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 William Lee Irwin III
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 <trace/sched.h>
22 static struct trace_array
*wakeup_trace
;
23 static int __read_mostly tracer_enabled
;
25 static struct task_struct
*wakeup_task
;
26 static int wakeup_cpu
;
27 static unsigned wakeup_prio
= -1;
29 static raw_spinlock_t wakeup_lock
=
30 (raw_spinlock_t
)__RAW_SPIN_LOCK_UNLOCKED
;
32 static void __wakeup_reset(struct trace_array
*tr
);
34 #ifdef CONFIG_FUNCTION_TRACER
36 * irqsoff uses its own tracer function to keep the overhead down:
39 wakeup_tracer_call(unsigned long ip
, unsigned long parent_ip
)
41 struct trace_array
*tr
= wakeup_trace
;
42 struct trace_array_cpu
*data
;
49 if (likely(!wakeup_task
))
53 resched
= ftrace_preempt_disable();
55 cpu
= raw_smp_processor_id();
57 disabled
= atomic_inc_return(&data
->disabled
);
58 if (unlikely(disabled
!= 1))
61 local_irq_save(flags
);
62 __raw_spin_lock(&wakeup_lock
);
64 if (unlikely(!wakeup_task
))
68 * The task can't disappear because it needs to
69 * wake up first, and we have the wakeup_lock.
71 if (task_cpu(wakeup_task
) != cpu
)
74 trace_function(tr
, data
, ip
, parent_ip
, flags
, pc
);
77 __raw_spin_unlock(&wakeup_lock
);
78 local_irq_restore(flags
);
81 atomic_dec(&data
->disabled
);
83 ftrace_preempt_enable(resched
);
86 static struct ftrace_ops trace_ops __read_mostly
=
88 .func
= wakeup_tracer_call
,
90 #endif /* CONFIG_FUNCTION_TRACER */
93 * Should this new latency be reported/recorded?
95 static int report_latency(cycle_t delta
)
98 if (delta
< tracing_thresh
)
101 if (delta
<= tracing_max_latency
)
108 probe_wakeup_sched_switch(struct rq
*rq
, struct task_struct
*prev
,
109 struct task_struct
*next
)
111 unsigned long latency
= 0, t0
= 0, t1
= 0;
112 struct trace_array_cpu
*data
;
113 cycle_t T0
, T1
, delta
;
119 tracing_record_cmdline(prev
);
121 if (unlikely(!tracer_enabled
))
125 * When we start a new trace, we set wakeup_task to NULL
126 * and then set tracer_enabled = 1. We want to make sure
127 * that another CPU does not see the tracer_enabled = 1
128 * and the wakeup_task with an older task, that might
129 * actually be the same as next.
133 if (next
!= wakeup_task
)
136 pc
= preempt_count();
138 /* The task we are waiting for is waking up */
139 data
= wakeup_trace
->data
[wakeup_cpu
];
141 /* disable local data, not wakeup_cpu data */
142 cpu
= raw_smp_processor_id();
143 disabled
= atomic_inc_return(&wakeup_trace
->data
[cpu
]->disabled
);
144 if (likely(disabled
!= 1))
147 local_irq_save(flags
);
148 __raw_spin_lock(&wakeup_lock
);
150 /* We could race with grabbing wakeup_lock */
151 if (unlikely(!tracer_enabled
|| next
!= wakeup_task
))
154 trace_function(wakeup_trace
, data
, CALLER_ADDR1
, CALLER_ADDR2
, flags
, pc
);
157 * usecs conversion is slow so we try to delay the conversion
158 * as long as possible:
160 T0
= data
->preempt_timestamp
;
161 T1
= ftrace_now(cpu
);
164 if (!report_latency(delta
))
167 latency
= nsecs_to_usecs(delta
);
169 tracing_max_latency
= delta
;
170 t0
= nsecs_to_usecs(T0
);
171 t1
= nsecs_to_usecs(T1
);
173 update_max_tr(wakeup_trace
, wakeup_task
, wakeup_cpu
);
176 __wakeup_reset(wakeup_trace
);
177 __raw_spin_unlock(&wakeup_lock
);
178 local_irq_restore(flags
);
180 atomic_dec(&wakeup_trace
->data
[cpu
]->disabled
);
183 static void __wakeup_reset(struct trace_array
*tr
)
185 struct trace_array_cpu
*data
;
188 for_each_possible_cpu(cpu
) {
189 data
= tr
->data
[cpu
];
190 tracing_reset(tr
, cpu
);
197 put_task_struct(wakeup_task
);
202 static void wakeup_reset(struct trace_array
*tr
)
206 local_irq_save(flags
);
207 __raw_spin_lock(&wakeup_lock
);
209 __raw_spin_unlock(&wakeup_lock
);
210 local_irq_restore(flags
);
214 probe_wakeup(struct rq
*rq
, struct task_struct
*p
, int success
)
216 int cpu
= smp_processor_id();
221 if (likely(!tracer_enabled
))
224 tracing_record_cmdline(p
);
225 tracing_record_cmdline(current
);
227 if (likely(!rt_task(p
)) ||
228 p
->prio
>= wakeup_prio
||
229 p
->prio
>= current
->prio
)
232 pc
= preempt_count();
233 disabled
= atomic_inc_return(&wakeup_trace
->data
[cpu
]->disabled
);
234 if (unlikely(disabled
!= 1))
237 /* interrupts should be off from try_to_wake_up */
238 __raw_spin_lock(&wakeup_lock
);
240 /* check for races. */
241 if (!tracer_enabled
|| p
->prio
>= wakeup_prio
)
244 /* reset the trace */
245 __wakeup_reset(wakeup_trace
);
247 wakeup_cpu
= task_cpu(p
);
248 wakeup_prio
= p
->prio
;
251 get_task_struct(wakeup_task
);
253 local_save_flags(flags
);
255 wakeup_trace
->data
[wakeup_cpu
]->preempt_timestamp
= ftrace_now(cpu
);
256 trace_function(wakeup_trace
, wakeup_trace
->data
[wakeup_cpu
],
257 CALLER_ADDR1
, CALLER_ADDR2
, flags
, pc
);
260 __raw_spin_unlock(&wakeup_lock
);
262 atomic_dec(&wakeup_trace
->data
[cpu
]->disabled
);
266 * save_tracer_enabled is used to save the state of the tracer_enabled
267 * variable when we disable it when we open a trace output file.
269 static int save_tracer_enabled
;
271 static void start_wakeup_tracer(struct trace_array
*tr
)
275 ret
= register_trace_sched_wakeup(probe_wakeup
);
277 pr_info("wakeup trace: Couldn't activate tracepoint"
278 " probe to kernel_sched_wakeup\n");
282 ret
= register_trace_sched_wakeup_new(probe_wakeup
);
284 pr_info("wakeup trace: Couldn't activate tracepoint"
285 " probe to kernel_sched_wakeup_new\n");
289 ret
= register_trace_sched_switch(probe_wakeup_sched_switch
);
291 pr_info("sched trace: Couldn't activate tracepoint"
292 " probe to kernel_sched_schedule\n");
293 goto fail_deprobe_wake_new
;
299 * Don't let the tracer_enabled = 1 show up before
300 * the wakeup_task is reset. This may be overkill since
301 * wakeup_reset does a spin_unlock after setting the
302 * wakeup_task to NULL, but I want to be safe.
303 * This is a slow path anyway.
307 register_ftrace_function(&trace_ops
);
309 if (tracing_is_enabled()) {
311 save_tracer_enabled
= 1;
314 save_tracer_enabled
= 0;
318 fail_deprobe_wake_new
:
319 unregister_trace_sched_wakeup_new(probe_wakeup
);
321 unregister_trace_sched_wakeup(probe_wakeup
);
324 static void stop_wakeup_tracer(struct trace_array
*tr
)
327 save_tracer_enabled
= 0;
328 unregister_ftrace_function(&trace_ops
);
329 unregister_trace_sched_switch(probe_wakeup_sched_switch
);
330 unregister_trace_sched_wakeup_new(probe_wakeup
);
331 unregister_trace_sched_wakeup(probe_wakeup
);
334 static int wakeup_tracer_init(struct trace_array
*tr
)
336 tracing_max_latency
= 0;
338 start_wakeup_tracer(tr
);
342 static void wakeup_tracer_reset(struct trace_array
*tr
)
344 stop_wakeup_tracer(tr
);
345 /* make sure we put back any tasks we are tracing */
349 static void wakeup_tracer_start(struct trace_array
*tr
)
353 save_tracer_enabled
= 1;
356 static void wakeup_tracer_stop(struct trace_array
*tr
)
359 save_tracer_enabled
= 0;
362 static void wakeup_tracer_open(struct trace_iterator
*iter
)
364 /* stop the trace while dumping */
368 static void wakeup_tracer_close(struct trace_iterator
*iter
)
370 /* forget about any processes we were recording */
371 if (save_tracer_enabled
) {
372 wakeup_reset(iter
->tr
);
377 static struct tracer wakeup_tracer __read_mostly
=
380 .init
= wakeup_tracer_init
,
381 .reset
= wakeup_tracer_reset
,
382 .start
= wakeup_tracer_start
,
383 .stop
= wakeup_tracer_stop
,
384 .open
= wakeup_tracer_open
,
385 .close
= wakeup_tracer_close
,
387 #ifdef CONFIG_FTRACE_SELFTEST
388 .selftest
= trace_selftest_startup_wakeup
,
392 __init
static int init_wakeup_tracer(void)
396 ret
= register_tracer(&wakeup_tracer
);
402 device_initcall(init_wakeup_tracer
);