1
2#ifdef CONFIG_SCHEDSTATS
3/*
4 * bump this up when changing the output format or the meaning of an existing
5 * format, so that tools can adapt (or abort)
6 */
7#define SCHEDSTAT_VERSION 15
8
9static int show_schedstat(struct seq_file *seq, void *v)
10{
11	int cpu;
12	int mask_len = DIV_ROUND_UP(NR_CPUS, 32) * 9;
13	char *mask_str = kmalloc(mask_len, GFP_KERNEL);
14
15	if (mask_str == NULL)
16		return -ENOMEM;
17
18	seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
19	seq_printf(seq, "timestamp %lu\n", jiffies);
20	for_each_online_cpu(cpu) {
21		struct rq *rq = cpu_rq(cpu);
22#ifdef CONFIG_SMP
23		struct sched_domain *sd;
24		int dcount = 0;
25#endif
26
27		/* runqueue-specific stats */
28		seq_printf(seq,
29		    "cpu%d %u %u %u %u %u %u %llu %llu %lu",
30		    cpu, rq->yld_count,
31		    rq->sched_switch, rq->sched_count, rq->sched_goidle,
32		    rq->ttwu_count, rq->ttwu_local,
33		    rq->rq_cpu_time,
34		    rq->rq_sched_info.run_delay, rq->rq_sched_info.pcount);
35
36		seq_printf(seq, "\n");
37
38#ifdef CONFIG_SMP
39		/* domain-specific stats */
40		preempt_disable();
41		for_each_domain(cpu, sd) {
42			enum cpu_idle_type itype;
43
44			cpumask_scnprintf(mask_str, mask_len,
45					  sched_domain_span(sd));
46			seq_printf(seq, "domain%d %s", dcount++, mask_str);
47			for (itype = CPU_IDLE; itype < CPU_MAX_IDLE_TYPES;
48					itype++) {
49				seq_printf(seq, " %u %u %u %u %u %u %u %u",
50				    sd->lb_count[itype],
51				    sd->lb_balanced[itype],
52				    sd->lb_failed[itype],
53				    sd->lb_imbalance[itype],
54				    sd->lb_gained[itype],
55				    sd->lb_hot_gained[itype],
56				    sd->lb_nobusyq[itype],
57				    sd->lb_nobusyg[itype]);
58			}
59			seq_printf(seq,
60				   " %u %u %u %u %u %u %u %u %u %u %u %u\n",
61			    sd->alb_count, sd->alb_failed, sd->alb_pushed,
62			    sd->sbe_count, sd->sbe_balanced, sd->sbe_pushed,
63			    sd->sbf_count, sd->sbf_balanced, sd->sbf_pushed,
64			    sd->ttwu_wake_remote, sd->ttwu_move_affine,
65			    sd->ttwu_move_balance);
66		}
67		preempt_enable();
68#endif
69	}
70	kfree(mask_str);
71	return 0;
72}
73
74static int schedstat_open(struct inode *inode, struct file *file)
75{
76	unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
77	char *buf = kmalloc(size, GFP_KERNEL);
78	struct seq_file *m;
79	int res;
80
81	if (!buf)
82		return -ENOMEM;
83	res = single_open(file, show_schedstat, NULL);
84	if (!res) {
85		m = file->private_data;
86		m->buf = buf;
87		m->size = size;
88	} else
89		kfree(buf);
90	return res;
91}
92
93static const struct file_operations proc_schedstat_operations = {
94	.open    = schedstat_open,
95	.read    = seq_read,
96	.llseek  = seq_lseek,
97	.release = single_release,
98};
99
100static int __init proc_schedstat_init(void)
101{
102	proc_create("schedstat", 0, NULL, &proc_schedstat_operations);
103	return 0;
104}
105module_init(proc_schedstat_init);
106
107/*
108 * Expects runqueue lock to be held for atomicity of update
109 */
110static inline void
111rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
112{
113	if (rq) {
114		rq->rq_sched_info.run_delay += delta;
115		rq->rq_sched_info.pcount++;
116	}
117}
118
119/*
120 * Expects runqueue lock to be held for atomicity of update
121 */
122static inline void
123rq_sched_info_depart(struct rq *rq, unsigned long long delta)
124{
125	if (rq)
126		rq->rq_cpu_time += delta;
127}
128
129static inline void
130rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
131{
132	if (rq)
133		rq->rq_sched_info.run_delay += delta;
134}
135# define schedstat_inc(rq, field)	do { (rq)->field++; } while (0)
136# define schedstat_add(rq, field, amt)	do { (rq)->field += (amt); } while (0)
137# define schedstat_set(var, val)	do { var = (val); } while (0)
138#else /* !CONFIG_SCHEDSTATS */
139static inline void
140rq_sched_info_arrive(struct rq *rq, unsigned long long delta)
141{}
142static inline void
143rq_sched_info_dequeued(struct rq *rq, unsigned long long delta)
144{}
145static inline void
146rq_sched_info_depart(struct rq *rq, unsigned long long delta)
147{}
148# define schedstat_inc(rq, field)	do { } while (0)
149# define schedstat_add(rq, field, amt)	do { } while (0)
150# define schedstat_set(var, val)	do { } while (0)
151#endif
152
153#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
154static inline void sched_info_reset_dequeued(struct task_struct *t)
155{
156	t->sched_info.last_queued = 0;
157}
158
159/*
160 * Called when a process is dequeued from the active array and given
161 * the cpu.  We should note that with the exception of interactive
162 * tasks, the expired queue will become the active queue after the active
163 * queue is empty, without explicitly dequeuing and requeuing tasks in the
164 * expired queue.  (Interactive tasks may be requeued directly to the
165 * active queue, thus delaying tasks in the expired queue from running;
166 * see scheduler_tick()).
167 *
168 * Though we are interested in knowing how long it was from the *first* time a
169 * task was queued to the time that it finally hit a cpu, we call this routine
170 * from dequeue_task() to account for possible rq->clock skew across cpus. The
171 * delta taken on each cpu would annul the skew.
172 */
173static inline void sched_info_dequeued(struct task_struct *t)
174{
175	unsigned long long now = task_rq(t)->clock, delta = 0;
176
177	if (unlikely(sched_info_on()))
178		if (t->sched_info.last_queued)
179			delta = now - t->sched_info.last_queued;
180	sched_info_reset_dequeued(t);
181	t->sched_info.run_delay += delta;
182
183	rq_sched_info_dequeued(task_rq(t), delta);
184}
185
186/*
187 * Called when a task finally hits the cpu.  We can now calculate how
188 * long it was waiting to run.  We also note when it began so that we
189 * can keep stats on how long its timeslice is.
190 */
191static void sched_info_arrive(struct task_struct *t)
192{
193	unsigned long long now = task_rq(t)->clock, delta = 0;
194
195	if (t->sched_info.last_queued)
196		delta = now - t->sched_info.last_queued;
197	sched_info_reset_dequeued(t);
198	t->sched_info.run_delay += delta;
199	t->sched_info.last_arrival = now;
200	t->sched_info.pcount++;
201
202	rq_sched_info_arrive(task_rq(t), delta);
203}
204
205/*
206 * Called when a process is queued into either the active or expired
207 * array.  The time is noted and later used to determine how long we
208 * had to wait for us to reach the cpu.  Since the expired queue will
209 * become the active queue after active queue is empty, without dequeuing
210 * and requeuing any tasks, we are interested in queuing to either. It
211 * is unusual but not impossible for tasks to be dequeued and immediately
212 * requeued in the same or another array: this can happen in sched_yield(),
213 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
214 * to runqueue.
215 *
216 * This function is only called from enqueue_task(), but also only updates
217 * the timestamp if it is already not set.  It's assumed that
218 * sched_info_dequeued() will clear that stamp when appropriate.
219 */
220static inline void sched_info_queued(struct task_struct *t)
221{
222	if (unlikely(sched_info_on()))
223		if (!t->sched_info.last_queued)
224			t->sched_info.last_queued = task_rq(t)->clock;
225}
226
227/*
228 * Called when a process ceases being the active-running process, either
229 * voluntarily or involuntarily.  Now we can calculate how long we ran.
230 * Also, if the process is still in the TASK_RUNNING state, call
231 * sched_info_queued() to mark that it has now again started waiting on
232 * the runqueue.
233 */
234static inline void sched_info_depart(struct task_struct *t)
235{
236	unsigned long long delta = task_rq(t)->clock -
237					t->sched_info.last_arrival;
238
239	rq_sched_info_depart(task_rq(t), delta);
240
241	if (t->state == TASK_RUNNING)
242		sched_info_queued(t);
243}
244
245/*
246 * Called when tasks are switched involuntarily due, typically, to expiring
247 * their time slice.  (This may also be called when switching to or from
248 * the idle task.)  We are only called when prev != next.
249 */
250static inline void
251__sched_info_switch(struct task_struct *prev, struct task_struct *next)
252{
253	struct rq *rq = task_rq(prev);
254
255	/*
256	 * prev now departs the cpu.  It's not interesting to record
257	 * stats about how efficient we were at scheduling the idle
258	 * process, however.
259	 */
260	if (prev != rq->idle)
261		sched_info_depart(prev);
262
263	if (next != rq->idle)
264		sched_info_arrive(next);
265}
266static inline void
267sched_info_switch(struct task_struct *prev, struct task_struct *next)
268{
269	if (unlikely(sched_info_on()))
270		__sched_info_switch(prev, next);
271}
272#else
273#define sched_info_queued(t)			do { } while (0)
274#define sched_info_reset_dequeued(t)	do { } while (0)
275#define sched_info_dequeued(t)			do { } while (0)
276#define sched_info_switch(t, next)		do { } while (0)
277#endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
278
279/*
280 * The following are functions that support scheduler-internal time accounting.
281 * These functions are generally called at the timer tick.  None of this depends
282 * on CONFIG_SCHEDSTATS.
283 */
284
285/**
286 * account_group_user_time - Maintain utime for a thread group.
287 *
288 * @tsk:	Pointer to task structure.
289 * @cputime:	Time value by which to increment the utime field of the
290 *		thread_group_cputime structure.
291 *
292 * If thread group time is being maintained, get the structure for the
293 * running CPU and update the utime field there.
294 */
295static inline void account_group_user_time(struct task_struct *tsk,
296					   cputime_t cputime)
297{
298	struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
299
300	if (!cputimer->running)
301		return;
302
303	spin_lock(&cputimer->lock);
304	cputimer->cputime.utime =
305		cputime_add(cputimer->cputime.utime, cputime);
306	spin_unlock(&cputimer->lock);
307}
308
309/**
310 * account_group_system_time - Maintain stime for a thread group.
311 *
312 * @tsk:	Pointer to task structure.
313 * @cputime:	Time value by which to increment the stime field of the
314 *		thread_group_cputime structure.
315 *
316 * If thread group time is being maintained, get the structure for the
317 * running CPU and update the stime field there.
318 */
319static inline void account_group_system_time(struct task_struct *tsk,
320					     cputime_t cputime)
321{
322	struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
323
324	if (!cputimer->running)
325		return;
326
327	spin_lock(&cputimer->lock);
328	cputimer->cputime.stime =
329		cputime_add(cputimer->cputime.stime, cputime);
330	spin_unlock(&cputimer->lock);
331}
332
333/**
334 * account_group_exec_runtime - Maintain exec runtime for a thread group.
335 *
336 * @tsk:	Pointer to task structure.
337 * @ns:		Time value by which to increment the sum_exec_runtime field
338 *		of the thread_group_cputime structure.
339 *
340 * If thread group time is being maintained, get the structure for the
341 * running CPU and update the sum_exec_runtime field there.
342 */
343static inline void account_group_exec_runtime(struct task_struct *tsk,
344					      unsigned long long ns)
345{
346	struct thread_group_cputimer *cputimer = &tsk->signal->cputimer;
347
348	if (!cputimer->running)
349		return;
350
351	spin_lock(&cputimer->lock);
352	cputimer->cputime.sum_exec_runtime += ns;
353	spin_unlock(&cputimer->lock);
354}
355