch_entry *entry;
384
385 event = trace_buffer_lock_reserve(buffer, TRACE_CTX,
386 sizeof(*entry), flags, pc);
387 if (!event)
388 return;
389 entry = ring_buffer_event_data(event);
390 entry->prev_pid = prev->pid;
391 entry->prev_prio = prev->prio;
392 entry->prev_state = task_state_index(prev);
393 entry->next_pid = next->pid;
394 entry->next_prio = next->prio;
395 entry->next_state = task_state_index(next);
396 entry->next_cpu = task_cpu(next);
397
398 if (!call_filter_check_discard(call, entry, buffer, event))
399 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
400}
401
402static void
403tracing_sched_wakeup_trace(struct trace_array *tr,
404 struct task_struct *wakee,
405 struct task_struct *curr,
406 unsigned long flags, int pc)
407{
408 struct trace_event_call *call = &event_wakeup;
409 struct ring_buffer_event *event;
410 struct ctx_switch_entry *entry;
411 struct ring_buffer *buffer = tr->trace_buffer.buffer;
412
413 event = trace_buffer_lock_reserve(buffer, TRACE_WAKE,
414 sizeof(*entry), flags, pc);
415 if (!event)
416 return;
417 entry = ring_buffer_event_data(event);
418 entry->prev_pid = curr->pid;
419 entry->prev_prio = curr->prio;
420 entry->prev_state = task_state_index(curr);
421 entry->next_pid = wakee->pid;
422 entry->next_prio = wakee->prio;
423 entry->next_state = task_state_index(wakee);
424 entry->next_cpu = task_cpu(wakee);
425
426 if (!call_filter_check_discard(call, entry, buffer, event))
427 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
428}
429
430static void notrace
431probe_wakeup_sched_switch(void *ignore, bool preempt,
432 struct task_struct *prev, struct task_struct *next)
433{
434 struct trace_array_cpu *data;
435 u64 T0, T1, delta;
436 unsigned long flags;
437 long disabled;
438 int cpu;
439 int pc;
440
441 tracing_record_cmdline(prev);
442
443 if (unlikely(!tracer_enabled))
444 return;
445
446 /*
447 * When we start a new trace, we set wakeup_task to NULL
448 * and then set tracer_enabled = 1. We want to make sure
449 * that another CPU does not see the tracer_enabled = 1
450 * and the wakeup_task with an older task, that might
451 * actually be the same as next.
452 */
453 smp_rmb();
454
455 if (next != wakeup_task)
456 return;
457
458 pc = preempt_count();
459
460 /* disable local data, not wakeup_cpu data */
461 cpu = raw_smp_processor_id();
462 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
463 if (likely(disabled != 1))
464 goto out;
465
466 local_irq_save(flags);
467 arch_spin_lock(&wakeup_lock);
468
469 /* We could race with grabbing wakeup_lock */
470 if (unlikely(!tracer_enabled || next != wakeup_task))
471 goto out_unlock;
472
473 /* The task we are waiting for is waking up */
474 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
475
476 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
477 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
478 __trace_stack(wakeup_trace, flags, 0, pc);
479
480 T0 = data->preempt_timestamp;
481 T1 = ftrace_now(cpu);
482 delta = T1-T0;
483
484 if (!report_latency(wakeup_trace, delta))
485 goto out_unlock;
486
487 if (likely(!is_tracing_stopped())) {
488 wakeup_trace->max_latency = delta;
489 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu, NULL);
490 }
491
492out_unlock:
493 __wakeup_reset(wakeup_trace);
494 arch_spin_unlock(&wakeup_lock);
495 local_irq_restore(flags);
496out:
497 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
498}
499
500static void __wakeup_reset(struct trace_array *tr)
501{
502 wakeup_cpu = -1;
503 wakeup_prio = -1;
504 tracing_dl = 0;
505
506 if (wakeup_task)
507 put_task_struct(wakeup_task);
508
509 wakeup_task = NULL;
510}
511
512static void wakeup_reset(struct trace_array *tr)
513{
514 unsigned long flags;
515
516 tracing_reset_online_cpus(&tr->trace_buffer);
517
518 local_irq_save(flags);
519 arch_spin_lock(&wakeup_lock);
520 __wakeup_reset(tr);
521 arch_spin_unlock(&wakeup_lock);
522 local_irq_restore(flags);
523}
524
525static void
526probe_wakeup(void *ignore, struct task_struct *p)
527{
528 struct trace_array_cpu *data;
529 int cpu = smp_processor_id();
530 unsigned long flags;
531 long disabled;
532 int pc;
533
534 if (likely(!tracer_enabled))
535 return;
536
537 tracing_record_cmdline(p);
538 tracing_record_cmdline(current);
539
540 /*
541 * Semantic is like this:
542 * - wakeup tracer handles all tasks in the system, independently
543 * from their scheduling class;
544 * - wakeup_rt tracer handles tasks belonging to sched_dl and
545 * sched_rt class;
546 * - wakeup_dl handles tasks belonging to sched_dl class only.
547 */
548 if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
549 (wakeup_rt && !dl_task(p) && !rt_task(p)) ||
550 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
551 return;
552
553 pc = preempt_count();
554 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
555 if (unlikely(disabled != 1))
556 goto out;
557
558 /* interrupts should be off from try_to_wake_up */
559 arch_spin_lock(&wakeup_lock);
560
561 /* check for races. */
562 if (!tracer_enabled || tracing_dl ||
563 (!dl_task(p) && p->prio >= wakeup_prio))
564 goto out_locked;
565
566 /* reset the trace */
567 __wakeup_reset(wakeup_trace);
568
569 wakeup_cpu = task_cpu(p);
570 wakeup_current_cpu = wakeup_cpu;
571 wakeup_prio = p->prio;
572
573 /*
574 * Once you start tracing a -deadline task, don't bother tracing
575 * another task until the first one wakes up.
576 */
577 if (dl_task(p))
578 tracing_dl = 1;
579 else
580 tracing_dl = 0;
581
582 wakeup_task = get_task_struct(p);
583
584 local_save_flags(flags);
585
586 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
587 data->preempt_timestamp = ftrace_now(cpu);
588 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
589 __trace_stack(wakeup_trace, flags, 0, pc);
590
591 /*
592 * We must be careful in using CALLER_ADDR2. But since wake_up
593 * is not called by an assembly function (where as schedule is)
594 * it should be safe to use it here.
595 */
596 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
597
598out_locked:
599 arch_spin_unlock(&wakeup_lock);
600out:
601 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
602}
603
604static void start_wakeup_tracer(struct trace_array *tr)
605{
606 int ret;
607
608 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
609 if (ret) {
610 pr_info("wakeup trace: Couldn't activate tracepoint"
611 " probe to kernel_sched_wakeup\n");
612 return;
613 }
614
615 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
616 if (ret) {
617 pr_info("wakeup trace: Couldn't activate tracepoint"
618 " probe to kernel_sched_wakeup_new\n");
619 goto fail_deprobe;
620 }
621
622 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
623 if (ret) {
624 pr_info("sched trace: Couldn't activate tracepoint"
625 " probe to kernel_sched_switch\n");
626 goto fail_deprobe_wake_new;
627 }
628
629 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
630 if (ret) {
631 pr_info("wakeup trace: Couldn't activate tracepoint"
632 " probe to kernel_sched_migrate_task\n");
633 return;
634 }
635
636 wakeup_reset(tr);
637
638 /*
639 * Don't let the tracer_enabled = 1 show up before
640 * the wakeup_task is reset. This may be overkill since
641 * wakeup_reset does a spin_unlock after setting the
642 * wakeup_task to NULL, but I want to be safe.
643 * This is a slow path anyway.
644 */
645 smp_wmb();
646
647 if (start_func_tracer(tr, is_graph(tr)))
648 printk(KERN_ERR "failed to start wakeup tracer\n");
649
650 return;
651fail_deprobe_wake_new:
652 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
653fail_deprobe:
654 unregister_trace_sched_wakeup(probe_wakeup, NULL);
655}
656
657static void stop_wakeup_tracer(struct trace_array *tr)
658{
659 tracer_enabled = 0;
660 stop_func_tracer(tr, is_graph(tr));
661 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
662 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
663 unregister_trace_sched_wakeup(probe_wakeup, NULL);
664 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
665}
666
667static bool wakeup_busy;
668
669static int __wakeup_tracer_init(struct trace_array *tr)
670{
671 save_flags = tr->trace_flags;
672
673 /* non overwrite screws up the latency tracers */
674 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
675 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
676
677 tr->max_latency = 0;
678 wakeup_trace = tr;
679 ftrace_init_array_ops(tr, wakeup_tracer_call);
680 start_wakeup_tracer(tr);
681
682 wakeup_busy = true;
683 return 0;
684}
685
686static int wakeup_tracer_init(struct trace_array *tr)
687{
688 if (wakeup_busy)
689 return -EBUSY;
690
691 wakeup_dl = 0;
692 wakeup_rt = 0;
693 return __wakeup_tracer_init(tr);
694}
695
696static int wakeup_rt_tracer_init(struct trace_array *tr)
697{
698 if (wakeup_busy)
699 return -EBUSY;
700
701 wakeup_dl = 0;
702 wakeup_rt = 1;
703 return __wakeup_tracer_init(tr);
704}
705
706static int wakeup_dl_tracer_init(struct trace_array *tr)
707{
708 if (wakeup_busy)
709 return -EBUSY;
710
711 wakeup_dl = 1;
712 wakeup_rt = 0;
713 return __wakeup_tracer_init(tr);
714}
715
716static void wakeup_tracer_reset(struct trace_array *tr)
717{
718 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
719 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
720
721 stop_wakeup_tracer(tr);
722 /* make sure we put back any tasks we are tracing */
723 wakeup_reset(tr);
724
725 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
726 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
727 ftrace_reset_array_ops(tr);
728 wakeup_busy = false;
729}
730
731static void wakeup_tracer_start(struct trace_array *tr)
732{
733 wakeup_reset(tr);
734 tracer_enabled = 1;
735}
736
737static void wakeup_tracer_stop(struct trace_array *tr)
738{
739 tracer_enabled = 0;
740}
741
742static struct tracer wakeup_tracer __read_mostly =
743{
744 .name = "wakeup",
745 .init = wakeup_tracer_init,
746 .reset = wakeup_tracer_reset,
747 .start = wakeup_tracer_start,
748 .stop = wakeup_tracer_stop,
749 .print_max = true,
750 .print_header = wakeup_print_header,
751 .print_line = wakeup_print_line,
752 .flag_changed = wakeup_flag_changed,
753#ifdef CONFIG_FTRACE_SELFTEST
754 .selftest = trace_selftest_startup_wakeup,
755#endif
756 .open = wakeup_trace_open,
757 .close = wakeup_trace_close,
758 .allow_instances = true,
759 .use_max_tr = true,
760};
761
762static struct tracer wakeup_rt_tracer __read_mostly =
763{
764 .name = "wakeup_rt",
765 .init = wakeup_rt_tracer_init,
766 .reset = wakeup_tracer_reset,
767 .start = wakeup_tracer_start,
768 .stop = wakeup_tracer_stop,
769 .print_max = true,
770 .print_header = wakeup_print_header,
771 .print_line = wakeup_print_line,
772 .flag_changed = wakeup_flag_changed,
773#ifdef CONFIG_FTRACE_SELFTEST
774 .selftest = trace_selftest_startup_wakeup,
775#endif
776 .open = wakeup_trace_open,
777 .close = wakeup_trace_close,
778 .allow_instances = true,
779 .use_max_tr = true,
780};
781
782static struct tracer wakeup_dl_tracer __read_mostly =
783{
784 .name = "wakeup_dl",
785 .init = wakeup_dl_tracer_init,
786 .reset = wakeup_tracer_reset,
787 .start = wakeup_tracer_start,
788 .stop = wakeup_tracer_stop,
789 .print_max = true,
790 .print_header = wakeup_print_header,
791 .print_line = wakeup_print_line,
792 .flag_changed = wakeup_flag_changed,
793#ifdef CONFIG_FTRACE_SELFTEST
794 .selftest = trace_selftest_startup_wakeup,
795#endif
796 .open = wakeup_trace_open,
797 .close = wakeup_trace_close,
798 .allow_instances = true,
799 .use_max_tr = true,
800};
801
802__init static int init_wakeup_tracer(void)
803{
804 int ret;
805
806 ret = register_tracer(&wakeup_tracer);
807 if (ret)
808 return ret;
809
810 ret = register_tracer(&wakeup_rt_tracer);
811 if (ret)
812 return ret;
813
814 ret = register_tracer(&wakeup_dl_tracer);
815 if (ret)
816 return ret;
817
818 return 0;
819}
820core_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);