trace_sched_wakeup.c 6.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
/*
 * trace task wakeup timings
 *
 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
 *
 * Based on code from the latency_tracer, that is:
 *
 *  Copyright (C) 2004-2006 Ingo Molnar
 *  Copyright (C) 2004 William Lee Irwin III
 */
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/debugfs.h>
#include <linux/kallsyms.h>
#include <linux/uaccess.h>
#include <linux/ftrace.h>

#include "trace.h"

static struct trace_array	*wakeup_trace;
static int __read_mostly	tracer_enabled;

static struct task_struct	*wakeup_task;
static int			wakeup_cpu;
static unsigned			wakeup_prio = -1;

static DEFINE_SPINLOCK(wakeup_lock);

I
Ingo Molnar 已提交
30
static void __wakeup_reset(struct trace_array *tr);
31 32 33 34

/*
 * Should this new latency be reported/recorded?
 */
I
Ingo Molnar 已提交
35
static int report_latency(cycle_t delta)
36 37 38 39 40 41 42 43 44 45 46
{
	if (tracing_thresh) {
		if (delta < tracing_thresh)
			return 0;
	} else {
		if (delta <= tracing_max_latency)
			return 0;
	}
	return 1;
}

I
Ingo Molnar 已提交
47
void
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
wakeup_sched_switch(struct task_struct *prev, struct task_struct *next)
{
	unsigned long latency = 0, t0 = 0, t1 = 0;
	struct trace_array *tr = wakeup_trace;
	struct trace_array_cpu *data;
	cycle_t T0, T1, delta;
	unsigned long flags;
	long disabled;
	int cpu;

	if (unlikely(!tracer_enabled))
		return;

	/*
	 * When we start a new trace, we set wakeup_task to NULL
	 * and then set tracer_enabled = 1. We want to make sure
	 * that another CPU does not see the tracer_enabled = 1
	 * and the wakeup_task with an older task, that might
	 * actually be the same as next.
	 */
	smp_rmb();

	if (next != wakeup_task)
		return;

	/* The task we are waitng for is waking up */
	data = tr->data[wakeup_cpu];

	/* disable local data, not wakeup_cpu data */
	cpu = raw_smp_processor_id();
	disabled = atomic_inc_return(&tr->data[cpu]->disabled);
	if (likely(disabled != 1))
		goto out;

	spin_lock_irqsave(&wakeup_lock, flags);

	/* We could race with grabbing wakeup_lock */
	if (unlikely(!tracer_enabled || next != wakeup_task))
		goto out_unlock;

88
	trace_function(tr, data, CALLER_ADDR1, CALLER_ADDR2, flags);
89 90 91 92 93 94

	/*
	 * usecs conversion is slow so we try to delay the conversion
	 * as long as possible:
	 */
	T0 = data->preempt_timestamp;
I
Ingo Molnar 已提交
95
	T1 = ftrace_now(cpu);
96 97 98 99 100 101 102 103 104 105 106 107 108 109
	delta = T1-T0;

	if (!report_latency(delta))
		goto out_unlock;

	latency = nsecs_to_usecs(delta);

	tracing_max_latency = delta;
	t0 = nsecs_to_usecs(T0);
	t1 = nsecs_to_usecs(T1);

	update_max_tr(tr, wakeup_task, wakeup_cpu);

	if (tracing_thresh) {
I
Ingo Molnar 已提交
110 111
		printk(KERN_INFO "(%16s-%-5d|#%d):"
			" %lu us wakeup latency violates %lu us threshold.\n",
112 113
				wakeup_task->comm, wakeup_task->pid,
				raw_smp_processor_id(),
I
Ingo Molnar 已提交
114
				latency, nsecs_to_usecs(tracing_thresh));
115
	} else {
I
Ingo Molnar 已提交
116 117
		printk(KERN_INFO "(%16s-%-5d|#%d):"
			" new %lu us maximum wakeup latency.\n",
118
				wakeup_task->comm, wakeup_task->pid,
I
Ingo Molnar 已提交
119
				cpu, latency);
120 121 122 123 124 125 126 127 128
	}

out_unlock:
	__wakeup_reset(tr);
	spin_unlock_irqrestore(&wakeup_lock, flags);
out:
	atomic_dec(&tr->data[cpu]->disabled);
}

I
Ingo Molnar 已提交
129
static void __wakeup_reset(struct trace_array *tr)
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
{
	struct trace_array_cpu *data;
	int cpu;

	assert_spin_locked(&wakeup_lock);

	for_each_possible_cpu(cpu) {
		data = tr->data[cpu];
		tracing_reset(data);
	}

	wakeup_cpu = -1;
	wakeup_prio = -1;

	if (wakeup_task)
		put_task_struct(wakeup_task);

	wakeup_task = NULL;
}

I
Ingo Molnar 已提交
150
static void wakeup_reset(struct trace_array *tr)
151 152 153 154 155 156 157 158
{
	unsigned long flags;

	spin_lock_irqsave(&wakeup_lock, flags);
	__wakeup_reset(tr);
	spin_unlock_irqrestore(&wakeup_lock, flags);
}

I
Ingo Molnar 已提交
159
static void
160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
wakeup_check_start(struct trace_array *tr, struct task_struct *p,
		   struct task_struct *curr)
{
	int cpu = smp_processor_id();
	unsigned long flags;
	long disabled;

	if (likely(!rt_task(p)) ||
			p->prio >= wakeup_prio ||
			p->prio >= curr->prio)
		return;

	disabled = atomic_inc_return(&tr->data[cpu]->disabled);
	if (unlikely(disabled != 1))
		goto out;

	/* interrupts should be off from try_to_wake_up */
	spin_lock(&wakeup_lock);

	/* check for races. */
	if (!tracer_enabled || p->prio >= wakeup_prio)
		goto out_locked;

	/* reset the trace */
	__wakeup_reset(tr);

	wakeup_cpu = task_cpu(p);
	wakeup_prio = p->prio;

	wakeup_task = p;
	get_task_struct(wakeup_task);

	local_save_flags(flags);

I
Ingo Molnar 已提交
194
	tr->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
195 196
	trace_function(tr, tr->data[wakeup_cpu],
		       CALLER_ADDR1, CALLER_ADDR2, flags);
197 198 199 200 201 202 203

out_locked:
	spin_unlock(&wakeup_lock);
out:
	atomic_dec(&tr->data[cpu]->disabled);
}

I
Ingo Molnar 已提交
204
void
205
wakeup_sched_wakeup(struct task_struct *wakee, struct task_struct *curr)
206 207 208 209 210 211 212
{
	if (likely(!tracer_enabled))
		return;

	wakeup_check_start(wakeup_trace, wakee, curr);
}

I
Ingo Molnar 已提交
213
void
214 215 216 217 218 219 220 221
ftrace_wake_up_new_task(struct task_struct *wakee, struct task_struct *curr)
{
	if (likely(!tracer_enabled))
		return;

	wakeup_check_start(wakeup_trace, wakee, curr);
}

I
Ingo Molnar 已提交
222
static void start_wakeup_tracer(struct trace_array *tr)
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
{
	wakeup_reset(tr);

	/*
	 * Don't let the tracer_enabled = 1 show up before
	 * the wakeup_task is reset. This may be overkill since
	 * wakeup_reset does a spin_unlock after setting the
	 * wakeup_task to NULL, but I want to be safe.
	 * This is a slow path anyway.
	 */
	smp_wmb();

	tracer_enabled = 1;

	return;
}

I
Ingo Molnar 已提交
240
static void stop_wakeup_tracer(struct trace_array *tr)
241 242 243 244
{
	tracer_enabled = 0;
}

I
Ingo Molnar 已提交
245
static void wakeup_tracer_init(struct trace_array *tr)
246 247 248 249 250 251 252
{
	wakeup_trace = tr;

	if (tr->ctrl)
		start_wakeup_tracer(tr);
}

I
Ingo Molnar 已提交
253
static void wakeup_tracer_reset(struct trace_array *tr)
254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269
{
	if (tr->ctrl) {
		stop_wakeup_tracer(tr);
		/* make sure we put back any tasks we are tracing */
		wakeup_reset(tr);
	}
}

static void wakeup_tracer_ctrl_update(struct trace_array *tr)
{
	if (tr->ctrl)
		start_wakeup_tracer(tr);
	else
		stop_wakeup_tracer(tr);
}

I
Ingo Molnar 已提交
270
static void wakeup_tracer_open(struct trace_iterator *iter)
271 272 273 274 275 276
{
	/* stop the trace while dumping */
	if (iter->tr->ctrl)
		stop_wakeup_tracer(iter->tr);
}

I
Ingo Molnar 已提交
277
static void wakeup_tracer_close(struct trace_iterator *iter)
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
{
	/* forget about any processes we were recording */
	if (iter->tr->ctrl)
		start_wakeup_tracer(iter->tr);
}

static struct tracer wakeup_tracer __read_mostly =
{
	.name		= "wakeup",
	.init		= wakeup_tracer_init,
	.reset		= wakeup_tracer_reset,
	.open		= wakeup_tracer_open,
	.close		= wakeup_tracer_close,
	.ctrl_update	= wakeup_tracer_ctrl_update,
	.print_max	= 1,
S
Steven Rostedt 已提交
293 294 295
#ifdef CONFIG_FTRACE_SELFTEST
	.selftest    = trace_selftest_startup_wakeup,
#endif
296 297 298 299 300 301 302 303 304 305 306 307 308
};

__init static int init_wakeup_tracer(void)
{
	int ret;

	ret = register_tracer(&wakeup_tracer);
	if (ret)
		return ret;

	return 0;
}
device_initcall(init_wakeup_tracer);