blob: 4230b15fe90e91a0cb8f02ce749c5b8d28bb0511 [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
Gregory Haskins398a1532009-01-14 09:10:04 -05006static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
7{
8 return container_of(rt_se, struct task_struct, rt);
9}
10
11#ifdef CONFIG_RT_GROUP_SCHED
12
13static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
14{
15 return rt_rq->rq;
16}
17
18static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
19{
20 return rt_se->rt_rq;
21}
22
23#else /* CONFIG_RT_GROUP_SCHED */
24
25static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
26{
27 return container_of(rt_rq, struct rq, rt);
28}
29
30static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
31{
32 struct task_struct *p = rt_task_of(rt_se);
33 struct rq *rq = task_rq(p);
34
35 return &rq->rt;
36}
37
38#endif /* CONFIG_RT_GROUP_SCHED */
39
Steven Rostedt4fd29172008-01-25 21:08:06 +010040#ifdef CONFIG_SMP
Ingo Molnar84de4272008-01-25 21:08:15 +010041
Gregory Haskins637f5082008-01-25 21:08:18 +010042static inline int rt_overloaded(struct rq *rq)
Steven Rostedt4fd29172008-01-25 21:08:06 +010043{
Gregory Haskins637f5082008-01-25 21:08:18 +010044 return atomic_read(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +010045}
Ingo Molnar84de4272008-01-25 21:08:15 +010046
Steven Rostedt4fd29172008-01-25 21:08:06 +010047static inline void rt_set_overload(struct rq *rq)
48{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -040049 if (!rq->online)
50 return;
51
Rusty Russellc6c49272008-11-25 02:35:05 +103052 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +010053 /*
54 * Make sure the mask is visible before we set
55 * the overload count. That is checked to determine
56 * if we should look at the mask. It would be a shame
57 * if we looked at the mask, but the mask was not
58 * updated yet.
59 */
60 wmb();
Gregory Haskins637f5082008-01-25 21:08:18 +010061 atomic_inc(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +010062}
Ingo Molnar84de4272008-01-25 21:08:15 +010063
Steven Rostedt4fd29172008-01-25 21:08:06 +010064static inline void rt_clear_overload(struct rq *rq)
65{
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -040066 if (!rq->online)
67 return;
68
Steven Rostedt4fd29172008-01-25 21:08:06 +010069 /* the order here really doesn't matter */
Gregory Haskins637f5082008-01-25 21:08:18 +010070 atomic_dec(&rq->rd->rto_count);
Rusty Russellc6c49272008-11-25 02:35:05 +103071 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +010072}
Gregory Haskins73fe6aa2008-01-25 21:08:07 +010073
Gregory Haskins398a1532009-01-14 09:10:04 -050074static void update_rt_migration(struct rt_rq *rt_rq)
Gregory Haskins73fe6aa2008-01-25 21:08:07 +010075{
Gregory Haskins398a1532009-01-14 09:10:04 -050076 if (rt_rq->rt_nr_migratory && (rt_rq->rt_nr_running > 1)) {
77 if (!rt_rq->overloaded) {
78 rt_set_overload(rq_of_rt_rq(rt_rq));
79 rt_rq->overloaded = 1;
Gregory Haskinscdc8eb92008-01-25 21:08:23 +010080 }
Gregory Haskins398a1532009-01-14 09:10:04 -050081 } else if (rt_rq->overloaded) {
82 rt_clear_overload(rq_of_rt_rq(rt_rq));
83 rt_rq->overloaded = 0;
Gregory Haskins637f5082008-01-25 21:08:18 +010084 }
Gregory Haskins73fe6aa2008-01-25 21:08:07 +010085}
Gregory Haskins917b6272008-12-29 09:39:53 -050086
Gregory Haskins398a1532009-01-14 09:10:04 -050087static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
88{
89 if (rt_se->nr_cpus_allowed > 1)
90 rt_rq->rt_nr_migratory++;
91
92 update_rt_migration(rt_rq);
93}
94
95static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
96{
97 if (rt_se->nr_cpus_allowed > 1)
98 rt_rq->rt_nr_migratory--;
99
100 update_rt_migration(rt_rq);
101}
102
Gregory Haskins917b6272008-12-29 09:39:53 -0500103static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
104{
105 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
106 plist_node_init(&p->pushable_tasks, p->prio);
107 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
108}
109
110static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
111{
112 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
113}
114
115#else
116
Gregory Haskinsb07430a2009-01-14 08:55:39 -0500117static inline
118void enqueue_pushable_task(struct rq *rq, struct task_struct *p) {}
119static inline
120void dequeue_pushable_task(struct rq *rq, struct task_struct *p) {}
Gregory Haskins398a1532009-01-14 09:10:04 -0500121static inline
122void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
123static inline
124void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
Gregory Haskins917b6272008-12-29 09:39:53 -0500125
Steven Rostedt4fd29172008-01-25 21:08:06 +0100126#endif /* CONFIG_SMP */
127
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100128static inline int on_rt_rq(struct sched_rt_entity *rt_se)
129{
130 return !list_empty(&rt_se->run_list);
131}
132
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100133#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100134
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100135static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100136{
137 if (!rt_rq->tg)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100138 return RUNTIME_INF;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100139
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200140 return rt_rq->rt_runtime;
141}
142
143static inline u64 sched_rt_period(struct rt_rq *rt_rq)
144{
145 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100146}
147
148#define for_each_leaf_rt_rq(rt_rq, rq) \
Bharata B Rao80f40ee2008-12-15 11:56:48 +0530149 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100150
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100151#define for_each_sched_rt_entity(rt_se) \
152 for (; rt_se; rt_se = rt_se->parent)
153
154static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
155{
156 return rt_se->my_q;
157}
158
159static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
160static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
161
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100162static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100163{
Dario Faggiolif6121f42008-10-03 17:40:46 +0200164 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100165 struct sched_rt_entity *rt_se = rt_rq->rt_se;
166
Dario Faggiolif6121f42008-10-03 17:40:46 +0200167 if (rt_rq->rt_nr_running) {
168 if (rt_se && !on_rt_rq(rt_se))
169 enqueue_rt_entity(rt_se);
Gregory Haskinse864c492008-12-29 09:39:49 -0500170 if (rt_rq->highest_prio.curr < curr->prio)
Peter Zijlstra10203872008-01-25 21:08:32 +0100171 resched_task(curr);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100172 }
173}
174
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100175static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100176{
177 struct sched_rt_entity *rt_se = rt_rq->rt_se;
178
179 if (rt_se && on_rt_rq(rt_se))
180 dequeue_rt_entity(rt_se);
181}
182
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100183static inline int rt_rq_throttled(struct rt_rq *rt_rq)
184{
185 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
186}
187
188static int rt_se_boosted(struct sched_rt_entity *rt_se)
189{
190 struct rt_rq *rt_rq = group_rt_rq(rt_se);
191 struct task_struct *p;
192
193 if (rt_rq)
194 return !!rt_rq->rt_nr_boosted;
195
196 p = rt_task_of(rt_se);
197 return p->prio != p->normal_prio;
198}
199
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200200#ifdef CONFIG_SMP
Rusty Russellc6c49272008-11-25 02:35:05 +1030201static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200202{
203 return cpu_rq(smp_processor_id())->rd->span;
204}
205#else
Rusty Russellc6c49272008-11-25 02:35:05 +1030206static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200207{
Rusty Russellc6c49272008-11-25 02:35:05 +1030208 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200209}
210#endif
211
212static inline
213struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
214{
215 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
216}
217
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200218static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
219{
220 return &rt_rq->tg->rt_bandwidth;
221}
222
Dhaval Giani55e12e52008-06-24 23:39:43 +0530223#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100224
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100225static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100226{
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200227 return rt_rq->rt_runtime;
228}
229
230static inline u64 sched_rt_period(struct rt_rq *rt_rq)
231{
232 return ktime_to_ns(def_rt_bandwidth.rt_period);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100233}
234
235#define for_each_leaf_rt_rq(rt_rq, rq) \
236 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
237
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100238#define for_each_sched_rt_entity(rt_se) \
239 for (; rt_se; rt_se = NULL)
240
241static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
242{
243 return NULL;
244}
245
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100246static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100247{
John Blackwoodf3ade832008-08-26 15:09:43 -0400248 if (rt_rq->rt_nr_running)
249 resched_task(rq_of_rt_rq(rt_rq)->curr);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100250}
251
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100252static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100253{
254}
255
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100256static inline int rt_rq_throttled(struct rt_rq *rt_rq)
257{
258 return rt_rq->rt_throttled;
259}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200260
Rusty Russellc6c49272008-11-25 02:35:05 +1030261static inline const struct cpumask *sched_rt_period_mask(void)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200262{
Rusty Russellc6c49272008-11-25 02:35:05 +1030263 return cpu_online_mask;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200264}
265
266static inline
267struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
268{
269 return &cpu_rq(cpu)->rt;
270}
271
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200272static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
273{
274 return &def_rt_bandwidth;
275}
276
Dhaval Giani55e12e52008-06-24 23:39:43 +0530277#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100278
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200279#ifdef CONFIG_SMP
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200280/*
281 * We ran out of runtime, see if we can borrow some from our neighbours.
282 */
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200283static int do_balance_runtime(struct rt_rq *rt_rq)
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200284{
285 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
286 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
287 int i, weight, more = 0;
288 u64 rt_period;
289
Rusty Russellc6c49272008-11-25 02:35:05 +1030290 weight = cpumask_weight(rd->span);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200291
292 spin_lock(&rt_b->rt_runtime_lock);
293 rt_period = ktime_to_ns(rt_b->rt_period);
Rusty Russellc6c49272008-11-25 02:35:05 +1030294 for_each_cpu(i, rd->span) {
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200295 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
296 s64 diff;
297
298 if (iter == rt_rq)
299 continue;
300
301 spin_lock(&iter->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200302 /*
303 * Either all rqs have inf runtime and there's nothing to steal
304 * or __disable_runtime() below sets a specific rq to inf to
305 * indicate its been disabled and disalow stealing.
306 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200307 if (iter->rt_runtime == RUNTIME_INF)
308 goto next;
309
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200310 /*
311 * From runqueues with spare time, take 1/n part of their
312 * spare time, but no more than our period.
313 */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200314 diff = iter->rt_runtime - iter->rt_time;
315 if (diff > 0) {
Peter Zijlstra58838cf2008-07-24 12:43:13 +0200316 diff = div_u64((u64)diff, weight);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200317 if (rt_rq->rt_runtime + diff > rt_period)
318 diff = rt_period - rt_rq->rt_runtime;
319 iter->rt_runtime -= diff;
320 rt_rq->rt_runtime += diff;
321 more = 1;
322 if (rt_rq->rt_runtime == rt_period) {
323 spin_unlock(&iter->rt_runtime_lock);
324 break;
325 }
326 }
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200327next:
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200328 spin_unlock(&iter->rt_runtime_lock);
329 }
330 spin_unlock(&rt_b->rt_runtime_lock);
331
332 return more;
333}
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200334
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200335/*
336 * Ensure this RQ takes back all the runtime it lend to its neighbours.
337 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200338static void __disable_runtime(struct rq *rq)
339{
340 struct root_domain *rd = rq->rd;
341 struct rt_rq *rt_rq;
342
343 if (unlikely(!scheduler_running))
344 return;
345
346 for_each_leaf_rt_rq(rt_rq, rq) {
347 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
348 s64 want;
349 int i;
350
351 spin_lock(&rt_b->rt_runtime_lock);
352 spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200353 /*
354 * Either we're all inf and nobody needs to borrow, or we're
355 * already disabled and thus have nothing to do, or we have
356 * exactly the right amount of runtime to take out.
357 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200358 if (rt_rq->rt_runtime == RUNTIME_INF ||
359 rt_rq->rt_runtime == rt_b->rt_runtime)
360 goto balanced;
361 spin_unlock(&rt_rq->rt_runtime_lock);
362
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200363 /*
364 * Calculate the difference between what we started out with
365 * and what we current have, that's the amount of runtime
366 * we lend and now have to reclaim.
367 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200368 want = rt_b->rt_runtime - rt_rq->rt_runtime;
369
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200370 /*
371 * Greedy reclaim, take back as much as we can.
372 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030373 for_each_cpu(i, rd->span) {
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200374 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
375 s64 diff;
376
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200377 /*
378 * Can't reclaim from ourselves or disabled runqueues.
379 */
Peter Zijlstraf1679d02008-08-14 15:49:00 +0200380 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200381 continue;
382
383 spin_lock(&iter->rt_runtime_lock);
384 if (want > 0) {
385 diff = min_t(s64, iter->rt_runtime, want);
386 iter->rt_runtime -= diff;
387 want -= diff;
388 } else {
389 iter->rt_runtime -= want;
390 want -= want;
391 }
392 spin_unlock(&iter->rt_runtime_lock);
393
394 if (!want)
395 break;
396 }
397
398 spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200399 /*
400 * We cannot be left wanting - that would mean some runtime
401 * leaked out of the system.
402 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200403 BUG_ON(want);
404balanced:
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200405 /*
406 * Disable all the borrow logic by pretending we have inf
407 * runtime - in which case borrowing doesn't make sense.
408 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200409 rt_rq->rt_runtime = RUNTIME_INF;
410 spin_unlock(&rt_rq->rt_runtime_lock);
411 spin_unlock(&rt_b->rt_runtime_lock);
412 }
413}
414
415static void disable_runtime(struct rq *rq)
416{
417 unsigned long flags;
418
419 spin_lock_irqsave(&rq->lock, flags);
420 __disable_runtime(rq);
421 spin_unlock_irqrestore(&rq->lock, flags);
422}
423
424static void __enable_runtime(struct rq *rq)
425{
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200426 struct rt_rq *rt_rq;
427
428 if (unlikely(!scheduler_running))
429 return;
430
Peter Zijlstra78333cd2008-09-23 15:33:43 +0200431 /*
432 * Reset each runqueue's bandwidth settings
433 */
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200434 for_each_leaf_rt_rq(rt_rq, rq) {
435 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
436
437 spin_lock(&rt_b->rt_runtime_lock);
438 spin_lock(&rt_rq->rt_runtime_lock);
439 rt_rq->rt_runtime = rt_b->rt_runtime;
440 rt_rq->rt_time = 0;
Zhang, Yanminbaf25732008-09-09 11:26:33 +0800441 rt_rq->rt_throttled = 0;
Peter Zijlstra7def2be2008-06-05 14:49:58 +0200442 spin_unlock(&rt_rq->rt_runtime_lock);
443 spin_unlock(&rt_b->rt_runtime_lock);
444 }
445}
446
447static void enable_runtime(struct rq *rq)
448{
449 unsigned long flags;
450
451 spin_lock_irqsave(&rq->lock, flags);
452 __enable_runtime(rq);
453 spin_unlock_irqrestore(&rq->lock, flags);
454}
455
Peter Zijlstraeff65492008-06-19 14:22:26 +0200456static int balance_runtime(struct rt_rq *rt_rq)
457{
458 int more = 0;
459
460 if (rt_rq->rt_time > rt_rq->rt_runtime) {
461 spin_unlock(&rt_rq->rt_runtime_lock);
462 more = do_balance_runtime(rt_rq);
463 spin_lock(&rt_rq->rt_runtime_lock);
464 }
465
466 return more;
467}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530468#else /* !CONFIG_SMP */
Peter Zijlstraeff65492008-06-19 14:22:26 +0200469static inline int balance_runtime(struct rt_rq *rt_rq)
470{
471 return 0;
472}
Dhaval Giani55e12e52008-06-24 23:39:43 +0530473#endif /* CONFIG_SMP */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100474
475static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
476{
477 int i, idle = 1;
Rusty Russellc6c49272008-11-25 02:35:05 +1030478 const struct cpumask *span;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200479
Peter Zijlstra0b148fa2008-08-19 12:33:04 +0200480 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200481 return 1;
482
483 span = sched_rt_period_mask();
Rusty Russellc6c49272008-11-25 02:35:05 +1030484 for_each_cpu(i, span) {
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200485 int enqueue = 0;
486 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
487 struct rq *rq = rq_of_rt_rq(rt_rq);
488
489 spin_lock(&rq->lock);
490 if (rt_rq->rt_time) {
491 u64 runtime;
492
493 spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraeff65492008-06-19 14:22:26 +0200494 if (rt_rq->rt_throttled)
495 balance_runtime(rt_rq);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200496 runtime = rt_rq->rt_runtime;
497 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
498 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
499 rt_rq->rt_throttled = 0;
500 enqueue = 1;
501 }
502 if (rt_rq->rt_time || rt_rq->rt_nr_running)
503 idle = 0;
504 spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstra8a8cde12008-06-19 14:22:28 +0200505 } else if (rt_rq->rt_nr_running)
506 idle = 0;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200507
508 if (enqueue)
509 sched_rt_rq_enqueue(rt_rq);
510 spin_unlock(&rq->lock);
511 }
512
513 return idle;
514}
515
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100516static inline int rt_se_prio(struct sched_rt_entity *rt_se)
517{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100518#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100519 struct rt_rq *rt_rq = group_rt_rq(rt_se);
520
521 if (rt_rq)
Gregory Haskinse864c492008-12-29 09:39:49 -0500522 return rt_rq->highest_prio.curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100523#endif
524
525 return rt_task_of(rt_se)->prio;
526}
527
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100528static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100529{
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100530 u64 runtime = sched_rt_runtime(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100531
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100532 if (rt_rq->rt_throttled)
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100533 return rt_rq_throttled(rt_rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100534
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200535 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
536 return 0;
537
Peter Zijlstrab79f3832008-06-19 14:22:25 +0200538 balance_runtime(rt_rq);
539 runtime = sched_rt_runtime(rt_rq);
540 if (runtime == RUNTIME_INF)
541 return 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200542
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100543 if (rt_rq->rt_time > runtime) {
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100544 rt_rq->rt_throttled = 1;
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100545 if (rt_rq_throttled(rt_rq)) {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100546 sched_rt_rq_dequeue(rt_rq);
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100547 return 1;
548 }
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100549 }
550
551 return 0;
552}
553
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200554/*
555 * Update the current task's runtime statistics. Skip current tasks that
556 * are not in our scheduling class.
557 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200558static void update_curr_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200559{
560 struct task_struct *curr = rq->curr;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100561 struct sched_rt_entity *rt_se = &curr->rt;
562 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200563 u64 delta_exec;
564
565 if (!task_has_rt_policy(curr))
566 return;
567
Ingo Molnard2819182007-08-09 11:16:47 +0200568 delta_exec = rq->clock - curr->se.exec_start;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200569 if (unlikely((s64)delta_exec < 0))
570 delta_exec = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +0200571
572 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200573
574 curr->se.sum_exec_runtime += delta_exec;
Frank Mayharf06febc2008-09-12 09:54:39 -0700575 account_group_exec_runtime(curr, delta_exec);
576
Ingo Molnard2819182007-08-09 11:16:47 +0200577 curr->se.exec_start = rq->clock;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100578 cpuacct_charge(curr, delta_exec);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100579
Peter Zijlstra0b148fa2008-08-19 12:33:04 +0200580 if (!rt_bandwidth_enabled())
581 return;
582
Dhaval Giani354d60c2008-04-19 19:44:59 +0200583 for_each_sched_rt_entity(rt_se) {
584 rt_rq = rt_rq_of_se(rt_se);
585
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200586 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
Dimitri Sivaniche113a742008-10-31 08:03:41 -0500587 spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200588 rt_rq->rt_time += delta_exec;
589 if (sched_rt_runtime_exceeded(rt_rq))
590 resched_task(curr);
Dimitri Sivaniche113a742008-10-31 08:03:41 -0500591 spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstracc2991c2008-08-19 12:33:03 +0200592 }
Dhaval Giani354d60c2008-04-19 19:44:59 +0200593 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200594}
595
Gregory Haskins398a1532009-01-14 09:10:04 -0500596#if defined CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500597
598static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
599
600static inline int next_prio(struct rq *rq)
601{
602 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
603
604 if (next && rt_prio(next->prio))
605 return next->prio;
606 else
607 return MAX_RT_PRIO;
608}
Gregory Haskinse864c492008-12-29 09:39:49 -0500609
Gregory Haskins398a1532009-01-14 09:10:04 -0500610static void
611inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +0100612{
Gregory Haskins4d984272008-12-29 09:39:49 -0500613 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins4d984272008-12-29 09:39:49 -0500614
Gregory Haskins398a1532009-01-14 09:10:04 -0500615 if (prio < prev_prio) {
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400616
Gregory Haskinse864c492008-12-29 09:39:49 -0500617 /*
618 * If the new task is higher in priority than anything on the
Gregory Haskins398a1532009-01-14 09:10:04 -0500619 * run-queue, we know that the previous high becomes our
620 * next-highest.
Gregory Haskinse864c492008-12-29 09:39:49 -0500621 */
Gregory Haskins398a1532009-01-14 09:10:04 -0500622 rt_rq->highest_prio.next = prev_prio;
623
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400624 if (rq->online)
Gregory Haskins4d984272008-12-29 09:39:49 -0500625 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
Gregory Haskins398a1532009-01-14 09:10:04 -0500626
Gregory Haskinse864c492008-12-29 09:39:49 -0500627 } else if (prio == rt_rq->highest_prio.curr)
628 /*
629 * If the next task is equal in priority to the highest on
630 * the run-queue, then we implicitly know that the next highest
631 * task cannot be any lower than current
632 */
633 rt_rq->highest_prio.next = prio;
634 else if (prio < rt_rq->highest_prio.next)
635 /*
636 * Otherwise, we need to recompute next-highest
637 */
638 rt_rq->highest_prio.next = next_prio(rq);
Steven Rostedt63489e42008-01-25 21:08:03 +0100639}
640
Gregory Haskins398a1532009-01-14 09:10:04 -0500641static void
642dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
Steven Rostedt63489e42008-01-25 21:08:03 +0100643{
Gregory Haskins4d984272008-12-29 09:39:49 -0500644 struct rq *rq = rq_of_rt_rq(rt_rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200645
Gregory Haskins398a1532009-01-14 09:10:04 -0500646 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
647 rt_rq->highest_prio.next = next_prio(rq);
648
649 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
650 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
651}
652
653#else /* CONFIG_SMP */
654
655static inline
656void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
657static inline
658void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
659
660#endif /* CONFIG_SMP */
661
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100662#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -0500663static void
664inc_rt_prio(struct rt_rq *rt_rq, int prio)
665{
666 int prev_prio = rt_rq->highest_prio.curr;
Steven Rostedt764a9d62008-01-25 21:08:04 +0100667
Gregory Haskins398a1532009-01-14 09:10:04 -0500668 if (prio < prev_prio)
669 rt_rq->highest_prio.curr = prio;
670
671 inc_rt_prio_smp(rt_rq, prio, prev_prio);
672}
673
674static void
675dec_rt_prio(struct rt_rq *rt_rq, int prio)
676{
677 int prev_prio = rt_rq->highest_prio.curr;
678
679 if (rt_rq->rt_nr_running) {
680
681 WARN_ON(prio < prev_prio);
Gregory Haskinse864c492008-12-29 09:39:49 -0500682
683 /*
Gregory Haskins398a1532009-01-14 09:10:04 -0500684 * This may have been our highest task, and therefore
685 * we may have some recomputation to do
Gregory Haskinse864c492008-12-29 09:39:49 -0500686 */
Gregory Haskins398a1532009-01-14 09:10:04 -0500687 if (prio == prev_prio) {
Gregory Haskinse864c492008-12-29 09:39:49 -0500688 struct rt_prio_array *array = &rt_rq->active;
689
690 rt_rq->highest_prio.curr =
Steven Rostedt764a9d62008-01-25 21:08:04 +0100691 sched_find_first_bit(array->bitmap);
Gregory Haskinse864c492008-12-29 09:39:49 -0500692 }
693
Steven Rostedt764a9d62008-01-25 21:08:04 +0100694 } else
Gregory Haskinse864c492008-12-29 09:39:49 -0500695 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100696
Gregory Haskins398a1532009-01-14 09:10:04 -0500697 dec_rt_prio_smp(rt_rq, prio, prev_prio);
698}
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400699
Gregory Haskins398a1532009-01-14 09:10:04 -0500700#else
701
702static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
703static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
704
705#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
706
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100707#ifdef CONFIG_RT_GROUP_SCHED
Gregory Haskins398a1532009-01-14 09:10:04 -0500708
709static void
710inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
711{
712 if (rt_se_boosted(rt_se))
713 rt_rq->rt_nr_boosted++;
714
715 if (rt_rq->tg)
716 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
717}
718
719static void
720dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
721{
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100722 if (rt_se_boosted(rt_se))
723 rt_rq->rt_nr_boosted--;
724
725 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
Gregory Haskins398a1532009-01-14 09:10:04 -0500726}
727
728#else /* CONFIG_RT_GROUP_SCHED */
729
730static void
731inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
732{
733 start_rt_bandwidth(&def_rt_bandwidth);
734}
735
736static inline
737void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
738
739#endif /* CONFIG_RT_GROUP_SCHED */
740
741static inline
742void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
743{
744 int prio = rt_se_prio(rt_se);
745
746 WARN_ON(!rt_prio(prio));
747 rt_rq->rt_nr_running++;
748
749 inc_rt_prio(rt_rq, prio);
750 inc_rt_migration(rt_se, rt_rq);
751 inc_rt_group(rt_se, rt_rq);
752}
753
754static inline
755void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
756{
757 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
758 WARN_ON(!rt_rq->rt_nr_running);
759 rt_rq->rt_nr_running--;
760
761 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
762 dec_rt_migration(rt_se, rt_rq);
763 dec_rt_group(rt_se, rt_rq);
Steven Rostedt63489e42008-01-25 21:08:03 +0100764}
765
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200766static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200767{
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100768 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
769 struct rt_prio_array *array = &rt_rq->active;
770 struct rt_rq *group_rq = group_rt_rq(rt_se);
Dmitry Adamushko20b63312008-06-11 00:58:30 +0200771 struct list_head *queue = array->queue + rt_se_prio(rt_se);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200772
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200773 /*
774 * Don't enqueue the group if its throttled, or when empty.
775 * The latter is a consequence of the former when a child group
776 * get throttled and the current group doesn't have any other
777 * active members.
778 */
779 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100780 return;
Steven Rostedt63489e42008-01-25 21:08:03 +0100781
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200782 list_add_tail(&rt_se->run_list, queue);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100783 __set_bit(rt_se_prio(rt_se), array->bitmap);
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +0100784
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100785 inc_rt_tasks(rt_se, rt_rq);
786}
787
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200788static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100789{
790 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
791 struct rt_prio_array *array = &rt_rq->active;
792
793 list_del_init(&rt_se->run_list);
794 if (list_empty(array->queue + rt_se_prio(rt_se)))
795 __clear_bit(rt_se_prio(rt_se), array->bitmap);
796
797 dec_rt_tasks(rt_se, rt_rq);
798}
799
800/*
801 * Because the prio of an upper entry depends on the lower
802 * entries, we must remove entries top - down.
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100803 */
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200804static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100805{
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200806 struct sched_rt_entity *back = NULL;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100807
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +0200808 for_each_sched_rt_entity(rt_se) {
809 rt_se->back = back;
810 back = rt_se;
811 }
812
813 for (rt_se = back; rt_se; rt_se = rt_se->back) {
814 if (on_rt_rq(rt_se))
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200815 __dequeue_rt_entity(rt_se);
816 }
817}
818
819static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
820{
821 dequeue_rt_stack(rt_se);
822 for_each_sched_rt_entity(rt_se)
823 __enqueue_rt_entity(rt_se);
824}
825
826static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
827{
828 dequeue_rt_stack(rt_se);
829
830 for_each_sched_rt_entity(rt_se) {
831 struct rt_rq *rt_rq = group_rt_rq(rt_se);
832
833 if (rt_rq && rt_rq->rt_nr_running)
834 __enqueue_rt_entity(rt_se);
Peter Zijlstra58d6c2d2008-04-19 19:45:00 +0200835 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200836}
837
838/*
839 * Adding/removing a task to/from a priority array:
840 */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100841static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
842{
843 struct sched_rt_entity *rt_se = &p->rt;
844
845 if (wakeup)
846 rt_se->timeout = 0;
847
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200848 enqueue_rt_entity(rt_se);
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200849
Gregory Haskins917b6272008-12-29 09:39:53 -0500850 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
851 enqueue_pushable_task(rq, p);
852
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200853 inc_cpu_load(rq, p->se.load.weight);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100854}
855
Ingo Molnarf02231e2007-08-09 11:16:48 +0200856static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200857{
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100858 struct sched_rt_entity *rt_se = &p->rt;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200859
Ingo Molnarf1e14ef2007-08-09 11:16:48 +0200860 update_curr_rt(rq);
Peter Zijlstraad2a3f12008-06-19 09:06:57 +0200861 dequeue_rt_entity(rt_se);
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200862
Gregory Haskins917b6272008-12-29 09:39:53 -0500863 dequeue_pushable_task(rq, p);
864
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200865 dec_cpu_load(rq, p->se.load.weight);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200866}
867
868/*
869 * Put task to the end of the run list without the overhead of dequeue
870 * followed by enqueue.
871 */
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200872static void
873requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200874{
Ingo Molnar1cdad712008-06-19 09:09:15 +0200875 if (on_rt_rq(rt_se)) {
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200876 struct rt_prio_array *array = &rt_rq->active;
877 struct list_head *queue = array->queue + rt_se_prio(rt_se);
878
879 if (head)
880 list_move(&rt_se->run_list, queue);
881 else
882 list_move_tail(&rt_se->run_list, queue);
Ingo Molnar1cdad712008-06-19 09:09:15 +0200883 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200884}
885
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200886static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100887{
888 struct sched_rt_entity *rt_se = &p->rt;
889 struct rt_rq *rt_rq;
890
891 for_each_sched_rt_entity(rt_se) {
892 rt_rq = rt_rq_of_se(rt_se);
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200893 requeue_rt_entity(rt_rq, rt_se, head);
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100894 }
895}
896
897static void yield_task_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200898{
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200899 requeue_task_rt(rq, rq->curr, 0);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200900}
901
Gregory Haskinse7693a32008-01-25 21:08:09 +0100902#ifdef CONFIG_SMP
Gregory Haskins318e0892008-01-25 21:08:10 +0100903static int find_lowest_rq(struct task_struct *task);
904
Gregory Haskinse7693a32008-01-25 21:08:09 +0100905static int select_task_rq_rt(struct task_struct *p, int sync)
906{
Gregory Haskins318e0892008-01-25 21:08:10 +0100907 struct rq *rq = task_rq(p);
908
909 /*
Steven Rostedte1f47d82008-01-25 21:08:12 +0100910 * If the current task is an RT task, then
911 * try to see if we can wake this RT task up on another
912 * runqueue. Otherwise simply start this RT task
913 * on its current runqueue.
914 *
915 * We want to avoid overloading runqueues. Even if
916 * the RT task is of higher priority than the current RT task.
917 * RT tasks behave differently than other tasks. If
918 * one gets preempted, we try to push it off to another queue.
919 * So trying to keep a preempting RT task on the same
920 * cache hot CPU will force the running RT task to
921 * a cold CPU. So we waste all the cache for the lower
922 * RT task in hopes of saving some of a RT task
923 * that is just being woken and probably will have
924 * cold cache anyway.
Gregory Haskins318e0892008-01-25 21:08:10 +0100925 */
Gregory Haskins17b32792008-01-25 21:08:13 +0100926 if (unlikely(rt_task(rq->curr)) &&
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100927 (p->rt.nr_cpus_allowed > 1)) {
Gregory Haskins318e0892008-01-25 21:08:10 +0100928 int cpu = find_lowest_rq(p);
929
930 return (cpu == -1) ? task_cpu(p) : cpu;
931 }
932
933 /*
934 * Otherwise, just let it ride on the affined RQ and the
935 * post-schedule router will push the preempted task away
936 */
Gregory Haskinse7693a32008-01-25 21:08:09 +0100937 return task_cpu(p);
938}
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200939
940static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
941{
Rusty Russell24600ce2008-11-25 02:35:13 +1030942 cpumask_var_t mask;
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200943
944 if (rq->curr->rt.nr_cpus_allowed == 1)
945 return;
946
Rusty Russell24600ce2008-11-25 02:35:13 +1030947 if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200948 return;
949
Rusty Russell24600ce2008-11-25 02:35:13 +1030950 if (p->rt.nr_cpus_allowed != 1
951 && cpupri_find(&rq->rd->cpupri, p, mask))
952 goto free;
953
954 if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
955 goto free;
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200956
957 /*
958 * There appears to be other cpus that can accept
959 * current and none to run 'p', so lets reschedule
960 * to try and push current away:
961 */
962 requeue_task_rt(rq, p, 1);
963 resched_task(rq->curr);
Rusty Russell24600ce2008-11-25 02:35:13 +1030964free:
965 free_cpumask_var(mask);
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200966}
967
Gregory Haskinse7693a32008-01-25 21:08:09 +0100968#endif /* CONFIG_SMP */
969
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200970/*
971 * Preempt the current task with a newly woken task if needed:
972 */
Peter Zijlstra15afe092008-09-20 23:38:02 +0200973static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200974{
Gregory Haskins45c01e82008-05-12 21:20:41 +0200975 if (p->prio < rq->curr->prio) {
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200976 resched_task(rq->curr);
Gregory Haskins45c01e82008-05-12 21:20:41 +0200977 return;
978 }
979
980#ifdef CONFIG_SMP
981 /*
982 * If:
983 *
984 * - the newly woken task is of equal priority to the current task
985 * - the newly woken task is non-migratable while current is migratable
986 * - current will be preempted on the next reschedule
987 *
988 * we should check to see if current can readily move to a different
989 * cpu. If so, we will reschedule to allow the push logic to try
990 * to move current somewhere else, making room for our non-migratable
991 * task.
992 */
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +0200993 if (p->prio == rq->curr->prio && !need_resched())
994 check_preempt_equal_prio(rq, p);
Gregory Haskins45c01e82008-05-12 21:20:41 +0200995#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200996}
997
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100998static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
999 struct rt_rq *rt_rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001000{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001001 struct rt_prio_array *array = &rt_rq->active;
1002 struct sched_rt_entity *next = NULL;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001003 struct list_head *queue;
1004 int idx;
1005
1006 idx = sched_find_first_bit(array->bitmap);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001007 BUG_ON(idx >= MAX_RT_PRIO);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001008
1009 queue = array->queue + idx;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001010 next = list_entry(queue->next, struct sched_rt_entity, run_list);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001011
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001012 return next;
1013}
1014
Gregory Haskins917b6272008-12-29 09:39:53 -05001015static struct task_struct *_pick_next_task_rt(struct rq *rq)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001016{
1017 struct sched_rt_entity *rt_se;
1018 struct task_struct *p;
1019 struct rt_rq *rt_rq;
1020
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001021 rt_rq = &rq->rt;
1022
1023 if (unlikely(!rt_rq->rt_nr_running))
1024 return NULL;
1025
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01001026 if (rt_rq_throttled(rt_rq))
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001027 return NULL;
1028
1029 do {
1030 rt_se = pick_next_rt_entity(rq, rt_rq);
Dmitry Adamushko326587b2008-01-25 21:08:34 +01001031 BUG_ON(!rt_se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001032 rt_rq = group_rt_rq(rt_se);
1033 } while (rt_rq);
1034
1035 p = rt_task_of(rt_se);
1036 p->se.exec_start = rq->clock;
Gregory Haskins917b6272008-12-29 09:39:53 -05001037
1038 return p;
1039}
1040
1041static struct task_struct *pick_next_task_rt(struct rq *rq)
1042{
1043 struct task_struct *p = _pick_next_task_rt(rq);
1044
1045 /* The running task is never eligible for pushing */
1046 if (p)
1047 dequeue_pushable_task(rq, p);
1048
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001049 return p;
1050}
1051
Ingo Molnar31ee5292007-08-09 11:16:49 +02001052static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001053{
Ingo Molnarf1e14ef2007-08-09 11:16:48 +02001054 update_curr_rt(rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001055 p->se.exec_start = 0;
Gregory Haskins917b6272008-12-29 09:39:53 -05001056
1057 /*
1058 * The previous task needs to be made eligible for pushing
1059 * if it is still active
1060 */
1061 if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1062 enqueue_pushable_task(rq, p);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001063}
1064
Peter Williams681f3e62007-10-24 18:23:51 +02001065#ifdef CONFIG_SMP
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001066
Steven Rostedte8fa1362008-01-25 21:08:05 +01001067/* Only try algorithms three times */
1068#define RT_MAX_TRIES 3
1069
Steven Rostedte8fa1362008-01-25 21:08:05 +01001070static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1071
Steven Rostedtf65eda42008-01-25 21:08:07 +01001072static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1073{
1074 if (!task_running(rq, p) &&
Rusty Russell96f874e22008-11-25 02:35:14 +10301075 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001076 (p->rt.nr_cpus_allowed > 1))
Steven Rostedtf65eda42008-01-25 21:08:07 +01001077 return 1;
1078 return 0;
1079}
1080
Steven Rostedte8fa1362008-01-25 21:08:05 +01001081/* Return the second highest RT task, NULL otherwise */
Ingo Molnar79064fb2008-01-25 21:08:14 +01001082static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001083{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001084 struct task_struct *next = NULL;
1085 struct sched_rt_entity *rt_se;
1086 struct rt_prio_array *array;
1087 struct rt_rq *rt_rq;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001088 int idx;
1089
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001090 for_each_leaf_rt_rq(rt_rq, rq) {
1091 array = &rt_rq->active;
1092 idx = sched_find_first_bit(array->bitmap);
1093 next_idx:
1094 if (idx >= MAX_RT_PRIO)
1095 continue;
1096 if (next && next->prio < idx)
1097 continue;
1098 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1099 struct task_struct *p = rt_task_of(rt_se);
1100 if (pick_rt_task(rq, p, cpu)) {
1101 next = p;
1102 break;
1103 }
1104 }
1105 if (!next) {
1106 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1107 goto next_idx;
1108 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01001109 }
1110
Steven Rostedte8fa1362008-01-25 21:08:05 +01001111 return next;
1112}
1113
Rusty Russell0e3900e2008-11-25 02:35:13 +10301114static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001115
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001116static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
1117{
1118 int first;
1119
1120 /* "this_cpu" is cheaper to preempt than a remote processor */
1121 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
1122 return this_cpu;
1123
1124 first = first_cpu(*mask);
1125 if (first != NR_CPUS)
1126 return first;
1127
1128 return -1;
1129}
1130
1131static int find_lowest_rq(struct task_struct *task)
1132{
1133 struct sched_domain *sd;
Rusty Russell96f874e22008-11-25 02:35:14 +10301134 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001135 int this_cpu = smp_processor_id();
1136 int cpu = task_cpu(task);
1137
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001138 if (task->rt.nr_cpus_allowed == 1)
1139 return -1; /* No other targets possible */
1140
1141 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
Gregory Haskins06f90db2008-01-25 21:08:13 +01001142 return -1; /* No targets found */
1143
1144 /*
Max Krasnyanskye761b772008-07-15 04:43:49 -07001145 * Only consider CPUs that are usable for migration.
1146 * I guess we might want to change cpupri_find() to ignore those
1147 * in the first place.
1148 */
Rusty Russell96f874e22008-11-25 02:35:14 +10301149 cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
Max Krasnyanskye761b772008-07-15 04:43:49 -07001150
1151 /*
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001152 * At this point we have built a mask of cpus representing the
1153 * lowest priority tasks in the system. Now we want to elect
1154 * the best one based on our affinity and topology.
1155 *
1156 * We prioritize the last cpu that the task executed on since
1157 * it is most likely cache-hot in that location.
1158 */
Rusty Russell96f874e22008-11-25 02:35:14 +10301159 if (cpumask_test_cpu(cpu, lowest_mask))
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001160 return cpu;
1161
1162 /*
1163 * Otherwise, we consult the sched_domains span maps to figure
1164 * out which cpu is logically closest to our hot cache data.
1165 */
1166 if (this_cpu == cpu)
1167 this_cpu = -1; /* Skip this_cpu opt if the same */
1168
1169 for_each_domain(cpu, sd) {
1170 if (sd->flags & SD_WAKE_AFFINE) {
1171 cpumask_t domain_mask;
1172 int best_cpu;
1173
Rusty Russell758b2cd2008-11-25 02:35:04 +10301174 cpumask_and(&domain_mask, sched_domain_span(sd),
1175 lowest_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +01001176
1177 best_cpu = pick_optimal_cpu(this_cpu,
1178 &domain_mask);
1179 if (best_cpu != -1)
1180 return best_cpu;
1181 }
1182 }
1183
1184 /*
1185 * And finally, if there were no matches within the domains
1186 * just give the caller *something* to work with from the compatible
1187 * locations.
1188 */
1189 return pick_optimal_cpu(this_cpu, lowest_mask);
Gregory Haskins07b40322008-01-25 21:08:10 +01001190}
1191
Steven Rostedte8fa1362008-01-25 21:08:05 +01001192/* Will lock the rq it finds */
Ingo Molnar4df64c02008-01-25 21:08:15 +01001193static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001194{
1195 struct rq *lowest_rq = NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001196 int tries;
Ingo Molnar4df64c02008-01-25 21:08:15 +01001197 int cpu;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001198
1199 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
Gregory Haskins07b40322008-01-25 21:08:10 +01001200 cpu = find_lowest_rq(task);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001201
Gregory Haskins2de0b462008-01-25 21:08:10 +01001202 if ((cpu == -1) || (cpu == rq->cpu))
Steven Rostedte8fa1362008-01-25 21:08:05 +01001203 break;
1204
Gregory Haskins07b40322008-01-25 21:08:10 +01001205 lowest_rq = cpu_rq(cpu);
1206
Steven Rostedte8fa1362008-01-25 21:08:05 +01001207 /* if the prio of this runqueue changed, try again */
Gregory Haskins07b40322008-01-25 21:08:10 +01001208 if (double_lock_balance(rq, lowest_rq)) {
Steven Rostedte8fa1362008-01-25 21:08:05 +01001209 /*
1210 * We had to unlock the run queue. In
1211 * the mean time, task could have
1212 * migrated already or had its affinity changed.
1213 * Also make sure that it wasn't scheduled on its rq.
1214 */
Gregory Haskins07b40322008-01-25 21:08:10 +01001215 if (unlikely(task_rq(task) != rq ||
Rusty Russell96f874e22008-11-25 02:35:14 +10301216 !cpumask_test_cpu(lowest_rq->cpu,
1217 &task->cpus_allowed) ||
Gregory Haskins07b40322008-01-25 21:08:10 +01001218 task_running(rq, task) ||
Steven Rostedte8fa1362008-01-25 21:08:05 +01001219 !task->se.on_rq)) {
Ingo Molnar4df64c02008-01-25 21:08:15 +01001220
Steven Rostedte8fa1362008-01-25 21:08:05 +01001221 spin_unlock(&lowest_rq->lock);
1222 lowest_rq = NULL;
1223 break;
1224 }
1225 }
1226
1227 /* If this rq is still suitable use it. */
Gregory Haskinse864c492008-12-29 09:39:49 -05001228 if (lowest_rq->rt.highest_prio.curr > task->prio)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001229 break;
1230
1231 /* try again */
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001232 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001233 lowest_rq = NULL;
1234 }
1235
1236 return lowest_rq;
1237}
1238
Gregory Haskins917b6272008-12-29 09:39:53 -05001239static inline int has_pushable_tasks(struct rq *rq)
1240{
1241 return !plist_head_empty(&rq->rt.pushable_tasks);
1242}
1243
1244static struct task_struct *pick_next_pushable_task(struct rq *rq)
1245{
1246 struct task_struct *p;
1247
1248 if (!has_pushable_tasks(rq))
1249 return NULL;
1250
1251 p = plist_first_entry(&rq->rt.pushable_tasks,
1252 struct task_struct, pushable_tasks);
1253
1254 BUG_ON(rq->cpu != task_cpu(p));
1255 BUG_ON(task_current(rq, p));
1256 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1257
1258 BUG_ON(!p->se.on_rq);
1259 BUG_ON(!rt_task(p));
1260
1261 return p;
1262}
1263
Steven Rostedte8fa1362008-01-25 21:08:05 +01001264/*
1265 * If the current CPU has more than one RT task, see if the non
1266 * running task can migrate over to a CPU that is running a task
1267 * of lesser priority.
1268 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001269static int push_rt_task(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001270{
1271 struct task_struct *next_task;
1272 struct rq *lowest_rq;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001273
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +01001274 if (!rq->rt.overloaded)
1275 return 0;
1276
Gregory Haskins917b6272008-12-29 09:39:53 -05001277 next_task = pick_next_pushable_task(rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001278 if (!next_task)
1279 return 0;
1280
1281 retry:
Gregory Haskins697f0a42008-01-25 21:08:09 +01001282 if (unlikely(next_task == rq->curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001283 WARN_ON(1);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001284 return 0;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001285 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01001286
1287 /*
1288 * It's possible that the next_task slipped in of
1289 * higher priority than current. If that's the case
1290 * just reschedule current.
1291 */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001292 if (unlikely(next_task->prio < rq->curr->prio)) {
1293 resched_task(rq->curr);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001294 return 0;
1295 }
1296
Gregory Haskins697f0a42008-01-25 21:08:09 +01001297 /* We might release rq lock */
Steven Rostedte8fa1362008-01-25 21:08:05 +01001298 get_task_struct(next_task);
1299
1300 /* find_lock_lowest_rq locks the rq if found */
Gregory Haskins697f0a42008-01-25 21:08:09 +01001301 lowest_rq = find_lock_lowest_rq(next_task, rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001302 if (!lowest_rq) {
1303 struct task_struct *task;
1304 /*
Gregory Haskins697f0a42008-01-25 21:08:09 +01001305 * find lock_lowest_rq releases rq->lock
Gregory Haskins15635132008-12-29 09:39:53 -05001306 * so it is possible that next_task has migrated.
1307 *
1308 * We need to make sure that the task is still on the same
1309 * run-queue and is also still the next task eligible for
1310 * pushing.
Steven Rostedte8fa1362008-01-25 21:08:05 +01001311 */
Gregory Haskins917b6272008-12-29 09:39:53 -05001312 task = pick_next_pushable_task(rq);
Gregory Haskins15635132008-12-29 09:39:53 -05001313 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1314 /*
1315 * If we get here, the task hasnt moved at all, but
1316 * it has failed to push. We will not try again,
1317 * since the other cpus will pull from us when they
1318 * are ready.
1319 */
1320 dequeue_pushable_task(rq, next_task);
1321 goto out;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001322 }
Gregory Haskins917b6272008-12-29 09:39:53 -05001323
Gregory Haskins15635132008-12-29 09:39:53 -05001324 if (!task)
1325 /* No more tasks, just exit */
1326 goto out;
1327
Gregory Haskins917b6272008-12-29 09:39:53 -05001328 /*
Gregory Haskins15635132008-12-29 09:39:53 -05001329 * Something has shifted, try again.
Gregory Haskins917b6272008-12-29 09:39:53 -05001330 */
Gregory Haskins15635132008-12-29 09:39:53 -05001331 put_task_struct(next_task);
1332 next_task = task;
1333 goto retry;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001334 }
1335
Gregory Haskins697f0a42008-01-25 21:08:09 +01001336 deactivate_task(rq, next_task, 0);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001337 set_task_cpu(next_task, lowest_rq->cpu);
1338 activate_task(lowest_rq, next_task, 0);
1339
1340 resched_task(lowest_rq->curr);
1341
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001342 double_unlock_balance(rq, lowest_rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001343
Steven Rostedte8fa1362008-01-25 21:08:05 +01001344out:
1345 put_task_struct(next_task);
1346
Gregory Haskins917b6272008-12-29 09:39:53 -05001347 return 1;
Steven Rostedte8fa1362008-01-25 21:08:05 +01001348}
1349
Steven Rostedte8fa1362008-01-25 21:08:05 +01001350static void push_rt_tasks(struct rq *rq)
1351{
1352 /* push_rt_task will return true if it moved an RT */
1353 while (push_rt_task(rq))
1354 ;
1355}
1356
Steven Rostedtf65eda42008-01-25 21:08:07 +01001357static int pull_rt_task(struct rq *this_rq)
1358{
Ingo Molnar80bf3172008-01-25 21:08:17 +01001359 int this_cpu = this_rq->cpu, ret = 0, cpu;
Gregory Haskinsa8728942008-12-29 09:39:49 -05001360 struct task_struct *p;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001361 struct rq *src_rq;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001362
Gregory Haskins637f5082008-01-25 21:08:18 +01001363 if (likely(!rt_overloaded(this_rq)))
Steven Rostedtf65eda42008-01-25 21:08:07 +01001364 return 0;
1365
Rusty Russellc6c49272008-11-25 02:35:05 +10301366 for_each_cpu(cpu, this_rq->rd->rto_mask) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001367 if (this_cpu == cpu)
1368 continue;
1369
1370 src_rq = cpu_rq(cpu);
Gregory Haskins74ab8e42008-12-29 09:39:50 -05001371
1372 /*
1373 * Don't bother taking the src_rq->lock if the next highest
1374 * task is known to be lower-priority than our current task.
1375 * This may look racy, but if this value is about to go
1376 * logically higher, the src_rq will push this task away.
1377 * And if its going logically lower, we do not care
1378 */
1379 if (src_rq->rt.highest_prio.next >=
1380 this_rq->rt.highest_prio.curr)
1381 continue;
1382
Steven Rostedtf65eda42008-01-25 21:08:07 +01001383 /*
1384 * We can potentially drop this_rq's lock in
1385 * double_lock_balance, and another CPU could
Gregory Haskinsa8728942008-12-29 09:39:49 -05001386 * alter this_rq
Steven Rostedtf65eda42008-01-25 21:08:07 +01001387 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001388 double_lock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001389
1390 /*
1391 * Are there still pullable RT tasks?
1392 */
Mike Galbraith614ee1f2008-01-25 21:08:30 +01001393 if (src_rq->rt.rt_nr_running <= 1)
1394 goto skip;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001395
Steven Rostedtf65eda42008-01-25 21:08:07 +01001396 p = pick_next_highest_task_rt(src_rq, this_cpu);
1397
1398 /*
1399 * Do we have an RT task that preempts
1400 * the to-be-scheduled task?
1401 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001402 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +01001403 WARN_ON(p == src_rq->curr);
1404 WARN_ON(!p->se.on_rq);
1405
1406 /*
1407 * There's a chance that p is higher in priority
1408 * than what's currently running on its cpu.
1409 * This is just that p is wakeing up and hasn't
1410 * had a chance to schedule. We only pull
1411 * p if it is lower in priority than the
Gregory Haskinsa8728942008-12-29 09:39:49 -05001412 * current task on the run queue
Steven Rostedtf65eda42008-01-25 21:08:07 +01001413 */
Gregory Haskinsa8728942008-12-29 09:39:49 -05001414 if (p->prio < src_rq->curr->prio)
Mike Galbraith614ee1f2008-01-25 21:08:30 +01001415 goto skip;
Steven Rostedtf65eda42008-01-25 21:08:07 +01001416
1417 ret = 1;
1418
1419 deactivate_task(src_rq, p, 0);
1420 set_task_cpu(p, this_cpu);
1421 activate_task(this_rq, p, 0);
1422 /*
1423 * We continue with the search, just in
1424 * case there's an even higher prio task
1425 * in another runqueue. (low likelyhood
1426 * but possible)
Steven Rostedtf65eda42008-01-25 21:08:07 +01001427 */
Steven Rostedtf65eda42008-01-25 21:08:07 +01001428 }
Mike Galbraith614ee1f2008-01-25 21:08:30 +01001429 skip:
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02001430 double_unlock_balance(this_rq, src_rq);
Steven Rostedtf65eda42008-01-25 21:08:07 +01001431 }
1432
1433 return ret;
1434}
1435
Steven Rostedt9a897c52008-01-25 21:08:22 +01001436static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
Steven Rostedtf65eda42008-01-25 21:08:07 +01001437{
1438 /* Try to pull RT tasks here if we lower this rq's prio */
Gregory Haskinse864c492008-12-29 09:39:49 -05001439 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
Steven Rostedtf65eda42008-01-25 21:08:07 +01001440 pull_rt_task(rq);
1441}
1442
Gregory Haskins967fc042008-12-29 09:39:52 -05001443/*
1444 * assumes rq->lock is held
1445 */
1446static int needs_post_schedule_rt(struct rq *rq)
1447{
Gregory Haskins917b6272008-12-29 09:39:53 -05001448 return has_pushable_tasks(rq);
Gregory Haskins967fc042008-12-29 09:39:52 -05001449}
1450
Steven Rostedt9a897c52008-01-25 21:08:22 +01001451static void post_schedule_rt(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +01001452{
1453 /*
Gregory Haskins967fc042008-12-29 09:39:52 -05001454 * This is only called if needs_post_schedule_rt() indicates that
1455 * we need to push tasks away
Steven Rostedte8fa1362008-01-25 21:08:05 +01001456 */
Gregory Haskins967fc042008-12-29 09:39:52 -05001457 spin_lock_irq(&rq->lock);
1458 push_rt_tasks(rq);
1459 spin_unlock_irq(&rq->lock);
Steven Rostedte8fa1362008-01-25 21:08:05 +01001460}
1461
Gregory Haskins8ae121a2008-04-23 07:13:29 -04001462/*
1463 * If we are not running and we are not going to reschedule soon, we should
1464 * try to push tasks away now
1465 */
Steven Rostedt9a897c52008-01-25 21:08:22 +01001466static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
Steven Rostedt4642daf2008-01-25 21:08:07 +01001467{
Steven Rostedt9a897c52008-01-25 21:08:22 +01001468 if (!task_running(rq, p) &&
Gregory Haskins8ae121a2008-04-23 07:13:29 -04001469 !test_tsk_need_resched(rq->curr) &&
Gregory Haskins917b6272008-12-29 09:39:53 -05001470 has_pushable_tasks(rq) &&
Gregory Haskins777c2f32008-12-29 09:39:50 -05001471 p->rt.nr_cpus_allowed > 1)
Steven Rostedt4642daf2008-01-25 21:08:07 +01001472 push_rt_tasks(rq);
1473}
1474
Peter Williams43010652007-08-09 11:16:46 +02001475static unsigned long
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001476load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02001477 unsigned long max_load_move,
1478 struct sched_domain *sd, enum cpu_idle_type idle,
1479 int *all_pinned, int *this_best_prio)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001480{
Steven Rostedtc7a1e462008-01-25 21:08:07 +01001481 /* don't touch RT tasks */
1482 return 0;
Peter Williamse1d14842007-10-24 18:23:51 +02001483}
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001484
Peter Williamse1d14842007-10-24 18:23:51 +02001485static int
1486move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1487 struct sched_domain *sd, enum cpu_idle_type idle)
1488{
Steven Rostedtc7a1e462008-01-25 21:08:07 +01001489 /* don't touch RT tasks */
1490 return 0;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001491}
Ingo Molnardeeeccd2008-01-25 21:08:15 +01001492
Mike Traviscd8ba7c2008-03-26 14:23:49 -07001493static void set_cpus_allowed_rt(struct task_struct *p,
Rusty Russell96f874e22008-11-25 02:35:14 +10301494 const struct cpumask *new_mask)
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001495{
Rusty Russell96f874e22008-11-25 02:35:14 +10301496 int weight = cpumask_weight(new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001497
1498 BUG_ON(!rt_task(p));
1499
1500 /*
1501 * Update the migration status of the RQ if we have an RT task
1502 * which is running AND changing its weight value.
1503 */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001504 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001505 struct rq *rq = task_rq(p);
1506
Gregory Haskins917b6272008-12-29 09:39:53 -05001507 if (!task_current(rq, p)) {
1508 /*
1509 * Make sure we dequeue this task from the pushable list
1510 * before going further. It will either remain off of
1511 * the list because we are no longer pushable, or it
1512 * will be requeued.
1513 */
1514 if (p->rt.nr_cpus_allowed > 1)
1515 dequeue_pushable_task(rq, p);
1516
1517 /*
1518 * Requeue if our weight is changing and still > 1
1519 */
1520 if (weight > 1)
1521 enqueue_pushable_task(rq, p);
1522
1523 }
1524
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001525 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001526 rq->rt.rt_nr_migratory++;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001527 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001528 BUG_ON(!rq->rt.rt_nr_migratory);
1529 rq->rt.rt_nr_migratory--;
1530 }
1531
Gregory Haskins398a1532009-01-14 09:10:04 -05001532 update_rt_migration(&rq->rt);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001533 }
1534
Rusty Russell96f874e22008-11-25 02:35:14 +10301535 cpumask_copy(&p->cpus_allowed, new_mask);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001536 p->rt.nr_cpus_allowed = weight;
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001537}
Ingo Molnardeeeccd2008-01-25 21:08:15 +01001538
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001539/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001540static void rq_online_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001541{
1542 if (rq->rt.overloaded)
1543 rt_set_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001544
Peter Zijlstra7def2be2008-06-05 14:49:58 +02001545 __enable_runtime(rq);
1546
Gregory Haskinse864c492008-12-29 09:39:49 -05001547 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001548}
1549
1550/* Assumes rq->lock is held */
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001551static void rq_offline_rt(struct rq *rq)
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001552{
1553 if (rq->rt.overloaded)
1554 rt_clear_overload(rq);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001555
Peter Zijlstra7def2be2008-06-05 14:49:58 +02001556 __disable_runtime(rq);
1557
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001558 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
Ingo Molnarbdd7c812008-01-25 21:08:18 +01001559}
Steven Rostedtcb469842008-01-25 21:08:22 +01001560
1561/*
1562 * When switch from the rt queue, we bring ourselves to a position
1563 * that we might want to pull RT tasks from other runqueues.
1564 */
1565static void switched_from_rt(struct rq *rq, struct task_struct *p,
1566 int running)
1567{
1568 /*
1569 * If there are other RT tasks then we will reschedule
1570 * and the scheduling of the other RT tasks will handle
1571 * the balancing. But if we are the last RT task
1572 * we may need to handle the pulling of RT tasks
1573 * now.
1574 */
1575 if (!rq->rt.rt_nr_running)
1576 pull_rt_task(rq);
1577}
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10301578
1579static inline void init_sched_rt_class(void)
1580{
1581 unsigned int i;
1582
1583 for_each_possible_cpu(i)
Mike Travis6ca09df2008-12-31 18:08:45 -08001584 alloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1585 GFP_KERNEL, cpu_to_node(i));
Rusty Russell3d8cbdf2008-11-25 09:58:41 +10301586}
Steven Rostedte8fa1362008-01-25 21:08:05 +01001587#endif /* CONFIG_SMP */
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001588
Steven Rostedtcb469842008-01-25 21:08:22 +01001589/*
1590 * When switching a task to RT, we may overload the runqueue
1591 * with RT tasks. In this case we try to push them off to
1592 * other runqueues.
1593 */
1594static void switched_to_rt(struct rq *rq, struct task_struct *p,
1595 int running)
1596{
1597 int check_resched = 1;
1598
1599 /*
1600 * If we are already running, then there's nothing
1601 * that needs to be done. But if we are not running
1602 * we may need to preempt the current running task.
1603 * If that current running task is also an RT task
1604 * then see if we can move to another run queue.
1605 */
1606 if (!running) {
1607#ifdef CONFIG_SMP
1608 if (rq->rt.overloaded && push_rt_task(rq) &&
1609 /* Don't resched if we changed runqueues */
1610 rq != task_rq(p))
1611 check_resched = 0;
1612#endif /* CONFIG_SMP */
1613 if (check_resched && p->prio < rq->curr->prio)
1614 resched_task(rq->curr);
1615 }
1616}
1617
1618/*
1619 * Priority of the task has changed. This may cause
1620 * us to initiate a push or pull.
1621 */
1622static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1623 int oldprio, int running)
1624{
1625 if (running) {
1626#ifdef CONFIG_SMP
1627 /*
1628 * If our priority decreases while running, we
1629 * may need to pull tasks to this runqueue.
1630 */
1631 if (oldprio < p->prio)
1632 pull_rt_task(rq);
1633 /*
1634 * If there's a higher priority task waiting to run
Steven Rostedt6fa46fa2008-03-05 10:00:12 -05001635 * then reschedule. Note, the above pull_rt_task
1636 * can release the rq lock and p could migrate.
1637 * Only reschedule if p is still on the same runqueue.
Steven Rostedtcb469842008-01-25 21:08:22 +01001638 */
Gregory Haskinse864c492008-12-29 09:39:49 -05001639 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
Steven Rostedtcb469842008-01-25 21:08:22 +01001640 resched_task(p);
1641#else
1642 /* For UP simply resched on drop of prio */
1643 if (oldprio < p->prio)
1644 resched_task(p);
1645#endif /* CONFIG_SMP */
1646 } else {
1647 /*
1648 * This task is not running, but if it is
1649 * greater than the current running task
1650 * then reschedule.
1651 */
1652 if (p->prio < rq->curr->prio)
1653 resched_task(rq->curr);
1654 }
1655}
1656
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001657static void watchdog(struct rq *rq, struct task_struct *p)
1658{
1659 unsigned long soft, hard;
1660
1661 if (!p->signal)
1662 return;
1663
1664 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1665 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1666
1667 if (soft != RLIM_INFINITY) {
1668 unsigned long next;
1669
1670 p->rt.timeout++;
1671 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
Peter Zijlstra5a52dd52008-01-25 21:08:32 +01001672 if (p->rt.timeout > next)
Frank Mayharf06febc2008-09-12 09:54:39 -07001673 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001674 }
1675}
Steven Rostedtcb469842008-01-25 21:08:22 +01001676
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001677static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001678{
Peter Zijlstra67e2be02007-12-20 15:01:17 +01001679 update_curr_rt(rq);
1680
Peter Zijlstra78f2c7d2008-01-25 21:08:27 +01001681 watchdog(rq, p);
1682
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001683 /*
1684 * RR tasks need a special form of timeslice management.
1685 * FIFO tasks have no timeslices.
1686 */
1687 if (p->policy != SCHED_RR)
1688 return;
1689
Peter Zijlstrafa717062008-01-25 21:08:27 +01001690 if (--p->rt.time_slice)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001691 return;
1692
Peter Zijlstrafa717062008-01-25 21:08:27 +01001693 p->rt.time_slice = DEF_TIMESLICE;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001694
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02001695 /*
1696 * Requeue to the end of queue if we are not the only element
1697 * on the queue:
1698 */
Peter Zijlstrafa717062008-01-25 21:08:27 +01001699 if (p->rt.run_list.prev != p->rt.run_list.next) {
Dmitry Adamushko7ebefa82008-07-01 23:32:15 +02001700 requeue_task_rt(rq, p, 0);
Dmitry Adamushko98fbc792007-08-24 20:39:10 +02001701 set_tsk_need_resched(p);
1702 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001703}
1704
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001705static void set_curr_task_rt(struct rq *rq)
1706{
1707 struct task_struct *p = rq->curr;
1708
1709 p->se.exec_start = rq->clock;
Gregory Haskins917b6272008-12-29 09:39:53 -05001710
1711 /* The running task is never eligible for pushing */
1712 dequeue_pushable_task(rq, p);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001713}
1714
Harvey Harrison2abdad02008-04-25 10:53:13 -07001715static const struct sched_class rt_sched_class = {
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001716 .next = &fair_sched_class,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001717 .enqueue_task = enqueue_task_rt,
1718 .dequeue_task = dequeue_task_rt,
1719 .yield_task = yield_task_rt,
1720
1721 .check_preempt_curr = check_preempt_curr_rt,
1722
1723 .pick_next_task = pick_next_task_rt,
1724 .put_prev_task = put_prev_task_rt,
1725
Peter Williams681f3e62007-10-24 18:23:51 +02001726#ifdef CONFIG_SMP
Li Zefan4ce72a22008-10-22 15:25:26 +08001727 .select_task_rq = select_task_rq_rt,
1728
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001729 .load_balance = load_balance_rt,
Peter Williamse1d14842007-10-24 18:23:51 +02001730 .move_one_task = move_one_task_rt,
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01001731 .set_cpus_allowed = set_cpus_allowed_rt,
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001732 .rq_online = rq_online_rt,
1733 .rq_offline = rq_offline_rt,
Steven Rostedt9a897c52008-01-25 21:08:22 +01001734 .pre_schedule = pre_schedule_rt,
Gregory Haskins967fc042008-12-29 09:39:52 -05001735 .needs_post_schedule = needs_post_schedule_rt,
Steven Rostedt9a897c52008-01-25 21:08:22 +01001736 .post_schedule = post_schedule_rt,
1737 .task_wake_up = task_wake_up_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01001738 .switched_from = switched_from_rt,
Peter Williams681f3e62007-10-24 18:23:51 +02001739#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001740
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02001741 .set_curr_task = set_curr_task_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001742 .task_tick = task_tick_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +01001743
1744 .prio_changed = prio_changed_rt,
1745 .switched_to = switched_to_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +02001746};
Peter Zijlstraada18de2008-06-19 14:22:24 +02001747
1748#ifdef CONFIG_SCHED_DEBUG
1749extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1750
1751static void print_rt_stats(struct seq_file *m, int cpu)
1752{
1753 struct rt_rq *rt_rq;
1754
1755 rcu_read_lock();
1756 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1757 print_rt_rq(m, cpu, rt_rq);
1758 rcu_read_unlock();
1759}
Dhaval Giani55e12e52008-06-24 23:39:43 +05301760#endif /* CONFIG_SCHED_DEBUG */
Rusty Russell0e3900e2008-11-25 02:35:13 +10301761