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 | { | |
a22d7fc1 | 19 | rq->rt.overloaded = 1; |
4fd29176 SR |
20 | cpu_set(rq->cpu, rt_overload_mask); |
21 | /* | |
22 | * Make sure the mask is visible before we set | |
23 | * the overload count. That is checked to determine | |
24 | * if we should look at the mask. It would be a shame | |
25 | * if we looked at the mask, but the mask was not | |
26 | * updated yet. | |
27 | */ | |
28 | wmb(); | |
29 | atomic_inc(&rto_count); | |
30 | } | |
31 | static inline void rt_clear_overload(struct rq *rq) | |
32 | { | |
33 | /* the order here really doesn't matter */ | |
34 | atomic_dec(&rto_count); | |
35 | cpu_clear(rq->cpu, rt_overload_mask); | |
a22d7fc1 | 36 | rq->rt.overloaded = 0; |
4fd29176 | 37 | } |
73fe6aae GH |
38 | |
39 | static void update_rt_migration(struct rq *rq) | |
40 | { | |
41 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) | |
42 | rt_set_overload(rq); | |
43 | else | |
44 | rt_clear_overload(rq); | |
45 | } | |
4fd29176 SR |
46 | #endif /* CONFIG_SMP */ |
47 | ||
bb44e5d1 IM |
48 | /* |
49 | * Update the current task's runtime statistics. Skip current tasks that | |
50 | * are not in our scheduling class. | |
51 | */ | |
a9957449 | 52 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
53 | { |
54 | struct task_struct *curr = rq->curr; | |
55 | u64 delta_exec; | |
56 | ||
57 | if (!task_has_rt_policy(curr)) | |
58 | return; | |
59 | ||
d281918d | 60 | delta_exec = rq->clock - curr->se.exec_start; |
bb44e5d1 IM |
61 | if (unlikely((s64)delta_exec < 0)) |
62 | delta_exec = 0; | |
6cfb0d5d IM |
63 | |
64 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); | |
bb44e5d1 IM |
65 | |
66 | curr->se.sum_exec_runtime += delta_exec; | |
d281918d | 67 | curr->se.exec_start = rq->clock; |
d842de87 | 68 | cpuacct_charge(curr, delta_exec); |
bb44e5d1 IM |
69 | } |
70 | ||
63489e45 SR |
71 | static 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++; | |
764a9d6f SR |
75 | #ifdef CONFIG_SMP |
76 | if (p->prio < rq->rt.highest_prio) | |
77 | rq->rt.highest_prio = p->prio; | |
73fe6aae GH |
78 | if (p->nr_cpus_allowed > 1) |
79 | rq->rt.rt_nr_migratory++; | |
80 | ||
81 | update_rt_migration(rq); | |
764a9d6f | 82 | #endif /* CONFIG_SMP */ |
63489e45 SR |
83 | } |
84 | ||
85 | static 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--; | |
764a9d6f SR |
90 | #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; | |
73fe6aae GH |
103 | if (p->nr_cpus_allowed > 1) |
104 | rq->rt.rt_nr_migratory--; | |
105 | ||
106 | update_rt_migration(rq); | |
764a9d6f | 107 | #endif /* CONFIG_SMP */ |
63489e45 SR |
108 | } |
109 | ||
fd390f6a | 110 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
bb44e5d1 IM |
111 | { |
112 | struct rt_prio_array *array = &rq->rt.active; | |
113 | ||
114 | list_add_tail(&p->run_list, array->queue + p->prio); | |
115 | __set_bit(p->prio, array->bitmap); | |
58e2d4ca | 116 | inc_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
117 | |
118 | inc_rt_tasks(p, rq); | |
bb44e5d1 IM |
119 | } |
120 | ||
121 | /* | |
122 | * Adding/removing a task to/from a priority array: | |
123 | */ | |
f02231e5 | 124 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
bb44e5d1 IM |
125 | { |
126 | struct rt_prio_array *array = &rq->rt.active; | |
127 | ||
f1e14ef6 | 128 | update_curr_rt(rq); |
bb44e5d1 IM |
129 | |
130 | list_del(&p->run_list); | |
131 | if (list_empty(array->queue + p->prio)) | |
132 | __clear_bit(p->prio, array->bitmap); | |
58e2d4ca | 133 | dec_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
134 | |
135 | dec_rt_tasks(p, rq); | |
bb44e5d1 IM |
136 | } |
137 | ||
138 | /* | |
139 | * Put task to the end of the run list without the overhead of dequeue | |
140 | * followed by enqueue. | |
141 | */ | |
142 | static void requeue_task_rt(struct rq *rq, struct task_struct *p) | |
143 | { | |
144 | struct rt_prio_array *array = &rq->rt.active; | |
145 | ||
146 | list_move_tail(&p->run_list, array->queue + p->prio); | |
147 | } | |
148 | ||
149 | static void | |
4530d7ab | 150 | yield_task_rt(struct rq *rq) |
bb44e5d1 | 151 | { |
4530d7ab | 152 | requeue_task_rt(rq, rq->curr); |
bb44e5d1 IM |
153 | } |
154 | ||
e7693a36 | 155 | #ifdef CONFIG_SMP |
318e0893 GH |
156 | static int find_lowest_rq(struct task_struct *task); |
157 | ||
e7693a36 GH |
158 | static int select_task_rq_rt(struct task_struct *p, int sync) |
159 | { | |
318e0893 GH |
160 | struct rq *rq = task_rq(p); |
161 | ||
162 | /* | |
e1f47d89 SR |
163 | * 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. | |
318e0893 | 178 | */ |
17b3279b GH |
179 | if (unlikely(rt_task(rq->curr)) && |
180 | (p->nr_cpus_allowed > 1)) { | |
318e0893 GH |
181 | 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 | */ | |
e7693a36 GH |
190 | return task_cpu(p); |
191 | } | |
192 | #endif /* CONFIG_SMP */ | |
193 | ||
bb44e5d1 IM |
194 | /* |
195 | * Preempt the current task with a newly woken task if needed: | |
196 | */ | |
197 | static 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 | ||
fb8d4724 | 203 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
bb44e5d1 IM |
204 | { |
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; | |
215 | next = list_entry(queue->next, struct task_struct, run_list); | |
216 | ||
d281918d | 217 | next->se.exec_start = rq->clock; |
bb44e5d1 IM |
218 | |
219 | return next; | |
220 | } | |
221 | ||
31ee529c | 222 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 223 | { |
f1e14ef6 | 224 | update_curr_rt(rq); |
bb44e5d1 IM |
225 | p->se.exec_start = 0; |
226 | } | |
227 | ||
681f3e68 | 228 | #ifdef CONFIG_SMP |
e8fa1362 SR |
229 | /* Only try algorithms three times */ |
230 | #define RT_MAX_TRIES 3 | |
231 | ||
232 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); | |
233 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); | |
234 | ||
f65eda4f SR |
235 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
236 | { | |
237 | if (!task_running(rq, p) && | |
73fe6aae GH |
238 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
239 | (p->nr_cpus_allowed > 1)) | |
f65eda4f SR |
240 | return 1; |
241 | return 0; | |
242 | } | |
243 | ||
e8fa1362 | 244 | /* Return the second highest RT task, NULL otherwise */ |
79064fbf | 245 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
e8fa1362 SR |
246 | { |
247 | struct rt_prio_array *array = &rq->rt.active; | |
248 | struct task_struct *next; | |
249 | struct list_head *queue; | |
250 | int idx; | |
251 | ||
252 | assert_spin_locked(&rq->lock); | |
253 | ||
254 | if (likely(rq->rt.rt_nr_running < 2)) | |
255 | return NULL; | |
256 | ||
257 | idx = sched_find_first_bit(array->bitmap); | |
258 | if (unlikely(idx >= MAX_RT_PRIO)) { | |
259 | WARN_ON(1); /* rt_nr_running is bad */ | |
260 | return NULL; | |
261 | } | |
262 | ||
263 | queue = array->queue + idx; | |
f65eda4f SR |
264 | BUG_ON(list_empty(queue)); |
265 | ||
e8fa1362 | 266 | next = list_entry(queue->next, struct task_struct, run_list); |
f65eda4f SR |
267 | if (unlikely(pick_rt_task(rq, next, cpu))) |
268 | goto out; | |
e8fa1362 SR |
269 | |
270 | if (queue->next->next != queue) { | |
271 | /* same prio task */ | |
79064fbf IM |
272 | next = list_entry(queue->next->next, struct task_struct, |
273 | run_list); | |
f65eda4f SR |
274 | if (pick_rt_task(rq, next, cpu)) |
275 | goto out; | |
e8fa1362 SR |
276 | } |
277 | ||
f65eda4f | 278 | retry: |
e8fa1362 SR |
279 | /* slower, but more flexible */ |
280 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); | |
f65eda4f | 281 | if (unlikely(idx >= MAX_RT_PRIO)) |
e8fa1362 | 282 | return NULL; |
e8fa1362 SR |
283 | |
284 | queue = array->queue + idx; | |
f65eda4f SR |
285 | BUG_ON(list_empty(queue)); |
286 | ||
287 | list_for_each_entry(next, queue, run_list) { | |
288 | if (pick_rt_task(rq, next, cpu)) | |
289 | goto out; | |
290 | } | |
291 | ||
292 | goto retry; | |
e8fa1362 | 293 | |
f65eda4f | 294 | out: |
e8fa1362 SR |
295 | return next; |
296 | } | |
297 | ||
298 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); | |
299 | ||
6e1254d2 | 300 | static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask) |
e8fa1362 | 301 | { |
6e1254d2 | 302 | int lowest_prio = -1; |
610bf056 | 303 | int lowest_cpu = -1; |
06f90dbd | 304 | int count = 0; |
610bf056 | 305 | int cpu; |
e8fa1362 | 306 | |
610bf056 | 307 | cpus_and(*lowest_mask, cpu_online_map, task->cpus_allowed); |
e8fa1362 | 308 | |
07b4032c GH |
309 | /* |
310 | * Scan each rq for the lowest prio. | |
311 | */ | |
610bf056 | 312 | for_each_cpu_mask(cpu, *lowest_mask) { |
07b4032c | 313 | struct rq *rq = cpu_rq(cpu); |
e8fa1362 | 314 | |
07b4032c GH |
315 | /* We look for lowest RT prio or non-rt CPU */ |
316 | if (rq->rt.highest_prio >= MAX_RT_PRIO) { | |
610bf056 SR |
317 | /* |
318 | * if we already found a low RT queue | |
319 | * and now we found this non-rt queue | |
320 | * clear the mask and set our bit. | |
321 | * Otherwise just return the queue as is | |
322 | * and the count==1 will cause the algorithm | |
323 | * to use the first bit found. | |
324 | */ | |
325 | if (lowest_cpu != -1) { | |
6e1254d2 | 326 | cpus_clear(*lowest_mask); |
610bf056 SR |
327 | cpu_set(rq->cpu, *lowest_mask); |
328 | } | |
6e1254d2 | 329 | return 1; |
07b4032c GH |
330 | } |
331 | ||
332 | /* no locking for now */ | |
6e1254d2 GH |
333 | if ((rq->rt.highest_prio > task->prio) |
334 | && (rq->rt.highest_prio >= lowest_prio)) { | |
335 | if (rq->rt.highest_prio > lowest_prio) { | |
336 | /* new low - clear old data */ | |
337 | lowest_prio = rq->rt.highest_prio; | |
610bf056 SR |
338 | lowest_cpu = cpu; |
339 | count = 0; | |
6e1254d2 | 340 | } |
06f90dbd | 341 | count++; |
610bf056 SR |
342 | } else |
343 | cpu_clear(cpu, *lowest_mask); | |
344 | } | |
345 | ||
346 | /* | |
347 | * Clear out all the set bits that represent | |
348 | * runqueues that were of higher prio than | |
349 | * the lowest_prio. | |
350 | */ | |
351 | if (lowest_cpu > 0) { | |
352 | /* | |
353 | * Perhaps we could add another cpumask op to | |
354 | * zero out bits. Like cpu_zero_bits(cpumask, nrbits); | |
355 | * Then that could be optimized to use memset and such. | |
356 | */ | |
357 | for_each_cpu_mask(cpu, *lowest_mask) { | |
358 | if (cpu >= lowest_cpu) | |
359 | break; | |
360 | cpu_clear(cpu, *lowest_mask); | |
e8fa1362 | 361 | } |
07b4032c GH |
362 | } |
363 | ||
06f90dbd | 364 | return count; |
6e1254d2 GH |
365 | } |
366 | ||
367 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) | |
368 | { | |
369 | int first; | |
370 | ||
371 | /* "this_cpu" is cheaper to preempt than a remote processor */ | |
372 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) | |
373 | return this_cpu; | |
374 | ||
375 | first = first_cpu(*mask); | |
376 | if (first != NR_CPUS) | |
377 | return first; | |
378 | ||
379 | return -1; | |
380 | } | |
381 | ||
382 | static int find_lowest_rq(struct task_struct *task) | |
383 | { | |
384 | struct sched_domain *sd; | |
385 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); | |
386 | int this_cpu = smp_processor_id(); | |
387 | int cpu = task_cpu(task); | |
06f90dbd GH |
388 | int count = find_lowest_cpus(task, lowest_mask); |
389 | ||
390 | if (!count) | |
391 | return -1; /* No targets found */ | |
6e1254d2 | 392 | |
06f90dbd GH |
393 | /* |
394 | * There is no sense in performing an optimal search if only one | |
395 | * target is found. | |
396 | */ | |
397 | if (count == 1) | |
398 | return first_cpu(*lowest_mask); | |
6e1254d2 GH |
399 | |
400 | /* | |
401 | * At this point we have built a mask of cpus representing the | |
402 | * lowest priority tasks in the system. Now we want to elect | |
403 | * the best one based on our affinity and topology. | |
404 | * | |
405 | * We prioritize the last cpu that the task executed on since | |
406 | * it is most likely cache-hot in that location. | |
407 | */ | |
408 | if (cpu_isset(cpu, *lowest_mask)) | |
409 | return cpu; | |
410 | ||
411 | /* | |
412 | * Otherwise, we consult the sched_domains span maps to figure | |
413 | * out which cpu is logically closest to our hot cache data. | |
414 | */ | |
415 | if (this_cpu == cpu) | |
416 | this_cpu = -1; /* Skip this_cpu opt if the same */ | |
417 | ||
418 | for_each_domain(cpu, sd) { | |
419 | if (sd->flags & SD_WAKE_AFFINE) { | |
420 | cpumask_t domain_mask; | |
421 | int best_cpu; | |
422 | ||
423 | cpus_and(domain_mask, sd->span, *lowest_mask); | |
424 | ||
425 | best_cpu = pick_optimal_cpu(this_cpu, | |
426 | &domain_mask); | |
427 | if (best_cpu != -1) | |
428 | return best_cpu; | |
429 | } | |
430 | } | |
431 | ||
432 | /* | |
433 | * And finally, if there were no matches within the domains | |
434 | * just give the caller *something* to work with from the compatible | |
435 | * locations. | |
436 | */ | |
437 | return pick_optimal_cpu(this_cpu, lowest_mask); | |
07b4032c GH |
438 | } |
439 | ||
440 | /* Will lock the rq it finds */ | |
4df64c0b | 441 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
442 | { |
443 | struct rq *lowest_rq = NULL; | |
07b4032c | 444 | int tries; |
4df64c0b | 445 | int cpu; |
e8fa1362 | 446 | |
07b4032c GH |
447 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
448 | cpu = find_lowest_rq(task); | |
449 | ||
2de0b463 | 450 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
451 | break; |
452 | ||
07b4032c GH |
453 | lowest_rq = cpu_rq(cpu); |
454 | ||
e8fa1362 | 455 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 456 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
457 | /* |
458 | * We had to unlock the run queue. In | |
459 | * the mean time, task could have | |
460 | * migrated already or had its affinity changed. | |
461 | * Also make sure that it wasn't scheduled on its rq. | |
462 | */ | |
07b4032c | 463 | if (unlikely(task_rq(task) != rq || |
4df64c0b IM |
464 | !cpu_isset(lowest_rq->cpu, |
465 | task->cpus_allowed) || | |
07b4032c | 466 | task_running(rq, task) || |
e8fa1362 | 467 | !task->se.on_rq)) { |
4df64c0b | 468 | |
e8fa1362 SR |
469 | spin_unlock(&lowest_rq->lock); |
470 | lowest_rq = NULL; | |
471 | break; | |
472 | } | |
473 | } | |
474 | ||
475 | /* If this rq is still suitable use it. */ | |
476 | if (lowest_rq->rt.highest_prio > task->prio) | |
477 | break; | |
478 | ||
479 | /* try again */ | |
480 | spin_unlock(&lowest_rq->lock); | |
481 | lowest_rq = NULL; | |
482 | } | |
483 | ||
484 | return lowest_rq; | |
485 | } | |
486 | ||
487 | /* | |
488 | * If the current CPU has more than one RT task, see if the non | |
489 | * running task can migrate over to a CPU that is running a task | |
490 | * of lesser priority. | |
491 | */ | |
697f0a48 | 492 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
493 | { |
494 | struct task_struct *next_task; | |
495 | struct rq *lowest_rq; | |
496 | int ret = 0; | |
497 | int paranoid = RT_MAX_TRIES; | |
498 | ||
697f0a48 | 499 | assert_spin_locked(&rq->lock); |
e8fa1362 | 500 | |
a22d7fc1 GH |
501 | if (!rq->rt.overloaded) |
502 | return 0; | |
503 | ||
697f0a48 | 504 | next_task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
505 | if (!next_task) |
506 | return 0; | |
507 | ||
508 | retry: | |
697f0a48 | 509 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 510 | WARN_ON(1); |
e8fa1362 | 511 | return 0; |
f65eda4f | 512 | } |
e8fa1362 SR |
513 | |
514 | /* | |
515 | * It's possible that the next_task slipped in of | |
516 | * higher priority than current. If that's the case | |
517 | * just reschedule current. | |
518 | */ | |
697f0a48 GH |
519 | if (unlikely(next_task->prio < rq->curr->prio)) { |
520 | resched_task(rq->curr); | |
e8fa1362 SR |
521 | return 0; |
522 | } | |
523 | ||
697f0a48 | 524 | /* We might release rq lock */ |
e8fa1362 SR |
525 | get_task_struct(next_task); |
526 | ||
527 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 528 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
529 | if (!lowest_rq) { |
530 | struct task_struct *task; | |
531 | /* | |
697f0a48 | 532 | * find lock_lowest_rq releases rq->lock |
e8fa1362 SR |
533 | * so it is possible that next_task has changed. |
534 | * If it has, then try again. | |
535 | */ | |
697f0a48 | 536 | task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
537 | if (unlikely(task != next_task) && task && paranoid--) { |
538 | put_task_struct(next_task); | |
539 | next_task = task; | |
540 | goto retry; | |
541 | } | |
542 | goto out; | |
543 | } | |
544 | ||
545 | assert_spin_locked(&lowest_rq->lock); | |
546 | ||
697f0a48 | 547 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
548 | set_task_cpu(next_task, lowest_rq->cpu); |
549 | activate_task(lowest_rq, next_task, 0); | |
550 | ||
551 | resched_task(lowest_rq->curr); | |
552 | ||
553 | spin_unlock(&lowest_rq->lock); | |
554 | ||
555 | ret = 1; | |
556 | out: | |
557 | put_task_struct(next_task); | |
558 | ||
559 | return ret; | |
560 | } | |
561 | ||
562 | /* | |
563 | * TODO: Currently we just use the second highest prio task on | |
564 | * the queue, and stop when it can't migrate (or there's | |
565 | * no more RT tasks). There may be a case where a lower | |
566 | * priority RT task has a different affinity than the | |
567 | * higher RT task. In this case the lower RT task could | |
568 | * possibly be able to migrate where as the higher priority | |
569 | * RT task could not. We currently ignore this issue. | |
570 | * Enhancements are welcome! | |
571 | */ | |
572 | static void push_rt_tasks(struct rq *rq) | |
573 | { | |
574 | /* push_rt_task will return true if it moved an RT */ | |
575 | while (push_rt_task(rq)) | |
576 | ; | |
577 | } | |
578 | ||
f65eda4f SR |
579 | static int pull_rt_task(struct rq *this_rq) |
580 | { | |
581 | struct task_struct *next; | |
582 | struct task_struct *p; | |
583 | struct rq *src_rq; | |
584 | cpumask_t *rto_cpumask; | |
585 | int this_cpu = this_rq->cpu; | |
586 | int cpu; | |
587 | int ret = 0; | |
588 | ||
589 | assert_spin_locked(&this_rq->lock); | |
590 | ||
591 | /* | |
592 | * If cpusets are used, and we have overlapping | |
593 | * run queue cpusets, then this algorithm may not catch all. | |
594 | * This is just the price you pay on trying to keep | |
595 | * dirtying caches down on large SMP machines. | |
596 | */ | |
597 | if (likely(!rt_overloaded())) | |
598 | return 0; | |
599 | ||
600 | next = pick_next_task_rt(this_rq); | |
601 | ||
602 | rto_cpumask = rt_overload(); | |
603 | ||
604 | for_each_cpu_mask(cpu, *rto_cpumask) { | |
605 | if (this_cpu == cpu) | |
606 | continue; | |
607 | ||
608 | src_rq = cpu_rq(cpu); | |
609 | if (unlikely(src_rq->rt.rt_nr_running <= 1)) { | |
610 | /* | |
611 | * It is possible that overlapping cpusets | |
612 | * will miss clearing a non overloaded runqueue. | |
613 | * Clear it now. | |
614 | */ | |
615 | if (double_lock_balance(this_rq, src_rq)) { | |
616 | /* unlocked our runqueue lock */ | |
617 | struct task_struct *old_next = next; | |
618 | next = pick_next_task_rt(this_rq); | |
619 | if (next != old_next) | |
620 | ret = 1; | |
621 | } | |
622 | if (likely(src_rq->rt.rt_nr_running <= 1)) | |
623 | /* | |
624 | * Small chance that this_rq->curr changed | |
625 | * but it's really harmless here. | |
626 | */ | |
627 | rt_clear_overload(this_rq); | |
628 | else | |
629 | /* | |
630 | * Heh, the src_rq is now overloaded, since | |
631 | * we already have the src_rq lock, go straight | |
632 | * to pulling tasks from it. | |
633 | */ | |
634 | goto try_pulling; | |
635 | spin_unlock(&src_rq->lock); | |
636 | continue; | |
637 | } | |
638 | ||
639 | /* | |
640 | * We can potentially drop this_rq's lock in | |
641 | * double_lock_balance, and another CPU could | |
642 | * steal our next task - hence we must cause | |
643 | * the caller to recalculate the next task | |
644 | * in that case: | |
645 | */ | |
646 | if (double_lock_balance(this_rq, src_rq)) { | |
647 | struct task_struct *old_next = next; | |
648 | next = pick_next_task_rt(this_rq); | |
649 | if (next != old_next) | |
650 | ret = 1; | |
651 | } | |
652 | ||
653 | /* | |
654 | * Are there still pullable RT tasks? | |
655 | */ | |
656 | if (src_rq->rt.rt_nr_running <= 1) { | |
657 | spin_unlock(&src_rq->lock); | |
658 | continue; | |
659 | } | |
660 | ||
661 | try_pulling: | |
662 | p = pick_next_highest_task_rt(src_rq, this_cpu); | |
663 | ||
664 | /* | |
665 | * Do we have an RT task that preempts | |
666 | * the to-be-scheduled task? | |
667 | */ | |
668 | if (p && (!next || (p->prio < next->prio))) { | |
669 | WARN_ON(p == src_rq->curr); | |
670 | WARN_ON(!p->se.on_rq); | |
671 | ||
672 | /* | |
673 | * There's a chance that p is higher in priority | |
674 | * than what's currently running on its cpu. | |
675 | * This is just that p is wakeing up and hasn't | |
676 | * had a chance to schedule. We only pull | |
677 | * p if it is lower in priority than the | |
678 | * current task on the run queue or | |
679 | * this_rq next task is lower in prio than | |
680 | * the current task on that rq. | |
681 | */ | |
682 | if (p->prio < src_rq->curr->prio || | |
683 | (next && next->prio < src_rq->curr->prio)) | |
684 | goto bail; | |
685 | ||
686 | ret = 1; | |
687 | ||
688 | deactivate_task(src_rq, p, 0); | |
689 | set_task_cpu(p, this_cpu); | |
690 | activate_task(this_rq, p, 0); | |
691 | /* | |
692 | * We continue with the search, just in | |
693 | * case there's an even higher prio task | |
694 | * in another runqueue. (low likelyhood | |
695 | * but possible) | |
696 | */ | |
697 | ||
698 | /* | |
699 | * Update next so that we won't pick a task | |
700 | * on another cpu with a priority lower (or equal) | |
701 | * than the one we just picked. | |
702 | */ | |
703 | next = p; | |
704 | ||
705 | } | |
706 | bail: | |
707 | spin_unlock(&src_rq->lock); | |
708 | } | |
709 | ||
710 | return ret; | |
711 | } | |
712 | ||
713 | static void schedule_balance_rt(struct rq *rq, | |
714 | struct task_struct *prev) | |
715 | { | |
716 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
717 | if (unlikely(rt_task(prev)) && | |
718 | rq->rt.highest_prio > prev->prio) | |
719 | pull_rt_task(rq); | |
720 | } | |
721 | ||
e8fa1362 SR |
722 | static void schedule_tail_balance_rt(struct rq *rq) |
723 | { | |
724 | /* | |
725 | * If we have more than one rt_task queued, then | |
726 | * see if we can push the other rt_tasks off to other CPUS. | |
727 | * Note we may release the rq lock, and since | |
728 | * the lock was owned by prev, we need to release it | |
729 | * first via finish_lock_switch and then reaquire it here. | |
730 | */ | |
a22d7fc1 | 731 | if (unlikely(rq->rt.overloaded)) { |
e8fa1362 SR |
732 | spin_lock_irq(&rq->lock); |
733 | push_rt_tasks(rq); | |
734 | spin_unlock_irq(&rq->lock); | |
735 | } | |
736 | } | |
737 | ||
4642dafd SR |
738 | |
739 | static void wakeup_balance_rt(struct rq *rq, struct task_struct *p) | |
740 | { | |
741 | if (unlikely(rt_task(p)) && | |
742 | !task_running(rq, p) && | |
a22d7fc1 GH |
743 | (p->prio >= rq->rt.highest_prio) && |
744 | rq->rt.overloaded) | |
4642dafd SR |
745 | push_rt_tasks(rq); |
746 | } | |
747 | ||
43010659 | 748 | static unsigned long |
bb44e5d1 | 749 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f PW |
750 | unsigned long max_load_move, |
751 | struct sched_domain *sd, enum cpu_idle_type idle, | |
752 | int *all_pinned, int *this_best_prio) | |
bb44e5d1 | 753 | { |
c7a1e46a SR |
754 | /* don't touch RT tasks */ |
755 | return 0; | |
e1d1484f PW |
756 | } |
757 | ||
758 | static int | |
759 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
760 | struct sched_domain *sd, enum cpu_idle_type idle) | |
761 | { | |
c7a1e46a SR |
762 | /* don't touch RT tasks */ |
763 | return 0; | |
bb44e5d1 | 764 | } |
73fe6aae GH |
765 | static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask) |
766 | { | |
767 | int weight = cpus_weight(*new_mask); | |
768 | ||
769 | BUG_ON(!rt_task(p)); | |
770 | ||
771 | /* | |
772 | * Update the migration status of the RQ if we have an RT task | |
773 | * which is running AND changing its weight value. | |
774 | */ | |
775 | if (p->se.on_rq && (weight != p->nr_cpus_allowed)) { | |
776 | struct rq *rq = task_rq(p); | |
777 | ||
778 | if ((p->nr_cpus_allowed <= 1) && (weight > 1)) | |
779 | rq->rt.rt_nr_migratory++; | |
780 | else if((p->nr_cpus_allowed > 1) && (weight <= 1)) { | |
781 | BUG_ON(!rq->rt.rt_nr_migratory); | |
782 | rq->rt.rt_nr_migratory--; | |
783 | } | |
784 | ||
785 | update_rt_migration(rq); | |
786 | } | |
787 | ||
788 | p->cpus_allowed = *new_mask; | |
789 | p->nr_cpus_allowed = weight; | |
790 | } | |
e8fa1362 SR |
791 | #else /* CONFIG_SMP */ |
792 | # define schedule_tail_balance_rt(rq) do { } while (0) | |
f65eda4f | 793 | # define schedule_balance_rt(rq, prev) do { } while (0) |
4642dafd | 794 | # define wakeup_balance_rt(rq, p) do { } while (0) |
e8fa1362 | 795 | #endif /* CONFIG_SMP */ |
bb44e5d1 IM |
796 | |
797 | static void task_tick_rt(struct rq *rq, struct task_struct *p) | |
798 | { | |
67e2be02 PZ |
799 | update_curr_rt(rq); |
800 | ||
bb44e5d1 IM |
801 | /* |
802 | * RR tasks need a special form of timeslice management. | |
803 | * FIFO tasks have no timeslices. | |
804 | */ | |
805 | if (p->policy != SCHED_RR) | |
806 | return; | |
807 | ||
808 | if (--p->time_slice) | |
809 | return; | |
810 | ||
a4ec24b4 | 811 | p->time_slice = DEF_TIMESLICE; |
bb44e5d1 | 812 | |
98fbc798 DA |
813 | /* |
814 | * Requeue to the end of queue if we are not the only element | |
815 | * on the queue: | |
816 | */ | |
817 | if (p->run_list.prev != p->run_list.next) { | |
818 | requeue_task_rt(rq, p); | |
819 | set_tsk_need_resched(p); | |
820 | } | |
bb44e5d1 IM |
821 | } |
822 | ||
83b699ed SV |
823 | static void set_curr_task_rt(struct rq *rq) |
824 | { | |
825 | struct task_struct *p = rq->curr; | |
826 | ||
827 | p->se.exec_start = rq->clock; | |
828 | } | |
829 | ||
5522d5d5 IM |
830 | const struct sched_class rt_sched_class = { |
831 | .next = &fair_sched_class, | |
bb44e5d1 IM |
832 | .enqueue_task = enqueue_task_rt, |
833 | .dequeue_task = dequeue_task_rt, | |
834 | .yield_task = yield_task_rt, | |
e7693a36 GH |
835 | #ifdef CONFIG_SMP |
836 | .select_task_rq = select_task_rq_rt, | |
837 | #endif /* CONFIG_SMP */ | |
bb44e5d1 IM |
838 | |
839 | .check_preempt_curr = check_preempt_curr_rt, | |
840 | ||
841 | .pick_next_task = pick_next_task_rt, | |
842 | .put_prev_task = put_prev_task_rt, | |
843 | ||
681f3e68 | 844 | #ifdef CONFIG_SMP |
bb44e5d1 | 845 | .load_balance = load_balance_rt, |
e1d1484f | 846 | .move_one_task = move_one_task_rt, |
73fe6aae | 847 | .set_cpus_allowed = set_cpus_allowed_rt, |
681f3e68 | 848 | #endif |
bb44e5d1 | 849 | |
83b699ed | 850 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 851 | .task_tick = task_tick_rt, |
bb44e5d1 | 852 | }; |