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 bool wakeup_rt;
30static bool wakeup_dl;
31static bool tracing_dl;
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 unsigned int *trace_ctx)
71{
72 long disabled;
73 int cpu;
74
75 if (likely(!wakeup_task))
76 return 0;
77
78 *trace_ctx = tracing_gen_ctx();
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->array_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 struct fgraph_ops *gops)
117{
118 struct trace_array *tr = wakeup_trace;
119 struct trace_array_cpu *data;
120 unsigned int trace_ctx;
121 u64 *calltime;
122 int ret = 0;
123
124 if (ftrace_graph_ignore_func(gops, trace))
125 return 0;
126 /*
127 * Do not trace a function if it's filtered by set_graph_notrace.
128 * Make the index of ret stack negative to indicate that it should
129 * ignore further functions. But it needs its own ret stack entry
130 * to recover the original index in order to continue tracing after
131 * returning from the function.
132 */
133 if (ftrace_graph_notrace_addr(trace->func))
134 return 1;
135
136 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
137 return 0;
138
139 calltime = fgraph_reserve_data(gops->idx, sizeof(*calltime));
140 if (!calltime)
141 return 0;
142
143 *calltime = trace_clock_local();
144
145 ret = __trace_graph_entry(tr, trace, trace_ctx);
146 atomic_dec(&data->disabled);
147 preempt_enable_notrace();
148
149 return ret;
150}
151
152static void wakeup_graph_return(struct ftrace_graph_ret *trace,
153 struct fgraph_ops *gops)
154{
155 struct trace_array *tr = wakeup_trace;
156 struct trace_array_cpu *data;
157 unsigned int trace_ctx;
158 u64 *calltime;
159 int size;
160
161 ftrace_graph_addr_finish(gops, trace);
162
163 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
164 return;
165
166 calltime = fgraph_retrieve_data(gops->idx, &size);
167 if (!calltime)
168 return;
169 trace->calltime = *calltime;
170
171 __trace_graph_return(tr, trace, trace_ctx);
172 atomic_dec(&data->disabled);
173
174 preempt_enable_notrace();
175 return;
176}
177
178static struct fgraph_ops fgraph_wakeup_ops = {
179 .entryfunc = &wakeup_graph_entry,
180 .retfunc = &wakeup_graph_return,
181};
182
183static void wakeup_trace_open(struct trace_iterator *iter)
184{
185 if (is_graph(iter->tr))
186 graph_trace_open(iter);
187 else
188 iter->private = NULL;
189}
190
191static void wakeup_trace_close(struct trace_iterator *iter)
192{
193 if (iter->private)
194 graph_trace_close(iter);
195}
196
197#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
198 TRACE_GRAPH_PRINT_CPU | \
199 TRACE_GRAPH_PRINT_REL_TIME | \
200 TRACE_GRAPH_PRINT_DURATION | \
201 TRACE_GRAPH_PRINT_OVERHEAD | \
202 TRACE_GRAPH_PRINT_IRQS)
203
204static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
205{
206 /*
207 * In graph mode call the graph tracer output function,
208 * otherwise go with the TRACE_FN event handler
209 */
210 if (is_graph(iter->tr))
211 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
212
213 return TRACE_TYPE_UNHANDLED;
214}
215
216static void wakeup_print_header(struct seq_file *s)
217{
218 if (is_graph(wakeup_trace))
219 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
220 else
221 trace_default_header(s);
222}
223#endif /* else CONFIG_FUNCTION_GRAPH_TRACER */
224
225/*
226 * wakeup uses its own tracer function to keep the overhead down:
227 */
228static void
229wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
230 struct ftrace_ops *op, struct ftrace_regs *fregs)
231{
232 struct trace_array *tr = wakeup_trace;
233 struct trace_array_cpu *data;
234 unsigned long flags;
235 unsigned int trace_ctx;
236
237 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
238 return;
239
240 local_irq_save(flags);
241 trace_function(tr, ip, parent_ip, trace_ctx);
242 local_irq_restore(flags);
243
244 atomic_dec(&data->disabled);
245 preempt_enable_notrace();
246}
247
248static int register_wakeup_function(struct trace_array *tr, int graph, int set)
249{
250 int ret;
251
252 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
253 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
254 return 0;
255
256 if (graph)
257 ret = register_ftrace_graph(&fgraph_wakeup_ops);
258 else
259 ret = register_ftrace_function(tr->ops);
260
261 if (!ret)
262 function_enabled = true;
263
264 return ret;
265}
266
267static void unregister_wakeup_function(struct trace_array *tr, int graph)
268{
269 if (!function_enabled)
270 return;
271
272 if (graph)
273 unregister_ftrace_graph(&fgraph_wakeup_ops);
274 else
275 unregister_ftrace_function(tr->ops);
276
277 function_enabled = false;
278}
279
280static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
281{
282 if (!(mask & TRACE_ITER_FUNCTION))
283 return 0;
284
285 if (set)
286 register_wakeup_function(tr, is_graph(tr), 1);
287 else
288 unregister_wakeup_function(tr, is_graph(tr));
289 return 1;
290}
291#else /* CONFIG_FUNCTION_TRACER */
292static int register_wakeup_function(struct trace_array *tr, int graph, int set)
293{
294 return 0;
295}
296static void unregister_wakeup_function(struct trace_array *tr, int graph) { }
297static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
298{
299 return 0;
300}
301#endif /* else CONFIG_FUNCTION_TRACER */
302
303#ifndef CONFIG_FUNCTION_GRAPH_TRACER
304static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
305{
306 return TRACE_TYPE_UNHANDLED;
307}
308
309static void wakeup_trace_open(struct trace_iterator *iter) { }
310static void wakeup_trace_close(struct trace_iterator *iter) { }
311
312static void wakeup_print_header(struct seq_file *s)
313{
314 trace_default_header(s);
315}
316#endif /* !CONFIG_FUNCTION_GRAPH_TRACER */
317
318static void
319__trace_function(struct trace_array *tr,
320 unsigned long ip, unsigned long parent_ip,
321 unsigned int trace_ctx)
322{
323 if (is_graph(tr))
324 trace_graph_function(tr, ip, parent_ip, trace_ctx);
325 else
326 trace_function(tr, ip, parent_ip, trace_ctx);
327}
328
329static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
330{
331 struct tracer *tracer = tr->current_trace;
332
333 if (wakeup_function_set(tr, mask, set))
334 return 0;
335
336#ifdef CONFIG_FUNCTION_GRAPH_TRACER
337 if (mask & TRACE_ITER_DISPLAY_GRAPH)
338 return wakeup_display_graph(tr, set);
339#endif
340
341 return trace_keep_overwrite(tracer, mask, set);
342}
343
344static int start_func_tracer(struct trace_array *tr, int graph)
345{
346 int ret;
347
348 ret = register_wakeup_function(tr, graph, 0);
349
350 if (!ret && tracing_is_enabled())
351 tracer_enabled = 1;
352 else
353 tracer_enabled = 0;
354
355 return ret;
356}
357
358static void stop_func_tracer(struct trace_array *tr, int graph)
359{
360 tracer_enabled = 0;
361
362 unregister_wakeup_function(tr, graph);
363}
364
365/*
366 * Should this new latency be reported/recorded?
367 */
368static bool report_latency(struct trace_array *tr, u64 delta)
369{
370 if (tracing_thresh) {
371 if (delta < tracing_thresh)
372 return false;
373 } else {
374 if (delta <= tr->max_latency)
375 return false;
376 }
377 return true;
378}
379
380static void
381probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
382{
383 if (task != wakeup_task)
384 return;
385
386 wakeup_current_cpu = cpu;
387}
388
389static void
390tracing_sched_switch_trace(struct trace_array *tr,
391 struct task_struct *prev,
392 struct task_struct *next,
393 unsigned int trace_ctx)
394{
395 struct trace_buffer *buffer = tr->array_buffer.buffer;
396 struct ring_buffer_event *event;
397 struct ctx_switch_entry *entry;
398
399 event = trace_buffer_lock_reserve(buffer, TRACE_CTX,
400 sizeof(*entry), trace_ctx);
401 if (!event)
402 return;
403 entry = ring_buffer_event_data(event);
404 entry->prev_pid = prev->pid;
405 entry->prev_prio = prev->prio;
406 entry->prev_state = task_state_index(prev);
407 entry->next_pid = next->pid;
408 entry->next_prio = next->prio;
409 entry->next_state = task_state_index(next);
410 entry->next_cpu = task_cpu(next);
411
412 trace_buffer_unlock_commit(tr, buffer, event, trace_ctx);
413}
414
415static void
416tracing_sched_wakeup_trace(struct trace_array *tr,
417 struct task_struct *wakee,
418 struct task_struct *curr,
419 unsigned int trace_ctx)
420{
421 struct ring_buffer_event *event;
422 struct ctx_switch_entry *entry;
423 struct trace_buffer *buffer = tr->array_buffer.buffer;
424
425 event = trace_buffer_lock_reserve(buffer, TRACE_WAKE,
426 sizeof(*entry), trace_ctx);
427 if (!event)
428 return;
429 entry = ring_buffer_event_data(event);
430 entry->prev_pid = curr->pid;
431 entry->prev_prio = curr->prio;
432 entry->prev_state = task_state_index(curr);
433 entry->next_pid = wakee->pid;
434 entry->next_prio = wakee->prio;
435 entry->next_state = task_state_index(wakee);
436 entry->next_cpu = task_cpu(wakee);
437
438 trace_buffer_unlock_commit(tr, buffer, event, trace_ctx);
439}
440
441static void notrace
442probe_wakeup_sched_switch(void *ignore, bool preempt,
443 struct task_struct *prev, struct task_struct *next,
444 unsigned int prev_state)
445{
446 struct trace_array_cpu *data;
447 u64 T0, T1, delta;
448 unsigned long flags;
449 long disabled;
450 int cpu;
451 unsigned int trace_ctx;
452
453 tracing_record_cmdline(prev);
454
455 if (unlikely(!tracer_enabled))
456 return;
457
458 /*
459 * When we start a new trace, we set wakeup_task to NULL
460 * and then set tracer_enabled = 1. We want to make sure
461 * that another CPU does not see the tracer_enabled = 1
462 * and the wakeup_task with an older task, that might
463 * actually be the same as next.
464 */
465 smp_rmb();
466
467 if (next != wakeup_task)
468 return;
469
470 /* disable local data, not wakeup_cpu data */
471 cpu = raw_smp_processor_id();
472 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
473 if (likely(disabled != 1))
474 goto out;
475
476 local_irq_save(flags);
477 trace_ctx = tracing_gen_ctx_flags(flags);
478
479 arch_spin_lock(&wakeup_lock);
480
481 /* We could race with grabbing wakeup_lock */
482 if (unlikely(!tracer_enabled || next != wakeup_task))
483 goto out_unlock;
484
485 /* The task we are waiting for is waking up */
486 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
487
488 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, trace_ctx);
489 tracing_sched_switch_trace(wakeup_trace, prev, next, trace_ctx);
490 __trace_stack(wakeup_trace, trace_ctx, 0);
491
492 T0 = data->preempt_timestamp;
493 T1 = ftrace_now(cpu);
494 delta = T1-T0;
495
496 if (!report_latency(wakeup_trace, delta))
497 goto out_unlock;
498
499 if (likely(!is_tracing_stopped())) {
500 wakeup_trace->max_latency = delta;
501 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu, NULL);
502 }
503
504out_unlock:
505 __wakeup_reset(wakeup_trace);
506 arch_spin_unlock(&wakeup_lock);
507 local_irq_restore(flags);
508out:
509 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
510}
511
512static void __wakeup_reset(struct trace_array *tr)
513{
514 wakeup_cpu = -1;
515 wakeup_prio = -1;
516 tracing_dl = false;
517
518 if (wakeup_task)
519 put_task_struct(wakeup_task);
520
521 wakeup_task = NULL;
522}
523
524static void wakeup_reset(struct trace_array *tr)
525{
526 unsigned long flags;
527
528 tracing_reset_online_cpus(&tr->array_buffer);
529
530 local_irq_save(flags);
531 arch_spin_lock(&wakeup_lock);
532 __wakeup_reset(tr);
533 arch_spin_unlock(&wakeup_lock);
534 local_irq_restore(flags);
535}
536
537static void
538probe_wakeup(void *ignore, struct task_struct *p)
539{
540 struct trace_array_cpu *data;
541 int cpu = smp_processor_id();
542 long disabled;
543 unsigned int trace_ctx;
544
545 if (likely(!tracer_enabled))
546 return;
547
548 tracing_record_cmdline(p);
549 tracing_record_cmdline(current);
550
551 /*
552 * Semantic is like this:
553 * - wakeup tracer handles all tasks in the system, independently
554 * from their scheduling class;
555 * - wakeup_rt tracer handles tasks belonging to sched_dl and
556 * sched_rt class;
557 * - wakeup_dl handles tasks belonging to sched_dl class only.
558 */
559 if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
560 (wakeup_rt && !rt_or_dl_task(p)) ||
561 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
562 return;
563
564 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
565 if (unlikely(disabled != 1))
566 goto out;
567
568 trace_ctx = tracing_gen_ctx();
569
570 /* interrupts should be off from try_to_wake_up */
571 arch_spin_lock(&wakeup_lock);
572
573 /* check for races. */
574 if (!tracer_enabled || tracing_dl ||
575 (!dl_task(p) && p->prio >= wakeup_prio))
576 goto out_locked;
577
578 /* reset the trace */
579 __wakeup_reset(wakeup_trace);
580
581 wakeup_cpu = task_cpu(p);
582 wakeup_current_cpu = wakeup_cpu;
583 wakeup_prio = p->prio;
584
585 /*
586 * Once you start tracing a -deadline task, don't bother tracing
587 * another task until the first one wakes up.
588 */
589 if (dl_task(p))
590 tracing_dl = true;
591 else
592 tracing_dl = false;
593
594 wakeup_task = get_task_struct(p);
595
596 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
597 data->preempt_timestamp = ftrace_now(cpu);
598 tracing_sched_wakeup_trace(wakeup_trace, p, current, trace_ctx);
599 __trace_stack(wakeup_trace, trace_ctx, 0);
600
601 /*
602 * We must be careful in using CALLER_ADDR2. But since wake_up
603 * is not called by an assembly function (where as schedule is)
604 * it should be safe to use it here.
605 */
606 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, trace_ctx);
607
608out_locked:
609 arch_spin_unlock(&wakeup_lock);
610out:
611 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
612}
613
614static void start_wakeup_tracer(struct trace_array *tr)
615{
616 int ret;
617
618 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
619 if (ret) {
620 pr_info("wakeup trace: Couldn't activate tracepoint"
621 " probe to kernel_sched_wakeup\n");
622 return;
623 }
624
625 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
626 if (ret) {
627 pr_info("wakeup trace: Couldn't activate tracepoint"
628 " probe to kernel_sched_wakeup_new\n");
629 goto fail_deprobe;
630 }
631
632 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
633 if (ret) {
634 pr_info("sched trace: Couldn't activate tracepoint"
635 " probe to kernel_sched_switch\n");
636 goto fail_deprobe_wake_new;
637 }
638
639 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
640 if (ret) {
641 pr_info("wakeup trace: Couldn't activate tracepoint"
642 " probe to kernel_sched_migrate_task\n");
643 goto fail_deprobe_sched_switch;
644 }
645
646 wakeup_reset(tr);
647
648 /*
649 * Don't let the tracer_enabled = 1 show up before
650 * the wakeup_task is reset. This may be overkill since
651 * wakeup_reset does a spin_unlock after setting the
652 * wakeup_task to NULL, but I want to be safe.
653 * This is a slow path anyway.
654 */
655 smp_wmb();
656
657 if (start_func_tracer(tr, is_graph(tr)))
658 printk(KERN_ERR "failed to start wakeup tracer\n");
659
660 return;
661fail_deprobe_sched_switch:
662 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
663fail_deprobe_wake_new:
664 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
665fail_deprobe:
666 unregister_trace_sched_wakeup(probe_wakeup, NULL);
667}
668
669static void stop_wakeup_tracer(struct trace_array *tr)
670{
671 tracer_enabled = 0;
672 stop_func_tracer(tr, is_graph(tr));
673 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
674 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
675 unregister_trace_sched_wakeup(probe_wakeup, NULL);
676 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
677}
678
679static bool wakeup_busy;
680
681static int __wakeup_tracer_init(struct trace_array *tr)
682{
683 save_flags = tr->trace_flags;
684
685 /* non overwrite screws up the latency tracers */
686 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
687 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
688
689 tr->max_latency = 0;
690 wakeup_trace = tr;
691 ftrace_init_array_ops(tr, wakeup_tracer_call);
692 start_wakeup_tracer(tr);
693
694 wakeup_busy = true;
695 return 0;
696}
697
698static int wakeup_tracer_init(struct trace_array *tr)
699{
700 if (wakeup_busy)
701 return -EBUSY;
702
703 wakeup_dl = false;
704 wakeup_rt = false;
705 return __wakeup_tracer_init(tr);
706}
707
708static int wakeup_rt_tracer_init(struct trace_array *tr)
709{
710 if (wakeup_busy)
711 return -EBUSY;
712
713 wakeup_dl = false;
714 wakeup_rt = true;
715 return __wakeup_tracer_init(tr);
716}
717
718static int wakeup_dl_tracer_init(struct trace_array *tr)
719{
720 if (wakeup_busy)
721 return -EBUSY;
722
723 wakeup_dl = true;
724 wakeup_rt = false;
725 return __wakeup_tracer_init(tr);
726}
727
728static void wakeup_tracer_reset(struct trace_array *tr)
729{
730 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
731 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
732
733 stop_wakeup_tracer(tr);
734 /* make sure we put back any tasks we are tracing */
735 wakeup_reset(tr);
736
737 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
738 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
739 ftrace_reset_array_ops(tr);
740 wakeup_busy = false;
741}
742
743static void wakeup_tracer_start(struct trace_array *tr)
744{
745 wakeup_reset(tr);
746 tracer_enabled = 1;
747}
748
749static void wakeup_tracer_stop(struct trace_array *tr)
750{
751 tracer_enabled = 0;
752}
753
754static struct tracer wakeup_tracer __read_mostly =
755{
756 .name = "wakeup",
757 .init = wakeup_tracer_init,
758 .reset = wakeup_tracer_reset,
759 .start = wakeup_tracer_start,
760 .stop = wakeup_tracer_stop,
761 .print_max = true,
762 .print_header = wakeup_print_header,
763 .print_line = wakeup_print_line,
764 .flag_changed = wakeup_flag_changed,
765#ifdef CONFIG_FTRACE_SELFTEST
766 .selftest = trace_selftest_startup_wakeup,
767#endif
768 .open = wakeup_trace_open,
769 .close = wakeup_trace_close,
770 .allow_instances = true,
771 .use_max_tr = true,
772};
773
774static struct tracer wakeup_rt_tracer __read_mostly =
775{
776 .name = "wakeup_rt",
777 .init = wakeup_rt_tracer_init,
778 .reset = wakeup_tracer_reset,
779 .start = wakeup_tracer_start,
780 .stop = wakeup_tracer_stop,
781 .print_max = true,
782 .print_header = wakeup_print_header,
783 .print_line = wakeup_print_line,
784 .flag_changed = wakeup_flag_changed,
785#ifdef CONFIG_FTRACE_SELFTEST
786 .selftest = trace_selftest_startup_wakeup,
787#endif
788 .open = wakeup_trace_open,
789 .close = wakeup_trace_close,
790 .allow_instances = true,
791 .use_max_tr = true,
792};
793
794static struct tracer wakeup_dl_tracer __read_mostly =
795{
796 .name = "wakeup_dl",
797 .init = wakeup_dl_tracer_init,
798 .reset = wakeup_tracer_reset,
799 .start = wakeup_tracer_start,
800 .stop = wakeup_tracer_stop,
801 .print_max = true,
802 .print_header = wakeup_print_header,
803 .print_line = wakeup_print_line,
804 .flag_changed = wakeup_flag_changed,
805#ifdef CONFIG_FTRACE_SELFTEST
806 .selftest = trace_selftest_startup_wakeup,
807#endif
808 .open = wakeup_trace_open,
809 .close = wakeup_trace_close,
810 .allow_instances = true,
811 .use_max_tr = true,
812};
813
814__init static int init_wakeup_tracer(void)
815{
816 int ret;
817
818 ret = register_tracer(&wakeup_tracer);
819 if (ret)
820 return ret;
821
822 ret = register_tracer(&wakeup_rt_tracer);
823 if (ret)
824 return ret;
825
826 ret = register_tracer(&wakeup_dl_tracer);
827 if (ret)
828 return ret;
829
830 return 0;
831}
832core_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_trace_open(struct trace_iterator *iter) { }
284static void wakeup_trace_close(struct trace_iterator *iter) { }
285
286#ifdef CONFIG_FUNCTION_TRACER
287static void wakeup_print_header(struct seq_file *s)
288{
289 trace_default_header(s);
290}
291#else
292static void wakeup_print_header(struct seq_file *s)
293{
294 trace_latency_header(s);
295}
296#endif /* CONFIG_FUNCTION_TRACER */
297#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
298
299/*
300 * Should this new latency be reported/recorded?
301 */
302static int report_latency(cycle_t delta)
303{
304 if (tracing_thresh) {
305 if (delta < tracing_thresh)
306 return 0;
307 } else {
308 if (delta <= tracing_max_latency)
309 return 0;
310 }
311 return 1;
312}
313
314static void
315probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
316{
317 if (task != wakeup_task)
318 return;
319
320 wakeup_current_cpu = cpu;
321}
322
323static void notrace
324probe_wakeup_sched_switch(void *ignore,
325 struct task_struct *prev, struct task_struct *next)
326{
327 struct trace_array_cpu *data;
328 cycle_t T0, T1, delta;
329 unsigned long flags;
330 long disabled;
331 int cpu;
332 int pc;
333
334 tracing_record_cmdline(prev);
335
336 if (unlikely(!tracer_enabled))
337 return;
338
339 /*
340 * When we start a new trace, we set wakeup_task to NULL
341 * and then set tracer_enabled = 1. We want to make sure
342 * that another CPU does not see the tracer_enabled = 1
343 * and the wakeup_task with an older task, that might
344 * actually be the same as next.
345 */
346 smp_rmb();
347
348 if (next != wakeup_task)
349 return;
350
351 pc = preempt_count();
352
353 /* disable local data, not wakeup_cpu data */
354 cpu = raw_smp_processor_id();
355 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
356 if (likely(disabled != 1))
357 goto out;
358
359 local_irq_save(flags);
360 arch_spin_lock(&wakeup_lock);
361
362 /* We could race with grabbing wakeup_lock */
363 if (unlikely(!tracer_enabled || next != wakeup_task))
364 goto out_unlock;
365
366 /* The task we are waiting for is waking up */
367 data = wakeup_trace->data[wakeup_cpu];
368
369 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
370 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
371
372 T0 = data->preempt_timestamp;
373 T1 = ftrace_now(cpu);
374 delta = T1-T0;
375
376 if (!report_latency(delta))
377 goto out_unlock;
378
379 if (likely(!is_tracing_stopped())) {
380 tracing_max_latency = delta;
381 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
382 }
383
384out_unlock:
385 __wakeup_reset(wakeup_trace);
386 arch_spin_unlock(&wakeup_lock);
387 local_irq_restore(flags);
388out:
389 atomic_dec(&wakeup_trace->data[cpu]->disabled);
390}
391
392static void __wakeup_reset(struct trace_array *tr)
393{
394 wakeup_cpu = -1;
395 wakeup_prio = -1;
396
397 if (wakeup_task)
398 put_task_struct(wakeup_task);
399
400 wakeup_task = NULL;
401}
402
403static void wakeup_reset(struct trace_array *tr)
404{
405 unsigned long flags;
406
407 tracing_reset_online_cpus(tr);
408
409 local_irq_save(flags);
410 arch_spin_lock(&wakeup_lock);
411 __wakeup_reset(tr);
412 arch_spin_unlock(&wakeup_lock);
413 local_irq_restore(flags);
414}
415
416static void
417probe_wakeup(void *ignore, struct task_struct *p, int success)
418{
419 struct trace_array_cpu *data;
420 int cpu = smp_processor_id();
421 unsigned long flags;
422 long disabled;
423 int pc;
424
425 if (likely(!tracer_enabled))
426 return;
427
428 tracing_record_cmdline(p);
429 tracing_record_cmdline(current);
430
431 if ((wakeup_rt && !rt_task(p)) ||
432 p->prio >= wakeup_prio ||
433 p->prio >= current->prio)
434 return;
435
436 pc = preempt_count();
437 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
438 if (unlikely(disabled != 1))
439 goto out;
440
441 /* interrupts should be off from try_to_wake_up */
442 arch_spin_lock(&wakeup_lock);
443
444 /* check for races. */
445 if (!tracer_enabled || p->prio >= wakeup_prio)
446 goto out_locked;
447
448 /* reset the trace */
449 __wakeup_reset(wakeup_trace);
450
451 wakeup_cpu = task_cpu(p);
452 wakeup_current_cpu = wakeup_cpu;
453 wakeup_prio = p->prio;
454
455 wakeup_task = p;
456 get_task_struct(wakeup_task);
457
458 local_save_flags(flags);
459
460 data = wakeup_trace->data[wakeup_cpu];
461 data->preempt_timestamp = ftrace_now(cpu);
462 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
463
464 /*
465 * We must be careful in using CALLER_ADDR2. But since wake_up
466 * is not called by an assembly function (where as schedule is)
467 * it should be safe to use it here.
468 */
469 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
470
471out_locked:
472 arch_spin_unlock(&wakeup_lock);
473out:
474 atomic_dec(&wakeup_trace->data[cpu]->disabled);
475}
476
477static void start_wakeup_tracer(struct trace_array *tr)
478{
479 int ret;
480
481 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
482 if (ret) {
483 pr_info("wakeup trace: Couldn't activate tracepoint"
484 " probe to kernel_sched_wakeup\n");
485 return;
486 }
487
488 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
489 if (ret) {
490 pr_info("wakeup trace: Couldn't activate tracepoint"
491 " probe to kernel_sched_wakeup_new\n");
492 goto fail_deprobe;
493 }
494
495 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
496 if (ret) {
497 pr_info("sched trace: Couldn't activate tracepoint"
498 " probe to kernel_sched_switch\n");
499 goto fail_deprobe_wake_new;
500 }
501
502 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
503 if (ret) {
504 pr_info("wakeup trace: Couldn't activate tracepoint"
505 " probe to kernel_sched_migrate_task\n");
506 return;
507 }
508
509 wakeup_reset(tr);
510
511 /*
512 * Don't let the tracer_enabled = 1 show up before
513 * the wakeup_task is reset. This may be overkill since
514 * wakeup_reset does a spin_unlock after setting the
515 * wakeup_task to NULL, but I want to be safe.
516 * This is a slow path anyway.
517 */
518 smp_wmb();
519
520 if (start_func_tracer(is_graph()))
521 printk(KERN_ERR "failed to start wakeup tracer\n");
522
523 return;
524fail_deprobe_wake_new:
525 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
526fail_deprobe:
527 unregister_trace_sched_wakeup(probe_wakeup, NULL);
528}
529
530static void stop_wakeup_tracer(struct trace_array *tr)
531{
532 tracer_enabled = 0;
533 stop_func_tracer(is_graph());
534 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
535 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
536 unregister_trace_sched_wakeup(probe_wakeup, NULL);
537 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
538}
539
540static int __wakeup_tracer_init(struct trace_array *tr)
541{
542 save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
543 trace_flags |= TRACE_ITER_LATENCY_FMT;
544
545 tracing_max_latency = 0;
546 wakeup_trace = tr;
547 start_wakeup_tracer(tr);
548 return 0;
549}
550
551static int wakeup_tracer_init(struct trace_array *tr)
552{
553 wakeup_rt = 0;
554 return __wakeup_tracer_init(tr);
555}
556
557static int wakeup_rt_tracer_init(struct trace_array *tr)
558{
559 wakeup_rt = 1;
560 return __wakeup_tracer_init(tr);
561}
562
563static void wakeup_tracer_reset(struct trace_array *tr)
564{
565 stop_wakeup_tracer(tr);
566 /* make sure we put back any tasks we are tracing */
567 wakeup_reset(tr);
568
569 if (!save_lat_flag)
570 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
571}
572
573static void wakeup_tracer_start(struct trace_array *tr)
574{
575 wakeup_reset(tr);
576 tracer_enabled = 1;
577}
578
579static void wakeup_tracer_stop(struct trace_array *tr)
580{
581 tracer_enabled = 0;
582}
583
584static struct tracer wakeup_tracer __read_mostly =
585{
586 .name = "wakeup",
587 .init = wakeup_tracer_init,
588 .reset = wakeup_tracer_reset,
589 .start = wakeup_tracer_start,
590 .stop = wakeup_tracer_stop,
591 .print_max = 1,
592 .print_header = wakeup_print_header,
593 .print_line = wakeup_print_line,
594 .flags = &tracer_flags,
595 .set_flag = wakeup_set_flag,
596#ifdef CONFIG_FTRACE_SELFTEST
597 .selftest = trace_selftest_startup_wakeup,
598#endif
599 .open = wakeup_trace_open,
600 .close = wakeup_trace_close,
601 .use_max_tr = 1,
602};
603
604static struct tracer wakeup_rt_tracer __read_mostly =
605{
606 .name = "wakeup_rt",
607 .init = wakeup_rt_tracer_init,
608 .reset = wakeup_tracer_reset,
609 .start = wakeup_tracer_start,
610 .stop = wakeup_tracer_stop,
611 .wait_pipe = poll_wait_pipe,
612 .print_max = 1,
613 .print_header = wakeup_print_header,
614 .print_line = wakeup_print_line,
615 .flags = &tracer_flags,
616 .set_flag = wakeup_set_flag,
617#ifdef CONFIG_FTRACE_SELFTEST
618 .selftest = trace_selftest_startup_wakeup,
619#endif
620 .open = wakeup_trace_open,
621 .close = wakeup_trace_close,
622 .use_max_tr = 1,
623};
624
625__init static int init_wakeup_tracer(void)
626{
627 int ret;
628
629 ret = register_tracer(&wakeup_tracer);
630 if (ret)
631 return ret;
632
633 ret = register_tracer(&wakeup_rt_tracer);
634 if (ret)
635 return ret;
636
637 return 0;
638}
639device_initcall(init_wakeup_tracer);