xref: /linux/kernel/trace/trace_sched_wakeup.c (revision f8324e20f8289dffc646d64366332e05eaacab25)
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
102 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
103 {
104 	if (task != wakeup_task)
105 		return;
106 
107 	wakeup_current_cpu = cpu;
108 }
109 
110 static void notrace
111 probe_wakeup_sched_switch(void *ignore,
112 			  struct task_struct *prev, struct task_struct *next)
113 {
114 	struct trace_array_cpu *data;
115 	cycle_t T0, T1, delta;
116 	unsigned long flags;
117 	long disabled;
118 	int cpu;
119 	int pc;
120 
121 	tracing_record_cmdline(prev);
122 
123 	if (unlikely(!tracer_enabled))
124 		return;
125 
126 	/*
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.
132 	 */
133 	smp_rmb();
134 
135 	if (next != wakeup_task)
136 		return;
137 
138 	pc = preempt_count();
139 
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))
144 		goto out;
145 
146 	local_irq_save(flags);
147 	arch_spin_lock(&wakeup_lock);
148 
149 	/* We could race with grabbing wakeup_lock */
150 	if (unlikely(!tracer_enabled || next != wakeup_task))
151 		goto out_unlock;
152 
153 	/* The task we are waiting for is waking up */
154 	data = wakeup_trace->data[wakeup_cpu];
155 
156 	trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
157 	tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
158 
159 	T0 = data->preempt_timestamp;
160 	T1 = ftrace_now(cpu);
161 	delta = T1-T0;
162 
163 	if (!report_latency(delta))
164 		goto out_unlock;
165 
166 	if (likely(!is_tracing_stopped())) {
167 		tracing_max_latency = delta;
168 		update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
169 	}
170 
171 out_unlock:
172 	__wakeup_reset(wakeup_trace);
173 	arch_spin_unlock(&wakeup_lock);
174 	local_irq_restore(flags);
175 out:
176 	atomic_dec(&wakeup_trace->data[cpu]->disabled);
177 }
178 
179 static void __wakeup_reset(struct trace_array *tr)
180 {
181 	wakeup_cpu = -1;
182 	wakeup_prio = -1;
183 
184 	if (wakeup_task)
185 		put_task_struct(wakeup_task);
186 
187 	wakeup_task = NULL;
188 }
189 
190 static void wakeup_reset(struct trace_array *tr)
191 {
192 	unsigned long flags;
193 
194 	tracing_reset_online_cpus(tr);
195 
196 	local_irq_save(flags);
197 	arch_spin_lock(&wakeup_lock);
198 	__wakeup_reset(tr);
199 	arch_spin_unlock(&wakeup_lock);
200 	local_irq_restore(flags);
201 }
202 
203 static void
204 probe_wakeup(void *ignore, struct task_struct *p, int success)
205 {
206 	struct trace_array_cpu *data;
207 	int cpu = smp_processor_id();
208 	unsigned long flags;
209 	long disabled;
210 	int pc;
211 
212 	if (likely(!tracer_enabled))
213 		return;
214 
215 	tracing_record_cmdline(p);
216 	tracing_record_cmdline(current);
217 
218 	if ((wakeup_rt && !rt_task(p)) ||
219 			p->prio >= wakeup_prio ||
220 			p->prio >= current->prio)
221 		return;
222 
223 	pc = preempt_count();
224 	disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
225 	if (unlikely(disabled != 1))
226 		goto out;
227 
228 	/* interrupts should be off from try_to_wake_up */
229 	arch_spin_lock(&wakeup_lock);
230 
231 	/* check for races. */
232 	if (!tracer_enabled || p->prio >= wakeup_prio)
233 		goto out_locked;
234 
235 	/* reset the trace */
236 	__wakeup_reset(wakeup_trace);
237 
238 	wakeup_cpu = task_cpu(p);
239 	wakeup_current_cpu = wakeup_cpu;
240 	wakeup_prio = p->prio;
241 
242 	wakeup_task = p;
243 	get_task_struct(wakeup_task);
244 
245 	local_save_flags(flags);
246 
247 	data = wakeup_trace->data[wakeup_cpu];
248 	data->preempt_timestamp = ftrace_now(cpu);
249 	tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
250 
251 	/*
252 	 * We must be careful in using CALLER_ADDR2. But since wake_up
253 	 * is not called by an assembly function  (where as schedule is)
254 	 * it should be safe to use it here.
255 	 */
256 	trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
257 
258 out_locked:
259 	arch_spin_unlock(&wakeup_lock);
260 out:
261 	atomic_dec(&wakeup_trace->data[cpu]->disabled);
262 }
263 
264 static void start_wakeup_tracer(struct trace_array *tr)
265 {
266 	int ret;
267 
268 	ret = register_trace_sched_wakeup(probe_wakeup, NULL);
269 	if (ret) {
270 		pr_info("wakeup trace: Couldn't activate tracepoint"
271 			" probe to kernel_sched_wakeup\n");
272 		return;
273 	}
274 
275 	ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
276 	if (ret) {
277 		pr_info("wakeup trace: Couldn't activate tracepoint"
278 			" probe to kernel_sched_wakeup_new\n");
279 		goto fail_deprobe;
280 	}
281 
282 	ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
283 	if (ret) {
284 		pr_info("sched trace: Couldn't activate tracepoint"
285 			" probe to kernel_sched_switch\n");
286 		goto fail_deprobe_wake_new;
287 	}
288 
289 	ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
290 	if (ret) {
291 		pr_info("wakeup trace: Couldn't activate tracepoint"
292 			" probe to kernel_sched_migrate_task\n");
293 		return;
294 	}
295 
296 	wakeup_reset(tr);
297 
298 	/*
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.
304 	 */
305 	smp_wmb();
306 
307 	register_ftrace_function(&trace_ops);
308 
309 	if (tracing_is_enabled())
310 		tracer_enabled = 1;
311 	else
312 		tracer_enabled = 0;
313 
314 	return;
315 fail_deprobe_wake_new:
316 	unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
317 fail_deprobe:
318 	unregister_trace_sched_wakeup(probe_wakeup, NULL);
319 }
320 
321 static void stop_wakeup_tracer(struct trace_array *tr)
322 {
323 	tracer_enabled = 0;
324 	unregister_ftrace_function(&trace_ops);
325 	unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
326 	unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
327 	unregister_trace_sched_wakeup(probe_wakeup, NULL);
328 	unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
329 }
330 
331 static int __wakeup_tracer_init(struct trace_array *tr)
332 {
333 	save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
334 	trace_flags |= TRACE_ITER_LATENCY_FMT;
335 
336 	tracing_max_latency = 0;
337 	wakeup_trace = tr;
338 	start_wakeup_tracer(tr);
339 	return 0;
340 }
341 
342 static int wakeup_tracer_init(struct trace_array *tr)
343 {
344 	wakeup_rt = 0;
345 	return __wakeup_tracer_init(tr);
346 }
347 
348 static int wakeup_rt_tracer_init(struct trace_array *tr)
349 {
350 	wakeup_rt = 1;
351 	return __wakeup_tracer_init(tr);
352 }
353 
354 static void wakeup_tracer_reset(struct trace_array *tr)
355 {
356 	stop_wakeup_tracer(tr);
357 	/* make sure we put back any tasks we are tracing */
358 	wakeup_reset(tr);
359 
360 	if (!save_lat_flag)
361 		trace_flags &= ~TRACE_ITER_LATENCY_FMT;
362 }
363 
364 static void wakeup_tracer_start(struct trace_array *tr)
365 {
366 	wakeup_reset(tr);
367 	tracer_enabled = 1;
368 }
369 
370 static void wakeup_tracer_stop(struct trace_array *tr)
371 {
372 	tracer_enabled = 0;
373 }
374 
375 static struct tracer wakeup_tracer __read_mostly =
376 {
377 	.name		= "wakeup",
378 	.init		= wakeup_tracer_init,
379 	.reset		= wakeup_tracer_reset,
380 	.start		= wakeup_tracer_start,
381 	.stop		= wakeup_tracer_stop,
382 	.print_max	= 1,
383 #ifdef CONFIG_FTRACE_SELFTEST
384 	.selftest    = trace_selftest_startup_wakeup,
385 #endif
386 };
387 
388 static struct tracer wakeup_rt_tracer __read_mostly =
389 {
390 	.name		= "wakeup_rt",
391 	.init		= wakeup_rt_tracer_init,
392 	.reset		= wakeup_tracer_reset,
393 	.start		= wakeup_tracer_start,
394 	.stop		= wakeup_tracer_stop,
395 	.wait_pipe	= poll_wait_pipe,
396 	.print_max	= 1,
397 #ifdef CONFIG_FTRACE_SELFTEST
398 	.selftest    = trace_selftest_startup_wakeup,
399 #endif
400 };
401 
402 __init static int init_wakeup_tracer(void)
403 {
404 	int ret;
405 
406 	ret = register_tracer(&wakeup_tracer);
407 	if (ret)
408 		return ret;
409 
410 	ret = register_tracer(&wakeup_rt_tracer);
411 	if (ret)
412 		return ret;
413 
414 	return 0;
415 }
416 device_initcall(init_wakeup_tracer);
417