blob: 0af5ca9e3e3f1d3613e69fabba32d909d0a15f3b [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>
Steven Rostedtb6366f02015-03-18 14:49:46 -04009#include <linux/irq_work.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +020010
Clark Williamsce0dbbb2013-02-07 09:47:04 -060011int sched_rr_timeslice = RR_TIMESLICE;
Shile Zhang975e1552017-01-28 22:00:49 +080012int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
Clark Williamsce0dbbb2013-02-07 09:47:04 -060013
Peter Zijlstra029632f2011-10-25 10:00:11 +020014static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
15
16struct rt_bandwidth def_rt_bandwidth;
17
18static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
19{
20 struct rt_bandwidth *rt_b =
21 container_of(timer, struct rt_bandwidth, rt_period_timer);
Peter Zijlstra029632f2011-10-25 10:00:11 +020022 int idle = 0;
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020023 int overrun;
Peter Zijlstra029632f2011-10-25 10:00:11 +020024
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020025 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +020026 for (;;) {
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020027 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
Peter Zijlstra029632f2011-10-25 10:00:11 +020028 if (!overrun)
29 break;
30
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020031 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +020032 idle = do_sched_rt_period_timer(rt_b, overrun);
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020033 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +020034 }
Peter Zijlstra4cfafd32015-05-14 12:23:11 +020035 if (idle)
36 rt_b->rt_period_active = 0;
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020037 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +020038
39 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
40}
41
42void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
43{
44 rt_b->rt_period = ns_to_ktime(period);
45 rt_b->rt_runtime = runtime;
46
47 raw_spin_lock_init(&rt_b->rt_runtime_lock);
48
49 hrtimer_init(&rt_b->rt_period_timer,
50 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
51 rt_b->rt_period_timer.function = sched_rt_period_timer;
52}
53
54static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
55{
56 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
57 return;
58
Peter Zijlstra029632f2011-10-25 10:00:11 +020059 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstra4cfafd32015-05-14 12:23:11 +020060 if (!rt_b->rt_period_active) {
61 rt_b->rt_period_active = 1;
Steven Rostedtc3a990d2016-02-16 18:37:46 -050062 /*
63 * SCHED_DEADLINE updates the bandwidth, as a run away
64 * RT task with a DL task could hog a CPU. But DL does
65 * not reset the period. If a deadline task was running
66 * without an RT task running, it can cause RT tasks to
67 * throttle when they start up. Kick the timer right away
68 * to update the period.
69 */
70 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
Peter Zijlstra4cfafd32015-05-14 12:23:11 +020071 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
72 }
Peter Zijlstra029632f2011-10-25 10:00:11 +020073 raw_spin_unlock(&rt_b->rt_runtime_lock);
74}
75
Arnd Bergmann89b41102015-11-12 17:19:58 +010076#if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
Steven Rostedtb6366f02015-03-18 14:49:46 -040077static void push_irq_work_func(struct irq_work *work);
78#endif
79
Abel Vesa07c54f72015-03-03 13:50:27 +020080void init_rt_rq(struct rt_rq *rt_rq)
Peter Zijlstra029632f2011-10-25 10:00:11 +020081{
82 struct rt_prio_array *array;
83 int i;
84
85 array = &rt_rq->active;
86 for (i = 0; i < MAX_RT_PRIO; i++) {
87 INIT_LIST_HEAD(array->queue + i);
88 __clear_bit(i, array->bitmap);
89 }
90 /* delimiter for bitsearch: */
91 __set_bit(MAX_RT_PRIO, array->bitmap);
92
93#if defined CONFIG_SMP
94 rt_rq->highest_prio.curr = MAX_RT_PRIO;
95 rt_rq->highest_prio.next = MAX_RT_PRIO;
96 rt_rq->rt_nr_migratory = 0;
97 rt_rq->overloaded = 0;
98 plist_head_init(&rt_rq->pushable_tasks);
Steven Rostedtb6366f02015-03-18 14:49:46 -040099
100#ifdef HAVE_RT_PUSH_IPI
101 rt_rq->push_flags = 0;
102 rt_rq->push_cpu = nr_cpu_ids;
103 raw_spin_lock_init(&rt_rq->push_lock);
104 init_irq_work(&rt_rq->push_work, push_irq_work_func);
Peter Zijlstra029632f2011-10-25 10:00:11 +0200105#endif
Steven Rostedtb6366f02015-03-18 14:49:46 -0400106#endif /* CONFIG_SMP */
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400107 /* We start is dequeued state, because no RT tasks are queued */
108 rt_rq->rt_queued = 0;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200109
110 rt_rq->rt_time = 0;
111 rt_rq->rt_throttled = 0;
112 rt_rq->rt_runtime = 0;
113 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
114}
115
Gregory Haskins398a1532009-01-14 09:10:04 -0500116#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra029632f2011-10-25 10:00:11 +0200117static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
118{
119 hrtimer_cancel(&rt_b->rt_period_timer);
120}
Gregory Haskins398a1532009-01-14 09:10:04 -0500121
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200122#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
123
Peter Zijlstra8f488942009-07-24 12:25:30 +0200124static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
125{
126#ifdef CONFIG_SCHED_DEBUG
127 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
128#endif
129 return container_of(rt_se, struct task_struct, rt);
130}
131
Gregory Haskins398a1532009-01-14 09:10:04 -0500132static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
133{
134 return rt_rq->rq;
135}
136
137static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
138{
139 return rt_se->rt_rq;
140}
141
Kirill Tkhai653d07a2014-03-15 02:14:55 +0400142static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
143{
144 struct rt_rq *rt_rq = rt_se->rt_rq;
145
146 return rt_rq->rq;
147}
148
Peter Zijlstra029632f2011-10-25 10:00:11 +0200149void free_rt_sched_group(struct task_group *tg)
150{
151 int i;
152
153 if (tg->rt_se)
154 destroy_rt_bandwidth(&tg->rt_bandwidth);
155
156 for_each_possible_cpu(i) {
157 if (tg->rt_rq)
158 kfree(tg->rt_rq[i]);
159 if (tg->rt_se)
160 kfree(tg->rt_se[i]);
161 }
162
163 kfree(tg->rt_rq);
164 kfree(tg->rt_se);
165}
166
167void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
168 struct sched_rt_entity *rt_se, int cpu,
169 struct sched_rt_entity *parent)
170{
171 struct rq *rq = cpu_rq(cpu);
172
173 rt_rq->highest_prio.curr = MAX_RT_PRIO;
174 rt_rq->rt_nr_boosted = 0;
175 rt_rq->rq = rq;
176 rt_rq->tg = tg;
177
178 tg->rt_rq[cpu] = rt_rq;
179 tg->rt_se[cpu] = rt_se;
180
181 if (!rt_se)
182 return;
183
184 if (!parent)
185 rt_se->rt_rq = &rq->rt;
186 else
187 rt_se->rt_rq = parent->my_q;
188
189 rt_se->my_q = rt_rq;
190 rt_se->parent = parent;
191 INIT_LIST_HEAD(&rt_se->run_list);
192}
193
194int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
195{
196 struct rt_rq *rt_rq;
197 struct sched_rt_entity *rt_se;
198 int i;
199
200 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
201 if (!tg->rt_rq)
202 goto err;
203 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
204 if (!tg->rt_se)
205 goto err;
206
207 init_rt_bandwidth(&tg->rt_bandwidth,
208 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
209
210 for_each_possible_cpu(i) {
211 rt_rq = kzalloc_node(sizeof(struct rt_rq),
212 GFP_KERNEL, cpu_to_node(i));
213 if (!rt_rq)
214 goto err;
215
216 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
217 GFP_KERNEL, cpu_to_node(i));
218 if (!rt_se)
219 goto err_free_rq;
220
Abel Vesa07c54f72015-03-03 13:50:27 +0200221 init_rt_rq(rt_rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +0200222 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
223 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
224 }
225
226 return 1;
227
228err_free_rq:
229 kfree(rt_rq);
230err:
231 return 0;
232}
233
Gregory Haskins398a1532009-01-14 09:10:04 -0500234#else /* CONFIG_RT_GROUP_SCHED */
235
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200236#define rt_entity_is_task(rt_se) (1)
237
Peter Zijlstra8f488942009-07-24 12:25:30 +0200238static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
239{
240 return container_of(rt_se, struct task_struct, rt);
241}
242
Gregory Haskins398a1532009-01-14 09:10:04 -0500243static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
244{
245 return container_of(rt_rq, struct rq, rt);
246}
247
Kirill Tkhai653d07a2014-03-15 02:14:55 +0400248static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
Gregory Haskins398a1532009-01-14 09:10:04 -0500249{
250 struct task_struct *p = rt_task_of(rt_se);
Kirill Tkhai653d07a2014-03-15 02:14:55 +0400251
252 return task_rq(p);
253}
254
255static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
256{
257 struct rq *rq = rq_of_rt_se(rt_se);
Gregory Haskins398a1532009-01-14 09:10:04 -0500258
259 return &rq->rt;
260}
261
Peter Zijlstra029632f2011-10-25 10:00:11 +0200262void free_rt_sched_group(struct task_group *tg) { }
263
264int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
265{
266 return 1;
267}
Gregory Haskins398a1532009-01-14 09:10:04 -0500268#endif /* CONFIG_RT_GROUP_SCHED */
269
Steven Rostedt4fd29172008-01-25 21:08:06 +0100270#ifdef CONFIG_SMP
Ingo Molnar84de4272008-01-25 21:08:15 +0100271
Peter Zijlstra8046d682015-06-11 14:46:40 +0200272static void pull_rt_task(struct rq *this_rq);
Peter Zijlstra38033c32014-01-23 20:32:21 +0100273
Peter Zijlstradc877342014-02-12 15:47:29 +0100274static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
275{
276 /* Try to pull RT tasks here if we lower this rq's prio */
277 return rq->rt.highest_prio.curr > prev->prio;
278}
279
Gregory Haskins637f5082008-01-25 21:08:18 +0100280static inline int rt_overloaded(struct rq *rq)
Steven Rostedt4fd29172008-01-25 21:08:06 +0100281{
Gregory Haskins637f5082008-01-25 21:08:18 +0100282 return atomic_read(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100283}
Ingo Molnar84de4272008-01-25 21:08:15 +0100284
Steven Rostedt4fd29172008-01-25 21:08:06 +0100285static inline void rt_set_overload(struct rq *rq)
286{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400287 if (!rq->online)
288 return;
289
Rusty Russellc6c49272008-11-25 02:35:05 +1030290 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100291 /*
292 * Make sure the mask is visible before we set
293 * the overload count. That is checked to determine
294 * if we should look at the mask. It would be a shame
295 * if we looked at the mask, but the mask was not
296 * updated yet.
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +0200297 *
298 * Matched by the barrier in pull_rt_task().
Steven Rostedt4fd29172008-01-25 21:08:06 +0100299 */
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +0200300 smp_wmb();
Gregory Haskins637f5082008-01-25 21:08:18 +0100301 atomic_inc(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100302}
Ingo Molnar84de4272008-01-25 21:08:15 +0100303
Steven Rostedt4fd29172008-01-25 21:08:06 +0100304static inline void rt_clear_overload(struct rq *rq)
305{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400306 if (!rq->online)
307 return;
308
Steven Rostedt4fd29172008-01-25 21:08:06 +0100309 /* the order here really doesn't matter */
Gregory Haskins637f5082008-01-25 21:08:18 +0100310 atomic_dec(&rq->rd->rto_count);
Rusty Russellc6c49272008-11-25 02:35:05 +1030311 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +0100312}
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100313
Gregory Haskins398a1532009-01-14 09:10:04 -0500314static void update_rt_migration(struct rt_rq *rt_rq)
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100315{
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200316 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
Gregory Haskins398a1532009-01-14 09:10:04 -0500317 if (!rt_rq->overloaded) {
318 rt_set_overload(rq_of_rt_rq(rt_rq));
319 rt_rq->overloaded = 1;
Gregory Haskinscdc8eb92008-01-25 21:08:23 +0100320 }
Gregory Haskins398a1532009-01-14 09:10:04 -0500321 } else if (rt_rq->overloaded) {
322 rt_clear_overload(rq_of_rt_rq(rt_rq));
323 rt_rq->overloaded = 0;
Gregory Haskins637f5082008-01-25 21:08:18 +0100324 }
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100325}
Steven Rostedt4fd29172008-01-25 21:08:06 +0100326
Gregory Haskins398a1532009-01-14 09:10:04 -0500327static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100328{
Peter Zijlstra29baa742012-04-23 12:11:21 +0200329 struct task_struct *p;
330
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200331 if (!rt_entity_is_task(rt_se))
332 return;
333
Peter Zijlstra29baa742012-04-23 12:11:21 +0200334 p = rt_task_of(rt_se);
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200335 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
336
337 rt_rq->rt_nr_total++;
Ingo Molnar4b53a342017-02-05 15:41:03 +0100338 if (p->nr_cpus_allowed > 1)
Gregory Haskins398a1532009-01-14 09:10:04 -0500339 rt_rq->rt_nr_migratory++;
340
341 update_rt_migration(rt_rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100342}
343
Gregory Haskins398a1532009-01-14 09:10:04 -0500344static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
345{
Peter Zijlstra29baa742012-04-23 12:11:21 +0200346 struct task_struct *p;
347
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200348 if (!rt_entity_is_task(rt_se))
349 return;
350
Peter Zijlstra29baa742012-04-23 12:11:21 +0200351 p = rt_task_of(rt_se);
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200352 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
353
354 rt_rq->rt_nr_total--;
Ingo Molnar4b53a342017-02-05 15:41:03 +0100355 if (p->nr_cpus_allowed > 1)
Gregory Haskins398a1532009-01-14 09:10:04 -0500356 rt_rq->rt_nr_migratory--;
357
358 update_rt_migration(rt_rq);
359}
360
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400361static inline int has_pushable_tasks(struct rq *rq)
362{
363 return !plist_head_empty(&rq->rt.pushable_tasks);
364}
365
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +0200366static DEFINE_PER_CPU(struct callback_head, rt_push_head);
367static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +0200368
369static void push_rt_tasks(struct rq *);
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +0200370static void pull_rt_task(struct rq *);
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +0200371
372static inline void queue_push_tasks(struct rq *rq)
Peter Zijlstradc877342014-02-12 15:47:29 +0100373{
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +0200374 if (!has_pushable_tasks(rq))
375 return;
376
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +0200377 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
378}
379
380static inline void queue_pull_task(struct rq *rq)
381{
382 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
Peter Zijlstradc877342014-02-12 15:47:29 +0100383}
384
Gregory Haskins917b6272008-12-29 09:39:53 -0500385static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
386{
387 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
388 plist_node_init(&p->pushable_tasks, p->prio);
389 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400390
391 /* Update the highest prio pushable task */
392 if (p->prio < rq->rt.highest_prio.next)
393 rq->rt.highest_prio.next = p->prio;
Gregory Haskins917b6272008-12-29 09:39:53 -0500394}
395
396static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
397{
398 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
Gregory Haskins917b6272008-12-29 09:39:53 -0500399
Steven Rostedt5181f4a42011-06-16 21:55:23 -0400400 /* Update the new highest prio pushable task */
401 if (has_pushable_tasks(rq)) {
402 p = plist_first_entry(&rq->rt.pushable_tasks,
403 struct task_struct, pushable_tasks);
404 rq->rt.highest_prio.next = p->prio;
405 } else
406 rq->rt.highest_prio.next = MAX_RT_PRIO;
Ingo Molnarbcf08df2008-04-19 12:11:10 +0200407}
408
Gregory Haskins917b6272008-12-29 09:39:53 -0500409#else
410
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100411static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
412{
413}
414
415static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
416{
417}
418
Gregory Haskinsb07430a2009-01-14 08:55:39 -0500419static inline
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100420void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
421{
422}
423
Gregory Haskinsb07430a2009-01-14 08:55:39 -0500424static inline
Peter Zijlstraceacc2c2009-01-16 14:46:40 +0100425void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
426{
427}
Gregory Haskins917b6272008-12-29 09:39:53 -0500428
Peter Zijlstradc877342014-02-12 15:47:29 +0100429static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
430{
431 return false;
432}
433
Peter Zijlstra8046d682015-06-11 14:46:40 +0200434static inline void pull_rt_task(struct rq *this_rq)
Peter Zijlstradc877342014-02-12 15:47:29 +0100435{
Peter Zijlstradc877342014-02-12 15:47:29 +0100436}
437
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +0200438static inline void queue_push_tasks(struct rq *rq)
Peter Zijlstradc877342014-02-12 15:47:29 +0100439{
440}
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200441#endif /* CONFIG_SMP */
442
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400443static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
444static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
445
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100446static inline int on_rt_rq(struct sched_rt_entity *rt_se)
447{
Peter Zijlstraff77e462016-01-18 15:27:07 +0100448 return rt_se->on_rq;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100449}
450
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100451#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100452
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100453static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100454{
455 if (!rt_rq->tg)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100456 return RUNTIME_INF;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100457
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200458 return rt_rq->rt_runtime;
459}
460
461static inline u64 sched_rt_period(struct rt_rq *rt_rq)
462{
463 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100464}
465
Cheng Xuec514c42011-05-14 14:20:02 +0800466typedef struct task_group *rt_rq_iter_t;
467
Yong Zhang1c09ab02011-06-28 10:51:31 +0800468static inline struct task_group *next_task_group(struct task_group *tg)
469{
470 do {
471 tg = list_entry_rcu(tg->list.next,
472 typeof(struct task_group), list);
473 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
474
475 if (&tg->list == &task_groups)
476 tg = NULL;
477
478 return tg;
479}
480
481#define for_each_rt_rq(rt_rq, iter, rq) \
482 for (iter = container_of(&task_groups, typeof(*iter), list); \
483 (iter = next_task_group(iter)) && \
484 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
Cheng Xuec514c42011-05-14 14:20:02 +0800485
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100486#define for_each_sched_rt_entity(rt_se) \
487 for (; rt_se; rt_se = rt_se->parent)
488
489static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
490{
491 return rt_se->my_q;
492}
493
Peter Zijlstraff77e462016-01-18 15:27:07 +0100494static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
495static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100496
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100497static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100498{
Dario Faggiolif6121f42008-10-03 17:40:46 +0200499 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
Kirill Tkhai88751252014-06-29 00:03:57 +0400500 struct rq *rq = rq_of_rt_rq(rt_rq);
Yong Zhang74b7eb52010-01-29 14:57:52 +0800501 struct sched_rt_entity *rt_se;
502
Kirill Tkhai88751252014-06-29 00:03:57 +0400503 int cpu = cpu_of(rq);
Balbir Singh0c3b9162011-03-03 17:04:35 +0530504
505 rt_se = rt_rq->tg->rt_se[cpu];
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100506
Dario Faggiolif6121f42008-10-03 17:40:46 +0200507 if (rt_rq->rt_nr_running) {
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400508 if (!rt_se)
509 enqueue_top_rt_rq(rt_rq);
510 else if (!on_rt_rq(rt_se))
Peter Zijlstraff77e462016-01-18 15:27:07 +0100511 enqueue_rt_entity(rt_se, 0);
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400512
Gregory Haskinse864c492008-12-29 09:39:49 -0500513 if (rt_rq->highest_prio.curr < curr->prio)
Kirill Tkhai88751252014-06-29 00:03:57 +0400514 resched_curr(rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100515 }
516}
517
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100518static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100519{
Yong Zhang74b7eb52010-01-29 14:57:52 +0800520 struct sched_rt_entity *rt_se;
Balbir Singh0c3b9162011-03-03 17:04:35 +0530521 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
Yong Zhang74b7eb52010-01-29 14:57:52 +0800522
Balbir Singh0c3b9162011-03-03 17:04:35 +0530523 rt_se = rt_rq->tg->rt_se[cpu];
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100524
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400525 if (!rt_se)
526 dequeue_top_rt_rq(rt_rq);
527 else if (on_rt_rq(rt_se))
Peter Zijlstraff77e462016-01-18 15:27:07 +0100528 dequeue_rt_entity(rt_se, 0);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100529}
530
Kirill Tkhai46383642014-03-15 02:15:07 +0400531static inline int rt_rq_throttled(struct rt_rq *rt_rq)
532{
533 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
534}
535
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100536static int rt_se_boosted(struct sched_rt_entity *rt_se)
537{
538 struct rt_rq *rt_rq = group_rt_rq(rt_se);
539 struct task_struct *p;
540
541 if (rt_rq)
542 return !!rt_rq->rt_nr_boosted;
543
544 p = rt_task_of(rt_se);
545 return p->prio != p->normal_prio;
546}
547
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200548#ifdef CONFIG_SMP
Rusty Russellc6c49272008-11-25 02:35:05 +1030549static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200550{
Nathan Zimmer424c93f2013-05-09 11:24:03 -0500551 return this_rq()->rd->span;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200552}
553#else
Rusty Russellc6c49272008-11-25 02:35:05 +1030554static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200555{
Rusty Russellc6c49272008-11-25 02:35:05 +1030556 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200557}
558#endif
559
560static inline
561struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
562{
563 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
564}
565
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200566static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
567{
568 return &rt_rq->tg->rt_bandwidth;
569}
570
Dhaval Giani55e12e52008-06-24 23:39:43 +0530571#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100572
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100573static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100574{
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200575 return rt_rq->rt_runtime;
576}
577
578static inline u64 sched_rt_period(struct rt_rq *rt_rq)
579{
580 return ktime_to_ns(def_rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100581}
582
Cheng Xuec514c42011-05-14 14:20:02 +0800583typedef struct rt_rq *rt_rq_iter_t;
584
585#define for_each_rt_rq(rt_rq, iter, rq) \
586 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
587
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100588#define for_each_sched_rt_entity(rt_se) \
589 for (; rt_se; rt_se = NULL)
590
591static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
592{
593 return NULL;
594}
595
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100596static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100597{
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400598 struct rq *rq = rq_of_rt_rq(rt_rq);
599
600 if (!rt_rq->rt_nr_running)
601 return;
602
603 enqueue_top_rt_rq(rt_rq);
Kirill Tkhai88751252014-06-29 00:03:57 +0400604 resched_curr(rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100605}
606
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100607static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100608{
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +0400609 dequeue_top_rt_rq(rt_rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100610}
611
Kirill Tkhai46383642014-03-15 02:15:07 +0400612static inline int rt_rq_throttled(struct rt_rq *rt_rq)
613{
614 return rt_rq->rt_throttled;
615}
616
Rusty Russellc6c49272008-11-25 02:35:05 +1030617static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200618{
Rusty Russellc6c49272008-11-25 02:35:05 +1030619 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200620}
621
622static inline
623struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
624{
625 return &cpu_rq(cpu)->rt;
626}
627
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200628static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
629{
630 return &def_rt_bandwidth;
631}
632
Dhaval Giani55e12e52008-06-24 23:39:43 +0530633#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100634
Juri Lellifaa59932014-02-21 11:37:15 +0100635bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
636{
637 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
638
639 return (hrtimer_active(&rt_b->rt_period_timer) ||
640 rt_rq->rt_time < rt_b->rt_runtime);
641}
642
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200643#ifdef CONFIG_SMP
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200644/*
645 * We ran out of runtime, see if we can borrow some from our neighbours.
646 */
Juri Lelli269b26a2015-09-02 11:01:36 +0100647static void do_balance_runtime(struct rt_rq *rt_rq)
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200648{
649 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
Shawn Bohreraa7f6732013-01-14 11:55:31 -0600650 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
Juri Lelli269b26a2015-09-02 11:01:36 +0100651 int i, weight;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200652 u64 rt_period;
653
Rusty Russellc6c49272008-11-25 02:35:05 +1030654 weight = cpumask_weight(rd->span);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200655
Thomas Gleixner0986b112009-11-17 15:32:06 +0100656 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200657 rt_period = ktime_to_ns(rt_b->rt_period);
Rusty Russellc6c49272008-11-25 02:35:05 +1030658 for_each_cpu(i, rd->span) {
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200659 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
660 s64 diff;
661
662 if (iter == rt_rq)
663 continue;
664
Thomas Gleixner0986b112009-11-17 15:32:06 +0100665 raw_spin_lock(&iter->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200666 /*
667 * Either all rqs have inf runtime and there's nothing to steal
668 * or __disable_runtime() below sets a specific rq to inf to
669 * indicate its been disabled and disalow stealing.
670 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200671 if (iter->rt_runtime == RUNTIME_INF)
672 goto next;
673
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200674 /*
675 * From runqueues with spare time, take 1/n part of their
676 * spare time, but no more than our period.
677 */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200678 diff = iter->rt_runtime - iter->rt_time;
679 if (diff > 0) {
Peter Zijlstra58838cf2008-07-24 12:43:13 +0200680 diff = div_u64((u64)diff, weight);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200681 if (rt_rq->rt_runtime + diff > rt_period)
682 diff = rt_period - rt_rq->rt_runtime;
683 iter->rt_runtime -= diff;
684 rt_rq->rt_runtime += diff;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200685 if (rt_rq->rt_runtime == rt_period) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100686 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200687 break;
688 }
689 }
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200690next:
Thomas Gleixner0986b112009-11-17 15:32:06 +0100691 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200692 }
Thomas Gleixner0986b112009-11-17 15:32:06 +0100693 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200694}
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200695
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200696/*
697 * Ensure this RQ takes back all the runtime it lend to its neighbours.
698 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200699static void __disable_runtime(struct rq *rq)
700{
701 struct root_domain *rd = rq->rd;
Cheng Xuec514c42011-05-14 14:20:02 +0800702 rt_rq_iter_t iter;
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200703 struct rt_rq *rt_rq;
704
705 if (unlikely(!scheduler_running))
706 return;
707
Cheng Xuec514c42011-05-14 14:20:02 +0800708 for_each_rt_rq(rt_rq, iter, rq) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200709 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
710 s64 want;
711 int i;
712
Thomas Gleixner0986b112009-11-17 15:32:06 +0100713 raw_spin_lock(&rt_b->rt_runtime_lock);
714 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200715 /*
716 * Either we're all inf and nobody needs to borrow, or we're
717 * already disabled and thus have nothing to do, or we have
718 * exactly the right amount of runtime to take out.
719 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200720 if (rt_rq->rt_runtime == RUNTIME_INF ||
721 rt_rq->rt_runtime == rt_b->rt_runtime)
722 goto balanced;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100723 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200724
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200725 /*
726 * Calculate the difference between what we started out with
727 * and what we current have, that's the amount of runtime
728 * we lend and now have to reclaim.
729 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200730 want = rt_b->rt_runtime - rt_rq->rt_runtime;
731
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200732 /*
733 * Greedy reclaim, take back as much as we can.
734 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030735 for_each_cpu(i, rd->span) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200736 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
737 s64 diff;
738
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200739 /*
740 * Can't reclaim from ourselves or disabled runqueues.
741 */
Peter Zijlstraf1679d02008-08-14 15:49:00 +0200742 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200743 continue;
744
Thomas Gleixner0986b112009-11-17 15:32:06 +0100745 raw_spin_lock(&iter->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200746 if (want > 0) {
747 diff = min_t(s64, iter->rt_runtime, want);
748 iter->rt_runtime -= diff;
749 want -= diff;
750 } else {
751 iter->rt_runtime -= want;
752 want -= want;
753 }
Thomas Gleixner0986b112009-11-17 15:32:06 +0100754 raw_spin_unlock(&iter->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200755
756 if (!want)
757 break;
758 }
759
Thomas Gleixner0986b112009-11-17 15:32:06 +0100760 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200761 /*
762 * We cannot be left wanting - that would mean some runtime
763 * leaked out of the system.
764 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200765 BUG_ON(want);
766balanced:
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200767 /*
768 * Disable all the borrow logic by pretending we have inf
769 * runtime - in which case borrowing doesn't make sense.
770 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200771 rt_rq->rt_runtime = RUNTIME_INF;
Peter Boonstoppela4c96ae2012-08-09 15:34:47 -0700772 rt_rq->rt_throttled = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100773 raw_spin_unlock(&rt_rq->rt_runtime_lock);
774 raw_spin_unlock(&rt_b->rt_runtime_lock);
Kirill Tkhai99b62562014-06-25 12:19:48 +0400775
776 /* Make rt_rq available for pick_next_task() */
777 sched_rt_rq_enqueue(rt_rq);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200778 }
779}
780
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200781static void __enable_runtime(struct rq *rq)
782{
Cheng Xuec514c42011-05-14 14:20:02 +0800783 rt_rq_iter_t iter;
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200784 struct rt_rq *rt_rq;
785
786 if (unlikely(!scheduler_running))
787 return;
788
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200789 /*
790 * Reset each runqueue's bandwidth settings
791 */
Cheng Xuec514c42011-05-14 14:20:02 +0800792 for_each_rt_rq(rt_rq, iter, rq) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200793 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
794
Thomas Gleixner0986b112009-11-17 15:32:06 +0100795 raw_spin_lock(&rt_b->rt_runtime_lock);
796 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200797 rt_rq->rt_runtime = rt_b->rt_runtime;
798 rt_rq->rt_time = 0;
Zhang, Yanminbaf25732008-09-09 11:26:33 +0800799 rt_rq->rt_throttled = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100800 raw_spin_unlock(&rt_rq->rt_runtime_lock);
801 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200802 }
803}
804
Juri Lelli269b26a2015-09-02 11:01:36 +0100805static void balance_runtime(struct rt_rq *rt_rq)
Peter Zijlstraeff65492008-06-19 14:22:26 +0200806{
Peter Zijlstra4a6184c2011-10-06 22:39:14 +0200807 if (!sched_feat(RT_RUNTIME_SHARE))
Juri Lelli269b26a2015-09-02 11:01:36 +0100808 return;
Peter Zijlstra4a6184c2011-10-06 22:39:14 +0200809
Peter Zijlstraeff65492008-06-19 14:22:26 +0200810 if (rt_rq->rt_time > rt_rq->rt_runtime) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100811 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Juri Lelli269b26a2015-09-02 11:01:36 +0100812 do_balance_runtime(rt_rq);
Thomas Gleixner0986b112009-11-17 15:32:06 +0100813 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200814 }
Peter Zijlstraeff65492008-06-19 14:22:26 +0200815}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530816#else /* !CONFIG_SMP */
Juri Lelli269b26a2015-09-02 11:01:36 +0100817static inline void balance_runtime(struct rt_rq *rt_rq) {}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530818#endif /* CONFIG_SMP */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100819
820static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
821{
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200822 int i, idle = 1, throttled = 0;
Rusty Russellc6c49272008-11-25 02:35:05 +1030823 const struct cpumask *span;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200824
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200825 span = sched_rt_period_mask();
Mike Galbraithe221d022012-08-07 10:02:38 +0200826#ifdef CONFIG_RT_GROUP_SCHED
827 /*
828 * FIXME: isolated CPUs should really leave the root task group,
829 * whether they are isolcpus or were isolated via cpusets, lest
830 * the timer run on a CPU which does not service all runqueues,
831 * potentially leaving other CPUs indefinitely throttled. If
832 * isolation is really required, the user will turn the throttle
833 * off to kill the perturbations it causes anyway. Meanwhile,
834 * this maintains functionality for boot and/or troubleshooting.
835 */
836 if (rt_b == &root_task_group.rt_bandwidth)
837 span = cpu_online_mask;
838#endif
Rusty Russellc6c49272008-11-25 02:35:05 +1030839 for_each_cpu(i, span) {
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200840 int enqueue = 0;
841 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
842 struct rq *rq = rq_of_rt_rq(rt_rq);
Dave Kleikampc249f252017-05-15 14:14:13 -0500843 int skip;
844
845 /*
846 * When span == cpu_online_mask, taking each rq->lock
847 * can be time-consuming. Try to avoid it when possible.
848 */
849 raw_spin_lock(&rt_rq->rt_runtime_lock);
850 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
851 raw_spin_unlock(&rt_rq->rt_runtime_lock);
852 if (skip)
853 continue;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200854
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100855 raw_spin_lock(&rq->lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200856 if (rt_rq->rt_time) {
857 u64 runtime;
858
Thomas Gleixner0986b112009-11-17 15:32:06 +0100859 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200860 if (rt_rq->rt_throttled)
861 balance_runtime(rt_rq);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200862 runtime = rt_rq->rt_runtime;
863 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
864 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
865 rt_rq->rt_throttled = 0;
866 enqueue = 1;
Mike Galbraith61eadef2011-04-29 08:36:50 +0200867
868 /*
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +0100869 * When we're idle and a woken (rt) task is
870 * throttled check_preempt_curr() will set
871 * skip_update and the time between the wakeup
872 * and this unthrottle will get accounted as
873 * 'runtime'.
Mike Galbraith61eadef2011-04-29 08:36:50 +0200874 */
875 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +0100876 rq_clock_skip_update(rq, false);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200877 }
878 if (rt_rq->rt_time || rt_rq->rt_nr_running)
879 idle = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +0100880 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Balbir Singh0c3b9162011-03-03 17:04:35 +0530881 } else if (rt_rq->rt_nr_running) {
Peter Zijlstra8a8cde12008-06-19 14:22:28 +0200882 idle = 0;
Balbir Singh0c3b9162011-03-03 17:04:35 +0530883 if (!rt_rq_throttled(rt_rq))
884 enqueue = 1;
885 }
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200886 if (rt_rq->rt_throttled)
887 throttled = 1;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200888
889 if (enqueue)
890 sched_rt_rq_enqueue(rt_rq);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100891 raw_spin_unlock(&rq->lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200892 }
893
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200894 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
895 return 1;
896
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200897 return idle;
898}
899
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100900static inline int rt_se_prio(struct sched_rt_entity *rt_se)
901{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100902#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100903 struct rt_rq *rt_rq = group_rt_rq(rt_se);
904
905 if (rt_rq)
Gregory Haskinse864c492008-12-29 09:39:49 -0500906 return rt_rq->highest_prio.curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100907#endif
908
909 return rt_task_of(rt_se)->prio;
910}
911
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100912static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100913{
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100914 u64 runtime = sched_rt_runtime(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100915
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100916 if (rt_rq->rt_throttled)
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100917 return rt_rq_throttled(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100918
Shan Hai5b680fd2011-11-29 11:03:56 +0800919 if (runtime >= sched_rt_period(rt_rq))
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200920 return 0;
921
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200922 balance_runtime(rt_rq);
923 runtime = sched_rt_runtime(rt_rq);
924 if (runtime == RUNTIME_INF)
925 return 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200926
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100927 if (rt_rq->rt_time > runtime) {
Peter Zijlstra7abc63b2011-10-18 22:03:48 +0200928 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
929
930 /*
931 * Don't actually throttle groups that have no runtime assigned
932 * but accrue some time due to boosting.
933 */
934 if (likely(rt_b->rt_runtime)) {
935 rt_rq->rt_throttled = 1;
John Stultzc2248152014-06-04 16:11:41 -0700936 printk_deferred_once("sched: RT throttling activated\n");
Peter Zijlstra7abc63b2011-10-18 22:03:48 +0200937 } else {
938 /*
939 * In case we did anyway, make it go away,
940 * replenishment is a joke, since it will replenish us
941 * with exactly 0 ns.
942 */
943 rt_rq->rt_time = 0;
944 }
945
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100946 if (rt_rq_throttled(rt_rq)) {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100947 sched_rt_rq_dequeue(rt_rq);
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100948 return 1;
949 }
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100950 }
951
952 return 0;
953}
954
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200955/*
956 * Update the current task's runtime statistics. Skip current tasks that
957 * are not in our scheduling class.
958 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200959static void update_curr_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200960{
961 struct task_struct *curr = rq->curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100962 struct sched_rt_entity *rt_se = &curr->rt;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200963 u64 delta_exec;
964
Peter Zijlstra06c3bc62011-02-02 13:19:48 +0100965 if (curr->sched_class != &rt_sched_class)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200966 return;
967
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200968 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
Kirill Tkhaifc79e242013-01-30 16:50:36 +0400969 if (unlikely((s64)delta_exec <= 0))
970 return;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +0200971
Rafael J. Wysocki58919e82016-08-16 22:14:55 +0200972 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
Viresh Kumar674e7542017-07-28 12:16:38 +0530973 cpufreq_update_util(rq, SCHED_CPUFREQ_RT);
Wanpeng Li594dd292016-04-22 17:07:24 +0800974
Peter Zijlstra42c62a52011-10-18 22:03:48 +0200975 schedstat_set(curr->se.statistics.exec_max,
976 max(curr->se.statistics.exec_max, delta_exec));
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200977
978 curr->se.sum_exec_runtime += delta_exec;
Frank Mayharf06febc2008-09-12 09:54:39 -0700979 account_group_exec_runtime(curr, delta_exec);
980
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200981 curr->se.exec_start = rq_clock_task(rq);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100982 cpuacct_charge(curr, delta_exec);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100983
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200984 sched_rt_avg_update(rq, delta_exec);
985
Peter Zijlstra0b148fa2008-08-19 12:33:04 +0200986 if (!rt_bandwidth_enabled())
987 return;
988
Dhaval Giani354d60c2008-04-19 19:44:59 +0200989 for_each_sched_rt_entity(rt_se) {
Giedrius Rekasius0b079392014-05-25 15:23:31 +0100990 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
Dhaval Giani354d60c2008-04-19 19:44:59 +0200991
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200992 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
Thomas Gleixner0986b112009-11-17 15:32:06 +0100993 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200994 rt_rq->rt_time += delta_exec;
995 if (sched_rt_runtime_exceeded(rt_rq))
Kirill Tkhai88751252014-06-29 00:03:57 +0400996 resched_curr(rq);
Thomas Gleixner0986b112009-11-17 15:32:06 +0100997 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200998 }
Dhaval Giani354d60c2008-04-19 19:44:59 +0200999 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001000}
1001
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001002static void
1003dequeue_top_rt_rq(struct rt_rq *rt_rq)
1004{
1005 struct rq *rq = rq_of_rt_rq(rt_rq);
1006
1007 BUG_ON(&rq->rt != rt_rq);
1008
1009 if (!rt_rq->rt_queued)
1010 return;
1011
1012 BUG_ON(!rq->nr_running);
1013
Kirill Tkhai72465442014-05-09 03:00:14 +04001014 sub_nr_running(rq, rt_rq->rt_nr_running);
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001015 rt_rq->rt_queued = 0;
1016}
1017
1018static void
1019enqueue_top_rt_rq(struct rt_rq *rt_rq)
1020{
1021 struct rq *rq = rq_of_rt_rq(rt_rq);
1022
1023 BUG_ON(&rq->rt != rt_rq);
1024
1025 if (rt_rq->rt_queued)
1026 return;
1027 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1028 return;
1029
Kirill Tkhai72465442014-05-09 03:00:14 +04001030 add_nr_running(rq, rt_rq->rt_nr_running);
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001031 rt_rq->rt_queued = 1;
1032}
1033
Gregory Haskins398a1532009-01-14 09:10:04 -05001034#if defined CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -05001035
Gregory Haskins398a1532009-01-14 09:10:04 -05001036static void
1037inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +01001038{
Gregory Haskins4d984272008-12-29 09:39:49 -05001039 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins4d984272008-12-29 09:39:49 -05001040
Kirill Tkhai757dfca2013-11-27 19:59:13 +04001041#ifdef CONFIG_RT_GROUP_SCHED
1042 /*
1043 * Change rq's cpupri only if rt_rq is the top queue.
1044 */
1045 if (&rq->rt != rt_rq)
1046 return;
1047#endif
Steven Rostedt5181f4a42011-06-16 21:55:23 -04001048 if (rq->online && prio < prev_prio)
1049 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
Steven Rostedt63489e42008-01-25 21:08:03 +01001050}
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001051
Gregory Haskins398a1532009-01-14 09:10:04 -05001052static void
1053dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +01001054{
Gregory Haskins4d984272008-12-29 09:39:49 -05001055 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001056
Kirill Tkhai757dfca2013-11-27 19:59:13 +04001057#ifdef CONFIG_RT_GROUP_SCHED
1058 /*
1059 * Change rq's cpupri only if rt_rq is the top queue.
1060 */
1061 if (&rq->rt != rt_rq)
1062 return;
1063#endif
Gregory Haskins398a1532009-01-14 09:10:04 -05001064 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1065 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1066}
1067
1068#else /* CONFIG_SMP */
1069
1070static inline
1071void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1072static inline
1073void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1074
1075#endif /* CONFIG_SMP */
1076
Steven Rostedt63489e42008-01-25 21:08:03 +01001077#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -05001078static void
1079inc_rt_prio(struct rt_rq *rt_rq, int prio)
1080{
1081 int prev_prio = rt_rq->highest_prio.curr;
Steven Rostedt63489e42008-01-25 21:08:03 +01001082
Gregory Haskins398a1532009-01-14 09:10:04 -05001083 if (prio < prev_prio)
1084 rt_rq->highest_prio.curr = prio;
1085
1086 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1087}
1088
1089static void
1090dec_rt_prio(struct rt_rq *rt_rq, int prio)
1091{
1092 int prev_prio = rt_rq->highest_prio.curr;
1093
1094 if (rt_rq->rt_nr_running) {
1095
1096 WARN_ON(prio < prev_prio);
Gregory Haskinse864c492008-12-29 09:39:49 -05001097
1098 /*
Gregory Haskins398a1532009-01-14 09:10:04 -05001099 * This may have been our highest task, and therefore
1100 * we may have some recomputation to do
Gregory Haskinse864c492008-12-29 09:39:49 -05001101 */
Gregory Haskins398a1532009-01-14 09:10:04 -05001102 if (prio == prev_prio) {
Gregory Haskinse864c492008-12-29 09:39:49 -05001103 struct rt_prio_array *array = &rt_rq->active;
1104
1105 rt_rq->highest_prio.curr =
Steven Rostedt764a9d62008-01-25 21:08:04 +01001106 sched_find_first_bit(array->bitmap);
Gregory Haskinse864c492008-12-29 09:39:49 -05001107 }
1108
Steven Rostedt764a9d62008-01-25 21:08:04 +01001109 } else
Gregory Haskinse864c492008-12-29 09:39:49 -05001110 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001111
Gregory Haskins398a1532009-01-14 09:10:04 -05001112 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1113}
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001114
Gregory Haskins398a1532009-01-14 09:10:04 -05001115#else
1116
1117static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1118static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1119
1120#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1121
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001122#ifdef CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -05001123
1124static void
1125inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1126{
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01001127 if (rt_se_boosted(rt_se))
Steven Rostedt764a9d62008-01-25 21:08:04 +01001128 rt_rq->rt_nr_boosted++;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01001129
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001130 if (rt_rq->tg)
1131 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
Gregory Haskins398a1532009-01-14 09:10:04 -05001132}
1133
1134static void
1135dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1136{
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001137 if (rt_se_boosted(rt_se))
1138 rt_rq->rt_nr_boosted--;
1139
1140 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
Gregory Haskins398a1532009-01-14 09:10:04 -05001141}
1142
1143#else /* CONFIG_RT_GROUP_SCHED */
1144
1145static void
1146inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1147{
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001148 start_rt_bandwidth(&def_rt_bandwidth);
Gregory Haskins398a1532009-01-14 09:10:04 -05001149}
1150
1151static inline
1152void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1153
1154#endif /* CONFIG_RT_GROUP_SCHED */
1155
1156static inline
Kirill Tkhai22abdef2014-03-15 02:14:49 +04001157unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1158{
1159 struct rt_rq *group_rq = group_rt_rq(rt_se);
1160
1161 if (group_rq)
1162 return group_rq->rt_nr_running;
1163 else
1164 return 1;
1165}
1166
1167static inline
Frederic Weisbecker01d36d02015-11-04 18:17:10 +01001168unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1169{
1170 struct rt_rq *group_rq = group_rt_rq(rt_se);
1171 struct task_struct *tsk;
1172
1173 if (group_rq)
1174 return group_rq->rr_nr_running;
1175
1176 tsk = rt_task_of(rt_se);
1177
1178 return (tsk->policy == SCHED_RR) ? 1 : 0;
1179}
1180
1181static inline
Gregory Haskins398a1532009-01-14 09:10:04 -05001182void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1183{
1184 int prio = rt_se_prio(rt_se);
1185
1186 WARN_ON(!rt_prio(prio));
Kirill Tkhai22abdef2014-03-15 02:14:49 +04001187 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
Frederic Weisbecker01d36d02015-11-04 18:17:10 +01001188 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
Gregory Haskins398a1532009-01-14 09:10:04 -05001189
1190 inc_rt_prio(rt_rq, prio);
1191 inc_rt_migration(rt_se, rt_rq);
1192 inc_rt_group(rt_se, rt_rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001193}
1194
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001195static inline
1196void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1197{
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001198 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001199 WARN_ON(!rt_rq->rt_nr_running);
Kirill Tkhai22abdef2014-03-15 02:14:49 +04001200 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
Frederic Weisbecker01d36d02015-11-04 18:17:10 +01001201 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001202
Gregory Haskins398a1532009-01-14 09:10:04 -05001203 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1204 dec_rt_migration(rt_se, rt_rq);
1205 dec_rt_group(rt_se, rt_rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001206}
1207
Peter Zijlstraff77e462016-01-18 15:27:07 +01001208/*
1209 * Change rt_se->run_list location unless SAVE && !MOVE
1210 *
1211 * assumes ENQUEUE/DEQUEUE flags match
1212 */
1213static inline bool move_entity(unsigned int flags)
1214{
1215 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1216 return false;
1217
1218 return true;
1219}
1220
1221static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1222{
1223 list_del_init(&rt_se->run_list);
1224
1225 if (list_empty(array->queue + rt_se_prio(rt_se)))
1226 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1227
1228 rt_se->on_list = 0;
1229}
1230
1231static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001232{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001233 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1234 struct rt_prio_array *array = &rt_rq->active;
1235 struct rt_rq *group_rq = group_rt_rq(rt_se);
Dmitry Adamushko20b63312008-06-11 00:58:30 +02001236 struct list_head *queue = array->queue + rt_se_prio(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001237
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001238 /*
1239 * Don't enqueue the group if its throttled, or when empty.
1240 * The latter is a consequence of the former when a child group
1241 * get throttled and the current group doesn't have any other
1242 * active members.
1243 */
Peter Zijlstraff77e462016-01-18 15:27:07 +01001244 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1245 if (rt_se->on_list)
1246 __delist_rt_entity(rt_se, array);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001247 return;
Peter Zijlstraff77e462016-01-18 15:27:07 +01001248 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001249
Peter Zijlstraff77e462016-01-18 15:27:07 +01001250 if (move_entity(flags)) {
1251 WARN_ON_ONCE(rt_se->on_list);
1252 if (flags & ENQUEUE_HEAD)
1253 list_add(&rt_se->run_list, queue);
1254 else
1255 list_add_tail(&rt_se->run_list, queue);
1256
1257 __set_bit(rt_se_prio(rt_se), array->bitmap);
1258 rt_se->on_list = 1;
1259 }
1260 rt_se->on_rq = 1;
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001261
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001262 inc_rt_tasks(rt_se, rt_rq);
1263}
1264
Peter Zijlstraff77e462016-01-18 15:27:07 +01001265static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001266{
1267 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1268 struct rt_prio_array *array = &rt_rq->active;
1269
Peter Zijlstraff77e462016-01-18 15:27:07 +01001270 if (move_entity(flags)) {
1271 WARN_ON_ONCE(!rt_se->on_list);
1272 __delist_rt_entity(rt_se, array);
1273 }
1274 rt_se->on_rq = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001275
1276 dec_rt_tasks(rt_se, rt_rq);
1277}
1278
1279/*
1280 * Because the prio of an upper entry depends on the lower
1281 * entries, we must remove entries top - down.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001282 */
Peter Zijlstraff77e462016-01-18 15:27:07 +01001283static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001284{
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001285 struct sched_rt_entity *back = NULL;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001286
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +02001287 for_each_sched_rt_entity(rt_se) {
1288 rt_se->back = back;
1289 back = rt_se;
1290 }
1291
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001292 dequeue_top_rt_rq(rt_rq_of_se(back));
1293
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +02001294 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1295 if (on_rt_rq(rt_se))
Peter Zijlstraff77e462016-01-18 15:27:07 +01001296 __dequeue_rt_entity(rt_se, flags);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001297 }
1298}
1299
Peter Zijlstraff77e462016-01-18 15:27:07 +01001300static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001301{
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001302 struct rq *rq = rq_of_rt_se(rt_se);
1303
Peter Zijlstraff77e462016-01-18 15:27:07 +01001304 dequeue_rt_stack(rt_se, flags);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001305 for_each_sched_rt_entity(rt_se)
Peter Zijlstraff77e462016-01-18 15:27:07 +01001306 __enqueue_rt_entity(rt_se, flags);
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001307 enqueue_top_rt_rq(&rq->rt);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001308}
1309
Peter Zijlstraff77e462016-01-18 15:27:07 +01001310static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001311{
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001312 struct rq *rq = rq_of_rt_se(rt_se);
1313
Peter Zijlstraff77e462016-01-18 15:27:07 +01001314 dequeue_rt_stack(rt_se, flags);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +02001315
1316 for_each_sched_rt_entity(rt_se) {
1317 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1318
1319 if (rt_rq && rt_rq->rt_nr_running)
Peter Zijlstraff77e462016-01-18 15:27:07 +01001320 __enqueue_rt_entity(rt_se, flags);
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +02001321 }
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001322 enqueue_top_rt_rq(&rq->rt);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001323}
1324
1325/*
1326 * Adding/removing a task to/from a priority array:
1327 */
Thomas Gleixnerea87bb72010-01-20 20:58:57 +00001328static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001329enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001330{
1331 struct sched_rt_entity *rt_se = &p->rt;
1332
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001333 if (flags & ENQUEUE_WAKEUP)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001334 rt_se->timeout = 0;
1335
Peter Zijlstraff77e462016-01-18 15:27:07 +01001336 enqueue_rt_entity(rt_se, flags);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001337
Ingo Molnar4b53a342017-02-05 15:41:03 +01001338 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
Gregory Haskins917b6272008-12-29 09:39:53 -05001339 enqueue_pushable_task(rq, p);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001340}
1341
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01001342static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001343{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001344 struct sched_rt_entity *rt_se = &p->rt;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001345
1346 update_curr_rt(rq);
Peter Zijlstraff77e462016-01-18 15:27:07 +01001347 dequeue_rt_entity(rt_se, flags);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001348
Gregory Haskins917b6272008-12-29 09:39:53 -05001349 dequeue_pushable_task(rq, p);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001350}
1351
1352/*
Richard Weinberger60686312011-11-12 18:07:57 +01001353 * Put task to the head or the end of the run list without the overhead of
1354 * dequeue followed by enqueue.
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001355 */
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001356static void
1357requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001358{
Ingo Molnar1cdad712008-06-19 09:09:15 +02001359 if (on_rt_rq(rt_se)) {
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001360 struct rt_prio_array *array = &rt_rq->active;
1361 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1362
1363 if (head)
1364 list_move(&rt_se->run_list, queue);
1365 else
1366 list_move_tail(&rt_se->run_list, queue);
Ingo Molnar1cdad712008-06-19 09:09:15 +02001367 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001368}
1369
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001370static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001371{
1372 struct sched_rt_entity *rt_se = &p->rt;
1373 struct rt_rq *rt_rq;
1374
1375 for_each_sched_rt_entity(rt_se) {
1376 rt_rq = rt_rq_of_se(rt_se);
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001377 requeue_rt_entity(rt_rq, rt_se, head);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001378 }
1379}
1380
1381static void yield_task_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001382{
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001383 requeue_task_rt(rq, rq->curr, 0);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001384}
1385
Gregory Haskinse7693a32008-01-25 21:08:09 +01001386#ifdef CONFIG_SMP
Gregory Haskins318e0892008-01-25 21:08:10 +01001387static int find_lowest_rq(struct task_struct *task);
1388
Peter Zijlstra0017d732010-03-24 18:34:10 +01001389static int
Peter Zijlstraac66f542013-10-07 11:29:16 +01001390select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
Gregory Haskinse7693a32008-01-25 21:08:09 +01001391{
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001392 struct task_struct *curr;
1393 struct rq *rq;
Steven Rostedtc37495f2011-06-16 21:55:22 -04001394
1395 /* For anything but wake ups, just return the task_cpu */
1396 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1397 goto out;
1398
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001399 rq = cpu_rq(cpu);
1400
1401 rcu_read_lock();
Jason Low316c1608d2015-04-28 13:00:20 -07001402 curr = READ_ONCE(rq->curr); /* unlocked access */
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001403
Gregory Haskins318e0892008-01-25 21:08:10 +01001404 /*
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001405 * If the current task on @p's runqueue is an RT task, then
Steven Rostedte1f47d82008-01-25 21:08:12 +01001406 * try to see if we can wake this RT task up on another
1407 * runqueue. Otherwise simply start this RT task
1408 * on its current runqueue.
1409 *
Steven Rostedt43fa5462010-09-20 22:40:03 -04001410 * We want to avoid overloading runqueues. If the woken
1411 * task is a higher priority, then it will stay on this CPU
1412 * and the lower prio task should be moved to another CPU.
1413 * Even though this will probably make the lower prio task
1414 * lose its cache, we do not want to bounce a higher task
1415 * around just because it gave up its CPU, perhaps for a
1416 * lock?
1417 *
1418 * For equal prio tasks, we just let the scheduler sort it out.
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001419 *
Gregory Haskins318e0892008-01-25 21:08:10 +01001420 * Otherwise, just let it ride on the affined RQ and the
1421 * post-schedule router will push the preempted task away
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001422 *
1423 * This test is optimistic, if we get it wrong the load-balancer
1424 * will have to sort it out.
Gregory Haskins318e0892008-01-25 21:08:10 +01001425 */
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001426 if (curr && unlikely(rt_task(curr)) &&
Ingo Molnar4b53a342017-02-05 15:41:03 +01001427 (curr->nr_cpus_allowed < 2 ||
Shawn Bohrer6bfa6872013-10-04 14:24:53 -05001428 curr->prio <= p->prio)) {
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001429 int target = find_lowest_rq(p);
1430
Tim Chen80e3d872014-12-12 15:38:12 -08001431 /*
1432 * Don't bother moving it if the destination CPU is
1433 * not running a lower priority task.
1434 */
1435 if (target != -1 &&
1436 p->prio < cpu_rq(target)->rt.highest_prio.curr)
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001437 cpu = target;
1438 }
1439 rcu_read_unlock();
1440
Steven Rostedtc37495f2011-06-16 21:55:22 -04001441out:
Peter Zijlstra7608dec2011-04-05 17:23:46 +02001442 return cpu;
Gregory Haskinse7693a32008-01-25 21:08:09 +01001443}
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001444
1445static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1446{
Wanpeng Li308a6232014-10-31 06:39:31 +08001447 /*
1448 * Current can't be migrated, useless to reschedule,
1449 * let's hope p can move out.
1450 */
Ingo Molnar4b53a342017-02-05 15:41:03 +01001451 if (rq->curr->nr_cpus_allowed == 1 ||
Wanpeng Li308a6232014-10-31 06:39:31 +08001452 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001453 return;
1454
Wanpeng Li308a6232014-10-31 06:39:31 +08001455 /*
1456 * p is migratable, so let's not schedule it and
1457 * see if it is pushed or pulled somewhere else.
1458 */
Ingo Molnar4b53a342017-02-05 15:41:03 +01001459 if (p->nr_cpus_allowed != 1
Rusty Russell13b8bd02009-03-25 15:01:22 +10301460 && cpupri_find(&rq->rd->cpupri, p, NULL))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001461 return;
1462
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001463 /*
1464 * There appears to be other cpus that can accept
1465 * current and none to run 'p', so lets reschedule
1466 * to try and push current away:
1467 */
1468 requeue_task_rt(rq, p, 1);
Kirill Tkhai88751252014-06-29 00:03:57 +04001469 resched_curr(rq);
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001470}
1471
Gregory Haskinse7693a32008-01-25 21:08:09 +01001472#endif /* CONFIG_SMP */
1473
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001474/*
1475 * Preempt the current task with a newly woken task if needed:
1476 */
Peter Zijlstra7d478722009-09-14 19:55:44 +02001477static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001478{
Gregory Haskins45c01e82008-05-12 21:20:41 +02001479 if (p->prio < rq->curr->prio) {
Kirill Tkhai88751252014-06-29 00:03:57 +04001480 resched_curr(rq);
Gregory Haskins45c01e82008-05-12 21:20:41 +02001481 return;
1482 }
1483
1484#ifdef CONFIG_SMP
1485 /*
1486 * If:
1487 *
1488 * - the newly woken task is of equal priority to the current task
1489 * - the newly woken task is non-migratable while current is migratable
1490 * - current will be preempted on the next reschedule
1491 *
1492 * we should check to see if current can readily move to a different
1493 * cpu. If so, we will reschedule to allow the push logic to try
1494 * to move current somewhere else, making room for our non-migratable
1495 * task.
1496 */
Hillf Danton8dd0de82011-06-14 18:36:24 -04001497 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001498 check_preempt_equal_prio(rq, p);
Gregory Haskins45c01e82008-05-12 21:20:41 +02001499#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001500}
1501
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001502static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1503 struct rt_rq *rt_rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001504{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001505 struct rt_prio_array *array = &rt_rq->active;
1506 struct sched_rt_entity *next = NULL;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001507 struct list_head *queue;
1508 int idx;
1509
1510 idx = sched_find_first_bit(array->bitmap);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001511 BUG_ON(idx >= MAX_RT_PRIO);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001512
1513 queue = array->queue + idx;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001514 next = list_entry(queue->next, struct sched_rt_entity, run_list);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001515
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001516 return next;
1517}
1518
Gregory Haskins917b6272008-12-29 09:39:53 -05001519static struct task_struct *_pick_next_task_rt(struct rq *rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001520{
1521 struct sched_rt_entity *rt_se;
1522 struct task_struct *p;
Peter Zijlstra606dba22012-02-11 06:05:00 +01001523 struct rt_rq *rt_rq = &rq->rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001524
1525 do {
1526 rt_se = pick_next_rt_entity(rq, rt_rq);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001527 BUG_ON(!rt_se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001528 rt_rq = group_rt_rq(rt_se);
1529 } while (rt_rq);
1530
1531 p = rt_task_of(rt_se);
Frederic Weisbecker78becc22013-04-12 01:51:02 +02001532 p->se.exec_start = rq_clock_task(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001533
1534 return p;
1535}
1536
Peter Zijlstra606dba22012-02-11 06:05:00 +01001537static struct task_struct *
Matt Flemingd8ac8972016-09-21 14:38:10 +01001538pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
Gregory Haskins917b6272008-12-29 09:39:53 -05001539{
Peter Zijlstra606dba22012-02-11 06:05:00 +01001540 struct task_struct *p;
1541 struct rt_rq *rt_rq = &rq->rt;
1542
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001543 if (need_pull_rt_task(rq, prev)) {
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001544 /*
1545 * This is OK, because current is on_cpu, which avoids it being
1546 * picked for load-balance and preemption/IRQs are still
1547 * disabled avoiding further scheduler activity on it and we're
1548 * being very careful to re-start the picking loop.
1549 */
Matt Flemingd8ac8972016-09-21 14:38:10 +01001550 rq_unpin_lock(rq, rf);
Peter Zijlstra38033c32014-01-23 20:32:21 +01001551 pull_rt_task(rq);
Matt Flemingd8ac8972016-09-21 14:38:10 +01001552 rq_repin_lock(rq, rf);
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001553 /*
1554 * pull_rt_task() can drop (and re-acquire) rq->lock; this
Kirill Tkhaia1d9a322014-04-10 17:38:36 +04001555 * means a dl or stop task can slip in, in which case we need
1556 * to re-start task selection.
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001557 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001558 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
Kirill Tkhaia1d9a322014-04-10 17:38:36 +04001559 rq->dl.dl_nr_running))
Peter Zijlstra37e117c2014-02-14 12:25:08 +01001560 return RETRY_TASK;
1561 }
Peter Zijlstra38033c32014-01-23 20:32:21 +01001562
Kirill Tkhai734ff2a2014-03-04 19:25:46 +04001563 /*
1564 * We may dequeue prev's rt_rq in put_prev_task().
1565 * So, we update time before rt_nr_running check.
1566 */
1567 if (prev->sched_class == &rt_sched_class)
1568 update_curr_rt(rq);
1569
Kirill Tkhaif4ebcbc2014-03-15 02:15:00 +04001570 if (!rt_rq->rt_queued)
Peter Zijlstra606dba22012-02-11 06:05:00 +01001571 return NULL;
1572
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01001573 put_prev_task(rq, prev);
Peter Zijlstra606dba22012-02-11 06:05:00 +01001574
1575 p = _pick_next_task_rt(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001576
1577 /* The running task is never eligible for pushing */
Kirill Tkhaif3f17682014-09-12 17:42:01 +04001578 dequeue_pushable_task(rq, p);
Gregory Haskins917b6272008-12-29 09:39:53 -05001579
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02001580 queue_push_tasks(rq);
Gregory Haskins3f029d32009-07-29 11:08:47 -04001581
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001582 return p;
1583}
1584
Ingo Molnar31ee5292007-08-09 11:16:49 +02001585static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001586{
Ingo Molnarf1e14ef2007-08-09 11:16:48 +02001587 update_curr_rt(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05001588
1589 /*
1590 * The previous task needs to be made eligible for pushing
1591 * if it is still active
1592 */
Ingo Molnar4b53a342017-02-05 15:41:03 +01001593 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
Gregory Haskins917b6272008-12-29 09:39:53 -05001594 enqueue_pushable_task(rq, p);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001595}
1596
Peter Williams681f3e62007-10-24 18:23:51 +02001597#ifdef CONFIG_SMP
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001598
Steven Rostedte8fa1362008-01-25 21:08:05 +01001599/* Only try algorithms three times */
1600#define RT_MAX_TRIES 3
1601
Steven Rostedtf65eda42008-01-25 21:08:07 +01001602static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1603{
1604 if (!task_running(rq, p) &&
Ingo Molnar0c98d342017-02-05 15:38:10 +01001605 cpumask_test_cpu(cpu, &p->cpus_allowed))
Steven Rostedtf65eda42008-01-25 21:08:07 +01001606 return 1;
1607 return 0;
1608}
1609
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001610/*
1611 * Return the highest pushable rq's task, which is suitable to be executed
1612 * on the cpu, NULL otherwise
1613 */
1614static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001615{
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001616 struct plist_head *head = &rq->rt.pushable_tasks;
1617 struct task_struct *p;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001618
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001619 if (!has_pushable_tasks(rq))
1620 return NULL;
Peter Zijlstra3d074672010-03-10 17:07:24 +01001621
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001622 plist_for_each_entry(p, head, pushable_tasks) {
1623 if (pick_rt_task(rq, p, cpu))
1624 return p;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001625 }
1626
Kirill Tkhaie23ee742013-06-07 15:37:43 -04001627 return NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001628}
1629
Rusty Russell0e3900e2008-11-25 02:35:13 +10301630static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001631
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001632static int find_lowest_rq(struct task_struct *task)
1633{
1634 struct sched_domain *sd;
Christoph Lameter4ba29682014-08-26 19:12:21 -05001635 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001636 int this_cpu = smp_processor_id();
1637 int cpu = task_cpu(task);
1638
Steven Rostedt0da938c2011-06-14 18:36:25 -04001639 /* Make sure the mask is initialized first */
1640 if (unlikely(!lowest_mask))
1641 return -1;
1642
Ingo Molnar4b53a342017-02-05 15:41:03 +01001643 if (task->nr_cpus_allowed == 1)
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001644 return -1; /* No other targets possible */
1645
1646 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
Gregory Haskins06f90db2008-01-25 21:08:13 +01001647 return -1; /* No targets found */
1648
1649 /*
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001650 * At this point we have built a mask of cpus representing the
1651 * lowest priority tasks in the system. Now we want to elect
1652 * the best one based on our affinity and topology.
1653 *
1654 * We prioritize the last cpu that the task executed on since
1655 * it is most likely cache-hot in that location.
1656 */
Rusty Russell96f874e22008-11-25 02:35:14 +10301657 if (cpumask_test_cpu(cpu, lowest_mask))
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001658 return cpu;
1659
1660 /*
1661 * Otherwise, we consult the sched_domains span maps to figure
1662 * out which cpu is logically closest to our hot cache data.
1663 */
Rusty Russelle2c88062009-11-03 14:53:15 +10301664 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1665 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001666
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001667 rcu_read_lock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301668 for_each_domain(cpu, sd) {
1669 if (sd->flags & SD_WAKE_AFFINE) {
1670 int best_cpu;
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001671
Rusty Russelle2c88062009-11-03 14:53:15 +10301672 /*
1673 * "this_cpu" is cheaper to preempt than a
1674 * remote processor.
1675 */
1676 if (this_cpu != -1 &&
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001677 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1678 rcu_read_unlock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301679 return this_cpu;
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001680 }
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001681
Rusty Russelle2c88062009-11-03 14:53:15 +10301682 best_cpu = cpumask_first_and(lowest_mask,
1683 sched_domain_span(sd));
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001684 if (best_cpu < nr_cpu_ids) {
1685 rcu_read_unlock();
Rusty Russelle2c88062009-11-03 14:53:15 +10301686 return best_cpu;
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001687 }
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001688 }
1689 }
Xiaotian Fengcd4ae6a2011-04-22 18:53:54 +08001690 rcu_read_unlock();
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001691
1692 /*
1693 * And finally, if there were no matches within the domains
1694 * just give the caller *something* to work with from the compatible
1695 * locations.
1696 */
Rusty Russelle2c88062009-11-03 14:53:15 +10301697 if (this_cpu != -1)
1698 return this_cpu;
1699
1700 cpu = cpumask_any(lowest_mask);
1701 if (cpu < nr_cpu_ids)
1702 return cpu;
1703 return -1;
Gregory Haskins07b40322008-01-25 21:08:10 +01001704}
1705
Steven Rostedte8fa1362008-01-25 21:08:05 +01001706/* Will lock the rq it finds */
Ingo Molnar4df64c02008-01-25 21:08:15 +01001707static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001708{
1709 struct rq *lowest_rq = NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001710 int tries;
Ingo Molnar4df64c02008-01-25 21:08:15 +01001711 int cpu;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001712
1713 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
Gregory Haskins07b40322008-01-25 21:08:10 +01001714 cpu = find_lowest_rq(task);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001715
Gregory Haskins2de0b462008-01-25 21:08:10 +01001716 if ((cpu == -1) || (cpu == rq->cpu))
Steven Rostedte8fa1362008-01-25 21:08:05 +01001717 break;
1718
Gregory Haskins07b40322008-01-25 21:08:10 +01001719 lowest_rq = cpu_rq(cpu);
1720
Tim Chen80e3d872014-12-12 15:38:12 -08001721 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1722 /*
1723 * Target rq has tasks of equal or higher priority,
1724 * retrying does not release any lock and is unlikely
1725 * to yield a different result.
1726 */
1727 lowest_rq = NULL;
1728 break;
1729 }
1730
Steven Rostedte8fa1362008-01-25 21:08:05 +01001731 /* if the prio of this runqueue changed, try again */
Gregory Haskins07b40322008-01-25 21:08:10 +01001732 if (double_lock_balance(rq, lowest_rq)) {
Steven Rostedte8fa1362008-01-25 21:08:05 +01001733 /*
1734 * We had to unlock the run queue. In
1735 * the mean time, task could have
1736 * migrated already or had its affinity changed.
1737 * Also make sure that it wasn't scheduled on its rq.
1738 */
Gregory Haskins07b40322008-01-25 21:08:10 +01001739 if (unlikely(task_rq(task) != rq ||
Ingo Molnar0c98d342017-02-05 15:38:10 +01001740 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
Gregory Haskins07b40322008-01-25 21:08:10 +01001741 task_running(rq, task) ||
Xunlei Pang13b5ab02016-05-09 12:11:31 +08001742 !rt_task(task) ||
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001743 !task_on_rq_queued(task))) {
Ingo Molnar4df64c02008-01-25 21:08:15 +01001744
Peter Zijlstra7f1b4392012-05-17 21:19:46 +02001745 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001746 lowest_rq = NULL;
1747 break;
1748 }
1749 }
1750
1751 /* If this rq is still suitable use it. */
Gregory Haskinse864c492008-12-29 09:39:49 -05001752 if (lowest_rq->rt.highest_prio.curr > task->prio)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001753 break;
1754
1755 /* try again */
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001756 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001757 lowest_rq = NULL;
1758 }
1759
1760 return lowest_rq;
1761}
1762
Gregory Haskins917b6272008-12-29 09:39:53 -05001763static struct task_struct *pick_next_pushable_task(struct rq *rq)
1764{
1765 struct task_struct *p;
1766
1767 if (!has_pushable_tasks(rq))
1768 return NULL;
1769
1770 p = plist_first_entry(&rq->rt.pushable_tasks,
1771 struct task_struct, pushable_tasks);
1772
1773 BUG_ON(rq->cpu != task_cpu(p));
1774 BUG_ON(task_current(rq, p));
Ingo Molnar4b53a342017-02-05 15:41:03 +01001775 BUG_ON(p->nr_cpus_allowed <= 1);
Gregory Haskins917b6272008-12-29 09:39:53 -05001776
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001777 BUG_ON(!task_on_rq_queued(p));
Gregory Haskins917b6272008-12-29 09:39:53 -05001778 BUG_ON(!rt_task(p));
1779
1780 return p;
1781}
1782
Steven Rostedte8fa1362008-01-25 21:08:05 +01001783/*
1784 * If the current CPU has more than one RT task, see if the non
1785 * running task can migrate over to a CPU that is running a task
1786 * of lesser priority.
1787 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001788static int push_rt_task(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001789{
1790 struct task_struct *next_task;
1791 struct rq *lowest_rq;
Hillf Danton311e8002011-06-16 21:55:20 -04001792 int ret = 0;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001793
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +01001794 if (!rq->rt.overloaded)
1795 return 0;
1796
Gregory Haskins917b6272008-12-29 09:39:53 -05001797 next_task = pick_next_pushable_task(rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001798 if (!next_task)
1799 return 0;
1800
Peter Zijlstra49246272010-10-17 21:46:10 +02001801retry:
Gregory Haskins697f0a42008-01-25 21:08:09 +01001802 if (unlikely(next_task == rq->curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001803 WARN_ON(1);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001804 return 0;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001805 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01001806
1807 /*
1808 * It's possible that the next_task slipped in of
1809 * higher priority than current. If that's the case
1810 * just reschedule current.
1811 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001812 if (unlikely(next_task->prio < rq->curr->prio)) {
Kirill Tkhai88751252014-06-29 00:03:57 +04001813 resched_curr(rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001814 return 0;
1815 }
1816
Gregory Haskins697f0a42008-01-25 21:08:09 +01001817 /* We might release rq lock */
Steven Rostedte8fa1362008-01-25 21:08:05 +01001818 get_task_struct(next_task);
1819
1820 /* find_lock_lowest_rq locks the rq if found */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001821 lowest_rq = find_lock_lowest_rq(next_task, rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001822 if (!lowest_rq) {
1823 struct task_struct *task;
1824 /*
Hillf Danton311e8002011-06-16 21:55:20 -04001825 * find_lock_lowest_rq releases rq->lock
Gregory Haskins15635132008-12-29 09:39:53 -05001826 * so it is possible that next_task has migrated.
1827 *
1828 * We need to make sure that the task is still on the same
1829 * run-queue and is also still the next task eligible for
1830 * pushing.
Steven Rostedte8fa1362008-01-25 21:08:05 +01001831 */
Gregory Haskins917b6272008-12-29 09:39:53 -05001832 task = pick_next_pushable_task(rq);
Byungchul Parkde16b912017-05-12 10:05:43 +09001833 if (task == next_task) {
Gregory Haskins15635132008-12-29 09:39:53 -05001834 /*
Hillf Danton311e8002011-06-16 21:55:20 -04001835 * The task hasn't migrated, and is still the next
1836 * eligible task, but we failed to find a run-queue
1837 * to push it to. Do not retry in this case, since
1838 * other cpus will pull from us when ready.
Gregory Haskins15635132008-12-29 09:39:53 -05001839 */
Gregory Haskins15635132008-12-29 09:39:53 -05001840 goto out;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001841 }
Gregory Haskins917b6272008-12-29 09:39:53 -05001842
Gregory Haskins15635132008-12-29 09:39:53 -05001843 if (!task)
1844 /* No more tasks, just exit */
1845 goto out;
1846
Gregory Haskins917b6272008-12-29 09:39:53 -05001847 /*
Gregory Haskins15635132008-12-29 09:39:53 -05001848 * Something has shifted, try again.
Gregory Haskins917b6272008-12-29 09:39:53 -05001849 */
Gregory Haskins15635132008-12-29 09:39:53 -05001850 put_task_struct(next_task);
1851 next_task = task;
1852 goto retry;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001853 }
1854
Gregory Haskins697f0a42008-01-25 21:08:09 +01001855 deactivate_task(rq, next_task, 0);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001856 set_task_cpu(next_task, lowest_rq->cpu);
1857 activate_task(lowest_rq, next_task, 0);
Hillf Danton311e8002011-06-16 21:55:20 -04001858 ret = 1;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001859
Kirill Tkhai88751252014-06-29 00:03:57 +04001860 resched_curr(lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001861
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001862 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001863
Steven Rostedte8fa1362008-01-25 21:08:05 +01001864out:
1865 put_task_struct(next_task);
1866
Hillf Danton311e8002011-06-16 21:55:20 -04001867 return ret;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001868}
1869
Steven Rostedte8fa1362008-01-25 21:08:05 +01001870static void push_rt_tasks(struct rq *rq)
1871{
1872 /* push_rt_task will return true if it moved an RT */
1873 while (push_rt_task(rq))
1874 ;
1875}
1876
Steven Rostedtb6366f02015-03-18 14:49:46 -04001877#ifdef HAVE_RT_PUSH_IPI
1878/*
1879 * The search for the next cpu always starts at rq->cpu and ends
1880 * when we reach rq->cpu again. It will never return rq->cpu.
1881 * This returns the next cpu to check, or nr_cpu_ids if the loop
1882 * is complete.
1883 *
1884 * rq->rt.push_cpu holds the last cpu returned by this function,
1885 * or if this is the first instance, it must hold rq->cpu.
1886 */
1887static int rto_next_cpu(struct rq *rq)
1888{
1889 int prev_cpu = rq->rt.push_cpu;
1890 int cpu;
1891
1892 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1893
1894 /*
1895 * If the previous cpu is less than the rq's CPU, then it already
1896 * passed the end of the mask, and has started from the beginning.
1897 * We end if the next CPU is greater or equal to rq's CPU.
1898 */
1899 if (prev_cpu < rq->cpu) {
1900 if (cpu >= rq->cpu)
1901 return nr_cpu_ids;
1902
1903 } else if (cpu >= nr_cpu_ids) {
1904 /*
1905 * We passed the end of the mask, start at the beginning.
1906 * If the result is greater or equal to the rq's CPU, then
1907 * the loop is finished.
1908 */
1909 cpu = cpumask_first(rq->rd->rto_mask);
1910 if (cpu >= rq->cpu)
1911 return nr_cpu_ids;
1912 }
1913 rq->rt.push_cpu = cpu;
1914
1915 /* Return cpu to let the caller know if the loop is finished or not */
1916 return cpu;
1917}
1918
1919static int find_next_push_cpu(struct rq *rq)
1920{
1921 struct rq *next_rq;
1922 int cpu;
1923
1924 while (1) {
1925 cpu = rto_next_cpu(rq);
1926 if (cpu >= nr_cpu_ids)
1927 break;
1928 next_rq = cpu_rq(cpu);
1929
1930 /* Make sure the next rq can push to this rq */
1931 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1932 break;
1933 }
1934
1935 return cpu;
1936}
1937
1938#define RT_PUSH_IPI_EXECUTING 1
1939#define RT_PUSH_IPI_RESTART 2
1940
Steven Rostedt (VMware)3e777f92017-02-28 15:50:30 -05001941/*
1942 * When a high priority task schedules out from a CPU and a lower priority
1943 * task is scheduled in, a check is made to see if there's any RT tasks
1944 * on other CPUs that are waiting to run because a higher priority RT task
1945 * is currently running on its CPU. In this case, the CPU with multiple RT
1946 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1947 * up that may be able to run one of its non-running queued RT tasks.
1948 *
1949 * On large CPU boxes, there's the case that several CPUs could schedule
1950 * a lower priority task at the same time, in which case it will look for
1951 * any overloaded CPUs that it could pull a task from. To do this, the runqueue
1952 * lock must be taken from that overloaded CPU. Having 10s of CPUs all fighting
1953 * for a single overloaded CPU's runqueue lock can produce a large latency.
1954 * (This has actually been observed on large boxes running cyclictest).
1955 * Instead of taking the runqueue lock of the overloaded CPU, each of the
1956 * CPUs that scheduled a lower priority task simply sends an IPI to the
1957 * overloaded CPU. An IPI is much cheaper than taking an runqueue lock with
1958 * lots of contention. The overloaded CPU will look to push its non-running
1959 * RT task off, and if it does, it can then ignore the other IPIs coming
1960 * in, and just pass those IPIs off to any other overloaded CPU.
1961 *
1962 * When a CPU schedules a lower priority task, it only sends an IPI to
1963 * the "next" CPU that has overloaded RT tasks. This prevents IPI storms,
1964 * as having 10 CPUs scheduling lower priority tasks and 10 CPUs with
1965 * RT overloaded tasks, would cause 100 IPIs to go out at once.
1966 *
1967 * The overloaded RT CPU, when receiving an IPI, will try to push off its
1968 * overloaded RT tasks and then send an IPI to the next CPU that has
1969 * overloaded RT tasks. This stops when all CPUs with overloaded RT tasks
1970 * have completed. Just because a CPU may have pushed off its own overloaded
1971 * RT task does not mean it should stop sending the IPI around to other
1972 * overloaded CPUs. There may be another RT task waiting to run on one of
1973 * those CPUs that are of higher priority than the one that was just
1974 * pushed.
1975 *
1976 * An optimization that could possibly be made is to make a CPU array similar
1977 * to the cpupri array mask of all running RT tasks, but for the overloaded
1978 * case, then the IPI could be sent to only the CPU with the highest priority
1979 * RT task waiting, and that CPU could send off further IPIs to the CPU with
1980 * the next highest waiting task. Since the overloaded case is much less likely
1981 * to happen, the complexity of this implementation may not be worth it.
1982 * Instead, just send an IPI around to all overloaded CPUs.
1983 *
1984 * The rq->rt.push_flags holds the status of the IPI that is going around.
1985 * A run queue can only send out a single IPI at a time. The possible flags
1986 * for rq->rt.push_flags are:
1987 *
1988 * (None or zero): No IPI is going around for the current rq
1989 * RT_PUSH_IPI_EXECUTING: An IPI for the rq is being passed around
1990 * RT_PUSH_IPI_RESTART: The priority of the running task for the rq
1991 * has changed, and the IPI should restart
1992 * circulating the overloaded CPUs again.
1993 *
1994 * rq->rt.push_cpu contains the CPU that is being sent the IPI. It is updated
1995 * before sending to the next CPU.
1996 *
1997 * Instead of having all CPUs that schedule a lower priority task send
1998 * an IPI to the same "first" CPU in the RT overload mask, they send it
1999 * to the next overloaded CPU after their own CPU. This helps distribute
2000 * the work when there's more than one overloaded CPU and multiple CPUs
2001 * scheduling in lower priority tasks.
2002 *
2003 * When a rq schedules a lower priority task than what was currently
2004 * running, the next CPU with overloaded RT tasks is examined first.
2005 * That is, if CPU 1 and 5 are overloaded, and CPU 3 schedules a lower
2006 * priority task, it will send an IPI first to CPU 5, then CPU 5 will
2007 * send to CPU 1 if it is still overloaded. CPU 1 will clear the
2008 * rq->rt.push_flags if RT_PUSH_IPI_RESTART is not set.
2009 *
2010 * The first CPU to notice IPI_RESTART is set, will clear that flag and then
2011 * send an IPI to the next overloaded CPU after the rq->cpu and not the next
2012 * CPU after push_cpu. That is, if CPU 1, 4 and 5 are overloaded when CPU 3
2013 * schedules a lower priority task, and the IPI_RESTART gets set while the
2014 * handling is being done on CPU 5, it will clear the flag and send it back to
2015 * CPU 4 instead of CPU 1.
2016 *
2017 * Note, the above logic can be disabled by turning off the sched_feature
2018 * RT_PUSH_IPI. Then the rq lock of the overloaded CPU will simply be
2019 * taken by the CPU requesting a pull and the waiting RT task will be pulled
2020 * by that CPU. This may be fine for machines with few CPUs.
2021 */
Steven Rostedtb6366f02015-03-18 14:49:46 -04002022static void tell_cpu_to_push(struct rq *rq)
2023{
2024 int cpu;
2025
2026 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2027 raw_spin_lock(&rq->rt.push_lock);
2028 /* Make sure it's still executing */
2029 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2030 /*
2031 * Tell the IPI to restart the loop as things have
2032 * changed since it started.
2033 */
2034 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
2035 raw_spin_unlock(&rq->rt.push_lock);
2036 return;
2037 }
2038 raw_spin_unlock(&rq->rt.push_lock);
2039 }
2040
2041 /* When here, there's no IPI going around */
2042
2043 rq->rt.push_cpu = rq->cpu;
2044 cpu = find_next_push_cpu(rq);
2045 if (cpu >= nr_cpu_ids)
2046 return;
2047
2048 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
2049
2050 irq_work_queue_on(&rq->rt.push_work, cpu);
2051}
2052
2053/* Called from hardirq context */
2054static void try_to_push_tasks(void *arg)
2055{
2056 struct rt_rq *rt_rq = arg;
2057 struct rq *rq, *src_rq;
2058 int this_cpu;
2059 int cpu;
2060
2061 this_cpu = rt_rq->push_cpu;
2062
2063 /* Paranoid check */
2064 BUG_ON(this_cpu != smp_processor_id());
2065
2066 rq = cpu_rq(this_cpu);
2067 src_rq = rq_of_rt_rq(rt_rq);
2068
2069again:
2070 if (has_pushable_tasks(rq)) {
2071 raw_spin_lock(&rq->lock);
2072 push_rt_task(rq);
2073 raw_spin_unlock(&rq->lock);
2074 }
2075
2076 /* Pass the IPI to the next rt overloaded queue */
2077 raw_spin_lock(&rt_rq->push_lock);
2078 /*
2079 * If the source queue changed since the IPI went out,
2080 * we need to restart the search from that CPU again.
2081 */
2082 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
2083 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
2084 rt_rq->push_cpu = src_rq->cpu;
2085 }
2086
2087 cpu = find_next_push_cpu(src_rq);
2088
2089 if (cpu >= nr_cpu_ids)
2090 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
2091 raw_spin_unlock(&rt_rq->push_lock);
2092
2093 if (cpu >= nr_cpu_ids)
2094 return;
2095
2096 /*
2097 * It is possible that a restart caused this CPU to be
2098 * chosen again. Don't bother with an IPI, just see if we
2099 * have more to push.
2100 */
2101 if (unlikely(cpu == rq->cpu))
2102 goto again;
2103
2104 /* Try the next RT overloaded CPU */
2105 irq_work_queue_on(&rt_rq->push_work, cpu);
2106}
2107
2108static void push_irq_work_func(struct irq_work *work)
2109{
2110 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2111
2112 try_to_push_tasks(rt_rq);
2113}
2114#endif /* HAVE_RT_PUSH_IPI */
2115
Peter Zijlstra8046d682015-06-11 14:46:40 +02002116static void pull_rt_task(struct rq *this_rq)
Steven Rostedtf65eda42008-01-25 21:08:07 +01002117{
Peter Zijlstra8046d682015-06-11 14:46:40 +02002118 int this_cpu = this_rq->cpu, cpu;
2119 bool resched = false;
Gregory Haskinsa8728942008-12-29 09:39:49 -05002120 struct task_struct *p;
Steven Rostedtf65eda42008-01-25 21:08:07 +01002121 struct rq *src_rq;
Steven Rostedtf65eda42008-01-25 21:08:07 +01002122
Gregory Haskins637f5082008-01-25 21:08:18 +01002123 if (likely(!rt_overloaded(this_rq)))
Peter Zijlstra8046d682015-06-11 14:46:40 +02002124 return;
Steven Rostedtf65eda42008-01-25 21:08:07 +01002125
Peter Zijlstra7c3f2ab2013-10-15 12:35:07 +02002126 /*
2127 * Match the barrier from rt_set_overloaded; this guarantees that if we
2128 * see overloaded we must also see the rto_mask bit.
2129 */
2130 smp_rmb();
2131
Steven Rostedtb6366f02015-03-18 14:49:46 -04002132#ifdef HAVE_RT_PUSH_IPI
2133 if (sched_feat(RT_PUSH_IPI)) {
2134 tell_cpu_to_push(this_rq);
Peter Zijlstra8046d682015-06-11 14:46:40 +02002135 return;
Steven Rostedtb6366f02015-03-18 14:49:46 -04002136 }
2137#endif
2138
Rusty Russellc6c49272008-11-25 02:35:05 +10302139 for_each_cpu(cpu, this_rq->rd->rto_mask) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01002140 if (this_cpu == cpu)
2141 continue;
2142
2143 src_rq = cpu_rq(cpu);
Gregory Haskins74ab8e42008-12-29 09:39:50 -05002144
2145 /*
2146 * Don't bother taking the src_rq->lock if the next highest
2147 * task is known to be lower-priority than our current task.
2148 * This may look racy, but if this value is about to go
2149 * logically higher, the src_rq will push this task away.
2150 * And if its going logically lower, we do not care
2151 */
2152 if (src_rq->rt.highest_prio.next >=
2153 this_rq->rt.highest_prio.curr)
2154 continue;
2155
Steven Rostedtf65eda42008-01-25 21:08:07 +01002156 /*
2157 * We can potentially drop this_rq's lock in
2158 * double_lock_balance, and another CPU could
Gregory Haskinsa8728942008-12-29 09:39:49 -05002159 * alter this_rq
Steven Rostedtf65eda42008-01-25 21:08:07 +01002160 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05002161 double_lock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01002162
2163 /*
Kirill Tkhaie23ee742013-06-07 15:37:43 -04002164 * We can pull only a task, which is pushable
2165 * on its rq, and no others.
Steven Rostedtf65eda42008-01-25 21:08:07 +01002166 */
Kirill Tkhaie23ee742013-06-07 15:37:43 -04002167 p = pick_highest_pushable_task(src_rq, this_cpu);
Steven Rostedtf65eda42008-01-25 21:08:07 +01002168
2169 /*
2170 * Do we have an RT task that preempts
2171 * the to-be-scheduled task?
2172 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05002173 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01002174 WARN_ON(p == src_rq->curr);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002175 WARN_ON(!task_on_rq_queued(p));
Steven Rostedtf65eda42008-01-25 21:08:07 +01002176
2177 /*
2178 * There's a chance that p is higher in priority
2179 * than what's currently running on its cpu.
2180 * This is just that p is wakeing up and hasn't
2181 * had a chance to schedule. We only pull
2182 * p if it is lower in priority than the
Gregory Haskinsa8728942008-12-29 09:39:49 -05002183 * current task on the run queue
Steven Rostedtf65eda42008-01-25 21:08:07 +01002184 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05002185 if (p->prio < src_rq->curr->prio)
Mike Galbraith614ee1f2008-01-25 21:08:30 +01002186 goto skip;
Steven Rostedtf65eda42008-01-25 21:08:07 +01002187
Peter Zijlstra8046d682015-06-11 14:46:40 +02002188 resched = true;
Steven Rostedtf65eda42008-01-25 21:08:07 +01002189
2190 deactivate_task(src_rq, p, 0);
2191 set_task_cpu(p, this_cpu);
2192 activate_task(this_rq, p, 0);
2193 /*
2194 * We continue with the search, just in
2195 * case there's an even higher prio task
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002196 * in another runqueue. (low likelihood
Steven Rostedtf65eda42008-01-25 21:08:07 +01002197 * but possible)
Steven Rostedtf65eda42008-01-25 21:08:07 +01002198 */
Steven Rostedtf65eda42008-01-25 21:08:07 +01002199 }
Peter Zijlstra49246272010-10-17 21:46:10 +02002200skip:
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02002201 double_unlock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01002202 }
2203
Peter Zijlstra8046d682015-06-11 14:46:40 +02002204 if (resched)
2205 resched_curr(this_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01002206}
2207
Gregory Haskins8ae121a2008-04-23 07:13:29 -04002208/*
2209 * If we are not running and we are not going to reschedule soon, we should
2210 * try to push tasks away now
2211 */
Peter Zijlstraefbbd052009-12-16 18:04:40 +01002212static void task_woken_rt(struct rq *rq, struct task_struct *p)
Steven Rostedt4642daf2008-01-25 21:08:07 +01002213{
Steven Rostedt9a897c52008-01-25 21:08:22 +01002214 if (!task_running(rq, p) &&
Gregory Haskins8ae121a2008-04-23 07:13:29 -04002215 !test_tsk_need_resched(rq->curr) &&
Ingo Molnar4b53a342017-02-05 15:41:03 +01002216 p->nr_cpus_allowed > 1 &&
Juri Lelli1baca4c2013-11-07 14:43:38 +01002217 (dl_task(rq->curr) || rt_task(rq->curr)) &&
Ingo Molnar4b53a342017-02-05 15:41:03 +01002218 (rq->curr->nr_cpus_allowed < 2 ||
Shawn Bohrer3be209a2011-09-12 09:28:04 -05002219 rq->curr->prio <= p->prio))
Steven Rostedt4642daf2008-01-25 21:08:07 +01002220 push_rt_tasks(rq);
2221}
2222
Ingo Molnarbdd7c812008-01-25 21:08:18 +01002223/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04002224static void rq_online_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01002225{
2226 if (rq->rt.overloaded)
2227 rt_set_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02002228
Peter Zijlstra7def2be2008-06-05 14:49:58 +02002229 __enable_runtime(rq);
2230
Gregory Haskinse864c492008-12-29 09:39:49 -05002231 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01002232}
2233
2234/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04002235static void rq_offline_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01002236{
2237 if (rq->rt.overloaded)
2238 rt_clear_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02002239
Peter Zijlstra7def2be2008-06-05 14:49:58 +02002240 __disable_runtime(rq);
2241
Gregory Haskins6e0534f2008-05-12 21:21:01 +02002242 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01002243}
Steven Rostedtcb469842008-01-25 21:08:22 +01002244
2245/*
2246 * When switch from the rt queue, we bring ourselves to a position
2247 * that we might want to pull RT tasks from other runqueues.
2248 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002249static void switched_from_rt(struct rq *rq, struct task_struct *p)
Steven Rostedtcb469842008-01-25 21:08:22 +01002250{
2251 /*
2252 * If there are other RT tasks then we will reschedule
2253 * and the scheduling of the other RT tasks will handle
2254 * the balancing. But if we are the last RT task
2255 * we may need to handle the pulling of RT tasks
2256 * now.
2257 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002258 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
Kirill Tkhai1158ddb2012-11-23 00:02:15 +04002259 return;
2260
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002261 queue_pull_task(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01002262}
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10302263
Li Zefan11c785b2014-02-08 14:17:45 +08002264void __init init_sched_rt_class(void)
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10302265{
2266 unsigned int i;
2267
Peter Zijlstra029632f2011-10-25 10:00:11 +02002268 for_each_possible_cpu(i) {
Yinghai Lueaa95842009-06-06 14:51:36 -07002269 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
Mike Travis6ca09df2008-12-31 18:08:45 -08002270 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra029632f2011-10-25 10:00:11 +02002271 }
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10302272}
Steven Rostedte8fa1362008-01-25 21:08:05 +01002273#endif /* CONFIG_SMP */
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002274
Steven Rostedtcb469842008-01-25 21:08:22 +01002275/*
2276 * When switching a task to RT, we may overload the runqueue
2277 * with RT tasks. In this case we try to push them off to
2278 * other runqueues.
2279 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002280static void switched_to_rt(struct rq *rq, struct task_struct *p)
Steven Rostedtcb469842008-01-25 21:08:22 +01002281{
Steven Rostedtcb469842008-01-25 21:08:22 +01002282 /*
2283 * If we are already running, then there's nothing
2284 * that needs to be done. But if we are not running
2285 * we may need to preempt the current running task.
2286 * If that current running task is also an RT task
2287 * then see if we can move to another run queue.
2288 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002289 if (task_on_rq_queued(p) && rq->curr != p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01002290#ifdef CONFIG_SMP
Ingo Molnar4b53a342017-02-05 15:41:03 +01002291 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002292 queue_push_tasks(rq);
Sebastian Andrzej Siewior619bd4a2017-01-24 15:40:06 +01002293#endif /* CONFIG_SMP */
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002294 if (p->prio < rq->curr->prio)
Kirill Tkhai88751252014-06-29 00:03:57 +04002295 resched_curr(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01002296 }
2297}
2298
2299/*
2300 * Priority of the task has changed. This may cause
2301 * us to initiate a push or pull.
2302 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002303static void
2304prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01002305{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002306 if (!task_on_rq_queued(p))
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002307 return;
2308
2309 if (rq->curr == p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01002310#ifdef CONFIG_SMP
2311 /*
2312 * If our priority decreases while running, we
2313 * may need to pull tasks to this runqueue.
2314 */
2315 if (oldprio < p->prio)
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002316 queue_pull_task(rq);
2317
Steven Rostedtcb469842008-01-25 21:08:22 +01002318 /*
2319 * If there's a higher priority task waiting to run
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002320 * then reschedule.
Steven Rostedtcb469842008-01-25 21:08:22 +01002321 */
Peter Zijlstrafd7a4be2015-06-11 14:46:41 +02002322 if (p->prio > rq->rt.highest_prio.curr)
Kirill Tkhai88751252014-06-29 00:03:57 +04002323 resched_curr(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01002324#else
2325 /* For UP simply resched on drop of prio */
2326 if (oldprio < p->prio)
Kirill Tkhai88751252014-06-29 00:03:57 +04002327 resched_curr(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01002328#endif /* CONFIG_SMP */
2329 } else {
2330 /*
2331 * This task is not running, but if it is
2332 * greater than the current running task
2333 * then reschedule.
2334 */
2335 if (p->prio < rq->curr->prio)
Kirill Tkhai88751252014-06-29 00:03:57 +04002336 resched_curr(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01002337 }
2338}
2339
Nicolas Pitreb18b6a92017-01-21 00:09:08 -05002340#ifdef CONFIG_POSIX_TIMERS
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01002341static void watchdog(struct rq *rq, struct task_struct *p)
2342{
2343 unsigned long soft, hard;
2344
Jiri Slaby78d7d402010-03-05 13:42:54 -08002345 /* max may change after cur was read, this will be fixed next tick */
2346 soft = task_rlimit(p, RLIMIT_RTTIME);
2347 hard = task_rlimit_max(p, RLIMIT_RTTIME);
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01002348
2349 if (soft != RLIM_INFINITY) {
2350 unsigned long next;
2351
Ying Xue57d2aa02012-07-17 15:03:43 +08002352 if (p->rt.watchdog_stamp != jiffies) {
2353 p->rt.timeout++;
2354 p->rt.watchdog_stamp = jiffies;
2355 }
2356
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01002357 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
Peter Zijlstra5a52dd52008-01-25 21:08:32 +01002358 if (p->rt.timeout > next)
Frank Mayharf06febc2008-09-12 09:54:39 -07002359 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01002360 }
2361}
Nicolas Pitreb18b6a92017-01-21 00:09:08 -05002362#else
2363static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2364#endif
Steven Rostedtcb469842008-01-25 21:08:22 +01002365
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01002366static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002367{
Colin Cross454c7992012-05-16 21:34:23 -07002368 struct sched_rt_entity *rt_se = &p->rt;
2369
Peter Zijlstra67e2be02007-12-20 15:01:17 +01002370 update_curr_rt(rq);
2371
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01002372 watchdog(rq, p);
2373
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002374 /*
2375 * RR tasks need a special form of timeslice management.
2376 * FIFO tasks have no timeslices.
2377 */
2378 if (p->policy != SCHED_RR)
2379 return;
2380
Peter Zijlstrafa717062008-01-25 21:08:27 +01002381 if (--p->rt.time_slice)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002382 return;
2383
Clark Williamsce0dbbb2013-02-07 09:47:04 -06002384 p->rt.time_slice = sched_rr_timeslice;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002385
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02002386 /*
Li Bine9aa39b2013-10-21 20:15:43 +08002387 * Requeue to the end of queue if we (and all of our ancestors) are not
2388 * the only element on the queue
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02002389 */
Colin Cross454c7992012-05-16 21:34:23 -07002390 for_each_sched_rt_entity(rt_se) {
2391 if (rt_se->run_list.prev != rt_se->run_list.next) {
2392 requeue_task_rt(rq, p, 0);
Kirill Tkhai8aa6f0e2014-09-22 22:36:43 +04002393 resched_curr(rq);
Colin Cross454c7992012-05-16 21:34:23 -07002394 return;
2395 }
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02002396 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002397}
2398
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002399static void set_curr_task_rt(struct rq *rq)
2400{
2401 struct task_struct *p = rq->curr;
2402
Frederic Weisbecker78becc22013-04-12 01:51:02 +02002403 p->se.exec_start = rq_clock_task(rq);
Gregory Haskins917b6272008-12-29 09:39:53 -05002404
2405 /* The running task is never eligible for pushing */
2406 dequeue_pushable_task(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002407}
2408
H Hartley Sweeten6d686f42010-01-13 20:21:52 -07002409static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
Peter Williams0d721ce2009-09-21 01:31:53 +00002410{
2411 /*
2412 * Time slice is 0 for SCHED_FIFO tasks
2413 */
2414 if (task->policy == SCHED_RR)
Clark Williamsce0dbbb2013-02-07 09:47:04 -06002415 return sched_rr_timeslice;
Peter Williams0d721ce2009-09-21 01:31:53 +00002416 else
2417 return 0;
2418}
2419
Peter Zijlstra029632f2011-10-25 10:00:11 +02002420const struct sched_class rt_sched_class = {
Ingo Molnar5522d5d2007-10-15 17:00:12 +02002421 .next = &fair_sched_class,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002422 .enqueue_task = enqueue_task_rt,
2423 .dequeue_task = dequeue_task_rt,
2424 .yield_task = yield_task_rt,
2425
2426 .check_preempt_curr = check_preempt_curr_rt,
2427
2428 .pick_next_task = pick_next_task_rt,
2429 .put_prev_task = put_prev_task_rt,
2430
Peter Williams681f3e62007-10-24 18:23:51 +02002431#ifdef CONFIG_SMP
Li Zefan4ce72a22008-10-22 15:25:26 +08002432 .select_task_rq = select_task_rq_rt,
2433
Peter Zijlstra6c370672015-05-15 17:43:36 +02002434 .set_cpus_allowed = set_cpus_allowed_common,
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04002435 .rq_online = rq_online_rt,
2436 .rq_offline = rq_offline_rt,
Peter Zijlstraefbbd052009-12-16 18:04:40 +01002437 .task_woken = task_woken_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01002438 .switched_from = switched_from_rt,
Peter Williams681f3e62007-10-24 18:23:51 +02002439#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002440
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02002441 .set_curr_task = set_curr_task_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002442 .task_tick = task_tick_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01002443
Peter Williams0d721ce2009-09-21 01:31:53 +00002444 .get_rr_interval = get_rr_interval_rt,
2445
Steven Rostedtcb469842008-01-25 21:08:22 +01002446 .prio_changed = prio_changed_rt,
2447 .switched_to = switched_to_rt,
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01002448
2449 .update_curr = update_curr_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02002450};
Peter Zijlstraada18de2008-06-19 14:22:24 +02002451
Nicolas Pitre8887cd92017-06-21 14:22:02 -04002452#ifdef CONFIG_RT_GROUP_SCHED
2453/*
2454 * Ensure that the real time constraints are schedulable.
2455 */
2456static DEFINE_MUTEX(rt_constraints_mutex);
2457
2458/* Must be called with tasklist_lock held */
2459static inline int tg_has_rt_tasks(struct task_group *tg)
2460{
2461 struct task_struct *g, *p;
2462
2463 /*
2464 * Autogroups do not have RT tasks; see autogroup_create().
2465 */
2466 if (task_group_is_autogroup(tg))
2467 return 0;
2468
2469 for_each_process_thread(g, p) {
2470 if (rt_task(p) && task_group(p) == tg)
2471 return 1;
2472 }
2473
2474 return 0;
2475}
2476
2477struct rt_schedulable_data {
2478 struct task_group *tg;
2479 u64 rt_period;
2480 u64 rt_runtime;
2481};
2482
2483static int tg_rt_schedulable(struct task_group *tg, void *data)
2484{
2485 struct rt_schedulable_data *d = data;
2486 struct task_group *child;
2487 unsigned long total, sum = 0;
2488 u64 period, runtime;
2489
2490 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2491 runtime = tg->rt_bandwidth.rt_runtime;
2492
2493 if (tg == d->tg) {
2494 period = d->rt_period;
2495 runtime = d->rt_runtime;
2496 }
2497
2498 /*
2499 * Cannot have more runtime than the period.
2500 */
2501 if (runtime > period && runtime != RUNTIME_INF)
2502 return -EINVAL;
2503
2504 /*
2505 * Ensure we don't starve existing RT tasks.
2506 */
2507 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2508 return -EBUSY;
2509
2510 total = to_ratio(period, runtime);
2511
2512 /*
2513 * Nobody can have more than the global setting allows.
2514 */
2515 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2516 return -EINVAL;
2517
2518 /*
2519 * The sum of our children's runtime should not exceed our own.
2520 */
2521 list_for_each_entry_rcu(child, &tg->children, siblings) {
2522 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2523 runtime = child->rt_bandwidth.rt_runtime;
2524
2525 if (child == d->tg) {
2526 period = d->rt_period;
2527 runtime = d->rt_runtime;
2528 }
2529
2530 sum += to_ratio(period, runtime);
2531 }
2532
2533 if (sum > total)
2534 return -EINVAL;
2535
2536 return 0;
2537}
2538
2539static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2540{
2541 int ret;
2542
2543 struct rt_schedulable_data data = {
2544 .tg = tg,
2545 .rt_period = period,
2546 .rt_runtime = runtime,
2547 };
2548
2549 rcu_read_lock();
2550 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2551 rcu_read_unlock();
2552
2553 return ret;
2554}
2555
2556static int tg_set_rt_bandwidth(struct task_group *tg,
2557 u64 rt_period, u64 rt_runtime)
2558{
2559 int i, err = 0;
2560
2561 /*
2562 * Disallowing the root group RT runtime is BAD, it would disallow the
2563 * kernel creating (and or operating) RT threads.
2564 */
2565 if (tg == &root_task_group && rt_runtime == 0)
2566 return -EINVAL;
2567
2568 /* No period doesn't make any sense. */
2569 if (rt_period == 0)
2570 return -EINVAL;
2571
2572 mutex_lock(&rt_constraints_mutex);
2573 read_lock(&tasklist_lock);
2574 err = __rt_schedulable(tg, rt_period, rt_runtime);
2575 if (err)
2576 goto unlock;
2577
2578 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2579 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2580 tg->rt_bandwidth.rt_runtime = rt_runtime;
2581
2582 for_each_possible_cpu(i) {
2583 struct rt_rq *rt_rq = tg->rt_rq[i];
2584
2585 raw_spin_lock(&rt_rq->rt_runtime_lock);
2586 rt_rq->rt_runtime = rt_runtime;
2587 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2588 }
2589 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2590unlock:
2591 read_unlock(&tasklist_lock);
2592 mutex_unlock(&rt_constraints_mutex);
2593
2594 return err;
2595}
2596
2597int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2598{
2599 u64 rt_runtime, rt_period;
2600
2601 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2602 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2603 if (rt_runtime_us < 0)
2604 rt_runtime = RUNTIME_INF;
2605
2606 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2607}
2608
2609long sched_group_rt_runtime(struct task_group *tg)
2610{
2611 u64 rt_runtime_us;
2612
2613 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2614 return -1;
2615
2616 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2617 do_div(rt_runtime_us, NSEC_PER_USEC);
2618 return rt_runtime_us;
2619}
2620
2621int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2622{
2623 u64 rt_runtime, rt_period;
2624
2625 rt_period = rt_period_us * NSEC_PER_USEC;
2626 rt_runtime = tg->rt_bandwidth.rt_runtime;
2627
2628 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2629}
2630
2631long sched_group_rt_period(struct task_group *tg)
2632{
2633 u64 rt_period_us;
2634
2635 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2636 do_div(rt_period_us, NSEC_PER_USEC);
2637 return rt_period_us;
2638}
2639
2640static int sched_rt_global_constraints(void)
2641{
2642 int ret = 0;
2643
2644 mutex_lock(&rt_constraints_mutex);
2645 read_lock(&tasklist_lock);
2646 ret = __rt_schedulable(NULL, 0, 0);
2647 read_unlock(&tasklist_lock);
2648 mutex_unlock(&rt_constraints_mutex);
2649
2650 return ret;
2651}
2652
2653int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2654{
2655 /* Don't accept realtime tasks when there is no way for them to run */
2656 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2657 return 0;
2658
2659 return 1;
2660}
2661
2662#else /* !CONFIG_RT_GROUP_SCHED */
2663static int sched_rt_global_constraints(void)
2664{
2665 unsigned long flags;
2666 int i;
2667
2668 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2669 for_each_possible_cpu(i) {
2670 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2671
2672 raw_spin_lock(&rt_rq->rt_runtime_lock);
2673 rt_rq->rt_runtime = global_rt_runtime();
2674 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2675 }
2676 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2677
2678 return 0;
2679}
2680#endif /* CONFIG_RT_GROUP_SCHED */
2681
2682static int sched_rt_global_validate(void)
2683{
2684 if (sysctl_sched_rt_period <= 0)
2685 return -EINVAL;
2686
2687 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2688 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2689 return -EINVAL;
2690
2691 return 0;
2692}
2693
2694static void sched_rt_do_global(void)
2695{
2696 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2697 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2698}
2699
2700int sched_rt_handler(struct ctl_table *table, int write,
2701 void __user *buffer, size_t *lenp,
2702 loff_t *ppos)
2703{
2704 int old_period, old_runtime;
2705 static DEFINE_MUTEX(mutex);
2706 int ret;
2707
2708 mutex_lock(&mutex);
2709 old_period = sysctl_sched_rt_period;
2710 old_runtime = sysctl_sched_rt_runtime;
2711
2712 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2713
2714 if (!ret && write) {
2715 ret = sched_rt_global_validate();
2716 if (ret)
2717 goto undo;
2718
2719 ret = sched_dl_global_validate();
2720 if (ret)
2721 goto undo;
2722
2723 ret = sched_rt_global_constraints();
2724 if (ret)
2725 goto undo;
2726
2727 sched_rt_do_global();
2728 sched_dl_do_global();
2729 }
2730 if (0) {
2731undo:
2732 sysctl_sched_rt_period = old_period;
2733 sysctl_sched_rt_runtime = old_runtime;
2734 }
2735 mutex_unlock(&mutex);
2736
2737 return ret;
2738}
2739
2740int sched_rr_handler(struct ctl_table *table, int write,
2741 void __user *buffer, size_t *lenp,
2742 loff_t *ppos)
2743{
2744 int ret;
2745 static DEFINE_MUTEX(mutex);
2746
2747 mutex_lock(&mutex);
2748 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2749 /*
2750 * Make sure that internally we keep jiffies.
2751 * Also, writing zero resets the timeslice to default:
2752 */
2753 if (!ret && write) {
2754 sched_rr_timeslice =
2755 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2756 msecs_to_jiffies(sysctl_sched_rr_timeslice);
2757 }
2758 mutex_unlock(&mutex);
2759 return ret;
2760}
2761
Peter Zijlstraada18de2008-06-19 14:22:24 +02002762#ifdef CONFIG_SCHED_DEBUG
2763extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2764
Peter Zijlstra029632f2011-10-25 10:00:11 +02002765void print_rt_stats(struct seq_file *m, int cpu)
Peter Zijlstraada18de2008-06-19 14:22:24 +02002766{
Cheng Xuec514c42011-05-14 14:20:02 +08002767 rt_rq_iter_t iter;
Peter Zijlstraada18de2008-06-19 14:22:24 +02002768 struct rt_rq *rt_rq;
2769
2770 rcu_read_lock();
Cheng Xuec514c42011-05-14 14:20:02 +08002771 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
Peter Zijlstraada18de2008-06-19 14:22:24 +02002772 print_rt_rq(m, cpu, rt_rq);
2773 rcu_read_unlock();
2774}
Dhaval Giani55e12e52008-06-24 23:39:43 +05302775#endif /* CONFIG_SCHED_DEBUG */