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 unsigned			wakeup_prio = -1;
28 static int			wakeup_rt;
29 
30 static raw_spinlock_t wakeup_lock =
31 	(raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
32 
33 static void __wakeup_reset(struct trace_array *tr);
34 
35 static int save_lat_flag;
36 
37 #ifdef CONFIG_FUNCTION_TRACER
38 /*
39  * irqsoff uses its own tracer function to keep the overhead down:
40  */
41 static void
42 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
43 {
44 	struct trace_array *tr = wakeup_trace;
45 	struct trace_array_cpu *data;
46 	unsigned long flags;
47 	long disabled;
48 	int resched;
49 	int cpu;
50 	int pc;
51 
52 	if (likely(!wakeup_task))
53 		return;
54 
55 	pc = preempt_count();
56 	resched = ftrace_preempt_disable();
57 
58 	cpu = raw_smp_processor_id();
59 	data = tr->data[cpu];
60 	disabled = atomic_inc_return(&data->disabled);
61 	if (unlikely(disabled != 1))
62 		goto out;
63 
64 	local_irq_save(flags);
65 	__raw_spin_lock(&wakeup_lock);
66 
67 	if (unlikely(!wakeup_task))
68 		goto unlock;
69 
70 	/*
71 	 * The task can't disappear because it needs to
72 	 * wake up first, and we have the wakeup_lock.
73 	 */
74 	if (task_cpu(wakeup_task) != cpu)
75 		goto unlock;
76 
77 	trace_function(tr, ip, parent_ip, flags, pc);
78 
79  unlock:
80 	__raw_spin_unlock(&wakeup_lock);
81 	local_irq_restore(flags);
82 
83  out:
84 	atomic_dec(&data->disabled);
85 
86 	ftrace_preempt_enable(resched);
87 }
88 
89 static struct ftrace_ops trace_ops __read_mostly =
90 {
91 	.func = wakeup_tracer_call,
92 };
93 #endif /* CONFIG_FUNCTION_TRACER */
94 
95 /*
96  * Should this new latency be reported/recorded?
97  */
98 static int report_latency(cycle_t delta)
99 {
100 	if (tracing_thresh) {
101 		if (delta < tracing_thresh)
102 			return 0;
103 	} else {
104 		if (delta <= tracing_max_latency)
105 			return 0;
106 	}
107 	return 1;
108 }
109 
110 static void notrace
111 probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev,
112 	struct task_struct *next)
113 {
114 	unsigned long latency = 0, t0 = 0, t1 = 0;
115 	struct trace_array_cpu *data;
116 	cycle_t T0, T1, delta;
117 	unsigned long flags;
118 	long disabled;
119 	int cpu;
120 	int pc;
121 
122 	tracing_record_cmdline(prev);
123 
124 	if (unlikely(!tracer_enabled))
125 		return;
126 
127 	/*
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.
133 	 */
134 	smp_rmb();
135 
136 	if (next != wakeup_task)
137 		return;
138 
139 	pc = preempt_count();
140 
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))
145 		goto out;
146 
147 	local_irq_save(flags);
148 	__raw_spin_lock(&wakeup_lock);
149 
150 	/* We could race with grabbing wakeup_lock */
151 	if (unlikely(!tracer_enabled || next != wakeup_task))
152 		goto out_unlock;
153 
154 	/* The task we are waiting for is waking up */
155 	data = wakeup_trace->data[wakeup_cpu];
156 
157 	trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
158 	tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
159 
160 	/*
161 	 * usecs conversion is slow so we try to delay the conversion
162 	 * as long as possible:
163 	 */
164 	T0 = data->preempt_timestamp;
165 	T1 = ftrace_now(cpu);
166 	delta = T1-T0;
167 
168 	if (!report_latency(delta))
169 		goto out_unlock;
170 
171 	latency = nsecs_to_usecs(delta);
172 
173 	tracing_max_latency = delta;
174 	t0 = nsecs_to_usecs(T0);
175 	t1 = nsecs_to_usecs(T1);
176 
177 	update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
178 
179 out_unlock:
180 	__wakeup_reset(wakeup_trace);
181 	__raw_spin_unlock(&wakeup_lock);
182 	local_irq_restore(flags);
183 out:
184 	atomic_dec(&wakeup_trace->data[cpu]->disabled);
185 }
186 
187 static void __wakeup_reset(struct trace_array *tr)
188 {
189 	int cpu;
190 
191 	for_each_possible_cpu(cpu)
192 		tracing_reset(tr, cpu);
193 
194 	wakeup_cpu = -1;
195 	wakeup_prio = -1;
196 
197 	if (wakeup_task)
198 		put_task_struct(wakeup_task);
199 
200 	wakeup_task = NULL;
201 }
202 
203 static void wakeup_reset(struct trace_array *tr)
204 {
205 	unsigned long flags;
206 
207 	local_irq_save(flags);
208 	__raw_spin_lock(&wakeup_lock);
209 	__wakeup_reset(tr);
210 	__raw_spin_unlock(&wakeup_lock);
211 	local_irq_restore(flags);
212 }
213 
214 static void
215 probe_wakeup(struct rq *rq, struct task_struct *p, int success)
216 {
217 	struct trace_array_cpu *data;
218 	int cpu = smp_processor_id();
219 	unsigned long flags;
220 	long disabled;
221 	int pc;
222 
223 	if (likely(!tracer_enabled))
224 		return;
225 
226 	tracing_record_cmdline(p);
227 	tracing_record_cmdline(current);
228 
229 	if ((wakeup_rt && !rt_task(p)) ||
230 			p->prio >= wakeup_prio ||
231 			p->prio >= current->prio)
232 		return;
233 
234 	pc = preempt_count();
235 	disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
236 	if (unlikely(disabled != 1))
237 		goto out;
238 
239 	/* interrupts should be off from try_to_wake_up */
240 	__raw_spin_lock(&wakeup_lock);
241 
242 	/* check for races. */
243 	if (!tracer_enabled || p->prio >= wakeup_prio)
244 		goto out_locked;
245 
246 	/* reset the trace */
247 	__wakeup_reset(wakeup_trace);
248 
249 	wakeup_cpu = task_cpu(p);
250 	wakeup_prio = p->prio;
251 
252 	wakeup_task = p;
253 	get_task_struct(wakeup_task);
254 
255 	local_save_flags(flags);
256 
257 	data = wakeup_trace->data[wakeup_cpu];
258 	data->preempt_timestamp = ftrace_now(cpu);
259 	tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
260 
261 	/*
262 	 * We must be careful in using CALLER_ADDR2. But since wake_up
263 	 * is not called by an assembly function  (where as schedule is)
264 	 * it should be safe to use it here.
265 	 */
266 	trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
267 
268 out_locked:
269 	__raw_spin_unlock(&wakeup_lock);
270 out:
271 	atomic_dec(&wakeup_trace->data[cpu]->disabled);
272 }
273 
274 static void start_wakeup_tracer(struct trace_array *tr)
275 {
276 	int ret;
277 
278 	ret = register_trace_sched_wakeup(probe_wakeup);
279 	if (ret) {
280 		pr_info("wakeup trace: Couldn't activate tracepoint"
281 			" probe to kernel_sched_wakeup\n");
282 		return;
283 	}
284 
285 	ret = register_trace_sched_wakeup_new(probe_wakeup);
286 	if (ret) {
287 		pr_info("wakeup trace: Couldn't activate tracepoint"
288 			" probe to kernel_sched_wakeup_new\n");
289 		goto fail_deprobe;
290 	}
291 
292 	ret = register_trace_sched_switch(probe_wakeup_sched_switch);
293 	if (ret) {
294 		pr_info("sched trace: Couldn't activate tracepoint"
295 			" probe to kernel_sched_switch\n");
296 		goto fail_deprobe_wake_new;
297 	}
298 
299 	wakeup_reset(tr);
300 
301 	/*
302 	 * Don't let the tracer_enabled = 1 show up before
303 	 * the wakeup_task is reset. This may be overkill since
304 	 * wakeup_reset does a spin_unlock after setting the
305 	 * wakeup_task to NULL, but I want to be safe.
306 	 * This is a slow path anyway.
307 	 */
308 	smp_wmb();
309 
310 	register_ftrace_function(&trace_ops);
311 
312 	if (tracing_is_enabled())
313 		tracer_enabled = 1;
314 	else
315 		tracer_enabled = 0;
316 
317 	return;
318 fail_deprobe_wake_new:
319 	unregister_trace_sched_wakeup_new(probe_wakeup);
320 fail_deprobe:
321 	unregister_trace_sched_wakeup(probe_wakeup);
322 }
323 
324 static void stop_wakeup_tracer(struct trace_array *tr)
325 {
326 	tracer_enabled = 0;
327 	unregister_ftrace_function(&trace_ops);
328 	unregister_trace_sched_switch(probe_wakeup_sched_switch);
329 	unregister_trace_sched_wakeup_new(probe_wakeup);
330 	unregister_trace_sched_wakeup(probe_wakeup);
331 }
332 
333 static int __wakeup_tracer_init(struct trace_array *tr)
334 {
335 	save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
336 	trace_flags |= TRACE_ITER_LATENCY_FMT;
337 
338 	tracing_max_latency = 0;
339 	wakeup_trace = tr;
340 	start_wakeup_tracer(tr);
341 	return 0;
342 }
343 
344 static int wakeup_tracer_init(struct trace_array *tr)
345 {
346 	wakeup_rt = 0;
347 	return __wakeup_tracer_init(tr);
348 }
349 
350 static int wakeup_rt_tracer_init(struct trace_array *tr)
351 {
352 	wakeup_rt = 1;
353 	return __wakeup_tracer_init(tr);
354 }
355 
356 static void wakeup_tracer_reset(struct trace_array *tr)
357 {
358 	stop_wakeup_tracer(tr);
359 	/* make sure we put back any tasks we are tracing */
360 	wakeup_reset(tr);
361 
362 	if (!save_lat_flag)
363 		trace_flags &= ~TRACE_ITER_LATENCY_FMT;
364 }
365 
366 static void wakeup_tracer_start(struct trace_array *tr)
367 {
368 	wakeup_reset(tr);
369 	tracer_enabled = 1;
370 }
371 
372 static void wakeup_tracer_stop(struct trace_array *tr)
373 {
374 	tracer_enabled = 0;
375 }
376 
377 static struct tracer wakeup_tracer __read_mostly =
378 {
379 	.name		= "wakeup",
380 	.init		= wakeup_tracer_init,
381 	.reset		= wakeup_tracer_reset,
382 	.start		= wakeup_tracer_start,
383 	.stop		= wakeup_tracer_stop,
384 	.print_max	= 1,
385 #ifdef CONFIG_FTRACE_SELFTEST
386 	.selftest    = trace_selftest_startup_wakeup,
387 #endif
388 };
389 
390 static struct tracer wakeup_rt_tracer __read_mostly =
391 {
392 	.name		= "wakeup_rt",
393 	.init		= wakeup_rt_tracer_init,
394 	.reset		= wakeup_tracer_reset,
395 	.start		= wakeup_tracer_start,
396 	.stop		= wakeup_tracer_stop,
397 	.wait_pipe	= poll_wait_pipe,
398 	.print_max	= 1,
399 #ifdef CONFIG_FTRACE_SELFTEST
400 	.selftest    = trace_selftest_startup_wakeup,
401 #endif
402 };
403 
404 __init static int init_wakeup_tracer(void)
405 {
406 	int ret;
407 
408 	ret = register_tracer(&wakeup_tracer);
409 	if (ret)
410 		return ret;
411 
412 	ret = register_tracer(&wakeup_rt_tracer);
413 	if (ret)
414 		return ret;
415 
416 	return 0;
417 }
418 device_initcall(init_wakeup_tracer);
419