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 int wakeup_current_cpu;
28 static unsigned wakeup_prio = -1;
31 static raw_spinlock_t wakeup_lock =
32 (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
34 static void __wakeup_reset(struct trace_array *tr);
36 static int save_lat_flag;
38 #ifdef CONFIG_FUNCTION_TRACER
40 * irqsoff uses its own tracer function to keep the overhead down:
43 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
45 struct trace_array *tr = wakeup_trace;
46 struct trace_array_cpu *data;
53 if (likely(!wakeup_task))
57 resched = ftrace_preempt_disable();
59 cpu = raw_smp_processor_id();
60 if (cpu != wakeup_current_cpu)
64 disabled = atomic_inc_return(&data->disabled);
65 if (unlikely(disabled != 1))
68 local_irq_save(flags);
70 trace_function(tr, ip, parent_ip, flags, pc);
72 local_irq_restore(flags);
75 atomic_dec(&data->disabled);
77 ftrace_preempt_enable(resched);
80 static struct ftrace_ops trace_ops __read_mostly =
82 .func = wakeup_tracer_call,
84 #endif /* CONFIG_FUNCTION_TRACER */
87 * Should this new latency be reported/recorded?
89 static int report_latency(cycle_t delta)
92 if (delta < tracing_thresh)
95 if (delta <= tracing_max_latency)
101 static void probe_wakeup_migrate_task(struct task_struct *task, int cpu)
103 if (task != wakeup_task)
106 wakeup_current_cpu = cpu;
110 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
111 struct task_struct *next)
113 unsigned long latency = 0, t0 = 0, t1 = 0;
114 struct trace_array_cpu *data;
115 cycle_t T0, T1, delta;
121 tracing_record_cmdline(prev);
123 if (unlikely(!tracer_enabled))
127 * When we start a new trace, we set wakeup_task to NULL
128 * and then set tracer_enabled = 1. We want to make sure
129 * that another CPU does not see the tracer_enabled = 1
130 * and the wakeup_task with an older task, that might
131 * actually be the same as next.
135 if (next != wakeup_task)
138 pc = preempt_count();
140 /* disable local data, not wakeup_cpu data */
141 cpu = raw_smp_processor_id();
142 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
143 if (likely(disabled != 1))
146 local_irq_save(flags);
147 __raw_spin_lock(&wakeup_lock);
149 /* We could race with grabbing wakeup_lock */
150 if (unlikely(!tracer_enabled || next != wakeup_task))
153 /* The task we are waiting for is waking up */
154 data = wakeup_trace->data[wakeup_cpu];
156 trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
157 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
160 * usecs conversion is slow so we try to delay the conversion
161 * as long as possible:
163 T0 = data->preempt_timestamp;
164 T1 = ftrace_now(cpu);
167 if (!report_latency(delta))
170 latency = nsecs_to_usecs(delta);
172 tracing_max_latency = delta;
173 t0 = nsecs_to_usecs(T0);
174 t1 = nsecs_to_usecs(T1);
176 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
179 __wakeup_reset(wakeup_trace);
180 __raw_spin_unlock(&wakeup_lock);
181 local_irq_restore(flags);
183 atomic_dec(&wakeup_trace->data[cpu]->disabled);
186 static void __wakeup_reset(struct trace_array *tr)
192 put_task_struct(wakeup_task);
197 static void wakeup_reset(struct trace_array *tr)
201 tracing_reset_online_cpus(tr);
203 local_irq_save(flags);
204 __raw_spin_lock(&wakeup_lock);
206 __raw_spin_unlock(&wakeup_lock);
207 local_irq_restore(flags);
211 probe_wakeup(struct rq *rq, struct task_struct *p, int success)
213 struct trace_array_cpu *data;
214 int cpu = smp_processor_id();
219 if (likely(!tracer_enabled))
222 tracing_record_cmdline(p);
223 tracing_record_cmdline(current);
225 if ((wakeup_rt && !rt_task(p)) ||
226 p->prio >= wakeup_prio ||
227 p->prio >= current->prio)
230 pc = preempt_count();
231 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
232 if (unlikely(disabled != 1))
235 /* interrupts should be off from try_to_wake_up */
236 __raw_spin_lock(&wakeup_lock);
238 /* check for races. */
239 if (!tracer_enabled || p->prio >= wakeup_prio)
242 /* reset the trace */
243 __wakeup_reset(wakeup_trace);
245 wakeup_cpu = task_cpu(p);
246 wakeup_current_cpu = wakeup_cpu;
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;
296 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task);
298 pr_info("wakeup trace: Couldn't activate tracepoint"
299 " probe to kernel_sched_migrate_task\n");
306 * Don't let the tracer_enabled = 1 show up before
307 * the wakeup_task is reset. This may be overkill since
308 * wakeup_reset does a spin_unlock after setting the
309 * wakeup_task to NULL, but I want to be safe.
310 * This is a slow path anyway.
314 register_ftrace_function(&trace_ops);
316 if (tracing_is_enabled())
322 fail_deprobe_wake_new:
323 unregister_trace_sched_wakeup_new(probe_wakeup);
325 unregister_trace_sched_wakeup(probe_wakeup);
328 static void stop_wakeup_tracer(struct trace_array *tr)
331 unregister_ftrace_function(&trace_ops);
332 unregister_trace_sched_switch(probe_wakeup_sched_switch);
333 unregister_trace_sched_wakeup_new(probe_wakeup);
334 unregister_trace_sched_wakeup(probe_wakeup);
335 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task);
338 static int __wakeup_tracer_init(struct trace_array *tr)
340 save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
341 trace_flags |= TRACE_ITER_LATENCY_FMT;
343 tracing_max_latency = 0;
345 start_wakeup_tracer(tr);
349 static int wakeup_tracer_init(struct trace_array *tr)
352 return __wakeup_tracer_init(tr);
355 static int wakeup_rt_tracer_init(struct trace_array *tr)
358 return __wakeup_tracer_init(tr);
361 static void wakeup_tracer_reset(struct trace_array *tr)
363 stop_wakeup_tracer(tr);
364 /* make sure we put back any tasks we are tracing */
368 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
371 static void wakeup_tracer_start(struct trace_array *tr)
377 static void wakeup_tracer_stop(struct trace_array *tr)
382 static struct tracer wakeup_tracer __read_mostly =
385 .init = wakeup_tracer_init,
386 .reset = wakeup_tracer_reset,
387 .start = wakeup_tracer_start,
388 .stop = wakeup_tracer_stop,
390 #ifdef CONFIG_FTRACE_SELFTEST
391 .selftest = trace_selftest_startup_wakeup,
395 static struct tracer wakeup_rt_tracer __read_mostly =
398 .init = wakeup_rt_tracer_init,
399 .reset = wakeup_tracer_reset,
400 .start = wakeup_tracer_start,
401 .stop = wakeup_tracer_stop,
402 .wait_pipe = poll_wait_pipe,
404 #ifdef CONFIG_FTRACE_SELFTEST
405 .selftest = trace_selftest_startup_wakeup,
409 __init static int init_wakeup_tracer(void)
413 ret = register_tracer(&wakeup_tracer);
417 ret = register_tracer(&wakeup_rt_tracer);
423 device_initcall(init_wakeup_tracer);