8052446ceeaa9333ae575edf6f762b665c76ac09
[linux-2.6.git] / kernel / trace / trace_sched_wakeup.c
1 /*
2  * trace task wakeup timings
3  *
4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6  *
7  * Based on code from the latency_tracer, that is:
8  *
9  *  Copyright (C) 2004-2006 Ingo Molnar
10  *  Copyright (C) 2004 William Lee Irwin III
11  */
12 #include <linux/module.h>
13 #include <linux/fs.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>
19
20 #include "trace.h"
21
22 static struct trace_array       *wakeup_trace;
23 static int __read_mostly        tracer_enabled;
24
25 static struct task_struct       *wakeup_task;
26 static int                      wakeup_cpu;
27 static int                      wakeup_current_cpu;
28 static unsigned                 wakeup_prio = -1;
29 static int                      wakeup_rt;
30
31 static arch_spinlock_t wakeup_lock =
32         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
33
34 static void __wakeup_reset(struct trace_array *tr);
35
36 static int save_lat_flag;
37
38 #ifdef CONFIG_FUNCTION_TRACER
39 /*
40  * irqsoff uses its own tracer function to keep the overhead down:
41  */
42 static void
43 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
44 {
45         struct trace_array *tr = wakeup_trace;
46         struct trace_array_cpu *data;
47         unsigned long flags;
48         long disabled;
49         int resched;
50         int cpu;
51         int pc;
52
53         if (likely(!wakeup_task))
54                 return;
55
56         pc = preempt_count();
57         resched = ftrace_preempt_disable();
58
59         cpu = raw_smp_processor_id();
60         if (cpu != wakeup_current_cpu)
61                 goto out_enable;
62
63         data = tr->data[cpu];
64         disabled = atomic_inc_return(&data->disabled);
65         if (unlikely(disabled != 1))
66                 goto out;
67
68         local_irq_save(flags);
69
70         trace_function(tr, ip, parent_ip, flags, pc);
71
72         local_irq_restore(flags);
73
74  out:
75         atomic_dec(&data->disabled);
76  out_enable:
77         ftrace_preempt_enable(resched);
78 }
79
80 static struct ftrace_ops trace_ops __read_mostly =
81 {
82         .func = wakeup_tracer_call,
83 };
84 #endif /* CONFIG_FUNCTION_TRACER */
85
86 /*
87  * Should this new latency be reported/recorded?
88  */
89 static int report_latency(cycle_t delta)
90 {
91         if (tracing_thresh) {
92                 if (delta < tracing_thresh)
93                         return 0;
94         } else {
95                 if (delta <= tracing_max_latency)
96                         return 0;
97         }
98         return 1;
99 }
100
101 static void probe_wakeup_migrate_task(struct task_struct *task, int cpu)
102 {
103         if (task != wakeup_task)
104                 return;
105
106         wakeup_current_cpu = cpu;
107 }
108
109 static void notrace
110 probe_wakeup_sched_switch(struct task_struct *prev, struct task_struct *next)
111 {
112         struct trace_array_cpu *data;
113         cycle_t T0, T1, delta;
114         unsigned long flags;
115         long disabled;
116         int cpu;
117         int pc;
118
119         tracing_record_cmdline(prev);
120
121         if (unlikely(!tracer_enabled))
122                 return;
123
124         /*
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.
130          */
131         smp_rmb();
132
133         if (next != wakeup_task)
134                 return;
135
136         pc = preempt_count();
137
138         /* disable local data, not wakeup_cpu data */
139         cpu = raw_smp_processor_id();
140         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
141         if (likely(disabled != 1))
142                 goto out;
143
144         local_irq_save(flags);
145         arch_spin_lock(&wakeup_lock);
146
147         /* We could race with grabbing wakeup_lock */
148         if (unlikely(!tracer_enabled || next != wakeup_task))
149                 goto out_unlock;
150
151         /* The task we are waiting for is waking up */
152         data = wakeup_trace->data[wakeup_cpu];
153
154         trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
155         tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
156
157         T0 = data->preempt_timestamp;
158         T1 = ftrace_now(cpu);
159         delta = T1-T0;
160
161         if (!report_latency(delta))
162                 goto out_unlock;
163
164         if (likely(!is_tracing_stopped())) {
165                 tracing_max_latency = delta;
166                 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
167         }
168
169 out_unlock:
170         __wakeup_reset(wakeup_trace);
171         arch_spin_unlock(&wakeup_lock);
172         local_irq_restore(flags);
173 out:
174         atomic_dec(&wakeup_trace->data[cpu]->disabled);
175 }
176
177 static void __wakeup_reset(struct trace_array *tr)
178 {
179         wakeup_cpu = -1;
180         wakeup_prio = -1;
181
182         if (wakeup_task)
183                 put_task_struct(wakeup_task);
184
185         wakeup_task = NULL;
186 }
187
188 static void wakeup_reset(struct trace_array *tr)
189 {
190         unsigned long flags;
191
192         tracing_reset_online_cpus(tr);
193
194         local_irq_save(flags);
195         arch_spin_lock(&wakeup_lock);
196         __wakeup_reset(tr);
197         arch_spin_unlock(&wakeup_lock);
198         local_irq_restore(flags);
199 }
200
201 static void
202 probe_wakeup(struct task_struct *p, int success)
203 {
204         struct trace_array_cpu *data;
205         int cpu = smp_processor_id();
206         unsigned long flags;
207         long disabled;
208         int pc;
209
210         if (likely(!tracer_enabled))
211                 return;
212
213         tracing_record_cmdline(p);
214         tracing_record_cmdline(current);
215
216         if ((wakeup_rt && !rt_task(p)) ||
217                         p->prio >= wakeup_prio ||
218                         p->prio >= current->prio)
219                 return;
220
221         pc = preempt_count();
222         disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
223         if (unlikely(disabled != 1))
224                 goto out;
225
226         /* interrupts should be off from try_to_wake_up */
227         arch_spin_lock(&wakeup_lock);
228
229         /* check for races. */
230         if (!tracer_enabled || p->prio >= wakeup_prio)
231                 goto out_locked;
232
233         /* reset the trace */
234         __wakeup_reset(wakeup_trace);
235
236         wakeup_cpu = task_cpu(p);
237         wakeup_current_cpu = wakeup_cpu;
238         wakeup_prio = p->prio;
239
240         wakeup_task = p;
241         get_task_struct(wakeup_task);
242
243         local_save_flags(flags);
244
245         data = wakeup_trace->data[wakeup_cpu];
246         data->preempt_timestamp = ftrace_now(cpu);
247         tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
248
249         /*
250          * We must be careful in using CALLER_ADDR2. But since wake_up
251          * is not called by an assembly function  (where as schedule is)
252          * it should be safe to use it here.
253          */
254         trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
255
256 out_locked:
257         arch_spin_unlock(&wakeup_lock);
258 out:
259         atomic_dec(&wakeup_trace->data[cpu]->disabled);
260 }
261
262 static void start_wakeup_tracer(struct trace_array *tr)
263 {
264         int ret;
265
266         ret = register_trace_sched_wakeup(probe_wakeup);
267         if (ret) {
268                 pr_info("wakeup trace: Couldn't activate tracepoint"
269                         " probe to kernel_sched_wakeup\n");
270                 return;
271         }
272
273         ret = register_trace_sched_wakeup_new(probe_wakeup);
274         if (ret) {
275                 pr_info("wakeup trace: Couldn't activate tracepoint"
276                         " probe to kernel_sched_wakeup_new\n");
277                 goto fail_deprobe;
278         }
279
280         ret = register_trace_sched_switch(probe_wakeup_sched_switch);
281         if (ret) {
282                 pr_info("sched trace: Couldn't activate tracepoint"
283                         " probe to kernel_sched_switch\n");
284                 goto fail_deprobe_wake_new;
285         }
286
287         ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task);
288         if (ret) {
289                 pr_info("wakeup trace: Couldn't activate tracepoint"
290                         " probe to kernel_sched_migrate_task\n");
291                 return;
292         }
293
294         wakeup_reset(tr);
295
296         /*
297          * Don't let the tracer_enabled = 1 show up before
298          * the wakeup_task is reset. This may be overkill since
299          * wakeup_reset does a spin_unlock after setting the
300          * wakeup_task to NULL, but I want to be safe.
301          * This is a slow path anyway.
302          */
303         smp_wmb();
304
305         register_ftrace_function(&trace_ops);
306
307         if (tracing_is_enabled())
308                 tracer_enabled = 1;
309         else
310                 tracer_enabled = 0;
311
312         return;
313 fail_deprobe_wake_new:
314         unregister_trace_sched_wakeup_new(probe_wakeup);
315 fail_deprobe:
316         unregister_trace_sched_wakeup(probe_wakeup);
317 }
318
319 static void stop_wakeup_tracer(struct trace_array *tr)
320 {
321         tracer_enabled = 0;
322         unregister_ftrace_function(&trace_ops);
323         unregister_trace_sched_switch(probe_wakeup_sched_switch);
324         unregister_trace_sched_wakeup_new(probe_wakeup);
325         unregister_trace_sched_wakeup(probe_wakeup);
326         unregister_trace_sched_migrate_task(probe_wakeup_migrate_task);
327 }
328
329 static int __wakeup_tracer_init(struct trace_array *tr)
330 {
331         save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
332         trace_flags |= TRACE_ITER_LATENCY_FMT;
333
334         tracing_max_latency = 0;
335         wakeup_trace = tr;
336         start_wakeup_tracer(tr);
337         return 0;
338 }
339
340 static int wakeup_tracer_init(struct trace_array *tr)
341 {
342         wakeup_rt = 0;
343         return __wakeup_tracer_init(tr);
344 }
345
346 static int wakeup_rt_tracer_init(struct trace_array *tr)
347 {
348         wakeup_rt = 1;
349         return __wakeup_tracer_init(tr);
350 }
351
352 static void wakeup_tracer_reset(struct trace_array *tr)
353 {
354         stop_wakeup_tracer(tr);
355         /* make sure we put back any tasks we are tracing */
356         wakeup_reset(tr);
357
358         if (!save_lat_flag)
359                 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
360 }
361
362 static void wakeup_tracer_start(struct trace_array *tr)
363 {
364         wakeup_reset(tr);
365         tracer_enabled = 1;
366 }
367
368 static void wakeup_tracer_stop(struct trace_array *tr)
369 {
370         tracer_enabled = 0;
371 }
372
373 static struct tracer wakeup_tracer __read_mostly =
374 {
375         .name           = "wakeup",
376         .init           = wakeup_tracer_init,
377         .reset          = wakeup_tracer_reset,
378         .start          = wakeup_tracer_start,
379         .stop           = wakeup_tracer_stop,
380         .print_max      = 1,
381 #ifdef CONFIG_FTRACE_SELFTEST
382         .selftest    = trace_selftest_startup_wakeup,
383 #endif
384 };
385
386 static struct tracer wakeup_rt_tracer __read_mostly =
387 {
388         .name           = "wakeup_rt",
389         .init           = wakeup_rt_tracer_init,
390         .reset          = wakeup_tracer_reset,
391         .start          = wakeup_tracer_start,
392         .stop           = wakeup_tracer_stop,
393         .wait_pipe      = poll_wait_pipe,
394         .print_max      = 1,
395 #ifdef CONFIG_FTRACE_SELFTEST
396         .selftest    = trace_selftest_startup_wakeup,
397 #endif
398 };
399
400 __init static int init_wakeup_tracer(void)
401 {
402         int ret;
403
404         ret = register_tracer(&wakeup_tracer);
405         if (ret)
406                 return ret;
407
408         ret = register_tracer(&wakeup_rt_tracer);
409         if (ret)
410                 return ret;
411
412         return 0;
413 }
414 device_initcall(init_wakeup_tracer);