Kernel

Kernel Git Source Tree

Root/kernel/sched_fair.c

1	/*
2	* Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3	*
4	* Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5	*
6	* Interactivity improvements by Mike Galbraith
7	* (C) 2007 Mike Galbraith <efault@gmx.de>
8	*
9	* Various enhancements by Dmitry Adamushko.
10	* (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11	*
12	* Group scheduling enhancements by Srivatsa Vaddagiri
13	* Copyright IBM Corporation, 2007
14	* Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15	*
16	* Scaled math optimizations by Thomas Gleixner
17	* Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18	*
19	* Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20	* Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
21	*/
22
23	#include <linux/latencytop.h>
24	#include <linux/sched.h>
25
26	/*
27	* Targeted preemption latency for CPU-bound tasks:
28	* (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds)
29	*
30	* NOTE: this latency value is not the same as the concept of
31	* 'timeslice length' - timeslices in CFS are of variable length
32	* and have no persistent notion like in traditional, time-slice
33	* based scheduling concepts.
34	*
35	* (to see the precise effective timeslice length of your workload,
36	* run vmstat and monitor the context-switches (cs) field)
37	*/
38	unsigned int sysctl_sched_latency = 5000000ULL;
39	unsigned int normalized_sysctl_sched_latency = 5000000ULL;
40
41	/*
42	* The initial- and re-scaling of tunables is configurable
43	* (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
44	*
45	* Options are:
46	* SCHED_TUNABLESCALING_NONE - unscaled, always *1
47	* SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48	* SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
49	*/
50	enum sched_tunable_scaling sysctl_sched_tunable_scaling
51	= SCHED_TUNABLESCALING_LOG;
52
53	/*
54	* Minimal preemption granularity for CPU-bound tasks:
55	* (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
56	*/
57	unsigned int sysctl_sched_min_granularity = 1000000ULL;
58	unsigned int normalized_sysctl_sched_min_granularity = 1000000ULL;
59
60	/*
61	* is kept at sysctl_sched_latency / sysctl_sched_min_granularity
62	*/
63	static unsigned int sched_nr_latency = 5;
64
65	/*
66	* After fork, child runs first. If set to 0 (default) then
67	* parent will (try to) run first.
68	*/
69	unsigned int sysctl_sched_child_runs_first __read_mostly;
70
71	/*
72	* sys_sched_yield() compat mode
73	*
74	* This option switches the agressive yield implementation of the
75	* old scheduler back on.
76	*/
77	unsigned int __read_mostly sysctl_sched_compat_yield;
78
79	/*
80	* SCHED_OTHER wake-up granularity.
81	* (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82	*
83	* This option delays the preemption effects of decoupled workloads
84	* and reduces their over-scheduling. Synchronous workloads will still
85	* have immediate wakeup/sleep latencies.
86	*/
87	unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88	unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
89
90	const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91
92	static const struct sched_class fair_sched_class;
93
94	/**************************************************************
95	* CFS operations on generic schedulable entities:
96	*/
97
98	#ifdef CONFIG_FAIR_GROUP_SCHED
99
100	/* cpu runqueue to which this cfs_rq is attached */
101	static inline struct rq rq_of(struct cfs_rq cfs_rq)
102	{
103	return cfs_rq->rq;
104	}
105
106	/* An entity is a task if it doesn't "own" a runqueue */
107	#define entity_is_task(se) (!se->my_q)
108
109	static inline struct task_struct task_of(struct sched_entity se)
110	{
111	#ifdef CONFIG_SCHED_DEBUG
112	WARN_ON_ONCE(!entity_is_task(se));
113	#endif
114	return container_of(se, struct task_struct, se);
115	}
116
117	/* Walk up scheduling entities hierarchy */
118	#define for_each_sched_entity(se) \
119	for (; se; se = se->parent)
120
121	static inline struct cfs_rq task_cfs_rq(struct task_struct p)
122	{
123	return p->se.cfs_rq;
124	}
125
126	/* runqueue on which this entity is (to be) queued */
127	static inline struct cfs_rq cfs_rq_of(struct sched_entity se)
128	{
129	return se->cfs_rq;
130	}
131
132	/* runqueue "owned" by this group */
133	static inline struct cfs_rq group_cfs_rq(struct sched_entity grp)
134	{
135	return grp->my_q;
136	}
137
138	/* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139	* another cpu ('this_cpu')
140	*/
141	static inline struct cfs_rq cpu_cfs_rq(struct cfs_rq cfs_rq, int this_cpu)
142	{
143	return cfs_rq->tg->cfs_rq[this_cpu];
144	}
145
146	/* Iterate thr' all leaf cfs_rq's on a runqueue */
147	#define for_each_leaf_cfs_rq(rq, cfs_rq) \
148	list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
149
150	/* Do the two (enqueued) entities belong to the same group ? */
151	static inline int
152	is_same_group(struct sched_entity se, struct sched_entity pse)
153	{
154	if (se->cfs_rq == pse->cfs_rq)
155	return 1;
156
157	return 0;
158	}
159
160	static inline struct sched_entity parent_entity(struct sched_entity se)
161	{
162	return se->parent;
163	}
164
165	/* return depth at which a sched entity is present in the hierarchy */
166	static inline int depth_se(struct sched_entity *se)
167	{
168	int depth = 0;
169
170	for_each_sched_entity(se)
171	depth++;
172
173	return depth;
174	}
175
176	static void
177	find_matching_se(struct sched_entity se, struct sched_entity pse)
178	{
179	int se_depth, pse_depth;
180
181	/*
182	* preemption test can be made between sibling entities who are in the
183	* same cfs_rq i.e who have a common parent. Walk up the hierarchy of
184	* both tasks until we find their ancestors who are siblings of common
185	* parent.
186	*/
187
188	/* First walk up until both entities are at same depth */
189	se_depth = depth_se(*se);
190	pse_depth = depth_se(*pse);
191
192	while (se_depth > pse_depth) {
193	se_depth--;
194	se = parent_entity(se);
195	}
196
197	while (pse_depth > se_depth) {
198	pse_depth--;
199	pse = parent_entity(pse);
200	}
201
202	while (!is_same_group(se, pse)) {
203	se = parent_entity(se);
204	pse = parent_entity(pse);
205	}
206	}
207
208	#else /* !CONFIG_FAIR_GROUP_SCHED */
209
210	static inline struct task_struct task_of(struct sched_entity se)
211	{
212	return container_of(se, struct task_struct, se);
213	}
214
215	static inline struct rq rq_of(struct cfs_rq cfs_rq)
216	{
217	return container_of(cfs_rq, struct rq, cfs);
218	}
219
220	#define entity_is_task(se) 1
221
222	#define for_each_sched_entity(se) \
223	for (; se; se = NULL)
224
225	static inline struct cfs_rq task_cfs_rq(struct task_struct p)
226	{
227	return &task_rq(p)->cfs;
228	}
229
230	static inline struct cfs_rq cfs_rq_of(struct sched_entity se)
231	{
232	struct task_struct *p = task_of(se);
233	struct rq *rq = task_rq(p);
234
235	return &rq->cfs;
236	}
237
238	/* runqueue "owned" by this group */
239	static inline struct cfs_rq group_cfs_rq(struct sched_entity grp)
240	{
241	return NULL;
242	}
243
244	static inline struct cfs_rq cpu_cfs_rq(struct cfs_rq cfs_rq, int this_cpu)
245	{
246	return &cpu_rq(this_cpu)->cfs;
247	}
248
249	#define for_each_leaf_cfs_rq(rq, cfs_rq) \
250	for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
251
252	static inline int
253	is_same_group(struct sched_entity se, struct sched_entity pse)
254	{
255	return 1;
256	}
257
258	static inline struct sched_entity parent_entity(struct sched_entity se)
259	{
260	return NULL;
261	}
262
263	static inline void
264	find_matching_se(struct sched_entity se, struct sched_entity pse)
265	{
266	}
267
268	#endif /* CONFIG_FAIR_GROUP_SCHED */
269
270
271	/**************************************************************
272	* Scheduling class tree data structure manipulation methods:
273	*/
274
275	static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
276	{
277	s64 delta = (s64)(vruntime - min_vruntime);
278	if (delta > 0)
279	min_vruntime = vruntime;
280
281	return min_vruntime;
282	}
283
284	static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
285	{
286	s64 delta = (s64)(vruntime - min_vruntime);
287	if (delta < 0)
288	min_vruntime = vruntime;
289
290	return min_vruntime;
291	}
292
293	static inline int entity_before(struct sched_entity *a,
294	struct sched_entity *b)
295	{
296	return (s64)(a->vruntime - b->vruntime) < 0;
297	}
298
299	static inline s64 entity_key(struct cfs_rq cfs_rq, struct sched_entity se)
300	{
301	return se->vruntime - cfs_rq->min_vruntime;
302	}
303
304	static void update_min_vruntime(struct cfs_rq *cfs_rq)
305	{
306	u64 vruntime = cfs_rq->min_vruntime;
307
308	if (cfs_rq->curr)
309	vruntime = cfs_rq->curr->vruntime;
310
311	if (cfs_rq->rb_leftmost) {
312	struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
313	struct sched_entity,
314	run_node);
315
316	if (!cfs_rq->curr)
317	vruntime = se->vruntime;
318	else
319	vruntime = min_vruntime(vruntime, se->vruntime);
320	}
321
322	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
323	}
324
325	/*
326	* Enqueue an entity into the rb-tree:
327	*/
328	static void __enqueue_entity(struct cfs_rq cfs_rq, struct sched_entity se)
329	{
330	struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
331	struct rb_node *parent = NULL;
332	struct sched_entity *entry;
333	s64 key = entity_key(cfs_rq, se);
334	int leftmost = 1;
335
336	/*
337	* Find the right place in the rbtree:
338	*/
339	while (*link) {
340	parent = *link;
341	entry = rb_entry(parent, struct sched_entity, run_node);
342	/*
343	* We dont care about collisions. Nodes with
344	* the same key stay together.
345	*/
346	if (key < entity_key(cfs_rq, entry)) {
347	link = &parent->rb_left;
348	} else {
349	link = &parent->rb_right;
350	leftmost = 0;
351	}
352	}
353
354	/*
355	* Maintain a cache of leftmost tree entries (it is frequently
356	* used):
357	*/
358	if (leftmost)
359	cfs_rq->rb_leftmost = &se->run_node;
360
361	rb_link_node(&se->run_node, parent, link);
362	rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
363	}
364
365	static void __dequeue_entity(struct cfs_rq cfs_rq, struct sched_entity se)
366	{
367	if (cfs_rq->rb_leftmost == &se->run_node) {
368	struct rb_node *next_node;
369
370	next_node = rb_next(&se->run_node);
371	cfs_rq->rb_leftmost = next_node;
372	}
373
374	rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
375	}
376
377	static struct sched_entity __pick_next_entity(struct cfs_rq cfs_rq)
378	{
379	struct rb_node *left = cfs_rq->rb_leftmost;
380
381	if (!left)
382	return NULL;
383
384	return rb_entry(left, struct sched_entity, run_node);
385	}
386
387	static struct sched_entity __pick_last_entity(struct cfs_rq cfs_rq)
388	{
389	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
390
391	if (!last)
392	return NULL;
393
394	return rb_entry(last, struct sched_entity, run_node);
395	}
396
397	/**************************************************************
398	* Scheduling class statistics methods:
399	*/
400
401	#ifdef CONFIG_SCHED_DEBUG
402	int sched_proc_update_handler(struct ctl_table *table, int write,
403	void __user buffer, size_t lenp,
404	loff_t *ppos)
405	{
406	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
407	int factor = get_update_sysctl_factor();
408
409	if (ret \|\| !write)
410	return ret;
411
412	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
413	sysctl_sched_min_granularity);
414
415	#define WRT_SYSCTL(name) \
416	(normalized_sysctl_##name = sysctl_##name / (factor))
417	WRT_SYSCTL(sched_min_granularity);
418	WRT_SYSCTL(sched_latency);
419	WRT_SYSCTL(sched_wakeup_granularity);
420	WRT_SYSCTL(sched_shares_ratelimit);
421	#undef WRT_SYSCTL
422
423	return 0;
424	}
425	#endif
426
427	/*
428	* delta /= w
429	*/
430	static inline unsigned long
431	calc_delta_fair(unsigned long delta, struct sched_entity *se)
432	{
433	if (unlikely(se->load.weight != NICE_0_LOAD))
434	delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
435
436	return delta;
437	}
438
439	/*
440	* The idea is to set a period in which each task runs once.
441	*
442	* When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
443	* this period because otherwise the slices get too small.
444	*
445	* p = (nr <= nl) ? l : l*nr/nl
446	*/
447	static u64 __sched_period(unsigned long nr_running)
448	{
449	u64 period = sysctl_sched_latency;
450	unsigned long nr_latency = sched_nr_latency;
451
452	if (unlikely(nr_running > nr_latency)) {
453	period = sysctl_sched_min_granularity;
454	period *= nr_running;
455	}
456
457	return period;
458	}
459
460	/*
461	* We calculate the wall-time slice from the period by taking a part
462	* proportional to the weight.
463	*
464	* s = p*P[w/rw]
465	*/
466	static u64 sched_slice(struct cfs_rq cfs_rq, struct sched_entity se)
467	{
468	u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
469
470	for_each_sched_entity(se) {
471	struct load_weight *load;
472	struct load_weight lw;
473
474	cfs_rq = cfs_rq_of(se);
475	load = &cfs_rq->load;
476
477	if (unlikely(!se->on_rq)) {
478	lw = cfs_rq->load;
479
480	update_load_add(&lw, se->load.weight);
481	load = &lw;
482	}
483	slice = calc_delta_mine(slice, se->load.weight, load);
484	}
485	return slice;
486	}
487
488	/*
489	* We calculate the vruntime slice of a to be inserted task
490	*
491	* vs = s/w
492	*/
493	static u64 sched_vslice(struct cfs_rq cfs_rq, struct sched_entity se)
494	{
495	return calc_delta_fair(sched_slice(cfs_rq, se), se);
496	}
497
498	/*
499	* Update the current task's runtime statistics. Skip current tasks that
500	* are not in our scheduling class.
501	*/
502	static inline void
503	__update_curr(struct cfs_rq cfs_rq, struct sched_entity curr,
504	unsigned long delta_exec)
505	{
506	unsigned long delta_exec_weighted;
507
508	schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
509
510	curr->sum_exec_runtime += delta_exec;
511	schedstat_add(cfs_rq, exec_clock, delta_exec);
512	delta_exec_weighted = calc_delta_fair(delta_exec, curr);
513
514	curr->vruntime += delta_exec_weighted;
515	update_min_vruntime(cfs_rq);
516	}
517
518	static void update_curr(struct cfs_rq *cfs_rq)
519	{
520	struct sched_entity *curr = cfs_rq->curr;
521	u64 now = rq_of(cfs_rq)->clock;
522	unsigned long delta_exec;
523
524	if (unlikely(!curr))
525	return;
526
527	/*
528	* Get the amount of time the current task was running
529	* since the last time we changed load (this cannot
530	* overflow on 32 bits):
531	*/
532	delta_exec = (unsigned long)(now - curr->exec_start);
533	if (!delta_exec)
534	return;
535
536	__update_curr(cfs_rq, curr, delta_exec);
537	curr->exec_start = now;
538
539	if (entity_is_task(curr)) {
540	struct task_struct *curtask = task_of(curr);
541
542	trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
543	cpuacct_charge(curtask, delta_exec);
544	account_group_exec_runtime(curtask, delta_exec);
545	}
546	}
547
548	static inline void
549	update_stats_wait_start(struct cfs_rq cfs_rq, struct sched_entity se)
550	{
551	schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
552	}
553
554	/*
555	* Task is being enqueued - update stats:
556	*/
557	static void update_stats_enqueue(struct cfs_rq cfs_rq, struct sched_entity se)
558	{
559	/*
560	* Are we enqueueing a waiting task? (for current tasks
561	* a dequeue/enqueue event is a NOP)
562	*/
563	if (se != cfs_rq->curr)
564	update_stats_wait_start(cfs_rq, se);
565	}
566
567	static void
568	update_stats_wait_end(struct cfs_rq cfs_rq, struct sched_entity se)
569	{
570	schedstat_set(se->wait_max, max(se->wait_max,
571	rq_of(cfs_rq)->clock - se->wait_start));
572	schedstat_set(se->wait_count, se->wait_count + 1);
573	schedstat_set(se->wait_sum, se->wait_sum +
574	rq_of(cfs_rq)->clock - se->wait_start);
575	#ifdef CONFIG_SCHEDSTATS
576	if (entity_is_task(se)) {
577	trace_sched_stat_wait(task_of(se),
578	rq_of(cfs_rq)->clock - se->wait_start);
579	}
580	#endif
581	schedstat_set(se->wait_start, 0);
582	}
583
584	static inline void
585	update_stats_dequeue(struct cfs_rq cfs_rq, struct sched_entity se)
586	{
587	/*
588	* Mark the end of the wait period if dequeueing a
589	* waiting task:
590	*/
591	if (se != cfs_rq->curr)
592	update_stats_wait_end(cfs_rq, se);
593	}
594
595	/*
596	* We are picking a new current task - update its stats:
597	*/
598	static inline void
599	update_stats_curr_start(struct cfs_rq cfs_rq, struct sched_entity se)
600	{
601	/*
602	* We are starting a new run period:
603	*/
604	se->exec_start = rq_of(cfs_rq)->clock;
605	}
606
607	/**************************************************
608	* Scheduling class queueing methods:
609	*/
610
611	#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
612	static void
613	add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
614	{
615	cfs_rq->task_weight += weight;
616	}
617	#else
618	static inline void
619	add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
620	{
621	}
622	#endif
623
624	static void
625	account_entity_enqueue(struct cfs_rq cfs_rq, struct sched_entity se)
626	{
627	update_load_add(&cfs_rq->load, se->load.weight);
628	if (!parent_entity(se))
629	inc_cpu_load(rq_of(cfs_rq), se->load.weight);
630	if (entity_is_task(se)) {
631	add_cfs_task_weight(cfs_rq, se->load.weight);
632	list_add(&se->group_node, &cfs_rq->tasks);
633	}
634	cfs_rq->nr_running++;
635	se->on_rq = 1;
636	}
637
638	static void
639	account_entity_dequeue(struct cfs_rq cfs_rq, struct sched_entity se)
640	{
641	update_load_sub(&cfs_rq->load, se->load.weight);
642	if (!parent_entity(se))
643	dec_cpu_load(rq_of(cfs_rq), se->load.weight);
644	if (entity_is_task(se)) {
645	add_cfs_task_weight(cfs_rq, -se->load.weight);
646	list_del_init(&se->group_node);
647	}
648	cfs_rq->nr_running--;
649	se->on_rq = 0;
650	}
651
652	static void enqueue_sleeper(struct cfs_rq cfs_rq, struct sched_entity se)
653	{
654	#ifdef CONFIG_SCHEDSTATS
655	struct task_struct *tsk = NULL;
656
657	if (entity_is_task(se))
658	tsk = task_of(se);
659
660	if (se->sleep_start) {
661	u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
662
663	if ((s64)delta < 0)
664	delta = 0;
665
666	if (unlikely(delta > se->sleep_max))
667	se->sleep_max = delta;
668
669	se->sleep_start = 0;
670	se->sum_sleep_runtime += delta;
671
672	if (tsk) {
673	account_scheduler_latency(tsk, delta >> 10, 1);
674	trace_sched_stat_sleep(tsk, delta);
675	}
676	}
677	if (se->block_start) {
678	u64 delta = rq_of(cfs_rq)->clock - se->block_start;
679
680	if ((s64)delta < 0)
681	delta = 0;
682
683	if (unlikely(delta > se->block_max))
684	se->block_max = delta;
685
686	se->block_start = 0;
687	se->sum_sleep_runtime += delta;
688
689	if (tsk) {
690	if (tsk->in_iowait) {
691	se->iowait_sum += delta;
692	se->iowait_count++;
693	trace_sched_stat_iowait(tsk, delta);
694	}
695
696	/*
697	* Blocking time is in units of nanosecs, so shift by
698	* 20 to get a milliseconds-range estimation of the
699	* amount of time that the task spent sleeping:
700	*/
701	if (unlikely(prof_on == SLEEP_PROFILING)) {
702	profile_hits(SLEEP_PROFILING,
703	(void *)get_wchan(tsk),
704	delta >> 20);
705	}
706	account_scheduler_latency(tsk, delta >> 10, 0);
707	}
708	}
709	#endif
710	}
711
712	static void check_spread(struct cfs_rq cfs_rq, struct sched_entity se)
713	{
714	#ifdef CONFIG_SCHED_DEBUG
715	s64 d = se->vruntime - cfs_rq->min_vruntime;
716
717	if (d < 0)
718	d = -d;
719
720	if (d > 3*sysctl_sched_latency)
721	schedstat_inc(cfs_rq, nr_spread_over);
722	#endif
723	}
724
725	static void
726	place_entity(struct cfs_rq cfs_rq, struct sched_entity se, int initial)
727	{
728	u64 vruntime = cfs_rq->min_vruntime;
729
730	/*
731	* The 'current' period is already promised to the current tasks,
732	* however the extra weight of the new task will slow them down a
733	* little, place the new task so that it fits in the slot that
734	* stays open at the end.
735	*/
736	if (initial && sched_feat(START_DEBIT))
737	vruntime += sched_vslice(cfs_rq, se);
738
739	/* sleeps up to a single latency don't count. */
740	if (!initial && sched_feat(FAIR_SLEEPERS)) {
741	unsigned long thresh = sysctl_sched_latency;
742
743	/*
744	* Convert the sleeper threshold into virtual time.
745	* SCHED_IDLE is a special sub-class. We care about
746	* fairness only relative to other SCHED_IDLE tasks,
747	* all of which have the same weight.
748	*/
749	if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) \|\|
750	task_of(se)->policy != SCHED_IDLE))
751	thresh = calc_delta_fair(thresh, se);
752
753	/*
754	* Halve their sleep time's effect, to allow
755	* for a gentler effect of sleepers:
756	*/
757	if (sched_feat(GENTLE_FAIR_SLEEPERS))
758	thresh >>= 1;
759
760	vruntime -= thresh;
761	}
762
763	/* ensure we never gain time by being placed backwards. */
764	vruntime = max_vruntime(se->vruntime, vruntime);
765
766	se->vruntime = vruntime;
767	}
768
769	#define ENQUEUE_WAKEUP 1
770	#define ENQUEUE_MIGRATE 2
771
772	static void
773	enqueue_entity(struct cfs_rq cfs_rq, struct sched_entity se, int flags)
774	{
775	/*
776	* Update the normalized vruntime before updating min_vruntime
777	* through callig update_curr().
778	*/
779	if (!(flags & ENQUEUE_WAKEUP) \|\| (flags & ENQUEUE_MIGRATE))
780	se->vruntime += cfs_rq->min_vruntime;
781
782	/*
783	* Update run-time statistics of the 'current'.
784	*/
785	update_curr(cfs_rq);
786	account_entity_enqueue(cfs_rq, se);
787
788	if (flags & ENQUEUE_WAKEUP) {
789	place_entity(cfs_rq, se, 0);
790	enqueue_sleeper(cfs_rq, se);
791	}
792
793	update_stats_enqueue(cfs_rq, se);
794	check_spread(cfs_rq, se);
795	if (se != cfs_rq->curr)
796	__enqueue_entity(cfs_rq, se);
797	}
798
799	static void __clear_buddies(struct cfs_rq cfs_rq, struct sched_entity se)
800	{
801	if (!se \|\| cfs_rq->last == se)
802	cfs_rq->last = NULL;
803
804	if (!se \|\| cfs_rq->next == se)
805	cfs_rq->next = NULL;
806	}
807
808	static void clear_buddies(struct cfs_rq cfs_rq, struct sched_entity se)
809	{
810	for_each_sched_entity(se)
811	__clear_buddies(cfs_rq_of(se), se);
812	}
813
814	static void
815	dequeue_entity(struct cfs_rq cfs_rq, struct sched_entity se, int sleep)
816	{
817	/*
818	* Update run-time statistics of the 'current'.
819	*/
820	update_curr(cfs_rq);
821
822	update_stats_dequeue(cfs_rq, se);
823	if (sleep) {
824	#ifdef CONFIG_SCHEDSTATS
825	if (entity_is_task(se)) {
826	struct task_struct *tsk = task_of(se);
827
828	if (tsk->state & TASK_INTERRUPTIBLE)
829	se->sleep_start = rq_of(cfs_rq)->clock;
830	if (tsk->state & TASK_UNINTERRUPTIBLE)
831	se->block_start = rq_of(cfs_rq)->clock;
832	}
833	#endif
834	}
835
836	clear_buddies(cfs_rq, se);
837
838	if (se != cfs_rq->curr)
839	__dequeue_entity(cfs_rq, se);
840	account_entity_dequeue(cfs_rq, se);
841	update_min_vruntime(cfs_rq);
842
843	/*
844	* Normalize the entity after updating the min_vruntime because the
845	* update can refer to the ->curr item and we need to reflect this
846	* movement in our normalized position.
847	*/
848	if (!sleep)
849	se->vruntime -= cfs_rq->min_vruntime;
850	}
851
852	/*
853	* Preempt the current task with a newly woken task if needed:
854	*/
855	static void
856	check_preempt_tick(struct cfs_rq cfs_rq, struct sched_entity curr)
857	{
858	unsigned long ideal_runtime, delta_exec;
859
860	ideal_runtime = sched_slice(cfs_rq, curr);
861	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
862	if (delta_exec > ideal_runtime) {
863	resched_task(rq_of(cfs_rq)->curr);
864	/*
865	* The current task ran long enough, ensure it doesn't get
866	* re-elected due to buddy favours.
867	*/
868	clear_buddies(cfs_rq, curr);
869	return;
870	}
871
872	/*
873	* Ensure that a task that missed wakeup preemption by a
874	* narrow margin doesn't have to wait for a full slice.
875	* This also mitigates buddy induced latencies under load.
876	*/
877	if (!sched_feat(WAKEUP_PREEMPT))
878	return;
879
880	if (delta_exec < sysctl_sched_min_granularity)
881	return;
882
883	if (cfs_rq->nr_running > 1) {
884	struct sched_entity *se = __pick_next_entity(cfs_rq);
885	s64 delta = curr->vruntime - se->vruntime;
886
887	if (delta > ideal_runtime)
888	resched_task(rq_of(cfs_rq)->curr);
889	}
890	}
891
892	static void
893	set_next_entity(struct cfs_rq cfs_rq, struct sched_entity se)
894	{
895	/* 'current' is not kept within the tree. */
896	if (se->on_rq) {
897	/*
898	* Any task has to be enqueued before it get to execute on
899	* a CPU. So account for the time it spent waiting on the
900	* runqueue.
901	*/
902	update_stats_wait_end(cfs_rq, se);
903	__dequeue_entity(cfs_rq, se);
904	}
905
906	update_stats_curr_start(cfs_rq, se);
907	cfs_rq->curr = se;
908	#ifdef CONFIG_SCHEDSTATS
909	/*
910	* Track our maximum slice length, if the CPU's load is at
911	* least twice that of our own weight (i.e. dont track it
912	* when there are only lesser-weight tasks around):
913	*/
914	if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
915	se->slice_max = max(se->slice_max,
916	se->sum_exec_runtime - se->prev_sum_exec_runtime);
917	}
918	#endif
919	se->prev_sum_exec_runtime = se->sum_exec_runtime;
920	}
921
922	static int
923	wakeup_preempt_entity(struct sched_entity curr, struct sched_entity se);
924
925	static struct sched_entity pick_next_entity(struct cfs_rq cfs_rq)
926	{
927	struct sched_entity *se = __pick_next_entity(cfs_rq);
928	struct sched_entity *left = se;
929
930	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
931	se = cfs_rq->next;
932
933	/*
934	* Prefer last buddy, try to return the CPU to a preempted task.
935	*/
936	if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
937	se = cfs_rq->last;
938
939	clear_buddies(cfs_rq, se);
940
941	return se;
942	}
943
944	static void put_prev_entity(struct cfs_rq cfs_rq, struct sched_entity prev)
945	{
946	/*
947	* If still on the runqueue then deactivate_task()
948	* was not called and update_curr() has to be done:
949	*/
950	if (prev->on_rq)
951	update_curr(cfs_rq);
952
953	check_spread(cfs_rq, prev);
954	if (prev->on_rq) {
955	update_stats_wait_start(cfs_rq, prev);
956	/* Put 'current' back into the tree. */
957	__enqueue_entity(cfs_rq, prev);
958	}
959	cfs_rq->curr = NULL;
960	}
961
962	static void
963	entity_tick(struct cfs_rq cfs_rq, struct sched_entity curr, int queued)
964	{
965	/*
966	* Update run-time statistics of the 'current'.
967	*/
968	update_curr(cfs_rq);
969
970	#ifdef CONFIG_SCHED_HRTICK
971	/*
972	* queued ticks are scheduled to match the slice, so don't bother
973	* validating it and just reschedule.
974	*/
975	if (queued) {
976	resched_task(rq_of(cfs_rq)->curr);
977	return;
978	}
979	/*
980	* don't let the period tick interfere with the hrtick preemption
981	*/
982	if (!sched_feat(DOUBLE_TICK) &&
983	hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
984	return;
985	#endif
986
987	if (cfs_rq->nr_running > 1 \|\| !sched_feat(WAKEUP_PREEMPT))
988	check_preempt_tick(cfs_rq, curr);
989	}
990
991	/**************************************************
992	* CFS operations on tasks:
993	*/
994
995	#ifdef CONFIG_SCHED_HRTICK
996	static void hrtick_start_fair(struct rq rq, struct task_struct p)
997	{
998	struct sched_entity *se = &p->se;
999	struct cfs_rq *cfs_rq = cfs_rq_of(se);
1000
1001	WARN_ON(task_rq(p) != rq);
1002
1003	if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1004	u64 slice = sched_slice(cfs_rq, se);
1005	u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1006	s64 delta = slice - ran;
1007
1008	if (delta < 0) {
1009	if (rq->curr == p)
1010	resched_task(p);
1011	return;
1012	}
1013
1014	/*
1015	* Don't schedule slices shorter than 10000ns, that just
1016	* doesn't make sense. Rely on vruntime for fairness.
1017	*/
1018	if (rq->curr != p)
1019	delta = max_t(s64, 10000LL, delta);
1020
1021	hrtick_start(rq, delta);
1022	}
1023	}
1024
1025	/*
1026	* called from enqueue/dequeue and updates the hrtick when the
1027	* current task is from our class and nr_running is low enough
1028	* to matter.
1029	*/
1030	static void hrtick_update(struct rq *rq)
1031	{
1032	struct task_struct *curr = rq->curr;
1033
1034	if (curr->sched_class != &fair_sched_class)
1035	return;
1036
1037	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1038	hrtick_start_fair(rq, curr);
1039	}
1040	#else /* !CONFIG_SCHED_HRTICK */
1041	static inline void
1042	hrtick_start_fair(struct rq rq, struct task_struct p)
1043	{
1044	}
1045
1046	static inline void hrtick_update(struct rq *rq)
1047	{
1048	}
1049	#endif
1050
1051	/*
1052	* The enqueue_task method is called before nr_running is
1053	* increased. Here we update the fair scheduling stats and
1054	* then put the task into the rbtree:
1055	*/
1056	static void
1057	enqueue_task_fair(struct rq rq, struct task_struct p, int wakeup, bool head)
1058	{
1059	struct cfs_rq *cfs_rq;
1060	struct sched_entity *se = &p->se;
1061	int flags = 0;
1062
1063	if (wakeup)
1064	flags \|= ENQUEUE_WAKEUP;
1065	if (p->state == TASK_WAKING)
1066	flags \|= ENQUEUE_MIGRATE;
1067
1068	for_each_sched_entity(se) {
1069	if (se->on_rq)
1070	break;
1071	cfs_rq = cfs_rq_of(se);
1072	enqueue_entity(cfs_rq, se, flags);
1073	flags = ENQUEUE_WAKEUP;
1074	}
1075
1076	hrtick_update(rq);
1077	}
1078
1079	/*
1080	* The dequeue_task method is called before nr_running is
1081	* decreased. We remove the task from the rbtree and
1082	* update the fair scheduling stats:
1083	*/
1084	static void dequeue_task_fair(struct rq rq, struct task_struct p, int sleep)
1085	{
1086	struct cfs_rq *cfs_rq;
1087	struct sched_entity *se = &p->se;
1088
1089	for_each_sched_entity(se) {
1090	cfs_rq = cfs_rq_of(se);
1091	dequeue_entity(cfs_rq, se, sleep);
1092	/* Don't dequeue parent if it has other entities besides us */
1093	if (cfs_rq->load.weight)
1094	break;
1095	sleep = 1;
1096	}
1097
1098	hrtick_update(rq);
1099	}
1100
1101	/*
1102	* sched_yield() support is very simple - we dequeue and enqueue.
1103	*
1104	* If compat_yield is turned on then we requeue to the end of the tree.
1105	*/
1106	static void yield_task_fair(struct rq *rq)
1107	{
1108	struct task_struct *curr = rq->curr;
1109	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1110	struct sched_entity rightmost, se = &curr->se;
1111
1112	/*
1113	* Are we the only task in the tree?
1114	*/
1115	if (unlikely(cfs_rq->nr_running == 1))
1116	return;
1117
1118	clear_buddies(cfs_rq, se);
1119
1120	if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1121	update_rq_clock(rq);
1122	/*
1123	* Update run-time statistics of the 'current'.
1124	*/
1125	update_curr(cfs_rq);
1126
1127	return;
1128	}
1129	/*
1130	* Find the rightmost entry in the rbtree:
1131	*/
1132	rightmost = __pick_last_entity(cfs_rq);
1133	/*
1134	* Already in the rightmost position?
1135	*/
1136	if (unlikely(!rightmost \|\| entity_before(rightmost, se)))
1137	return;
1138
1139	/*
1140	* Minimally necessary key value to be last in the tree:
1141	* Upon rescheduling, sched_class::put_prev_task() will place
1142	* 'current' within the tree based on its new key value.
1143	*/
1144	se->vruntime = rightmost->vruntime + 1;
1145	}
1146
1147	#ifdef CONFIG_SMP
1148
1149	static void task_waking_fair(struct rq rq, struct task_struct p)
1150	{
1151	struct sched_entity *se = &p->se;
1152	struct cfs_rq *cfs_rq = cfs_rq_of(se);
1153
1154	se->vruntime -= cfs_rq->min_vruntime;
1155	}
1156
1157	#ifdef CONFIG_FAIR_GROUP_SCHED
1158	/*
1159	* effective_load() calculates the load change as seen from the root_task_group
1160	*
1161	* Adding load to a group doesn't make a group heavier, but can cause movement
1162	* of group shares between cpus. Assuming the shares were perfectly aligned one
1163	* can calculate the shift in shares.
1164	*
1165	* The problem is that perfectly aligning the shares is rather expensive, hence
1166	* we try to avoid doing that too often - see update_shares(), which ratelimits
1167	* this change.
1168	*
1169	* We compensate this by not only taking the current delta into account, but
1170	* also considering the delta between when the shares were last adjusted and
1171	* now.
1172	*
1173	* We still saw a performance dip, some tracing learned us that between
1174	* cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1175	* significantly. Therefore try to bias the error in direction of failing
1176	* the affine wakeup.
1177	*
1178	*/
1179	static long effective_load(struct task_group *tg, int cpu,
1180	long wl, long wg)
1181	{
1182	struct sched_entity *se = tg->se[cpu];
1183
1184	if (!tg->parent)
1185	return wl;
1186
1187	/*
1188	* By not taking the decrease of shares on the other cpu into
1189	* account our error leans towards reducing the affine wakeups.
1190	*/
1191	if (!wl && sched_feat(ASYM_EFF_LOAD))
1192	return wl;
1193
1194	for_each_sched_entity(se) {
1195	long S, rw, s, a, b;
1196	long more_w;
1197
1198	/*
1199	* Instead of using this increment, also add the difference
1200	* between when the shares were last updated and now.
1201	*/
1202	more_w = se->my_q->load.weight - se->my_q->rq_weight;
1203	wl += more_w;
1204	wg += more_w;
1205
1206	S = se->my_q->tg->shares;
1207	s = se->my_q->shares;
1208	rw = se->my_q->rq_weight;
1209
1210	a = S*(rw + wl);
1211	b = Srw + swg;
1212
1213	wl = s*(a-b);
1214
1215	if (likely(b))
1216	wl /= b;
1217
1218	/*
1219	* Assume the group is already running and will
1220	* thus already be accounted for in the weight.
1221	*
1222	* That is, moving shares between CPUs, does not
1223	* alter the group weight.
1224	*/
1225	wg = 0;
1226	}
1227
1228	return wl;
1229	}
1230
1231	#else
1232
1233	static inline unsigned long effective_load(struct task_group *tg, int cpu,
1234	unsigned long wl, unsigned long wg)
1235	{
1236	return wl;
1237	}
1238
1239	#endif
1240
1241	static int wake_affine(struct sched_domain sd, struct task_struct p, int sync)
1242	{
1243	struct task_struct *curr = current;
1244	unsigned long this_load, load;
1245	int idx, this_cpu, prev_cpu;
1246	unsigned long tl_per_task;
1247	unsigned int imbalance;
1248	struct task_group *tg;
1249	unsigned long weight;
1250	int balanced;
1251
1252	idx = sd->wake_idx;
1253	this_cpu = smp_processor_id();
1254	prev_cpu = task_cpu(p);
1255	load = source_load(prev_cpu, idx);
1256	this_load = target_load(this_cpu, idx);
1257
1258	if (sync) {
1259	if (sched_feat(SYNC_LESS) &&
1260	(curr->se.avg_overlap > sysctl_sched_migration_cost \|\|
1261	p->se.avg_overlap > sysctl_sched_migration_cost))
1262	sync = 0;
1263	} else {
1264	if (sched_feat(SYNC_MORE) &&
1265	(curr->se.avg_overlap < sysctl_sched_migration_cost &&
1266	p->se.avg_overlap < sysctl_sched_migration_cost))
1267	sync = 1;
1268	}
1269
1270	/*
1271	* If sync wakeup then subtract the (maximum possible)
1272	* effect of the currently running task from the load
1273	* of the current CPU:
1274	*/
1275	if (sync) {
1276	tg = task_group(current);
1277	weight = current->se.load.weight;
1278
1279	this_load += effective_load(tg, this_cpu, -weight, -weight);
1280	load += effective_load(tg, prev_cpu, 0, -weight);
1281	}
1282
1283	tg = task_group(p);
1284	weight = p->se.load.weight;
1285
1286	imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1287
1288	/*
1289	* In low-load situations, where prev_cpu is idle and this_cpu is idle
1290	* due to the sync cause above having dropped this_load to 0, we'll
1291	* always have an imbalance, but there's really nothing you can do
1292	* about that, so that's good too.
1293	*
1294	* Otherwise check if either cpus are near enough in load to allow this
1295	* task to be woken on this_cpu.
1296	*/
1297	balanced = !this_load \|\|
1298	100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1299	imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1300
1301	/*
1302	* If the currently running task will sleep within
1303	* a reasonable amount of time then attract this newly
1304	* woken task:
1305	*/
1306	if (sync && balanced)
1307	return 1;
1308
1309	schedstat_inc(p, se.nr_wakeups_affine_attempts);
1310	tl_per_task = cpu_avg_load_per_task(this_cpu);
1311
1312	if (balanced \|\|
1313	(this_load <= load &&
1314	this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1315	/*
1316	* This domain has SD_WAKE_AFFINE and
1317	* p is cache cold in this domain, and
1318	* there is no bad imbalance.
1319	*/
1320	schedstat_inc(sd, ttwu_move_affine);
1321	schedstat_inc(p, se.nr_wakeups_affine);
1322
1323	return 1;
1324	}
1325	return 0;
1326	}
1327
1328	/*
1329	* find_idlest_group finds and returns the least busy CPU group within the
1330	* domain.
1331	*/
1332	static struct sched_group *
1333	find_idlest_group(struct sched_domain sd, struct task_struct p,
1334	int this_cpu, int load_idx)
1335	{
1336	struct sched_group idlest = NULL, this = NULL, *group = sd->groups;
1337	unsigned long min_load = ULONG_MAX, this_load = 0;
1338	int imbalance = 100 + (sd->imbalance_pct-100)/2;
1339
1340	do {
1341	unsigned long load, avg_load;
1342	int local_group;
1343	int i;
1344
1345	/* Skip over this group if it has no CPUs allowed */
1346	if (!cpumask_intersects(sched_group_cpus(group),
1347	&p->cpus_allowed))
1348	continue;
1349
1350	local_group = cpumask_test_cpu(this_cpu,
1351	sched_group_cpus(group));
1352
1353	/* Tally up the load of all CPUs in the group */
1354	avg_load = 0;
1355
1356	for_each_cpu(i, sched_group_cpus(group)) {
1357	/* Bias balancing toward cpus of our domain */
1358	if (local_group)
1359	load = source_load(i, load_idx);
1360	else
1361	load = target_load(i, load_idx);
1362
1363	avg_load += load;
1364	}
1365
1366	/* Adjust by relative CPU power of the group */
1367	avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1368
1369	if (local_group) {
1370	this_load = avg_load;
1371	this = group;
1372	} else if (avg_load < min_load) {
1373	min_load = avg_load;
1374	idlest = group;
1375	}
1376	} while (group = group->next, group != sd->groups);
1377
1378	if (!idlest \|\| 100this_load < imbalancemin_load)
1379	return NULL;
1380	return idlest;
1381	}
1382
1383	/*
1384	* find_idlest_cpu - find the idlest cpu among the cpus in group.
1385	*/
1386	static int
1387	find_idlest_cpu(struct sched_group group, struct task_struct p, int this_cpu)
1388	{
1389	unsigned long load, min_load = ULONG_MAX;
1390	int idlest = -1;
1391	int i;
1392
1393	/* Traverse only the allowed CPUs */
1394	for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1395	load = weighted_cpuload(i);
1396
1397	if (load < min_load \|\| (load == min_load && i == this_cpu)) {
1398	min_load = load;
1399	idlest = i;
1400	}
1401	}
1402
1403	return idlest;
1404	}
1405
1406	/*
1407	* Try and locate an idle CPU in the sched_domain.
1408	*/
1409	static int
1410	select_idle_sibling(struct task_struct p, struct sched_domain sd, int target)
1411	{
1412	int cpu = smp_processor_id();
1413	int prev_cpu = task_cpu(p);
1414	int i;
1415
1416	/*
1417	* If this domain spans both cpu and prev_cpu (see the SD_WAKE_AFFINE
1418	* test in select_task_rq_fair) and the prev_cpu is idle then that's
1419	* always a better target than the current cpu.
1420	*/
1421	if (target == cpu && !cpu_rq(prev_cpu)->cfs.nr_running)
1422	return prev_cpu;
1423
1424	/*
1425	* Otherwise, iterate the domain and find an elegible idle cpu.
1426	*/
1427	for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1428	if (!cpu_rq(i)->cfs.nr_running) {
1429	target = i;
1430	break;
1431	}
1432	}
1433
1434	return target;
1435	}
1436
1437	/*
1438	* sched_balance_self: balance the current task (running on cpu) in domains
1439	* that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1440	* SD_BALANCE_EXEC.
1441	*
1442	* Balance, ie. select the least loaded group.
1443	*
1444	* Returns the target CPU number, or the same CPU if no balancing is needed.
1445	*
1446	* preempt must be disabled.
1447	*/
1448	static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1449	{
1450	struct sched_domain tmp, affine_sd = NULL, *sd = NULL;
1451	int cpu = smp_processor_id();
1452	int prev_cpu = task_cpu(p);
1453	int new_cpu = cpu;
1454	int want_affine = 0;
1455	int want_sd = 1;
1456	int sync = wake_flags & WF_SYNC;
1457
1458	if (sd_flag & SD_BALANCE_WAKE) {
1459	if (sched_feat(AFFINE_WAKEUPS) &&
1460	cpumask_test_cpu(cpu, &p->cpus_allowed))
1461	want_affine = 1;
1462	new_cpu = prev_cpu;
1463	}
1464
1465	for_each_domain(cpu, tmp) {
1466	if (!(tmp->flags & SD_LOAD_BALANCE))
1467	continue;
1468
1469	/*
1470	* If power savings logic is enabled for a domain, see if we
1471	* are not overloaded, if so, don't balance wider.
1472	*/
1473	if (tmp->flags & (SD_POWERSAVINGS_BALANCE\|SD_PREFER_LOCAL)) {
1474	unsigned long power = 0;
1475	unsigned long nr_running = 0;
1476	unsigned long capacity;
1477	int i;
1478
1479	for_each_cpu(i, sched_domain_span(tmp)) {
1480	power += power_of(i);
1481	nr_running += cpu_rq(i)->cfs.nr_running;
1482	}
1483
1484	capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1485
1486	if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1487	nr_running /= 2;
1488
1489	if (nr_running < capacity)
1490	want_sd = 0;
1491	}
1492
1493	/*
1494	* While iterating the domains looking for a spanning
1495	* WAKE_AFFINE domain, adjust the affine target to any idle cpu
1496	* in cache sharing domains along the way.
1497	*/
1498	if (want_affine) {
1499	int target = -1;
1500
1501	/*
1502	* If both cpu and prev_cpu are part of this domain,
1503	* cpu is a valid SD_WAKE_AFFINE target.
1504	*/
1505	if (cpumask_test_cpu(prev_cpu, sched_domain_span(tmp)))
1506	target = cpu;
1507
1508	/*
1509	* If there's an idle sibling in this domain, make that
1510	* the wake_affine target instead of the current cpu.
1511	*/
1512	if (tmp->flags & SD_SHARE_PKG_RESOURCES)
1513	target = select_idle_sibling(p, tmp, target);
1514
1515	if (target >= 0) {
1516	if (tmp->flags & SD_WAKE_AFFINE) {
1517	affine_sd = tmp;
1518	want_affine = 0;
1519	}
1520	cpu = target;
1521	}
1522	}
1523
1524	if (!want_sd && !want_affine)
1525	break;
1526
1527	if (!(tmp->flags & sd_flag))
1528	continue;
1529
1530	if (want_sd)
1531	sd = tmp;
1532	}
1533
1534	if (sched_feat(LB_SHARES_UPDATE)) {
1535	/*
1536	* Pick the largest domain to update shares over
1537	*/
1538	tmp = sd;
1539	if (affine_sd && (!tmp \|\|
1540	cpumask_weight(sched_domain_span(affine_sd)) >
1541	cpumask_weight(sched_domain_span(sd))))
1542	tmp = affine_sd;
1543
1544	if (tmp)
1545	update_shares(tmp);
1546	}
1547
1548	if (affine_sd && wake_affine(affine_sd, p, sync))
1549	return cpu;
1550
1551	while (sd) {
1552	int load_idx = sd->forkexec_idx;
1553	struct sched_group *group;
1554	int weight;
1555
1556	if (!(sd->flags & sd_flag)) {
1557	sd = sd->child;
1558	continue;
1559	}
1560
1561	if (sd_flag & SD_BALANCE_WAKE)
1562	load_idx = sd->wake_idx;
1563
1564	group = find_idlest_group(sd, p, cpu, load_idx);
1565	if (!group) {
1566	sd = sd->child;
1567	continue;
1568	}
1569
1570	new_cpu = find_idlest_cpu(group, p, cpu);
1571	if (new_cpu == -1 \|\| new_cpu == cpu) {
1572	/* Now try balancing at a lower domain level of cpu */
1573	sd = sd->child;
1574	continue;
1575	}
1576
1577	/* Now try balancing at a lower domain level of new_cpu */
1578	cpu = new_cpu;
1579	weight = cpumask_weight(sched_domain_span(sd));
1580	sd = NULL;
1581	for_each_domain(cpu, tmp) {
1582	if (weight <= cpumask_weight(sched_domain_span(tmp)))
1583	break;
1584	if (tmp->flags & sd_flag)
1585	sd = tmp;
1586	}
1587	/* while loop will break here if sd == NULL */
1588	}
1589
1590	return new_cpu;
1591	}
1592	#endif /* CONFIG_SMP */
1593
1594	/*
1595	* Adaptive granularity
1596	*
1597	* se->avg_wakeup gives the average time a task runs until it does a wakeup,
1598	* with the limit of wakeup_gran -- when it never does a wakeup.
1599	*
1600	* So the smaller avg_wakeup is the faster we want this task to preempt,
1601	* but we don't want to treat the preemptee unfairly and therefore allow it
1602	* to run for at least the amount of time we'd like to run.
1603	*
1604	* NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1605	*
1606	* NOTE: we use *nr_running to scale with load, this nicely matches the
1607	* degrading latency on load.
1608	*/
1609	static unsigned long
1610	adaptive_gran(struct sched_entity curr, struct sched_entity se)
1611	{
1612	u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1613	u64 expected_wakeup = 2se->avg_wakeup cfs_rq_of(se)->nr_running;
1614	u64 gran = 0;
1615
1616	if (this_run < expected_wakeup)
1617	gran = expected_wakeup - this_run;
1618
1619	return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1620	}
1621
1622	static unsigned long
1623	wakeup_gran(struct sched_entity curr, struct sched_entity se)
1624	{
1625	unsigned long gran = sysctl_sched_wakeup_granularity;
1626
1627	if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1628	gran = adaptive_gran(curr, se);
1629
1630	/*
1631	* Since its curr running now, convert the gran from real-time
1632	* to virtual-time in his units.
1633	*/
1634	if (sched_feat(ASYM_GRAN)) {
1635	/*
1636	* By using 'se' instead of 'curr' we penalize light tasks, so
1637	* they get preempted easier. That is, if 'se' < 'curr' then
1638	* the resulting gran will be larger, therefore penalizing the
1639	* lighter, if otoh 'se' > 'curr' then the resulting gran will
1640	* be smaller, again penalizing the lighter task.
1641	*
1642	* This is especially important for buddies when the leftmost
1643	* task is higher priority than the buddy.
1644	*/
1645	if (unlikely(se->load.weight != NICE_0_LOAD))
1646	gran = calc_delta_fair(gran, se);
1647	} else {
1648	if (unlikely(curr->load.weight != NICE_0_LOAD))
1649	gran = calc_delta_fair(gran, curr);
1650	}
1651
1652	return gran;
1653	}
1654
1655	/*
1656	* Should 'se' preempt 'curr'.
1657	*
1658	* \|s1
1659	* \|s2
1660	* \|s3
1661	* g
1662	* \|<--->\|c
1663	*
1664	* w(c, s1) = -1
1665	* w(c, s2) = 0
1666	* w(c, s3) = 1
1667	*
1668	*/
1669	static int
1670	wakeup_preempt_entity(struct sched_entity curr, struct sched_entity se)
1671	{
1672	s64 gran, vdiff = curr->vruntime - se->vruntime;
1673
1674	if (vdiff <= 0)
1675	return -1;
1676
1677	gran = wakeup_gran(curr, se);
1678	if (vdiff > gran)
1679	return 1;
1680
1681	return 0;
1682	}
1683
1684	static void set_last_buddy(struct sched_entity *se)
1685	{
1686	if (likely(task_of(se)->policy != SCHED_IDLE)) {
1687	for_each_sched_entity(se)
1688	cfs_rq_of(se)->last = se;
1689	}
1690	}
1691
1692	static void set_next_buddy(struct sched_entity *se)
1693	{
1694	if (likely(task_of(se)->policy != SCHED_IDLE)) {
1695	for_each_sched_entity(se)
1696	cfs_rq_of(se)->next = se;
1697	}
1698	}
1699
1700	/*
1701	* Preempt the current task with a newly woken task if needed:
1702	*/
1703	static void check_preempt_wakeup(struct rq rq, struct task_struct p, int wake_flags)
1704	{
1705	struct task_struct *curr = rq->curr;
1706	struct sched_entity se = &curr->se, pse = &p->se;
1707	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1708	int sync = wake_flags & WF_SYNC;
1709	int scale = cfs_rq->nr_running >= sched_nr_latency;
1710
1711	if (unlikely(rt_prio(p->prio)))
1712	goto preempt;
1713
1714	if (unlikely(p->sched_class != &fair_sched_class))
1715	return;
1716
1717	if (unlikely(se == pse))
1718	return;
1719
1720	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1721	set_next_buddy(pse);
1722
1723	/*
1724	* We can come here with TIF_NEED_RESCHED already set from new task
1725	* wake up path.
1726	*/
1727	if (test_tsk_need_resched(curr))
1728	return;
1729
1730	/*
1731	* Batch and idle tasks do not preempt (their preemption is driven by
1732	* the tick):
1733	*/
1734	if (unlikely(p->policy != SCHED_NORMAL))
1735	return;
1736
1737	/* Idle tasks are by definition preempted by everybody. */
1738	if (unlikely(curr->policy == SCHED_IDLE))
1739	goto preempt;
1740
1741	if (sched_feat(WAKEUP_SYNC) && sync)
1742	goto preempt;
1743
1744	if (sched_feat(WAKEUP_OVERLAP) &&
1745	se->avg_overlap < sysctl_sched_migration_cost &&
1746	pse->avg_overlap < sysctl_sched_migration_cost)
1747	goto preempt;
1748
1749	if (!sched_feat(WAKEUP_PREEMPT))
1750	return;
1751
1752	update_curr(cfs_rq);
1753	find_matching_se(&se, &pse);
1754	BUG_ON(!pse);
1755	if (wakeup_preempt_entity(se, pse) == 1)
1756	goto preempt;
1757
1758	return;
1759
1760	preempt:
1761	resched_task(curr);
1762	/*
1763	* Only set the backward buddy when the current task is still
1764	* on the rq. This can happen when a wakeup gets interleaved
1765	* with schedule on the ->pre_schedule() or idle_balance()
1766	* point, either of which can * drop the rq lock.
1767	*
1768	* Also, during early boot the idle thread is in the fair class,
1769	* for obvious reasons its a bad idea to schedule back to it.
1770	*/
1771	if (unlikely(!se->on_rq \|\| curr == rq->idle))
1772	return;
1773
1774	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1775	set_last_buddy(se);
1776	}
1777
1778	static struct task_struct pick_next_task_fair(struct rq rq)
1779	{
1780	struct task_struct *p;
1781	struct cfs_rq *cfs_rq = &rq->cfs;
1782	struct sched_entity *se;
1783
1784	if (!cfs_rq->nr_running)
1785	return NULL;
1786
1787	do {
1788	se = pick_next_entity(cfs_rq);
1789	set_next_entity(cfs_rq, se);
1790	cfs_rq = group_cfs_rq(se);
1791	} while (cfs_rq);
1792
1793	p = task_of(se);
1794	hrtick_start_fair(rq, p);
1795
1796	return p;
1797	}
1798
1799	/*
1800	* Account for a descheduled task:
1801	*/
1802	static void put_prev_task_fair(struct rq rq, struct task_struct prev)
1803	{
1804	struct sched_entity *se = &prev->se;
1805	struct cfs_rq *cfs_rq;
1806
1807	for_each_sched_entity(se) {
1808	cfs_rq = cfs_rq_of(se);
1809	put_prev_entity(cfs_rq, se);
1810	}
1811	}
1812
1813	#ifdef CONFIG_SMP
1814	/**************************************************
1815	* Fair scheduling class load-balancing methods:
1816	*/
1817
1818	/*
1819	* pull_task - move a task from a remote runqueue to the local runqueue.
1820	* Both runqueues must be locked.
1821	*/
1822	static void pull_task(struct rq src_rq, struct task_struct p,
1823	struct rq *this_rq, int this_cpu)
1824	{
1825	deactivate_task(src_rq, p, 0);
1826	set_task_cpu(p, this_cpu);
1827	activate_task(this_rq, p, 0);
1828	check_preempt_curr(this_rq, p, 0);
1829	}
1830
1831	/*
1832	* can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1833	*/
1834	static
1835	int can_migrate_task(struct task_struct p, struct rq rq, int this_cpu,
1836	struct sched_domain *sd, enum cpu_idle_type idle,
1837	int *all_pinned)
1838	{
1839	int tsk_cache_hot = 0;
1840	/*
1841	* We do not migrate tasks that are:
1842	* 1) running (obviously), or
1843	* 2) cannot be migrated to this CPU due to cpus_allowed, or
1844	* 3) are cache-hot on their current CPU.
1845	*/
1846	if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1847	schedstat_inc(p, se.nr_failed_migrations_affine);
1848	return 0;
1849	}
1850	*all_pinned = 0;
1851
1852	if (task_running(rq, p)) {
1853	schedstat_inc(p, se.nr_failed_migrations_running);
1854	return 0;
1855	}
1856
1857	/*
1858	* Aggressive migration if:
1859	* 1) task is cache cold, or
1860	* 2) too many balance attempts have failed.
1861	*/
1862
1863	tsk_cache_hot = task_hot(p, rq->clock, sd);
1864	if (!tsk_cache_hot \|\|
1865	sd->nr_balance_failed > sd->cache_nice_tries) {
1866	#ifdef CONFIG_SCHEDSTATS
1867	if (tsk_cache_hot) {
1868	schedstat_inc(sd, lb_hot_gained[idle]);
1869	schedstat_inc(p, se.nr_forced_migrations);
1870	}
1871	#endif
1872	return 1;
1873	}
1874
1875	if (tsk_cache_hot) {
1876	schedstat_inc(p, se.nr_failed_migrations_hot);
1877	return 0;
1878	}
1879	return 1;
1880	}
1881
1882	/*
1883	* move_one_task tries to move exactly one task from busiest to this_rq, as
1884	* part of active balancing operations within "domain".
1885	* Returns 1 if successful and 0 otherwise.
1886	*
1887	* Called with both runqueues locked.
1888	*/
1889	static int
1890	move_one_task(struct rq this_rq, int this_cpu, struct rq busiest,
1891	struct sched_domain *sd, enum cpu_idle_type idle)
1892	{
1893	struct task_struct p, n;
1894	struct cfs_rq *cfs_rq;
1895	int pinned = 0;
1896
1897	for_each_leaf_cfs_rq(busiest, cfs_rq) {
1898	list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1899
1900	if (!can_migrate_task(p, busiest, this_cpu,
1901	sd, idle, &pinned))
1902	continue;
1903
1904	pull_task(busiest, p, this_rq, this_cpu);
1905	/*
1906	* Right now, this is only the second place pull_task()
1907	* is called, so we can safely collect pull_task()
1908	* stats here rather than inside pull_task().
1909	*/
1910	schedstat_inc(sd, lb_gained[idle]);
1911	return 1;
1912	}
1913	}
1914
1915	return 0;
1916	}
1917
1918	static unsigned long
1919	balance_tasks(struct rq this_rq, int this_cpu, struct rq busiest,
1920	unsigned long max_load_move, struct sched_domain *sd,
1921	enum cpu_idle_type idle, int *all_pinned,
1922	int this_best_prio, struct cfs_rq busiest_cfs_rq)
1923	{
1924	int loops = 0, pulled = 0, pinned = 0;
1925	long rem_load_move = max_load_move;
1926	struct task_struct p, n;
1927
1928	if (max_load_move == 0)
1929	goto out;
1930
1931	pinned = 1;
1932
1933	list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1934	if (loops++ > sysctl_sched_nr_migrate)
1935	break;
1936
1937	if ((p->se.load.weight >> 1) > rem_load_move \|\|
1938	!can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1939	continue;
1940
1941	pull_task(busiest, p, this_rq, this_cpu);
1942	pulled++;
1943	rem_load_move -= p->se.load.weight;
1944
1945	#ifdef CONFIG_PREEMPT
1946	/*
1947	* NEWIDLE balancing is a source of latency, so preemptible
1948	* kernels will stop after the first task is pulled to minimize
1949	* the critical section.
1950	*/
1951	if (idle == CPU_NEWLY_IDLE)
1952	break;
1953	#endif
1954
1955	/*
1956	* We only want to steal up to the prescribed amount of
1957	* weighted load.
1958	*/
1959	if (rem_load_move <= 0)
1960	break;
1961
1962	if (p->prio < *this_best_prio)
1963	*this_best_prio = p->prio;
1964	}
1965	out:
1966	/*
1967	* Right now, this is one of only two places pull_task() is called,
1968	* so we can safely collect pull_task() stats here rather than
1969	* inside pull_task().
1970	*/
1971	schedstat_add(sd, lb_gained[idle], pulled);
1972
1973	if (all_pinned)
1974	*all_pinned = pinned;
1975
1976	return max_load_move - rem_load_move;
1977	}
1978
1979	#ifdef CONFIG_FAIR_GROUP_SCHED
1980	static unsigned long
1981	load_balance_fair(struct rq this_rq, int this_cpu, struct rq busiest,
1982	unsigned long max_load_move,
1983	struct sched_domain *sd, enum cpu_idle_type idle,
1984	int all_pinned, int this_best_prio)
1985	{
1986	long rem_load_move = max_load_move;
1987	int busiest_cpu = cpu_of(busiest);
1988	struct task_group *tg;
1989
1990	rcu_read_lock();
1991	update_h_load(busiest_cpu);
1992
1993	list_for_each_entry_rcu(tg, &task_groups, list) {
1994	struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1995	unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1996	unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1997	u64 rem_load, moved_load;
1998
1999	/*
2000	* empty group
2001	*/
2002	if (!busiest_cfs_rq->task_weight)
2003	continue;
2004
2005	rem_load = (u64)rem_load_move * busiest_weight;
2006	rem_load = div_u64(rem_load, busiest_h_load + 1);
2007
2008	moved_load = balance_tasks(this_rq, this_cpu, busiest,
2009	rem_load, sd, idle, all_pinned, this_best_prio,
2010	busiest_cfs_rq);
2011
2012	if (!moved_load)
2013	continue;
2014
2015	moved_load *= busiest_h_load;
2016	moved_load = div_u64(moved_load, busiest_weight + 1);
2017
2018	rem_load_move -= moved_load;
2019	if (rem_load_move < 0)
2020	break;
2021	}
2022	rcu_read_unlock();
2023
2024	return max_load_move - rem_load_move;
2025	}
2026	#else
2027	static unsigned long
2028	load_balance_fair(struct rq this_rq, int this_cpu, struct rq busiest,
2029	unsigned long max_load_move,
2030	struct sched_domain *sd, enum cpu_idle_type idle,
2031	int all_pinned, int this_best_prio)
2032	{
2033	return balance_tasks(this_rq, this_cpu, busiest,
2034	max_load_move, sd, idle, all_pinned,
2035	this_best_prio, &busiest->cfs);
2036	}
2037	#endif
2038
2039	/*
2040	* move_tasks tries to move up to max_load_move weighted load from busiest to
2041	* this_rq, as part of a balancing operation within domain "sd".
2042	* Returns 1 if successful and 0 otherwise.
2043	*
2044	* Called with both runqueues locked.
2045	*/
2046	static int move_tasks(struct rq this_rq, int this_cpu, struct rq busiest,
2047	unsigned long max_load_move,
2048	struct sched_domain *sd, enum cpu_idle_type idle,
2049	int *all_pinned)
2050	{
2051	unsigned long total_load_moved = 0, load_moved;
2052	int this_best_prio = this_rq->curr->prio;
2053
2054	do {
2055	load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2056	max_load_move - total_load_moved,
2057	sd, idle, all_pinned, &this_best_prio);
2058
2059	total_load_moved += load_moved;
2060
2061	#ifdef CONFIG_PREEMPT
2062	/*
2063	* NEWIDLE balancing is a source of latency, so preemptible
2064	* kernels will stop after the first task is pulled to minimize
2065	* the critical section.
2066	*/
2067	if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2068	break;
2069
2070	if (raw_spin_is_contended(&this_rq->lock) \|\|
2071	raw_spin_is_contended(&busiest->lock))
2072	break;
2073	#endif
2074	} while (load_moved && max_load_move > total_load_moved);
2075
2076	return total_load_moved > 0;
2077	}
2078
2079	/******** Helpers for find_busiest_group **********************/
2080	/*
2081	* sd_lb_stats - Structure to store the statistics of a sched_domain
2082	* during load balancing.
2083	*/
2084	struct sd_lb_stats {
2085	struct sched_group busiest; / Busiest group in this sd */
2086	struct sched_group this; / Local group in this sd */
2087	unsigned long total_load; /* Total load of all groups in sd */
2088	unsigned long total_pwr; /* Total power of all groups in sd */
2089	unsigned long avg_load; /* Average load across all groups in sd */
2090
2091	/** Statistics of this group */
2092	unsigned long this_load;
2093	unsigned long this_load_per_task;
2094	unsigned long this_nr_running;
2095
2096	/* Statistics of the busiest group */
2097	unsigned long max_load;
2098	unsigned long busiest_load_per_task;
2099	unsigned long busiest_nr_running;
2100	unsigned long busiest_group_capacity;
2101
2102	int group_imb; /* Is there imbalance in this sd */
2103	#if defined(CONFIG_SCHED_MC) \|\| defined(CONFIG_SCHED_SMT)
2104	int power_savings_balance; /* Is powersave balance needed for this sd */
2105	struct sched_group group_min; / Least loaded group in sd */
2106	struct sched_group group_leader; / Group which relieves group_min */
2107	unsigned long min_load_per_task; /* load_per_task in group_min */
2108	unsigned long leader_nr_running; /* Nr running of group_leader */
2109	unsigned long min_nr_running; /* Nr running of group_min */
2110	#endif
2111	};
2112
2113	/*
2114	* sg_lb_stats - stats of a sched_group required for load_balancing
2115	*/
2116	struct sg_lb_stats {
2117	unsigned long avg_load; /Avg load across the CPUs of the group /
2118	unsigned long group_load; /* Total load over the CPUs of the group */
2119	unsigned long sum_nr_running; /* Nr tasks running in the group */
2120	unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2121	unsigned long group_capacity;
2122	int group_imb; /* Is there an imbalance in the group ? */
2123	};
2124
2125	/**
2126	* group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2127	* @group: The group whose first cpu is to be returned.
2128	*/
2129	static inline unsigned int group_first_cpu(struct sched_group *group)
2130	{
2131	return cpumask_first(sched_group_cpus(group));
2132	}
2133
2134	/**
2135	* get_sd_load_idx - Obtain the load index for a given sched domain.
2136	* @sd: The sched_domain whose load_idx is to be obtained.
2137	* @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2138	*/
2139	static inline int get_sd_load_idx(struct sched_domain *sd,
2140	enum cpu_idle_type idle)
2141	{
2142	int load_idx;
2143
2144	switch (idle) {
2145	case CPU_NOT_IDLE:
2146	load_idx = sd->busy_idx;
2147	break;
2148
2149	case CPU_NEWLY_IDLE:
2150	load_idx = sd->newidle_idx;
2151	break;
2152	default:
2153	load_idx = sd->idle_idx;
2154	break;
2155	}
2156
2157	return load_idx;
2158	}
2159
2160
2161	#if defined(CONFIG_SCHED_MC) \|\| defined(CONFIG_SCHED_SMT)
2162	/**
2163	* init_sd_power_savings_stats - Initialize power savings statistics for
2164	* the given sched_domain, during load balancing.
2165	*
2166	* @sd: Sched domain whose power-savings statistics are to be initialized.
2167	* @sds: Variable containing the statistics for sd.
2168	* @idle: Idle status of the CPU at which we're performing load-balancing.
2169	*/
2170	static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2171	struct sd_lb_stats *sds, enum cpu_idle_type idle)
2172	{
2173	/*
2174	* Busy processors will not participate in power savings
2175	* balance.
2176	*/
2177	if (idle == CPU_NOT_IDLE \|\| !(sd->flags & SD_POWERSAVINGS_BALANCE))
2178	sds->power_savings_balance = 0;
2179	else {
2180	sds->power_savings_balance = 1;
2181	sds->min_nr_running = ULONG_MAX;
2182	sds->leader_nr_running = 0;
2183	}
2184	}
2185
2186	/**
2187	* update_sd_power_savings_stats - Update the power saving stats for a
2188	* sched_domain while performing load balancing.
2189	*
2190	* @group: sched_group belonging to the sched_domain under consideration.
2191	* @sds: Variable containing the statistics of the sched_domain
2192	* @local_group: Does group contain the CPU for which we're performing
2193	* load balancing ?
2194	* @sgs: Variable containing the statistics of the group.
2195	*/
2196	static inline void update_sd_power_savings_stats(struct sched_group *group,
2197	struct sd_lb_stats sds, int local_group, struct sg_lb_stats sgs)
2198	{
2199
2200	if (!sds->power_savings_balance)
2201	return;
2202
2203	/*
2204	* If the local group is idle or completely loaded
2205	* no need to do power savings balance at this domain
2206	*/
2207	if (local_group && (sds->this_nr_running >= sgs->group_capacity \|\|
2208	!sds->this_nr_running))
2209	sds->power_savings_balance = 0;
2210
2211	/*
2212	* If a group is already running at full capacity or idle,
2213	* don't include that group in power savings calculations
2214	*/
2215	if (!sds->power_savings_balance \|\|
2216	sgs->sum_nr_running >= sgs->group_capacity \|\|
2217	!sgs->sum_nr_running)
2218	return;
2219
2220	/*
2221	* Calculate the group which has the least non-idle load.
2222	* This is the group from where we need to pick up the load
2223	* for saving power
2224	*/
2225	if ((sgs->sum_nr_running < sds->min_nr_running) \|\|
2226	(sgs->sum_nr_running == sds->min_nr_running &&
2227	group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2228	sds->group_min = group;
2229	sds->min_nr_running = sgs->sum_nr_running;
2230	sds->min_load_per_task = sgs->sum_weighted_load /
2231	sgs->sum_nr_running;
2232	}
2233
2234	/*
2235	* Calculate the group which is almost near its
2236	* capacity but still has some space to pick up some load
2237	* from other group and save more power
2238	*/
2239	if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2240	return;
2241
2242	if (sgs->sum_nr_running > sds->leader_nr_running \|\|
2243	(sgs->sum_nr_running == sds->leader_nr_running &&
2244	group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2245	sds->group_leader = group;
2246	sds->leader_nr_running = sgs->sum_nr_running;
2247	}
2248	}
2249
2250	/**
2251	* check_power_save_busiest_group - see if there is potential for some power-savings balance
2252	* @sds: Variable containing the statistics of the sched_domain
2253	* under consideration.
2254	* @this_cpu: Cpu at which we're currently performing load-balancing.
2255	* @imbalance: Variable to store the imbalance.
2256	*
2257	* Description:
2258	* Check if we have potential to perform some power-savings balance.
2259	* If yes, set the busiest group to be the least loaded group in the
2260	* sched_domain, so that it's CPUs can be put to idle.
2261	*
2262	* Returns 1 if there is potential to perform power-savings balance.
2263	* Else returns 0.
2264	*/
2265	static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2266	int this_cpu, unsigned long *imbalance)
2267	{
2268	if (!sds->power_savings_balance)
2269	return 0;
2270
2271	if (sds->this != sds->group_leader \|\|
2272	sds->group_leader == sds->group_min)
2273	return 0;
2274
2275	*imbalance = sds->min_load_per_task;
2276	sds->busiest = sds->group_min;
2277
2278	return 1;
2279
2280	}
2281	#else /* CONFIG_SCHED_MC \|\| CONFIG_SCHED_SMT */
2282	static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2283	struct sd_lb_stats *sds, enum cpu_idle_type idle)
2284	{
2285	return;
2286	}
2287
2288	static inline void update_sd_power_savings_stats(struct sched_group *group,
2289	struct sd_lb_stats sds, int local_group, struct sg_lb_stats sgs)
2290	{
2291	return;
2292	}
2293
2294	static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2295	int this_cpu, unsigned long *imbalance)
2296	{
2297	return 0;
2298	}
2299	#endif /* CONFIG_SCHED_MC \|\| CONFIG_SCHED_SMT */
2300
2301
2302	unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2303	{
2304	return SCHED_LOAD_SCALE;
2305	}
2306
2307	unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2308	{
2309	return default_scale_freq_power(sd, cpu);
2310	}
2311
2312	unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2313	{
2314	unsigned long weight = cpumask_weight(sched_domain_span(sd));
2315	unsigned long smt_gain = sd->smt_gain;
2316
2317	smt_gain /= weight;
2318
2319	return smt_gain;
2320	}
2321
2322	unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2323	{
2324	return default_scale_smt_power(sd, cpu);
2325	}
2326
2327	unsigned long scale_rt_power(int cpu)
2328	{
2329	struct rq *rq = cpu_rq(cpu);
2330	u64 total, available;
2331
2332	sched_avg_update(rq);
2333
2334	total = sched_avg_period() + (rq->clock - rq->age_stamp);
2335	available = total - rq->rt_avg;
2336
2337	if (unlikely((s64)total < SCHED_LOAD_SCALE))
2338	total = SCHED_LOAD_SCALE;
2339
2340	total >>= SCHED_LOAD_SHIFT;
2341
2342	return div_u64(available, total);
2343	}
2344
2345	static void update_cpu_power(struct sched_domain *sd, int cpu)
2346	{
2347	unsigned long weight = cpumask_weight(sched_domain_span(sd));
2348	unsigned long power = SCHED_LOAD_SCALE;
2349	struct sched_group *sdg = sd->groups;
2350
2351	if (sched_feat(ARCH_POWER))
2352	power *= arch_scale_freq_power(sd, cpu);
2353	else
2354	power *= default_scale_freq_power(sd, cpu);
2355
2356	power >>= SCHED_LOAD_SHIFT;
2357
2358	if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2359	if (sched_feat(ARCH_POWER))
2360	power *= arch_scale_smt_power(sd, cpu);
2361	else
2362	power *= default_scale_smt_power(sd, cpu);
2363
2364	power >>= SCHED_LOAD_SHIFT;
2365	}
2366
2367	power *= scale_rt_power(cpu);
2368	power >>= SCHED_LOAD_SHIFT;
2369
2370	if (!power)
2371	power = 1;
2372
2373	sdg->cpu_power = power;
2374	}
2375
2376	static void update_group_power(struct sched_domain *sd, int cpu)
2377	{
2378	struct sched_domain *child = sd->child;
2379	struct sched_group group, sdg = sd->groups;
2380	unsigned long power;
2381
2382	if (!child) {
2383	update_cpu_power(sd, cpu);
2384	return;
2385	}
2386
2387	power = 0;
2388
2389	group = child->groups;
2390	do {
2391	power += group->cpu_power;
2392	group = group->next;
2393	} while (group != child->groups);
2394
2395	sdg->cpu_power = power;
2396	}
2397
2398	/**
2399	* update_sg_lb_stats - Update sched_group's statistics for load balancing.
2400	* @sd: The sched_domain whose statistics are to be updated.
2401	* @group: sched_group whose statistics are to be updated.
2402	* @this_cpu: Cpu for which load balance is currently performed.
2403	* @idle: Idle status of this_cpu
2404	* @load_idx: Load index of sched_domain of this_cpu for load calc.
2405	* @sd_idle: Idle status of the sched_domain containing group.
2406	* @local_group: Does group contain this_cpu.
2407	* @cpus: Set of cpus considered for load balancing.
2408	* @balance: Should we balance.
2409	* @sgs: variable to hold the statistics for this group.
2410	*/
2411	static inline void update_sg_lb_stats(struct sched_domain *sd,
2412	struct sched_group *group, int this_cpu,
2413	enum cpu_idle_type idle, int load_idx, int *sd_idle,
2414	int local_group, const struct cpumask *cpus,
2415	int balance, struct sg_lb_stats sgs)
2416	{
2417	unsigned long load, max_cpu_load, min_cpu_load;
2418	int i;
2419	unsigned int balance_cpu = -1, first_idle_cpu = 0;
2420	unsigned long avg_load_per_task = 0;
2421
2422	if (local_group)
2423	balance_cpu = group_first_cpu(group);
2424
2425	/* Tally up the load of all CPUs in the group */
2426	max_cpu_load = 0;
2427	min_cpu_load = ~0UL;
2428
2429	for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2430	struct rq *rq = cpu_rq(i);
2431
2432	if (*sd_idle && rq->nr_running)
2433	*sd_idle = 0;
2434
2435	/* Bias balancing toward cpus of our domain */
2436	if (local_group) {
2437	if (idle_cpu(i) && !first_idle_cpu) {
2438	first_idle_cpu = 1;
2439	balance_cpu = i;
2440	}
2441
2442	load = target_load(i, load_idx);
2443	} else {
2444	load = source_load(i, load_idx);
2445	if (load > max_cpu_load)
2446	max_cpu_load = load;
2447	if (min_cpu_load > load)
2448	min_cpu_load = load;
2449	}
2450
2451	sgs->group_load += load;
2452	sgs->sum_nr_running += rq->nr_running;
2453	sgs->sum_weighted_load += weighted_cpuload(i);
2454
2455	}
2456
2457	/*
2458	* First idle cpu or the first cpu(busiest) in this sched group
2459	* is eligible for doing load balancing at this and above
2460	* domains. In the newly idle case, we will allow all the cpu's
2461	* to do the newly idle load balance.
2462	*/
2463	if (idle != CPU_NEWLY_IDLE && local_group &&
2464	balance_cpu != this_cpu) {
2465	*balance = 0;
2466	return;
2467	}
2468
2469	update_group_power(sd, this_cpu);
2470
2471	/* Adjust by relative CPU power of the group */
2472	sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2473
2474	/*
2475	* Consider the group unbalanced when the imbalance is larger
2476	* than the average weight of two tasks.
2477	*
2478	* APZ: with cgroup the avg task weight can vary wildly and
2479	* might not be a suitable number - should we keep a
2480	* normalized nr_running number somewhere that negates
2481	* the hierarchy?
2482	*/
2483	if (sgs->sum_nr_running)
2484	avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2485
2486	if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
2487	sgs->group_imb = 1;
2488
2489	sgs->group_capacity =
2490	DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2491	}
2492
2493	/**
2494	* update_sd_lb_stats - Update sched_group's statistics for load balancing.
2495	* @sd: sched_domain whose statistics are to be updated.
2496	* @this_cpu: Cpu for which load balance is currently performed.
2497	* @idle: Idle status of this_cpu
2498	* @sd_idle: Idle status of the sched_domain containing group.
2499	* @cpus: Set of cpus considered for load balancing.
2500	* @balance: Should we balance.
2501	* @sds: variable to hold the statistics for this sched_domain.
2502	*/
2503	static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2504	enum cpu_idle_type idle, int *sd_idle,
2505	const struct cpumask cpus, int balance,
2506	struct sd_lb_stats *sds)
2507	{
2508	struct sched_domain *child = sd->child;
2509	struct sched_group *group = sd->groups;
2510	struct sg_lb_stats sgs;
2511	int load_idx, prefer_sibling = 0;
2512
2513	if (child && child->flags & SD_PREFER_SIBLING)
2514	prefer_sibling = 1;
2515
2516	init_sd_power_savings_stats(sd, sds, idle);
2517	load_idx = get_sd_load_idx(sd, idle);
2518
2519	do {
2520	int local_group;
2521
2522	local_group = cpumask_test_cpu(this_cpu,
2523	sched_group_cpus(group));
2524	memset(&sgs, 0, sizeof(sgs));
2525	update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
2526	local_group, cpus, balance, &sgs);
2527
2528	if (local_group && !(*balance))
2529	return;
2530
2531	sds->total_load += sgs.group_load;
2532	sds->total_pwr += group->cpu_power;
2533
2534	/*
2535	* In case the child domain prefers tasks go to siblings
2536	* first, lower the group capacity to one so that we'll try
2537	* and move all the excess tasks away.
2538	*/
2539	if (prefer_sibling)
2540	sgs.group_capacity = min(sgs.group_capacity, 1UL);
2541
2542	if (local_group) {
2543	sds->this_load = sgs.avg_load;
2544	sds->this = group;
2545	sds->this_nr_running = sgs.sum_nr_running;
2546	sds->this_load_per_task = sgs.sum_weighted_load;
2547	} else if (sgs.avg_load > sds->max_load &&
2548	(sgs.sum_nr_running > sgs.group_capacity \|\|
2549	sgs.group_imb)) {
2550	sds->max_load = sgs.avg_load;
2551	sds->busiest = group;
2552	sds->busiest_nr_running = sgs.sum_nr_running;
2553	sds->busiest_group_capacity = sgs.group_capacity;
2554	sds->busiest_load_per_task = sgs.sum_weighted_load;
2555	sds->group_imb = sgs.group_imb;
2556	}
2557
2558	update_sd_power_savings_stats(group, sds, local_group, &sgs);
2559	group = group->next;
2560	} while (group != sd->groups);
2561	}
2562
2563	/**
2564	* fix_small_imbalance - Calculate the minor imbalance that exists
2565	* amongst the groups of a sched_domain, during
2566	* load balancing.
2567	* @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2568	* @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2569	* @imbalance: Variable to store the imbalance.
2570	*/
2571	static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2572	int this_cpu, unsigned long *imbalance)
2573	{
2574	unsigned long tmp, pwr_now = 0, pwr_move = 0;
2575	unsigned int imbn = 2;
2576	unsigned long scaled_busy_load_per_task;
2577
2578	if (sds->this_nr_running) {
2579	sds->this_load_per_task /= sds->this_nr_running;
2580	if (sds->busiest_load_per_task >
2581	sds->this_load_per_task)
2582	imbn = 1;
2583	} else
2584	sds->this_load_per_task =
2585	cpu_avg_load_per_task(this_cpu);
2586
2587	scaled_busy_load_per_task = sds->busiest_load_per_task
2588	* SCHED_LOAD_SCALE;
2589	scaled_busy_load_per_task /= sds->busiest->cpu_power;
2590
2591	if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2592	(scaled_busy_load_per_task * imbn)) {
2593	*imbalance = sds->busiest_load_per_task;
2594	return;
2595	}
2596
2597	/*
2598	* OK, we don't have enough imbalance to justify moving tasks,
2599	* however we may be able to increase total CPU power used by
2600	* moving them.
2601	*/
2602
2603	pwr_now += sds->busiest->cpu_power *
2604	min(sds->busiest_load_per_task, sds->max_load);
2605	pwr_now += sds->this->cpu_power *
2606	min(sds->this_load_per_task, sds->this_load);
2607	pwr_now /= SCHED_LOAD_SCALE;
2608
2609	/* Amount of load we'd subtract */
2610	tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2611	sds->busiest->cpu_power;
2612	if (sds->max_load > tmp)
2613	pwr_move += sds->busiest->cpu_power *
2614	min(sds->busiest_load_per_task, sds->max_load - tmp);
2615
2616	/* Amount of load we'd add */
2617	if (sds->max_load * sds->busiest->cpu_power <
2618	sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2619	tmp = (sds->max_load * sds->busiest->cpu_power) /
2620	sds->this->cpu_power;
2621	else
2622	tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2623	sds->this->cpu_power;
2624	pwr_move += sds->this->cpu_power *
2625	min(sds->this_load_per_task, sds->this_load + tmp);
2626	pwr_move /= SCHED_LOAD_SCALE;
2627
2628	/* Move if we gain throughput */
2629	if (pwr_move > pwr_now)
2630	*imbalance = sds->busiest_load_per_task;
2631	}
2632
2633	/**
2634	* calculate_imbalance - Calculate the amount of imbalance present within the
2635	* groups of a given sched_domain during load balance.
2636	* @sds: statistics of the sched_domain whose imbalance is to be calculated.
2637	* @this_cpu: Cpu for which currently load balance is being performed.
2638	* @imbalance: The variable to store the imbalance.
2639	*/
2640	static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2641	unsigned long *imbalance)
2642	{
2643	unsigned long max_pull, load_above_capacity = ~0UL;
2644
2645	sds->busiest_load_per_task /= sds->busiest_nr_running;
2646	if (sds->group_imb) {
2647	sds->busiest_load_per_task =
2648	min(sds->busiest_load_per_task, sds->avg_load);
2649	}
2650
2651	/*
2652	* In the presence of smp nice balancing, certain scenarios can have
2653	* max load less than avg load(as we skip the groups at or below
2654	* its cpu_power, while calculating max_load..)
2655	*/
2656	if (sds->max_load < sds->avg_load) {
2657	*imbalance = 0;
2658	return fix_small_imbalance(sds, this_cpu, imbalance);
2659	}
2660
2661	if (!sds->group_imb) {
2662	/*
2663	* Don't want to pull so many tasks that a group would go idle.
2664	*/
2665	load_above_capacity = (sds->busiest_nr_running -
2666	sds->busiest_group_capacity);
2667
2668	load_above_capacity = (SCHED_LOAD_SCALE SCHED_LOAD_SCALE);
2669
2670	load_above_capacity /= sds->busiest->cpu_power;
2671	}
2672
2673	/*
2674	* We're trying to get all the cpus to the average_load, so we don't
2675	* want to push ourselves above the average load, nor do we wish to
2676	* reduce the max loaded cpu below the average load. At the same time,
2677	* we also don't want to reduce the group load below the group capacity
2678	* (so that we can implement power-savings policies etc). Thus we look
2679	* for the minimum possible imbalance.
2680	* Be careful of negative numbers as they'll appear as very large values
2681	* with unsigned longs.
2682	*/
2683	max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2684
2685	/* How much load to actually move to equalise the imbalance */
2686	imbalance = min(max_pull sds->busiest->cpu_power,
2687	(sds->avg_load - sds->this_load) * sds->this->cpu_power)
2688	/ SCHED_LOAD_SCALE;
2689
2690	/*
2691	* if *imbalance is less than the average load per runnable task
2692	* there is no gaurantee that any tasks will be moved so we'll have
2693	* a think about bumping its value to force at least one task to be
2694	* moved
2695	*/
2696	if (*imbalance < sds->busiest_load_per_task)
2697	return fix_small_imbalance(sds, this_cpu, imbalance);
2698
2699	}
2700	/***** find_busiest_group() helpers end here *******************/
2701
2702	/**
2703	* find_busiest_group - Returns the busiest group within the sched_domain
2704	* if there is an imbalance. If there isn't an imbalance, and
2705	* the user has opted for power-savings, it returns a group whose
2706	* CPUs can be put to idle by rebalancing those tasks elsewhere, if
2707	* such a group exists.
2708	*
2709	* Also calculates the amount of weighted load which should be moved
2710	* to restore balance.
2711	*
2712	* @sd: The sched_domain whose busiest group is to be returned.
2713	* @this_cpu: The cpu for which load balancing is currently being performed.
2714	* @imbalance: Variable which stores amount of weighted load which should
2715	* be moved to restore balance/put a group to idle.
2716	* @idle: The idle status of this_cpu.
2717	* @sd_idle: The idleness of sd
2718	* @cpus: The set of CPUs under consideration for load-balancing.
2719	* @balance: Pointer to a variable indicating if this_cpu
2720	* is the appropriate cpu to perform load balancing at this_level.
2721	*
2722	* Returns: - the busiest group if imbalance exists.
2723	* - If no imbalance and user has opted for power-savings balance,
2724	* return the least loaded group whose CPUs can be
2725	* put to idle by rebalancing its tasks onto our group.
2726	*/
2727	static struct sched_group *
2728	find_busiest_group(struct sched_domain *sd, int this_cpu,
2729	unsigned long *imbalance, enum cpu_idle_type idle,
2730	int sd_idle, const struct cpumask cpus, int *balance)
2731	{
2732	struct sd_lb_stats sds;
2733
2734	memset(&sds, 0, sizeof(sds));
2735
2736	/*
2737	* Compute the various statistics relavent for load balancing at
2738	* this level.
2739	*/
2740	update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2741	balance, &sds);
2742
2743	/* Cases where imbalance does not exist from POV of this_cpu */
2744	/* 1) this_cpu is not the appropriate cpu to perform load balancing
2745	* at this level.
2746	* 2) There is no busy sibling group to pull from.
2747	* 3) This group is the busiest group.
2748	* 4) This group is more busy than the avg busieness at this
2749	* sched_domain.
2750	* 5) The imbalance is within the specified limit.
2751	*/
2752	if (!(*balance))
2753	goto ret;
2754
2755	if (!sds.busiest \|\| sds.busiest_nr_running == 0)
2756	goto out_balanced;
2757
2758	if (sds.this_load >= sds.max_load)
2759	goto out_balanced;
2760
2761	sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
2762
2763	if (sds.this_load >= sds.avg_load)
2764	goto out_balanced;
2765
2766	if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
2767	goto out_balanced;
2768
2769	/* Looks like there is an imbalance. Compute it */
2770	calculate_imbalance(&sds, this_cpu, imbalance);
2771	return sds.busiest;
2772
2773	out_balanced:
2774	/*
2775	* There is no obvious imbalance. But check if we can do some balancing
2776	* to save power.
2777	*/
2778	if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
2779	return sds.busiest;
2780	ret:
2781	*imbalance = 0;
2782	return NULL;
2783	}
2784
2785	/*
2786	* find_busiest_queue - find the busiest runqueue among the cpus in group.
2787	*/
2788	static struct rq *
2789	find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2790	unsigned long imbalance, const struct cpumask *cpus)
2791	{
2792	struct rq busiest = NULL, rq;
2793	unsigned long max_load = 0;
2794	int i;
2795
2796	for_each_cpu(i, sched_group_cpus(group)) {
2797	unsigned long power = power_of(i);
2798	unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
2799	unsigned long wl;
2800
2801	if (!cpumask_test_cpu(i, cpus))
2802	continue;
2803
2804	rq = cpu_rq(i);
2805	wl = weighted_cpuload(i);
2806
2807	/*
2808	* When comparing with imbalance, use weighted_cpuload()
2809	* which is not scaled with the cpu power.
2810	*/
2811	if (capacity && rq->nr_running == 1 && wl > imbalance)
2812	continue;
2813
2814	/*
2815	* For the load comparisons with the other cpu's, consider
2816	* the weighted_cpuload() scaled with the cpu power, so that
2817	* the load can be moved away from the cpu that is potentially
2818	* running at a lower capacity.
2819	*/
2820	wl = (wl * SCHED_LOAD_SCALE) / power;
2821
2822	if (wl > max_load) {
2823	max_load = wl;
2824	busiest = rq;
2825	}
2826	}
2827
2828	return busiest;
2829	}
2830
2831	/*
2832	* Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2833	* so long as it is large enough.
2834	*/
2835	#define MAX_PINNED_INTERVAL 512
2836
2837	/* Working cpumask for load_balance and load_balance_newidle. */
2838	static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
2839
2840	static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle)
2841	{
2842	if (idle == CPU_NEWLY_IDLE) {
2843	/*
2844	* The only task running in a non-idle cpu can be moved to this
2845	* cpu in an attempt to completely freeup the other CPU
2846	* package.
2847	*
2848	* The package power saving logic comes from
2849	* find_busiest_group(). If there are no imbalance, then
2850	* f_b_g() will return NULL. However when sched_mc={1,2} then
2851	* f_b_g() will select a group from which a running task may be
2852	* pulled to this cpu in order to make the other package idle.
2853	* If there is no opportunity to make a package idle and if
2854	* there are no imbalance, then f_b_g() will return NULL and no
2855	* action will be taken in load_balance_newidle().
2856	*
2857	* Under normal task pull operation due to imbalance, there
2858	* will be more than one task in the source run queue and
2859	* move_tasks() will succeed. ld_moved will be true and this
2860	* active balance code will not be triggered.
2861	*/
2862	if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2863	!test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2864	return 0;
2865
2866	if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
2867	return 0;
2868	}
2869
2870	return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
2871	}
2872
2873	/*
2874	* Check this_cpu to ensure it is balanced within domain. Attempt to move
2875	* tasks if there is an imbalance.
2876	*/
2877	static int load_balance(int this_cpu, struct rq *this_rq,
2878	struct sched_domain *sd, enum cpu_idle_type idle,
2879	int *balance)
2880	{
2881	int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2882	struct sched_group *group;
2883	unsigned long imbalance;
2884	struct rq *busiest;
2885	unsigned long flags;
2886	struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
2887
2888	cpumask_copy(cpus, cpu_active_mask);
2889
2890	/*
2891	* When power savings policy is enabled for the parent domain, idle
2892	* sibling can pick up load irrespective of busy siblings. In this case,
2893	* let the state of idle sibling percolate up as CPU_IDLE, instead of
2894	* portraying it as CPU_NOT_IDLE.
2895	*/
2896	if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2897	!test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2898	sd_idle = 1;
2899
2900	schedstat_inc(sd, lb_count[idle]);
2901
2902	redo:
2903	update_shares(sd);
2904	group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2905	cpus, balance);
2906
2907	if (*balance == 0)
2908	goto out_balanced;
2909
2910	if (!group) {
2911	schedstat_inc(sd, lb_nobusyg[idle]);
2912	goto out_balanced;
2913	}
2914
2915	busiest = find_busiest_queue(group, idle, imbalance, cpus);
2916	if (!busiest) {
2917	schedstat_inc(sd, lb_nobusyq[idle]);
2918	goto out_balanced;
2919	}
2920
2921	BUG_ON(busiest == this_rq);
2922
2923	schedstat_add(sd, lb_imbalance[idle], imbalance);
2924
2925	ld_moved = 0;
2926	if (busiest->nr_running > 1) {
2927	/*
2928	* Attempt to move tasks. If find_busiest_group has found
2929	* an imbalance but busiest->nr_running <= 1, the group is
2930	* still unbalanced. ld_moved simply stays zero, so it is
2931	* correctly treated as an imbalance.
2932	*/
2933	local_irq_save(flags);
2934	double_rq_lock(this_rq, busiest);
2935	ld_moved = move_tasks(this_rq, this_cpu, busiest,
2936	imbalance, sd, idle, &all_pinned);
2937	double_rq_unlock(this_rq, busiest);
2938	local_irq_restore(flags);
2939
2940	/*
2941	* some other cpu did the load balance for us.
2942	*/
2943	if (ld_moved && this_cpu != smp_processor_id())
2944	resched_cpu(this_cpu);
2945
2946	/* All tasks on this runqueue were pinned by CPU affinity */
2947	if (unlikely(all_pinned)) {
2948	cpumask_clear_cpu(cpu_of(busiest), cpus);
2949	if (!cpumask_empty(cpus))
2950	goto redo;
2951	goto out_balanced;
2952	}
2953	}
2954
2955	if (!ld_moved) {
2956	schedstat_inc(sd, lb_failed[idle]);
2957	sd->nr_balance_failed++;
2958
2959	if (need_active_balance(sd, sd_idle, idle)) {
2960	raw_spin_lock_irqsave(&busiest->lock, flags);
2961
2962	/* don't kick the migration_thread, if the curr
2963	* task on busiest cpu can't be moved to this_cpu
2964	*/
2965	if (!cpumask_test_cpu(this_cpu,
2966	&busiest->curr->cpus_allowed)) {
2967	raw_spin_unlock_irqrestore(&busiest->lock,
2968	flags);
2969	all_pinned = 1;
2970	goto out_one_pinned;
2971	}
2972
2973	if (!busiest->active_balance) {
2974	busiest->active_balance = 1;
2975	busiest->push_cpu = this_cpu;
2976	active_balance = 1;
2977	}
2978	raw_spin_unlock_irqrestore(&busiest->lock, flags);
2979	if (active_balance)
2980	wake_up_process(busiest->migration_thread);
2981
2982	/*
2983	* We've kicked active balancing, reset the failure
2984	* counter.
2985	*/
2986	sd->nr_balance_failed = sd->cache_nice_tries+1;
2987	}
2988	} else
2989	sd->nr_balance_failed = 0;
2990
2991	if (likely(!active_balance)) {
2992	/* We were unbalanced, so reset the balancing interval */
2993	sd->balance_interval = sd->min_interval;
2994	} else {
2995	/*
2996	* If we've begun active balancing, start to back off. This
2997	* case may not be covered by the all_pinned logic if there
2998	* is only 1 task on the busy runqueue (because we don't call
2999	* move_tasks).
3000	*/
3001	if (sd->balance_interval < sd->max_interval)
3002	sd->balance_interval *= 2;
3003	}
3004
3005	if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3006	!test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3007	ld_moved = -1;
3008
3009	goto out;
3010
3011	out_balanced:
3012	schedstat_inc(sd, lb_balanced[idle]);
3013
3014	sd->nr_balance_failed = 0;
3015
3016	out_one_pinned:
3017	/* tune up the balancing interval */
3018	if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) \|\|
3019	(sd->balance_interval < sd->max_interval))
3020	sd->balance_interval *= 2;
3021
3022	if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3023	!test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3024	ld_moved = -1;
3025	else
3026	ld_moved = 0;
3027	out:
3028	if (ld_moved)
3029	update_shares(sd);
3030	return ld_moved;
3031	}
3032
3033	/*
3034	* idle_balance is called by schedule() if this_cpu is about to become
3035	* idle. Attempts to pull tasks from other CPUs.
3036	*/
3037	static void idle_balance(int this_cpu, struct rq *this_rq)
3038	{
3039	struct sched_domain *sd;
3040	int pulled_task = 0;
3041	unsigned long next_balance = jiffies + HZ;
3042
3043	this_rq->idle_stamp = this_rq->clock;
3044
3045	if (this_rq->avg_idle < sysctl_sched_migration_cost)
3046	return;
3047
3048	/*
3049	* Drop the rq->lock, but keep IRQ/preempt disabled.
3050	*/
3051	raw_spin_unlock(&this_rq->lock);
3052
3053	for_each_domain(this_cpu, sd) {
3054	unsigned long interval;
3055	int balance = 1;
3056
3057	if (!(sd->flags & SD_LOAD_BALANCE))
3058	continue;
3059
3060	if (sd->flags & SD_BALANCE_NEWIDLE) {
3061	/* If we've pulled tasks over stop searching: */
3062	pulled_task = load_balance(this_cpu, this_rq,
3063	sd, CPU_NEWLY_IDLE, &balance);
3064	}
3065
3066	interval = msecs_to_jiffies(sd->balance_interval);
3067	if (time_after(next_balance, sd->last_balance + interval))
3068	next_balance = sd->last_balance + interval;
3069	if (pulled_task) {
3070	this_rq->idle_stamp = 0;
3071	break;
3072	}
3073	}
3074
3075	raw_spin_lock(&this_rq->lock);
3076
3077	if (pulled_task \|\| time_after(jiffies, this_rq->next_balance)) {
3078	/*
3079	* We are going idle. next_balance may be set based on
3080	* a busy processor. So reset next_balance.
3081	*/
3082	this_rq->next_balance = next_balance;
3083	}
3084	}
3085
3086	/*
3087	* active_load_balance is run by migration threads. It pushes running tasks
3088	* off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3089	* running on each physical CPU where possible, and avoids physical /
3090	* logical imbalances.
3091	*
3092	* Called with busiest_rq locked.
3093	*/
3094	static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3095	{
3096	int target_cpu = busiest_rq->push_cpu;
3097	struct sched_domain *sd;
3098	struct rq *target_rq;
3099
3100	/* Is there any task to move? */
3101	if (busiest_rq->nr_running <= 1)
3102	return;
3103
3104	target_rq = cpu_rq(target_cpu);
3105
3106	/*
3107	* This condition is "impossible", if it occurs
3108	* we need to fix it. Originally reported by
3109	* Bjorn Helgaas on a 128-cpu setup.
3110	*/
3111	BUG_ON(busiest_rq == target_rq);
3112
3113	/* move a task from busiest_rq to target_rq */
3114	double_lock_balance(busiest_rq, target_rq);
3115	update_rq_clock(busiest_rq);
3116	update_rq_clock(target_rq);
3117
3118	/* Search for an sd spanning us and the target CPU. */
3119	for_each_domain(target_cpu, sd) {
3120	if ((sd->flags & SD_LOAD_BALANCE) &&
3121	cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3122	break;
3123	}
3124
3125	if (likely(sd)) {
3126	schedstat_inc(sd, alb_count);
3127
3128	if (move_one_task(target_rq, target_cpu, busiest_rq,
3129	sd, CPU_IDLE))
3130	schedstat_inc(sd, alb_pushed);
3131	else
3132	schedstat_inc(sd, alb_failed);
3133	}
3134	double_unlock_balance(busiest_rq, target_rq);
3135	}
3136
3137	#ifdef CONFIG_NO_HZ
3138	static struct {
3139	atomic_t load_balancer;
3140	cpumask_var_t cpu_mask;
3141	cpumask_var_t ilb_grp_nohz_mask;
3142	} nohz ____cacheline_aligned = {
3143	.load_balancer = ATOMIC_INIT(-1),
3144	};
3145
3146	int get_nohz_load_balancer(void)
3147	{
3148	return atomic_read(&nohz.load_balancer);
3149	}
3150
3151	#if defined(CONFIG_SCHED_MC) \|\| defined(CONFIG_SCHED_SMT)
3152	/**
3153	* lowest_flag_domain - Return lowest sched_domain containing flag.
3154	* @cpu: The cpu whose lowest level of sched domain is to
3155	* be returned.
3156	* @flag: The flag to check for the lowest sched_domain
3157	* for the given cpu.
3158	*
3159	* Returns the lowest sched_domain of a cpu which contains the given flag.
3160	*/
3161	static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3162	{
3163	struct sched_domain *sd;
3164
3165	for_each_domain(cpu, sd)
3166	if (sd && (sd->flags & flag))
3167	break;
3168
3169	return sd;
3170	}
3171
3172	/**
3173	* for_each_flag_domain - Iterates over sched_domains containing the flag.
3174	* @cpu: The cpu whose domains we're iterating over.
3175	* @sd: variable holding the value of the power_savings_sd
3176	* for cpu.
3177	* @flag: The flag to filter the sched_domains to be iterated.
3178	*
3179	* Iterates over all the scheduler domains for a given cpu that has the 'flag'
3180	* set, starting from the lowest sched_domain to the highest.
3181	*/
3182	#define for_each_flag_domain(cpu, sd, flag) \
3183	for (sd = lowest_flag_domain(cpu, flag); \
3184	(sd && (sd->flags & flag)); sd = sd->parent)
3185
3186	/**
3187	* is_semi_idle_group - Checks if the given sched_group is semi-idle.
3188	* @ilb_group: group to be checked for semi-idleness
3189	*
3190	* Returns: 1 if the group is semi-idle. 0 otherwise.
3191	*
3192	* We define a sched_group to be semi idle if it has atleast one idle-CPU
3193	* and atleast one non-idle CPU. This helper function checks if the given
3194	* sched_group is semi-idle or not.
3195	*/
3196	static inline int is_semi_idle_group(struct sched_group *ilb_group)
3197	{
3198	cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
3199	sched_group_cpus(ilb_group));
3200
3201	/*
3202	* A sched_group is semi-idle when it has atleast one busy cpu
3203	* and atleast one idle cpu.
3204	*/
3205	if (cpumask_empty(nohz.ilb_grp_nohz_mask))
3206	return 0;
3207
3208	if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
3209	return 0;
3210
3211	return 1;
3212	}
3213	/**
3214	* find_new_ilb - Finds the optimum idle load balancer for nomination.
3215	* @cpu: The cpu which is nominating a new idle_load_balancer.
3216	*
3217	* Returns: Returns the id of the idle load balancer if it exists,
3218	* Else, returns >= nr_cpu_ids.
3219	*
3220	* This algorithm picks the idle load balancer such that it belongs to a
3221	* semi-idle powersavings sched_domain. The idea is to try and avoid
3222	* completely idle packages/cores just for the purpose of idle load balancing
3223	* when there are other idle cpu's which are better suited for that job.
3224	*/
3225	static int find_new_ilb(int cpu)
3226	{
3227	struct sched_domain *sd;
3228	struct sched_group *ilb_group;
3229
3230	/*
3231	* Have idle load balancer selection from semi-idle packages only
3232	* when power-aware load balancing is enabled
3233	*/
3234	if (!(sched_smt_power_savings \|\| sched_mc_power_savings))
3235	goto out_done;
3236
3237	/*
3238	* Optimize for the case when we have no idle CPUs or only one
3239	* idle CPU. Don't walk the sched_domain hierarchy in such cases
3240	*/
3241	if (cpumask_weight(nohz.cpu_mask) < 2)
3242	goto out_done;
3243
3244	for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3245	ilb_group = sd->groups;
3246
3247	do {
3248	if (is_semi_idle_group(ilb_group))
3249	return cpumask_first(nohz.ilb_grp_nohz_mask);
3250
3251	ilb_group = ilb_group->next;
3252
3253	} while (ilb_group != sd->groups);
3254	}
3255
3256	out_done:
3257	return cpumask_first(nohz.cpu_mask);
3258	}
3259	#else /* (CONFIG_SCHED_MC \|\| CONFIG_SCHED_SMT) */
3260	static inline int find_new_ilb(int call_cpu)
3261	{
3262	return cpumask_first(nohz.cpu_mask);
3263	}
3264	#endif
3265
3266	/*
3267	* This routine will try to nominate the ilb (idle load balancing)
3268	* owner among the cpus whose ticks are stopped. ilb owner will do the idle
3269	* load balancing on behalf of all those cpus. If all the cpus in the system
3270	* go into this tickless mode, then there will be no ilb owner (as there is
3271	* no need for one) and all the cpus will sleep till the next wakeup event
3272	* arrives...
3273	*
3274	* For the ilb owner, tick is not stopped. And this tick will be used
3275	* for idle load balancing. ilb owner will still be part of
3276	* nohz.cpu_mask..
3277	*
3278	* While stopping the tick, this cpu will become the ilb owner if there
3279	* is no other owner. And will be the owner till that cpu becomes busy
3280	* or if all cpus in the system stop their ticks at which point
3281	* there is no need for ilb owner.
3282	*
3283	* When the ilb owner becomes busy, it nominates another owner, during the
3284	* next busy scheduler_tick()
3285	*/
3286	int select_nohz_load_balancer(int stop_tick)
3287	{
3288	int cpu = smp_processor_id();
3289
3290	if (stop_tick) {
3291	cpu_rq(cpu)->in_nohz_recently = 1;
3292
3293	if (!cpu_active(cpu)) {
3294	if (atomic_read(&nohz.load_balancer) != cpu)
3295	return 0;
3296
3297	/*
3298	* If we are going offline and still the leader,
3299	* give up!
3300	*/
3301	if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3302	BUG();
3303
3304	return 0;
3305	}
3306
3307	cpumask_set_cpu(cpu, nohz.cpu_mask);
3308
3309	/* time for ilb owner also to sleep */
3310	if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
3311	if (atomic_read(&nohz.load_balancer) == cpu)
3312	atomic_set(&nohz.load_balancer, -1);
3313	return 0;
3314	}
3315
3316	if (atomic_read(&nohz.load_balancer) == -1) {
3317	/* make me the ilb owner */
3318	if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3319	return 1;
3320	} else if (atomic_read(&nohz.load_balancer) == cpu) {
3321	int new_ilb;
3322
3323	if (!(sched_smt_power_savings \|\|
3324	sched_mc_power_savings))
3325	return 1;
3326	/*
3327	* Check to see if there is a more power-efficient
3328	* ilb.
3329	*/
3330	new_ilb = find_new_ilb(cpu);
3331	if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3332	atomic_set(&nohz.load_balancer, -1);
3333	resched_cpu(new_ilb);
3334	return 0;
3335	}
3336	return 1;
3337	}
3338	} else {
3339	if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3340	return 0;
3341
3342	cpumask_clear_cpu(cpu, nohz.cpu_mask);
3343
3344	if (atomic_read(&nohz.load_balancer) == cpu)
3345	if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3346	BUG();
3347	}
3348	return 0;
3349	}
3350	#endif
3351
3352	static DEFINE_SPINLOCK(balancing);
3353
3354	/*
3355	* It checks each scheduling domain to see if it is due to be balanced,
3356	* and initiates a balancing operation if so.
3357	*
3358	* Balancing parameters are set up in arch_init_sched_domains.
3359	*/
3360	static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3361	{
3362	int balance = 1;
3363	struct rq *rq = cpu_rq(cpu);
3364	unsigned long interval;
3365	struct sched_domain *sd;
3366	/* Earliest time when we have to do rebalance again */
3367	unsigned long next_balance = jiffies + 60*HZ;
3368	int update_next_balance = 0;
3369	int need_serialize;
3370
3371	for_each_domain(cpu, sd) {
3372	if (!(sd->flags & SD_LOAD_BALANCE))
3373	continue;
3374
3375	interval = sd->balance_interval;
3376	if (idle != CPU_IDLE)
3377	interval *= sd->busy_factor;
3378
3379	/* scale ms to jiffies */
3380	interval = msecs_to_jiffies(interval);
3381	if (unlikely(!interval))
3382	interval = 1;
3383	if (interval > HZ*NR_CPUS/10)
3384	interval = HZ*NR_CPUS/10;
3385
3386	need_serialize = sd->flags & SD_SERIALIZE;
3387
3388	if (need_serialize) {
3389	if (!spin_trylock(&balancing))
3390	goto out;
3391	}
3392
3393	if (time_after_eq(jiffies, sd->last_balance + interval)) {
3394	if (load_balance(cpu, rq, sd, idle, &balance)) {
3395	/*
3396	* We've pulled tasks over so either we're no
3397	* longer idle, or one of our SMT siblings is
3398	* not idle.
3399	*/
3400	idle = CPU_NOT_IDLE;
3401	}
3402	sd->last_balance = jiffies;
3403	}
3404	if (need_serialize)
3405	spin_unlock(&balancing);
3406	out:
3407	if (time_after(next_balance, sd->last_balance + interval)) {
3408	next_balance = sd->last_balance + interval;
3409	update_next_balance = 1;
3410	}
3411
3412	/*
3413	* Stop the load balance at this level. There is another
3414	* CPU in our sched group which is doing load balancing more
3415	* actively.
3416	*/
3417	if (!balance)
3418	break;
3419	}
3420
3421	/*
3422	* next_balance will be updated only when there is a need.
3423	* When the cpu is attached to null domain for ex, it will not be
3424	* updated.
3425	*/
3426	if (likely(update_next_balance))
3427	rq->next_balance = next_balance;
3428	}
3429
3430	/*
3431	* run_rebalance_domains is triggered when needed from the scheduler tick.
3432	* In CONFIG_NO_HZ case, the idle load balance owner will do the
3433	* rebalancing for all the cpus for whom scheduler ticks are stopped.
3434	*/
3435	static void run_rebalance_domains(struct softirq_action *h)
3436	{
3437	int this_cpu = smp_processor_id();
3438	struct rq *this_rq = cpu_rq(this_cpu);
3439	enum cpu_idle_type idle = this_rq->idle_at_tick ?
3440	CPU_IDLE : CPU_NOT_IDLE;
3441
3442	rebalance_domains(this_cpu, idle);
3443
3444	#ifdef CONFIG_NO_HZ
3445	/*
3446	* If this cpu is the owner for idle load balancing, then do the
3447	* balancing on behalf of the other idle cpus whose ticks are
3448	* stopped.
3449	*/
3450	if (this_rq->idle_at_tick &&
3451	atomic_read(&nohz.load_balancer) == this_cpu) {
3452	struct rq *rq;
3453	int balance_cpu;
3454
3455	for_each_cpu(balance_cpu, nohz.cpu_mask) {
3456	if (balance_cpu == this_cpu)
3457	continue;
3458
3459	/*
3460	* If this cpu gets work to do, stop the load balancing
3461	* work being done for other cpus. Next load
3462	* balancing owner will pick it up.
3463	*/
3464	if (need_resched())
3465	break;
3466
3467	rebalance_domains(balance_cpu, CPU_IDLE);
3468
3469	rq = cpu_rq(balance_cpu);
3470	if (time_after(this_rq->next_balance, rq->next_balance))
3471	this_rq->next_balance = rq->next_balance;
3472	}
3473	}
3474	#endif
3475	}
3476
3477	static inline int on_null_domain(int cpu)
3478	{
3479	return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3480	}
3481
3482	/*
3483	* Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3484	*
3485	* In case of CONFIG_NO_HZ, this is the place where we nominate a new
3486	* idle load balancing owner or decide to stop the periodic load balancing,
3487	* if the whole system is idle.
3488	*/
3489	static inline void trigger_load_balance(struct rq *rq, int cpu)
3490	{
3491	#ifdef CONFIG_NO_HZ
3492	/*
3493	* If we were in the nohz mode recently and busy at the current
3494	* scheduler tick, then check if we need to nominate new idle
3495	* load balancer.
3496	*/
3497	if (rq->in_nohz_recently && !rq->idle_at_tick) {
3498	rq->in_nohz_recently = 0;
3499
3500	if (atomic_read(&nohz.load_balancer) == cpu) {
3501	cpumask_clear_cpu(cpu, nohz.cpu_mask);
3502	atomic_set(&nohz.load_balancer, -1);
3503	}
3504
3505	if (atomic_read(&nohz.load_balancer) == -1) {
3506	int ilb = find_new_ilb(cpu);
3507
3508	if (ilb < nr_cpu_ids)
3509	resched_cpu(ilb);
3510	}
3511	}
3512
3513	/*
3514	* If this cpu is idle and doing idle load balancing for all the
3515	* cpus with ticks stopped, is it time for that to stop?
3516	*/
3517	if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3518	cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3519	resched_cpu(cpu);
3520	return;
3521	}
3522
3523	/*
3524	* If this cpu is idle and the idle load balancing is done by
3525	* someone else, then no need raise the SCHED_SOFTIRQ
3526	*/
3527	if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3528	cpumask_test_cpu(cpu, nohz.cpu_mask))
3529	return;
3530	#endif
3531	/* Don't need to rebalance while attached to NULL domain */
3532	if (time_after_eq(jiffies, rq->next_balance) &&
3533	likely(!on_null_domain(cpu)))
3534	raise_softirq(SCHED_SOFTIRQ);
3535	}
3536
3537	static void rq_online_fair(struct rq *rq)
3538	{
3539	update_sysctl();
3540	}
3541
3542	static void rq_offline_fair(struct rq *rq)
3543	{
3544	update_sysctl();
3545	}
3546
3547	#else /* CONFIG_SMP */
3548
3549	/*
3550	* on UP we do not need to balance between CPUs:
3551	*/
3552	static inline void idle_balance(int cpu, struct rq *rq)
3553	{
3554	}
3555
3556	#endif /* CONFIG_SMP */
3557
3558	/*
3559	* scheduler tick hitting a task of our scheduling class:
3560	*/
3561	static void task_tick_fair(struct rq rq, struct task_struct curr, int queued)
3562	{
3563	struct cfs_rq *cfs_rq;
3564	struct sched_entity *se = &curr->se;
3565
3566	for_each_sched_entity(se) {
3567	cfs_rq = cfs_rq_of(se);
3568	entity_tick(cfs_rq, se, queued);
3569	}
3570	}
3571
3572	/*
3573	* called on fork with the child task as argument from the parent's context
3574	* - child not yet on the tasklist
3575	* - preemption disabled
3576	*/
3577	static void task_fork_fair(struct task_struct *p)
3578	{
3579	struct cfs_rq *cfs_rq = task_cfs_rq(current);
3580	struct sched_entity se = &p->se, curr = cfs_rq->curr;
3581	int this_cpu = smp_processor_id();
3582	struct rq *rq = this_rq();
3583	unsigned long flags;
3584
3585	raw_spin_lock_irqsave(&rq->lock, flags);
3586
3587	if (unlikely(task_cpu(p) != this_cpu))
3588	__set_task_cpu(p, this_cpu);
3589
3590	update_curr(cfs_rq);
3591
3592	if (curr)
3593	se->vruntime = curr->vruntime;
3594	place_entity(cfs_rq, se, 1);
3595
3596	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3597	/*
3598	* Upon rescheduling, sched_class::put_prev_task() will place
3599	* 'current' within the tree based on its new key value.
3600	*/
3601	swap(curr->vruntime, se->vruntime);
3602	resched_task(rq->curr);
3603	}
3604
3605	se->vruntime -= cfs_rq->min_vruntime;
3606
3607	raw_spin_unlock_irqrestore(&rq->lock, flags);
3608	}
3609
3610	/*
3611	* Priority of the task has changed. Check to see if we preempt
3612	* the current task.
3613	*/
3614	static void prio_changed_fair(struct rq rq, struct task_struct p,
3615	int oldprio, int running)
3616	{
3617	/*
3618	* Reschedule if we are currently running on this runqueue and
3619	* our priority decreased, or if we are not currently running on
3620	* this runqueue and our priority is higher than the current's
3621	*/
3622	if (running) {
3623	if (p->prio > oldprio)
3624	resched_task(rq->curr);
3625	} else
3626	check_preempt_curr(rq, p, 0);
3627	}
3628
3629	/*
3630	* We switched to the sched_fair class.
3631	*/
3632	static void switched_to_fair(struct rq rq, struct task_struct p,
3633	int running)
3634	{
3635	/*
3636	* We were most likely switched from sched_rt, so
3637	* kick off the schedule if running, otherwise just see
3638	* if we can still preempt the current task.
3639	*/
3640	if (running)
3641	resched_task(rq->curr);
3642	else
3643	check_preempt_curr(rq, p, 0);
3644	}
3645
3646	/* Account for a task changing its policy or group.
3647	*
3648	* This routine is mostly called to set cfs_rq->curr field when a task
3649	* migrates between groups/classes.
3650	*/
3651	static void set_curr_task_fair(struct rq *rq)
3652	{
3653	struct sched_entity *se = &rq->curr->se;
3654
3655	for_each_sched_entity(se)
3656	set_next_entity(cfs_rq_of(se), se);
3657	}
3658
3659	#ifdef CONFIG_FAIR_GROUP_SCHED
3660	static void moved_group_fair(struct task_struct *p, int on_rq)
3661	{
3662	struct cfs_rq *cfs_rq = task_cfs_rq(p);
3663
3664	update_curr(cfs_rq);
3665	if (!on_rq)
3666	place_entity(cfs_rq, &p->se, 1);
3667	}
3668	#endif
3669
3670	static unsigned int get_rr_interval_fair(struct rq rq, struct task_struct task)
3671	{
3672	struct sched_entity *se = &task->se;
3673	unsigned int rr_interval = 0;
3674
3675	/*
3676	* Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3677	* idle runqueue:
3678	*/
3679	if (rq->cfs.load.weight)
3680	rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
3681
3682	return rr_interval;
3683	}
3684
3685	/*
3686	* All the scheduling class methods:
3687	*/
3688	static const struct sched_class fair_sched_class = {
3689	.next = &idle_sched_class,
3690	.enqueue_task = enqueue_task_fair,
3691	.dequeue_task = dequeue_task_fair,
3692	.yield_task = yield_task_fair,
3693
3694	.check_preempt_curr = check_preempt_wakeup,
3695
3696	.pick_next_task = pick_next_task_fair,
3697	.put_prev_task = put_prev_task_fair,
3698
3699	#ifdef CONFIG_SMP
3700	.select_task_rq = select_task_rq_fair,
3701
3702	.rq_online = rq_online_fair,
3703	.rq_offline = rq_offline_fair,
3704
3705	.task_waking = task_waking_fair,
3706	#endif
3707
3708	.set_curr_task = set_curr_task_fair,
3709	.task_tick = task_tick_fair,
3710	.task_fork = task_fork_fair,
3711
3712	.prio_changed = prio_changed_fair,
3713	.switched_to = switched_to_fair,
3714
3715	.get_rr_interval = get_rr_interval_fair,
3716
3717	#ifdef CONFIG_FAIR_GROUP_SCHED
3718	.moved_group = moved_group_fair,
3719	#endif
3720	};
3721
3722	#ifdef CONFIG_SCHED_DEBUG
3723	static void print_cfs_stats(struct seq_file *m, int cpu)
3724	{
3725	struct cfs_rq *cfs_rq;
3726
3727	rcu_read_lock();
3728	for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
3729	print_cfs_rq(m, cpu, cfs_rq);
3730	rcu_read_unlock();
3731	}
3732	#endif
3733

Download this file

Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master

Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9

Kernel

Kernel Git Source Tree

Root/kernel/sched_fair.c

interactive