blob: 29963af782aedb669a41a3cfce538bf0713cceea [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
Steven Rostedt4fd29172008-01-25 21:08:06 +01006#ifdef CONFIG_SMP
Ingo Molnar84de4272008-01-25 21:08:15 +01007
Gregory Haskins637f5082008-01-25 21:08:18 +01008static inline int rt_overloaded(struct rq *rq)
Steven Rostedt4fd29172008-01-25 21:08:06 +01009{
Gregory Haskins637f5082008-01-25 21:08:18 +010010 return atomic_read(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +010011}
Ingo Molnar84de4272008-01-25 21:08:15 +010012
Steven Rostedt4fd29172008-01-25 21:08:06 +010013static inline void rt_set_overload(struct rq *rq)
14{
Gregory Haskins637f5082008-01-25 21:08:18 +010015 cpu_set(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +010016 /*
17 * Make sure the mask is visible before we set
18 * the overload count. That is checked to determine
19 * if we should look at the mask. It would be a shame
20 * if we looked at the mask, but the mask was not
21 * updated yet.
22 */
23 wmb();
Gregory Haskins637f5082008-01-25 21:08:18 +010024 atomic_inc(&rq->rd->rto_count);
Steven Rostedt4fd29172008-01-25 21:08:06 +010025}
Ingo Molnar84de4272008-01-25 21:08:15 +010026
Steven Rostedt4fd29172008-01-25 21:08:06 +010027static inline void rt_clear_overload(struct rq *rq)
28{
29 /* the order here really doesn't matter */
Gregory Haskins637f5082008-01-25 21:08:18 +010030 atomic_dec(&rq->rd->rto_count);
31 cpu_clear(rq->cpu, rq->rd->rto_mask);
Steven Rostedt4fd29172008-01-25 21:08:06 +010032}
Gregory Haskins73fe6aae2008-01-25 21:08:07 +010033
34static void update_rt_migration(struct rq *rq)
35{
Gregory Haskins637f5082008-01-25 21:08:18 +010036 if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
Gregory Haskinscdc8eb92008-01-25 21:08:23 +010037 if (!rq->rt.overloaded) {
38 rt_set_overload(rq);
39 rq->rt.overloaded = 1;
40 }
41 } else if (rq->rt.overloaded) {
Gregory Haskins73fe6aae2008-01-25 21:08:07 +010042 rt_clear_overload(rq);
Gregory Haskins637f5082008-01-25 21:08:18 +010043 rq->rt.overloaded = 0;
44 }
Gregory Haskins73fe6aae2008-01-25 21:08:07 +010045}
Steven Rostedt4fd29172008-01-25 21:08:06 +010046#endif /* CONFIG_SMP */
47
Ingo Molnarbb44e5d2007-07-09 18:51:58 +020048/*
49 * Update the current task's runtime statistics. Skip current tasks that
50 * are not in our scheduling class.
51 */
Alexey Dobriyana9957442007-10-15 17:00:13 +020052static void update_curr_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +020053{
54 struct task_struct *curr = rq->curr;
55 u64 delta_exec;
56
57 if (!task_has_rt_policy(curr))
58 return;
59
Ingo Molnard2819182007-08-09 11:16:47 +020060 delta_exec = rq->clock - curr->se.exec_start;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +020061 if (unlikely((s64)delta_exec < 0))
62 delta_exec = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +020063
64 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
Ingo Molnarbb44e5d2007-07-09 18:51:58 +020065
66 curr->se.sum_exec_runtime += delta_exec;
Ingo Molnard2819182007-08-09 11:16:47 +020067 curr->se.exec_start = rq->clock;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010068 cpuacct_charge(curr, delta_exec);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +020069}
70
Steven Rostedt63489e42008-01-25 21:08:03 +010071static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq)
72{
73 WARN_ON(!rt_task(p));
74 rq->rt.rt_nr_running++;
Steven Rostedt764a9d62008-01-25 21:08:04 +010075#ifdef CONFIG_SMP
76 if (p->prio < rq->rt.highest_prio)
77 rq->rt.highest_prio = p->prio;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +010078 if (p->nr_cpus_allowed > 1)
79 rq->rt.rt_nr_migratory++;
80
81 update_rt_migration(rq);
Steven Rostedt764a9d62008-01-25 21:08:04 +010082#endif /* CONFIG_SMP */
Steven Rostedt63489e42008-01-25 21:08:03 +010083}
84
85static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq)
86{
87 WARN_ON(!rt_task(p));
88 WARN_ON(!rq->rt.rt_nr_running);
89 rq->rt.rt_nr_running--;
Steven Rostedt764a9d62008-01-25 21:08:04 +010090#ifdef CONFIG_SMP
91 if (rq->rt.rt_nr_running) {
92 struct rt_prio_array *array;
93
94 WARN_ON(p->prio < rq->rt.highest_prio);
95 if (p->prio == rq->rt.highest_prio) {
96 /* recalculate */
97 array = &rq->rt.active;
98 rq->rt.highest_prio =
99 sched_find_first_bit(array->bitmap);
100 } /* otherwise leave rq->highest prio alone */
101 } else
102 rq->rt.highest_prio = MAX_RT_PRIO;
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100103 if (p->nr_cpus_allowed > 1)
104 rq->rt.rt_nr_migratory--;
105
106 update_rt_migration(rq);
Steven Rostedt764a9d62008-01-25 21:08:04 +0100107#endif /* CONFIG_SMP */
Steven Rostedt63489e42008-01-25 21:08:03 +0100108}
109
Ingo Molnarfd390f62007-08-09 11:16:48 +0200110static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200111{
112 struct rt_prio_array *array = &rq->rt.active;
113
Peter Zijlstrafa717062008-01-25 21:08:27 +0100114 list_add_tail(&p->rt.run_list, array->queue + p->prio);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200115 __set_bit(p->prio, array->bitmap);
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +0100116 inc_cpu_load(rq, p->se.load.weight);
Steven Rostedt63489e42008-01-25 21:08:03 +0100117
118 inc_rt_tasks(p, rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200119}
120
121/*
122 * Adding/removing a task to/from a priority array:
123 */
Ingo Molnarf02231e2007-08-09 11:16:48 +0200124static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200125{
126 struct rt_prio_array *array = &rq->rt.active;
127
Ingo Molnarf1e14ef2007-08-09 11:16:48 +0200128 update_curr_rt(rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200129
Peter Zijlstrafa717062008-01-25 21:08:27 +0100130 list_del(&p->rt.run_list);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200131 if (list_empty(array->queue + p->prio))
132 __clear_bit(p->prio, array->bitmap);
Srivatsa Vaddagiri58e2d4c2008-01-25 21:08:00 +0100133 dec_cpu_load(rq, p->se.load.weight);
Steven Rostedt63489e42008-01-25 21:08:03 +0100134
135 dec_rt_tasks(p, rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200136}
137
138/*
139 * Put task to the end of the run list without the overhead of dequeue
140 * followed by enqueue.
141 */
142static void requeue_task_rt(struct rq *rq, struct task_struct *p)
143{
144 struct rt_prio_array *array = &rq->rt.active;
145
Peter Zijlstrafa717062008-01-25 21:08:27 +0100146 list_move_tail(&p->rt.run_list, array->queue + p->prio);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200147}
148
149static void
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +0200150yield_task_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200151{
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +0200152 requeue_task_rt(rq, rq->curr);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200153}
154
Gregory Haskinse7693a32008-01-25 21:08:09 +0100155#ifdef CONFIG_SMP
Gregory Haskins318e0892008-01-25 21:08:10 +0100156static int find_lowest_rq(struct task_struct *task);
157
Gregory Haskinse7693a32008-01-25 21:08:09 +0100158static int select_task_rq_rt(struct task_struct *p, int sync)
159{
Gregory Haskins318e0892008-01-25 21:08:10 +0100160 struct rq *rq = task_rq(p);
161
162 /*
Steven Rostedte1f47d82008-01-25 21:08:12 +0100163 * If the current task is an RT task, then
164 * try to see if we can wake this RT task up on another
165 * runqueue. Otherwise simply start this RT task
166 * on its current runqueue.
167 *
168 * We want to avoid overloading runqueues. Even if
169 * the RT task is of higher priority than the current RT task.
170 * RT tasks behave differently than other tasks. If
171 * one gets preempted, we try to push it off to another queue.
172 * So trying to keep a preempting RT task on the same
173 * cache hot CPU will force the running RT task to
174 * a cold CPU. So we waste all the cache for the lower
175 * RT task in hopes of saving some of a RT task
176 * that is just being woken and probably will have
177 * cold cache anyway.
Gregory Haskins318e0892008-01-25 21:08:10 +0100178 */
Gregory Haskins17b32792008-01-25 21:08:13 +0100179 if (unlikely(rt_task(rq->curr)) &&
180 (p->nr_cpus_allowed > 1)) {
Gregory Haskins318e0892008-01-25 21:08:10 +0100181 int cpu = find_lowest_rq(p);
182
183 return (cpu == -1) ? task_cpu(p) : cpu;
184 }
185
186 /*
187 * Otherwise, just let it ride on the affined RQ and the
188 * post-schedule router will push the preempted task away
189 */
Gregory Haskinse7693a32008-01-25 21:08:09 +0100190 return task_cpu(p);
191}
192#endif /* CONFIG_SMP */
193
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200194/*
195 * Preempt the current task with a newly woken task if needed:
196 */
197static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
198{
199 if (p->prio < rq->curr->prio)
200 resched_task(rq->curr);
201}
202
Ingo Molnarfb8d4722007-08-09 11:16:48 +0200203static struct task_struct *pick_next_task_rt(struct rq *rq)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200204{
205 struct rt_prio_array *array = &rq->rt.active;
206 struct task_struct *next;
207 struct list_head *queue;
208 int idx;
209
210 idx = sched_find_first_bit(array->bitmap);
211 if (idx >= MAX_RT_PRIO)
212 return NULL;
213
214 queue = array->queue + idx;
Peter Zijlstrafa717062008-01-25 21:08:27 +0100215 next = list_entry(queue->next, struct task_struct, rt.run_list);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200216
Ingo Molnard2819182007-08-09 11:16:47 +0200217 next->se.exec_start = rq->clock;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200218
219 return next;
220}
221
Ingo Molnar31ee5292007-08-09 11:16:49 +0200222static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200223{
Ingo Molnarf1e14ef2007-08-09 11:16:48 +0200224 update_curr_rt(rq);
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200225 p->se.exec_start = 0;
226}
227
Peter Williams681f3e62007-10-24 18:23:51 +0200228#ifdef CONFIG_SMP
Steven Rostedte8fa1362008-01-25 21:08:05 +0100229/* Only try algorithms three times */
230#define RT_MAX_TRIES 3
231
232static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
233static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
234
Steven Rostedtf65eda42008-01-25 21:08:07 +0100235static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
236{
237 if (!task_running(rq, p) &&
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100238 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
239 (p->nr_cpus_allowed > 1))
Steven Rostedtf65eda42008-01-25 21:08:07 +0100240 return 1;
241 return 0;
242}
243
Steven Rostedte8fa1362008-01-25 21:08:05 +0100244/* Return the second highest RT task, NULL otherwise */
Ingo Molnar79064fb2008-01-25 21:08:14 +0100245static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
Steven Rostedte8fa1362008-01-25 21:08:05 +0100246{
247 struct rt_prio_array *array = &rq->rt.active;
248 struct task_struct *next;
249 struct list_head *queue;
250 int idx;
251
Steven Rostedte8fa1362008-01-25 21:08:05 +0100252 if (likely(rq->rt.rt_nr_running < 2))
253 return NULL;
254
255 idx = sched_find_first_bit(array->bitmap);
256 if (unlikely(idx >= MAX_RT_PRIO)) {
257 WARN_ON(1); /* rt_nr_running is bad */
258 return NULL;
259 }
260
261 queue = array->queue + idx;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100262 BUG_ON(list_empty(queue));
263
Peter Zijlstrafa717062008-01-25 21:08:27 +0100264 next = list_entry(queue->next, struct task_struct, rt.run_list);
Steven Rostedtf65eda42008-01-25 21:08:07 +0100265 if (unlikely(pick_rt_task(rq, next, cpu)))
266 goto out;
Steven Rostedte8fa1362008-01-25 21:08:05 +0100267
268 if (queue->next->next != queue) {
269 /* same prio task */
Ingo Molnar79064fb2008-01-25 21:08:14 +0100270 next = list_entry(queue->next->next, struct task_struct,
Peter Zijlstrafa717062008-01-25 21:08:27 +0100271 rt.run_list);
Steven Rostedtf65eda42008-01-25 21:08:07 +0100272 if (pick_rt_task(rq, next, cpu))
273 goto out;
Steven Rostedte8fa1362008-01-25 21:08:05 +0100274 }
275
Steven Rostedtf65eda42008-01-25 21:08:07 +0100276 retry:
Steven Rostedte8fa1362008-01-25 21:08:05 +0100277 /* slower, but more flexible */
278 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
Steven Rostedtf65eda42008-01-25 21:08:07 +0100279 if (unlikely(idx >= MAX_RT_PRIO))
Steven Rostedte8fa1362008-01-25 21:08:05 +0100280 return NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +0100281
282 queue = array->queue + idx;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100283 BUG_ON(list_empty(queue));
Steven Rostedte8fa1362008-01-25 21:08:05 +0100284
Peter Zijlstrafa717062008-01-25 21:08:27 +0100285 list_for_each_entry(next, queue, rt.run_list) {
Steven Rostedtf65eda42008-01-25 21:08:07 +0100286 if (pick_rt_task(rq, next, cpu))
287 goto out;
288 }
289
290 goto retry;
291
292 out:
Steven Rostedte8fa1362008-01-25 21:08:05 +0100293 return next;
294}
295
296static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
297
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100298static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
Gregory Haskins07b40322008-01-25 21:08:10 +0100299{
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100300 int lowest_prio = -1;
Steven Rostedt610bf052008-01-25 21:08:13 +0100301 int lowest_cpu = -1;
Gregory Haskins06f90db2008-01-25 21:08:13 +0100302 int count = 0;
Steven Rostedt610bf052008-01-25 21:08:13 +0100303 int cpu;
Gregory Haskins07b40322008-01-25 21:08:10 +0100304
Gregory Haskins637f5082008-01-25 21:08:18 +0100305 cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
Gregory Haskins07b40322008-01-25 21:08:10 +0100306
307 /*
308 * Scan each rq for the lowest prio.
309 */
Steven Rostedt610bf052008-01-25 21:08:13 +0100310 for_each_cpu_mask(cpu, *lowest_mask) {
Gregory Haskins07b40322008-01-25 21:08:10 +0100311 struct rq *rq = cpu_rq(cpu);
312
Gregory Haskins07b40322008-01-25 21:08:10 +0100313 /* We look for lowest RT prio or non-rt CPU */
314 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
Steven Rostedt610bf052008-01-25 21:08:13 +0100315 /*
316 * if we already found a low RT queue
317 * and now we found this non-rt queue
318 * clear the mask and set our bit.
319 * Otherwise just return the queue as is
320 * and the count==1 will cause the algorithm
321 * to use the first bit found.
322 */
323 if (lowest_cpu != -1) {
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100324 cpus_clear(*lowest_mask);
Steven Rostedt610bf052008-01-25 21:08:13 +0100325 cpu_set(rq->cpu, *lowest_mask);
326 }
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100327 return 1;
Gregory Haskins07b40322008-01-25 21:08:10 +0100328 }
329
330 /* no locking for now */
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100331 if ((rq->rt.highest_prio > task->prio)
332 && (rq->rt.highest_prio >= lowest_prio)) {
333 if (rq->rt.highest_prio > lowest_prio) {
334 /* new low - clear old data */
335 lowest_prio = rq->rt.highest_prio;
Steven Rostedt610bf052008-01-25 21:08:13 +0100336 lowest_cpu = cpu;
337 count = 0;
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100338 }
Gregory Haskins06f90db2008-01-25 21:08:13 +0100339 count++;
Steven Rostedt610bf052008-01-25 21:08:13 +0100340 } else
341 cpu_clear(cpu, *lowest_mask);
342 }
343
344 /*
345 * Clear out all the set bits that represent
346 * runqueues that were of higher prio than
347 * the lowest_prio.
348 */
349 if (lowest_cpu > 0) {
350 /*
351 * Perhaps we could add another cpumask op to
352 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
353 * Then that could be optimized to use memset and such.
354 */
355 for_each_cpu_mask(cpu, *lowest_mask) {
356 if (cpu >= lowest_cpu)
357 break;
358 cpu_clear(cpu, *lowest_mask);
Gregory Haskins07b40322008-01-25 21:08:10 +0100359 }
360 }
361
Gregory Haskins06f90db2008-01-25 21:08:13 +0100362 return count;
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100363}
364
365static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
366{
367 int first;
368
369 /* "this_cpu" is cheaper to preempt than a remote processor */
370 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
371 return this_cpu;
372
373 first = first_cpu(*mask);
374 if (first != NR_CPUS)
375 return first;
376
377 return -1;
378}
379
380static int find_lowest_rq(struct task_struct *task)
381{
382 struct sched_domain *sd;
383 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
384 int this_cpu = smp_processor_id();
385 int cpu = task_cpu(task);
Gregory Haskins06f90db2008-01-25 21:08:13 +0100386 int count = find_lowest_cpus(task, lowest_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100387
Gregory Haskins06f90db2008-01-25 21:08:13 +0100388 if (!count)
389 return -1; /* No targets found */
390
391 /*
392 * There is no sense in performing an optimal search if only one
393 * target is found.
394 */
395 if (count == 1)
396 return first_cpu(*lowest_mask);
Gregory Haskins6e1254d2008-01-25 21:08:11 +0100397
398 /*
399 * At this point we have built a mask of cpus representing the
400 * lowest priority tasks in the system. Now we want to elect
401 * the best one based on our affinity and topology.
402 *
403 * We prioritize the last cpu that the task executed on since
404 * it is most likely cache-hot in that location.
405 */
406 if (cpu_isset(cpu, *lowest_mask))
407 return cpu;
408
409 /*
410 * Otherwise, we consult the sched_domains span maps to figure
411 * out which cpu is logically closest to our hot cache data.
412 */
413 if (this_cpu == cpu)
414 this_cpu = -1; /* Skip this_cpu opt if the same */
415
416 for_each_domain(cpu, sd) {
417 if (sd->flags & SD_WAKE_AFFINE) {
418 cpumask_t domain_mask;
419 int best_cpu;
420
421 cpus_and(domain_mask, sd->span, *lowest_mask);
422
423 best_cpu = pick_optimal_cpu(this_cpu,
424 &domain_mask);
425 if (best_cpu != -1)
426 return best_cpu;
427 }
428 }
429
430 /*
431 * And finally, if there were no matches within the domains
432 * just give the caller *something* to work with from the compatible
433 * locations.
434 */
435 return pick_optimal_cpu(this_cpu, lowest_mask);
Gregory Haskins07b40322008-01-25 21:08:10 +0100436}
437
Steven Rostedte8fa1362008-01-25 21:08:05 +0100438/* Will lock the rq it finds */
Ingo Molnar4df64c02008-01-25 21:08:15 +0100439static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +0100440{
441 struct rq *lowest_rq = NULL;
Steven Rostedte8fa1362008-01-25 21:08:05 +0100442 int tries;
Ingo Molnar4df64c02008-01-25 21:08:15 +0100443 int cpu;
Steven Rostedte8fa1362008-01-25 21:08:05 +0100444
445 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
Gregory Haskins07b40322008-01-25 21:08:10 +0100446 cpu = find_lowest_rq(task);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100447
Gregory Haskins2de0b462008-01-25 21:08:10 +0100448 if ((cpu == -1) || (cpu == rq->cpu))
Steven Rostedte8fa1362008-01-25 21:08:05 +0100449 break;
450
Gregory Haskins07b40322008-01-25 21:08:10 +0100451 lowest_rq = cpu_rq(cpu);
452
Steven Rostedte8fa1362008-01-25 21:08:05 +0100453 /* if the prio of this runqueue changed, try again */
Gregory Haskins07b40322008-01-25 21:08:10 +0100454 if (double_lock_balance(rq, lowest_rq)) {
Steven Rostedte8fa1362008-01-25 21:08:05 +0100455 /*
456 * We had to unlock the run queue. In
457 * the mean time, task could have
458 * migrated already or had its affinity changed.
459 * Also make sure that it wasn't scheduled on its rq.
460 */
Gregory Haskins07b40322008-01-25 21:08:10 +0100461 if (unlikely(task_rq(task) != rq ||
Ingo Molnar4df64c02008-01-25 21:08:15 +0100462 !cpu_isset(lowest_rq->cpu,
463 task->cpus_allowed) ||
Gregory Haskins07b40322008-01-25 21:08:10 +0100464 task_running(rq, task) ||
Steven Rostedte8fa1362008-01-25 21:08:05 +0100465 !task->se.on_rq)) {
Ingo Molnar4df64c02008-01-25 21:08:15 +0100466
Steven Rostedte8fa1362008-01-25 21:08:05 +0100467 spin_unlock(&lowest_rq->lock);
468 lowest_rq = NULL;
469 break;
470 }
471 }
472
473 /* If this rq is still suitable use it. */
474 if (lowest_rq->rt.highest_prio > task->prio)
475 break;
476
477 /* try again */
478 spin_unlock(&lowest_rq->lock);
479 lowest_rq = NULL;
480 }
481
482 return lowest_rq;
483}
484
485/*
486 * If the current CPU has more than one RT task, see if the non
487 * running task can migrate over to a CPU that is running a task
488 * of lesser priority.
489 */
Gregory Haskins697f0a42008-01-25 21:08:09 +0100490static int push_rt_task(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +0100491{
492 struct task_struct *next_task;
493 struct rq *lowest_rq;
494 int ret = 0;
495 int paranoid = RT_MAX_TRIES;
496
Gregory Haskinsa22d7fc12008-01-25 21:08:12 +0100497 if (!rq->rt.overloaded)
498 return 0;
499
Gregory Haskins697f0a42008-01-25 21:08:09 +0100500 next_task = pick_next_highest_task_rt(rq, -1);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100501 if (!next_task)
502 return 0;
503
504 retry:
Gregory Haskins697f0a42008-01-25 21:08:09 +0100505 if (unlikely(next_task == rq->curr)) {
Steven Rostedtf65eda42008-01-25 21:08:07 +0100506 WARN_ON(1);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100507 return 0;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100508 }
Steven Rostedte8fa1362008-01-25 21:08:05 +0100509
510 /*
511 * It's possible that the next_task slipped in of
512 * higher priority than current. If that's the case
513 * just reschedule current.
514 */
Gregory Haskins697f0a42008-01-25 21:08:09 +0100515 if (unlikely(next_task->prio < rq->curr->prio)) {
516 resched_task(rq->curr);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100517 return 0;
518 }
519
Gregory Haskins697f0a42008-01-25 21:08:09 +0100520 /* We might release rq lock */
Steven Rostedte8fa1362008-01-25 21:08:05 +0100521 get_task_struct(next_task);
522
523 /* find_lock_lowest_rq locks the rq if found */
Gregory Haskins697f0a42008-01-25 21:08:09 +0100524 lowest_rq = find_lock_lowest_rq(next_task, rq);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100525 if (!lowest_rq) {
526 struct task_struct *task;
527 /*
Gregory Haskins697f0a42008-01-25 21:08:09 +0100528 * find lock_lowest_rq releases rq->lock
Steven Rostedte8fa1362008-01-25 21:08:05 +0100529 * so it is possible that next_task has changed.
530 * If it has, then try again.
531 */
Gregory Haskins697f0a42008-01-25 21:08:09 +0100532 task = pick_next_highest_task_rt(rq, -1);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100533 if (unlikely(task != next_task) && task && paranoid--) {
534 put_task_struct(next_task);
535 next_task = task;
536 goto retry;
537 }
538 goto out;
539 }
540
Gregory Haskins697f0a42008-01-25 21:08:09 +0100541 deactivate_task(rq, next_task, 0);
Steven Rostedte8fa1362008-01-25 21:08:05 +0100542 set_task_cpu(next_task, lowest_rq->cpu);
543 activate_task(lowest_rq, next_task, 0);
544
545 resched_task(lowest_rq->curr);
546
547 spin_unlock(&lowest_rq->lock);
548
549 ret = 1;
550out:
551 put_task_struct(next_task);
552
553 return ret;
554}
555
556/*
557 * TODO: Currently we just use the second highest prio task on
558 * the queue, and stop when it can't migrate (or there's
559 * no more RT tasks). There may be a case where a lower
560 * priority RT task has a different affinity than the
561 * higher RT task. In this case the lower RT task could
562 * possibly be able to migrate where as the higher priority
563 * RT task could not. We currently ignore this issue.
564 * Enhancements are welcome!
565 */
566static void push_rt_tasks(struct rq *rq)
567{
568 /* push_rt_task will return true if it moved an RT */
569 while (push_rt_task(rq))
570 ;
571}
572
Steven Rostedtf65eda42008-01-25 21:08:07 +0100573static int pull_rt_task(struct rq *this_rq)
574{
Ingo Molnar80bf3172008-01-25 21:08:17 +0100575 int this_cpu = this_rq->cpu, ret = 0, cpu;
576 struct task_struct *p, *next;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100577 struct rq *src_rq;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100578
Gregory Haskins637f5082008-01-25 21:08:18 +0100579 if (likely(!rt_overloaded(this_rq)))
Steven Rostedtf65eda42008-01-25 21:08:07 +0100580 return 0;
581
582 next = pick_next_task_rt(this_rq);
583
Gregory Haskins637f5082008-01-25 21:08:18 +0100584 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
Steven Rostedtf65eda42008-01-25 21:08:07 +0100585 if (this_cpu == cpu)
586 continue;
587
588 src_rq = cpu_rq(cpu);
Steven Rostedtf65eda42008-01-25 21:08:07 +0100589 /*
590 * We can potentially drop this_rq's lock in
591 * double_lock_balance, and another CPU could
592 * steal our next task - hence we must cause
593 * the caller to recalculate the next task
594 * in that case:
595 */
596 if (double_lock_balance(this_rq, src_rq)) {
597 struct task_struct *old_next = next;
Ingo Molnar80bf3172008-01-25 21:08:17 +0100598
Steven Rostedtf65eda42008-01-25 21:08:07 +0100599 next = pick_next_task_rt(this_rq);
600 if (next != old_next)
601 ret = 1;
602 }
603
604 /*
605 * Are there still pullable RT tasks?
606 */
607 if (src_rq->rt.rt_nr_running <= 1) {
608 spin_unlock(&src_rq->lock);
609 continue;
610 }
611
Steven Rostedtf65eda42008-01-25 21:08:07 +0100612 p = pick_next_highest_task_rt(src_rq, this_cpu);
613
614 /*
615 * Do we have an RT task that preempts
616 * the to-be-scheduled task?
617 */
618 if (p && (!next || (p->prio < next->prio))) {
619 WARN_ON(p == src_rq->curr);
620 WARN_ON(!p->se.on_rq);
621
622 /*
623 * There's a chance that p is higher in priority
624 * than what's currently running on its cpu.
625 * This is just that p is wakeing up and hasn't
626 * had a chance to schedule. We only pull
627 * p if it is lower in priority than the
628 * current task on the run queue or
629 * this_rq next task is lower in prio than
630 * the current task on that rq.
631 */
632 if (p->prio < src_rq->curr->prio ||
633 (next && next->prio < src_rq->curr->prio))
Ingo Molnar80bf3172008-01-25 21:08:17 +0100634 goto out;
Steven Rostedtf65eda42008-01-25 21:08:07 +0100635
636 ret = 1;
637
638 deactivate_task(src_rq, p, 0);
639 set_task_cpu(p, this_cpu);
640 activate_task(this_rq, p, 0);
641 /*
642 * We continue with the search, just in
643 * case there's an even higher prio task
644 * in another runqueue. (low likelyhood
645 * but possible)
Ingo Molnar80bf3172008-01-25 21:08:17 +0100646 *
Steven Rostedtf65eda42008-01-25 21:08:07 +0100647 * Update next so that we won't pick a task
648 * on another cpu with a priority lower (or equal)
649 * than the one we just picked.
650 */
651 next = p;
652
653 }
Ingo Molnar80bf3172008-01-25 21:08:17 +0100654 out:
Steven Rostedtf65eda42008-01-25 21:08:07 +0100655 spin_unlock(&src_rq->lock);
656 }
657
658 return ret;
659}
660
Steven Rostedt9a897c52008-01-25 21:08:22 +0100661static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
Steven Rostedtf65eda42008-01-25 21:08:07 +0100662{
663 /* Try to pull RT tasks here if we lower this rq's prio */
Ingo Molnar7f51f292008-01-25 21:08:17 +0100664 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
Steven Rostedtf65eda42008-01-25 21:08:07 +0100665 pull_rt_task(rq);
666}
667
Steven Rostedt9a897c52008-01-25 21:08:22 +0100668static void post_schedule_rt(struct rq *rq)
Steven Rostedte8fa1362008-01-25 21:08:05 +0100669{
670 /*
671 * If we have more than one rt_task queued, then
672 * see if we can push the other rt_tasks off to other CPUS.
673 * Note we may release the rq lock, and since
674 * the lock was owned by prev, we need to release it
675 * first via finish_lock_switch and then reaquire it here.
676 */
Gregory Haskinsa22d7fc12008-01-25 21:08:12 +0100677 if (unlikely(rq->rt.overloaded)) {
Steven Rostedte8fa1362008-01-25 21:08:05 +0100678 spin_lock_irq(&rq->lock);
679 push_rt_tasks(rq);
680 spin_unlock_irq(&rq->lock);
681 }
682}
683
Steven Rostedt4642daf2008-01-25 21:08:07 +0100684
Steven Rostedt9a897c52008-01-25 21:08:22 +0100685static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
Steven Rostedt4642daf2008-01-25 21:08:07 +0100686{
Steven Rostedt9a897c52008-01-25 21:08:22 +0100687 if (!task_running(rq, p) &&
Gregory Haskinsa22d7fc12008-01-25 21:08:12 +0100688 (p->prio >= rq->rt.highest_prio) &&
689 rq->rt.overloaded)
Steven Rostedt4642daf2008-01-25 21:08:07 +0100690 push_rt_tasks(rq);
691}
692
Peter Williams43010652007-08-09 11:16:46 +0200693static unsigned long
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200694load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williamse1d14842007-10-24 18:23:51 +0200695 unsigned long max_load_move,
696 struct sched_domain *sd, enum cpu_idle_type idle,
697 int *all_pinned, int *this_best_prio)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200698{
Steven Rostedtc7a1e462008-01-25 21:08:07 +0100699 /* don't touch RT tasks */
700 return 0;
Peter Williamse1d14842007-10-24 18:23:51 +0200701}
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200702
Peter Williamse1d14842007-10-24 18:23:51 +0200703static int
704move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
705 struct sched_domain *sd, enum cpu_idle_type idle)
706{
Steven Rostedtc7a1e462008-01-25 21:08:07 +0100707 /* don't touch RT tasks */
708 return 0;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200709}
Ingo Molnardeeeccd2008-01-25 21:08:15 +0100710
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100711static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
712{
713 int weight = cpus_weight(*new_mask);
714
715 BUG_ON(!rt_task(p));
716
717 /*
718 * Update the migration status of the RQ if we have an RT task
719 * which is running AND changing its weight value.
720 */
721 if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
722 struct rq *rq = task_rq(p);
723
Ingo Molnardeeeccd2008-01-25 21:08:15 +0100724 if ((p->nr_cpus_allowed <= 1) && (weight > 1)) {
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100725 rq->rt.rt_nr_migratory++;
Ingo Molnardeeeccd2008-01-25 21:08:15 +0100726 } else if ((p->nr_cpus_allowed > 1) && (weight <= 1)) {
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100727 BUG_ON(!rq->rt.rt_nr_migratory);
728 rq->rt.rt_nr_migratory--;
729 }
730
731 update_rt_migration(rq);
732 }
733
734 p->cpus_allowed = *new_mask;
735 p->nr_cpus_allowed = weight;
736}
Ingo Molnardeeeccd2008-01-25 21:08:15 +0100737
Ingo Molnarbdd7c812008-01-25 21:08:18 +0100738/* Assumes rq->lock is held */
739static void join_domain_rt(struct rq *rq)
740{
741 if (rq->rt.overloaded)
742 rt_set_overload(rq);
743}
744
745/* Assumes rq->lock is held */
746static void leave_domain_rt(struct rq *rq)
747{
748 if (rq->rt.overloaded)
749 rt_clear_overload(rq);
750}
Steven Rostedtcb469842008-01-25 21:08:22 +0100751
752/*
753 * When switch from the rt queue, we bring ourselves to a position
754 * that we might want to pull RT tasks from other runqueues.
755 */
756static void switched_from_rt(struct rq *rq, struct task_struct *p,
757 int running)
758{
759 /*
760 * If there are other RT tasks then we will reschedule
761 * and the scheduling of the other RT tasks will handle
762 * the balancing. But if we are the last RT task
763 * we may need to handle the pulling of RT tasks
764 * now.
765 */
766 if (!rq->rt.rt_nr_running)
767 pull_rt_task(rq);
768}
Steven Rostedte8fa1362008-01-25 21:08:05 +0100769#endif /* CONFIG_SMP */
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200770
Steven Rostedtcb469842008-01-25 21:08:22 +0100771/*
772 * When switching a task to RT, we may overload the runqueue
773 * with RT tasks. In this case we try to push them off to
774 * other runqueues.
775 */
776static void switched_to_rt(struct rq *rq, struct task_struct *p,
777 int running)
778{
779 int check_resched = 1;
780
781 /*
782 * If we are already running, then there's nothing
783 * that needs to be done. But if we are not running
784 * we may need to preempt the current running task.
785 * If that current running task is also an RT task
786 * then see if we can move to another run queue.
787 */
788 if (!running) {
789#ifdef CONFIG_SMP
790 if (rq->rt.overloaded && push_rt_task(rq) &&
791 /* Don't resched if we changed runqueues */
792 rq != task_rq(p))
793 check_resched = 0;
794#endif /* CONFIG_SMP */
795 if (check_resched && p->prio < rq->curr->prio)
796 resched_task(rq->curr);
797 }
798}
799
800/*
801 * Priority of the task has changed. This may cause
802 * us to initiate a push or pull.
803 */
804static void prio_changed_rt(struct rq *rq, struct task_struct *p,
805 int oldprio, int running)
806{
807 if (running) {
808#ifdef CONFIG_SMP
809 /*
810 * If our priority decreases while running, we
811 * may need to pull tasks to this runqueue.
812 */
813 if (oldprio < p->prio)
814 pull_rt_task(rq);
815 /*
816 * If there's a higher priority task waiting to run
817 * then reschedule.
818 */
819 if (p->prio > rq->rt.highest_prio)
820 resched_task(p);
821#else
822 /* For UP simply resched on drop of prio */
823 if (oldprio < p->prio)
824 resched_task(p);
825#endif /* CONFIG_SMP */
826 } else {
827 /*
828 * This task is not running, but if it is
829 * greater than the current running task
830 * then reschedule.
831 */
832 if (p->prio < rq->curr->prio)
833 resched_task(rq->curr);
834 }
835}
836
837
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200838static void task_tick_rt(struct rq *rq, struct task_struct *p)
839{
Peter Zijlstra67e2be02007-12-20 15:01:17 +0100840 update_curr_rt(rq);
841
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200842 /*
843 * RR tasks need a special form of timeslice management.
844 * FIFO tasks have no timeslices.
845 */
846 if (p->policy != SCHED_RR)
847 return;
848
Peter Zijlstrafa717062008-01-25 21:08:27 +0100849 if (--p->rt.time_slice)
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200850 return;
851
Peter Zijlstrafa717062008-01-25 21:08:27 +0100852 p->rt.time_slice = DEF_TIMESLICE;
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200853
Dmitry Adamushko98fbc792007-08-24 20:39:10 +0200854 /*
855 * Requeue to the end of queue if we are not the only element
856 * on the queue:
857 */
Peter Zijlstrafa717062008-01-25 21:08:27 +0100858 if (p->rt.run_list.prev != p->rt.run_list.next) {
Dmitry Adamushko98fbc792007-08-24 20:39:10 +0200859 requeue_task_rt(rq, p);
860 set_tsk_need_resched(p);
861 }
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200862}
863
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +0200864static void set_curr_task_rt(struct rq *rq)
865{
866 struct task_struct *p = rq->curr;
867
868 p->se.exec_start = rq->clock;
869}
870
Ingo Molnar5522d5d2007-10-15 17:00:12 +0200871const struct sched_class rt_sched_class = {
872 .next = &fair_sched_class,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200873 .enqueue_task = enqueue_task_rt,
874 .dequeue_task = dequeue_task_rt,
875 .yield_task = yield_task_rt,
Gregory Haskinse7693a32008-01-25 21:08:09 +0100876#ifdef CONFIG_SMP
877 .select_task_rq = select_task_rq_rt,
878#endif /* CONFIG_SMP */
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200879
880 .check_preempt_curr = check_preempt_curr_rt,
881
882 .pick_next_task = pick_next_task_rt,
883 .put_prev_task = put_prev_task_rt,
884
Peter Williams681f3e62007-10-24 18:23:51 +0200885#ifdef CONFIG_SMP
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200886 .load_balance = load_balance_rt,
Peter Williamse1d14842007-10-24 18:23:51 +0200887 .move_one_task = move_one_task_rt,
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100888 .set_cpus_allowed = set_cpus_allowed_rt,
Ingo Molnarbdd7c812008-01-25 21:08:18 +0100889 .join_domain = join_domain_rt,
890 .leave_domain = leave_domain_rt,
Steven Rostedt9a897c52008-01-25 21:08:22 +0100891 .pre_schedule = pre_schedule_rt,
892 .post_schedule = post_schedule_rt,
893 .task_wake_up = task_wake_up_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +0100894 .switched_from = switched_from_rt,
Peter Williams681f3e62007-10-24 18:23:51 +0200895#endif
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200896
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +0200897 .set_curr_task = set_curr_task_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200898 .task_tick = task_tick_rt,
Steven Rostedtcb469842008-01-25 21:08:22 +0100899
900 .prio_changed = prio_changed_rt,
901 .switched_to = switched_to_rt,
Ingo Molnarbb44e5d2007-07-09 18:51:58 +0200902};