2 * trace task wakeup timings
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
7 * Based on code from the latency_tracer, that is:
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 Nadia Yvette Chambers
12 #include <linux/module.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>
23 static struct trace_array
*wakeup_trace
;
24 static int __read_mostly tracer_enabled
;
26 static struct task_struct
*wakeup_task
;
27 static int wakeup_cpu
;
28 static int wakeup_current_cpu
;
29 static unsigned wakeup_prio
= -1;
32 static int tracing_dl
= 0;
34 static arch_spinlock_t wakeup_lock
=
35 (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
37 static void wakeup_reset(struct trace_array
*tr
);
38 static void __wakeup_reset(struct trace_array
*tr
);
39 static int wakeup_graph_entry(struct ftrace_graph_ent
*trace
);
40 static void wakeup_graph_return(struct ftrace_graph_ret
*trace
);
42 static int save_flags
;
43 static bool function_enabled
;
45 #define TRACE_DISPLAY_GRAPH 1
47 static 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
) },
55 static struct tracer_flags tracer_flags
= {
60 #define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH)
62 #ifdef CONFIG_FUNCTION_TRACER
65 * Prologue for the wakeup function tracers.
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
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.
79 func_prolog_preempt_disable(struct trace_array
*tr
,
80 struct trace_array_cpu
**data
,
86 if (likely(!wakeup_task
))
89 *pc
= preempt_count();
90 preempt_disable_notrace();
92 cpu
= raw_smp_processor_id();
93 if (cpu
!= wakeup_current_cpu
)
96 *data
= per_cpu_ptr(tr
->trace_buffer
.data
, cpu
);
97 disabled
= atomic_inc_return(&(*data
)->disabled
);
98 if (unlikely(disabled
!= 1))
104 atomic_dec(&(*data
)->disabled
);
107 preempt_enable_notrace();
112 * wakeup uses its own tracer function to keep the overhead down:
115 wakeup_tracer_call(unsigned long ip
, unsigned long parent_ip
,
116 struct ftrace_ops
*op
, struct pt_regs
*pt_regs
)
118 struct trace_array
*tr
= wakeup_trace
;
119 struct trace_array_cpu
*data
;
123 if (!func_prolog_preempt_disable(tr
, &data
, &pc
))
126 local_irq_save(flags
);
127 trace_function(tr
, ip
, parent_ip
, flags
, pc
);
128 local_irq_restore(flags
);
130 atomic_dec(&data
->disabled
);
131 preempt_enable_notrace();
134 static struct ftrace_ops trace_ops __read_mostly
=
136 .func
= wakeup_tracer_call
,
137 .flags
= FTRACE_OPS_FL_GLOBAL
| FTRACE_OPS_FL_RECURSION_SAFE
,
139 #endif /* CONFIG_FUNCTION_TRACER */
141 static int register_wakeup_function(int graph
, int set
)
145 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
146 if (function_enabled
|| (!set
&& !(trace_flags
& TRACE_ITER_FUNCTION
)))
150 ret
= register_ftrace_graph(&wakeup_graph_return
,
151 &wakeup_graph_entry
);
153 ret
= register_ftrace_function(&trace_ops
);
156 function_enabled
= true;
161 static void unregister_wakeup_function(int graph
)
163 if (!function_enabled
)
167 unregister_ftrace_graph();
169 unregister_ftrace_function(&trace_ops
);
171 function_enabled
= false;
174 static void wakeup_function_set(int set
)
177 register_wakeup_function(is_graph(), 1);
179 unregister_wakeup_function(is_graph());
182 static int wakeup_flag_changed(struct trace_array
*tr
, u32 mask
, int set
)
184 struct tracer
*tracer
= tr
->current_trace
;
186 if (mask
& TRACE_ITER_FUNCTION
)
187 wakeup_function_set(set
);
189 return trace_keep_overwrite(tracer
, mask
, set
);
192 static int start_func_tracer(int graph
)
196 ret
= register_wakeup_function(graph
, 0);
198 if (!ret
&& tracing_is_enabled())
206 static void stop_func_tracer(int graph
)
210 unregister_wakeup_function(graph
);
213 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
215 wakeup_set_flag(struct trace_array
*tr
, u32 old_flags
, u32 bit
, int set
)
218 if (!(bit
& TRACE_DISPLAY_GRAPH
))
221 if (!(is_graph() ^ set
))
224 stop_func_tracer(!set
);
226 wakeup_reset(wakeup_trace
);
227 tracing_max_latency
= 0;
229 return start_func_tracer(set
);
232 static int wakeup_graph_entry(struct ftrace_graph_ent
*trace
)
234 struct trace_array
*tr
= wakeup_trace
;
235 struct trace_array_cpu
*data
;
239 if (!func_prolog_preempt_disable(tr
, &data
, &pc
))
242 local_save_flags(flags
);
243 ret
= __trace_graph_entry(tr
, trace
, flags
, pc
);
244 atomic_dec(&data
->disabled
);
245 preempt_enable_notrace();
250 static void wakeup_graph_return(struct ftrace_graph_ret
*trace
)
252 struct trace_array
*tr
= wakeup_trace
;
253 struct trace_array_cpu
*data
;
257 if (!func_prolog_preempt_disable(tr
, &data
, &pc
))
260 local_save_flags(flags
);
261 __trace_graph_return(tr
, trace
, flags
, pc
);
262 atomic_dec(&data
->disabled
);
264 preempt_enable_notrace();
268 static void wakeup_trace_open(struct trace_iterator
*iter
)
271 graph_trace_open(iter
);
274 static void wakeup_trace_close(struct trace_iterator
*iter
)
277 graph_trace_close(iter
);
280 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
281 TRACE_GRAPH_PRINT_ABS_TIME | \
282 TRACE_GRAPH_PRINT_DURATION)
284 static enum print_line_t
wakeup_print_line(struct trace_iterator
*iter
)
287 * In graph mode call the graph tracer output function,
288 * otherwise go with the TRACE_FN event handler
291 return print_graph_function_flags(iter
, GRAPH_TRACER_FLAGS
);
293 return TRACE_TYPE_UNHANDLED
;
296 static void wakeup_print_header(struct seq_file
*s
)
299 print_graph_headers_flags(s
, GRAPH_TRACER_FLAGS
);
301 trace_default_header(s
);
305 __trace_function(struct trace_array
*tr
,
306 unsigned long ip
, unsigned long parent_ip
,
307 unsigned long flags
, int pc
)
310 trace_graph_function(tr
, ip
, parent_ip
, flags
, pc
);
312 trace_function(tr
, ip
, parent_ip
, flags
, pc
);
315 #define __trace_function trace_function
318 wakeup_set_flag(struct trace_array
*tr
, u32 old_flags
, u32 bit
, int set
)
323 static int wakeup_graph_entry(struct ftrace_graph_ent
*trace
)
328 static enum print_line_t
wakeup_print_line(struct trace_iterator
*iter
)
330 return TRACE_TYPE_UNHANDLED
;
333 static void wakeup_graph_return(struct ftrace_graph_ret
*trace
) { }
334 static void wakeup_trace_open(struct trace_iterator
*iter
) { }
335 static void wakeup_trace_close(struct trace_iterator
*iter
) { }
337 #ifdef CONFIG_FUNCTION_TRACER
338 static void wakeup_print_header(struct seq_file
*s
)
340 trace_default_header(s
);
343 static void wakeup_print_header(struct seq_file
*s
)
345 trace_latency_header(s
);
347 #endif /* CONFIG_FUNCTION_TRACER */
348 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
351 * Should this new latency be reported/recorded?
353 static int report_latency(cycle_t delta
)
355 if (tracing_thresh
) {
356 if (delta
< tracing_thresh
)
359 if (delta
<= tracing_max_latency
)
366 probe_wakeup_migrate_task(void *ignore
, struct task_struct
*task
, int cpu
)
368 if (task
!= wakeup_task
)
371 wakeup_current_cpu
= cpu
;
375 probe_wakeup_sched_switch(void *ignore
,
376 struct task_struct
*prev
, struct task_struct
*next
)
378 struct trace_array_cpu
*data
;
379 cycle_t T0
, T1
, delta
;
385 tracing_record_cmdline(prev
);
387 if (unlikely(!tracer_enabled
))
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.
399 if (next
!= wakeup_task
)
402 pc
= preempt_count();
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))
410 local_irq_save(flags
);
411 arch_spin_lock(&wakeup_lock
);
413 /* We could race with grabbing wakeup_lock */
414 if (unlikely(!tracer_enabled
|| next
!= wakeup_task
))
417 /* The task we are waiting for is waking up */
418 data
= per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, wakeup_cpu
);
420 __trace_function(wakeup_trace
, CALLER_ADDR0
, CALLER_ADDR1
, flags
, pc
);
421 tracing_sched_switch_trace(wakeup_trace
, prev
, next
, flags
, pc
);
423 T0
= data
->preempt_timestamp
;
424 T1
= ftrace_now(cpu
);
427 if (!report_latency(delta
))
430 if (likely(!is_tracing_stopped())) {
431 tracing_max_latency
= delta
;
432 update_max_tr(wakeup_trace
, wakeup_task
, wakeup_cpu
);
436 __wakeup_reset(wakeup_trace
);
437 arch_spin_unlock(&wakeup_lock
);
438 local_irq_restore(flags
);
440 atomic_dec(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
443 static void __wakeup_reset(struct trace_array
*tr
)
450 put_task_struct(wakeup_task
);
455 static void wakeup_reset(struct trace_array
*tr
)
459 tracing_reset_online_cpus(&tr
->trace_buffer
);
461 local_irq_save(flags
);
462 arch_spin_lock(&wakeup_lock
);
464 arch_spin_unlock(&wakeup_lock
);
465 local_irq_restore(flags
);
469 probe_wakeup(void *ignore
, struct task_struct
*p
, int success
)
471 struct trace_array_cpu
*data
;
472 int cpu
= smp_processor_id();
477 if (likely(!tracer_enabled
))
480 tracing_record_cmdline(p
);
481 tracing_record_cmdline(current
);
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
489 * - wakeup_dl handles tasks belonging to sched_dl class only.
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
)))
496 pc
= preempt_count();
497 disabled
= atomic_inc_return(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
498 if (unlikely(disabled
!= 1))
501 /* interrupts should be off from try_to_wake_up */
502 arch_spin_lock(&wakeup_lock
);
504 /* check for races. */
505 if (!tracer_enabled
|| tracing_dl
||
506 (!dl_task(p
) && p
->prio
>= wakeup_prio
))
509 /* reset the trace */
510 __wakeup_reset(wakeup_trace
);
512 wakeup_cpu
= task_cpu(p
);
513 wakeup_current_cpu
= wakeup_cpu
;
514 wakeup_prio
= p
->prio
;
517 * Once you start tracing a -deadline task, don't bother tracing
518 * another task until the first one wakes up.
526 get_task_struct(wakeup_task
);
528 local_save_flags(flags
);
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
);
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.
539 __trace_function(wakeup_trace
, CALLER_ADDR1
, CALLER_ADDR2
, flags
, pc
);
542 arch_spin_unlock(&wakeup_lock
);
544 atomic_dec(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
547 static void start_wakeup_tracer(struct trace_array
*tr
)
551 ret
= register_trace_sched_wakeup(probe_wakeup
, NULL
);
553 pr_info("wakeup trace: Couldn't activate tracepoint"
554 " probe to kernel_sched_wakeup\n");
558 ret
= register_trace_sched_wakeup_new(probe_wakeup
, NULL
);
560 pr_info("wakeup trace: Couldn't activate tracepoint"
561 " probe to kernel_sched_wakeup_new\n");
565 ret
= register_trace_sched_switch(probe_wakeup_sched_switch
, NULL
);
567 pr_info("sched trace: Couldn't activate tracepoint"
568 " probe to kernel_sched_switch\n");
569 goto fail_deprobe_wake_new
;
572 ret
= register_trace_sched_migrate_task(probe_wakeup_migrate_task
, NULL
);
574 pr_info("wakeup trace: Couldn't activate tracepoint"
575 " probe to kernel_sched_migrate_task\n");
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.
590 if (start_func_tracer(is_graph()))
591 printk(KERN_ERR
"failed to start wakeup tracer\n");
594 fail_deprobe_wake_new
:
595 unregister_trace_sched_wakeup_new(probe_wakeup
, NULL
);
597 unregister_trace_sched_wakeup(probe_wakeup
, NULL
);
600 static void stop_wakeup_tracer(struct trace_array
*tr
)
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
);
610 static int __wakeup_tracer_init(struct trace_array
*tr
)
612 save_flags
= trace_flags
;
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);
618 tracing_max_latency
= 0;
620 start_wakeup_tracer(tr
);
624 static int wakeup_tracer_init(struct trace_array
*tr
)
628 return __wakeup_tracer_init(tr
);
631 static int wakeup_rt_tracer_init(struct trace_array
*tr
)
635 return __wakeup_tracer_init(tr
);
638 static int wakeup_dl_tracer_init(struct trace_array
*tr
)
642 return __wakeup_tracer_init(tr
);
645 static void wakeup_tracer_reset(struct trace_array
*tr
)
647 int lat_flag
= save_flags
& TRACE_ITER_LATENCY_FMT
;
648 int overwrite_flag
= save_flags
& TRACE_ITER_OVERWRITE
;
650 stop_wakeup_tracer(tr
);
651 /* make sure we put back any tasks we are tracing */
654 set_tracer_flag(tr
, TRACE_ITER_LATENCY_FMT
, lat_flag
);
655 set_tracer_flag(tr
, TRACE_ITER_OVERWRITE
, overwrite_flag
);
658 static void wakeup_tracer_start(struct trace_array
*tr
)
664 static void wakeup_tracer_stop(struct trace_array
*tr
)
669 static struct tracer wakeup_tracer __read_mostly
=
672 .init
= wakeup_tracer_init
,
673 .reset
= wakeup_tracer_reset
,
674 .start
= wakeup_tracer_start
,
675 .stop
= wakeup_tracer_stop
,
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
,
685 .open
= wakeup_trace_open
,
686 .close
= wakeup_trace_close
,
690 static struct tracer wakeup_rt_tracer __read_mostly
=
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
,
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
,
707 .open
= wakeup_trace_open
,
708 .close
= wakeup_trace_close
,
712 static struct tracer wakeup_dl_tracer __read_mostly
=
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
,
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
,
729 .open
= wakeup_trace_open
,
730 .close
= wakeup_trace_close
,
734 __init
static int init_wakeup_tracer(void)
738 ret
= register_tracer(&wakeup_tracer
);
742 ret
= register_tracer(&wakeup_rt_tracer
);
746 ret
= register_tracer(&wakeup_dl_tracer
);
752 core_initcall(init_wakeup_tracer
);