blob: bc8ee999381437515c7664c5e4b7fddb00656841 [file] [log] [blame]
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001/*
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>
Peter Zijlstra21805082007-08-25 18:41:53 +020018 *
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020021 */
22
Arjan van de Ven97455122008-01-25 21:08:34 +010023#include <linux/latencytop.h>
Christian Ehrhardt1983a922009-11-30 12:16:47 +010024#include <linux/sched.h>
Sisir Koppaka3436ae12011-03-26 18:22:55 +053025#include <linux/cpumask.h>
Arjan van de Ven97455122008-01-25 21:08:34 +010026
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020027/*
Peter Zijlstra21805082007-08-25 18:41:53 +020028 * Targeted preemption latency for CPU-bound tasks:
Takuya Yoshikawa864616e2010-10-14 16:09:13 +090029 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020030 *
Peter Zijlstra21805082007-08-25 18:41:53 +020031 * NOTE: this latency value is not the same as the concept of
Ingo Molnard274a4c2007-10-15 17:00:14 +020032 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020035 *
Ingo Molnard274a4c2007-10-15 17:00:14 +020036 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020038 */
Mike Galbraith21406922010-03-11 17:17:15 +010039unsigned int sysctl_sched_latency = 6000000ULL;
40unsigned int normalized_sysctl_sched_latency = 6000000ULL;
Ingo Molnar2bd8e6d2007-10-15 17:00:02 +020041
42/*
Christian Ehrhardt1983a922009-11-30 12:16:47 +010043 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
45 *
46 * Options are:
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 */
51enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
53
54/*
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010055 * Minimal preemption granularity for CPU-bound tasks:
Takuya Yoshikawa864616e2010-10-14 16:09:13 +090056 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010057 */
Ingo Molnar0bf377b2010-09-12 08:14:52 +020058unsigned int sysctl_sched_min_granularity = 750000ULL;
59unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010060
61/*
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 */
Ingo Molnar0bf377b2010-09-12 08:14:52 +020064static unsigned int sched_nr_latency = 8;
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010065
66/*
Mike Galbraith2bba22c2009-09-09 15:41:37 +020067 * After fork, child runs first. If set to 0 (default) then
Ingo Molnar2bd8e6d2007-10-15 17:00:02 +020068 * parent will (try to) run first.
69 */
Mike Galbraith2bba22c2009-09-09 15:41:37 +020070unsigned int sysctl_sched_child_runs_first __read_mostly;
Peter Zijlstra21805082007-08-25 18:41:53 +020071
72/*
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020073 * SCHED_OTHER wake-up granularity.
Mike Galbraith172e0822009-09-09 15:41:37 +020074 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020075 *
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
79 */
Mike Galbraith172e0822009-09-09 15:41:37 +020080unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +010081unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020082
Ingo Molnarda84d962007-10-15 17:00:18 +020083const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
84
Paul Turnera7a4f8a2010-11-15 15:47:06 -080085/*
86 * The exponential sliding window over which load is averaged for shares
87 * distribution.
88 * (default: 10msec)
89 */
90unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
91
Peter Zijlstraa4c2f002008-10-17 19:27:03 +020092static const struct sched_class fair_sched_class;
93
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020094/**************************************************************
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 */
101static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
102{
103 return cfs_rq->rq;
104}
105
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200106/* An entity is a task if it doesn't "own" a runqueue */
107#define entity_is_task(se) (!se->my_q)
108
Peter Zijlstra8f488942009-07-24 12:25:30 +0200109static 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
Peter Zijlstrab7581492008-04-19 19:45:00 +0200117/* Walk up scheduling entities hierarchy */
118#define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
120
121static 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 */
127static 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 */
133static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
134{
135 return grp->my_q;
136}
137
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800138static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
139{
140 if (!cfs_rq->on_list) {
Paul Turner67e86252010-11-15 15:47:05 -0800141 /*
142 * Ensure we either appear before our parent (if already
143 * enqueued) or force our parent to appear after us when it is
144 * enqueued. The fact that we always enqueue bottom-up
145 * reduces this to two cases.
146 */
147 if (cfs_rq->tg->parent &&
148 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
149 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800150 &rq_of(cfs_rq)->leaf_cfs_rq_list);
Paul Turner67e86252010-11-15 15:47:05 -0800151 } else {
152 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
153 &rq_of(cfs_rq)->leaf_cfs_rq_list);
154 }
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800155
156 cfs_rq->on_list = 1;
157 }
158}
159
160static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
161{
162 if (cfs_rq->on_list) {
163 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
164 cfs_rq->on_list = 0;
165 }
166}
167
Peter Zijlstrab7581492008-04-19 19:45:00 +0200168/* Iterate thr' all leaf cfs_rq's on a runqueue */
169#define for_each_leaf_cfs_rq(rq, cfs_rq) \
170 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
171
172/* Do the two (enqueued) entities belong to the same group ? */
173static inline int
174is_same_group(struct sched_entity *se, struct sched_entity *pse)
175{
176 if (se->cfs_rq == pse->cfs_rq)
177 return 1;
178
179 return 0;
180}
181
182static inline struct sched_entity *parent_entity(struct sched_entity *se)
183{
184 return se->parent;
185}
186
Peter Zijlstra464b7522008-10-24 11:06:15 +0200187/* return depth at which a sched entity is present in the hierarchy */
188static inline int depth_se(struct sched_entity *se)
189{
190 int depth = 0;
191
192 for_each_sched_entity(se)
193 depth++;
194
195 return depth;
196}
197
198static void
199find_matching_se(struct sched_entity **se, struct sched_entity **pse)
200{
201 int se_depth, pse_depth;
202
203 /*
204 * preemption test can be made between sibling entities who are in the
205 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
206 * both tasks until we find their ancestors who are siblings of common
207 * parent.
208 */
209
210 /* First walk up until both entities are at same depth */
211 se_depth = depth_se(*se);
212 pse_depth = depth_se(*pse);
213
214 while (se_depth > pse_depth) {
215 se_depth--;
216 *se = parent_entity(*se);
217 }
218
219 while (pse_depth > se_depth) {
220 pse_depth--;
221 *pse = parent_entity(*pse);
222 }
223
224 while (!is_same_group(*se, *pse)) {
225 *se = parent_entity(*se);
226 *pse = parent_entity(*pse);
227 }
228}
229
Peter Zijlstra8f488942009-07-24 12:25:30 +0200230#else /* !CONFIG_FAIR_GROUP_SCHED */
231
232static inline struct task_struct *task_of(struct sched_entity *se)
233{
234 return container_of(se, struct task_struct, se);
235}
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200236
237static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
238{
239 return container_of(cfs_rq, struct rq, cfs);
240}
241
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200242#define entity_is_task(se) 1
243
Peter Zijlstrab7581492008-04-19 19:45:00 +0200244#define for_each_sched_entity(se) \
245 for (; se; se = NULL)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200246
Peter Zijlstrab7581492008-04-19 19:45:00 +0200247static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200248{
Peter Zijlstrab7581492008-04-19 19:45:00 +0200249 return &task_rq(p)->cfs;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200250}
251
Peter Zijlstrab7581492008-04-19 19:45:00 +0200252static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
253{
254 struct task_struct *p = task_of(se);
255 struct rq *rq = task_rq(p);
256
257 return &rq->cfs;
258}
259
260/* runqueue "owned" by this group */
261static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
262{
263 return NULL;
264}
265
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800266static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
267{
268}
269
270static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
271{
272}
273
Peter Zijlstrab7581492008-04-19 19:45:00 +0200274#define for_each_leaf_cfs_rq(rq, cfs_rq) \
275 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
276
277static inline int
278is_same_group(struct sched_entity *se, struct sched_entity *pse)
279{
280 return 1;
281}
282
283static inline struct sched_entity *parent_entity(struct sched_entity *se)
284{
285 return NULL;
286}
287
Peter Zijlstra464b7522008-10-24 11:06:15 +0200288static inline void
289find_matching_se(struct sched_entity **se, struct sched_entity **pse)
290{
291}
292
Peter Zijlstrab7581492008-04-19 19:45:00 +0200293#endif /* CONFIG_FAIR_GROUP_SCHED */
294
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200295
296/**************************************************************
297 * Scheduling class tree data structure manipulation methods:
298 */
299
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200300static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
Peter Zijlstra02e04312007-10-15 17:00:07 +0200301{
Peter Zijlstra368059a2007-10-15 17:00:11 +0200302 s64 delta = (s64)(vruntime - min_vruntime);
303 if (delta > 0)
Peter Zijlstra02e04312007-10-15 17:00:07 +0200304 min_vruntime = vruntime;
305
306 return min_vruntime;
307}
308
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200309static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
Peter Zijlstrab0ffd242007-10-15 17:00:12 +0200310{
311 s64 delta = (s64)(vruntime - min_vruntime);
312 if (delta < 0)
313 min_vruntime = vruntime;
314
315 return min_vruntime;
316}
317
Fabio Checconi54fdc582009-07-16 12:32:27 +0200318static inline int entity_before(struct sched_entity *a,
319 struct sched_entity *b)
320{
321 return (s64)(a->vruntime - b->vruntime) < 0;
322}
323
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200324static void update_min_vruntime(struct cfs_rq *cfs_rq)
325{
326 u64 vruntime = cfs_rq->min_vruntime;
327
328 if (cfs_rq->curr)
329 vruntime = cfs_rq->curr->vruntime;
330
331 if (cfs_rq->rb_leftmost) {
332 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
333 struct sched_entity,
334 run_node);
335
Peter Zijlstrae17036d2009-01-15 14:53:39 +0100336 if (!cfs_rq->curr)
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200337 vruntime = se->vruntime;
338 else
339 vruntime = min_vruntime(vruntime, se->vruntime);
340 }
341
342 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
Peter Zijlstra3fe16982011-04-05 17:23:48 +0200343#ifndef CONFIG_64BIT
344 smp_wmb();
345 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
346#endif
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200347}
348
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200349/*
350 * Enqueue an entity into the rb-tree:
351 */
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200352static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200353{
354 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
355 struct rb_node *parent = NULL;
356 struct sched_entity *entry;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200357 int leftmost = 1;
358
359 /*
360 * Find the right place in the rbtree:
361 */
362 while (*link) {
363 parent = *link;
364 entry = rb_entry(parent, struct sched_entity, run_node);
365 /*
366 * We dont care about collisions. Nodes with
367 * the same key stay together.
368 */
Stephan Baerwolf2bd2d6f2011-07-20 14:46:59 +0200369 if (entity_before(se, entry)) {
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200370 link = &parent->rb_left;
371 } else {
372 link = &parent->rb_right;
373 leftmost = 0;
374 }
375 }
376
377 /*
378 * Maintain a cache of leftmost tree entries (it is frequently
379 * used):
380 */
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200381 if (leftmost)
Ingo Molnar57cb4992007-10-15 17:00:11 +0200382 cfs_rq->rb_leftmost = &se->run_node;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200383
384 rb_link_node(&se->run_node, parent, link);
385 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200386}
387
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200388static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200389{
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100390 if (cfs_rq->rb_leftmost == &se->run_node) {
391 struct rb_node *next_node;
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100392
393 next_node = rb_next(&se->run_node);
394 cfs_rq->rb_leftmost = next_node;
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100395 }
Ingo Molnare9acbff2007-10-15 17:00:04 +0200396
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200397 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200398}
399
Rik van Rielac53db52011-02-01 09:51:03 -0500400static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200401{
Peter Zijlstraf4b67552008-11-04 21:25:07 +0100402 struct rb_node *left = cfs_rq->rb_leftmost;
403
404 if (!left)
405 return NULL;
406
407 return rb_entry(left, struct sched_entity, run_node);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200408}
409
Rik van Rielac53db52011-02-01 09:51:03 -0500410static struct sched_entity *__pick_next_entity(struct sched_entity *se)
411{
412 struct rb_node *next = rb_next(&se->run_node);
413
414 if (!next)
415 return NULL;
416
417 return rb_entry(next, struct sched_entity, run_node);
418}
419
420#ifdef CONFIG_SCHED_DEBUG
Peter Zijlstraf4b67552008-11-04 21:25:07 +0100421static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200422{
Ingo Molnar7eee3e62008-02-22 10:32:21 +0100423 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200424
Balbir Singh70eee742008-02-22 13:25:53 +0530425 if (!last)
426 return NULL;
Ingo Molnar7eee3e62008-02-22 10:32:21 +0100427
428 return rb_entry(last, struct sched_entity, run_node);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200429}
430
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200431/**************************************************************
432 * Scheduling class statistics methods:
433 */
434
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100435int sched_proc_update_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700436 void __user *buffer, size_t *lenp,
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100437 loff_t *ppos)
438{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700439 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100440 int factor = get_update_sysctl_factor();
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100441
442 if (ret || !write)
443 return ret;
444
445 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
446 sysctl_sched_min_granularity);
447
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100448#define WRT_SYSCTL(name) \
449 (normalized_sysctl_##name = sysctl_##name / (factor))
450 WRT_SYSCTL(sched_min_granularity);
451 WRT_SYSCTL(sched_latency);
452 WRT_SYSCTL(sched_wakeup_granularity);
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100453#undef WRT_SYSCTL
454
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100455 return 0;
456}
457#endif
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200458
459/*
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200460 * delta /= w
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200461 */
462static inline unsigned long
463calc_delta_fair(unsigned long delta, struct sched_entity *se)
464{
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200465 if (unlikely(se->load.weight != NICE_0_LOAD))
466 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200467
468 return delta;
469}
470
471/*
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200472 * The idea is to set a period in which each task runs once.
473 *
474 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
475 * this period because otherwise the slices get too small.
476 *
477 * p = (nr <= nl) ? l : l*nr/nl
478 */
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200479static u64 __sched_period(unsigned long nr_running)
480{
481 u64 period = sysctl_sched_latency;
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100482 unsigned long nr_latency = sched_nr_latency;
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200483
484 if (unlikely(nr_running > nr_latency)) {
Peter Zijlstra4bf0b772008-01-25 21:08:21 +0100485 period = sysctl_sched_min_granularity;
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200486 period *= nr_running;
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200487 }
488
489 return period;
490}
491
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200492/*
493 * We calculate the wall-time slice from the period by taking a part
494 * proportional to the weight.
495 *
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200496 * s = p*P[w/rw]
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200497 */
Peter Zijlstra6d0f0eb2007-10-15 17:00:05 +0200498static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
Peter Zijlstra21805082007-08-25 18:41:53 +0200499{
Mike Galbraith0a582442009-01-02 12:16:42 +0100500 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200501
Mike Galbraith0a582442009-01-02 12:16:42 +0100502 for_each_sched_entity(se) {
Lin Ming6272d682009-01-15 17:17:15 +0100503 struct load_weight *load;
Christian Engelmayer3104bf02009-06-16 10:35:12 +0200504 struct load_weight lw;
Lin Ming6272d682009-01-15 17:17:15 +0100505
506 cfs_rq = cfs_rq_of(se);
507 load = &cfs_rq->load;
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200508
Mike Galbraith0a582442009-01-02 12:16:42 +0100509 if (unlikely(!se->on_rq)) {
Christian Engelmayer3104bf02009-06-16 10:35:12 +0200510 lw = cfs_rq->load;
Mike Galbraith0a582442009-01-02 12:16:42 +0100511
512 update_load_add(&lw, se->load.weight);
513 load = &lw;
514 }
515 slice = calc_delta_mine(slice, se->load.weight, load);
516 }
517 return slice;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200518}
519
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200520/*
Peter Zijlstraac884de2008-04-19 19:45:00 +0200521 * We calculate the vruntime slice of a to be inserted task
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200522 *
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200523 * vs = s/w
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200524 */
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200525static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200526{
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200527 return calc_delta_fair(sched_slice(cfs_rq, se), se);
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200528}
529
Paul Turnerd6b55912010-11-15 15:47:09 -0800530static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
Paul Turner6d5ab292011-01-21 20:45:01 -0800531static void update_cfs_shares(struct cfs_rq *cfs_rq);
Paul Turner3b3d1902010-11-15 15:47:08 -0800532
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200533/*
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200534 * Update the current task's runtime statistics. Skip current tasks that
535 * are not in our scheduling class.
536 */
537static inline void
Ingo Molnar8ebc91d2007-10-15 17:00:03 +0200538__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
539 unsigned long delta_exec)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200540{
Ingo Molnarbbdba7c2007-10-15 17:00:06 +0200541 unsigned long delta_exec_weighted;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200542
Lucas De Marchi41acab82010-03-10 23:37:45 -0300543 schedstat_set(curr->statistics.exec_max,
544 max((u64)delta_exec, curr->statistics.exec_max));
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200545
546 curr->sum_exec_runtime += delta_exec;
Ingo Molnar7a62eab2007-10-15 17:00:06 +0200547 schedstat_add(cfs_rq, exec_clock, delta_exec);
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200548 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
Peter Zijlstra88ec22d2009-12-16 18:04:41 +0100549
Ingo Molnare9acbff2007-10-15 17:00:04 +0200550 curr->vruntime += delta_exec_weighted;
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200551 update_min_vruntime(cfs_rq);
Paul Turner3b3d1902010-11-15 15:47:08 -0800552
Peter Zijlstra70caf8a2010-11-20 00:53:51 +0100553#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
Paul Turner3b3d1902010-11-15 15:47:08 -0800554 cfs_rq->load_unacc_exec_time += delta_exec;
Paul Turner3b3d1902010-11-15 15:47:08 -0800555#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200556}
557
Ingo Molnarb7cc0892007-08-09 11:16:47 +0200558static void update_curr(struct cfs_rq *cfs_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200559{
Ingo Molnar429d43b2007-10-15 17:00:03 +0200560 struct sched_entity *curr = cfs_rq->curr;
Venkatesh Pallipadi305e6832010-10-04 17:03:21 -0700561 u64 now = rq_of(cfs_rq)->clock_task;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200562 unsigned long delta_exec;
563
564 if (unlikely(!curr))
565 return;
566
567 /*
568 * Get the amount of time the current task was running
569 * since the last time we changed load (this cannot
570 * overflow on 32 bits):
571 */
Ingo Molnar8ebc91d2007-10-15 17:00:03 +0200572 delta_exec = (unsigned long)(now - curr->exec_start);
Peter Zijlstra34f28ec2008-12-16 08:45:31 +0100573 if (!delta_exec)
574 return;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200575
Ingo Molnar8ebc91d2007-10-15 17:00:03 +0200576 __update_curr(cfs_rq, curr, delta_exec);
577 curr->exec_start = now;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100578
579 if (entity_is_task(curr)) {
580 struct task_struct *curtask = task_of(curr);
581
Ingo Molnarf977bb42009-09-13 18:15:54 +0200582 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100583 cpuacct_charge(curtask, delta_exec);
Frank Mayharf06febc2008-09-12 09:54:39 -0700584 account_group_exec_runtime(curtask, delta_exec);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100585 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200586}
587
588static inline void
Ingo Molnar5870db52007-08-09 11:16:47 +0200589update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200590{
Lucas De Marchi41acab82010-03-10 23:37:45 -0300591 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200592}
593
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200594/*
595 * Task is being enqueued - update stats:
596 */
Ingo Molnard2417e52007-08-09 11:16:47 +0200597static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200598{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200599 /*
600 * Are we enqueueing a waiting task? (for current tasks
601 * a dequeue/enqueue event is a NOP)
602 */
Ingo Molnar429d43b2007-10-15 17:00:03 +0200603 if (se != cfs_rq->curr)
Ingo Molnar5870db52007-08-09 11:16:47 +0200604 update_stats_wait_start(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200605}
606
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200607static void
Ingo Molnar9ef0a962007-08-09 11:16:47 +0200608update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200609{
Lucas De Marchi41acab82010-03-10 23:37:45 -0300610 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
611 rq_of(cfs_rq)->clock - se->statistics.wait_start));
612 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
613 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
614 rq_of(cfs_rq)->clock - se->statistics.wait_start);
Peter Zijlstra768d0c22009-07-23 20:13:26 +0200615#ifdef CONFIG_SCHEDSTATS
616 if (entity_is_task(se)) {
617 trace_sched_stat_wait(task_of(se),
Lucas De Marchi41acab82010-03-10 23:37:45 -0300618 rq_of(cfs_rq)->clock - se->statistics.wait_start);
Peter Zijlstra768d0c22009-07-23 20:13:26 +0200619 }
620#endif
Lucas De Marchi41acab82010-03-10 23:37:45 -0300621 schedstat_set(se->statistics.wait_start, 0);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200622}
623
624static inline void
Ingo Molnar19b6a2e2007-08-09 11:16:48 +0200625update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200626{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200627 /*
628 * Mark the end of the wait period if dequeueing a
629 * waiting task:
630 */
Ingo Molnar429d43b2007-10-15 17:00:03 +0200631 if (se != cfs_rq->curr)
Ingo Molnar9ef0a962007-08-09 11:16:47 +0200632 update_stats_wait_end(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200633}
634
635/*
636 * We are picking a new current task - update its stats:
637 */
638static inline void
Ingo Molnar79303e92007-08-09 11:16:47 +0200639update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200640{
641 /*
642 * We are starting a new run period:
643 */
Venkatesh Pallipadi305e6832010-10-04 17:03:21 -0700644 se->exec_start = rq_of(cfs_rq)->clock_task;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200645}
646
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200647/**************************************************
648 * Scheduling class queueing methods:
649 */
650
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200651#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
652static void
653add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
654{
655 cfs_rq->task_weight += weight;
656}
657#else
658static inline void
659add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
660{
661}
662#endif
663
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +0200664static void
665account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
666{
667 update_load_add(&cfs_rq->load, se->load.weight);
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200668 if (!parent_entity(se))
669 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
Bharata B Raob87f1722008-09-25 09:53:54 +0530670 if (entity_is_task(se)) {
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200671 add_cfs_task_weight(cfs_rq, se->load.weight);
Bharata B Raob87f1722008-09-25 09:53:54 +0530672 list_add(&se->group_node, &cfs_rq->tasks);
673 }
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +0200674 cfs_rq->nr_running++;
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +0200675}
676
677static void
678account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
679{
680 update_load_sub(&cfs_rq->load, se->load.weight);
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200681 if (!parent_entity(se))
682 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
Bharata B Raob87f1722008-09-25 09:53:54 +0530683 if (entity_is_task(se)) {
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200684 add_cfs_task_weight(cfs_rq, -se->load.weight);
Bharata B Raob87f1722008-09-25 09:53:54 +0530685 list_del_init(&se->group_node);
686 }
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +0200687 cfs_rq->nr_running--;
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +0200688}
689
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800690#ifdef CONFIG_FAIR_GROUP_SCHED
691# ifdef CONFIG_SMP
Paul Turnerd6b55912010-11-15 15:47:09 -0800692static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
693 int global_update)
694{
695 struct task_group *tg = cfs_rq->tg;
696 long load_avg;
697
698 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
699 load_avg -= cfs_rq->load_contribution;
700
701 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
702 atomic_add(load_avg, &tg->load_weight);
703 cfs_rq->load_contribution += load_avg;
704 }
705}
706
707static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800708{
Paul Turnera7a4f8a2010-11-15 15:47:06 -0800709 u64 period = sysctl_sched_shares_window;
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800710 u64 now, delta;
Paul Turnere33078b2010-11-15 15:47:04 -0800711 unsigned long load = cfs_rq->load.weight;
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800712
Paul Turnerb815f192011-01-21 20:45:00 -0800713 if (cfs_rq->tg == &root_task_group)
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800714 return;
715
Paul Turner05ca62c2011-01-21 20:45:02 -0800716 now = rq_of(cfs_rq)->clock_task;
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800717 delta = now - cfs_rq->load_stamp;
718
Paul Turnere33078b2010-11-15 15:47:04 -0800719 /* truncate load history at 4 idle periods */
720 if (cfs_rq->load_stamp > cfs_rq->load_last &&
721 now - cfs_rq->load_last > 4 * period) {
722 cfs_rq->load_period = 0;
723 cfs_rq->load_avg = 0;
Paul Turnerf07333b2011-01-21 20:45:03 -0800724 delta = period - 1;
Paul Turnere33078b2010-11-15 15:47:04 -0800725 }
726
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800727 cfs_rq->load_stamp = now;
Paul Turner3b3d1902010-11-15 15:47:08 -0800728 cfs_rq->load_unacc_exec_time = 0;
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800729 cfs_rq->load_period += delta;
Paul Turnere33078b2010-11-15 15:47:04 -0800730 if (load) {
731 cfs_rq->load_last = now;
732 cfs_rq->load_avg += delta * load;
733 }
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800734
Paul Turnerd6b55912010-11-15 15:47:09 -0800735 /* consider updating load contribution on each fold or truncate */
736 if (global_update || cfs_rq->load_period > period
737 || !cfs_rq->load_period)
738 update_cfs_rq_load_contribution(cfs_rq, global_update);
739
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800740 while (cfs_rq->load_period > period) {
741 /*
742 * Inline assembly required to prevent the compiler
743 * optimising this loop into a divmod call.
744 * See __iter_div_u64_rem() for another example of this.
745 */
746 asm("" : "+rm" (cfs_rq->load_period));
747 cfs_rq->load_period /= 2;
748 cfs_rq->load_avg /= 2;
749 }
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800750
Paul Turnere33078b2010-11-15 15:47:04 -0800751 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
752 list_del_leaf_cfs_rq(cfs_rq);
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800753}
754
Paul Turner6d5ab292011-01-21 20:45:01 -0800755static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800756{
757 long load_weight, load, shares;
758
Paul Turner6d5ab292011-01-21 20:45:01 -0800759 load = cfs_rq->load.weight;
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800760
761 load_weight = atomic_read(&tg->load_weight);
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800762 load_weight += load;
Paul Turner6d5ab292011-01-21 20:45:01 -0800763 load_weight -= cfs_rq->load_contribution;
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800764
765 shares = (tg->shares * load);
766 if (load_weight)
767 shares /= load_weight;
768
769 if (shares < MIN_SHARES)
770 shares = MIN_SHARES;
771 if (shares > tg->shares)
772 shares = tg->shares;
773
774 return shares;
775}
776
777static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
778{
779 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
780 update_cfs_load(cfs_rq, 0);
Paul Turner6d5ab292011-01-21 20:45:01 -0800781 update_cfs_shares(cfs_rq);
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800782 }
783}
784# else /* CONFIG_SMP */
785static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
786{
787}
788
Paul Turner6d5ab292011-01-21 20:45:01 -0800789static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800790{
791 return tg->shares;
792}
793
794static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
795{
796}
797# endif /* CONFIG_SMP */
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800798static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
799 unsigned long weight)
800{
Paul Turner19e5eeb2010-12-15 19:10:18 -0800801 if (se->on_rq) {
802 /* commit outstanding execution time */
803 if (cfs_rq->curr == se)
804 update_curr(cfs_rq);
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800805 account_entity_dequeue(cfs_rq, se);
Paul Turner19e5eeb2010-12-15 19:10:18 -0800806 }
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800807
808 update_load_set(&se->load, weight);
809
810 if (se->on_rq)
811 account_entity_enqueue(cfs_rq, se);
812}
813
Paul Turner6d5ab292011-01-21 20:45:01 -0800814static void update_cfs_shares(struct cfs_rq *cfs_rq)
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800815{
816 struct task_group *tg;
817 struct sched_entity *se;
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800818 long shares;
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800819
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800820 tg = cfs_rq->tg;
821 se = tg->se[cpu_of(rq_of(cfs_rq))];
822 if (!se)
823 return;
Yong Zhang3ff6dca2011-01-24 15:33:52 +0800824#ifndef CONFIG_SMP
825 if (likely(se->load.weight == tg->shares))
826 return;
827#endif
Paul Turner6d5ab292011-01-21 20:45:01 -0800828 shares = calc_cfs_shares(cfs_rq, tg);
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800829
830 reweight_entity(cfs_rq_of(se), se, shares);
831}
832#else /* CONFIG_FAIR_GROUP_SCHED */
Paul Turnerd6b55912010-11-15 15:47:09 -0800833static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800834{
835}
836
Paul Turner6d5ab292011-01-21 20:45:01 -0800837static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800838{
839}
Paul Turner43365bd2010-12-15 19:10:17 -0800840
841static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
842{
843}
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800844#endif /* CONFIG_FAIR_GROUP_SCHED */
845
Ingo Molnar2396af62007-08-09 11:16:48 +0200846static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200847{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200848#ifdef CONFIG_SCHEDSTATS
Peter Zijlstrae4143142009-07-23 20:13:26 +0200849 struct task_struct *tsk = NULL;
850
851 if (entity_is_task(se))
852 tsk = task_of(se);
853
Lucas De Marchi41acab82010-03-10 23:37:45 -0300854 if (se->statistics.sleep_start) {
855 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200856
857 if ((s64)delta < 0)
858 delta = 0;
859
Lucas De Marchi41acab82010-03-10 23:37:45 -0300860 if (unlikely(delta > se->statistics.sleep_max))
861 se->statistics.sleep_max = delta;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200862
Lucas De Marchi41acab82010-03-10 23:37:45 -0300863 se->statistics.sleep_start = 0;
864 se->statistics.sum_sleep_runtime += delta;
Arjan van de Ven97455122008-01-25 21:08:34 +0100865
Peter Zijlstra768d0c22009-07-23 20:13:26 +0200866 if (tsk) {
Peter Zijlstrae4143142009-07-23 20:13:26 +0200867 account_scheduler_latency(tsk, delta >> 10, 1);
Peter Zijlstra768d0c22009-07-23 20:13:26 +0200868 trace_sched_stat_sleep(tsk, delta);
869 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200870 }
Lucas De Marchi41acab82010-03-10 23:37:45 -0300871 if (se->statistics.block_start) {
872 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200873
874 if ((s64)delta < 0)
875 delta = 0;
876
Lucas De Marchi41acab82010-03-10 23:37:45 -0300877 if (unlikely(delta > se->statistics.block_max))
878 se->statistics.block_max = delta;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200879
Lucas De Marchi41acab82010-03-10 23:37:45 -0300880 se->statistics.block_start = 0;
881 se->statistics.sum_sleep_runtime += delta;
Ingo Molnar30084fb2007-10-02 14:13:08 +0200882
Peter Zijlstrae4143142009-07-23 20:13:26 +0200883 if (tsk) {
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -0700884 if (tsk->in_iowait) {
Lucas De Marchi41acab82010-03-10 23:37:45 -0300885 se->statistics.iowait_sum += delta;
886 se->statistics.iowait_count++;
Peter Zijlstra768d0c22009-07-23 20:13:26 +0200887 trace_sched_stat_iowait(tsk, delta);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -0700888 }
889
Peter Zijlstrae4143142009-07-23 20:13:26 +0200890 /*
891 * Blocking time is in units of nanosecs, so shift by
892 * 20 to get a milliseconds-range estimation of the
893 * amount of time that the task spent sleeping:
894 */
895 if (unlikely(prof_on == SLEEP_PROFILING)) {
896 profile_hits(SLEEP_PROFILING,
897 (void *)get_wchan(tsk),
898 delta >> 20);
899 }
900 account_scheduler_latency(tsk, delta >> 10, 0);
Ingo Molnar30084fb2007-10-02 14:13:08 +0200901 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200902 }
903#endif
904}
905
Peter Zijlstraddc97292007-10-15 17:00:10 +0200906static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
907{
908#ifdef CONFIG_SCHED_DEBUG
909 s64 d = se->vruntime - cfs_rq->min_vruntime;
910
911 if (d < 0)
912 d = -d;
913
914 if (d > 3*sysctl_sched_latency)
915 schedstat_inc(cfs_rq, nr_spread_over);
916#endif
917}
918
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200919static void
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200920place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
921{
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200922 u64 vruntime = cfs_rq->min_vruntime;
Peter Zijlstra94dfb5e2007-10-15 17:00:05 +0200923
Peter Zijlstra2cb86002007-11-09 22:39:37 +0100924 /*
925 * The 'current' period is already promised to the current tasks,
926 * however the extra weight of the new task will slow them down a
927 * little, place the new task so that it fits in the slot that
928 * stays open at the end.
929 */
Peter Zijlstra94dfb5e2007-10-15 17:00:05 +0200930 if (initial && sched_feat(START_DEBIT))
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200931 vruntime += sched_vslice(cfs_rq, se);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200932
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +0200933 /* sleeps up to a single latency don't count. */
Mike Galbraith5ca98802010-03-11 17:17:17 +0100934 if (!initial) {
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +0200935 unsigned long thresh = sysctl_sched_latency;
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200936
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +0200937 /*
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +0200938 * Halve their sleep time's effect, to allow
939 * for a gentler effect of sleepers:
940 */
941 if (sched_feat(GENTLE_FAIR_SLEEPERS))
942 thresh >>= 1;
Ingo Molnar51e03042009-09-16 08:54:45 +0200943
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +0200944 vruntime -= thresh;
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200945 }
946
Mike Galbraithb5d9d732009-09-08 11:12:28 +0200947 /* ensure we never gain time by being placed backwards. */
948 vruntime = max_vruntime(se->vruntime, vruntime);
949
Peter Zijlstra67e9fb22007-10-15 17:00:10 +0200950 se->vruntime = vruntime;
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200951}
952
953static void
Peter Zijlstra88ec22d2009-12-16 18:04:41 +0100954enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200955{
956 /*
Peter Zijlstra88ec22d2009-12-16 18:04:41 +0100957 * Update the normalized vruntime before updating min_vruntime
958 * through callig update_curr().
959 */
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100960 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
Peter Zijlstra88ec22d2009-12-16 18:04:41 +0100961 se->vruntime += cfs_rq->min_vruntime;
962
963 /*
Dmitry Adamushkoa2a2d682007-10-15 17:00:13 +0200964 * Update run-time statistics of the 'current'.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200965 */
Ingo Molnarb7cc0892007-08-09 11:16:47 +0200966 update_curr(cfs_rq);
Paul Turnerd6b55912010-11-15 15:47:09 -0800967 update_cfs_load(cfs_rq, 0);
Peter Zijlstraa9922412008-05-05 23:56:17 +0200968 account_entity_enqueue(cfs_rq, se);
Paul Turner6d5ab292011-01-21 20:45:01 -0800969 update_cfs_shares(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200970
Peter Zijlstra88ec22d2009-12-16 18:04:41 +0100971 if (flags & ENQUEUE_WAKEUP) {
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200972 place_entity(cfs_rq, se, 0);
Ingo Molnar2396af62007-08-09 11:16:48 +0200973 enqueue_sleeper(cfs_rq, se);
Ingo Molnare9acbff2007-10-15 17:00:04 +0200974 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200975
Ingo Molnard2417e52007-08-09 11:16:47 +0200976 update_stats_enqueue(cfs_rq, se);
Peter Zijlstraddc97292007-10-15 17:00:10 +0200977 check_spread(cfs_rq, se);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +0200978 if (se != cfs_rq->curr)
979 __enqueue_entity(cfs_rq, se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -0800980 se->on_rq = 1;
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800981
982 if (cfs_rq->nr_running == 1)
983 list_add_leaf_cfs_rq(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200984}
985
Rik van Riel2c13c9192011-02-01 09:48:37 -0500986static void __clear_buddies_last(struct sched_entity *se)
Peter Zijlstra2002c692008-11-11 11:52:33 +0100987{
Rik van Riel2c13c9192011-02-01 09:48:37 -0500988 for_each_sched_entity(se) {
989 struct cfs_rq *cfs_rq = cfs_rq_of(se);
990 if (cfs_rq->last == se)
991 cfs_rq->last = NULL;
992 else
993 break;
994 }
995}
Peter Zijlstra2002c692008-11-11 11:52:33 +0100996
Rik van Riel2c13c9192011-02-01 09:48:37 -0500997static void __clear_buddies_next(struct sched_entity *se)
998{
999 for_each_sched_entity(se) {
1000 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1001 if (cfs_rq->next == se)
1002 cfs_rq->next = NULL;
1003 else
1004 break;
1005 }
Peter Zijlstra2002c692008-11-11 11:52:33 +01001006}
1007
Rik van Rielac53db52011-02-01 09:51:03 -05001008static void __clear_buddies_skip(struct sched_entity *se)
1009{
1010 for_each_sched_entity(se) {
1011 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1012 if (cfs_rq->skip == se)
1013 cfs_rq->skip = NULL;
1014 else
1015 break;
1016 }
1017}
1018
Peter Zijlstraa571bbe2009-01-28 14:51:40 +01001019static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1020{
Rik van Riel2c13c9192011-02-01 09:48:37 -05001021 if (cfs_rq->last == se)
1022 __clear_buddies_last(se);
1023
1024 if (cfs_rq->next == se)
1025 __clear_buddies_next(se);
Rik van Rielac53db52011-02-01 09:51:03 -05001026
1027 if (cfs_rq->skip == se)
1028 __clear_buddies_skip(se);
Peter Zijlstraa571bbe2009-01-28 14:51:40 +01001029}
1030
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001031static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001032dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001033{
Dmitry Adamushkoa2a2d682007-10-15 17:00:13 +02001034 /*
1035 * Update run-time statistics of the 'current'.
1036 */
1037 update_curr(cfs_rq);
1038
Ingo Molnar19b6a2e2007-08-09 11:16:48 +02001039 update_stats_dequeue(cfs_rq, se);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001040 if (flags & DEQUEUE_SLEEP) {
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02001041#ifdef CONFIG_SCHEDSTATS
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001042 if (entity_is_task(se)) {
1043 struct task_struct *tsk = task_of(se);
1044
1045 if (tsk->state & TASK_INTERRUPTIBLE)
Lucas De Marchi41acab82010-03-10 23:37:45 -03001046 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001047 if (tsk->state & TASK_UNINTERRUPTIBLE)
Lucas De Marchi41acab82010-03-10 23:37:45 -03001048 se->statistics.block_start = rq_of(cfs_rq)->clock;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001049 }
Dmitry Adamushkodb36cc72007-10-15 17:00:06 +02001050#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02001051 }
1052
Peter Zijlstra2002c692008-11-11 11:52:33 +01001053 clear_buddies(cfs_rq, se);
Peter Zijlstra47932412008-11-04 21:25:09 +01001054
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001055 if (se != cfs_rq->curr)
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001056 __dequeue_entity(cfs_rq, se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001057 se->on_rq = 0;
Paul Turnerd6b55912010-11-15 15:47:09 -08001058 update_cfs_load(cfs_rq, 0);
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001059 account_entity_dequeue(cfs_rq, se);
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001060
1061 /*
1062 * Normalize the entity after updating the min_vruntime because the
1063 * update can refer to the ->curr item and we need to reflect this
1064 * movement in our normalized position.
1065 */
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001066 if (!(flags & DEQUEUE_SLEEP))
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001067 se->vruntime -= cfs_rq->min_vruntime;
Peter Zijlstra1e876232011-05-17 16:21:10 -07001068
1069 update_min_vruntime(cfs_rq);
1070 update_cfs_shares(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001071}
1072
1073/*
1074 * Preempt the current task with a newly woken task if needed:
1075 */
Peter Zijlstra7c92e542007-09-05 14:32:49 +02001076static void
Ingo Molnar2e09bf52007-10-15 17:00:05 +02001077check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001078{
Peter Zijlstra11697832007-09-05 14:32:49 +02001079 unsigned long ideal_runtime, delta_exec;
1080
Peter Zijlstra6d0f0eb2007-10-15 17:00:05 +02001081 ideal_runtime = sched_slice(cfs_rq, curr);
Peter Zijlstra11697832007-09-05 14:32:49 +02001082 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
Mike Galbraitha9f3e2b2009-01-28 14:51:39 +01001083 if (delta_exec > ideal_runtime) {
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001084 resched_task(rq_of(cfs_rq)->curr);
Mike Galbraitha9f3e2b2009-01-28 14:51:39 +01001085 /*
1086 * The current task ran long enough, ensure it doesn't get
1087 * re-elected due to buddy favours.
1088 */
1089 clear_buddies(cfs_rq, curr);
Mike Galbraithf685cea2009-10-23 23:09:22 +02001090 return;
1091 }
1092
1093 /*
1094 * Ensure that a task that missed wakeup preemption by a
1095 * narrow margin doesn't have to wait for a full slice.
1096 * This also mitigates buddy induced latencies under load.
1097 */
1098 if (!sched_feat(WAKEUP_PREEMPT))
1099 return;
1100
1101 if (delta_exec < sysctl_sched_min_granularity)
1102 return;
1103
1104 if (cfs_rq->nr_running > 1) {
Rik van Rielac53db52011-02-01 09:51:03 -05001105 struct sched_entity *se = __pick_first_entity(cfs_rq);
Mike Galbraithf685cea2009-10-23 23:09:22 +02001106 s64 delta = curr->vruntime - se->vruntime;
1107
Mike Galbraithd7d82942011-01-05 05:41:17 +01001108 if (delta < 0)
1109 return;
1110
Mike Galbraithf685cea2009-10-23 23:09:22 +02001111 if (delta > ideal_runtime)
1112 resched_task(rq_of(cfs_rq)->curr);
Mike Galbraitha9f3e2b2009-01-28 14:51:39 +01001113 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001114}
1115
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001116static void
Ingo Molnar8494f412007-08-09 11:16:48 +02001117set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001118{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001119 /* 'current' is not kept within the tree. */
1120 if (se->on_rq) {
1121 /*
1122 * Any task has to be enqueued before it get to execute on
1123 * a CPU. So account for the time it spent waiting on the
1124 * runqueue.
1125 */
1126 update_stats_wait_end(cfs_rq, se);
1127 __dequeue_entity(cfs_rq, se);
1128 }
1129
Ingo Molnar79303e92007-08-09 11:16:47 +02001130 update_stats_curr_start(cfs_rq, se);
Ingo Molnar429d43b2007-10-15 17:00:03 +02001131 cfs_rq->curr = se;
Ingo Molnareba1ed42007-10-15 17:00:02 +02001132#ifdef CONFIG_SCHEDSTATS
1133 /*
1134 * Track our maximum slice length, if the CPU's load is at
1135 * least twice that of our own weight (i.e. dont track it
1136 * when there are only lesser-weight tasks around):
1137 */
Dmitry Adamushko495eca42007-10-15 17:00:06 +02001138 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
Lucas De Marchi41acab82010-03-10 23:37:45 -03001139 se->statistics.slice_max = max(se->statistics.slice_max,
Ingo Molnareba1ed42007-10-15 17:00:02 +02001140 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1141 }
1142#endif
Peter Zijlstra4a55b452007-09-05 14:32:49 +02001143 se->prev_sum_exec_runtime = se->sum_exec_runtime;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001144}
1145
Peter Zijlstra3f3a4902008-10-24 11:06:16 +02001146static int
1147wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1148
Rik van Rielac53db52011-02-01 09:51:03 -05001149/*
1150 * Pick the next process, keeping these things in mind, in this order:
1151 * 1) keep things fair between processes/task groups
1152 * 2) pick the "next" process, since someone really wants that to run
1153 * 3) pick the "last" process, for cache locality
1154 * 4) do not run the "skip" process, if something else is available
1155 */
Peter Zijlstraf4b67552008-11-04 21:25:07 +01001156static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01001157{
Rik van Rielac53db52011-02-01 09:51:03 -05001158 struct sched_entity *se = __pick_first_entity(cfs_rq);
Mike Galbraithf685cea2009-10-23 23:09:22 +02001159 struct sched_entity *left = se;
Peter Zijlstraf4b67552008-11-04 21:25:07 +01001160
Rik van Rielac53db52011-02-01 09:51:03 -05001161 /*
1162 * Avoid running the skip buddy, if running something else can
1163 * be done without getting too unfair.
1164 */
1165 if (cfs_rq->skip == se) {
1166 struct sched_entity *second = __pick_next_entity(se);
1167 if (second && wakeup_preempt_entity(second, left) < 1)
1168 se = second;
1169 }
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01001170
Mike Galbraithf685cea2009-10-23 23:09:22 +02001171 /*
1172 * Prefer last buddy, try to return the CPU to a preempted task.
1173 */
1174 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1175 se = cfs_rq->last;
1176
Rik van Rielac53db52011-02-01 09:51:03 -05001177 /*
1178 * Someone really wants this to run. If it's not unfair, run it.
1179 */
1180 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1181 se = cfs_rq->next;
1182
Mike Galbraithf685cea2009-10-23 23:09:22 +02001183 clear_buddies(cfs_rq, se);
Peter Zijlstra47932412008-11-04 21:25:09 +01001184
1185 return se;
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01001186}
1187
Ingo Molnarab6cde22007-08-09 11:16:48 +02001188static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001189{
1190 /*
1191 * If still on the runqueue then deactivate_task()
1192 * was not called and update_curr() has to be done:
1193 */
1194 if (prev->on_rq)
Ingo Molnarb7cc0892007-08-09 11:16:47 +02001195 update_curr(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001196
Peter Zijlstraddc97292007-10-15 17:00:10 +02001197 check_spread(cfs_rq, prev);
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001198 if (prev->on_rq) {
Ingo Molnar5870db52007-08-09 11:16:47 +02001199 update_stats_wait_start(cfs_rq, prev);
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001200 /* Put 'current' back into the tree. */
1201 __enqueue_entity(cfs_rq, prev);
1202 }
Ingo Molnar429d43b2007-10-15 17:00:03 +02001203 cfs_rq->curr = NULL;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001204}
1205
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001206static void
1207entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001208{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001209 /*
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001210 * Update run-time statistics of the 'current'.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001211 */
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02001212 update_curr(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001213
Paul Turner43365bd2010-12-15 19:10:17 -08001214 /*
1215 * Update share accounting for long-running entities.
1216 */
1217 update_entity_shares_tick(cfs_rq);
1218
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001219#ifdef CONFIG_SCHED_HRTICK
1220 /*
1221 * queued ticks are scheduled to match the slice, so don't bother
1222 * validating it and just reschedule.
1223 */
Harvey Harrison983ed7a2008-04-24 18:17:55 -07001224 if (queued) {
1225 resched_task(rq_of(cfs_rq)->curr);
1226 return;
1227 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001228 /*
1229 * don't let the period tick interfere with the hrtick preemption
1230 */
1231 if (!sched_feat(DOUBLE_TICK) &&
1232 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1233 return;
1234#endif
1235
Peter Zijlstrace6c1312007-10-15 17:00:14 +02001236 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
Ingo Molnar2e09bf52007-10-15 17:00:05 +02001237 check_preempt_tick(cfs_rq, curr);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001238}
1239
1240/**************************************************
1241 * CFS operations on tasks:
1242 */
1243
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001244#ifdef CONFIG_SCHED_HRTICK
1245static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1246{
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001247 struct sched_entity *se = &p->se;
1248 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1249
1250 WARN_ON(task_rq(p) != rq);
1251
1252 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1253 u64 slice = sched_slice(cfs_rq, se);
1254 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1255 s64 delta = slice - ran;
1256
1257 if (delta < 0) {
1258 if (rq->curr == p)
1259 resched_task(p);
1260 return;
1261 }
1262
1263 /*
1264 * Don't schedule slices shorter than 10000ns, that just
1265 * doesn't make sense. Rely on vruntime for fairness.
1266 */
Peter Zijlstra31656512008-07-18 18:01:23 +02001267 if (rq->curr != p)
Peter Zijlstra157124c2008-07-28 11:53:11 +02001268 delta = max_t(s64, 10000LL, delta);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001269
Peter Zijlstra31656512008-07-18 18:01:23 +02001270 hrtick_start(rq, delta);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001271 }
1272}
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02001273
1274/*
1275 * called from enqueue/dequeue and updates the hrtick when the
1276 * current task is from our class and nr_running is low enough
1277 * to matter.
1278 */
1279static void hrtick_update(struct rq *rq)
1280{
1281 struct task_struct *curr = rq->curr;
1282
1283 if (curr->sched_class != &fair_sched_class)
1284 return;
1285
1286 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1287 hrtick_start_fair(rq, curr);
1288}
Dhaval Giani55e12e52008-06-24 23:39:43 +05301289#else /* !CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001290static inline void
1291hrtick_start_fair(struct rq *rq, struct task_struct *p)
1292{
1293}
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02001294
1295static inline void hrtick_update(struct rq *rq)
1296{
1297}
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001298#endif
1299
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001300/*
1301 * The enqueue_task method is called before nr_running is
1302 * increased. Here we update the fair scheduling stats and
1303 * then put the task into the rbtree:
1304 */
Thomas Gleixnerea87bb72010-01-20 20:58:57 +00001305static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001306enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001307{
1308 struct cfs_rq *cfs_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001309 struct sched_entity *se = &p->se;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001310
1311 for_each_sched_entity(se) {
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001312 if (se->on_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001313 break;
1314 cfs_rq = cfs_rq_of(se);
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001315 enqueue_entity(cfs_rq, se, flags);
1316 flags = ENQUEUE_WAKEUP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001317 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001318
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001319 for_each_sched_entity(se) {
Lin Ming0f317142011-07-22 09:14:31 +08001320 cfs_rq = cfs_rq_of(se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001321
Paul Turnerd6b55912010-11-15 15:47:09 -08001322 update_cfs_load(cfs_rq, 0);
Paul Turner6d5ab292011-01-21 20:45:01 -08001323 update_cfs_shares(cfs_rq);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001324 }
1325
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02001326 hrtick_update(rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001327}
1328
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001329static void set_next_buddy(struct sched_entity *se);
1330
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001331/*
1332 * The dequeue_task method is called before nr_running is
1333 * decreased. We remove the task from the rbtree and
1334 * update the fair scheduling stats:
1335 */
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001336static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001337{
1338 struct cfs_rq *cfs_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01001339 struct sched_entity *se = &p->se;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001340 int task_sleep = flags & DEQUEUE_SLEEP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001341
1342 for_each_sched_entity(se) {
1343 cfs_rq = cfs_rq_of(se);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001344 dequeue_entity(cfs_rq, se, flags);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001345
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001346 /* Don't dequeue parent if it has other entities besides us */
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001347 if (cfs_rq->load.weight) {
1348 /*
1349 * Bias pick_next to pick a task from this cfs_rq, as
1350 * p is sleeping when it is within its sched_slice.
1351 */
1352 if (task_sleep && parent_entity(se))
1353 set_next_buddy(parent_entity(se));
Paul Turner9598c822011-07-06 22:30:37 -07001354
1355 /* avoid re-evaluating load for this entity */
1356 se = parent_entity(se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001357 break;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001358 }
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001359 flags |= DEQUEUE_SLEEP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001360 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001361
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001362 for_each_sched_entity(se) {
Lin Ming0f317142011-07-22 09:14:31 +08001363 cfs_rq = cfs_rq_of(se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001364
Paul Turnerd6b55912010-11-15 15:47:09 -08001365 update_cfs_load(cfs_rq, 0);
Paul Turner6d5ab292011-01-21 20:45:01 -08001366 update_cfs_shares(cfs_rq);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001367 }
1368
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02001369 hrtick_update(rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001370}
1371
Gregory Haskinse7693a32008-01-25 21:08:09 +01001372#ifdef CONFIG_SMP
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001373
Peter Zijlstra74f8e4b2011-04-05 17:23:47 +02001374static void task_waking_fair(struct task_struct *p)
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001375{
1376 struct sched_entity *se = &p->se;
1377 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Peter Zijlstra3fe16982011-04-05 17:23:48 +02001378 u64 min_vruntime;
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001379
Peter Zijlstra3fe16982011-04-05 17:23:48 +02001380#ifndef CONFIG_64BIT
1381 u64 min_vruntime_copy;
Peter Zijlstra74f8e4b2011-04-05 17:23:47 +02001382
Peter Zijlstra3fe16982011-04-05 17:23:48 +02001383 do {
1384 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1385 smp_rmb();
1386 min_vruntime = cfs_rq->min_vruntime;
1387 } while (min_vruntime != min_vruntime_copy);
1388#else
1389 min_vruntime = cfs_rq->min_vruntime;
1390#endif
1391
1392 se->vruntime -= min_vruntime;
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01001393}
1394
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001395#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstraf5bfb7d2008-06-27 13:41:39 +02001396/*
1397 * effective_load() calculates the load change as seen from the root_task_group
1398 *
1399 * Adding load to a group doesn't make a group heavier, but can cause movement
1400 * of group shares between cpus. Assuming the shares were perfectly aligned one
1401 * can calculate the shift in shares.
Peter Zijlstraf5bfb7d2008-06-27 13:41:39 +02001402 */
Peter Zijlstra2069dd72010-11-15 15:47:00 -08001403static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001404{
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001405 struct sched_entity *se = tg->se[cpu];
Peter Zijlstraf1d239f2008-06-27 13:41:38 +02001406
1407 if (!tg->parent)
1408 return wl;
1409
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001410 for_each_sched_entity(se) {
Paul Turner977dda72011-01-14 17:57:50 -08001411 long lw, w;
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001412
Paul Turner977dda72011-01-14 17:57:50 -08001413 tg = se->my_q->tg;
1414 w = se->my_q->load.weight;
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001415
Paul Turner977dda72011-01-14 17:57:50 -08001416 /* use this cpu's instantaneous contribution */
1417 lw = atomic_read(&tg->load_weight);
1418 lw -= se->my_q->load_contribution;
1419 lw += w + wg;
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001420
Paul Turner977dda72011-01-14 17:57:50 -08001421 wl += w;
Peter Zijlstra940959e2008-09-23 15:33:42 +02001422
Paul Turner977dda72011-01-14 17:57:50 -08001423 if (lw > 0 && wl < lw)
1424 wl = (wl * tg->shares) / lw;
1425 else
1426 wl = tg->shares;
Peter Zijlstra940959e2008-09-23 15:33:42 +02001427
Paul Turner977dda72011-01-14 17:57:50 -08001428 /* zero point is MIN_SHARES */
1429 if (wl < MIN_SHARES)
1430 wl = MIN_SHARES;
1431 wl -= se->load.weight;
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001432 wg = 0;
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001433 }
1434
1435 return wl;
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001436}
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001437
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001438#else
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001439
Peter Zijlstra83378262008-06-27 13:41:37 +02001440static inline unsigned long effective_load(struct task_group *tg, int cpu,
1441 unsigned long wl, unsigned long wg)
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001442{
Peter Zijlstra83378262008-06-27 13:41:37 +02001443 return wl;
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001444}
Peter Zijlstra4be9daa2008-06-27 13:41:30 +02001445
Peter Zijlstrabb3469a2008-06-27 13:41:27 +02001446#endif
1447
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001448static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001449{
Paul Turnere37b6a72011-01-21 20:44:59 -08001450 s64 this_load, load;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001451 int idx, this_cpu, prev_cpu;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001452 unsigned long tl_per_task;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001453 struct task_group *tg;
Peter Zijlstra83378262008-06-27 13:41:37 +02001454 unsigned long weight;
Mike Galbraithb3137bc2008-05-29 11:11:41 +02001455 int balanced;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001456
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001457 idx = sd->wake_idx;
1458 this_cpu = smp_processor_id();
1459 prev_cpu = task_cpu(p);
1460 load = source_load(prev_cpu, idx);
1461 this_load = target_load(this_cpu, idx);
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001462
1463 /*
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001464 * If sync wakeup then subtract the (maximum possible)
1465 * effect of the currently running task from the load
1466 * of the current CPU:
1467 */
Peter Zijlstra83378262008-06-27 13:41:37 +02001468 if (sync) {
1469 tg = task_group(current);
1470 weight = current->se.load.weight;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001471
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001472 this_load += effective_load(tg, this_cpu, -weight, -weight);
Peter Zijlstra83378262008-06-27 13:41:37 +02001473 load += effective_load(tg, prev_cpu, 0, -weight);
1474 }
1475
1476 tg = task_group(p);
1477 weight = p->se.load.weight;
1478
Peter Zijlstra71a29aa2009-09-07 18:28:05 +02001479 /*
1480 * In low-load situations, where prev_cpu is idle and this_cpu is idle
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001481 * due to the sync cause above having dropped this_load to 0, we'll
1482 * always have an imbalance, but there's really nothing you can do
1483 * about that, so that's good too.
Peter Zijlstra71a29aa2009-09-07 18:28:05 +02001484 *
1485 * Otherwise check if either cpus are near enough in load to allow this
1486 * task to be woken on this_cpu.
1487 */
Paul Turnere37b6a72011-01-21 20:44:59 -08001488 if (this_load > 0) {
1489 s64 this_eff_load, prev_eff_load;
Peter Zijlstrae51fd5e2010-05-31 12:37:30 +02001490
1491 this_eff_load = 100;
1492 this_eff_load *= power_of(prev_cpu);
1493 this_eff_load *= this_load +
1494 effective_load(tg, this_cpu, weight, weight);
1495
1496 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1497 prev_eff_load *= power_of(this_cpu);
1498 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1499
1500 balanced = this_eff_load <= prev_eff_load;
1501 } else
1502 balanced = true;
Mike Galbraithb3137bc2008-05-29 11:11:41 +02001503
1504 /*
1505 * If the currently running task will sleep within
1506 * a reasonable amount of time then attract this newly
1507 * woken task:
1508 */
Peter Zijlstra2fb76352008-10-08 09:16:04 +02001509 if (sync && balanced)
1510 return 1;
Mike Galbraithb3137bc2008-05-29 11:11:41 +02001511
Lucas De Marchi41acab82010-03-10 23:37:45 -03001512 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
Mike Galbraithb3137bc2008-05-29 11:11:41 +02001513 tl_per_task = cpu_avg_load_per_task(this_cpu);
1514
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001515 if (balanced ||
1516 (this_load <= load &&
1517 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001518 /*
1519 * This domain has SD_WAKE_AFFINE and
1520 * p is cache cold in this domain, and
1521 * there is no bad imbalance.
1522 */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001523 schedstat_inc(sd, ttwu_move_affine);
Lucas De Marchi41acab82010-03-10 23:37:45 -03001524 schedstat_inc(p, se.statistics.nr_wakeups_affine);
Ingo Molnar098fb9d2008-03-16 20:36:10 +01001525
1526 return 1;
1527 }
1528 return 0;
1529}
1530
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001531/*
1532 * find_idlest_group finds and returns the least busy CPU group within the
1533 * domain.
1534 */
1535static struct sched_group *
Peter Zijlstra78e7ed52009-09-03 13:16:51 +02001536find_idlest_group(struct sched_domain *sd, struct task_struct *p,
Peter Zijlstra5158f4e2009-09-16 13:46:59 +02001537 int this_cpu, int load_idx)
Gregory Haskinse7693a32008-01-25 21:08:09 +01001538{
Andi Kleenb3bd3de2010-08-10 14:17:51 -07001539 struct sched_group *idlest = NULL, *group = sd->groups;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001540 unsigned long min_load = ULONG_MAX, this_load = 0;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001541 int imbalance = 100 + (sd->imbalance_pct-100)/2;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001542
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001543 do {
1544 unsigned long load, avg_load;
1545 int local_group;
1546 int i;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001547
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001548 /* Skip over this group if it has no CPUs allowed */
1549 if (!cpumask_intersects(sched_group_cpus(group),
1550 &p->cpus_allowed))
1551 continue;
1552
1553 local_group = cpumask_test_cpu(this_cpu,
1554 sched_group_cpus(group));
1555
1556 /* Tally up the load of all CPUs in the group */
1557 avg_load = 0;
1558
1559 for_each_cpu(i, sched_group_cpus(group)) {
1560 /* Bias balancing toward cpus of our domain */
1561 if (local_group)
1562 load = source_load(i, load_idx);
1563 else
1564 load = target_load(i, load_idx);
1565
1566 avg_load += load;
1567 }
1568
1569 /* Adjust by relative CPU power of the group */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02001570 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001571
1572 if (local_group) {
1573 this_load = avg_load;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001574 } else if (avg_load < min_load) {
1575 min_load = avg_load;
1576 idlest = group;
1577 }
1578 } while (group = group->next, group != sd->groups);
1579
1580 if (!idlest || 100*this_load < imbalance*min_load)
1581 return NULL;
1582 return idlest;
1583}
1584
1585/*
1586 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1587 */
1588static int
1589find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1590{
1591 unsigned long load, min_load = ULONG_MAX;
1592 int idlest = -1;
1593 int i;
1594
1595 /* Traverse only the allowed CPUs */
1596 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1597 load = weighted_cpuload(i);
1598
1599 if (load < min_load || (load == min_load && i == this_cpu)) {
1600 min_load = load;
1601 idlest = i;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001602 }
1603 }
1604
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001605 return idlest;
1606}
Gregory Haskinse7693a32008-01-25 21:08:09 +01001607
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001608/*
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001609 * Try and locate an idle CPU in the sched_domain.
1610 */
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001611static int select_idle_sibling(struct task_struct *p, int target)
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001612{
1613 int cpu = smp_processor_id();
1614 int prev_cpu = task_cpu(p);
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001615 struct sched_domain *sd;
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001616 int i;
1617
1618 /*
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001619 * If the task is going to be woken-up on this cpu and if it is
1620 * already idle, then it is the right target.
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001621 */
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001622 if (target == cpu && idle_cpu(cpu))
1623 return cpu;
1624
1625 /*
1626 * If the task is going to be woken-up on the cpu where it previously
1627 * ran and if it is currently idle, then it the right target.
1628 */
1629 if (target == prev_cpu && idle_cpu(prev_cpu))
Peter Zijlstrafe3bcfe2009-11-12 15:55:29 +01001630 return prev_cpu;
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001631
1632 /*
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001633 * Otherwise, iterate the domains and find an elegible idle cpu.
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001634 */
Peter Zijlstradce840a2011-04-07 14:09:50 +02001635 rcu_read_lock();
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001636 for_each_domain(target, sd) {
1637 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
Peter Zijlstrafe3bcfe2009-11-12 15:55:29 +01001638 break;
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001639
1640 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1641 if (idle_cpu(i)) {
1642 target = i;
1643 break;
1644 }
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001645 }
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001646
1647 /*
1648 * Lets stop looking for an idle sibling when we reached
1649 * the domain that spans the current cpu and prev_cpu.
1650 */
1651 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1652 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1653 break;
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001654 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02001655 rcu_read_unlock();
Peter Zijlstraa50bde52009-11-12 15:55:28 +01001656
1657 return target;
1658}
1659
1660/*
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001661 * sched_balance_self: balance the current task (running on cpu) in domains
1662 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1663 * SD_BALANCE_EXEC.
1664 *
1665 * Balance, ie. select the least loaded group.
1666 *
1667 * Returns the target CPU number, or the same CPU if no balancing is needed.
1668 *
1669 * preempt must be disabled.
1670 */
Peter Zijlstra0017d732010-03-24 18:34:10 +01001671static int
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001672select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001673{
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02001674 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001675 int cpu = smp_processor_id();
1676 int prev_cpu = task_cpu(p);
1677 int new_cpu = cpu;
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001678 int want_affine = 0;
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02001679 int want_sd = 1;
Peter Zijlstra5158f4e2009-09-16 13:46:59 +02001680 int sync = wake_flags & WF_SYNC;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001681
Peter Zijlstra0763a662009-09-14 19:37:39 +02001682 if (sd_flag & SD_BALANCE_WAKE) {
Mike Galbraithbeac4c72010-03-11 17:17:20 +01001683 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001684 want_affine = 1;
1685 new_cpu = prev_cpu;
1686 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01001687
Peter Zijlstradce840a2011-04-07 14:09:50 +02001688 rcu_read_lock();
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001689 for_each_domain(cpu, tmp) {
Peter Zijlstrae4f42882009-12-16 18:04:34 +01001690 if (!(tmp->flags & SD_LOAD_BALANCE))
1691 continue;
1692
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001693 /*
Peter Zijlstraae154be2009-09-10 14:40:57 +02001694 * If power savings logic is enabled for a domain, see if we
1695 * are not overloaded, if so, don't balance wider.
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001696 */
Peter Zijlstra59abf022009-09-16 08:28:30 +02001697 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
Peter Zijlstraae154be2009-09-10 14:40:57 +02001698 unsigned long power = 0;
1699 unsigned long nr_running = 0;
1700 unsigned long capacity;
1701 int i;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001702
Peter Zijlstraae154be2009-09-10 14:40:57 +02001703 for_each_cpu(i, sched_domain_span(tmp)) {
1704 power += power_of(i);
1705 nr_running += cpu_rq(i)->cfs.nr_running;
1706 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01001707
Nikhil Rao1399fa72011-05-18 10:09:39 -07001708 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01001709
Peter Zijlstra59abf022009-09-16 08:28:30 +02001710 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1711 nr_running /= 2;
1712
1713 if (nr_running < capacity)
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02001714 want_sd = 0;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001715 }
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001716
Peter Zijlstrafe3bcfe2009-11-12 15:55:29 +01001717 /*
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001718 * If both cpu and prev_cpu are part of this domain,
1719 * cpu is a valid SD_WAKE_AFFINE target.
Peter Zijlstrafe3bcfe2009-11-12 15:55:29 +01001720 */
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001721 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1722 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1723 affine_sd = tmp;
1724 want_affine = 0;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001725 }
1726
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02001727 if (!want_sd && !want_affine)
1728 break;
1729
Peter Zijlstra0763a662009-09-14 19:37:39 +02001730 if (!(tmp->flags & sd_flag))
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001731 continue;
1732
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02001733 if (want_sd)
1734 sd = tmp;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001735 }
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001736
Mike Galbraith8b911ac2010-03-11 17:17:16 +01001737 if (affine_sd) {
Suresh Siddha99bd5e22010-03-31 16:47:45 -07001738 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
Peter Zijlstradce840a2011-04-07 14:09:50 +02001739 prev_cpu = cpu;
1740
1741 new_cpu = select_idle_sibling(p, prev_cpu);
1742 goto unlock;
Mike Galbraith8b911ac2010-03-11 17:17:16 +01001743 }
Peter Zijlstra3b640892009-09-16 13:44:33 +02001744
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001745 while (sd) {
Peter Zijlstra5158f4e2009-09-16 13:46:59 +02001746 int load_idx = sd->forkexec_idx;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001747 struct sched_group *group;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001748 int weight;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001749
Peter Zijlstra0763a662009-09-14 19:37:39 +02001750 if (!(sd->flags & sd_flag)) {
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001751 sd = sd->child;
1752 continue;
1753 }
1754
Peter Zijlstra5158f4e2009-09-16 13:46:59 +02001755 if (sd_flag & SD_BALANCE_WAKE)
1756 load_idx = sd->wake_idx;
1757
1758 group = find_idlest_group(sd, p, cpu, load_idx);
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001759 if (!group) {
1760 sd = sd->child;
1761 continue;
1762 }
1763
Peter Zijlstrad7c33c42009-09-11 12:45:38 +02001764 new_cpu = find_idlest_cpu(group, p, cpu);
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001765 if (new_cpu == -1 || new_cpu == cpu) {
1766 /* Now try balancing at a lower domain level of cpu */
1767 sd = sd->child;
1768 continue;
1769 }
1770
1771 /* Now try balancing at a lower domain level of new_cpu */
1772 cpu = new_cpu;
Peter Zijlstra669c55e2010-04-16 14:59:29 +02001773 weight = sd->span_weight;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001774 sd = NULL;
1775 for_each_domain(cpu, tmp) {
Peter Zijlstra669c55e2010-04-16 14:59:29 +02001776 if (weight <= tmp->span_weight)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001777 break;
Peter Zijlstra0763a662009-09-14 19:37:39 +02001778 if (tmp->flags & sd_flag)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02001779 sd = tmp;
1780 }
1781 /* while loop will break here if sd == NULL */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001782 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02001783unlock:
1784 rcu_read_unlock();
Gregory Haskinse7693a32008-01-25 21:08:09 +01001785
Peter Zijlstrac88d5912009-09-10 13:50:02 +02001786 return new_cpu;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001787}
1788#endif /* CONFIG_SMP */
1789
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01001790static unsigned long
1791wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02001792{
1793 unsigned long gran = sysctl_sched_wakeup_granularity;
1794
1795 /*
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01001796 * Since its curr running now, convert the gran from real-time
1797 * to virtual-time in his units.
Mike Galbraith13814d42010-03-11 17:17:04 +01001798 *
1799 * By using 'se' instead of 'curr' we penalize light tasks, so
1800 * they get preempted easier. That is, if 'se' < 'curr' then
1801 * the resulting gran will be larger, therefore penalizing the
1802 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1803 * be smaller, again penalizing the lighter task.
1804 *
1805 * This is especially important for buddies when the leftmost
1806 * task is higher priority than the buddy.
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02001807 */
Shaohua Lif4ad9bd2011-04-08 12:53:09 +08001808 return calc_delta_fair(gran, se);
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02001809}
1810
1811/*
Peter Zijlstra464b7522008-10-24 11:06:15 +02001812 * Should 'se' preempt 'curr'.
1813 *
1814 * |s1
1815 * |s2
1816 * |s3
1817 * g
1818 * |<--->|c
1819 *
1820 * w(c, s1) = -1
1821 * w(c, s2) = 0
1822 * w(c, s3) = 1
1823 *
1824 */
1825static int
1826wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1827{
1828 s64 gran, vdiff = curr->vruntime - se->vruntime;
1829
1830 if (vdiff <= 0)
1831 return -1;
1832
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01001833 gran = wakeup_gran(curr, se);
Peter Zijlstra464b7522008-10-24 11:06:15 +02001834 if (vdiff > gran)
1835 return 1;
1836
1837 return 0;
1838}
1839
Peter Zijlstra02479092008-11-04 21:25:10 +01001840static void set_last_buddy(struct sched_entity *se)
1841{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07001842 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1843 return;
1844
1845 for_each_sched_entity(se)
1846 cfs_rq_of(se)->last = se;
Peter Zijlstra02479092008-11-04 21:25:10 +01001847}
1848
1849static void set_next_buddy(struct sched_entity *se)
1850{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07001851 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1852 return;
1853
1854 for_each_sched_entity(se)
1855 cfs_rq_of(se)->next = se;
Peter Zijlstra02479092008-11-04 21:25:10 +01001856}
1857
Rik van Rielac53db52011-02-01 09:51:03 -05001858static void set_skip_buddy(struct sched_entity *se)
1859{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07001860 for_each_sched_entity(se)
1861 cfs_rq_of(se)->skip = se;
Rik van Rielac53db52011-02-01 09:51:03 -05001862}
1863
Peter Zijlstra464b7522008-10-24 11:06:15 +02001864/*
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001865 * Preempt the current task with a newly woken task if needed:
1866 */
Peter Zijlstra5a9b86f2009-09-16 13:47:58 +02001867static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001868{
1869 struct task_struct *curr = rq->curr;
Srivatsa Vaddagiri8651a862007-10-15 17:00:12 +02001870 struct sched_entity *se = &curr->se, *pse = &p->se;
Mike Galbraith03e89e42008-12-16 08:45:30 +01001871 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
Mike Galbraithf685cea2009-10-23 23:09:22 +02001872 int scale = cfs_rq->nr_running >= sched_nr_latency;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001873 int next_buddy_marked = 0;
Mike Galbraith03e89e42008-12-16 08:45:30 +01001874
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01001875 if (unlikely(se == pse))
1876 return;
1877
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001878 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
Mike Galbraith3cb63d52009-09-11 12:01:17 +02001879 set_next_buddy(pse);
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001880 next_buddy_marked = 1;
1881 }
Peter Zijlstra57fdc262008-09-23 15:33:45 +02001882
Bharata B Raoaec0a512008-08-28 14:42:49 +05301883 /*
1884 * We can come here with TIF_NEED_RESCHED already set from new task
1885 * wake up path.
1886 */
1887 if (test_tsk_need_resched(curr))
1888 return;
1889
Darren Harta2f5c9a2011-02-22 13:04:33 -08001890 /* Idle tasks are by definition preempted by non-idle tasks. */
1891 if (unlikely(curr->policy == SCHED_IDLE) &&
1892 likely(p->policy != SCHED_IDLE))
1893 goto preempt;
1894
Ingo Molnar91c234b2007-10-15 17:00:18 +02001895 /*
Darren Harta2f5c9a2011-02-22 13:04:33 -08001896 * Batch and idle tasks do not preempt non-idle tasks (their preemption
1897 * is driven by the tick):
Ingo Molnar91c234b2007-10-15 17:00:18 +02001898 */
Peter Zijlstra6bc912b2009-01-15 14:53:38 +01001899 if (unlikely(p->policy != SCHED_NORMAL))
Ingo Molnar91c234b2007-10-15 17:00:18 +02001900 return;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001901
Peter Zijlstra6bc912b2009-01-15 14:53:38 +01001902
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02001903 if (!sched_feat(WAKEUP_PREEMPT))
1904 return;
1905
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01001906 find_matching_se(&se, &pse);
Paul Turner9bbd7372011-07-05 19:07:21 -07001907 update_curr(cfs_rq_of(se));
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01001908 BUG_ON(!pse);
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001909 if (wakeup_preempt_entity(se, pse) == 1) {
1910 /*
1911 * Bias pick_next to pick the sched entity that is
1912 * triggering this preemption.
1913 */
1914 if (!next_buddy_marked)
1915 set_next_buddy(pse);
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01001916 goto preempt;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07001917 }
Jupyung Leea65ac742009-11-17 18:51:40 +09001918
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01001919 return;
1920
1921preempt:
1922 resched_task(curr);
1923 /*
1924 * Only set the backward buddy when the current task is still
1925 * on the rq. This can happen when a wakeup gets interleaved
1926 * with schedule on the ->pre_schedule() or idle_balance()
1927 * point, either of which can * drop the rq lock.
1928 *
1929 * Also, during early boot the idle thread is in the fair class,
1930 * for obvious reasons its a bad idea to schedule back to it.
1931 */
1932 if (unlikely(!se->on_rq || curr == rq->idle))
1933 return;
1934
1935 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1936 set_last_buddy(se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001937}
1938
Ingo Molnarfb8d4722007-08-09 11:16:48 +02001939static struct task_struct *pick_next_task_fair(struct rq *rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001940{
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001941 struct task_struct *p;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001942 struct cfs_rq *cfs_rq = &rq->cfs;
1943 struct sched_entity *se;
1944
Tim Blechmann36ace272009-11-24 11:55:45 +01001945 if (!cfs_rq->nr_running)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001946 return NULL;
1947
1948 do {
Ingo Molnar9948f4b2007-08-09 11:16:48 +02001949 se = pick_next_entity(cfs_rq);
Peter Zijlstraf4b67552008-11-04 21:25:07 +01001950 set_next_entity(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001951 cfs_rq = group_cfs_rq(se);
1952 } while (cfs_rq);
1953
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001954 p = task_of(se);
1955 hrtick_start_fair(rq, p);
1956
1957 return p;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001958}
1959
1960/*
1961 * Account for a descheduled task:
1962 */
Ingo Molnar31ee5292007-08-09 11:16:49 +02001963static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001964{
1965 struct sched_entity *se = &prev->se;
1966 struct cfs_rq *cfs_rq;
1967
1968 for_each_sched_entity(se) {
1969 cfs_rq = cfs_rq_of(se);
Ingo Molnarab6cde22007-08-09 11:16:48 +02001970 put_prev_entity(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001971 }
1972}
1973
Rik van Rielac53db52011-02-01 09:51:03 -05001974/*
1975 * sched_yield() is very simple
1976 *
1977 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1978 */
1979static void yield_task_fair(struct rq *rq)
1980{
1981 struct task_struct *curr = rq->curr;
1982 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1983 struct sched_entity *se = &curr->se;
1984
1985 /*
1986 * Are we the only task in the tree?
1987 */
1988 if (unlikely(rq->nr_running == 1))
1989 return;
1990
1991 clear_buddies(cfs_rq, se);
1992
1993 if (curr->policy != SCHED_BATCH) {
1994 update_rq_clock(rq);
1995 /*
1996 * Update run-time statistics of the 'current'.
1997 */
1998 update_curr(cfs_rq);
1999 }
2000
2001 set_skip_buddy(se);
2002}
2003
Mike Galbraithd95f4122011-02-01 09:50:51 -05002004static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2005{
2006 struct sched_entity *se = &p->se;
2007
2008 if (!se->on_rq)
2009 return false;
2010
2011 /* Tell the scheduler that we'd really like pse to run next. */
2012 set_next_buddy(se);
2013
Mike Galbraithd95f4122011-02-01 09:50:51 -05002014 yield_task_fair(rq);
2015
2016 return true;
2017}
2018
Peter Williams681f3e62007-10-24 18:23:51 +02002019#ifdef CONFIG_SMP
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02002020/**************************************************
2021 * Fair scheduling class load-balancing methods:
2022 */
2023
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002024/*
2025 * pull_task - move a task from a remote runqueue to the local runqueue.
2026 * Both runqueues must be locked.
2027 */
2028static void pull_task(struct rq *src_rq, struct task_struct *p,
2029 struct rq *this_rq, int this_cpu)
2030{
2031 deactivate_task(src_rq, p, 0);
2032 set_task_cpu(p, this_cpu);
2033 activate_task(this_rq, p, 0);
2034 check_preempt_curr(this_rq, p, 0);
2035}
2036
2037/*
2038 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2039 */
2040static
2041int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2042 struct sched_domain *sd, enum cpu_idle_type idle,
2043 int *all_pinned)
2044{
2045 int tsk_cache_hot = 0;
2046 /*
2047 * We do not migrate tasks that are:
2048 * 1) running (obviously), or
2049 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2050 * 3) are cache-hot on their current CPU.
2051 */
2052 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
Lucas De Marchi41acab82010-03-10 23:37:45 -03002053 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002054 return 0;
2055 }
2056 *all_pinned = 0;
2057
2058 if (task_running(rq, p)) {
Lucas De Marchi41acab82010-03-10 23:37:45 -03002059 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002060 return 0;
2061 }
2062
2063 /*
2064 * Aggressive migration if:
2065 * 1) task is cache cold, or
2066 * 2) too many balance attempts have failed.
2067 */
2068
Venkatesh Pallipadi305e6832010-10-04 17:03:21 -07002069 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002070 if (!tsk_cache_hot ||
2071 sd->nr_balance_failed > sd->cache_nice_tries) {
2072#ifdef CONFIG_SCHEDSTATS
2073 if (tsk_cache_hot) {
2074 schedstat_inc(sd, lb_hot_gained[idle]);
Lucas De Marchi41acab82010-03-10 23:37:45 -03002075 schedstat_inc(p, se.statistics.nr_forced_migrations);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002076 }
2077#endif
2078 return 1;
2079 }
2080
2081 if (tsk_cache_hot) {
Lucas De Marchi41acab82010-03-10 23:37:45 -03002082 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002083 return 0;
2084 }
2085 return 1;
2086}
2087
Peter Zijlstra897c3952009-12-17 17:45:42 +01002088/*
2089 * move_one_task tries to move exactly one task from busiest to this_rq, as
2090 * part of active balancing operations within "domain".
2091 * Returns 1 if successful and 0 otherwise.
2092 *
2093 * Called with both runqueues locked.
2094 */
2095static int
2096move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2097 struct sched_domain *sd, enum cpu_idle_type idle)
2098{
2099 struct task_struct *p, *n;
2100 struct cfs_rq *cfs_rq;
2101 int pinned = 0;
2102
2103 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2104 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2105
2106 if (!can_migrate_task(p, busiest, this_cpu,
2107 sd, idle, &pinned))
2108 continue;
2109
2110 pull_task(busiest, p, this_rq, this_cpu);
2111 /*
2112 * Right now, this is only the second place pull_task()
2113 * is called, so we can safely collect pull_task()
2114 * stats here rather than inside pull_task().
2115 */
2116 schedstat_inc(sd, lb_gained[idle]);
2117 return 1;
2118 }
2119 }
2120
2121 return 0;
2122}
2123
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002124static unsigned long
2125balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2126 unsigned long max_load_move, struct sched_domain *sd,
2127 enum cpu_idle_type idle, int *all_pinned,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002128 struct cfs_rq *busiest_cfs_rq)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002129{
Ken Chenb30aef12011-04-08 12:20:16 -07002130 int loops = 0, pulled = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002131 long rem_load_move = max_load_move;
Peter Zijlstraee00e662009-12-17 17:25:20 +01002132 struct task_struct *p, *n;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002133
2134 if (max_load_move == 0)
2135 goto out;
2136
Peter Zijlstraee00e662009-12-17 17:25:20 +01002137 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2138 if (loops++ > sysctl_sched_nr_migrate)
2139 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002140
Peter Zijlstraee00e662009-12-17 17:25:20 +01002141 if ((p->se.load.weight >> 1) > rem_load_move ||
Ken Chenb30aef12011-04-08 12:20:16 -07002142 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2143 all_pinned))
Peter Zijlstraee00e662009-12-17 17:25:20 +01002144 continue;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002145
Peter Zijlstraee00e662009-12-17 17:25:20 +01002146 pull_task(busiest, p, this_rq, this_cpu);
2147 pulled++;
2148 rem_load_move -= p->se.load.weight;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002149
2150#ifdef CONFIG_PREEMPT
Peter Zijlstraee00e662009-12-17 17:25:20 +01002151 /*
2152 * NEWIDLE balancing is a source of latency, so preemptible
2153 * kernels will stop after the first task is pulled to minimize
2154 * the critical section.
2155 */
2156 if (idle == CPU_NEWLY_IDLE)
2157 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002158#endif
2159
Peter Zijlstraee00e662009-12-17 17:25:20 +01002160 /*
2161 * We only want to steal up to the prescribed amount of
2162 * weighted load.
2163 */
2164 if (rem_load_move <= 0)
2165 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002166 }
2167out:
2168 /*
2169 * Right now, this is one of only two places pull_task() is called,
2170 * so we can safely collect pull_task() stats here rather than
2171 * inside pull_task().
2172 */
2173 schedstat_add(sd, lb_gained[idle], pulled);
2174
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002175 return max_load_move - rem_load_move;
2176}
2177
Peter Zijlstra230059de2009-12-17 17:47:12 +01002178#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002179/*
2180 * update tg->load_weight by folding this cpu's load_avg
2181 */
Paul Turner67e86252010-11-15 15:47:05 -08002182static int update_shares_cpu(struct task_group *tg, int cpu)
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002183{
2184 struct cfs_rq *cfs_rq;
2185 unsigned long flags;
2186 struct rq *rq;
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002187
2188 if (!tg->se[cpu])
2189 return 0;
2190
2191 rq = cpu_rq(cpu);
2192 cfs_rq = tg->cfs_rq[cpu];
2193
2194 raw_spin_lock_irqsave(&rq->lock, flags);
2195
2196 update_rq_clock(rq);
Paul Turnerd6b55912010-11-15 15:47:09 -08002197 update_cfs_load(cfs_rq, 1);
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002198
2199 /*
2200 * We need to update shares after updating tg->load_weight in
2201 * order to adjust the weight of groups with long running tasks.
2202 */
Paul Turner6d5ab292011-01-21 20:45:01 -08002203 update_cfs_shares(cfs_rq);
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002204
2205 raw_spin_unlock_irqrestore(&rq->lock, flags);
2206
2207 return 0;
2208}
2209
2210static void update_shares(int cpu)
2211{
2212 struct cfs_rq *cfs_rq;
2213 struct rq *rq = cpu_rq(cpu);
2214
2215 rcu_read_lock();
Peter Zijlstra9763b672011-07-13 13:09:25 +02002216 /*
2217 * Iterates the task_group tree in a bottom up fashion, see
2218 * list_add_leaf_cfs_rq() for details.
2219 */
Paul Turner67e86252010-11-15 15:47:05 -08002220 for_each_leaf_cfs_rq(rq, cfs_rq)
2221 update_shares_cpu(cfs_rq->tg, cpu);
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002222 rcu_read_unlock();
2223}
2224
Peter Zijlstra9763b672011-07-13 13:09:25 +02002225/*
2226 * Compute the cpu's hierarchical load factor for each task group.
2227 * This needs to be done in a top-down fashion because the load of a child
2228 * group is a fraction of its parents load.
2229 */
2230static int tg_load_down(struct task_group *tg, void *data)
2231{
2232 unsigned long load;
2233 long cpu = (long)data;
2234
2235 if (!tg->parent) {
2236 load = cpu_rq(cpu)->load.weight;
2237 } else {
2238 load = tg->parent->cfs_rq[cpu]->h_load;
2239 load *= tg->se[cpu]->load.weight;
2240 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2241 }
2242
2243 tg->cfs_rq[cpu]->h_load = load;
2244
2245 return 0;
2246}
2247
2248static void update_h_load(long cpu)
2249{
2250 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2251}
2252
Peter Zijlstra230059de2009-12-17 17:47:12 +01002253static unsigned long
2254load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2255 unsigned long max_load_move,
2256 struct sched_domain *sd, enum cpu_idle_type idle,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002257 int *all_pinned)
Peter Zijlstra230059de2009-12-17 17:47:12 +01002258{
2259 long rem_load_move = max_load_move;
Peter Zijlstra9763b672011-07-13 13:09:25 +02002260 struct cfs_rq *busiest_cfs_rq;
Peter Zijlstra230059de2009-12-17 17:47:12 +01002261
2262 rcu_read_lock();
Peter Zijlstra9763b672011-07-13 13:09:25 +02002263 update_h_load(cpu_of(busiest));
Peter Zijlstra230059de2009-12-17 17:47:12 +01002264
Peter Zijlstra9763b672011-07-13 13:09:25 +02002265 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
Peter Zijlstra230059de2009-12-17 17:47:12 +01002266 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2267 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2268 u64 rem_load, moved_load;
2269
2270 /*
2271 * empty group
2272 */
2273 if (!busiest_cfs_rq->task_weight)
2274 continue;
2275
2276 rem_load = (u64)rem_load_move * busiest_weight;
2277 rem_load = div_u64(rem_load, busiest_h_load + 1);
2278
2279 moved_load = balance_tasks(this_rq, this_cpu, busiest,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002280 rem_load, sd, idle, all_pinned,
Peter Zijlstra230059de2009-12-17 17:47:12 +01002281 busiest_cfs_rq);
2282
2283 if (!moved_load)
2284 continue;
2285
2286 moved_load *= busiest_h_load;
2287 moved_load = div_u64(moved_load, busiest_weight + 1);
2288
2289 rem_load_move -= moved_load;
2290 if (rem_load_move < 0)
2291 break;
2292 }
2293 rcu_read_unlock();
2294
2295 return max_load_move - rem_load_move;
2296}
2297#else
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08002298static inline void update_shares(int cpu)
2299{
2300}
2301
Peter Zijlstra230059de2009-12-17 17:47:12 +01002302static unsigned long
2303load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2304 unsigned long max_load_move,
2305 struct sched_domain *sd, enum cpu_idle_type idle,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002306 int *all_pinned)
Peter Zijlstra230059de2009-12-17 17:47:12 +01002307{
2308 return balance_tasks(this_rq, this_cpu, busiest,
2309 max_load_move, sd, idle, all_pinned,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002310 &busiest->cfs);
Peter Zijlstra230059de2009-12-17 17:47:12 +01002311}
2312#endif
2313
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002314/*
2315 * move_tasks tries to move up to max_load_move weighted load from busiest to
2316 * this_rq, as part of a balancing operation within domain "sd".
2317 * Returns 1 if successful and 0 otherwise.
2318 *
2319 * Called with both runqueues locked.
2320 */
2321static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2322 unsigned long max_load_move,
2323 struct sched_domain *sd, enum cpu_idle_type idle,
2324 int *all_pinned)
2325{
Peter Zijlstra3d45fd82009-12-17 17:12:46 +01002326 unsigned long total_load_moved = 0, load_moved;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002327
2328 do {
Peter Zijlstra3d45fd82009-12-17 17:12:46 +01002329 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002330 max_load_move - total_load_moved,
Vladimir Davydov931aeed2011-05-03 22:31:07 +04002331 sd, idle, all_pinned);
Peter Zijlstra3d45fd82009-12-17 17:12:46 +01002332
2333 total_load_moved += load_moved;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002334
2335#ifdef CONFIG_PREEMPT
2336 /*
2337 * NEWIDLE balancing is a source of latency, so preemptible
2338 * kernels will stop after the first task is pulled to minimize
2339 * the critical section.
2340 */
2341 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2342 break;
Peter Zijlstrabaa8c112009-12-17 18:10:09 +01002343
2344 if (raw_spin_is_contended(&this_rq->lock) ||
2345 raw_spin_is_contended(&busiest->lock))
2346 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002347#endif
Peter Zijlstra3d45fd82009-12-17 17:12:46 +01002348 } while (load_moved && max_load_move > total_load_moved);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002349
2350 return total_load_moved > 0;
2351}
2352
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002353/********** Helpers for find_busiest_group ************************/
2354/*
2355 * sd_lb_stats - Structure to store the statistics of a sched_domain
2356 * during load balancing.
2357 */
2358struct sd_lb_stats {
2359 struct sched_group *busiest; /* Busiest group in this sd */
2360 struct sched_group *this; /* Local group in this sd */
2361 unsigned long total_load; /* Total load of all groups in sd */
2362 unsigned long total_pwr; /* Total power of all groups in sd */
2363 unsigned long avg_load; /* Average load across all groups in sd */
2364
2365 /** Statistics of this group */
2366 unsigned long this_load;
2367 unsigned long this_load_per_task;
2368 unsigned long this_nr_running;
Nikhil Raofab47622010-10-15 13:12:29 -07002369 unsigned long this_has_capacity;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002370 unsigned int this_idle_cpus;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002371
2372 /* Statistics of the busiest group */
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002373 unsigned int busiest_idle_cpus;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002374 unsigned long max_load;
2375 unsigned long busiest_load_per_task;
2376 unsigned long busiest_nr_running;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08002377 unsigned long busiest_group_capacity;
Nikhil Raofab47622010-10-15 13:12:29 -07002378 unsigned long busiest_has_capacity;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002379 unsigned int busiest_group_weight;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002380
2381 int group_imb; /* Is there imbalance in this sd */
2382#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2383 int power_savings_balance; /* Is powersave balance needed for this sd */
2384 struct sched_group *group_min; /* Least loaded group in sd */
2385 struct sched_group *group_leader; /* Group which relieves group_min */
2386 unsigned long min_load_per_task; /* load_per_task in group_min */
2387 unsigned long leader_nr_running; /* Nr running of group_leader */
2388 unsigned long min_nr_running; /* Nr running of group_min */
2389#endif
2390};
2391
2392/*
2393 * sg_lb_stats - stats of a sched_group required for load_balancing
2394 */
2395struct sg_lb_stats {
2396 unsigned long avg_load; /*Avg load across the CPUs of the group */
2397 unsigned long group_load; /* Total load over the CPUs of the group */
2398 unsigned long sum_nr_running; /* Nr tasks running in the group */
2399 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2400 unsigned long group_capacity;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002401 unsigned long idle_cpus;
2402 unsigned long group_weight;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002403 int group_imb; /* Is there an imbalance in the group ? */
Nikhil Raofab47622010-10-15 13:12:29 -07002404 int group_has_capacity; /* Is there extra capacity in the group? */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002405};
2406
2407/**
2408 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2409 * @group: The group whose first cpu is to be returned.
2410 */
2411static inline unsigned int group_first_cpu(struct sched_group *group)
2412{
2413 return cpumask_first(sched_group_cpus(group));
2414}
2415
2416/**
2417 * get_sd_load_idx - Obtain the load index for a given sched domain.
2418 * @sd: The sched_domain whose load_idx is to be obtained.
2419 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2420 */
2421static inline int get_sd_load_idx(struct sched_domain *sd,
2422 enum cpu_idle_type idle)
2423{
2424 int load_idx;
2425
2426 switch (idle) {
2427 case CPU_NOT_IDLE:
2428 load_idx = sd->busy_idx;
2429 break;
2430
2431 case CPU_NEWLY_IDLE:
2432 load_idx = sd->newidle_idx;
2433 break;
2434 default:
2435 load_idx = sd->idle_idx;
2436 break;
2437 }
2438
2439 return load_idx;
2440}
2441
2442
2443#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2444/**
2445 * init_sd_power_savings_stats - Initialize power savings statistics for
2446 * the given sched_domain, during load balancing.
2447 *
2448 * @sd: Sched domain whose power-savings statistics are to be initialized.
2449 * @sds: Variable containing the statistics for sd.
2450 * @idle: Idle status of the CPU at which we're performing load-balancing.
2451 */
2452static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2453 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2454{
2455 /*
2456 * Busy processors will not participate in power savings
2457 * balance.
2458 */
2459 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2460 sds->power_savings_balance = 0;
2461 else {
2462 sds->power_savings_balance = 1;
2463 sds->min_nr_running = ULONG_MAX;
2464 sds->leader_nr_running = 0;
2465 }
2466}
2467
2468/**
2469 * update_sd_power_savings_stats - Update the power saving stats for a
2470 * sched_domain while performing load balancing.
2471 *
2472 * @group: sched_group belonging to the sched_domain under consideration.
2473 * @sds: Variable containing the statistics of the sched_domain
2474 * @local_group: Does group contain the CPU for which we're performing
2475 * load balancing ?
2476 * @sgs: Variable containing the statistics of the group.
2477 */
2478static inline void update_sd_power_savings_stats(struct sched_group *group,
2479 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2480{
2481
2482 if (!sds->power_savings_balance)
2483 return;
2484
2485 /*
2486 * If the local group is idle or completely loaded
2487 * no need to do power savings balance at this domain
2488 */
2489 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2490 !sds->this_nr_running))
2491 sds->power_savings_balance = 0;
2492
2493 /*
2494 * If a group is already running at full capacity or idle,
2495 * don't include that group in power savings calculations
2496 */
2497 if (!sds->power_savings_balance ||
2498 sgs->sum_nr_running >= sgs->group_capacity ||
2499 !sgs->sum_nr_running)
2500 return;
2501
2502 /*
2503 * Calculate the group which has the least non-idle load.
2504 * This is the group from where we need to pick up the load
2505 * for saving power
2506 */
2507 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2508 (sgs->sum_nr_running == sds->min_nr_running &&
2509 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2510 sds->group_min = group;
2511 sds->min_nr_running = sgs->sum_nr_running;
2512 sds->min_load_per_task = sgs->sum_weighted_load /
2513 sgs->sum_nr_running;
2514 }
2515
2516 /*
2517 * Calculate the group which is almost near its
2518 * capacity but still has some space to pick up some load
2519 * from other group and save more power
2520 */
2521 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2522 return;
2523
2524 if (sgs->sum_nr_running > sds->leader_nr_running ||
2525 (sgs->sum_nr_running == sds->leader_nr_running &&
2526 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2527 sds->group_leader = group;
2528 sds->leader_nr_running = sgs->sum_nr_running;
2529 }
2530}
2531
2532/**
2533 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2534 * @sds: Variable containing the statistics of the sched_domain
2535 * under consideration.
2536 * @this_cpu: Cpu at which we're currently performing load-balancing.
2537 * @imbalance: Variable to store the imbalance.
2538 *
2539 * Description:
2540 * Check if we have potential to perform some power-savings balance.
2541 * If yes, set the busiest group to be the least loaded group in the
2542 * sched_domain, so that it's CPUs can be put to idle.
2543 *
2544 * Returns 1 if there is potential to perform power-savings balance.
2545 * Else returns 0.
2546 */
2547static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2548 int this_cpu, unsigned long *imbalance)
2549{
2550 if (!sds->power_savings_balance)
2551 return 0;
2552
2553 if (sds->this != sds->group_leader ||
2554 sds->group_leader == sds->group_min)
2555 return 0;
2556
2557 *imbalance = sds->min_load_per_task;
2558 sds->busiest = sds->group_min;
2559
2560 return 1;
2561
2562}
2563#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2564static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2565 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2566{
2567 return;
2568}
2569
2570static inline void update_sd_power_savings_stats(struct sched_group *group,
2571 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2572{
2573 return;
2574}
2575
2576static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2577 int this_cpu, unsigned long *imbalance)
2578{
2579 return 0;
2580}
2581#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2582
2583
2584unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2585{
Nikhil Rao1399fa72011-05-18 10:09:39 -07002586 return SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002587}
2588
2589unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2590{
2591 return default_scale_freq_power(sd, cpu);
2592}
2593
2594unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2595{
Peter Zijlstra669c55e2010-04-16 14:59:29 +02002596 unsigned long weight = sd->span_weight;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002597 unsigned long smt_gain = sd->smt_gain;
2598
2599 smt_gain /= weight;
2600
2601 return smt_gain;
2602}
2603
2604unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2605{
2606 return default_scale_smt_power(sd, cpu);
2607}
2608
2609unsigned long scale_rt_power(int cpu)
2610{
2611 struct rq *rq = cpu_rq(cpu);
2612 u64 total, available;
2613
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002614 total = sched_avg_period() + (rq->clock - rq->age_stamp);
Venkatesh Pallipadiaa483802010-10-04 17:03:22 -07002615
2616 if (unlikely(total < rq->rt_avg)) {
2617 /* Ensures that power won't end up being negative */
2618 available = 0;
2619 } else {
2620 available = total - rq->rt_avg;
2621 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002622
Nikhil Rao1399fa72011-05-18 10:09:39 -07002623 if (unlikely((s64)total < SCHED_POWER_SCALE))
2624 total = SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002625
Nikhil Rao1399fa72011-05-18 10:09:39 -07002626 total >>= SCHED_POWER_SHIFT;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002627
2628 return div_u64(available, total);
2629}
2630
2631static void update_cpu_power(struct sched_domain *sd, int cpu)
2632{
Peter Zijlstra669c55e2010-04-16 14:59:29 +02002633 unsigned long weight = sd->span_weight;
Nikhil Rao1399fa72011-05-18 10:09:39 -07002634 unsigned long power = SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002635 struct sched_group *sdg = sd->groups;
2636
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002637 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2638 if (sched_feat(ARCH_POWER))
2639 power *= arch_scale_smt_power(sd, cpu);
2640 else
2641 power *= default_scale_smt_power(sd, cpu);
2642
Nikhil Rao1399fa72011-05-18 10:09:39 -07002643 power >>= SCHED_POWER_SHIFT;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002644 }
2645
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002646 sdg->sgp->power_orig = power;
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002647
2648 if (sched_feat(ARCH_POWER))
2649 power *= arch_scale_freq_power(sd, cpu);
2650 else
2651 power *= default_scale_freq_power(sd, cpu);
2652
Nikhil Rao1399fa72011-05-18 10:09:39 -07002653 power >>= SCHED_POWER_SHIFT;
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002654
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002655 power *= scale_rt_power(cpu);
Nikhil Rao1399fa72011-05-18 10:09:39 -07002656 power >>= SCHED_POWER_SHIFT;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002657
2658 if (!power)
2659 power = 1;
2660
Peter Zijlstrae51fd5e2010-05-31 12:37:30 +02002661 cpu_rq(cpu)->cpu_power = power;
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002662 sdg->sgp->power = power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002663}
2664
2665static void update_group_power(struct sched_domain *sd, int cpu)
2666{
2667 struct sched_domain *child = sd->child;
2668 struct sched_group *group, *sdg = sd->groups;
2669 unsigned long power;
2670
2671 if (!child) {
2672 update_cpu_power(sd, cpu);
2673 return;
2674 }
2675
2676 power = 0;
2677
2678 group = child->groups;
2679 do {
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002680 power += group->sgp->power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002681 group = group->next;
2682 } while (group != child->groups);
2683
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002684 sdg->sgp->power = power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002685}
2686
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002687/*
2688 * Try and fix up capacity for tiny siblings, this is needed when
2689 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2690 * which on its own isn't powerful enough.
2691 *
2692 * See update_sd_pick_busiest() and check_asym_packing().
2693 */
2694static inline int
2695fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2696{
2697 /*
Nikhil Rao1399fa72011-05-18 10:09:39 -07002698 * Only siblings can have significantly less than SCHED_POWER_SCALE
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002699 */
Peter Zijlstraa6c75f22011-04-07 14:09:52 +02002700 if (!(sd->flags & SD_SHARE_CPUPOWER))
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002701 return 0;
2702
2703 /*
2704 * If ~90% of the cpu_power is still there, we're good.
2705 */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002706 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002707 return 1;
2708
2709 return 0;
2710}
2711
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002712/**
2713 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2714 * @sd: The sched_domain whose statistics are to be updated.
2715 * @group: sched_group whose statistics are to be updated.
2716 * @this_cpu: Cpu for which load balance is currently performed.
2717 * @idle: Idle status of this_cpu
2718 * @load_idx: Load index of sched_domain of this_cpu for load calc.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002719 * @local_group: Does group contain this_cpu.
2720 * @cpus: Set of cpus considered for load balancing.
2721 * @balance: Should we balance.
2722 * @sgs: variable to hold the statistics for this group.
2723 */
2724static inline void update_sg_lb_stats(struct sched_domain *sd,
2725 struct sched_group *group, int this_cpu,
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08002726 enum cpu_idle_type idle, int load_idx,
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002727 int local_group, const struct cpumask *cpus,
2728 int *balance, struct sg_lb_stats *sgs)
2729{
Nikhil Rao2582f0e2010-10-13 12:09:36 -07002730 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002731 int i;
2732 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08002733 unsigned long avg_load_per_task = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002734
Gautham R Shenoy871e35b2010-01-20 14:02:44 -06002735 if (local_group)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002736 balance_cpu = group_first_cpu(group);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002737
2738 /* Tally up the load of all CPUs in the group */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002739 max_cpu_load = 0;
2740 min_cpu_load = ~0UL;
Nikhil Rao2582f0e2010-10-13 12:09:36 -07002741 max_nr_running = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002742
2743 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2744 struct rq *rq = cpu_rq(i);
2745
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002746 /* Bias balancing toward cpus of our domain */
2747 if (local_group) {
2748 if (idle_cpu(i) && !first_idle_cpu) {
2749 first_idle_cpu = 1;
2750 balance_cpu = i;
2751 }
2752
2753 load = target_load(i, load_idx);
2754 } else {
2755 load = source_load(i, load_idx);
Nikhil Rao2582f0e2010-10-13 12:09:36 -07002756 if (load > max_cpu_load) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002757 max_cpu_load = load;
Nikhil Rao2582f0e2010-10-13 12:09:36 -07002758 max_nr_running = rq->nr_running;
2759 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002760 if (min_cpu_load > load)
2761 min_cpu_load = load;
2762 }
2763
2764 sgs->group_load += load;
2765 sgs->sum_nr_running += rq->nr_running;
2766 sgs->sum_weighted_load += weighted_cpuload(i);
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002767 if (idle_cpu(i))
2768 sgs->idle_cpus++;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002769 }
2770
2771 /*
2772 * First idle cpu or the first cpu(busiest) in this sched group
2773 * is eligible for doing load balancing at this and above
2774 * domains. In the newly idle case, we will allow all the cpu's
2775 * to do the newly idle load balance.
2776 */
Peter Zijlstrabbc8cb52010-07-09 15:15:43 +02002777 if (idle != CPU_NEWLY_IDLE && local_group) {
2778 if (balance_cpu != this_cpu) {
2779 *balance = 0;
2780 return;
2781 }
2782 update_group_power(sd, this_cpu);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002783 }
2784
2785 /* Adjust by relative CPU power of the group */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002786 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002787
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002788 /*
2789 * Consider the group unbalanced when the imbalance is larger
Peter Zijlstra866ab432011-02-21 18:56:47 +01002790 * than the average weight of a task.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002791 *
2792 * APZ: with cgroup the avg task weight can vary wildly and
2793 * might not be a suitable number - should we keep a
2794 * normalized nr_running number somewhere that negates
2795 * the hierarchy?
2796 */
Suresh Siddhadd5feea2010-02-23 16:13:52 -08002797 if (sgs->sum_nr_running)
2798 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002799
Peter Zijlstra866ab432011-02-21 18:56:47 +01002800 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002801 sgs->group_imb = 1;
2802
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002803 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
Nikhil Rao1399fa72011-05-18 10:09:39 -07002804 SCHED_POWER_SCALE);
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10002805 if (!sgs->group_capacity)
2806 sgs->group_capacity = fix_small_capacity(sd, group);
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002807 sgs->group_weight = group->group_weight;
Nikhil Raofab47622010-10-15 13:12:29 -07002808
2809 if (sgs->group_capacity > sgs->sum_nr_running)
2810 sgs->group_has_capacity = 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002811}
2812
2813/**
Michael Neuling532cb4c2010-06-08 14:57:02 +10002814 * update_sd_pick_busiest - return 1 on busiest group
2815 * @sd: sched_domain whose statistics are to be checked
2816 * @sds: sched_domain statistics
2817 * @sg: sched_group candidate to be checked for being the busiest
Michael Neulingb6b12292010-06-10 12:06:21 +10002818 * @sgs: sched_group statistics
2819 * @this_cpu: the current cpu
Michael Neuling532cb4c2010-06-08 14:57:02 +10002820 *
2821 * Determine if @sg is a busier group than the previously selected
2822 * busiest group.
2823 */
2824static bool update_sd_pick_busiest(struct sched_domain *sd,
2825 struct sd_lb_stats *sds,
2826 struct sched_group *sg,
2827 struct sg_lb_stats *sgs,
2828 int this_cpu)
2829{
2830 if (sgs->avg_load <= sds->max_load)
2831 return false;
2832
2833 if (sgs->sum_nr_running > sgs->group_capacity)
2834 return true;
2835
2836 if (sgs->group_imb)
2837 return true;
2838
2839 /*
2840 * ASYM_PACKING needs to move all the work to the lowest
2841 * numbered CPUs in the group, therefore mark all groups
2842 * higher than ourself as busy.
2843 */
2844 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2845 this_cpu < group_first_cpu(sg)) {
2846 if (!sds->busiest)
2847 return true;
2848
2849 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2850 return true;
2851 }
2852
2853 return false;
2854}
2855
2856/**
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002857 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2858 * @sd: sched_domain whose statistics are to be updated.
2859 * @this_cpu: Cpu for which load balance is currently performed.
2860 * @idle: Idle status of this_cpu
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002861 * @cpus: Set of cpus considered for load balancing.
2862 * @balance: Should we balance.
2863 * @sds: variable to hold the statistics for this sched_domain.
2864 */
2865static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08002866 enum cpu_idle_type idle, const struct cpumask *cpus,
2867 int *balance, struct sd_lb_stats *sds)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002868{
2869 struct sched_domain *child = sd->child;
Michael Neuling532cb4c2010-06-08 14:57:02 +10002870 struct sched_group *sg = sd->groups;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002871 struct sg_lb_stats sgs;
2872 int load_idx, prefer_sibling = 0;
2873
2874 if (child && child->flags & SD_PREFER_SIBLING)
2875 prefer_sibling = 1;
2876
2877 init_sd_power_savings_stats(sd, sds, idle);
2878 load_idx = get_sd_load_idx(sd, idle);
2879
2880 do {
2881 int local_group;
2882
Michael Neuling532cb4c2010-06-08 14:57:02 +10002883 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002884 memset(&sgs, 0, sizeof(sgs));
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08002885 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002886 local_group, cpus, balance, &sgs);
2887
Peter Zijlstra8f190fb2009-12-24 14:18:21 +01002888 if (local_group && !(*balance))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002889 return;
2890
2891 sds->total_load += sgs.group_load;
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002892 sds->total_pwr += sg->sgp->power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002893
2894 /*
2895 * In case the child domain prefers tasks go to siblings
Michael Neuling532cb4c2010-06-08 14:57:02 +10002896 * first, lower the sg capacity to one so that we'll try
Nikhil Rao75dd3212010-10-15 13:12:30 -07002897 * and move all the excess tasks away. We lower the capacity
2898 * of a group only if the local group has the capacity to fit
2899 * these excess tasks, i.e. nr_running < group_capacity. The
2900 * extra check prevents the case where you always pull from the
2901 * heaviest group when it is already under-utilized (possible
2902 * with a large weight task outweighs the tasks on the system).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002903 */
Nikhil Rao75dd3212010-10-15 13:12:30 -07002904 if (prefer_sibling && !local_group && sds->this_has_capacity)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002905 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2906
2907 if (local_group) {
2908 sds->this_load = sgs.avg_load;
Michael Neuling532cb4c2010-06-08 14:57:02 +10002909 sds->this = sg;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002910 sds->this_nr_running = sgs.sum_nr_running;
2911 sds->this_load_per_task = sgs.sum_weighted_load;
Nikhil Raofab47622010-10-15 13:12:29 -07002912 sds->this_has_capacity = sgs.group_has_capacity;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002913 sds->this_idle_cpus = sgs.idle_cpus;
Michael Neuling532cb4c2010-06-08 14:57:02 +10002914 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002915 sds->max_load = sgs.avg_load;
Michael Neuling532cb4c2010-06-08 14:57:02 +10002916 sds->busiest = sg;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002917 sds->busiest_nr_running = sgs.sum_nr_running;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002918 sds->busiest_idle_cpus = sgs.idle_cpus;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08002919 sds->busiest_group_capacity = sgs.group_capacity;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002920 sds->busiest_load_per_task = sgs.sum_weighted_load;
Nikhil Raofab47622010-10-15 13:12:29 -07002921 sds->busiest_has_capacity = sgs.group_has_capacity;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07002922 sds->busiest_group_weight = sgs.group_weight;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002923 sds->group_imb = sgs.group_imb;
2924 }
2925
Michael Neuling532cb4c2010-06-08 14:57:02 +10002926 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2927 sg = sg->next;
2928 } while (sg != sd->groups);
2929}
2930
Michael Neuling2ec57d42010-06-29 12:02:01 +10002931int __weak arch_sd_sibling_asym_packing(void)
Michael Neuling532cb4c2010-06-08 14:57:02 +10002932{
2933 return 0*SD_ASYM_PACKING;
2934}
2935
2936/**
2937 * check_asym_packing - Check to see if the group is packed into the
2938 * sched doman.
2939 *
2940 * This is primarily intended to used at the sibling level. Some
2941 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2942 * case of POWER7, it can move to lower SMT modes only when higher
2943 * threads are idle. When in lower SMT modes, the threads will
2944 * perform better since they share less core resources. Hence when we
2945 * have idle threads, we want them to be the higher ones.
2946 *
2947 * This packing function is run on idle threads. It checks to see if
2948 * the busiest CPU in this domain (core in the P7 case) has a higher
2949 * CPU number than the packing function is being run on. Here we are
2950 * assuming lower CPU number will be equivalent to lower a SMT thread
2951 * number.
2952 *
Michael Neulingb6b12292010-06-10 12:06:21 +10002953 * Returns 1 when packing is required and a task should be moved to
2954 * this CPU. The amount of the imbalance is returned in *imbalance.
2955 *
Michael Neuling532cb4c2010-06-08 14:57:02 +10002956 * @sd: The sched_domain whose packing is to be checked.
2957 * @sds: Statistics of the sched_domain which is to be packed
2958 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2959 * @imbalance: returns amount of imbalanced due to packing.
Michael Neuling532cb4c2010-06-08 14:57:02 +10002960 */
2961static int check_asym_packing(struct sched_domain *sd,
2962 struct sd_lb_stats *sds,
2963 int this_cpu, unsigned long *imbalance)
2964{
2965 int busiest_cpu;
2966
2967 if (!(sd->flags & SD_ASYM_PACKING))
2968 return 0;
2969
2970 if (!sds->busiest)
2971 return 0;
2972
2973 busiest_cpu = group_first_cpu(sds->busiest);
2974 if (this_cpu > busiest_cpu)
2975 return 0;
2976
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02002977 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
Nikhil Rao1399fa72011-05-18 10:09:39 -07002978 SCHED_POWER_SCALE);
Michael Neuling532cb4c2010-06-08 14:57:02 +10002979 return 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002980}
2981
2982/**
2983 * fix_small_imbalance - Calculate the minor imbalance that exists
2984 * amongst the groups of a sched_domain, during
2985 * load balancing.
2986 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2987 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2988 * @imbalance: Variable to store the imbalance.
2989 */
2990static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2991 int this_cpu, unsigned long *imbalance)
2992{
2993 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2994 unsigned int imbn = 2;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08002995 unsigned long scaled_busy_load_per_task;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002996
2997 if (sds->this_nr_running) {
2998 sds->this_load_per_task /= sds->this_nr_running;
2999 if (sds->busiest_load_per_task >
3000 sds->this_load_per_task)
3001 imbn = 1;
3002 } else
3003 sds->this_load_per_task =
3004 cpu_avg_load_per_task(this_cpu);
3005
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003006 scaled_busy_load_per_task = sds->busiest_load_per_task
Nikhil Rao1399fa72011-05-18 10:09:39 -07003007 * SCHED_POWER_SCALE;
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003008 scaled_busy_load_per_task /= sds->busiest->sgp->power;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003009
3010 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3011 (scaled_busy_load_per_task * imbn)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003012 *imbalance = sds->busiest_load_per_task;
3013 return;
3014 }
3015
3016 /*
3017 * OK, we don't have enough imbalance to justify moving tasks,
3018 * however we may be able to increase total CPU power used by
3019 * moving them.
3020 */
3021
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003022 pwr_now += sds->busiest->sgp->power *
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003023 min(sds->busiest_load_per_task, sds->max_load);
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003024 pwr_now += sds->this->sgp->power *
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003025 min(sds->this_load_per_task, sds->this_load);
Nikhil Rao1399fa72011-05-18 10:09:39 -07003026 pwr_now /= SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003027
3028 /* Amount of load we'd subtract */
Nikhil Rao1399fa72011-05-18 10:09:39 -07003029 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003030 sds->busiest->sgp->power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003031 if (sds->max_load > tmp)
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003032 pwr_move += sds->busiest->sgp->power *
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003033 min(sds->busiest_load_per_task, sds->max_load - tmp);
3034
3035 /* Amount of load we'd add */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003036 if (sds->max_load * sds->busiest->sgp->power <
Nikhil Rao1399fa72011-05-18 10:09:39 -07003037 sds->busiest_load_per_task * SCHED_POWER_SCALE)
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003038 tmp = (sds->max_load * sds->busiest->sgp->power) /
3039 sds->this->sgp->power;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003040 else
Nikhil Rao1399fa72011-05-18 10:09:39 -07003041 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003042 sds->this->sgp->power;
3043 pwr_move += sds->this->sgp->power *
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003044 min(sds->this_load_per_task, sds->this_load + tmp);
Nikhil Rao1399fa72011-05-18 10:09:39 -07003045 pwr_move /= SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003046
3047 /* Move if we gain throughput */
3048 if (pwr_move > pwr_now)
3049 *imbalance = sds->busiest_load_per_task;
3050}
3051
3052/**
3053 * calculate_imbalance - Calculate the amount of imbalance present within the
3054 * groups of a given sched_domain during load balance.
3055 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3056 * @this_cpu: Cpu for which currently load balance is being performed.
3057 * @imbalance: The variable to store the imbalance.
3058 */
3059static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3060 unsigned long *imbalance)
3061{
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003062 unsigned long max_pull, load_above_capacity = ~0UL;
3063
3064 sds->busiest_load_per_task /= sds->busiest_nr_running;
3065 if (sds->group_imb) {
3066 sds->busiest_load_per_task =
3067 min(sds->busiest_load_per_task, sds->avg_load);
3068 }
3069
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003070 /*
3071 * In the presence of smp nice balancing, certain scenarios can have
3072 * max load less than avg load(as we skip the groups at or below
3073 * its cpu_power, while calculating max_load..)
3074 */
3075 if (sds->max_load < sds->avg_load) {
3076 *imbalance = 0;
3077 return fix_small_imbalance(sds, this_cpu, imbalance);
3078 }
3079
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003080 if (!sds->group_imb) {
3081 /*
3082 * Don't want to pull so many tasks that a group would go idle.
3083 */
3084 load_above_capacity = (sds->busiest_nr_running -
3085 sds->busiest_group_capacity);
3086
Nikhil Rao1399fa72011-05-18 10:09:39 -07003087 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003088
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003089 load_above_capacity /= sds->busiest->sgp->power;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08003090 }
3091
3092 /*
3093 * We're trying to get all the cpus to the average_load, so we don't
3094 * want to push ourselves above the average load, nor do we wish to
3095 * reduce the max loaded cpu below the average load. At the same time,
3096 * we also don't want to reduce the group load below the group capacity
3097 * (so that we can implement power-savings policies etc). Thus we look
3098 * for the minimum possible imbalance.
3099 * Be careful of negative numbers as they'll appear as very large values
3100 * with unsigned longs.
3101 */
3102 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003103
3104 /* How much load to actually move to equalise the imbalance */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02003105 *imbalance = min(max_pull * sds->busiest->sgp->power,
3106 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
Nikhil Rao1399fa72011-05-18 10:09:39 -07003107 / SCHED_POWER_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003108
3109 /*
3110 * if *imbalance is less than the average load per runnable task
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003111 * there is no guarantee that any tasks will be moved so we'll have
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003112 * a think about bumping its value to force at least one task to be
3113 * moved
3114 */
3115 if (*imbalance < sds->busiest_load_per_task)
3116 return fix_small_imbalance(sds, this_cpu, imbalance);
3117
3118}
Nikhil Raofab47622010-10-15 13:12:29 -07003119
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003120/******* find_busiest_group() helpers end here *********************/
3121
3122/**
3123 * find_busiest_group - Returns the busiest group within the sched_domain
3124 * if there is an imbalance. If there isn't an imbalance, and
3125 * the user has opted for power-savings, it returns a group whose
3126 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3127 * such a group exists.
3128 *
3129 * Also calculates the amount of weighted load which should be moved
3130 * to restore balance.
3131 *
3132 * @sd: The sched_domain whose busiest group is to be returned.
3133 * @this_cpu: The cpu for which load balancing is currently being performed.
3134 * @imbalance: Variable which stores amount of weighted load which should
3135 * be moved to restore balance/put a group to idle.
3136 * @idle: The idle status of this_cpu.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003137 * @cpus: The set of CPUs under consideration for load-balancing.
3138 * @balance: Pointer to a variable indicating if this_cpu
3139 * is the appropriate cpu to perform load balancing at this_level.
3140 *
3141 * Returns: - the busiest group if imbalance exists.
3142 * - If no imbalance and user has opted for power-savings balance,
3143 * return the least loaded group whose CPUs can be
3144 * put to idle by rebalancing its tasks onto our group.
3145 */
3146static struct sched_group *
3147find_busiest_group(struct sched_domain *sd, int this_cpu,
3148 unsigned long *imbalance, enum cpu_idle_type idle,
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003149 const struct cpumask *cpus, int *balance)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003150{
3151 struct sd_lb_stats sds;
3152
3153 memset(&sds, 0, sizeof(sds));
3154
3155 /*
3156 * Compute the various statistics relavent for load balancing at
3157 * this level.
3158 */
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003159 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003160
Peter Zijlstracc57aa82011-02-21 18:55:32 +01003161 /*
3162 * this_cpu is not the appropriate cpu to perform load balancing at
3163 * this level.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003164 */
Peter Zijlstra8f190fb2009-12-24 14:18:21 +01003165 if (!(*balance))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003166 goto ret;
3167
Michael Neuling532cb4c2010-06-08 14:57:02 +10003168 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3169 check_asym_packing(sd, &sds, this_cpu, imbalance))
3170 return sds.busiest;
3171
Peter Zijlstracc57aa82011-02-21 18:55:32 +01003172 /* There is no busy sibling group to pull tasks from */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003173 if (!sds.busiest || sds.busiest_nr_running == 0)
3174 goto out_balanced;
3175
Nikhil Rao1399fa72011-05-18 10:09:39 -07003176 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
Ken Chenb0432d82011-04-07 17:23:22 -07003177
Peter Zijlstra866ab432011-02-21 18:56:47 +01003178 /*
3179 * If the busiest group is imbalanced the below checks don't
3180 * work because they assumes all things are equal, which typically
3181 * isn't true due to cpus_allowed constraints and the like.
3182 */
3183 if (sds.group_imb)
3184 goto force_balance;
3185
Peter Zijlstracc57aa82011-02-21 18:55:32 +01003186 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
Nikhil Raofab47622010-10-15 13:12:29 -07003187 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3188 !sds.busiest_has_capacity)
3189 goto force_balance;
3190
Peter Zijlstracc57aa82011-02-21 18:55:32 +01003191 /*
3192 * If the local group is more busy than the selected busiest group
3193 * don't try and pull any tasks.
3194 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003195 if (sds.this_load >= sds.max_load)
3196 goto out_balanced;
3197
Peter Zijlstracc57aa82011-02-21 18:55:32 +01003198 /*
3199 * Don't pull any tasks if this group is already above the domain
3200 * average load.
3201 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003202 if (sds.this_load >= sds.avg_load)
3203 goto out_balanced;
3204
Peter Zijlstrac186faf2011-02-21 18:52:53 +01003205 if (idle == CPU_IDLE) {
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07003206 /*
3207 * This cpu is idle. If the busiest group load doesn't
3208 * have more tasks than the number of available cpu's and
3209 * there is no imbalance between this and busiest group
3210 * wrt to idle cpu's, it is balanced.
3211 */
Peter Zijlstrac186faf2011-02-21 18:52:53 +01003212 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07003213 sds.busiest_nr_running <= sds.busiest_group_weight)
3214 goto out_balanced;
Peter Zijlstrac186faf2011-02-21 18:52:53 +01003215 } else {
3216 /*
3217 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3218 * imbalance_pct to be conservative.
3219 */
3220 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3221 goto out_balanced;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07003222 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003223
Nikhil Raofab47622010-10-15 13:12:29 -07003224force_balance:
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003225 /* Looks like there is an imbalance. Compute it */
3226 calculate_imbalance(&sds, this_cpu, imbalance);
3227 return sds.busiest;
3228
3229out_balanced:
3230 /*
3231 * There is no obvious imbalance. But check if we can do some balancing
3232 * to save power.
3233 */
3234 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3235 return sds.busiest;
3236ret:
3237 *imbalance = 0;
3238 return NULL;
3239}
3240
3241/*
3242 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3243 */
3244static struct rq *
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10003245find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3246 enum cpu_idle_type idle, unsigned long imbalance,
3247 const struct cpumask *cpus)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003248{
3249 struct rq *busiest = NULL, *rq;
3250 unsigned long max_load = 0;
3251 int i;
3252
3253 for_each_cpu(i, sched_group_cpus(group)) {
3254 unsigned long power = power_of(i);
Nikhil Rao1399fa72011-05-18 10:09:39 -07003255 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3256 SCHED_POWER_SCALE);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003257 unsigned long wl;
3258
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10003259 if (!capacity)
3260 capacity = fix_small_capacity(sd, group);
3261
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003262 if (!cpumask_test_cpu(i, cpus))
3263 continue;
3264
3265 rq = cpu_rq(i);
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01003266 wl = weighted_cpuload(i);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003267
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01003268 /*
3269 * When comparing with imbalance, use weighted_cpuload()
3270 * which is not scaled with the cpu power.
3271 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003272 if (capacity && rq->nr_running == 1 && wl > imbalance)
3273 continue;
3274
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01003275 /*
3276 * For the load comparisons with the other cpu's, consider
3277 * the weighted_cpuload() scaled with the cpu power, so that
3278 * the load can be moved away from the cpu that is potentially
3279 * running at a lower capacity.
3280 */
Nikhil Rao1399fa72011-05-18 10:09:39 -07003281 wl = (wl * SCHED_POWER_SCALE) / power;
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01003282
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003283 if (wl > max_load) {
3284 max_load = wl;
3285 busiest = rq;
3286 }
3287 }
3288
3289 return busiest;
3290}
3291
3292/*
3293 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3294 * so long as it is large enough.
3295 */
3296#define MAX_PINNED_INTERVAL 512
3297
3298/* Working cpumask for load_balance and load_balance_newidle. */
3299static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3300
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003301static int need_active_balance(struct sched_domain *sd, int idle,
Michael Neuling532cb4c2010-06-08 14:57:02 +10003302 int busiest_cpu, int this_cpu)
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01003303{
3304 if (idle == CPU_NEWLY_IDLE) {
Michael Neuling532cb4c2010-06-08 14:57:02 +10003305
3306 /*
3307 * ASYM_PACKING needs to force migrate tasks from busy but
3308 * higher numbered CPUs in order to pack all tasks in the
3309 * lowest numbered CPUs.
3310 */
3311 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3312 return 1;
3313
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01003314 /*
3315 * The only task running in a non-idle cpu can be moved to this
3316 * cpu in an attempt to completely freeup the other CPU
3317 * package.
3318 *
3319 * The package power saving logic comes from
3320 * find_busiest_group(). If there are no imbalance, then
3321 * f_b_g() will return NULL. However when sched_mc={1,2} then
3322 * f_b_g() will select a group from which a running task may be
3323 * pulled to this cpu in order to make the other package idle.
3324 * If there is no opportunity to make a package idle and if
3325 * there are no imbalance, then f_b_g() will return NULL and no
3326 * action will be taken in load_balance_newidle().
3327 *
3328 * Under normal task pull operation due to imbalance, there
3329 * will be more than one task in the source run queue and
3330 * move_tasks() will succeed. ld_moved will be true and this
3331 * active balance code will not be triggered.
3332 */
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01003333 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3334 return 0;
3335 }
3336
3337 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3338}
3339
Tejun Heo969c7922010-05-06 18:49:21 +02003340static int active_load_balance_cpu_stop(void *data);
3341
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003342/*
3343 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3344 * tasks if there is an imbalance.
3345 */
3346static int load_balance(int this_cpu, struct rq *this_rq,
3347 struct sched_domain *sd, enum cpu_idle_type idle,
3348 int *balance)
3349{
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003350 int ld_moved, all_pinned = 0, active_balance = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003351 struct sched_group *group;
3352 unsigned long imbalance;
3353 struct rq *busiest;
3354 unsigned long flags;
3355 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3356
3357 cpumask_copy(cpus, cpu_active_mask);
3358
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003359 schedstat_inc(sd, lb_count[idle]);
3360
3361redo:
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003362 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003363 cpus, balance);
3364
3365 if (*balance == 0)
3366 goto out_balanced;
3367
3368 if (!group) {
3369 schedstat_inc(sd, lb_nobusyg[idle]);
3370 goto out_balanced;
3371 }
3372
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10003373 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003374 if (!busiest) {
3375 schedstat_inc(sd, lb_nobusyq[idle]);
3376 goto out_balanced;
3377 }
3378
3379 BUG_ON(busiest == this_rq);
3380
3381 schedstat_add(sd, lb_imbalance[idle], imbalance);
3382
3383 ld_moved = 0;
3384 if (busiest->nr_running > 1) {
3385 /*
3386 * Attempt to move tasks. If find_busiest_group has found
3387 * an imbalance but busiest->nr_running <= 1, the group is
3388 * still unbalanced. ld_moved simply stays zero, so it is
3389 * correctly treated as an imbalance.
3390 */
Ken Chenb30aef12011-04-08 12:20:16 -07003391 all_pinned = 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003392 local_irq_save(flags);
3393 double_rq_lock(this_rq, busiest);
3394 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3395 imbalance, sd, idle, &all_pinned);
3396 double_rq_unlock(this_rq, busiest);
3397 local_irq_restore(flags);
3398
3399 /*
3400 * some other cpu did the load balance for us.
3401 */
3402 if (ld_moved && this_cpu != smp_processor_id())
3403 resched_cpu(this_cpu);
3404
3405 /* All tasks on this runqueue were pinned by CPU affinity */
3406 if (unlikely(all_pinned)) {
3407 cpumask_clear_cpu(cpu_of(busiest), cpus);
3408 if (!cpumask_empty(cpus))
3409 goto redo;
3410 goto out_balanced;
3411 }
3412 }
3413
3414 if (!ld_moved) {
3415 schedstat_inc(sd, lb_failed[idle]);
Venkatesh Pallipadi58b26c42010-09-10 18:19:17 -07003416 /*
3417 * Increment the failure counter only on periodic balance.
3418 * We do not want newidle balance, which can be very
3419 * frequent, pollute the failure counter causing
3420 * excessive cache_hot migrations and active balances.
3421 */
3422 if (idle != CPU_NEWLY_IDLE)
3423 sd->nr_balance_failed++;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003424
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003425 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003426 raw_spin_lock_irqsave(&busiest->lock, flags);
3427
Tejun Heo969c7922010-05-06 18:49:21 +02003428 /* don't kick the active_load_balance_cpu_stop,
3429 * if the curr task on busiest cpu can't be
3430 * moved to this_cpu
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003431 */
3432 if (!cpumask_test_cpu(this_cpu,
3433 &busiest->curr->cpus_allowed)) {
3434 raw_spin_unlock_irqrestore(&busiest->lock,
3435 flags);
3436 all_pinned = 1;
3437 goto out_one_pinned;
3438 }
3439
Tejun Heo969c7922010-05-06 18:49:21 +02003440 /*
3441 * ->active_balance synchronizes accesses to
3442 * ->active_balance_work. Once set, it's cleared
3443 * only after active load balance is finished.
3444 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003445 if (!busiest->active_balance) {
3446 busiest->active_balance = 1;
3447 busiest->push_cpu = this_cpu;
3448 active_balance = 1;
3449 }
3450 raw_spin_unlock_irqrestore(&busiest->lock, flags);
Tejun Heo969c7922010-05-06 18:49:21 +02003451
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003452 if (active_balance)
Tejun Heo969c7922010-05-06 18:49:21 +02003453 stop_one_cpu_nowait(cpu_of(busiest),
3454 active_load_balance_cpu_stop, busiest,
3455 &busiest->active_balance_work);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003456
3457 /*
3458 * We've kicked active balancing, reset the failure
3459 * counter.
3460 */
3461 sd->nr_balance_failed = sd->cache_nice_tries+1;
3462 }
3463 } else
3464 sd->nr_balance_failed = 0;
3465
3466 if (likely(!active_balance)) {
3467 /* We were unbalanced, so reset the balancing interval */
3468 sd->balance_interval = sd->min_interval;
3469 } else {
3470 /*
3471 * If we've begun active balancing, start to back off. This
3472 * case may not be covered by the all_pinned logic if there
3473 * is only 1 task on the busy runqueue (because we don't call
3474 * move_tasks).
3475 */
3476 if (sd->balance_interval < sd->max_interval)
3477 sd->balance_interval *= 2;
3478 }
3479
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003480 goto out;
3481
3482out_balanced:
3483 schedstat_inc(sd, lb_balanced[idle]);
3484
3485 sd->nr_balance_failed = 0;
3486
3487out_one_pinned:
3488 /* tune up the balancing interval */
3489 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3490 (sd->balance_interval < sd->max_interval))
3491 sd->balance_interval *= 2;
3492
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08003493 ld_moved = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003494out:
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003495 return ld_moved;
3496}
3497
3498/*
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003499 * idle_balance is called by schedule() if this_cpu is about to become
3500 * idle. Attempts to pull tasks from other CPUs.
3501 */
3502static void idle_balance(int this_cpu, struct rq *this_rq)
3503{
3504 struct sched_domain *sd;
3505 int pulled_task = 0;
3506 unsigned long next_balance = jiffies + HZ;
3507
3508 this_rq->idle_stamp = this_rq->clock;
3509
3510 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3511 return;
3512
Peter Zijlstraf492e122009-12-23 15:29:42 +01003513 /*
3514 * Drop the rq->lock, but keep IRQ/preempt disabled.
3515 */
3516 raw_spin_unlock(&this_rq->lock);
3517
Paul Turnerc66eaf62010-11-15 15:47:07 -08003518 update_shares(this_cpu);
Peter Zijlstradce840a2011-04-07 14:09:50 +02003519 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003520 for_each_domain(this_cpu, sd) {
3521 unsigned long interval;
Peter Zijlstraf492e122009-12-23 15:29:42 +01003522 int balance = 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003523
3524 if (!(sd->flags & SD_LOAD_BALANCE))
3525 continue;
3526
Peter Zijlstraf492e122009-12-23 15:29:42 +01003527 if (sd->flags & SD_BALANCE_NEWIDLE) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003528 /* If we've pulled tasks over stop searching: */
Peter Zijlstraf492e122009-12-23 15:29:42 +01003529 pulled_task = load_balance(this_cpu, this_rq,
3530 sd, CPU_NEWLY_IDLE, &balance);
3531 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003532
3533 interval = msecs_to_jiffies(sd->balance_interval);
3534 if (time_after(next_balance, sd->last_balance + interval))
3535 next_balance = sd->last_balance + interval;
Nikhil Raod5ad1402010-11-17 11:42:04 -08003536 if (pulled_task) {
3537 this_rq->idle_stamp = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003538 break;
Nikhil Raod5ad1402010-11-17 11:42:04 -08003539 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003540 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02003541 rcu_read_unlock();
Peter Zijlstraf492e122009-12-23 15:29:42 +01003542
3543 raw_spin_lock(&this_rq->lock);
3544
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003545 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3546 /*
3547 * We are going idle. next_balance may be set based on
3548 * a busy processor. So reset next_balance.
3549 */
3550 this_rq->next_balance = next_balance;
3551 }
3552}
3553
3554/*
Tejun Heo969c7922010-05-06 18:49:21 +02003555 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3556 * running tasks off the busiest CPU onto idle CPUs. It requires at
3557 * least 1 task to be running on each physical CPU where possible, and
3558 * avoids physical / logical imbalances.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003559 */
Tejun Heo969c7922010-05-06 18:49:21 +02003560static int active_load_balance_cpu_stop(void *data)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003561{
Tejun Heo969c7922010-05-06 18:49:21 +02003562 struct rq *busiest_rq = data;
3563 int busiest_cpu = cpu_of(busiest_rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003564 int target_cpu = busiest_rq->push_cpu;
Tejun Heo969c7922010-05-06 18:49:21 +02003565 struct rq *target_rq = cpu_rq(target_cpu);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003566 struct sched_domain *sd;
Tejun Heo969c7922010-05-06 18:49:21 +02003567
3568 raw_spin_lock_irq(&busiest_rq->lock);
3569
3570 /* make sure the requested cpu hasn't gone down in the meantime */
3571 if (unlikely(busiest_cpu != smp_processor_id() ||
3572 !busiest_rq->active_balance))
3573 goto out_unlock;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003574
3575 /* Is there any task to move? */
3576 if (busiest_rq->nr_running <= 1)
Tejun Heo969c7922010-05-06 18:49:21 +02003577 goto out_unlock;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003578
3579 /*
3580 * This condition is "impossible", if it occurs
3581 * we need to fix it. Originally reported by
3582 * Bjorn Helgaas on a 128-cpu setup.
3583 */
3584 BUG_ON(busiest_rq == target_rq);
3585
3586 /* move a task from busiest_rq to target_rq */
3587 double_lock_balance(busiest_rq, target_rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003588
3589 /* Search for an sd spanning us and the target CPU. */
Peter Zijlstradce840a2011-04-07 14:09:50 +02003590 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003591 for_each_domain(target_cpu, sd) {
3592 if ((sd->flags & SD_LOAD_BALANCE) &&
3593 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3594 break;
3595 }
3596
3597 if (likely(sd)) {
3598 schedstat_inc(sd, alb_count);
3599
3600 if (move_one_task(target_rq, target_cpu, busiest_rq,
3601 sd, CPU_IDLE))
3602 schedstat_inc(sd, alb_pushed);
3603 else
3604 schedstat_inc(sd, alb_failed);
3605 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02003606 rcu_read_unlock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003607 double_unlock_balance(busiest_rq, target_rq);
Tejun Heo969c7922010-05-06 18:49:21 +02003608out_unlock:
3609 busiest_rq->active_balance = 0;
3610 raw_spin_unlock_irq(&busiest_rq->lock);
3611 return 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003612}
3613
3614#ifdef CONFIG_NO_HZ
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003615
3616static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3617
3618static void trigger_sched_softirq(void *data)
3619{
3620 raise_softirq_irqoff(SCHED_SOFTIRQ);
3621}
3622
3623static inline void init_sched_softirq_csd(struct call_single_data *csd)
3624{
3625 csd->func = trigger_sched_softirq;
3626 csd->info = NULL;
3627 csd->flags = 0;
3628 csd->priv = 0;
3629}
3630
3631/*
3632 * idle load balancing details
3633 * - One of the idle CPUs nominates itself as idle load_balancer, while
3634 * entering idle.
3635 * - This idle load balancer CPU will also go into tickless mode when
3636 * it is idle, just like all other idle CPUs
3637 * - When one of the busy CPUs notice that there may be an idle rebalancing
3638 * needed, they will kick the idle load balancer, which then does idle
3639 * load balancing for all the idle CPUs.
3640 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003641static struct {
3642 atomic_t load_balancer;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003643 atomic_t first_pick_cpu;
3644 atomic_t second_pick_cpu;
3645 cpumask_var_t idle_cpus_mask;
3646 cpumask_var_t grp_idle_mask;
3647 unsigned long next_balance; /* in jiffy units */
3648} nohz ____cacheline_aligned;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003649
3650int get_nohz_load_balancer(void)
3651{
3652 return atomic_read(&nohz.load_balancer);
3653}
3654
3655#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3656/**
3657 * lowest_flag_domain - Return lowest sched_domain containing flag.
3658 * @cpu: The cpu whose lowest level of sched domain is to
3659 * be returned.
3660 * @flag: The flag to check for the lowest sched_domain
3661 * for the given cpu.
3662 *
3663 * Returns the lowest sched_domain of a cpu which contains the given flag.
3664 */
3665static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3666{
3667 struct sched_domain *sd;
3668
3669 for_each_domain(cpu, sd)
3670 if (sd && (sd->flags & flag))
3671 break;
3672
3673 return sd;
3674}
3675
3676/**
3677 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3678 * @cpu: The cpu whose domains we're iterating over.
3679 * @sd: variable holding the value of the power_savings_sd
3680 * for cpu.
3681 * @flag: The flag to filter the sched_domains to be iterated.
3682 *
3683 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3684 * set, starting from the lowest sched_domain to the highest.
3685 */
3686#define for_each_flag_domain(cpu, sd, flag) \
3687 for (sd = lowest_flag_domain(cpu, flag); \
3688 (sd && (sd->flags & flag)); sd = sd->parent)
3689
3690/**
3691 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3692 * @ilb_group: group to be checked for semi-idleness
3693 *
3694 * Returns: 1 if the group is semi-idle. 0 otherwise.
3695 *
3696 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3697 * and atleast one non-idle CPU. This helper function checks if the given
3698 * sched_group is semi-idle or not.
3699 */
3700static inline int is_semi_idle_group(struct sched_group *ilb_group)
3701{
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003702 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003703 sched_group_cpus(ilb_group));
3704
3705 /*
3706 * A sched_group is semi-idle when it has atleast one busy cpu
3707 * and atleast one idle cpu.
3708 */
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003709 if (cpumask_empty(nohz.grp_idle_mask))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003710 return 0;
3711
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003712 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003713 return 0;
3714
3715 return 1;
3716}
3717/**
3718 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3719 * @cpu: The cpu which is nominating a new idle_load_balancer.
3720 *
3721 * Returns: Returns the id of the idle load balancer if it exists,
3722 * Else, returns >= nr_cpu_ids.
3723 *
3724 * This algorithm picks the idle load balancer such that it belongs to a
3725 * semi-idle powersavings sched_domain. The idea is to try and avoid
3726 * completely idle packages/cores just for the purpose of idle load balancing
3727 * when there are other idle cpu's which are better suited for that job.
3728 */
3729static int find_new_ilb(int cpu)
3730{
3731 struct sched_domain *sd;
3732 struct sched_group *ilb_group;
Peter Zijlstradce840a2011-04-07 14:09:50 +02003733 int ilb = nr_cpu_ids;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003734
3735 /*
3736 * Have idle load balancer selection from semi-idle packages only
3737 * when power-aware load balancing is enabled
3738 */
3739 if (!(sched_smt_power_savings || sched_mc_power_savings))
3740 goto out_done;
3741
3742 /*
3743 * Optimize for the case when we have no idle CPUs or only one
3744 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3745 */
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003746 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003747 goto out_done;
3748
Peter Zijlstradce840a2011-04-07 14:09:50 +02003749 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003750 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3751 ilb_group = sd->groups;
3752
3753 do {
Peter Zijlstradce840a2011-04-07 14:09:50 +02003754 if (is_semi_idle_group(ilb_group)) {
3755 ilb = cpumask_first(nohz.grp_idle_mask);
3756 goto unlock;
3757 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003758
3759 ilb_group = ilb_group->next;
3760
3761 } while (ilb_group != sd->groups);
3762 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02003763unlock:
3764 rcu_read_unlock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003765
3766out_done:
Peter Zijlstradce840a2011-04-07 14:09:50 +02003767 return ilb;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003768}
3769#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3770static inline int find_new_ilb(int call_cpu)
3771{
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003772 return nr_cpu_ids;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003773}
3774#endif
3775
3776/*
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003777 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3778 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3779 * CPU (if there is one).
3780 */
3781static void nohz_balancer_kick(int cpu)
3782{
3783 int ilb_cpu;
3784
3785 nohz.next_balance++;
3786
3787 ilb_cpu = get_nohz_load_balancer();
3788
3789 if (ilb_cpu >= nr_cpu_ids) {
3790 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3791 if (ilb_cpu >= nr_cpu_ids)
3792 return;
3793 }
3794
3795 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3796 struct call_single_data *cp;
3797
3798 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3799 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3800 __smp_call_function_single(ilb_cpu, cp, 0);
3801 }
3802 return;
3803}
3804
3805/*
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003806 * This routine will try to nominate the ilb (idle load balancing)
3807 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003808 * load balancing on behalf of all those cpus.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003809 *
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003810 * When the ilb owner becomes busy, we will not have new ilb owner until some
3811 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3812 * idle load balancing by kicking one of the idle CPUs.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003813 *
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003814 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3815 * ilb owner CPU in future (when there is a need for idle load balancing on
3816 * behalf of all idle CPUs).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003817 */
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003818void select_nohz_load_balancer(int stop_tick)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003819{
3820 int cpu = smp_processor_id();
3821
3822 if (stop_tick) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003823 if (!cpu_active(cpu)) {
3824 if (atomic_read(&nohz.load_balancer) != cpu)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003825 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003826
3827 /*
3828 * If we are going offline and still the leader,
3829 * give up!
3830 */
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003831 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3832 nr_cpu_ids) != cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003833 BUG();
3834
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003835 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003836 }
3837
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003838 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003839
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003840 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3841 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3842 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3843 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003844
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003845 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003846 int new_ilb;
3847
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003848 /* make me the ilb owner */
3849 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3850 cpu) != nr_cpu_ids)
3851 return;
3852
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003853 /*
3854 * Check to see if there is a more power-efficient
3855 * ilb.
3856 */
3857 new_ilb = find_new_ilb(cpu);
3858 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003859 atomic_set(&nohz.load_balancer, nr_cpu_ids);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003860 resched_cpu(new_ilb);
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003861 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003862 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003863 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003864 }
3865 } else {
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003866 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3867 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003868
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003869 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003870
3871 if (atomic_read(&nohz.load_balancer) == cpu)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003872 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3873 nr_cpu_ids) != cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003874 BUG();
3875 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003876 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003877}
3878#endif
3879
3880static DEFINE_SPINLOCK(balancing);
3881
Peter Zijlstra49c022e2011-04-05 10:14:25 +02003882static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3883
3884/*
3885 * Scale the max load_balance interval with the number of CPUs in the system.
3886 * This trades load-balance latency on larger machines for less cross talk.
3887 */
3888static void update_max_interval(void)
3889{
3890 max_load_balance_interval = HZ*num_online_cpus()/10;
3891}
3892
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003893/*
3894 * It checks each scheduling domain to see if it is due to be balanced,
3895 * and initiates a balancing operation if so.
3896 *
3897 * Balancing parameters are set up in arch_init_sched_domains.
3898 */
3899static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3900{
3901 int balance = 1;
3902 struct rq *rq = cpu_rq(cpu);
3903 unsigned long interval;
3904 struct sched_domain *sd;
3905 /* Earliest time when we have to do rebalance again */
3906 unsigned long next_balance = jiffies + 60*HZ;
3907 int update_next_balance = 0;
3908 int need_serialize;
3909
Peter Zijlstra2069dd72010-11-15 15:47:00 -08003910 update_shares(cpu);
3911
Peter Zijlstradce840a2011-04-07 14:09:50 +02003912 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003913 for_each_domain(cpu, sd) {
3914 if (!(sd->flags & SD_LOAD_BALANCE))
3915 continue;
3916
3917 interval = sd->balance_interval;
3918 if (idle != CPU_IDLE)
3919 interval *= sd->busy_factor;
3920
3921 /* scale ms to jiffies */
3922 interval = msecs_to_jiffies(interval);
Peter Zijlstra49c022e2011-04-05 10:14:25 +02003923 interval = clamp(interval, 1UL, max_load_balance_interval);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003924
3925 need_serialize = sd->flags & SD_SERIALIZE;
3926
3927 if (need_serialize) {
3928 if (!spin_trylock(&balancing))
3929 goto out;
3930 }
3931
3932 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3933 if (load_balance(cpu, rq, sd, idle, &balance)) {
3934 /*
3935 * We've pulled tasks over so either we're no
Peter Zijlstrac186faf2011-02-21 18:52:53 +01003936 * longer idle.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003937 */
3938 idle = CPU_NOT_IDLE;
3939 }
3940 sd->last_balance = jiffies;
3941 }
3942 if (need_serialize)
3943 spin_unlock(&balancing);
3944out:
3945 if (time_after(next_balance, sd->last_balance + interval)) {
3946 next_balance = sd->last_balance + interval;
3947 update_next_balance = 1;
3948 }
3949
3950 /*
3951 * Stop the load balance at this level. There is another
3952 * CPU in our sched group which is doing load balancing more
3953 * actively.
3954 */
3955 if (!balance)
3956 break;
3957 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02003958 rcu_read_unlock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01003959
3960 /*
3961 * next_balance will be updated only when there is a need.
3962 * When the cpu is attached to null domain for ex, it will not be
3963 * updated.
3964 */
3965 if (likely(update_next_balance))
3966 rq->next_balance = next_balance;
3967}
3968
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003969#ifdef CONFIG_NO_HZ
3970/*
3971 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3972 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3973 */
3974static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3975{
3976 struct rq *this_rq = cpu_rq(this_cpu);
3977 struct rq *rq;
3978 int balance_cpu;
3979
3980 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3981 return;
3982
3983 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3984 if (balance_cpu == this_cpu)
3985 continue;
3986
3987 /*
3988 * If this cpu gets work to do, stop the load balancing
3989 * work being done for other cpus. Next load
3990 * balancing owner will pick it up.
3991 */
3992 if (need_resched()) {
3993 this_rq->nohz_balance_kick = 0;
3994 break;
3995 }
3996
3997 raw_spin_lock_irq(&this_rq->lock);
Suresh Siddha5343bdb2010-07-09 15:19:54 +02003998 update_rq_clock(this_rq);
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07003999 update_cpu_load(this_rq);
4000 raw_spin_unlock_irq(&this_rq->lock);
4001
4002 rebalance_domains(balance_cpu, CPU_IDLE);
4003
4004 rq = cpu_rq(balance_cpu);
4005 if (time_after(this_rq->next_balance, rq->next_balance))
4006 this_rq->next_balance = rq->next_balance;
4007 }
4008 nohz.next_balance = this_rq->next_balance;
4009 this_rq->nohz_balance_kick = 0;
4010}
4011
4012/*
4013 * Current heuristic for kicking the idle load balancer
4014 * - first_pick_cpu is the one of the busy CPUs. It will kick
4015 * idle load balancer when it has more than one process active. This
4016 * eliminates the need for idle load balancing altogether when we have
4017 * only one running process in the system (common case).
4018 * - If there are more than one busy CPU, idle load balancer may have
4019 * to run for active_load_balance to happen (i.e., two busy CPUs are
4020 * SMT or core siblings and can run better if they move to different
4021 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4022 * which will kick idle load balancer as soon as it has any load.
4023 */
4024static inline int nohz_kick_needed(struct rq *rq, int cpu)
4025{
4026 unsigned long now = jiffies;
4027 int ret;
4028 int first_pick_cpu, second_pick_cpu;
4029
4030 if (time_before(now, nohz.next_balance))
4031 return 0;
4032
Suresh Siddhaf6c3f162010-09-13 11:02:21 -07004033 if (rq->idle_at_tick)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07004034 return 0;
4035
4036 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4037 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4038
4039 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4040 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4041 return 0;
4042
4043 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4044 if (ret == nr_cpu_ids || ret == cpu) {
4045 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4046 if (rq->nr_running > 1)
4047 return 1;
4048 } else {
4049 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4050 if (ret == nr_cpu_ids || ret == cpu) {
4051 if (rq->nr_running)
4052 return 1;
4053 }
4054 }
4055 return 0;
4056}
4057#else
4058static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4059#endif
4060
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004061/*
4062 * run_rebalance_domains is triggered when needed from the scheduler tick.
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07004063 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004064 */
4065static void run_rebalance_domains(struct softirq_action *h)
4066{
4067 int this_cpu = smp_processor_id();
4068 struct rq *this_rq = cpu_rq(this_cpu);
4069 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4070 CPU_IDLE : CPU_NOT_IDLE;
4071
4072 rebalance_domains(this_cpu, idle);
4073
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004074 /*
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07004075 * If this cpu has a pending nohz_balance_kick, then do the
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004076 * balancing on behalf of the other idle cpus whose ticks are
4077 * stopped.
4078 */
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07004079 nohz_idle_balance(this_cpu, idle);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004080}
4081
4082static inline int on_null_domain(int cpu)
4083{
Paul E. McKenney90a65012010-02-28 08:32:18 -08004084 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004085}
4086
4087/*
4088 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004089 */
4090static inline void trigger_load_balance(struct rq *rq, int cpu)
4091{
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004092 /* Don't need to rebalance while attached to NULL domain */
4093 if (time_after_eq(jiffies, rq->next_balance) &&
4094 likely(!on_null_domain(cpu)))
4095 raise_softirq(SCHED_SOFTIRQ);
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07004096#ifdef CONFIG_NO_HZ
4097 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4098 nohz_balancer_kick(cpu);
4099#endif
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004100}
4101
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01004102static void rq_online_fair(struct rq *rq)
4103{
4104 update_sysctl();
4105}
4106
4107static void rq_offline_fair(struct rq *rq)
4108{
4109 update_sysctl();
4110}
4111
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01004112#else /* CONFIG_SMP */
4113
4114/*
4115 * on UP we do not need to balance between CPUs:
4116 */
4117static inline void idle_balance(int cpu, struct rq *rq)
4118{
4119}
4120
Dhaval Giani55e12e52008-06-24 23:39:43 +05304121#endif /* CONFIG_SMP */
Peter Williamse1d14842007-10-24 18:23:51 +02004122
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004123/*
4124 * scheduler tick hitting a task of our scheduling class:
4125 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004126static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004127{
4128 struct cfs_rq *cfs_rq;
4129 struct sched_entity *se = &curr->se;
4130
4131 for_each_sched_entity(se) {
4132 cfs_rq = cfs_rq_of(se);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004133 entity_tick(cfs_rq, se, queued);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004134 }
4135}
4136
4137/*
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004138 * called on fork with the child task as argument from the parent's context
4139 * - child not yet on the tasklist
4140 * - preemption disabled
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004141 */
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004142static void task_fork_fair(struct task_struct *p)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004143{
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004144 struct cfs_rq *cfs_rq = task_cfs_rq(current);
Ingo Molnar429d43b2007-10-15 17:00:03 +02004145 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
Ingo Molnar00bf7bf2007-10-15 17:00:14 +02004146 int this_cpu = smp_processor_id();
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004147 struct rq *rq = this_rq();
4148 unsigned long flags;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004149
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004150 raw_spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004151
Peter Zijlstra861d0342010-08-19 13:31:43 +02004152 update_rq_clock(rq);
4153
Paul E. McKenneyb0a0f662010-10-06 17:32:51 -07004154 if (unlikely(task_cpu(p) != this_cpu)) {
4155 rcu_read_lock();
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004156 __set_task_cpu(p, this_cpu);
Paul E. McKenneyb0a0f662010-10-06 17:32:51 -07004157 rcu_read_unlock();
4158 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004159
Ting Yang7109c4422007-08-28 12:53:24 +02004160 update_curr(cfs_rq);
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004161
Mike Galbraithb5d9d732009-09-08 11:12:28 +02004162 if (curr)
4163 se->vruntime = curr->vruntime;
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02004164 place_entity(cfs_rq, se, 1);
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02004165
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004166 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
Dmitry Adamushko87fefa32007-10-15 17:00:08 +02004167 /*
Ingo Molnaredcb60a2007-10-15 17:00:08 +02004168 * Upon rescheduling, sched_class::put_prev_task() will place
4169 * 'current' within the tree based on its new key value.
4170 */
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02004171 swap(curr->vruntime, se->vruntime);
Bharata B Raoaec0a512008-08-28 14:42:49 +05304172 resched_task(rq->curr);
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02004173 }
4174
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01004175 se->vruntime -= cfs_rq->min_vruntime;
4176
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004177 raw_spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004178}
4179
Steven Rostedtcb469842008-01-25 21:08:22 +01004180/*
4181 * Priority of the task has changed. Check to see if we preempt
4182 * the current task.
4183 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004184static void
4185prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01004186{
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004187 if (!p->se.on_rq)
4188 return;
4189
Steven Rostedtcb469842008-01-25 21:08:22 +01004190 /*
4191 * Reschedule if we are currently running on this runqueue and
4192 * our priority decreased, or if we are not currently running on
4193 * this runqueue and our priority is higher than the current's
4194 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004195 if (rq->curr == p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01004196 if (p->prio > oldprio)
4197 resched_task(rq->curr);
4198 } else
Peter Zijlstra15afe092008-09-20 23:38:02 +02004199 check_preempt_curr(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004200}
4201
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004202static void switched_from_fair(struct rq *rq, struct task_struct *p)
4203{
4204 struct sched_entity *se = &p->se;
4205 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4206
4207 /*
4208 * Ensure the task's vruntime is normalized, so that when its
4209 * switched back to the fair class the enqueue_entity(.flags=0) will
4210 * do the right thing.
4211 *
4212 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4213 * have normalized the vruntime, if it was !on_rq, then only when
4214 * the task is sleeping will it still have non-normalized vruntime.
4215 */
4216 if (!se->on_rq && p->state != TASK_RUNNING) {
4217 /*
4218 * Fix up our vruntime so that the current sleep doesn't
4219 * cause 'unlimited' sleep bonus.
4220 */
4221 place_entity(cfs_rq, se, 0);
4222 se->vruntime -= cfs_rq->min_vruntime;
4223 }
4224}
4225
Steven Rostedtcb469842008-01-25 21:08:22 +01004226/*
4227 * We switched to the sched_fair class.
4228 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004229static void switched_to_fair(struct rq *rq, struct task_struct *p)
Steven Rostedtcb469842008-01-25 21:08:22 +01004230{
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004231 if (!p->se.on_rq)
4232 return;
4233
Steven Rostedtcb469842008-01-25 21:08:22 +01004234 /*
4235 * We were most likely switched from sched_rt, so
4236 * kick off the schedule if running, otherwise just see
4237 * if we can still preempt the current task.
4238 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004239 if (rq->curr == p)
Steven Rostedtcb469842008-01-25 21:08:22 +01004240 resched_task(rq->curr);
4241 else
Peter Zijlstra15afe092008-09-20 23:38:02 +02004242 check_preempt_curr(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004243}
4244
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004245/* Account for a task changing its policy or group.
4246 *
4247 * This routine is mostly called to set cfs_rq->curr field when a task
4248 * migrates between groups/classes.
4249 */
4250static void set_curr_task_fair(struct rq *rq)
4251{
4252 struct sched_entity *se = &rq->curr->se;
4253
4254 for_each_sched_entity(se)
4255 set_next_entity(cfs_rq_of(se), se);
4256}
4257
Peter Zijlstra810b3812008-02-29 15:21:01 -05004258#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02004259static void task_move_group_fair(struct task_struct *p, int on_rq)
Peter Zijlstra810b3812008-02-29 15:21:01 -05004260{
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02004261 /*
4262 * If the task was not on the rq at the time of this cgroup movement
4263 * it must have been asleep, sleeping tasks keep their ->vruntime
4264 * absolute on their old rq until wakeup (needed for the fair sleeper
4265 * bonus in place_entity()).
4266 *
4267 * If it was on the rq, we've just 'preempted' it, which does convert
4268 * ->vruntime to a relative base.
4269 *
4270 * Make sure both cases convert their relative position when migrating
4271 * to another cgroup's rq. This does somewhat interfere with the
4272 * fair sleeper stuff for the first placement, but who cares.
4273 */
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01004274 if (!on_rq)
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02004275 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4276 set_task_rq(p, task_cpu(p));
4277 if (!on_rq)
4278 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
Peter Zijlstra810b3812008-02-29 15:21:01 -05004279}
4280#endif
4281
H Hartley Sweeten6d686f42010-01-13 20:21:52 -07004282static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
Peter Williams0d721ce2009-09-21 01:31:53 +00004283{
4284 struct sched_entity *se = &task->se;
Peter Williams0d721ce2009-09-21 01:31:53 +00004285 unsigned int rr_interval = 0;
4286
4287 /*
4288 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4289 * idle runqueue:
4290 */
Peter Williams0d721ce2009-09-21 01:31:53 +00004291 if (rq->cfs.load.weight)
4292 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Peter Williams0d721ce2009-09-21 01:31:53 +00004293
4294 return rr_interval;
4295}
4296
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004297/*
4298 * All the scheduling class methods:
4299 */
Ingo Molnar5522d5d2007-10-15 17:00:12 +02004300static const struct sched_class fair_sched_class = {
4301 .next = &idle_sched_class,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004302 .enqueue_task = enqueue_task_fair,
4303 .dequeue_task = dequeue_task_fair,
4304 .yield_task = yield_task_fair,
Mike Galbraithd95f4122011-02-01 09:50:51 -05004305 .yield_to_task = yield_to_task_fair,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004306
Ingo Molnar2e09bf52007-10-15 17:00:05 +02004307 .check_preempt_curr = check_preempt_wakeup,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004308
4309 .pick_next_task = pick_next_task_fair,
4310 .put_prev_task = put_prev_task_fair,
4311
Peter Williams681f3e62007-10-24 18:23:51 +02004312#ifdef CONFIG_SMP
Li Zefan4ce72a22008-10-22 15:25:26 +08004313 .select_task_rq = select_task_rq_fair,
4314
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01004315 .rq_online = rq_online_fair,
4316 .rq_offline = rq_offline_fair,
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01004317
4318 .task_waking = task_waking_fair,
Peter Williams681f3e62007-10-24 18:23:51 +02004319#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004320
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004321 .set_curr_task = set_curr_task_fair,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004322 .task_tick = task_tick_fair,
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004323 .task_fork = task_fork_fair,
Steven Rostedtcb469842008-01-25 21:08:22 +01004324
4325 .prio_changed = prio_changed_fair,
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004326 .switched_from = switched_from_fair,
Steven Rostedtcb469842008-01-25 21:08:22 +01004327 .switched_to = switched_to_fair,
Peter Zijlstra810b3812008-02-29 15:21:01 -05004328
Peter Williams0d721ce2009-09-21 01:31:53 +00004329 .get_rr_interval = get_rr_interval_fair,
4330
Peter Zijlstra810b3812008-02-29 15:21:01 -05004331#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02004332 .task_move_group = task_move_group_fair,
Peter Zijlstra810b3812008-02-29 15:21:01 -05004333#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004334};
4335
4336#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar5cef9ec2007-08-09 11:16:47 +02004337static void print_cfs_stats(struct seq_file *m, int cpu)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004338{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004339 struct cfs_rq *cfs_rq;
4340
Peter Zijlstra5973e5b2008-01-25 21:08:34 +01004341 rcu_read_lock();
Ingo Molnarc3b64f12007-08-09 11:16:51 +02004342 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
Ingo Molnar5cef9ec2007-08-09 11:16:47 +02004343 print_cfs_rq(m, cpu, cfs_rq);
Peter Zijlstra5973e5b2008-01-25 21:08:34 +01004344 rcu_read_unlock();
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004345}
4346#endif