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/events/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;
30 static raw_spinlock_t wakeup_lock =
31 (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
33 static void __wakeup_reset(struct trace_array *tr);
35 static int save_lat_flag;
37 #ifdef CONFIG_FUNCTION_TRACER
39 * irqsoff uses its own tracer function to keep the overhead down:
42 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
44 struct trace_array *tr = wakeup_trace;
45 struct trace_array_cpu *data;
52 if (likely(!wakeup_task))
56 resched = ftrace_preempt_disable();
58 cpu = raw_smp_processor_id();
60 disabled = atomic_inc_return(&data->disabled);
61 if (unlikely(disabled != 1))
64 local_irq_save(flags);
65 __raw_spin_lock(&wakeup_lock);
67 if (unlikely(!wakeup_task))
71 * The task can't disappear because it needs to
72 * wake up first, and we have the wakeup_lock.
74 if (task_cpu(wakeup_task) != cpu)
77 trace_function(tr, ip, parent_ip, flags, pc);
80 __raw_spin_unlock(&wakeup_lock);
81 local_irq_restore(flags);
84 atomic_dec(&data->disabled);
86 ftrace_preempt_enable(resched);
89 static struct ftrace_ops trace_ops __read_mostly =
91 .func = wakeup_tracer_call,
93 #endif /* CONFIG_FUNCTION_TRACER */
96 * Should this new latency be reported/recorded?
98 static int report_latency(cycle_t delta)
100 if (tracing_thresh) {
101 if (delta < tracing_thresh)
104 if (delta <= tracing_max_latency)
111 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
112 struct task_struct *next)
114 unsigned long latency = 0, t0 = 0, t1 = 0;
115 struct trace_array_cpu *data;
116 cycle_t T0, T1, delta;
122 tracing_record_cmdline(prev);
124 if (unlikely(!tracer_enabled))
128 * When we start a new trace, we set wakeup_task to NULL
129 * and then set tracer_enabled = 1. We want to make sure
130 * that another CPU does not see the tracer_enabled = 1
131 * and the wakeup_task with an older task, that might
132 * actually be the same as next.
136 if (next != wakeup_task)
139 pc = preempt_count();
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 /* The task we are waiting for is waking up */
155 data = wakeup_trace->data[wakeup_cpu];
157 trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
158 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
161 * usecs conversion is slow so we try to delay the conversion
162 * as long as possible:
164 T0 = data->preempt_timestamp;
165 T1 = ftrace_now(cpu);
168 if (!report_latency(delta))
171 latency = nsecs_to_usecs(delta);
173 tracing_max_latency = delta;
174 t0 = nsecs_to_usecs(T0);
175 t1 = nsecs_to_usecs(T1);
177 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
180 __wakeup_reset(wakeup_trace);
181 __raw_spin_unlock(&wakeup_lock);
182 local_irq_restore(flags);
184 atomic_dec(&wakeup_trace->data[cpu]->disabled);
187 static void __wakeup_reset(struct trace_array *tr)
193 put_task_struct(wakeup_task);
198 static void wakeup_reset(struct trace_array *tr)
202 tracing_reset_online_cpus(tr);
204 local_irq_save(flags);
205 __raw_spin_lock(&wakeup_lock);
207 __raw_spin_unlock(&wakeup_lock);
208 local_irq_restore(flags);
212 probe_wakeup(struct rq *rq, struct task_struct *p, int success)
214 struct trace_array_cpu *data;
215 int cpu = smp_processor_id();
220 if (likely(!tracer_enabled))
223 tracing_record_cmdline(p);
224 tracing_record_cmdline(current);
226 if ((wakeup_rt && !rt_task(p)) ||
227 p->prio >= wakeup_prio ||
228 p->prio >= current->prio)
231 pc = preempt_count();
232 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
233 if (unlikely(disabled != 1))
236 /* interrupts should be off from try_to_wake_up */
237 __raw_spin_lock(&wakeup_lock);
239 /* check for races. */
240 if (!tracer_enabled || p->prio >= wakeup_prio)
243 /* reset the trace */
244 __wakeup_reset(wakeup_trace);
246 wakeup_cpu = task_cpu(p);
247 wakeup_prio = p->prio;
250 get_task_struct(wakeup_task);
252 local_save_flags(flags);
254 data = wakeup_trace->data[wakeup_cpu];
255 data->preempt_timestamp = ftrace_now(cpu);
256 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
259 * We must be careful in using CALLER_ADDR2. But since wake_up
260 * is not called by an assembly function (where as schedule is)
261 * it should be safe to use it here.
263 trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
266 __raw_spin_unlock(&wakeup_lock);
268 atomic_dec(&wakeup_trace->data[cpu]->disabled);
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_switch\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())
315 fail_deprobe_wake_new:
316 unregister_trace_sched_wakeup_new(probe_wakeup);
318 unregister_trace_sched_wakeup(probe_wakeup);
321 static void stop_wakeup_tracer(struct trace_array *tr)
324 unregister_ftrace_function(&trace_ops);
325 unregister_trace_sched_switch(probe_wakeup_sched_switch);
326 unregister_trace_sched_wakeup_new(probe_wakeup);
327 unregister_trace_sched_wakeup(probe_wakeup);
330 static int __wakeup_tracer_init(struct trace_array *tr)
332 save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
333 trace_flags |= TRACE_ITER_LATENCY_FMT;
335 tracing_max_latency = 0;
337 start_wakeup_tracer(tr);
341 static int wakeup_tracer_init(struct trace_array *tr)
344 return __wakeup_tracer_init(tr);
347 static int wakeup_rt_tracer_init(struct trace_array *tr)
350 return __wakeup_tracer_init(tr);
353 static void wakeup_tracer_reset(struct trace_array *tr)
355 stop_wakeup_tracer(tr);
356 /* make sure we put back any tasks we are tracing */
360 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
363 static void wakeup_tracer_start(struct trace_array *tr)
369 static void wakeup_tracer_stop(struct trace_array *tr)
374 static struct tracer wakeup_tracer __read_mostly =
377 .init = wakeup_tracer_init,
378 .reset = wakeup_tracer_reset,
379 .start = wakeup_tracer_start,
380 .stop = wakeup_tracer_stop,
382 #ifdef CONFIG_FTRACE_SELFTEST
383 .selftest = trace_selftest_startup_wakeup,
387 static struct tracer wakeup_rt_tracer __read_mostly =
390 .init = wakeup_rt_tracer_init,
391 .reset = wakeup_tracer_reset,
392 .start = wakeup_tracer_start,
393 .stop = wakeup_tracer_stop,
394 .wait_pipe = poll_wait_pipe,
396 #ifdef CONFIG_FTRACE_SELFTEST
397 .selftest = trace_selftest_startup_wakeup,
401 __init static int init_wakeup_tracer(void)
405 ret = register_tracer(&wakeup_tracer);
409 ret = register_tracer(&wakeup_rt_tracer);
415 device_initcall(init_wakeup_tracer);