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 tracer
*tracer
, u32 mask
, int set
)
184 if (mask
& TRACE_ITER_FUNCTION
)
185 wakeup_function_set(set
);
187 return trace_keep_overwrite(tracer
, mask
, set
);
190 static int start_func_tracer(int graph
)
194 ret
= register_wakeup_function(graph
, 0);
196 if (!ret
&& tracing_is_enabled())
204 static void stop_func_tracer(int graph
)
208 unregister_wakeup_function(graph
);
211 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
212 static int wakeup_set_flag(u32 old_flags
, u32 bit
, int set
)
215 if (!(bit
& TRACE_DISPLAY_GRAPH
))
218 if (!(is_graph() ^ set
))
221 stop_func_tracer(!set
);
223 wakeup_reset(wakeup_trace
);
224 tracing_max_latency
= 0;
226 return start_func_tracer(set
);
229 static int wakeup_graph_entry(struct ftrace_graph_ent
*trace
)
231 struct trace_array
*tr
= wakeup_trace
;
232 struct trace_array_cpu
*data
;
236 if (!func_prolog_preempt_disable(tr
, &data
, &pc
))
239 local_save_flags(flags
);
240 ret
= __trace_graph_entry(tr
, trace
, flags
, pc
);
241 atomic_dec(&data
->disabled
);
242 preempt_enable_notrace();
247 static void wakeup_graph_return(struct ftrace_graph_ret
*trace
)
249 struct trace_array
*tr
= wakeup_trace
;
250 struct trace_array_cpu
*data
;
254 if (!func_prolog_preempt_disable(tr
, &data
, &pc
))
257 local_save_flags(flags
);
258 __trace_graph_return(tr
, trace
, flags
, pc
);
259 atomic_dec(&data
->disabled
);
261 preempt_enable_notrace();
265 static void wakeup_trace_open(struct trace_iterator
*iter
)
268 graph_trace_open(iter
);
271 static void wakeup_trace_close(struct trace_iterator
*iter
)
274 graph_trace_close(iter
);
277 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
278 TRACE_GRAPH_PRINT_ABS_TIME | \
279 TRACE_GRAPH_PRINT_DURATION)
281 static enum print_line_t
wakeup_print_line(struct trace_iterator
*iter
)
284 * In graph mode call the graph tracer output function,
285 * otherwise go with the TRACE_FN event handler
288 return print_graph_function_flags(iter
, GRAPH_TRACER_FLAGS
);
290 return TRACE_TYPE_UNHANDLED
;
293 static void wakeup_print_header(struct seq_file
*s
)
296 print_graph_headers_flags(s
, GRAPH_TRACER_FLAGS
);
298 trace_default_header(s
);
302 __trace_function(struct trace_array
*tr
,
303 unsigned long ip
, unsigned long parent_ip
,
304 unsigned long flags
, int pc
)
307 trace_graph_function(tr
, ip
, parent_ip
, flags
, pc
);
309 trace_function(tr
, ip
, parent_ip
, flags
, pc
);
312 #define __trace_function trace_function
314 static int wakeup_set_flag(u32 old_flags
, u32 bit
, int set
)
319 static int wakeup_graph_entry(struct ftrace_graph_ent
*trace
)
324 static enum print_line_t
wakeup_print_line(struct trace_iterator
*iter
)
326 return TRACE_TYPE_UNHANDLED
;
329 static void wakeup_graph_return(struct ftrace_graph_ret
*trace
) { }
330 static void wakeup_trace_open(struct trace_iterator
*iter
) { }
331 static void wakeup_trace_close(struct trace_iterator
*iter
) { }
333 #ifdef CONFIG_FUNCTION_TRACER
334 static void wakeup_print_header(struct seq_file
*s
)
336 trace_default_header(s
);
339 static void wakeup_print_header(struct seq_file
*s
)
341 trace_latency_header(s
);
343 #endif /* CONFIG_FUNCTION_TRACER */
344 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
347 * Should this new latency be reported/recorded?
349 static int report_latency(cycle_t delta
)
351 if (tracing_thresh
) {
352 if (delta
< tracing_thresh
)
355 if (delta
<= tracing_max_latency
)
362 probe_wakeup_migrate_task(void *ignore
, struct task_struct
*task
, int cpu
)
364 if (task
!= wakeup_task
)
367 wakeup_current_cpu
= cpu
;
371 probe_wakeup_sched_switch(void *ignore
,
372 struct task_struct
*prev
, struct task_struct
*next
)
374 struct trace_array_cpu
*data
;
375 cycle_t T0
, T1
, delta
;
381 tracing_record_cmdline(prev
);
383 if (unlikely(!tracer_enabled
))
387 * When we start a new trace, we set wakeup_task to NULL
388 * and then set tracer_enabled = 1. We want to make sure
389 * that another CPU does not see the tracer_enabled = 1
390 * and the wakeup_task with an older task, that might
391 * actually be the same as next.
395 if (next
!= wakeup_task
)
398 pc
= preempt_count();
400 /* disable local data, not wakeup_cpu data */
401 cpu
= raw_smp_processor_id();
402 disabled
= atomic_inc_return(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
403 if (likely(disabled
!= 1))
406 local_irq_save(flags
);
407 arch_spin_lock(&wakeup_lock
);
409 /* We could race with grabbing wakeup_lock */
410 if (unlikely(!tracer_enabled
|| next
!= wakeup_task
))
413 /* The task we are waiting for is waking up */
414 data
= per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, wakeup_cpu
);
416 __trace_function(wakeup_trace
, CALLER_ADDR0
, CALLER_ADDR1
, flags
, pc
);
417 tracing_sched_switch_trace(wakeup_trace
, prev
, next
, flags
, pc
);
419 T0
= data
->preempt_timestamp
;
420 T1
= ftrace_now(cpu
);
423 if (!report_latency(delta
))
426 if (likely(!is_tracing_stopped())) {
427 tracing_max_latency
= delta
;
428 update_max_tr(wakeup_trace
, wakeup_task
, wakeup_cpu
);
432 __wakeup_reset(wakeup_trace
);
433 arch_spin_unlock(&wakeup_lock
);
434 local_irq_restore(flags
);
436 atomic_dec(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
439 static void __wakeup_reset(struct trace_array
*tr
)
446 put_task_struct(wakeup_task
);
451 static void wakeup_reset(struct trace_array
*tr
)
455 tracing_reset_online_cpus(&tr
->trace_buffer
);
457 local_irq_save(flags
);
458 arch_spin_lock(&wakeup_lock
);
460 arch_spin_unlock(&wakeup_lock
);
461 local_irq_restore(flags
);
465 probe_wakeup(void *ignore
, struct task_struct
*p
, int success
)
467 struct trace_array_cpu
*data
;
468 int cpu
= smp_processor_id();
473 if (likely(!tracer_enabled
))
476 tracing_record_cmdline(p
);
477 tracing_record_cmdline(current
);
480 * Semantic is like this:
481 * - wakeup tracer handles all tasks in the system, independently
482 * from their scheduling class;
483 * - wakeup_rt tracer handles tasks belonging to sched_dl and
485 * - wakeup_dl handles tasks belonging to sched_dl class only.
487 if (tracing_dl
|| (wakeup_dl
&& !dl_task(p
)) ||
488 (wakeup_rt
&& !dl_task(p
) && !rt_task(p
)) ||
489 (!dl_task(p
) && (p
->prio
>= wakeup_prio
|| p
->prio
>= current
->prio
)))
492 pc
= preempt_count();
493 disabled
= atomic_inc_return(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
494 if (unlikely(disabled
!= 1))
497 /* interrupts should be off from try_to_wake_up */
498 arch_spin_lock(&wakeup_lock
);
500 /* check for races. */
501 if (!tracer_enabled
|| tracing_dl
||
502 (!dl_task(p
) && p
->prio
>= wakeup_prio
))
505 /* reset the trace */
506 __wakeup_reset(wakeup_trace
);
508 wakeup_cpu
= task_cpu(p
);
509 wakeup_current_cpu
= wakeup_cpu
;
510 wakeup_prio
= p
->prio
;
513 * Once you start tracing a -deadline task, don't bother tracing
514 * another task until the first one wakes up.
522 get_task_struct(wakeup_task
);
524 local_save_flags(flags
);
526 data
= per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, wakeup_cpu
);
527 data
->preempt_timestamp
= ftrace_now(cpu
);
528 tracing_sched_wakeup_trace(wakeup_trace
, p
, current
, flags
, pc
);
531 * We must be careful in using CALLER_ADDR2. But since wake_up
532 * is not called by an assembly function (where as schedule is)
533 * it should be safe to use it here.
535 __trace_function(wakeup_trace
, CALLER_ADDR1
, CALLER_ADDR2
, flags
, pc
);
538 arch_spin_unlock(&wakeup_lock
);
540 atomic_dec(&per_cpu_ptr(wakeup_trace
->trace_buffer
.data
, cpu
)->disabled
);
543 static void start_wakeup_tracer(struct trace_array
*tr
)
547 ret
= register_trace_sched_wakeup(probe_wakeup
, NULL
);
549 pr_info("wakeup trace: Couldn't activate tracepoint"
550 " probe to kernel_sched_wakeup\n");
554 ret
= register_trace_sched_wakeup_new(probe_wakeup
, NULL
);
556 pr_info("wakeup trace: Couldn't activate tracepoint"
557 " probe to kernel_sched_wakeup_new\n");
561 ret
= register_trace_sched_switch(probe_wakeup_sched_switch
, NULL
);
563 pr_info("sched trace: Couldn't activate tracepoint"
564 " probe to kernel_sched_switch\n");
565 goto fail_deprobe_wake_new
;
568 ret
= register_trace_sched_migrate_task(probe_wakeup_migrate_task
, NULL
);
570 pr_info("wakeup trace: Couldn't activate tracepoint"
571 " probe to kernel_sched_migrate_task\n");
578 * Don't let the tracer_enabled = 1 show up before
579 * the wakeup_task is reset. This may be overkill since
580 * wakeup_reset does a spin_unlock after setting the
581 * wakeup_task to NULL, but I want to be safe.
582 * This is a slow path anyway.
586 if (start_func_tracer(is_graph()))
587 printk(KERN_ERR
"failed to start wakeup tracer\n");
590 fail_deprobe_wake_new
:
591 unregister_trace_sched_wakeup_new(probe_wakeup
, NULL
);
593 unregister_trace_sched_wakeup(probe_wakeup
, NULL
);
596 static void stop_wakeup_tracer(struct trace_array
*tr
)
599 stop_func_tracer(is_graph());
600 unregister_trace_sched_switch(probe_wakeup_sched_switch
, NULL
);
601 unregister_trace_sched_wakeup_new(probe_wakeup
, NULL
);
602 unregister_trace_sched_wakeup(probe_wakeup
, NULL
);
603 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task
, NULL
);
606 static int __wakeup_tracer_init(struct trace_array
*tr
)
608 save_flags
= trace_flags
;
610 /* non overwrite screws up the latency tracers */
611 set_tracer_flag(tr
, TRACE_ITER_OVERWRITE
, 1);
612 set_tracer_flag(tr
, TRACE_ITER_LATENCY_FMT
, 1);
614 tracing_max_latency
= 0;
616 start_wakeup_tracer(tr
);
620 static int wakeup_tracer_init(struct trace_array
*tr
)
624 return __wakeup_tracer_init(tr
);
627 static int wakeup_rt_tracer_init(struct trace_array
*tr
)
631 return __wakeup_tracer_init(tr
);
634 static int wakeup_dl_tracer_init(struct trace_array
*tr
)
638 return __wakeup_tracer_init(tr
);
641 static void wakeup_tracer_reset(struct trace_array
*tr
)
643 int lat_flag
= save_flags
& TRACE_ITER_LATENCY_FMT
;
644 int overwrite_flag
= save_flags
& TRACE_ITER_OVERWRITE
;
646 stop_wakeup_tracer(tr
);
647 /* make sure we put back any tasks we are tracing */
650 set_tracer_flag(tr
, TRACE_ITER_LATENCY_FMT
, lat_flag
);
651 set_tracer_flag(tr
, TRACE_ITER_OVERWRITE
, overwrite_flag
);
654 static void wakeup_tracer_start(struct trace_array
*tr
)
660 static void wakeup_tracer_stop(struct trace_array
*tr
)
665 static struct tracer wakeup_tracer __read_mostly
=
668 .init
= wakeup_tracer_init
,
669 .reset
= wakeup_tracer_reset
,
670 .start
= wakeup_tracer_start
,
671 .stop
= wakeup_tracer_stop
,
673 .print_header
= wakeup_print_header
,
674 .print_line
= wakeup_print_line
,
675 .flags
= &tracer_flags
,
676 .set_flag
= wakeup_set_flag
,
677 .flag_changed
= wakeup_flag_changed
,
678 #ifdef CONFIG_FTRACE_SELFTEST
679 .selftest
= trace_selftest_startup_wakeup
,
681 .open
= wakeup_trace_open
,
682 .close
= wakeup_trace_close
,
686 static struct tracer wakeup_rt_tracer __read_mostly
=
689 .init
= wakeup_rt_tracer_init
,
690 .reset
= wakeup_tracer_reset
,
691 .start
= wakeup_tracer_start
,
692 .stop
= wakeup_tracer_stop
,
693 .wait_pipe
= poll_wait_pipe
,
695 .print_header
= wakeup_print_header
,
696 .print_line
= wakeup_print_line
,
697 .flags
= &tracer_flags
,
698 .set_flag
= wakeup_set_flag
,
699 .flag_changed
= wakeup_flag_changed
,
700 #ifdef CONFIG_FTRACE_SELFTEST
701 .selftest
= trace_selftest_startup_wakeup
,
703 .open
= wakeup_trace_open
,
704 .close
= wakeup_trace_close
,
708 static struct tracer wakeup_dl_tracer __read_mostly
=
711 .init
= wakeup_dl_tracer_init
,
712 .reset
= wakeup_tracer_reset
,
713 .start
= wakeup_tracer_start
,
714 .stop
= wakeup_tracer_stop
,
715 .wait_pipe
= poll_wait_pipe
,
717 .print_header
= wakeup_print_header
,
718 .print_line
= wakeup_print_line
,
719 .flags
= &tracer_flags
,
720 .set_flag
= wakeup_set_flag
,
721 .flag_changed
= wakeup_flag_changed
,
722 #ifdef CONFIG_FTRACE_SELFTEST
723 .selftest
= trace_selftest_startup_wakeup
,
725 .open
= wakeup_trace_open
,
726 .close
= wakeup_trace_close
,
730 __init
static int init_wakeup_tracer(void)
734 ret
= register_tracer(&wakeup_tracer
);
738 ret
= register_tracer(&wakeup_rt_tracer
);
742 ret
= register_tracer(&wakeup_dl_tracer
);
748 core_initcall(init_wakeup_tracer
);