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