]>
Commit | Line | Data |
---|---|---|
bb44e5d1 IM |
1 | /* |
2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR | |
3 | * policies) | |
4 | */ | |
5 | ||
4fd29176 SR |
6 | #ifdef CONFIG_SMP |
7 | static cpumask_t rt_overload_mask; | |
8 | static atomic_t rto_count; | |
9 | static inline int rt_overloaded(void) | |
10 | { | |
11 | return atomic_read(&rto_count); | |
12 | } | |
13 | static inline cpumask_t *rt_overload(void) | |
14 | { | |
15 | return &rt_overload_mask; | |
16 | } | |
17 | static inline void rt_set_overload(struct rq *rq) | |
18 | { | |
19 | cpu_set(rq->cpu, rt_overload_mask); | |
20 | /* | |
21 | * Make sure the mask is visible before we set | |
22 | * the overload count. That is checked to determine | |
23 | * if we should look at the mask. It would be a shame | |
24 | * if we looked at the mask, but the mask was not | |
25 | * updated yet. | |
26 | */ | |
27 | wmb(); | |
28 | atomic_inc(&rto_count); | |
29 | } | |
30 | static inline void rt_clear_overload(struct rq *rq) | |
31 | { | |
32 | /* the order here really doesn't matter */ | |
33 | atomic_dec(&rto_count); | |
34 | cpu_clear(rq->cpu, rt_overload_mask); | |
35 | } | |
73fe6aae GH |
36 | |
37 | static void update_rt_migration(struct rq *rq) | |
38 | { | |
39 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) | |
40 | rt_set_overload(rq); | |
41 | else | |
42 | rt_clear_overload(rq); | |
43 | } | |
4fd29176 SR |
44 | #endif /* CONFIG_SMP */ |
45 | ||
bb44e5d1 IM |
46 | /* |
47 | * Update the current task's runtime statistics. Skip current tasks that | |
48 | * are not in our scheduling class. | |
49 | */ | |
a9957449 | 50 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
51 | { |
52 | struct task_struct *curr = rq->curr; | |
53 | u64 delta_exec; | |
54 | ||
55 | if (!task_has_rt_policy(curr)) | |
56 | return; | |
57 | ||
d281918d | 58 | delta_exec = rq->clock - curr->se.exec_start; |
bb44e5d1 IM |
59 | if (unlikely((s64)delta_exec < 0)) |
60 | delta_exec = 0; | |
6cfb0d5d IM |
61 | |
62 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); | |
bb44e5d1 IM |
63 | |
64 | curr->se.sum_exec_runtime += delta_exec; | |
d281918d | 65 | curr->se.exec_start = rq->clock; |
d842de87 | 66 | cpuacct_charge(curr, delta_exec); |
bb44e5d1 IM |
67 | } |
68 | ||
63489e45 SR |
69 | static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq) |
70 | { | |
71 | WARN_ON(!rt_task(p)); | |
72 | rq->rt.rt_nr_running++; | |
764a9d6f SR |
73 | #ifdef CONFIG_SMP |
74 | if (p->prio < rq->rt.highest_prio) | |
75 | rq->rt.highest_prio = p->prio; | |
73fe6aae GH |
76 | if (p->nr_cpus_allowed > 1) |
77 | rq->rt.rt_nr_migratory++; | |
78 | ||
79 | update_rt_migration(rq); | |
764a9d6f | 80 | #endif /* CONFIG_SMP */ |
63489e45 SR |
81 | } |
82 | ||
83 | static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq) | |
84 | { | |
85 | WARN_ON(!rt_task(p)); | |
86 | WARN_ON(!rq->rt.rt_nr_running); | |
87 | rq->rt.rt_nr_running--; | |
764a9d6f SR |
88 | #ifdef CONFIG_SMP |
89 | if (rq->rt.rt_nr_running) { | |
90 | struct rt_prio_array *array; | |
91 | ||
92 | WARN_ON(p->prio < rq->rt.highest_prio); | |
93 | if (p->prio == rq->rt.highest_prio) { | |
94 | /* recalculate */ | |
95 | array = &rq->rt.active; | |
96 | rq->rt.highest_prio = | |
97 | sched_find_first_bit(array->bitmap); | |
98 | } /* otherwise leave rq->highest prio alone */ | |
99 | } else | |
100 | rq->rt.highest_prio = MAX_RT_PRIO; | |
73fe6aae GH |
101 | if (p->nr_cpus_allowed > 1) |
102 | rq->rt.rt_nr_migratory--; | |
103 | ||
104 | update_rt_migration(rq); | |
764a9d6f | 105 | #endif /* CONFIG_SMP */ |
63489e45 SR |
106 | } |
107 | ||
fd390f6a | 108 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
bb44e5d1 IM |
109 | { |
110 | struct rt_prio_array *array = &rq->rt.active; | |
111 | ||
112 | list_add_tail(&p->run_list, array->queue + p->prio); | |
113 | __set_bit(p->prio, array->bitmap); | |
58e2d4ca | 114 | inc_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
115 | |
116 | inc_rt_tasks(p, rq); | |
bb44e5d1 IM |
117 | } |
118 | ||
119 | /* | |
120 | * Adding/removing a task to/from a priority array: | |
121 | */ | |
f02231e5 | 122 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
bb44e5d1 IM |
123 | { |
124 | struct rt_prio_array *array = &rq->rt.active; | |
125 | ||
f1e14ef6 | 126 | update_curr_rt(rq); |
bb44e5d1 IM |
127 | |
128 | list_del(&p->run_list); | |
129 | if (list_empty(array->queue + p->prio)) | |
130 | __clear_bit(p->prio, array->bitmap); | |
58e2d4ca | 131 | dec_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
132 | |
133 | dec_rt_tasks(p, rq); | |
bb44e5d1 IM |
134 | } |
135 | ||
136 | /* | |
137 | * Put task to the end of the run list without the overhead of dequeue | |
138 | * followed by enqueue. | |
139 | */ | |
140 | static void requeue_task_rt(struct rq *rq, struct task_struct *p) | |
141 | { | |
142 | struct rt_prio_array *array = &rq->rt.active; | |
143 | ||
144 | list_move_tail(&p->run_list, array->queue + p->prio); | |
145 | } | |
146 | ||
147 | static void | |
4530d7ab | 148 | yield_task_rt(struct rq *rq) |
bb44e5d1 | 149 | { |
4530d7ab | 150 | requeue_task_rt(rq, rq->curr); |
bb44e5d1 IM |
151 | } |
152 | ||
e7693a36 | 153 | #ifdef CONFIG_SMP |
318e0893 GH |
154 | static int find_lowest_rq(struct task_struct *task); |
155 | ||
e7693a36 GH |
156 | static int select_task_rq_rt(struct task_struct *p, int sync) |
157 | { | |
318e0893 GH |
158 | struct rq *rq = task_rq(p); |
159 | ||
160 | /* | |
161 | * If the task will not preempt the RQ, try to find a better RQ | |
162 | * before we even activate the task | |
163 | */ | |
164 | if ((p->prio >= rq->rt.highest_prio) | |
165 | && (p->nr_cpus_allowed > 1)) { | |
166 | int cpu = find_lowest_rq(p); | |
167 | ||
168 | return (cpu == -1) ? task_cpu(p) : cpu; | |
169 | } | |
170 | ||
171 | /* | |
172 | * Otherwise, just let it ride on the affined RQ and the | |
173 | * post-schedule router will push the preempted task away | |
174 | */ | |
e7693a36 GH |
175 | return task_cpu(p); |
176 | } | |
177 | #endif /* CONFIG_SMP */ | |
178 | ||
bb44e5d1 IM |
179 | /* |
180 | * Preempt the current task with a newly woken task if needed: | |
181 | */ | |
182 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p) | |
183 | { | |
184 | if (p->prio < rq->curr->prio) | |
185 | resched_task(rq->curr); | |
186 | } | |
187 | ||
fb8d4724 | 188 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
bb44e5d1 IM |
189 | { |
190 | struct rt_prio_array *array = &rq->rt.active; | |
191 | struct task_struct *next; | |
192 | struct list_head *queue; | |
193 | int idx; | |
194 | ||
195 | idx = sched_find_first_bit(array->bitmap); | |
196 | if (idx >= MAX_RT_PRIO) | |
197 | return NULL; | |
198 | ||
199 | queue = array->queue + idx; | |
200 | next = list_entry(queue->next, struct task_struct, run_list); | |
201 | ||
d281918d | 202 | next->se.exec_start = rq->clock; |
bb44e5d1 IM |
203 | |
204 | return next; | |
205 | } | |
206 | ||
31ee529c | 207 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 208 | { |
f1e14ef6 | 209 | update_curr_rt(rq); |
bb44e5d1 IM |
210 | p->se.exec_start = 0; |
211 | } | |
212 | ||
681f3e68 | 213 | #ifdef CONFIG_SMP |
e8fa1362 SR |
214 | /* Only try algorithms three times */ |
215 | #define RT_MAX_TRIES 3 | |
216 | ||
217 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); | |
218 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); | |
219 | ||
f65eda4f SR |
220 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
221 | { | |
222 | if (!task_running(rq, p) && | |
73fe6aae GH |
223 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
224 | (p->nr_cpus_allowed > 1)) | |
f65eda4f SR |
225 | return 1; |
226 | return 0; | |
227 | } | |
228 | ||
e8fa1362 | 229 | /* Return the second highest RT task, NULL otherwise */ |
f65eda4f SR |
230 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, |
231 | int cpu) | |
e8fa1362 SR |
232 | { |
233 | struct rt_prio_array *array = &rq->rt.active; | |
234 | struct task_struct *next; | |
235 | struct list_head *queue; | |
236 | int idx; | |
237 | ||
238 | assert_spin_locked(&rq->lock); | |
239 | ||
240 | if (likely(rq->rt.rt_nr_running < 2)) | |
241 | return NULL; | |
242 | ||
243 | idx = sched_find_first_bit(array->bitmap); | |
244 | if (unlikely(idx >= MAX_RT_PRIO)) { | |
245 | WARN_ON(1); /* rt_nr_running is bad */ | |
246 | return NULL; | |
247 | } | |
248 | ||
249 | queue = array->queue + idx; | |
f65eda4f SR |
250 | BUG_ON(list_empty(queue)); |
251 | ||
e8fa1362 | 252 | next = list_entry(queue->next, struct task_struct, run_list); |
f65eda4f SR |
253 | if (unlikely(pick_rt_task(rq, next, cpu))) |
254 | goto out; | |
e8fa1362 SR |
255 | |
256 | if (queue->next->next != queue) { | |
257 | /* same prio task */ | |
258 | next = list_entry(queue->next->next, struct task_struct, run_list); | |
f65eda4f SR |
259 | if (pick_rt_task(rq, next, cpu)) |
260 | goto out; | |
e8fa1362 SR |
261 | } |
262 | ||
f65eda4f | 263 | retry: |
e8fa1362 SR |
264 | /* slower, but more flexible */ |
265 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); | |
f65eda4f | 266 | if (unlikely(idx >= MAX_RT_PRIO)) |
e8fa1362 | 267 | return NULL; |
e8fa1362 SR |
268 | |
269 | queue = array->queue + idx; | |
f65eda4f SR |
270 | BUG_ON(list_empty(queue)); |
271 | ||
272 | list_for_each_entry(next, queue, run_list) { | |
273 | if (pick_rt_task(rq, next, cpu)) | |
274 | goto out; | |
275 | } | |
276 | ||
277 | goto retry; | |
e8fa1362 | 278 | |
f65eda4f | 279 | out: |
e8fa1362 SR |
280 | return next; |
281 | } | |
282 | ||
283 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); | |
6e1254d2 | 284 | static DEFINE_PER_CPU(cpumask_t, valid_cpu_mask); |
e8fa1362 | 285 | |
6e1254d2 | 286 | static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask) |
e8fa1362 | 287 | { |
6e1254d2 GH |
288 | int cpu; |
289 | cpumask_t *valid_mask = &__get_cpu_var(valid_cpu_mask); | |
290 | int lowest_prio = -1; | |
291 | int ret = 0; | |
e8fa1362 | 292 | |
6e1254d2 GH |
293 | cpus_clear(*lowest_mask); |
294 | cpus_and(*valid_mask, cpu_online_map, task->cpus_allowed); | |
e8fa1362 | 295 | |
07b4032c GH |
296 | /* |
297 | * Scan each rq for the lowest prio. | |
298 | */ | |
6e1254d2 | 299 | for_each_cpu_mask(cpu, *valid_mask) { |
07b4032c | 300 | struct rq *rq = cpu_rq(cpu); |
e8fa1362 | 301 | |
07b4032c GH |
302 | /* We look for lowest RT prio or non-rt CPU */ |
303 | if (rq->rt.highest_prio >= MAX_RT_PRIO) { | |
6e1254d2 GH |
304 | if (ret) |
305 | cpus_clear(*lowest_mask); | |
306 | cpu_set(rq->cpu, *lowest_mask); | |
307 | return 1; | |
07b4032c GH |
308 | } |
309 | ||
310 | /* no locking for now */ | |
6e1254d2 GH |
311 | if ((rq->rt.highest_prio > task->prio) |
312 | && (rq->rt.highest_prio >= lowest_prio)) { | |
313 | if (rq->rt.highest_prio > lowest_prio) { | |
314 | /* new low - clear old data */ | |
315 | lowest_prio = rq->rt.highest_prio; | |
316 | cpus_clear(*lowest_mask); | |
317 | } | |
318 | cpu_set(rq->cpu, *lowest_mask); | |
319 | ret = 1; | |
e8fa1362 | 320 | } |
07b4032c GH |
321 | } |
322 | ||
6e1254d2 GH |
323 | return ret; |
324 | } | |
325 | ||
326 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) | |
327 | { | |
328 | int first; | |
329 | ||
330 | /* "this_cpu" is cheaper to preempt than a remote processor */ | |
331 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) | |
332 | return this_cpu; | |
333 | ||
334 | first = first_cpu(*mask); | |
335 | if (first != NR_CPUS) | |
336 | return first; | |
337 | ||
338 | return -1; | |
339 | } | |
340 | ||
341 | static int find_lowest_rq(struct task_struct *task) | |
342 | { | |
343 | struct sched_domain *sd; | |
344 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); | |
345 | int this_cpu = smp_processor_id(); | |
346 | int cpu = task_cpu(task); | |
347 | ||
348 | if (!find_lowest_cpus(task, lowest_mask)) | |
349 | return -1; | |
350 | ||
351 | /* | |
352 | * At this point we have built a mask of cpus representing the | |
353 | * lowest priority tasks in the system. Now we want to elect | |
354 | * the best one based on our affinity and topology. | |
355 | * | |
356 | * We prioritize the last cpu that the task executed on since | |
357 | * it is most likely cache-hot in that location. | |
358 | */ | |
359 | if (cpu_isset(cpu, *lowest_mask)) | |
360 | return cpu; | |
361 | ||
362 | /* | |
363 | * Otherwise, we consult the sched_domains span maps to figure | |
364 | * out which cpu is logically closest to our hot cache data. | |
365 | */ | |
366 | if (this_cpu == cpu) | |
367 | this_cpu = -1; /* Skip this_cpu opt if the same */ | |
368 | ||
369 | for_each_domain(cpu, sd) { | |
370 | if (sd->flags & SD_WAKE_AFFINE) { | |
371 | cpumask_t domain_mask; | |
372 | int best_cpu; | |
373 | ||
374 | cpus_and(domain_mask, sd->span, *lowest_mask); | |
375 | ||
376 | best_cpu = pick_optimal_cpu(this_cpu, | |
377 | &domain_mask); | |
378 | if (best_cpu != -1) | |
379 | return best_cpu; | |
380 | } | |
381 | } | |
382 | ||
383 | /* | |
384 | * And finally, if there were no matches within the domains | |
385 | * just give the caller *something* to work with from the compatible | |
386 | * locations. | |
387 | */ | |
388 | return pick_optimal_cpu(this_cpu, lowest_mask); | |
07b4032c GH |
389 | } |
390 | ||
391 | /* Will lock the rq it finds */ | |
392 | static struct rq *find_lock_lowest_rq(struct task_struct *task, | |
393 | struct rq *rq) | |
394 | { | |
395 | struct rq *lowest_rq = NULL; | |
396 | int cpu; | |
397 | int tries; | |
e8fa1362 | 398 | |
07b4032c GH |
399 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
400 | cpu = find_lowest_rq(task); | |
401 | ||
2de0b463 | 402 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
403 | break; |
404 | ||
07b4032c GH |
405 | lowest_rq = cpu_rq(cpu); |
406 | ||
e8fa1362 | 407 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 408 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
409 | /* |
410 | * We had to unlock the run queue. In | |
411 | * the mean time, task could have | |
412 | * migrated already or had its affinity changed. | |
413 | * Also make sure that it wasn't scheduled on its rq. | |
414 | */ | |
07b4032c | 415 | if (unlikely(task_rq(task) != rq || |
e8fa1362 | 416 | !cpu_isset(lowest_rq->cpu, task->cpus_allowed) || |
07b4032c | 417 | task_running(rq, task) || |
e8fa1362 SR |
418 | !task->se.on_rq)) { |
419 | spin_unlock(&lowest_rq->lock); | |
420 | lowest_rq = NULL; | |
421 | break; | |
422 | } | |
423 | } | |
424 | ||
425 | /* If this rq is still suitable use it. */ | |
426 | if (lowest_rq->rt.highest_prio > task->prio) | |
427 | break; | |
428 | ||
429 | /* try again */ | |
430 | spin_unlock(&lowest_rq->lock); | |
431 | lowest_rq = NULL; | |
432 | } | |
433 | ||
434 | return lowest_rq; | |
435 | } | |
436 | ||
437 | /* | |
438 | * If the current CPU has more than one RT task, see if the non | |
439 | * running task can migrate over to a CPU that is running a task | |
440 | * of lesser priority. | |
441 | */ | |
697f0a48 | 442 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
443 | { |
444 | struct task_struct *next_task; | |
445 | struct rq *lowest_rq; | |
446 | int ret = 0; | |
447 | int paranoid = RT_MAX_TRIES; | |
448 | ||
697f0a48 | 449 | assert_spin_locked(&rq->lock); |
e8fa1362 | 450 | |
697f0a48 | 451 | next_task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
452 | if (!next_task) |
453 | return 0; | |
454 | ||
455 | retry: | |
697f0a48 | 456 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 457 | WARN_ON(1); |
e8fa1362 | 458 | return 0; |
f65eda4f | 459 | } |
e8fa1362 SR |
460 | |
461 | /* | |
462 | * It's possible that the next_task slipped in of | |
463 | * higher priority than current. If that's the case | |
464 | * just reschedule current. | |
465 | */ | |
697f0a48 GH |
466 | if (unlikely(next_task->prio < rq->curr->prio)) { |
467 | resched_task(rq->curr); | |
e8fa1362 SR |
468 | return 0; |
469 | } | |
470 | ||
697f0a48 | 471 | /* We might release rq lock */ |
e8fa1362 SR |
472 | get_task_struct(next_task); |
473 | ||
474 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 475 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
476 | if (!lowest_rq) { |
477 | struct task_struct *task; | |
478 | /* | |
697f0a48 | 479 | * find lock_lowest_rq releases rq->lock |
e8fa1362 SR |
480 | * so it is possible that next_task has changed. |
481 | * If it has, then try again. | |
482 | */ | |
697f0a48 | 483 | task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
484 | if (unlikely(task != next_task) && task && paranoid--) { |
485 | put_task_struct(next_task); | |
486 | next_task = task; | |
487 | goto retry; | |
488 | } | |
489 | goto out; | |
490 | } | |
491 | ||
492 | assert_spin_locked(&lowest_rq->lock); | |
493 | ||
697f0a48 | 494 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
495 | set_task_cpu(next_task, lowest_rq->cpu); |
496 | activate_task(lowest_rq, next_task, 0); | |
497 | ||
498 | resched_task(lowest_rq->curr); | |
499 | ||
500 | spin_unlock(&lowest_rq->lock); | |
501 | ||
502 | ret = 1; | |
503 | out: | |
504 | put_task_struct(next_task); | |
505 | ||
506 | return ret; | |
507 | } | |
508 | ||
509 | /* | |
510 | * TODO: Currently we just use the second highest prio task on | |
511 | * the queue, and stop when it can't migrate (or there's | |
512 | * no more RT tasks). There may be a case where a lower | |
513 | * priority RT task has a different affinity than the | |
514 | * higher RT task. In this case the lower RT task could | |
515 | * possibly be able to migrate where as the higher priority | |
516 | * RT task could not. We currently ignore this issue. | |
517 | * Enhancements are welcome! | |
518 | */ | |
519 | static void push_rt_tasks(struct rq *rq) | |
520 | { | |
521 | /* push_rt_task will return true if it moved an RT */ | |
522 | while (push_rt_task(rq)) | |
523 | ; | |
524 | } | |
525 | ||
f65eda4f SR |
526 | static int pull_rt_task(struct rq *this_rq) |
527 | { | |
528 | struct task_struct *next; | |
529 | struct task_struct *p; | |
530 | struct rq *src_rq; | |
531 | cpumask_t *rto_cpumask; | |
532 | int this_cpu = this_rq->cpu; | |
533 | int cpu; | |
534 | int ret = 0; | |
535 | ||
536 | assert_spin_locked(&this_rq->lock); | |
537 | ||
538 | /* | |
539 | * If cpusets are used, and we have overlapping | |
540 | * run queue cpusets, then this algorithm may not catch all. | |
541 | * This is just the price you pay on trying to keep | |
542 | * dirtying caches down on large SMP machines. | |
543 | */ | |
544 | if (likely(!rt_overloaded())) | |
545 | return 0; | |
546 | ||
547 | next = pick_next_task_rt(this_rq); | |
548 | ||
549 | rto_cpumask = rt_overload(); | |
550 | ||
551 | for_each_cpu_mask(cpu, *rto_cpumask) { | |
552 | if (this_cpu == cpu) | |
553 | continue; | |
554 | ||
555 | src_rq = cpu_rq(cpu); | |
556 | if (unlikely(src_rq->rt.rt_nr_running <= 1)) { | |
557 | /* | |
558 | * It is possible that overlapping cpusets | |
559 | * will miss clearing a non overloaded runqueue. | |
560 | * Clear it now. | |
561 | */ | |
562 | if (double_lock_balance(this_rq, src_rq)) { | |
563 | /* unlocked our runqueue lock */ | |
564 | struct task_struct *old_next = next; | |
565 | next = pick_next_task_rt(this_rq); | |
566 | if (next != old_next) | |
567 | ret = 1; | |
568 | } | |
569 | if (likely(src_rq->rt.rt_nr_running <= 1)) | |
570 | /* | |
571 | * Small chance that this_rq->curr changed | |
572 | * but it's really harmless here. | |
573 | */ | |
574 | rt_clear_overload(this_rq); | |
575 | else | |
576 | /* | |
577 | * Heh, the src_rq is now overloaded, since | |
578 | * we already have the src_rq lock, go straight | |
579 | * to pulling tasks from it. | |
580 | */ | |
581 | goto try_pulling; | |
582 | spin_unlock(&src_rq->lock); | |
583 | continue; | |
584 | } | |
585 | ||
586 | /* | |
587 | * We can potentially drop this_rq's lock in | |
588 | * double_lock_balance, and another CPU could | |
589 | * steal our next task - hence we must cause | |
590 | * the caller to recalculate the next task | |
591 | * in that case: | |
592 | */ | |
593 | if (double_lock_balance(this_rq, src_rq)) { | |
594 | struct task_struct *old_next = next; | |
595 | next = pick_next_task_rt(this_rq); | |
596 | if (next != old_next) | |
597 | ret = 1; | |
598 | } | |
599 | ||
600 | /* | |
601 | * Are there still pullable RT tasks? | |
602 | */ | |
603 | if (src_rq->rt.rt_nr_running <= 1) { | |
604 | spin_unlock(&src_rq->lock); | |
605 | continue; | |
606 | } | |
607 | ||
608 | try_pulling: | |
609 | p = pick_next_highest_task_rt(src_rq, this_cpu); | |
610 | ||
611 | /* | |
612 | * Do we have an RT task that preempts | |
613 | * the to-be-scheduled task? | |
614 | */ | |
615 | if (p && (!next || (p->prio < next->prio))) { | |
616 | WARN_ON(p == src_rq->curr); | |
617 | WARN_ON(!p->se.on_rq); | |
618 | ||
619 | /* | |
620 | * There's a chance that p is higher in priority | |
621 | * than what's currently running on its cpu. | |
622 | * This is just that p is wakeing up and hasn't | |
623 | * had a chance to schedule. We only pull | |
624 | * p if it is lower in priority than the | |
625 | * current task on the run queue or | |
626 | * this_rq next task is lower in prio than | |
627 | * the current task on that rq. | |
628 | */ | |
629 | if (p->prio < src_rq->curr->prio || | |
630 | (next && next->prio < src_rq->curr->prio)) | |
631 | goto bail; | |
632 | ||
633 | ret = 1; | |
634 | ||
635 | deactivate_task(src_rq, p, 0); | |
636 | set_task_cpu(p, this_cpu); | |
637 | activate_task(this_rq, p, 0); | |
638 | /* | |
639 | * We continue with the search, just in | |
640 | * case there's an even higher prio task | |
641 | * in another runqueue. (low likelyhood | |
642 | * but possible) | |
643 | */ | |
644 | ||
645 | /* | |
646 | * Update next so that we won't pick a task | |
647 | * on another cpu with a priority lower (or equal) | |
648 | * than the one we just picked. | |
649 | */ | |
650 | next = p; | |
651 | ||
652 | } | |
653 | bail: | |
654 | spin_unlock(&src_rq->lock); | |
655 | } | |
656 | ||
657 | return ret; | |
658 | } | |
659 | ||
660 | static void schedule_balance_rt(struct rq *rq, | |
661 | struct task_struct *prev) | |
662 | { | |
663 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
664 | if (unlikely(rt_task(prev)) && | |
665 | rq->rt.highest_prio > prev->prio) | |
666 | pull_rt_task(rq); | |
667 | } | |
668 | ||
e8fa1362 SR |
669 | static void schedule_tail_balance_rt(struct rq *rq) |
670 | { | |
671 | /* | |
672 | * If we have more than one rt_task queued, then | |
673 | * see if we can push the other rt_tasks off to other CPUS. | |
674 | * Note we may release the rq lock, and since | |
675 | * the lock was owned by prev, we need to release it | |
676 | * first via finish_lock_switch and then reaquire it here. | |
677 | */ | |
678 | if (unlikely(rq->rt.rt_nr_running > 1)) { | |
679 | spin_lock_irq(&rq->lock); | |
680 | push_rt_tasks(rq); | |
681 | spin_unlock_irq(&rq->lock); | |
682 | } | |
683 | } | |
684 | ||
4642dafd SR |
685 | |
686 | static void wakeup_balance_rt(struct rq *rq, struct task_struct *p) | |
687 | { | |
688 | if (unlikely(rt_task(p)) && | |
689 | !task_running(rq, p) && | |
690 | (p->prio >= rq->curr->prio)) | |
691 | push_rt_tasks(rq); | |
692 | } | |
693 | ||
43010659 | 694 | static unsigned long |
bb44e5d1 | 695 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f PW |
696 | unsigned long max_load_move, |
697 | struct sched_domain *sd, enum cpu_idle_type idle, | |
698 | int *all_pinned, int *this_best_prio) | |
bb44e5d1 | 699 | { |
c7a1e46a SR |
700 | /* don't touch RT tasks */ |
701 | return 0; | |
e1d1484f PW |
702 | } |
703 | ||
704 | static int | |
705 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
706 | struct sched_domain *sd, enum cpu_idle_type idle) | |
707 | { | |
c7a1e46a SR |
708 | /* don't touch RT tasks */ |
709 | return 0; | |
bb44e5d1 | 710 | } |
73fe6aae GH |
711 | static 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 | ||
724 | if ((p->nr_cpus_allowed <= 1) && (weight > 1)) | |
725 | rq->rt.rt_nr_migratory++; | |
726 | else if((p->nr_cpus_allowed > 1) && (weight <= 1)) { | |
727 | 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 | } | |
e8fa1362 SR |
737 | #else /* CONFIG_SMP */ |
738 | # define schedule_tail_balance_rt(rq) do { } while (0) | |
f65eda4f | 739 | # define schedule_balance_rt(rq, prev) do { } while (0) |
4642dafd | 740 | # define wakeup_balance_rt(rq, p) do { } while (0) |
e8fa1362 | 741 | #endif /* CONFIG_SMP */ |
bb44e5d1 IM |
742 | |
743 | static void task_tick_rt(struct rq *rq, struct task_struct *p) | |
744 | { | |
67e2be02 PZ |
745 | update_curr_rt(rq); |
746 | ||
bb44e5d1 IM |
747 | /* |
748 | * RR tasks need a special form of timeslice management. | |
749 | * FIFO tasks have no timeslices. | |
750 | */ | |
751 | if (p->policy != SCHED_RR) | |
752 | return; | |
753 | ||
754 | if (--p->time_slice) | |
755 | return; | |
756 | ||
a4ec24b4 | 757 | p->time_slice = DEF_TIMESLICE; |
bb44e5d1 | 758 | |
98fbc798 DA |
759 | /* |
760 | * Requeue to the end of queue if we are not the only element | |
761 | * on the queue: | |
762 | */ | |
763 | if (p->run_list.prev != p->run_list.next) { | |
764 | requeue_task_rt(rq, p); | |
765 | set_tsk_need_resched(p); | |
766 | } | |
bb44e5d1 IM |
767 | } |
768 | ||
83b699ed SV |
769 | static void set_curr_task_rt(struct rq *rq) |
770 | { | |
771 | struct task_struct *p = rq->curr; | |
772 | ||
773 | p->se.exec_start = rq->clock; | |
774 | } | |
775 | ||
5522d5d5 IM |
776 | const struct sched_class rt_sched_class = { |
777 | .next = &fair_sched_class, | |
bb44e5d1 IM |
778 | .enqueue_task = enqueue_task_rt, |
779 | .dequeue_task = dequeue_task_rt, | |
780 | .yield_task = yield_task_rt, | |
e7693a36 GH |
781 | #ifdef CONFIG_SMP |
782 | .select_task_rq = select_task_rq_rt, | |
783 | #endif /* CONFIG_SMP */ | |
bb44e5d1 IM |
784 | |
785 | .check_preempt_curr = check_preempt_curr_rt, | |
786 | ||
787 | .pick_next_task = pick_next_task_rt, | |
788 | .put_prev_task = put_prev_task_rt, | |
789 | ||
681f3e68 | 790 | #ifdef CONFIG_SMP |
bb44e5d1 | 791 | .load_balance = load_balance_rt, |
e1d1484f | 792 | .move_one_task = move_one_task_rt, |
73fe6aae | 793 | .set_cpus_allowed = set_cpus_allowed_rt, |
681f3e68 | 794 | #endif |
bb44e5d1 | 795 | |
83b699ed | 796 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 797 | .task_tick = task_tick_rt, |
bb44e5d1 | 798 | }; |