blob: bd2267ad404fa78de092b8bc8f228cfea2dca87b [file] [log] [blame]
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001/*
2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3 * policies)
4 */
5
Peter Zijlstra029632f2011-10-25 10:00:11 +02006#include "sched.h"
7
8#include <linux/slab.h>
9
Clark Williamsce0dbbb2013-02-07 09:47:04 -060010int sched_rr_timeslice = RR_TIMESLICE;
11
Peter Zijlstra029632f2011-10-25 10:00:11 +020012static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
13
14struct rt_bandwidth def_rt_bandwidth;
15
16static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
17{
18 struct rt_bandwidth *rt_b =
19 container_of(timer, struct rt_bandwidth, rt_period_timer);
20 ktime_t now;
21 int overrun;
22 int idle = 0;
23
24 for (;;) {
25 now = hrtimer_cb_get_time(timer);
26 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
27
28 if (!overrun)
29 break;
30
31 idle = do_sched_rt_period_timer(rt_b, overrun);
32 }
33
34 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
35}
36
37void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
38{
39 rt_b->rt_period = ns_to_ktime(period);
40 rt_b->rt_runtime = runtime;
41
42 raw_spin_lock_init(&rt_b->rt_runtime_lock);
43
44 hrtimer_init(&rt_b->rt_period_timer,
45 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
46 rt_b->rt_period_timer.function = sched_rt_period_timer;
47}
48
49static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
50{
51 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
52 return;
53
54 if (hrtimer_active(&rt_b->rt_period_timer))
55 return;
56
57 raw_spin_lock(&rt_b->rt_runtime_lock);
58 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
59 raw_spin_unlock(&rt_b->rt_runtime_lock);
60}
61
62void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
63{
64 struct rt_prio_array *array;
65 int i;
66
67 array = &rt_rq->active;
68 for (i = 0; i < MAX_RT_PRIO; i++) {
69 INIT_LIST_HEAD(array->queue + i);
70 __clear_bit(i, array->bitmap);
71 }
72 /* delimiter for bitsearch: */
73 __set_bit(MAX_RT_PRIO, array->bitmap);
74
75#if defined CONFIG_SMP
76 rt_rq->highest_prio.curr = MAX_RT_PRIO;
77 rt_rq->highest_prio.next = MAX_RT_PRIO;
78 rt_rq->rt_nr_migratory = 0;
79 rt_rq->overloaded = 0;
80 plist_head_init(&rt_rq->pushable_tasks);
81#endif
82
83 rt_rq->rt_time = 0;
84 rt_rq->rt_throttled = 0;
85 rt_rq->rt_runtime = 0;
86 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
87}
88
Gregory Haskins398a1532009-01-14 09:10:04 -050089#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra029632f2011-10-25 10:00:11 +020090static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
91{
92 hrtimer_cancel(&rt_b->rt_period_timer);
93}
Gregory Haskins398a1532009-01-14 09:10:04 -050094
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +020095#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
96
Peter Zijlstra8f488942009-07-24 12:25:30 +020097static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
98{
99#ifdef CONFIG_SCHED_DEBUG
100 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
101#endif
102 return container_of(rt_se, struct task_struct, rt);
103}
104
Gregory Haskins398a1532009-01-14 09:10:04 -0500105static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
106{
107 return rt_rq->rq;
108}
109
110static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
111{
112 return rt_se->rt_rq;
113}
114
Peter Zijlstra029632f2011-10-25 10:00:11 +0200115void free_rt_sched_group(struct task_group *tg)
116{
117 int i;
118
119 if (tg->rt_se)
120 destroy_rt_bandwidth(&tg->rt_bandwidth);
121
122 for_each_possible_cpu(i) {
123 if (tg->rt_rq)
124 kfree(tg->rt_rq[i]);
125 if (tg->rt_se)
126 kfree(tg->rt_se[i]);
127 }
128
129 kfree(tg->rt_rq);
130 kfree(tg->rt_se);
131}
132
133void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
134 struct sched_rt_entity *rt_se, int cpu,
135 struct sched_rt_entity *parent)
136{
137 struct rq *rq = cpu_rq(cpu);
138
139 rt_rq->highest_prio.curr = MAX_RT_PRIO;
140 rt_rq->rt_nr_boosted = 0;
141 rt_rq->rq = rq;
142 rt_rq->tg = tg;
143
144 tg->rt_rq[cpu] = rt_rq;
145 tg->rt_se[cpu] = rt_se;
146
147 if (!rt_se)
148 return;
149
150 if (!parent)
151 rt_se->rt_rq = &rq->rt;
152 else
153 rt_se->rt_rq = parent->my_q;
154
155 rt_se->my_q = rt_rq;
156 rt_se->parent = parent;
157 INIT_LIST_HEAD(&rt_se->run_list);
158}
159
160int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
161{
162 struct rt_rq *rt_rq;
163 struct sched_rt_entity *rt_se;
164 int i;
165
166 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
167 if (!tg->rt_rq)
168 goto err;
169 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
170 if (!tg->rt_se)
171 goto err;
172
173 init_rt_bandwidth(&tg->rt_bandwidth,
174 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
175
176 for_each_possible_cpu(i) {
177 rt_rq = kzalloc_node(sizeof(struct rt_rq),
178 GFP_KERNEL, cpu_to_node(i));
179 if (!rt_rq)
180 goto err;
181
182 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
183 GFP_KERNEL, cpu_to_node(i));
184 if (!rt_se)
185 goto err_free_rq;
186
187 init_rt_rq(rt_rq, cpu_rq(i));
188 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
189 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
190 }
191
192 return 1;
193
194err_free_rq:
195 kfree(rt_rq);
196err:
197 return 0;
198}
199
Gregory Haskins398a1532009-01-14 09:10:04 -0500200#else /* CONFIG_RT_GROUP_SCHED */
201
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200202#define rt_entity_is_task(rt_se) (1)
203
Peter Zijlstra8f488942009-07-24 12:25:30 +0200204static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
205{
206 return container_of(rt_se, struct task_struct, rt);
207}
208
Gregory Haskins398a1532009-01-14 09:10:04 -0500209static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
210{
211 return container_of(rt_rq, struct rq, rt);
212}
213
214static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
215{
216 struct task_struct *p = rt_task_of(rt_se);
217 struct rq *rq = task_rq(p);
218
219 return &rq->rt;
220}
221
Peter Zijlstra029632f2011-10-25 10:00:11 +0200222void free_rt_sched_group(struct task_group *tg) { }
223
224int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
225{
226 return 1;
227}
Gregory Haskins398a1532009-01-14 09:10:04 -0500228#endif /* CONFIG_RT_GROUP_SCHED */
229
Steven Rostedt4fd29172008-01-25 21:08:06 +0100230#ifdef CONFIG_SMP
Ingo Molnar84de4272008-01-25 21:08:15 +0100231
Peter Zijlstra38033c32014-01-23 20:32:21 +0100232static int pull_rt_task(struct rq *this_rq);
233
Peter Zijlstradc877342014-02-12 15:47:29 +0100234static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
235{
236 /* Try to pull RT tasks here if we lower this rq's prio */
237 return rq->rt.highest_prio.curr > prev->prio;
238}
239
Gregory Haskins637f5082008-01-25 21:08:18 +0100240static inline int rt_overloaded(struct rq *rq)
Steven Rostedt4fd29172008-01-25 21:08:06 +0100241{
Gregory Haskins637f5082008-01-25 21:08:18 +0100242 return atomic_read(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100243}
Ingo Molnar84de4272008-01-25 21:08:15 +0100244
Steven Rostedt4fd29172008-01-25 21:08:06 +0100245static inline void rt_set_overload(struct rq *rq)
246{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400247 if (!rq->online)
248 return;
249
Rusty Russellc6c49272008-11-25 02:35:05 +1030250 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100251 /*
252 * Make sure the mask is visible before we set
253 * the overload count. That is checked to determine
254 * if we should look at the mask. It would be a shame
255 * if we looked at the mask, but the mask was not
256 * updated yet.
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +0200257 *
258 * Matched by the barrier in pull_rt_task().
Steven Rostedt4fd29172008-01-25 21:08:06 +0100259 */
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +0200260 smp_wmb();
Gregory Haskins637f5082008-01-25 21:08:18 +0100261 atomic_inc(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100262}
Ingo Molnar84de4272008-01-25 21:08:15 +0100263
Steven Rostedt4fd29172008-01-25 21:08:06 +0100264static inline void rt_clear_overload(struct rq *rq)
265{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400266 if (!rq->online)
267 return;
268
Steven Rostedt4fd29172008-01-25 21:08:06 +0100269 /* the order here really doesn't matter */
Gregory Haskins637f5082008-01-25 21:08:18 +0100270 atomic_dec(&rq->rd->rto_count);
Rusty Russellc6c49272008-11-25 02:35:05 +1030271 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100272}
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100273
Gregory Haskins398a1532009-01-14 09:10:04 -0500274static void update_rt_migration(struct rt_rq *rt_rq)
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100275{
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200276 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
Gregory Haskins398a1532009-01-14 09:10:04 -0500277 if (!rt_rq->overloaded) {
278 rt_set_overload(rq_of_rt_rq(rt_rq));
279 rt_rq->overloaded = 1;
Gregory Haskinscdc8eb92008-01-25 21:08:23 +0100280 }
Gregory Haskins398a1532009-01-14 09:10:04 -0500281 } else if (rt_rq->overloaded) {
282 rt_clear_overload(rq_of_rt_rq(rt_rq));
283 rt_rq->overloaded = 0;
Gregory Haskins637f5082008-01-25 21:08:18 +0100284 }
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100285}
Steven Rostedt4fd29172008-01-25 21:08:06 +0100286
Gregory Haskins398a1532009-01-14 09:10:04 -0500287static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100288{
Peter Zijlstra29baa742012-04-23 12:11:21 +0200289 struct task_struct *p;
290
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200291 if (!rt_entity_is_task(rt_se))
292 return;
293
Peter Zijlstra29baa742012-04-23 12:11:21 +0200294 p = rt_task_of(rt_se);
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200295 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
296
297 rt_rq->rt_nr_total++;
Peter Zijlstra29baa742012-04-23 12:11:21 +0200298 if (p->nr_cpus_allowed > 1)
Gregory Haskins398a1532009-01-14 09:10:04 -0500299 rt_rq->rt_nr_migratory++;
300
301 update_rt_migration(rt_rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100302}
303
Gregory Haskins398a1532009-01-14 09:10:04 -0500304static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
305{
Peter Zijlstra29baa742012-04-23 12:11:21 +0200306 struct task_struct *p;
307
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200308 if (!rt_entity_is_task(rt_se))
309 return;
310
Peter Zijlstra29baa742012-04-23 12:11:21 +0200311 p = rt_task_of(rt_se);
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200312 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
313
314 rt_rq->rt_nr_total--;
Peter Zijlstra29baa742012-04-23 12:11:21 +0200315 if (p->nr_cpus_allowed > 1)
Gregory Haskins398a1532009-01-14 09:10:04 -0500316 rt_rq->rt_nr_migratory--;
317
318 update_rt_migration(rt_rq);
319}
320
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400321static inline int has_pushable_tasks(struct rq *rq)
322{
323 return !plist_head_empty(&rq->rt.pushable_tasks);
324}
325
Peter Zijlstradc877342014-02-12 15:47:29 +0100326static inline void set_post_schedule(struct rq *rq)
327{
328 /*
329 * We detect this state here so that we can avoid taking the RQ
330 * lock again later if there is no need to push
331 */
332 rq->post_schedule = has_pushable_tasks(rq);
333}
334
Gregory Haskins917b6272008-12-29 09:39:53 -0500335static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
336{
337 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
338 plist_node_init(&p->pushable_tasks, p->prio);
339 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400340
341 /* Update the highest prio pushable task */
342 if (p->prio < rq->rt.highest_prio.next)
343 rq->rt.highest_prio.next = p->prio;
Gregory Haskins917b6272008-12-29 09:39:53 -0500344}
345
346static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
347{
348 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
Gregory Haskins917b6272008-12-29 09:39:53 -0500349
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400350 /* Update the new highest prio pushable task */
351 if (has_pushable_tasks(rq)) {
352 p = plist_first_entry(&rq->rt.pushable_tasks,
353 struct task_struct, pushable_tasks);
354 rq->rt.highest_prio.next = p->prio;
355 } else
356 rq->rt.highest_prio.next = MAX_RT_PRIO;
Ingo Molnarbcf08df2008-04-19 12:11:10 +0200357}
358
Gregory Haskins917b6272008-12-29 09:39:53 -0500359#else
360
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100361static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
362{
363}
364
365static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
366{
367}
368
Gregory Haskinsb07430a2009-01-14 08:55:39 -0500369static inline
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100370void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
371{
372}
373
Gregory Haskinsb07430a2009-01-14 08:55:39 -0500374static inline
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100375void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
376{
377}
Gregory Haskins917b6272008-12-29 09:39:53 -0500378
Peter Zijlstradc877342014-02-12 15:47:29 +0100379static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
380{
381 return false;
382}
383
384static inline int pull_rt_task(struct rq *this_rq)
385{
386 return 0;
387}
388
389static inline void set_post_schedule(struct rq *rq)
390{
391}
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200392#endif /* CONFIG_SMP */
393
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100394static inline int on_rt_rq(struct sched_rt_entity *rt_se)
395{
396 return !list_empty(&rt_se->run_list);
397}
398
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100399#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100400
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100401static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100402{
403 if (!rt_rq->tg)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100404 return RUNTIME_INF;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100405
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200406 return rt_rq->rt_runtime;
407}
408
409static inline u64 sched_rt_period(struct rt_rq *rt_rq)
410{
411 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100412}
413
Cheng Xuec514c42011-05-14 14:20:02 +0800414typedef struct task_group *rt_rq_iter_t;
415
Yong Zhang1c09ab02011-06-28 10:51:31 +0800416static inline struct task_group *next_task_group(struct task_group *tg)
417{
418 do {
419 tg = list_entry_rcu(tg->list.next,
420 typeof(struct task_group), list);
421 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
422
423 if (&tg->list == &task_groups)
424 tg = NULL;
425
426 return tg;
427}
428
429#define for_each_rt_rq(rt_rq, iter, rq) \
430 for (iter = container_of(&task_groups, typeof(*iter), list); \
431 (iter = next_task_group(iter)) && \
432 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
Cheng Xuec514c42011-05-14 14:20:02 +0800433
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100434#define for_each_sched_rt_entity(rt_se) \
435 for (; rt_se; rt_se = rt_se->parent)
436
437static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
438{
439 return rt_se->my_q;
440}
441
Thomas Gleixner37dad3f2010-01-20 20:59:01 +0000442static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100443static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
444
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100445static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100446{
Dario Faggiolif6121f42008-10-03 17:40:46 +0200447 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
Yong Zhang74b7eb52010-01-29 14:57:52 +0800448 struct sched_rt_entity *rt_se;
449
Balbir Singh0c3b9162011-03-03 17:04:35 +0530450 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
451
452 rt_se = rt_rq->tg->rt_se[cpu];
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100453
Dario Faggiolif6121f42008-10-03 17:40:46 +0200454 if (rt_rq->rt_nr_running) {
455 if (rt_se && !on_rt_rq(rt_se))
Thomas Gleixner37dad3f2010-01-20 20:59:01 +0000456 enqueue_rt_entity(rt_se, false);
Gregory Haskinse864c492008-12-29 09:39:49 -0500457 if (rt_rq->highest_prio.curr < curr->prio)
Peter Zijlstra10203872008-01-25 21:08:32 +0100458 resched_task(curr);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100459 }
460}
461
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100462static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100463{
Yong Zhang74b7eb52010-01-29 14:57:52 +0800464 struct sched_rt_entity *rt_se;
Balbir Singh0c3b9162011-03-03 17:04:35 +0530465 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
Yong Zhang74b7eb52010-01-29 14:57:52 +0800466
Balbir Singh0c3b9162011-03-03 17:04:35 +0530467 rt_se = rt_rq->tg->rt_se[cpu];
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100468
469 if (rt_se && on_rt_rq(rt_se))
470 dequeue_rt_entity(rt_se);
471}
472
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100473static int rt_se_boosted(struct sched_rt_entity *rt_se)
474{
475 struct rt_rq *rt_rq = group_rt_rq(rt_se);
476 struct task_struct *p;
477
478 if (rt_rq)
479 return !!rt_rq->rt_nr_boosted;
480
481 p = rt_task_of(rt_se);
482 return p->prio != p->normal_prio;
483}
484
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200485#ifdef CONFIG_SMP
Rusty Russellc6c49272008-11-25 02:35:05 +1030486static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200487{
Nathan Zimmer424c93f2013-05-09 11:24:03 -0500488 return this_rq()->rd->span;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200489}
490#else
Rusty Russellc6c49272008-11-25 02:35:05 +1030491static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200492{
Rusty Russellc6c49272008-11-25 02:35:05 +1030493 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200494}
495#endif
496
497static inline
498struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
499{
500 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
501}
502
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200503static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
504{
505 return &rt_rq->tg->rt_bandwidth;
506}
507
Dhaval Giani55e12e52008-06-24 23:39:43 +0530508#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100509
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100510static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100511{
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200512 return rt_rq->rt_runtime;
513}
514
515static inline u64 sched_rt_period(struct rt_rq *rt_rq)
516{
517 return ktime_to_ns(def_rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100518}
519
Cheng Xuec514c42011-05-14 14:20:02 +0800520typedef struct rt_rq *rt_rq_iter_t;
521
522#define for_each_rt_rq(rt_rq, iter, rq) \
523 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
524
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100525#define for_each_sched_rt_entity(rt_se) \
526 for (; rt_se; rt_se = NULL)
527
528static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
529{
530 return NULL;
531}
532
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100533static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100534{
John Blackwoodf3ade832008-08-26 15:09:43 -0400535 if (rt_rq->rt_nr_running)
536 resched_task(rq_of_rt_rq(rt_rq)->curr);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100537}
538
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100539static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100540{
541}
542
Rusty Russellc6c49272008-11-25 02:35:05 +1030543static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200544{
Rusty Russellc6c49272008-11-25 02:35:05 +1030545 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200546}
547
548static inline
549struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
550{
551 return &cpu_rq(cpu)->rt;
552}
553
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200554static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
555{
556 return &def_rt_bandwidth;
557}
558
Dhaval Giani55e12e52008-06-24 23:39:43 +0530559#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100560
Juri Lellifaa59932014-02-21 11:37:15 +0100561bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
562{
563 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
564
565 return (hrtimer_active(&rt_b->rt_period_timer) ||
566 rt_rq->rt_time < rt_b->rt_runtime);
567}
568
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200569#ifdef CONFIG_SMP
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200570/*
571 * We ran out of runtime, see if we can borrow some from our neighbours.
572 */
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200573static int do_balance_runtime(struct rt_rq *rt_rq)
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200574{
575 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
Shawn Bohreraa7f6732013-01-14 11:55:31 -0600576 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200577 int i, weight, more = 0;
578 u64 rt_period;
579
Rusty Russellc6c49272008-11-25 02:35:05 +1030580 weight = cpumask_weight(rd->span);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200581
Thomas Gleixner0986b112009-11-17 15:32:06 +0100582 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200583 rt_period = ktime_to_ns(rt_b->rt_period);
Rusty Russellc6c49272008-11-25 02:35:05 +1030584 for_each_cpu(i, rd->span) {
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200585 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
586 s64 diff;
587
588 if (iter == rt_rq)
589 continue;
590
Thomas Gleixner0986b112009-11-17 15:32:06 +0100591 raw_spin_lock(&iter->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200592 /*
593 * Either all rqs have inf runtime and there's nothing to steal
594 * or __disable_runtime() below sets a specific rq to inf to
595 * indicate its been disabled and disalow stealing.
596 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200597 if (iter->rt_runtime == RUNTIME_INF)
598 goto next;
599
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200600 /*
601 * From runqueues with spare time, take 1/n part of their
602 * spare time, but no more than our period.
603 */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200604 diff = iter->rt_runtime - iter->rt_time;
605 if (diff > 0) {
Peter Zijlstra58838cf2008-07-24 12:43:13 +0200606 diff = div_u64((u64)diff, weight);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200607 if (rt_rq->rt_runtime + diff > rt_period)
608 diff = rt_period - rt_rq->rt_runtime;
609 iter->rt_runtime -= diff;
610 rt_rq->rt_runtime += diff;
611 more = 1;
612 if (rt_rq->rt_runtime == rt_period) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100613 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200614 break;
615 }
616 }
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200617next:
Thomas Gleixner0986b112009-11-17 15:32:06 +0100618 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200619 }
Thomas Gleixner0986b112009-11-17 15:32:06 +0100620 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200621
622 return more;
623}
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200624
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200625/*
626 * Ensure this RQ takes back all the runtime it lend to its neighbours.
627 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200628static void __disable_runtime(struct rq *rq)
629{
630 struct root_domain *rd = rq->rd;
Cheng Xuec514c42011-05-14 14:20:02 +0800631 rt_rq_iter_t iter;
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200632 struct rt_rq *rt_rq;
633
634 if (unlikely(!scheduler_running))
635 return;
636
Cheng Xuec514c42011-05-14 14:20:02 +0800637 for_each_rt_rq(rt_rq, iter, rq) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200638 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
639 s64 want;
640 int i;
641
Thomas Gleixner0986b112009-11-17 15:32:06 +0100642 raw_spin_lock(&rt_b->rt_runtime_lock);
643 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200644 /*
645 * Either we're all inf and nobody needs to borrow, or we're
646 * already disabled and thus have nothing to do, or we have
647 * exactly the right amount of runtime to take out.
648 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200649 if (rt_rq->rt_runtime == RUNTIME_INF ||
650 rt_rq->rt_runtime == rt_b->rt_runtime)
651 goto balanced;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100652 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200653
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200654 /*
655 * Calculate the difference between what we started out with
656 * and what we current have, that's the amount of runtime
657 * we lend and now have to reclaim.
658 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200659 want = rt_b->rt_runtime - rt_rq->rt_runtime;
660
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200661 /*
662 * Greedy reclaim, take back as much as we can.
663 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030664 for_each_cpu(i, rd->span) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200665 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
666 s64 diff;
667
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200668 /*
669 * Can't reclaim from ourselves or disabled runqueues.
670 */
Peter Zijlstraf1679d02008-08-14 15:49:00 +0200671 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200672 continue;
673
Thomas Gleixner0986b112009-11-17 15:32:06 +0100674 raw_spin_lock(&iter->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200675 if (want > 0) {
676 diff = min_t(s64, iter->rt_runtime, want);
677 iter->rt_runtime -= diff;
678 want -= diff;
679 } else {
680 iter->rt_runtime -= want;
681 want -= want;
682 }
Thomas Gleixner0986b112009-11-17 15:32:06 +0100683 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200684
685 if (!want)
686 break;
687 }
688
Thomas Gleixner0986b112009-11-17 15:32:06 +0100689 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200690 /*
691 * We cannot be left wanting - that would mean some runtime
692 * leaked out of the system.
693 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200694 BUG_ON(want);
695balanced:
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200696 /*
697 * Disable all the borrow logic by pretending we have inf
698 * runtime - in which case borrowing doesn't make sense.
699 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200700 rt_rq->rt_runtime = RUNTIME_INF;
Peter Boonstoppela4c96ae2012-08-09 15:34:47 -0700701 rt_rq->rt_throttled = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100702 raw_spin_unlock(&rt_rq->rt_runtime_lock);
703 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200704 }
705}
706
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200707static void __enable_runtime(struct rq *rq)
708{
Cheng Xuec514c42011-05-14 14:20:02 +0800709 rt_rq_iter_t iter;
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200710 struct rt_rq *rt_rq;
711
712 if (unlikely(!scheduler_running))
713 return;
714
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200715 /*
716 * Reset each runqueue's bandwidth settings
717 */
Cheng Xuec514c42011-05-14 14:20:02 +0800718 for_each_rt_rq(rt_rq, iter, rq) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200719 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
720
Thomas Gleixner0986b112009-11-17 15:32:06 +0100721 raw_spin_lock(&rt_b->rt_runtime_lock);
722 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200723 rt_rq->rt_runtime = rt_b->rt_runtime;
724 rt_rq->rt_time = 0;
Zhang, Yanminbaf25732008-09-09 11:26:33 +0800725 rt_rq->rt_throttled = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100726 raw_spin_unlock(&rt_rq->rt_runtime_lock);
727 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200728 }
729}
730
Peter Zijlstraeff65492008-06-19 14:22:26 +0200731static int balance_runtime(struct rt_rq *rt_rq)
732{
733 int more = 0;
734
Peter Zijlstra4a6184c2011-10-06 22:39:14 +0200735 if (!sched_feat(RT_RUNTIME_SHARE))
736 return more;
737
Peter Zijlstraeff65492008-06-19 14:22:26 +0200738 if (rt_rq->rt_time > rt_rq->rt_runtime) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100739 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200740 more = do_balance_runtime(rt_rq);
Thomas Gleixner0986b112009-11-17 15:32:06 +0100741 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200742 }
743
744 return more;
745}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530746#else /* !CONFIG_SMP */
Peter Zijlstraeff65492008-06-19 14:22:26 +0200747static inline int balance_runtime(struct rt_rq *rt_rq)
748{
749 return 0;
750}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530751#endif /* CONFIG_SMP */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100752
753static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
754{
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200755 int i, idle = 1, throttled = 0;
Rusty Russellc6c49272008-11-25 02:35:05 +1030756 const struct cpumask *span;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200757
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200758 span = sched_rt_period_mask();
Mike Galbraithe221d022012-08-07 10:02:38 +0200759#ifdef CONFIG_RT_GROUP_SCHED
760 /*
761 * FIXME: isolated CPUs should really leave the root task group,
762 * whether they are isolcpus or were isolated via cpusets, lest
763 * the timer run on a CPU which does not service all runqueues,
764 * potentially leaving other CPUs indefinitely throttled. If
765 * isolation is really required, the user will turn the throttle
766 * off to kill the perturbations it causes anyway. Meanwhile,
767 * this maintains functionality for boot and/or troubleshooting.
768 */
769 if (rt_b == &root_task_group.rt_bandwidth)
770 span = cpu_online_mask;
771#endif
Rusty Russellc6c49272008-11-25 02:35:05 +1030772 for_each_cpu(i, span) {
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200773 int enqueue = 0;
774 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
775 struct rq *rq = rq_of_rt_rq(rt_rq);
776
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100777 raw_spin_lock(&rq->lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200778 if (rt_rq->rt_time) {
779 u64 runtime;
780
Thomas Gleixner0986b112009-11-17 15:32:06 +0100781 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200782 if (rt_rq->rt_throttled)
783 balance_runtime(rt_rq);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200784 runtime = rt_rq->rt_runtime;
785 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
786 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
787 rt_rq->rt_throttled = 0;
788 enqueue = 1;
Mike Galbraith61eadef2011-04-29 08:36:50 +0200789
790 /*
791 * Force a clock update if the CPU was idle,
792 * lest wakeup -> unthrottle time accumulate.
793 */
794 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
795 rq->skip_clock_update = -1;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200796 }
797 if (rt_rq->rt_time || rt_rq->rt_nr_running)
798 idle = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Balbir Singh0c3b9162011-03-03 17:04:35 +0530800 } else if (rt_rq->rt_nr_running) {
Peter Zijlstra8a8cde12008-06-19 14:22:28 +0200801 idle = 0;
Balbir Singh0c3b9162011-03-03 17:04:35 +0530802 if (!rt_rq_throttled(rt_rq))
803 enqueue = 1;
804 }
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200805 if (rt_rq->rt_throttled)
806 throttled = 1;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200807
808 if (enqueue)
809 sched_rt_rq_enqueue(rt_rq);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100810 raw_spin_unlock(&rq->lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200811 }
812
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200813 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
814 return 1;
815
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200816 return idle;
817}
818
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100819static inline int rt_se_prio(struct sched_rt_entity *rt_se)
820{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100821#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100822 struct rt_rq *rt_rq = group_rt_rq(rt_se);
823
824 if (rt_rq)
Gregory Haskinse864c492008-12-29 09:39:49 -0500825 return rt_rq->highest_prio.curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100826#endif
827
828 return rt_task_of(rt_se)->prio;
829}
830
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100831static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100832{
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100833 u64 runtime = sched_rt_runtime(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100834
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100835 if (rt_rq->rt_throttled)
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100836 return rt_rq_throttled(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100837
Shan Hai5b680fd2011-11-29 11:03:56 +0800838 if (runtime >= sched_rt_period(rt_rq))
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200839 return 0;
840
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200841 balance_runtime(rt_rq);
842 runtime = sched_rt_runtime(rt_rq);
843 if (runtime == RUNTIME_INF)
844 return 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200845
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100846 if (rt_rq->rt_time > runtime) {
Peter Zijlstra7abc63b2011-10-18 22:03:48 +0200847 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
848
849 /*
850 * Don't actually throttle groups that have no runtime assigned
851 * but accrue some time due to boosting.
852 */
853 if (likely(rt_b->rt_runtime)) {
Peter Zijlstra3ccf3e82012-02-27 10:47:00 +0100854 static bool once = false;
855
Peter Zijlstra7abc63b2011-10-18 22:03:48 +0200856 rt_rq->rt_throttled = 1;
Peter Zijlstra3ccf3e82012-02-27 10:47:00 +0100857
858 if (!once) {
859 once = true;
860 printk_sched("sched: RT throttling activated\n");
861 }
Peter Zijlstra7abc63b2011-10-18 22:03:48 +0200862 } else {
863 /*
864 * In case we did anyway, make it go away,
865 * replenishment is a joke, since it will replenish us
866 * with exactly 0 ns.
867 */
868 rt_rq->rt_time = 0;
869 }
870
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100871 if (rt_rq_throttled(rt_rq)) {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100872 sched_rt_rq_dequeue(rt_rq);
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100873 return 1;
874 }
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100875 }
876
877 return 0;
878}
879
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200880/*
881 * Update the current task's runtime statistics. Skip current tasks that
882 * are not in our scheduling class.
883 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200884static void update_curr_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200885{
886 struct task_struct *curr = rq->curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100887 struct sched_rt_entity *rt_se = &curr->rt;
888 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200889 u64 delta_exec;
890
Peter Zijlstra06c3bc62011-02-02 13:19:48 +0100891 if (curr->sched_class != &rt_sched_class)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200892 return;
893
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200894 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
Kirill Tkhaifc79e242013-01-30 16:50:36 +0400895 if (unlikely((s64)delta_exec <= 0))
896 return;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +0200897
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200898 schedstat_set(curr->se.statistics.exec_max,
899 max(curr->se.statistics.exec_max, delta_exec));
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200900
901 curr->se.sum_exec_runtime += delta_exec;
Frank Mayharf06febc2008-09-12 09:54:39 -0700902 account_group_exec_runtime(curr, delta_exec);
903
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200904 curr->se.exec_start = rq_clock_task(rq);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100905 cpuacct_charge(curr, delta_exec);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100906
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200907 sched_rt_avg_update(rq, delta_exec);
908
Peter Zijlstra0b148fa2008-08-19 12:33:04 +0200909 if (!rt_bandwidth_enabled())
910 return;
911
Dhaval Giani354d60c2008-04-19 19:44:59 +0200912 for_each_sched_rt_entity(rt_se) {
913 rt_rq = rt_rq_of_se(rt_se);
914
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200915 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100916 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200917 rt_rq->rt_time += delta_exec;
918 if (sched_rt_runtime_exceeded(rt_rq))
919 resched_task(curr);
Thomas Gleixner0986b112009-11-17 15:32:06 +0100920 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200921 }
Dhaval Giani354d60c2008-04-19 19:44:59 +0200922 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200923}
924
Gregory Haskins398a1532009-01-14 09:10:04 -0500925#if defined CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500926
Gregory Haskins398a1532009-01-14 09:10:04 -0500927static void
928inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +0100929{
Gregory Haskins4d984272008-12-29 09:39:49 -0500930 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins4d984272008-12-29 09:39:49 -0500931
Kirill Tkhai757dfca2013-11-27 19:59:13 +0400932#ifdef CONFIG_RT_GROUP_SCHED
933 /*
934 * Change rq's cpupri only if rt_rq is the top queue.
935 */
936 if (&rq->rt != rt_rq)
937 return;
938#endif
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400939 if (rq->online && prio < prev_prio)
940 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
Steven Rostedt63489e42008-01-25 21:08:03 +0100941}
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100942
Gregory Haskins398a1532009-01-14 09:10:04 -0500943static void
944dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +0100945{
Gregory Haskins4d984272008-12-29 09:39:49 -0500946 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200947
Kirill Tkhai757dfca2013-11-27 19:59:13 +0400948#ifdef CONFIG_RT_GROUP_SCHED
949 /*
950 * Change rq's cpupri only if rt_rq is the top queue.
951 */
952 if (&rq->rt != rt_rq)
953 return;
954#endif
Gregory Haskins398a1532009-01-14 09:10:04 -0500955 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
956 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
957}
958
959#else /* CONFIG_SMP */
960
961static inline
962void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
963static inline
964void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
965
966#endif /* CONFIG_SMP */
967
Steven Rostedt63489e42008-01-25 21:08:03 +0100968#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -0500969static void
970inc_rt_prio(struct rt_rq *rt_rq, int prio)
971{
972 int prev_prio = rt_rq->highest_prio.curr;
Steven Rostedt63489e42008-01-25 21:08:03 +0100973
Gregory Haskins398a1532009-01-14 09:10:04 -0500974 if (prio < prev_prio)
975 rt_rq->highest_prio.curr = prio;
976
977 inc_rt_prio_smp(rt_rq, prio, prev_prio);
978}
979
980static void
981dec_rt_prio(struct rt_rq *rt_rq, int prio)
982{
983 int prev_prio = rt_rq->highest_prio.curr;
984
985 if (rt_rq->rt_nr_running) {
986
987 WARN_ON(prio < prev_prio);
Gregory Haskinse864c492008-12-29 09:39:49 -0500988
989 /*
Gregory Haskins398a1532009-01-14 09:10:04 -0500990 * This may have been our highest task, and therefore
991 * we may have some recomputation to do
Gregory Haskinse864c492008-12-29 09:39:49 -0500992 */
Gregory Haskins398a1532009-01-14 09:10:04 -0500993 if (prio == prev_prio) {
Gregory Haskinse864c492008-12-29 09:39:49 -0500994 struct rt_prio_array *array = &rt_rq->active;
995
996 rt_rq->highest_prio.curr =
Steven Rostedt764a9d62008-01-25 21:08:04 +0100997 sched_find_first_bit(array->bitmap);
Gregory Haskinse864c492008-12-29 09:39:49 -0500998 }
999
Steven Rostedt764a9d62008-01-25 21:08:04 +01001000 } else
Gregory Haskinse864c492008-12-29 09:39:49 -05001001 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001002
Gregory Haskins398a1532009-01-14 09:10:04 -05001003 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1004}
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001005
Gregory Haskins398a1532009-01-14 09:10:04 -05001006#else
1007
1008static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1009static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1010
1011#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1012
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001013#ifdef CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -05001014
1015static void
1016inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1017{
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001018 if (rt_se_boosted(rt_se))
Steven Rostedt764a9d62008-01-25 21:08:04 +01001019 rt_rq->rt_nr_boosted++;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01001020
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001021 if (rt_rq->tg)
1022 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
Gregory Haskins398a1532009-01-14 09:10:04 -05001023}
1024
1025static void
1026dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1027{
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001028 if (rt_se_boosted(rt_se))
1029 rt_rq->rt_nr_boosted--;
1030
1031 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
Gregory Haskins398a1532009-01-14 09:10:04 -05001032}
1033
1034#else /* CONFIG_RT_GROUP_SCHED */
1035
1036static void
1037inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1038{
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001039 start_rt_bandwidth(&def_rt_bandwidth);
Gregory Haskins398a1532009-01-14 09:10:04 -05001040}
1041
1042static inline
1043void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1044
1045#endif /* CONFIG_RT_GROUP_SCHED */
1046
1047static inline
1048void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1049{
1050 int prio = rt_se_prio(rt_se);
1051
1052 WARN_ON(!rt_prio(prio));
1053 rt_rq->rt_nr_running++;
1054
1055 inc_rt_prio(rt_rq, prio);
1056 inc_rt_migration(rt_se, rt_rq);
1057 inc_rt_group(rt_se, rt_rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001058}
1059
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001060static inline
1061void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1062{
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001063 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001064 WARN_ON(!rt_rq->rt_nr_running);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001065 rt_rq->rt_nr_running--;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001066
Gregory Haskins398a1532009-01-14 09:10:04 -05001067 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1068 dec_rt_migration(rt_se, rt_rq);
1069 dec_rt_group(rt_se, rt_rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001070}
1071
Thomas Gleixner37dad3f2010-01-20 20:59:01 +00001072static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001073{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001074 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1075 struct rt_prio_array *array = &rt_rq->active;
1076 struct rt_rq *group_rq = group_rt_rq(rt_se);
Dmitry Adamushko20b63312008-06-11 00:58:30 +02001077 struct list_head *queue = array->queue + rt_se_prio(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001078
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001079 /*
1080 * Don't enqueue the group if its throttled, or when empty.
1081 * The latter is a consequence of the former when a child group
1082 * get throttled and the current group doesn't have any other
1083 * active members.
1084 */
1085 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001086 return;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001087
Thomas Gleixner37dad3f2010-01-20 20:59:01 +00001088 if (head)
1089 list_add(&rt_se->run_list, queue);
1090 else
1091 list_add_tail(&rt_se->run_list, queue);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001092 __set_bit(rt_se_prio(rt_se), array->bitmap);
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001093
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001094 inc_rt_tasks(rt_se, rt_rq);
1095}
1096
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001097static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001098{
1099 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1100 struct rt_prio_array *array = &rt_rq->active;
1101
1102 list_del_init(&rt_se->run_list);
1103 if (list_empty(array->queue + rt_se_prio(rt_se)))
1104 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1105
1106 dec_rt_tasks(rt_se, rt_rq);
1107}
1108
1109/*
1110 * Because the prio of an upper entry depends on the lower
1111 * entries, we must remove entries top - down.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001112 */
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001113static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001114{
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001115 struct sched_rt_entity *back = NULL;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001116
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +02001117 for_each_sched_rt_entity(rt_se) {
1118 rt_se->back = back;
1119 back = rt_se;
1120 }
1121
1122 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1123 if (on_rt_rq(rt_se))
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001124 __dequeue_rt_entity(rt_se);
1125 }
1126}
1127
Thomas Gleixner37dad3f2010-01-20 20:59:01 +00001128static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001129{
1130 dequeue_rt_stack(rt_se);
1131 for_each_sched_rt_entity(rt_se)
Thomas Gleixner37dad3f2010-01-20 20:59:01 +00001132 __enqueue_rt_entity(rt_se, head);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001133}
1134
1135static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1136{
1137 dequeue_rt_stack(rt_se);
1138
1139 for_each_sched_rt_entity(rt_se) {
1140 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1141
1142 if (rt_rq && rt_rq->rt_nr_running)
Thomas Gleixner37dad3f2010-01-20 20:59:01 +00001143 __enqueue_rt_entity(rt_se, false);
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +02001144 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001145}
1146
1147/*
1148 * Adding/removing a task to/from a priority array:
1149 */
Thomas Gleixnerea87bb72010-01-20 20:58:57 +00001150static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001151enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001152{
1153 struct sched_rt_entity *rt_se = &p->rt;
1154
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001155 if (flags & ENQUEUE_WAKEUP)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001156 rt_se->timeout = 0;
1157
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001158 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001159
Peter Zijlstra29baa742012-04-23 12:11:21 +02001160 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
Gregory Haskins917b6272008-12-29 09:39:53 -05001161 enqueue_pushable_task(rq, p);
Paul Turner953bfcd12011-07-21 09:43:27 -07001162
1163 inc_nr_running(rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001164}
1165
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001166static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001167{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001168 struct sched_rt_entity *rt_se = &p->rt;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001169
1170 update_curr_rt(rq);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001171 dequeue_rt_entity(rt_se);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001172
Gregory Haskins917b6272008-12-29 09:39:53 -05001173 dequeue_pushable_task(rq, p);
Paul Turner953bfcd12011-07-21 09:43:27 -07001174
1175 dec_nr_running(rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001176}
1177
1178/*
Richard Weinberger60686312011-11-12 18:07:57 +01001179 * Put task to the head or the end of the run list without the overhead of
1180 * dequeue followed by enqueue.
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001181 */
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001182static void
1183requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001184{
Ingo Molnar1cdad712008-06-19 09:09:15 +02001185 if (on_rt_rq(rt_se)) {
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001186 struct rt_prio_array *array = &rt_rq->active;
1187 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1188
1189 if (head)
1190 list_move(&rt_se->run_list, queue);
1191 else
1192 list_move_tail(&rt_se->run_list, queue);
Ingo Molnar1cdad712008-06-19 09:09:15 +02001193 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001194}
1195
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001196static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001197{
1198 struct sched_rt_entity *rt_se = &p->rt;
1199 struct rt_rq *rt_rq;
1200
1201 for_each_sched_rt_entity(rt_se) {
1202 rt_rq = rt_rq_of_se(rt_se);
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001203 requeue_rt_entity(rt_rq, rt_se, head);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001204 }
1205}
1206
1207static void yield_task_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001208{
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001209 requeue_task_rt(rq, rq->curr, 0);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001210}
1211
Gregory Haskinse7693a32008-01-25 21:08:09 +01001212#ifdef CONFIG_SMP
Gregory Haskins318e0892008-01-25 21:08:10 +01001213static int find_lowest_rq(struct task_struct *task);
1214
Peter Zijlstra0017d732010-03-24 18:34:10 +01001215static int
Peter Zijlstraac66f542013-10-07 11:29:16 +01001216select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
Gregory Haskinse7693a32008-01-25 21:08:09 +01001217{
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001218 struct task_struct *curr;
1219 struct rq *rq;
Steven Rostedtc37495f2011-06-16 21:55:22 -04001220
Peter Zijlstra29baa742012-04-23 12:11:21 +02001221 if (p->nr_cpus_allowed == 1)
Mike Galbraith76854c72011-11-22 15:18:24 +01001222 goto out;
1223
Steven Rostedtc37495f2011-06-16 21:55:22 -04001224 /* For anything but wake ups, just return the task_cpu */
1225 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1226 goto out;
1227
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001228 rq = cpu_rq(cpu);
1229
1230 rcu_read_lock();
1231 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1232
Gregory Haskins318e0892008-01-25 21:08:10 +01001233 /*
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001234 * If the current task on @p's runqueue is an RT task, then
Steven Rostedte1f47d82008-01-25 21:08:12 +01001235 * try to see if we can wake this RT task up on another
1236 * runqueue. Otherwise simply start this RT task
1237 * on its current runqueue.
1238 *
Steven Rostedt43fa5462010-09-20 22:40:03 -04001239 * We want to avoid overloading runqueues. If the woken
1240 * task is a higher priority, then it will stay on this CPU
1241 * and the lower prio task should be moved to another CPU.
1242 * Even though this will probably make the lower prio task
1243 * lose its cache, we do not want to bounce a higher task
1244 * around just because it gave up its CPU, perhaps for a
1245 * lock?
1246 *
1247 * For equal prio tasks, we just let the scheduler sort it out.
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001248 *
Gregory Haskins318e0892008-01-25 21:08:10 +01001249 * Otherwise, just let it ride on the affined RQ and the
1250 * post-schedule router will push the preempted task away
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001251 *
1252 * This test is optimistic, if we get it wrong the load-balancer
1253 * will have to sort it out.
Gregory Haskins318e0892008-01-25 21:08:10 +01001254 */
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001255 if (curr && unlikely(rt_task(curr)) &&
Peter Zijlstra29baa742012-04-23 12:11:21 +02001256 (curr->nr_cpus_allowed < 2 ||
Shawn Bohrer6bfa6872013-10-04 14:24:53 -05001257 curr->prio <= p->prio)) {
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001258 int target = find_lowest_rq(p);
1259
1260 if (target != -1)
1261 cpu = target;
1262 }
1263 rcu_read_unlock();
1264
Steven Rostedtc37495f2011-06-16 21:55:22 -04001265out:
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001266 return cpu;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001267}
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001268
1269static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1270{
Peter Zijlstra29baa742012-04-23 12:11:21 +02001271 if (rq->curr->nr_cpus_allowed == 1)
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001272 return;
1273
Peter Zijlstra29baa742012-04-23 12:11:21 +02001274 if (p->nr_cpus_allowed != 1
Rusty Russell13b8bd02009-03-25 15:01:22 +10301275 && cpupri_find(&rq->rd->cpupri, p, NULL))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001276 return;
1277
Rusty Russell13b8bd02009-03-25 15:01:22 +10301278 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1279 return;
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001280
1281 /*
1282 * There appears to be other cpus that can accept
1283 * current and none to run 'p', so lets reschedule
1284 * to try and push current away:
1285 */
1286 requeue_task_rt(rq, p, 1);
1287 resched_task(rq->curr);
1288}
1289
Gregory Haskinse7693a32008-01-25 21:08:09 +01001290#endif /* CONFIG_SMP */
1291
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001292/*
1293 * Preempt the current task with a newly woken task if needed:
1294 */
Peter Zijlstra7d478722009-09-14 19:55:44 +02001295static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001296{
Gregory Haskins45c01e82008-05-12 21:20:41 +02001297 if (p->prio < rq->curr->prio) {
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001298 resched_task(rq->curr);
Gregory Haskins45c01e82008-05-12 21:20:41 +02001299 return;
1300 }
1301
1302#ifdef CONFIG_SMP
1303 /*
1304 * If:
1305 *
1306 * - the newly woken task is of equal priority to the current task
1307 * - the newly woken task is non-migratable while current is migratable
1308 * - current will be preempted on the next reschedule
1309 *
1310 * we should check to see if current can readily move to a different
1311 * cpu. If so, we will reschedule to allow the push logic to try
1312 * to move current somewhere else, making room for our non-migratable
1313 * task.
1314 */
Hillf Danton8dd0de8b2011-06-14 18:36:24 -04001315 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001316 check_preempt_equal_prio(rq, p);
Gregory Haskins45c01e82008-05-12 21:20:41 +02001317#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001318}
1319
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001320static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1321 struct rt_rq *rt_rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001322{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001323 struct rt_prio_array *array = &rt_rq->active;
1324 struct sched_rt_entity *next = NULL;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001325 struct list_head *queue;
1326 int idx;
1327
1328 idx = sched_find_first_bit(array->bitmap);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001329 BUG_ON(idx >= MAX_RT_PRIO);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001330
1331 queue = array->queue + idx;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001332 next = list_entry(queue->next, struct sched_rt_entity, run_list);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001333
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001334 return next;
1335}
1336
Gregory Haskins917b6272008-12-29 09:39:53 -05001337static struct task_struct *_pick_next_task_rt(struct rq *rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001338{
1339 struct sched_rt_entity *rt_se;
1340 struct task_struct *p;
Peter Zijlstra606dba22012-02-11 06:05:00 +01001341 struct rt_rq *rt_rq = &rq->rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001342
1343 do {
1344 rt_se = pick_next_rt_entity(rq, rt_rq);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001345 BUG_ON(!rt_se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001346 rt_rq = group_rt_rq(rt_se);
1347 } while (rt_rq);
1348
1349 p = rt_task_of(rt_se);
Frederic Weisbecker78becc22013-04-12 01:51:02 +02001350 p->se.exec_start = rq_clock_task(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001351
1352 return p;
1353}
1354
Peter Zijlstra606dba22012-02-11 06:05:00 +01001355static struct task_struct *
1356pick_next_task_rt(struct rq *rq, struct task_struct *prev)
Gregory Haskins917b6272008-12-29 09:39:53 -05001357{
Peter Zijlstra606dba22012-02-11 06:05:00 +01001358 struct task_struct *p;
1359 struct rt_rq *rt_rq = &rq->rt;
1360
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001361 if (need_pull_rt_task(rq, prev)) {
Peter Zijlstra38033c32014-01-23 20:32:21 +01001362 pull_rt_task(rq);
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001363 /*
1364 * pull_rt_task() can drop (and re-acquire) rq->lock; this
Kirill Tkhaia1d9a322014-04-10 17:38:36 +04001365 * means a dl or stop task can slip in, in which case we need
1366 * to re-start task selection.
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001367 */
Kirill Tkhaia1d9a322014-04-10 17:38:36 +04001368 if (unlikely((rq->stop && rq->stop->on_rq) ||
1369 rq->dl.dl_nr_running))
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001370 return RETRY_TASK;
1371 }
Peter Zijlstra38033c32014-01-23 20:32:21 +01001372
Kirill Tkhai734ff2a2014-03-04 19:25:46 +04001373 /*
1374 * We may dequeue prev's rt_rq in put_prev_task().
1375 * So, we update time before rt_nr_running check.
1376 */
1377 if (prev->sched_class == &rt_sched_class)
1378 update_curr_rt(rq);
1379
Peter Zijlstra606dba22012-02-11 06:05:00 +01001380 if (!rt_rq->rt_nr_running)
1381 return NULL;
1382
1383 if (rt_rq_throttled(rt_rq))
1384 return NULL;
1385
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01001386 put_prev_task(rq, prev);
Peter Zijlstra606dba22012-02-11 06:05:00 +01001387
1388 p = _pick_next_task_rt(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001389
1390 /* The running task is never eligible for pushing */
1391 if (p)
1392 dequeue_pushable_task(rq, p);
1393
Peter Zijlstradc877342014-02-12 15:47:29 +01001394 set_post_schedule(rq);
Gregory Haskins3f029d32009-07-29 11:08:47 -04001395
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001396 return p;
1397}
1398
Ingo Molnar31ee5292007-08-09 11:16:49 +02001399static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001400{
Ingo Molnarf1e14ef2007-08-09 11:16:48 +02001401 update_curr_rt(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001402
1403 /*
1404 * The previous task needs to be made eligible for pushing
1405 * if it is still active
1406 */
Peter Zijlstra29baa742012-04-23 12:11:21 +02001407 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
Gregory Haskins917b6272008-12-29 09:39:53 -05001408 enqueue_pushable_task(rq, p);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001409}
1410
Peter Williams681f3e62007-10-24 18:23:51 +02001411#ifdef CONFIG_SMP
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001412
Steven Rostedte8fa1362008-01-25 21:08:05 +01001413/* Only try algorithms three times */
1414#define RT_MAX_TRIES 3
1415
Steven Rostedtf65eda42008-01-25 21:08:07 +01001416static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1417{
1418 if (!task_running(rq, p) &&
Kirill Tkhai60334ca2013-01-31 18:56:17 +04001419 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
Steven Rostedtf65eda42008-01-25 21:08:07 +01001420 return 1;
1421 return 0;
1422}
1423
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001424/*
1425 * Return the highest pushable rq's task, which is suitable to be executed
1426 * on the cpu, NULL otherwise
1427 */
1428static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001429{
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001430 struct plist_head *head = &rq->rt.pushable_tasks;
1431 struct task_struct *p;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001432
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001433 if (!has_pushable_tasks(rq))
1434 return NULL;
Peter Zijlstra3d074672010-03-10 17:07:24 +01001435
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001436 plist_for_each_entry(p, head, pushable_tasks) {
1437 if (pick_rt_task(rq, p, cpu))
1438 return p;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001439 }
1440
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001441 return NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001442}
1443
Rusty Russell0e3900e2008-11-25 02:35:13 +10301444static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001445
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001446static int find_lowest_rq(struct task_struct *task)
1447{
1448 struct sched_domain *sd;
Rusty Russell96f874e22008-11-25 02:35:14 +10301449 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001450 int this_cpu = smp_processor_id();
1451 int cpu = task_cpu(task);
1452
Steven Rostedt0da938c2011-06-14 18:36:25 -04001453 /* Make sure the mask is initialized first */
1454 if (unlikely(!lowest_mask))
1455 return -1;
1456
Peter Zijlstra29baa742012-04-23 12:11:21 +02001457 if (task->nr_cpus_allowed == 1)
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001458 return -1; /* No other targets possible */
1459
1460 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
Gregory Haskins06f90db2008-01-25 21:08:13 +01001461 return -1; /* No targets found */
1462
1463 /*
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001464 * At this point we have built a mask of cpus representing the
1465 * lowest priority tasks in the system. Now we want to elect
1466 * the best one based on our affinity and topology.
1467 *
1468 * We prioritize the last cpu that the task executed on since
1469 * it is most likely cache-hot in that location.
1470 */
Rusty Russell96f874e22008-11-25 02:35:14 +10301471 if (cpumask_test_cpu(cpu, lowest_mask))
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001472 return cpu;
1473
1474 /*
1475 * Otherwise, we consult the sched_domains span maps to figure
1476 * out which cpu is logically closest to our hot cache data.
1477 */
Rusty Russelle2c88062009-11-03 14:53:15 +10301478 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1479 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001480
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001481 rcu_read_lock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301482 for_each_domain(cpu, sd) {
1483 if (sd->flags & SD_WAKE_AFFINE) {
1484 int best_cpu;
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001485
Rusty Russelle2c88062009-11-03 14:53:15 +10301486 /*
1487 * "this_cpu" is cheaper to preempt than a
1488 * remote processor.
1489 */
1490 if (this_cpu != -1 &&
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001491 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1492 rcu_read_unlock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301493 return this_cpu;
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001494 }
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001495
Rusty Russelle2c88062009-11-03 14:53:15 +10301496 best_cpu = cpumask_first_and(lowest_mask,
1497 sched_domain_span(sd));
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001498 if (best_cpu < nr_cpu_ids) {
1499 rcu_read_unlock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301500 return best_cpu;
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001501 }
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001502 }
1503 }
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001504 rcu_read_unlock();
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001505
1506 /*
1507 * And finally, if there were no matches within the domains
1508 * just give the caller *something* to work with from the compatible
1509 * locations.
1510 */
Rusty Russelle2c88062009-11-03 14:53:15 +10301511 if (this_cpu != -1)
1512 return this_cpu;
1513
1514 cpu = cpumask_any(lowest_mask);
1515 if (cpu < nr_cpu_ids)
1516 return cpu;
1517 return -1;
Gregory Haskins07b40322008-01-25 21:08:10 +01001518}
1519
Steven Rostedte8fa1362008-01-25 21:08:05 +01001520/* Will lock the rq it finds */
Ingo Molnar4df64c02008-01-25 21:08:15 +01001521static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001522{
1523 struct rq *lowest_rq = NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001524 int tries;
Ingo Molnar4df64c02008-01-25 21:08:15 +01001525 int cpu;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001526
1527 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
Gregory Haskins07b40322008-01-25 21:08:10 +01001528 cpu = find_lowest_rq(task);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001529
Gregory Haskins2de0b462008-01-25 21:08:10 +01001530 if ((cpu == -1) || (cpu == rq->cpu))
Steven Rostedte8fa1362008-01-25 21:08:05 +01001531 break;
1532
Gregory Haskins07b40322008-01-25 21:08:10 +01001533 lowest_rq = cpu_rq(cpu);
1534
Steven Rostedte8fa1362008-01-25 21:08:05 +01001535 /* if the prio of this runqueue changed, try again */
Gregory Haskins07b40322008-01-25 21:08:10 +01001536 if (double_lock_balance(rq, lowest_rq)) {
Steven Rostedte8fa1362008-01-25 21:08:05 +01001537 /*
1538 * We had to unlock the run queue. In
1539 * the mean time, task could have
1540 * migrated already or had its affinity changed.
1541 * Also make sure that it wasn't scheduled on its rq.
1542 */
Gregory Haskins07b40322008-01-25 21:08:10 +01001543 if (unlikely(task_rq(task) != rq ||
Rusty Russell96f874e22008-11-25 02:35:14 +10301544 !cpumask_test_cpu(lowest_rq->cpu,
Peter Zijlstrafa17b502011-06-16 12:23:22 +02001545 tsk_cpus_allowed(task)) ||
Gregory Haskins07b40322008-01-25 21:08:10 +01001546 task_running(rq, task) ||
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02001547 !task->on_rq)) {
Ingo Molnar4df64c02008-01-25 21:08:15 +01001548
Peter Zijlstra7f1b4392012-05-17 21:19:46 +02001549 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001550 lowest_rq = NULL;
1551 break;
1552 }
1553 }
1554
1555 /* If this rq is still suitable use it. */
Gregory Haskinse864c492008-12-29 09:39:49 -05001556 if (lowest_rq->rt.highest_prio.curr > task->prio)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001557 break;
1558
1559 /* try again */
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001560 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001561 lowest_rq = NULL;
1562 }
1563
1564 return lowest_rq;
1565}
1566
Gregory Haskins917b6272008-12-29 09:39:53 -05001567static struct task_struct *pick_next_pushable_task(struct rq *rq)
1568{
1569 struct task_struct *p;
1570
1571 if (!has_pushable_tasks(rq))
1572 return NULL;
1573
1574 p = plist_first_entry(&rq->rt.pushable_tasks,
1575 struct task_struct, pushable_tasks);
1576
1577 BUG_ON(rq->cpu != task_cpu(p));
1578 BUG_ON(task_current(rq, p));
Peter Zijlstra29baa742012-04-23 12:11:21 +02001579 BUG_ON(p->nr_cpus_allowed <= 1);
Gregory Haskins917b6272008-12-29 09:39:53 -05001580
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02001581 BUG_ON(!p->on_rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001582 BUG_ON(!rt_task(p));
1583
1584 return p;
1585}
1586
Steven Rostedte8fa1362008-01-25 21:08:05 +01001587/*
1588 * If the current CPU has more than one RT task, see if the non
1589 * running task can migrate over to a CPU that is running a task
1590 * of lesser priority.
1591 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001592static int push_rt_task(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001593{
1594 struct task_struct *next_task;
1595 struct rq *lowest_rq;
Hillf Danton311e8002011-06-16 21:55:20 -04001596 int ret = 0;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001597
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +01001598 if (!rq->rt.overloaded)
1599 return 0;
1600
Gregory Haskins917b6272008-12-29 09:39:53 -05001601 next_task = pick_next_pushable_task(rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001602 if (!next_task)
1603 return 0;
1604
Peter Zijlstra49246272010-10-17 21:46:10 +02001605retry:
Gregory Haskins697f0a42008-01-25 21:08:09 +01001606 if (unlikely(next_task == rq->curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001607 WARN_ON(1);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001608 return 0;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001609 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01001610
1611 /*
1612 * It's possible that the next_task slipped in of
1613 * higher priority than current. If that's the case
1614 * just reschedule current.
1615 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001616 if (unlikely(next_task->prio < rq->curr->prio)) {
1617 resched_task(rq->curr);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001618 return 0;
1619 }
1620
Gregory Haskins697f0a42008-01-25 21:08:09 +01001621 /* We might release rq lock */
Steven Rostedte8fa1362008-01-25 21:08:05 +01001622 get_task_struct(next_task);
1623
1624 /* find_lock_lowest_rq locks the rq if found */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001625 lowest_rq = find_lock_lowest_rq(next_task, rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001626 if (!lowest_rq) {
1627 struct task_struct *task;
1628 /*
Hillf Danton311e8002011-06-16 21:55:20 -04001629 * find_lock_lowest_rq releases rq->lock
Gregory Haskins15635132008-12-29 09:39:53 -05001630 * so it is possible that next_task has migrated.
1631 *
1632 * We need to make sure that the task is still on the same
1633 * run-queue and is also still the next task eligible for
1634 * pushing.
Steven Rostedte8fa1362008-01-25 21:08:05 +01001635 */
Gregory Haskins917b6272008-12-29 09:39:53 -05001636 task = pick_next_pushable_task(rq);
Gregory Haskins15635132008-12-29 09:39:53 -05001637 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1638 /*
Hillf Danton311e8002011-06-16 21:55:20 -04001639 * The task hasn't migrated, and is still the next
1640 * eligible task, but we failed to find a run-queue
1641 * to push it to. Do not retry in this case, since
1642 * other cpus will pull from us when ready.
Gregory Haskins15635132008-12-29 09:39:53 -05001643 */
Gregory Haskins15635132008-12-29 09:39:53 -05001644 goto out;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001645 }
Gregory Haskins917b6272008-12-29 09:39:53 -05001646
Gregory Haskins15635132008-12-29 09:39:53 -05001647 if (!task)
1648 /* No more tasks, just exit */
1649 goto out;
1650
Gregory Haskins917b6272008-12-29 09:39:53 -05001651 /*
Gregory Haskins15635132008-12-29 09:39:53 -05001652 * Something has shifted, try again.
Gregory Haskins917b6272008-12-29 09:39:53 -05001653 */
Gregory Haskins15635132008-12-29 09:39:53 -05001654 put_task_struct(next_task);
1655 next_task = task;
1656 goto retry;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001657 }
1658
Gregory Haskins697f0a42008-01-25 21:08:09 +01001659 deactivate_task(rq, next_task, 0);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001660 set_task_cpu(next_task, lowest_rq->cpu);
1661 activate_task(lowest_rq, next_task, 0);
Hillf Danton311e8002011-06-16 21:55:20 -04001662 ret = 1;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001663
1664 resched_task(lowest_rq->curr);
1665
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001666 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001667
Steven Rostedte8fa1362008-01-25 21:08:05 +01001668out:
1669 put_task_struct(next_task);
1670
Hillf Danton311e8002011-06-16 21:55:20 -04001671 return ret;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001672}
1673
Steven Rostedte8fa1362008-01-25 21:08:05 +01001674static void push_rt_tasks(struct rq *rq)
1675{
1676 /* push_rt_task will return true if it moved an RT */
1677 while (push_rt_task(rq))
1678 ;
1679}
1680
Steven Rostedtf65eda42008-01-25 21:08:07 +01001681static int pull_rt_task(struct rq *this_rq)
1682{
Ingo Molnar80bf3172008-01-25 21:08:17 +01001683 int this_cpu = this_rq->cpu, ret = 0, cpu;
Gregory Haskinsa8728942008-12-29 09:39:49 -05001684 struct task_struct *p;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001685 struct rq *src_rq;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001686
Gregory Haskins637f5082008-01-25 21:08:18 +01001687 if (likely(!rt_overloaded(this_rq)))
Steven Rostedtf65eda42008-01-25 21:08:07 +01001688 return 0;
1689
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +02001690 /*
1691 * Match the barrier from rt_set_overloaded; this guarantees that if we
1692 * see overloaded we must also see the rto_mask bit.
1693 */
1694 smp_rmb();
1695
Rusty Russellc6c49272008-11-25 02:35:05 +10301696 for_each_cpu(cpu, this_rq->rd->rto_mask) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001697 if (this_cpu == cpu)
1698 continue;
1699
1700 src_rq = cpu_rq(cpu);
Gregory Haskins74ab8e42008-12-29 09:39:50 -05001701
1702 /*
1703 * Don't bother taking the src_rq->lock if the next highest
1704 * task is known to be lower-priority than our current task.
1705 * This may look racy, but if this value is about to go
1706 * logically higher, the src_rq will push this task away.
1707 * And if its going logically lower, we do not care
1708 */
1709 if (src_rq->rt.highest_prio.next >=
1710 this_rq->rt.highest_prio.curr)
1711 continue;
1712
Steven Rostedtf65eda42008-01-25 21:08:07 +01001713 /*
1714 * We can potentially drop this_rq's lock in
1715 * double_lock_balance, and another CPU could
Gregory Haskinsa8728942008-12-29 09:39:49 -05001716 * alter this_rq
Steven Rostedtf65eda42008-01-25 21:08:07 +01001717 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001718 double_lock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001719
1720 /*
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001721 * We can pull only a task, which is pushable
1722 * on its rq, and no others.
Steven Rostedtf65eda42008-01-25 21:08:07 +01001723 */
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001724 p = pick_highest_pushable_task(src_rq, this_cpu);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001725
1726 /*
1727 * Do we have an RT task that preempts
1728 * the to-be-scheduled task?
1729 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001730 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001731 WARN_ON(p == src_rq->curr);
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02001732 WARN_ON(!p->on_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001733
1734 /*
1735 * There's a chance that p is higher in priority
1736 * than what's currently running on its cpu.
1737 * This is just that p is wakeing up and hasn't
1738 * had a chance to schedule. We only pull
1739 * p if it is lower in priority than the
Gregory Haskinsa8728942008-12-29 09:39:49 -05001740 * current task on the run queue
Steven Rostedtf65eda42008-01-25 21:08:07 +01001741 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001742 if (p->prio < src_rq->curr->prio)
Mike Galbraith614ee1f2008-01-25 21:08:30 +01001743 goto skip;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001744
1745 ret = 1;
1746
1747 deactivate_task(src_rq, p, 0);
1748 set_task_cpu(p, this_cpu);
1749 activate_task(this_rq, p, 0);
1750 /*
1751 * We continue with the search, just in
1752 * case there's an even higher prio task
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001753 * in another runqueue. (low likelihood
Steven Rostedtf65eda42008-01-25 21:08:07 +01001754 * but possible)
Steven Rostedtf65eda42008-01-25 21:08:07 +01001755 */
Steven Rostedtf65eda42008-01-25 21:08:07 +01001756 }
Peter Zijlstra49246272010-10-17 21:46:10 +02001757skip:
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001758 double_unlock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001759 }
1760
1761 return ret;
1762}
1763
Steven Rostedt9a897c52008-01-25 21:08:22 +01001764static void post_schedule_rt(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001765{
Gregory Haskins967fc042008-12-29 09:39:52 -05001766 push_rt_tasks(rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001767}
1768
Gregory Haskins8ae121a2008-04-23 07:13:29 -04001769/*
1770 * If we are not running and we are not going to reschedule soon, we should
1771 * try to push tasks away now
1772 */
Peter Zijlstraefbbd052009-12-16 18:04:40 +01001773static void task_woken_rt(struct rq *rq, struct task_struct *p)
Steven Rostedt4642daf2008-01-25 21:08:07 +01001774{
Steven Rostedt9a897c52008-01-25 21:08:22 +01001775 if (!task_running(rq, p) &&
Gregory Haskins8ae121a2008-04-23 07:13:29 -04001776 !test_tsk_need_resched(rq->curr) &&
Gregory Haskins917b6272008-12-29 09:39:53 -05001777 has_pushable_tasks(rq) &&
Peter Zijlstra29baa742012-04-23 12:11:21 +02001778 p->nr_cpus_allowed > 1 &&
Juri Lelli1baca4c2013-11-07 14:43:38 +01001779 (dl_task(rq->curr) || rt_task(rq->curr)) &&
Peter Zijlstra29baa742012-04-23 12:11:21 +02001780 (rq->curr->nr_cpus_allowed < 2 ||
Shawn Bohrer3be209a2011-09-12 09:28:04 -05001781 rq->curr->prio <= p->prio))
Steven Rostedt4642daf2008-01-25 21:08:07 +01001782 push_rt_tasks(rq);
1783}
1784
Mike Traviscd8ba7c2008-03-26 14:23:49 -07001785static void set_cpus_allowed_rt(struct task_struct *p,
Rusty Russell96f874e22008-11-25 02:35:14 +10301786 const struct cpumask *new_mask)
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001787{
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001788 struct rq *rq;
1789 int weight;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001790
1791 BUG_ON(!rt_task(p));
1792
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001793 if (!p->on_rq)
1794 return;
1795
1796 weight = cpumask_weight(new_mask);
1797
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001798 /*
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001799 * Only update if the process changes its state from whether it
1800 * can migrate or not.
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001801 */
Peter Zijlstra29baa742012-04-23 12:11:21 +02001802 if ((p->nr_cpus_allowed > 1) == (weight > 1))
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001803 return;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001804
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001805 rq = task_rq(p);
Gregory Haskins917b6272008-12-29 09:39:53 -05001806
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001807 /*
1808 * The process used to be able to migrate OR it can now migrate
1809 */
1810 if (weight <= 1) {
1811 if (!task_current(rq, p))
1812 dequeue_pushable_task(rq, p);
1813 BUG_ON(!rq->rt.rt_nr_migratory);
1814 rq->rt.rt_nr_migratory--;
1815 } else {
1816 if (!task_current(rq, p))
1817 enqueue_pushable_task(rq, p);
1818 rq->rt.rt_nr_migratory++;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001819 }
Kirill Tkhai8d3d5ad2012-04-11 09:06:04 +04001820
1821 update_rt_migration(&rq->rt);
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001822}
Ingo Molnardeeeccd2008-01-25 21:08:15 +01001823
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001824/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001825static void rq_online_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001826{
1827 if (rq->rt.overloaded)
1828 rt_set_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001829
Peter Zijlstra7def2be2008-06-05 14:49:58 +02001830 __enable_runtime(rq);
1831
Gregory Haskinse864c492008-12-29 09:39:49 -05001832 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001833}
1834
1835/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001836static void rq_offline_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001837{
1838 if (rq->rt.overloaded)
1839 rt_clear_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001840
Peter Zijlstra7def2be2008-06-05 14:49:58 +02001841 __disable_runtime(rq);
1842
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001843 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001844}
Steven Rostedtcb469842008-01-25 21:08:22 +01001845
1846/*
1847 * When switch from the rt queue, we bring ourselves to a position
1848 * that we might want to pull RT tasks from other runqueues.
1849 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001850static void switched_from_rt(struct rq *rq, struct task_struct *p)
Steven Rostedtcb469842008-01-25 21:08:22 +01001851{
1852 /*
1853 * If there are other RT tasks then we will reschedule
1854 * and the scheduling of the other RT tasks will handle
1855 * the balancing. But if we are the last RT task
1856 * we may need to handle the pulling of RT tasks
1857 * now.
1858 */
Kirill Tkhai1158ddb2012-11-23 00:02:15 +04001859 if (!p->on_rq || rq->rt.rt_nr_running)
1860 return;
1861
1862 if (pull_rt_task(rq))
1863 resched_task(rq->curr);
Steven Rostedtcb469842008-01-25 21:08:22 +01001864}
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10301865
Li Zefan11c785b2014-02-08 14:17:45 +08001866void __init init_sched_rt_class(void)
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10301867{
1868 unsigned int i;
1869
Peter Zijlstra029632f2011-10-25 10:00:11 +02001870 for_each_possible_cpu(i) {
Yinghai Lueaa95842009-06-06 14:51:36 -07001871 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
Mike Travis6ca09df2008-12-31 18:08:45 -08001872 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra029632f2011-10-25 10:00:11 +02001873 }
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10301874}
Steven Rostedte8fa1362008-01-25 21:08:05 +01001875#endif /* CONFIG_SMP */
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001876
Steven Rostedtcb469842008-01-25 21:08:22 +01001877/*
1878 * When switching a task to RT, we may overload the runqueue
1879 * with RT tasks. In this case we try to push them off to
1880 * other runqueues.
1881 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001882static void switched_to_rt(struct rq *rq, struct task_struct *p)
Steven Rostedtcb469842008-01-25 21:08:22 +01001883{
1884 int check_resched = 1;
1885
1886 /*
1887 * If we are already running, then there's nothing
1888 * that needs to be done. But if we are not running
1889 * we may need to preempt the current running task.
1890 * If that current running task is also an RT task
1891 * then see if we can move to another run queue.
1892 */
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02001893 if (p->on_rq && rq->curr != p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01001894#ifdef CONFIG_SMP
1895 if (rq->rt.overloaded && push_rt_task(rq) &&
1896 /* Don't resched if we changed runqueues */
1897 rq != task_rq(p))
1898 check_resched = 0;
1899#endif /* CONFIG_SMP */
1900 if (check_resched && p->prio < rq->curr->prio)
1901 resched_task(rq->curr);
1902 }
1903}
1904
1905/*
1906 * Priority of the task has changed. This may cause
1907 * us to initiate a push or pull.
1908 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001909static void
1910prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01001911{
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02001912 if (!p->on_rq)
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001913 return;
1914
1915 if (rq->curr == p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01001916#ifdef CONFIG_SMP
1917 /*
1918 * If our priority decreases while running, we
1919 * may need to pull tasks to this runqueue.
1920 */
1921 if (oldprio < p->prio)
1922 pull_rt_task(rq);
1923 /*
1924 * If there's a higher priority task waiting to run
Steven Rostedt6fa46fa2008-03-05 10:00:12 -05001925 * then reschedule. Note, the above pull_rt_task
1926 * can release the rq lock and p could migrate.
1927 * Only reschedule if p is still on the same runqueue.
Steven Rostedtcb469842008-01-25 21:08:22 +01001928 */
Gregory Haskinse864c492008-12-29 09:39:49 -05001929 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
Steven Rostedtcb469842008-01-25 21:08:22 +01001930 resched_task(p);
1931#else
1932 /* For UP simply resched on drop of prio */
1933 if (oldprio < p->prio)
1934 resched_task(p);
1935#endif /* CONFIG_SMP */
1936 } else {
1937 /*
1938 * This task is not running, but if it is
1939 * greater than the current running task
1940 * then reschedule.
1941 */
1942 if (p->prio < rq->curr->prio)
1943 resched_task(rq->curr);
1944 }
1945}
1946
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001947static void watchdog(struct rq *rq, struct task_struct *p)
1948{
1949 unsigned long soft, hard;
1950
Jiri Slaby78d7d402010-03-05 13:42:54 -08001951 /* max may change after cur was read, this will be fixed next tick */
1952 soft = task_rlimit(p, RLIMIT_RTTIME);
1953 hard = task_rlimit_max(p, RLIMIT_RTTIME);
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001954
1955 if (soft != RLIM_INFINITY) {
1956 unsigned long next;
1957
Ying Xue57d2aa02012-07-17 15:03:43 +08001958 if (p->rt.watchdog_stamp != jiffies) {
1959 p->rt.timeout++;
1960 p->rt.watchdog_stamp = jiffies;
1961 }
1962
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001963 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
Peter Zijlstra5a52dd52008-01-25 21:08:32 +01001964 if (p->rt.timeout > next)
Frank Mayharf06febc2008-09-12 09:54:39 -07001965 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001966 }
1967}
Steven Rostedtcb469842008-01-25 21:08:22 +01001968
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001969static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001970{
Colin Cross454c7992012-05-16 21:34:23 -07001971 struct sched_rt_entity *rt_se = &p->rt;
1972
Peter Zijlstra67e2be02007-12-20 15:01:17 +01001973 update_curr_rt(rq);
1974
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001975 watchdog(rq, p);
1976
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001977 /*
1978 * RR tasks need a special form of timeslice management.
1979 * FIFO tasks have no timeslices.
1980 */
1981 if (p->policy != SCHED_RR)
1982 return;
1983
Peter Zijlstrafa717062008-01-25 21:08:27 +01001984 if (--p->rt.time_slice)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001985 return;
1986
Clark Williamsce0dbbb2013-02-07 09:47:04 -06001987 p->rt.time_slice = sched_rr_timeslice;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001988
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02001989 /*
Li Bine9aa39b2013-10-21 20:15:43 +08001990 * Requeue to the end of queue if we (and all of our ancestors) are not
1991 * the only element on the queue
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02001992 */
Colin Cross454c7992012-05-16 21:34:23 -07001993 for_each_sched_rt_entity(rt_se) {
1994 if (rt_se->run_list.prev != rt_se->run_list.next) {
1995 requeue_task_rt(rq, p, 0);
1996 set_tsk_need_resched(p);
1997 return;
1998 }
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02001999 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002000}
2001
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002002static void set_curr_task_rt(struct rq *rq)
2003{
2004 struct task_struct *p = rq->curr;
2005
Frederic Weisbecker78becc22013-04-12 01:51:02 +02002006 p->se.exec_start = rq_clock_task(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05002007
2008 /* The running task is never eligible for pushing */
2009 dequeue_pushable_task(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002010}
2011
H Hartley Sweeten6d686f42010-01-13 20:21:52 -07002012static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
Peter Williams0d721ce2009-09-21 01:31:53 +00002013{
2014 /*
2015 * Time slice is 0 for SCHED_FIFO tasks
2016 */
2017 if (task->policy == SCHED_RR)
Clark Williamsce0dbbb2013-02-07 09:47:04 -06002018 return sched_rr_timeslice;
Peter Williams0d721ce2009-09-21 01:31:53 +00002019 else
2020 return 0;
2021}
2022
Peter Zijlstra029632f2011-10-25 10:00:11 +02002023const struct sched_class rt_sched_class = {
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002024 .next = &fair_sched_class,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002025 .enqueue_task = enqueue_task_rt,
2026 .dequeue_task = dequeue_task_rt,
2027 .yield_task = yield_task_rt,
2028
2029 .check_preempt_curr = check_preempt_curr_rt,
2030
2031 .pick_next_task = pick_next_task_rt,
2032 .put_prev_task = put_prev_task_rt,
2033
Peter Williams681f3e62007-10-24 18:23:51 +02002034#ifdef CONFIG_SMP
Li Zefan4ce72a22008-10-22 15:25:26 +08002035 .select_task_rq = select_task_rq_rt,
2036
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01002037 .set_cpus_allowed = set_cpus_allowed_rt,
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04002038 .rq_online = rq_online_rt,
2039 .rq_offline = rq_offline_rt,
Steven Rostedt9a897c52008-01-25 21:08:22 +01002040 .post_schedule = post_schedule_rt,
Peter Zijlstraefbbd052009-12-16 18:04:40 +01002041 .task_woken = task_woken_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01002042 .switched_from = switched_from_rt,
Peter Williams681f3e62007-10-24 18:23:51 +02002043#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002044
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002045 .set_curr_task = set_curr_task_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002046 .task_tick = task_tick_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01002047
Peter Williams0d721ce2009-09-21 01:31:53 +00002048 .get_rr_interval = get_rr_interval_rt,
2049
Steven Rostedtcb469842008-01-25 21:08:22 +01002050 .prio_changed = prio_changed_rt,
2051 .switched_to = switched_to_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002052};
Peter Zijlstraada18de22008-06-19 14:22:24 +02002053
2054#ifdef CONFIG_SCHED_DEBUG
2055extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2056
Peter Zijlstra029632f2011-10-25 10:00:11 +02002057void print_rt_stats(struct seq_file *m, int cpu)
Peter Zijlstraada18de22008-06-19 14:22:24 +02002058{
Cheng Xuec514c42011-05-14 14:20:02 +08002059 rt_rq_iter_t iter;
Peter Zijlstraada18de22008-06-19 14:22:24 +02002060 struct rt_rq *rt_rq;
2061
2062 rcu_read_lock();
Cheng Xuec514c42011-05-14 14:20:02 +08002063 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
Peter Zijlstraada18de22008-06-19 14:22:24 +02002064 print_rt_rq(m, cpu, rt_rq);
2065 rcu_read_unlock();
2066}
Dhaval Giani55e12e52008-06-24 23:39:43 +05302067#endif /* CONFIG_SCHED_DEBUG */