]>
Commit | Line | Data |
---|---|---|
bb44e5d1 IM |
1 | /* |
2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR | |
3 | * policies) | |
4 | */ | |
5 | ||
4fd29176 | 6 | #ifdef CONFIG_SMP |
84de4274 | 7 | |
637f5085 | 8 | static inline int rt_overloaded(struct rq *rq) |
4fd29176 | 9 | { |
637f5085 | 10 | return atomic_read(&rq->rd->rto_count); |
4fd29176 | 11 | } |
84de4274 | 12 | |
4fd29176 SR |
13 | static inline void rt_set_overload(struct rq *rq) |
14 | { | |
637f5085 | 15 | cpu_set(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
16 | /* |
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(); | |
637f5085 | 24 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 25 | } |
84de4274 | 26 | |
4fd29176 SR |
27 | static inline void rt_clear_overload(struct rq *rq) |
28 | { | |
29 | /* the order here really doesn't matter */ | |
637f5085 GH |
30 | atomic_dec(&rq->rd->rto_count); |
31 | cpu_clear(rq->cpu, rq->rd->rto_mask); | |
4fd29176 | 32 | } |
73fe6aae GH |
33 | |
34 | static void update_rt_migration(struct rq *rq) | |
35 | { | |
637f5085 | 36 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) { |
cdc8eb98 GH |
37 | if (!rq->rt.overloaded) { |
38 | rt_set_overload(rq); | |
39 | rq->rt.overloaded = 1; | |
40 | } | |
41 | } else if (rq->rt.overloaded) { | |
73fe6aae | 42 | rt_clear_overload(rq); |
637f5085 GH |
43 | rq->rt.overloaded = 0; |
44 | } | |
73fe6aae | 45 | } |
4fd29176 SR |
46 | #endif /* CONFIG_SMP */ |
47 | ||
fa85ae24 PZ |
48 | static int sched_rt_ratio_exceeded(struct rq *rq, struct rt_rq *rt_rq) |
49 | { | |
50 | u64 period, ratio; | |
51 | ||
52 | if (sysctl_sched_rt_ratio == SCHED_RT_FRAC) | |
53 | return 0; | |
54 | ||
55 | if (rt_rq->rt_throttled) | |
56 | return 1; | |
57 | ||
58 | period = (u64)sysctl_sched_rt_period * NSEC_PER_MSEC; | |
59 | ratio = (period * sysctl_sched_rt_ratio) >> SCHED_RT_FRAC_SHIFT; | |
60 | ||
61 | if (rt_rq->rt_time > ratio) { | |
62 | rt_rq->rt_throttled = rq->clock + period - rt_rq->rt_time; | |
63 | return 1; | |
64 | } | |
65 | ||
66 | return 0; | |
67 | } | |
68 | ||
69 | static void update_sched_rt_period(struct rq *rq) | |
70 | { | |
71 | while (rq->clock > rq->rt_period_expire) { | |
72 | u64 period, ratio; | |
73 | ||
74 | period = (u64)sysctl_sched_rt_period * NSEC_PER_MSEC; | |
75 | ratio = (period * sysctl_sched_rt_ratio) >> SCHED_RT_FRAC_SHIFT; | |
76 | ||
77 | rq->rt.rt_time -= min(rq->rt.rt_time, ratio); | |
78 | rq->rt_period_expire += period; | |
79 | } | |
80 | ||
81 | /* | |
82 | * When the rt throttle is expired, let them rip. | |
83 | * (XXX: use hrtick when available) | |
84 | */ | |
85 | if (rq->rt.rt_throttled && rq->clock > rq->rt.rt_throttled) { | |
86 | rq->rt.rt_throttled = 0; | |
87 | if (!sched_rt_ratio_exceeded(rq, &rq->rt)) | |
88 | resched_task(rq->curr); | |
89 | } | |
90 | } | |
91 | ||
bb44e5d1 IM |
92 | /* |
93 | * Update the current task's runtime statistics. Skip current tasks that | |
94 | * are not in our scheduling class. | |
95 | */ | |
a9957449 | 96 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
97 | { |
98 | struct task_struct *curr = rq->curr; | |
99 | u64 delta_exec; | |
100 | ||
101 | if (!task_has_rt_policy(curr)) | |
102 | return; | |
103 | ||
d281918d | 104 | delta_exec = rq->clock - curr->se.exec_start; |
bb44e5d1 IM |
105 | if (unlikely((s64)delta_exec < 0)) |
106 | delta_exec = 0; | |
6cfb0d5d IM |
107 | |
108 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); | |
bb44e5d1 IM |
109 | |
110 | curr->se.sum_exec_runtime += delta_exec; | |
d281918d | 111 | curr->se.exec_start = rq->clock; |
d842de87 | 112 | cpuacct_charge(curr, delta_exec); |
fa85ae24 PZ |
113 | |
114 | rq->rt.rt_time += delta_exec; | |
115 | update_sched_rt_period(rq); | |
116 | if (sched_rt_ratio_exceeded(rq, &rq->rt)) | |
117 | resched_task(curr); | |
bb44e5d1 IM |
118 | } |
119 | ||
63489e45 SR |
120 | static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq) |
121 | { | |
122 | WARN_ON(!rt_task(p)); | |
123 | rq->rt.rt_nr_running++; | |
764a9d6f SR |
124 | #ifdef CONFIG_SMP |
125 | if (p->prio < rq->rt.highest_prio) | |
126 | rq->rt.highest_prio = p->prio; | |
73fe6aae GH |
127 | if (p->nr_cpus_allowed > 1) |
128 | rq->rt.rt_nr_migratory++; | |
129 | ||
130 | update_rt_migration(rq); | |
764a9d6f | 131 | #endif /* CONFIG_SMP */ |
63489e45 SR |
132 | } |
133 | ||
134 | static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq) | |
135 | { | |
136 | WARN_ON(!rt_task(p)); | |
137 | WARN_ON(!rq->rt.rt_nr_running); | |
138 | rq->rt.rt_nr_running--; | |
764a9d6f SR |
139 | #ifdef CONFIG_SMP |
140 | if (rq->rt.rt_nr_running) { | |
141 | struct rt_prio_array *array; | |
142 | ||
143 | WARN_ON(p->prio < rq->rt.highest_prio); | |
144 | if (p->prio == rq->rt.highest_prio) { | |
145 | /* recalculate */ | |
146 | array = &rq->rt.active; | |
147 | rq->rt.highest_prio = | |
148 | sched_find_first_bit(array->bitmap); | |
149 | } /* otherwise leave rq->highest prio alone */ | |
150 | } else | |
151 | rq->rt.highest_prio = MAX_RT_PRIO; | |
73fe6aae GH |
152 | if (p->nr_cpus_allowed > 1) |
153 | rq->rt.rt_nr_migratory--; | |
154 | ||
155 | update_rt_migration(rq); | |
764a9d6f | 156 | #endif /* CONFIG_SMP */ |
63489e45 SR |
157 | } |
158 | ||
fd390f6a | 159 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
bb44e5d1 IM |
160 | { |
161 | struct rt_prio_array *array = &rq->rt.active; | |
162 | ||
fa717060 | 163 | list_add_tail(&p->rt.run_list, array->queue + p->prio); |
bb44e5d1 | 164 | __set_bit(p->prio, array->bitmap); |
58e2d4ca | 165 | inc_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
166 | |
167 | inc_rt_tasks(p, rq); | |
78f2c7db PZ |
168 | |
169 | if (wakeup) | |
170 | p->rt.timeout = 0; | |
bb44e5d1 IM |
171 | } |
172 | ||
173 | /* | |
174 | * Adding/removing a task to/from a priority array: | |
175 | */ | |
f02231e5 | 176 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
bb44e5d1 IM |
177 | { |
178 | struct rt_prio_array *array = &rq->rt.active; | |
179 | ||
f1e14ef6 | 180 | update_curr_rt(rq); |
bb44e5d1 | 181 | |
fa717060 | 182 | list_del(&p->rt.run_list); |
bb44e5d1 IM |
183 | if (list_empty(array->queue + p->prio)) |
184 | __clear_bit(p->prio, array->bitmap); | |
58e2d4ca | 185 | dec_cpu_load(rq, p->se.load.weight); |
63489e45 SR |
186 | |
187 | dec_rt_tasks(p, rq); | |
bb44e5d1 IM |
188 | } |
189 | ||
190 | /* | |
191 | * Put task to the end of the run list without the overhead of dequeue | |
192 | * followed by enqueue. | |
193 | */ | |
194 | static void requeue_task_rt(struct rq *rq, struct task_struct *p) | |
195 | { | |
196 | struct rt_prio_array *array = &rq->rt.active; | |
197 | ||
fa717060 | 198 | list_move_tail(&p->rt.run_list, array->queue + p->prio); |
bb44e5d1 IM |
199 | } |
200 | ||
201 | static void | |
4530d7ab | 202 | yield_task_rt(struct rq *rq) |
bb44e5d1 | 203 | { |
4530d7ab | 204 | requeue_task_rt(rq, rq->curr); |
bb44e5d1 IM |
205 | } |
206 | ||
e7693a36 | 207 | #ifdef CONFIG_SMP |
318e0893 GH |
208 | static int find_lowest_rq(struct task_struct *task); |
209 | ||
e7693a36 GH |
210 | static int select_task_rq_rt(struct task_struct *p, int sync) |
211 | { | |
318e0893 GH |
212 | struct rq *rq = task_rq(p); |
213 | ||
214 | /* | |
e1f47d89 SR |
215 | * If the current task is an RT task, then |
216 | * try to see if we can wake this RT task up on another | |
217 | * runqueue. Otherwise simply start this RT task | |
218 | * on its current runqueue. | |
219 | * | |
220 | * We want to avoid overloading runqueues. Even if | |
221 | * the RT task is of higher priority than the current RT task. | |
222 | * RT tasks behave differently than other tasks. If | |
223 | * one gets preempted, we try to push it off to another queue. | |
224 | * So trying to keep a preempting RT task on the same | |
225 | * cache hot CPU will force the running RT task to | |
226 | * a cold CPU. So we waste all the cache for the lower | |
227 | * RT task in hopes of saving some of a RT task | |
228 | * that is just being woken and probably will have | |
229 | * cold cache anyway. | |
318e0893 | 230 | */ |
17b3279b GH |
231 | if (unlikely(rt_task(rq->curr)) && |
232 | (p->nr_cpus_allowed > 1)) { | |
318e0893 GH |
233 | int cpu = find_lowest_rq(p); |
234 | ||
235 | return (cpu == -1) ? task_cpu(p) : cpu; | |
236 | } | |
237 | ||
238 | /* | |
239 | * Otherwise, just let it ride on the affined RQ and the | |
240 | * post-schedule router will push the preempted task away | |
241 | */ | |
e7693a36 GH |
242 | return task_cpu(p); |
243 | } | |
244 | #endif /* CONFIG_SMP */ | |
245 | ||
bb44e5d1 IM |
246 | /* |
247 | * Preempt the current task with a newly woken task if needed: | |
248 | */ | |
249 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p) | |
250 | { | |
251 | if (p->prio < rq->curr->prio) | |
252 | resched_task(rq->curr); | |
253 | } | |
254 | ||
fb8d4724 | 255 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
bb44e5d1 IM |
256 | { |
257 | struct rt_prio_array *array = &rq->rt.active; | |
258 | struct task_struct *next; | |
259 | struct list_head *queue; | |
fa85ae24 | 260 | struct rt_rq *rt_rq = &rq->rt; |
bb44e5d1 IM |
261 | int idx; |
262 | ||
fa85ae24 PZ |
263 | if (sched_rt_ratio_exceeded(rq, rt_rq)) |
264 | return NULL; | |
265 | ||
bb44e5d1 IM |
266 | idx = sched_find_first_bit(array->bitmap); |
267 | if (idx >= MAX_RT_PRIO) | |
268 | return NULL; | |
269 | ||
270 | queue = array->queue + idx; | |
fa717060 | 271 | next = list_entry(queue->next, struct task_struct, rt.run_list); |
bb44e5d1 | 272 | |
d281918d | 273 | next->se.exec_start = rq->clock; |
bb44e5d1 IM |
274 | |
275 | return next; | |
276 | } | |
277 | ||
31ee529c | 278 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 279 | { |
f1e14ef6 | 280 | update_curr_rt(rq); |
bb44e5d1 IM |
281 | p->se.exec_start = 0; |
282 | } | |
283 | ||
681f3e68 | 284 | #ifdef CONFIG_SMP |
e8fa1362 SR |
285 | /* Only try algorithms three times */ |
286 | #define RT_MAX_TRIES 3 | |
287 | ||
288 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); | |
289 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); | |
290 | ||
f65eda4f SR |
291 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
292 | { | |
293 | if (!task_running(rq, p) && | |
73fe6aae GH |
294 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
295 | (p->nr_cpus_allowed > 1)) | |
f65eda4f SR |
296 | return 1; |
297 | return 0; | |
298 | } | |
299 | ||
e8fa1362 | 300 | /* Return the second highest RT task, NULL otherwise */ |
79064fbf | 301 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
e8fa1362 SR |
302 | { |
303 | struct rt_prio_array *array = &rq->rt.active; | |
304 | struct task_struct *next; | |
305 | struct list_head *queue; | |
306 | int idx; | |
307 | ||
e8fa1362 SR |
308 | if (likely(rq->rt.rt_nr_running < 2)) |
309 | return NULL; | |
310 | ||
311 | idx = sched_find_first_bit(array->bitmap); | |
312 | if (unlikely(idx >= MAX_RT_PRIO)) { | |
313 | WARN_ON(1); /* rt_nr_running is bad */ | |
314 | return NULL; | |
315 | } | |
316 | ||
317 | queue = array->queue + idx; | |
f65eda4f SR |
318 | BUG_ON(list_empty(queue)); |
319 | ||
fa717060 | 320 | next = list_entry(queue->next, struct task_struct, rt.run_list); |
f65eda4f SR |
321 | if (unlikely(pick_rt_task(rq, next, cpu))) |
322 | goto out; | |
e8fa1362 SR |
323 | |
324 | if (queue->next->next != queue) { | |
325 | /* same prio task */ | |
79064fbf | 326 | next = list_entry(queue->next->next, struct task_struct, |
fa717060 | 327 | rt.run_list); |
f65eda4f SR |
328 | if (pick_rt_task(rq, next, cpu)) |
329 | goto out; | |
e8fa1362 SR |
330 | } |
331 | ||
f65eda4f | 332 | retry: |
e8fa1362 SR |
333 | /* slower, but more flexible */ |
334 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); | |
f65eda4f | 335 | if (unlikely(idx >= MAX_RT_PRIO)) |
e8fa1362 | 336 | return NULL; |
e8fa1362 SR |
337 | |
338 | queue = array->queue + idx; | |
f65eda4f SR |
339 | BUG_ON(list_empty(queue)); |
340 | ||
fa717060 | 341 | list_for_each_entry(next, queue, rt.run_list) { |
f65eda4f SR |
342 | if (pick_rt_task(rq, next, cpu)) |
343 | goto out; | |
344 | } | |
345 | ||
346 | goto retry; | |
e8fa1362 | 347 | |
f65eda4f | 348 | out: |
e8fa1362 SR |
349 | return next; |
350 | } | |
351 | ||
352 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); | |
353 | ||
6e1254d2 | 354 | static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask) |
e8fa1362 | 355 | { |
6e1254d2 | 356 | int lowest_prio = -1; |
610bf056 | 357 | int lowest_cpu = -1; |
06f90dbd | 358 | int count = 0; |
610bf056 | 359 | int cpu; |
e8fa1362 | 360 | |
637f5085 | 361 | cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed); |
e8fa1362 | 362 | |
07b4032c GH |
363 | /* |
364 | * Scan each rq for the lowest prio. | |
365 | */ | |
610bf056 | 366 | for_each_cpu_mask(cpu, *lowest_mask) { |
07b4032c | 367 | struct rq *rq = cpu_rq(cpu); |
e8fa1362 | 368 | |
07b4032c GH |
369 | /* We look for lowest RT prio or non-rt CPU */ |
370 | if (rq->rt.highest_prio >= MAX_RT_PRIO) { | |
610bf056 SR |
371 | /* |
372 | * if we already found a low RT queue | |
373 | * and now we found this non-rt queue | |
374 | * clear the mask and set our bit. | |
375 | * Otherwise just return the queue as is | |
376 | * and the count==1 will cause the algorithm | |
377 | * to use the first bit found. | |
378 | */ | |
379 | if (lowest_cpu != -1) { | |
6e1254d2 | 380 | cpus_clear(*lowest_mask); |
610bf056 SR |
381 | cpu_set(rq->cpu, *lowest_mask); |
382 | } | |
6e1254d2 | 383 | return 1; |
07b4032c GH |
384 | } |
385 | ||
386 | /* no locking for now */ | |
6e1254d2 GH |
387 | if ((rq->rt.highest_prio > task->prio) |
388 | && (rq->rt.highest_prio >= lowest_prio)) { | |
389 | if (rq->rt.highest_prio > lowest_prio) { | |
390 | /* new low - clear old data */ | |
391 | lowest_prio = rq->rt.highest_prio; | |
610bf056 SR |
392 | lowest_cpu = cpu; |
393 | count = 0; | |
6e1254d2 | 394 | } |
06f90dbd | 395 | count++; |
610bf056 SR |
396 | } else |
397 | cpu_clear(cpu, *lowest_mask); | |
398 | } | |
399 | ||
400 | /* | |
401 | * Clear out all the set bits that represent | |
402 | * runqueues that were of higher prio than | |
403 | * the lowest_prio. | |
404 | */ | |
405 | if (lowest_cpu > 0) { | |
406 | /* | |
407 | * Perhaps we could add another cpumask op to | |
408 | * zero out bits. Like cpu_zero_bits(cpumask, nrbits); | |
409 | * Then that could be optimized to use memset and such. | |
410 | */ | |
411 | for_each_cpu_mask(cpu, *lowest_mask) { | |
412 | if (cpu >= lowest_cpu) | |
413 | break; | |
414 | cpu_clear(cpu, *lowest_mask); | |
e8fa1362 | 415 | } |
07b4032c GH |
416 | } |
417 | ||
06f90dbd | 418 | return count; |
6e1254d2 GH |
419 | } |
420 | ||
421 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) | |
422 | { | |
423 | int first; | |
424 | ||
425 | /* "this_cpu" is cheaper to preempt than a remote processor */ | |
426 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) | |
427 | return this_cpu; | |
428 | ||
429 | first = first_cpu(*mask); | |
430 | if (first != NR_CPUS) | |
431 | return first; | |
432 | ||
433 | return -1; | |
434 | } | |
435 | ||
436 | static int find_lowest_rq(struct task_struct *task) | |
437 | { | |
438 | struct sched_domain *sd; | |
439 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); | |
440 | int this_cpu = smp_processor_id(); | |
441 | int cpu = task_cpu(task); | |
06f90dbd GH |
442 | int count = find_lowest_cpus(task, lowest_mask); |
443 | ||
444 | if (!count) | |
445 | return -1; /* No targets found */ | |
6e1254d2 | 446 | |
06f90dbd GH |
447 | /* |
448 | * There is no sense in performing an optimal search if only one | |
449 | * target is found. | |
450 | */ | |
451 | if (count == 1) | |
452 | return first_cpu(*lowest_mask); | |
6e1254d2 GH |
453 | |
454 | /* | |
455 | * At this point we have built a mask of cpus representing the | |
456 | * lowest priority tasks in the system. Now we want to elect | |
457 | * the best one based on our affinity and topology. | |
458 | * | |
459 | * We prioritize the last cpu that the task executed on since | |
460 | * it is most likely cache-hot in that location. | |
461 | */ | |
462 | if (cpu_isset(cpu, *lowest_mask)) | |
463 | return cpu; | |
464 | ||
465 | /* | |
466 | * Otherwise, we consult the sched_domains span maps to figure | |
467 | * out which cpu is logically closest to our hot cache data. | |
468 | */ | |
469 | if (this_cpu == cpu) | |
470 | this_cpu = -1; /* Skip this_cpu opt if the same */ | |
471 | ||
472 | for_each_domain(cpu, sd) { | |
473 | if (sd->flags & SD_WAKE_AFFINE) { | |
474 | cpumask_t domain_mask; | |
475 | int best_cpu; | |
476 | ||
477 | cpus_and(domain_mask, sd->span, *lowest_mask); | |
478 | ||
479 | best_cpu = pick_optimal_cpu(this_cpu, | |
480 | &domain_mask); | |
481 | if (best_cpu != -1) | |
482 | return best_cpu; | |
483 | } | |
484 | } | |
485 | ||
486 | /* | |
487 | * And finally, if there were no matches within the domains | |
488 | * just give the caller *something* to work with from the compatible | |
489 | * locations. | |
490 | */ | |
491 | return pick_optimal_cpu(this_cpu, lowest_mask); | |
07b4032c GH |
492 | } |
493 | ||
494 | /* Will lock the rq it finds */ | |
4df64c0b | 495 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
496 | { |
497 | struct rq *lowest_rq = NULL; | |
07b4032c | 498 | int tries; |
4df64c0b | 499 | int cpu; |
e8fa1362 | 500 | |
07b4032c GH |
501 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
502 | cpu = find_lowest_rq(task); | |
503 | ||
2de0b463 | 504 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
505 | break; |
506 | ||
07b4032c GH |
507 | lowest_rq = cpu_rq(cpu); |
508 | ||
e8fa1362 | 509 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 510 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
511 | /* |
512 | * We had to unlock the run queue. In | |
513 | * the mean time, task could have | |
514 | * migrated already or had its affinity changed. | |
515 | * Also make sure that it wasn't scheduled on its rq. | |
516 | */ | |
07b4032c | 517 | if (unlikely(task_rq(task) != rq || |
4df64c0b IM |
518 | !cpu_isset(lowest_rq->cpu, |
519 | task->cpus_allowed) || | |
07b4032c | 520 | task_running(rq, task) || |
e8fa1362 | 521 | !task->se.on_rq)) { |
4df64c0b | 522 | |
e8fa1362 SR |
523 | spin_unlock(&lowest_rq->lock); |
524 | lowest_rq = NULL; | |
525 | break; | |
526 | } | |
527 | } | |
528 | ||
529 | /* If this rq is still suitable use it. */ | |
530 | if (lowest_rq->rt.highest_prio > task->prio) | |
531 | break; | |
532 | ||
533 | /* try again */ | |
534 | spin_unlock(&lowest_rq->lock); | |
535 | lowest_rq = NULL; | |
536 | } | |
537 | ||
538 | return lowest_rq; | |
539 | } | |
540 | ||
541 | /* | |
542 | * If the current CPU has more than one RT task, see if the non | |
543 | * running task can migrate over to a CPU that is running a task | |
544 | * of lesser priority. | |
545 | */ | |
697f0a48 | 546 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
547 | { |
548 | struct task_struct *next_task; | |
549 | struct rq *lowest_rq; | |
550 | int ret = 0; | |
551 | int paranoid = RT_MAX_TRIES; | |
552 | ||
a22d7fc1 GH |
553 | if (!rq->rt.overloaded) |
554 | return 0; | |
555 | ||
697f0a48 | 556 | next_task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
557 | if (!next_task) |
558 | return 0; | |
559 | ||
560 | retry: | |
697f0a48 | 561 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 562 | WARN_ON(1); |
e8fa1362 | 563 | return 0; |
f65eda4f | 564 | } |
e8fa1362 SR |
565 | |
566 | /* | |
567 | * It's possible that the next_task slipped in of | |
568 | * higher priority than current. If that's the case | |
569 | * just reschedule current. | |
570 | */ | |
697f0a48 GH |
571 | if (unlikely(next_task->prio < rq->curr->prio)) { |
572 | resched_task(rq->curr); | |
e8fa1362 SR |
573 | return 0; |
574 | } | |
575 | ||
697f0a48 | 576 | /* We might release rq lock */ |
e8fa1362 SR |
577 | get_task_struct(next_task); |
578 | ||
579 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 580 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
581 | if (!lowest_rq) { |
582 | struct task_struct *task; | |
583 | /* | |
697f0a48 | 584 | * find lock_lowest_rq releases rq->lock |
e8fa1362 SR |
585 | * so it is possible that next_task has changed. |
586 | * If it has, then try again. | |
587 | */ | |
697f0a48 | 588 | task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
589 | if (unlikely(task != next_task) && task && paranoid--) { |
590 | put_task_struct(next_task); | |
591 | next_task = task; | |
592 | goto retry; | |
593 | } | |
594 | goto out; | |
595 | } | |
596 | ||
697f0a48 | 597 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
598 | set_task_cpu(next_task, lowest_rq->cpu); |
599 | activate_task(lowest_rq, next_task, 0); | |
600 | ||
601 | resched_task(lowest_rq->curr); | |
602 | ||
603 | spin_unlock(&lowest_rq->lock); | |
604 | ||
605 | ret = 1; | |
606 | out: | |
607 | put_task_struct(next_task); | |
608 | ||
609 | return ret; | |
610 | } | |
611 | ||
612 | /* | |
613 | * TODO: Currently we just use the second highest prio task on | |
614 | * the queue, and stop when it can't migrate (or there's | |
615 | * no more RT tasks). There may be a case where a lower | |
616 | * priority RT task has a different affinity than the | |
617 | * higher RT task. In this case the lower RT task could | |
618 | * possibly be able to migrate where as the higher priority | |
619 | * RT task could not. We currently ignore this issue. | |
620 | * Enhancements are welcome! | |
621 | */ | |
622 | static void push_rt_tasks(struct rq *rq) | |
623 | { | |
624 | /* push_rt_task will return true if it moved an RT */ | |
625 | while (push_rt_task(rq)) | |
626 | ; | |
627 | } | |
628 | ||
f65eda4f SR |
629 | static int pull_rt_task(struct rq *this_rq) |
630 | { | |
80bf3171 IM |
631 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
632 | struct task_struct *p, *next; | |
f65eda4f | 633 | struct rq *src_rq; |
f65eda4f | 634 | |
637f5085 | 635 | if (likely(!rt_overloaded(this_rq))) |
f65eda4f SR |
636 | return 0; |
637 | ||
638 | next = pick_next_task_rt(this_rq); | |
639 | ||
637f5085 | 640 | for_each_cpu_mask(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
641 | if (this_cpu == cpu) |
642 | continue; | |
643 | ||
644 | src_rq = cpu_rq(cpu); | |
f65eda4f SR |
645 | /* |
646 | * We can potentially drop this_rq's lock in | |
647 | * double_lock_balance, and another CPU could | |
648 | * steal our next task - hence we must cause | |
649 | * the caller to recalculate the next task | |
650 | * in that case: | |
651 | */ | |
652 | if (double_lock_balance(this_rq, src_rq)) { | |
653 | struct task_struct *old_next = next; | |
80bf3171 | 654 | |
f65eda4f SR |
655 | next = pick_next_task_rt(this_rq); |
656 | if (next != old_next) | |
657 | ret = 1; | |
658 | } | |
659 | ||
660 | /* | |
661 | * Are there still pullable RT tasks? | |
662 | */ | |
663 | if (src_rq->rt.rt_nr_running <= 1) { | |
664 | spin_unlock(&src_rq->lock); | |
665 | continue; | |
666 | } | |
667 | ||
f65eda4f SR |
668 | p = pick_next_highest_task_rt(src_rq, this_cpu); |
669 | ||
670 | /* | |
671 | * Do we have an RT task that preempts | |
672 | * the to-be-scheduled task? | |
673 | */ | |
674 | if (p && (!next || (p->prio < next->prio))) { | |
675 | WARN_ON(p == src_rq->curr); | |
676 | WARN_ON(!p->se.on_rq); | |
677 | ||
678 | /* | |
679 | * There's a chance that p is higher in priority | |
680 | * than what's currently running on its cpu. | |
681 | * This is just that p is wakeing up and hasn't | |
682 | * had a chance to schedule. We only pull | |
683 | * p if it is lower in priority than the | |
684 | * current task on the run queue or | |
685 | * this_rq next task is lower in prio than | |
686 | * the current task on that rq. | |
687 | */ | |
688 | if (p->prio < src_rq->curr->prio || | |
689 | (next && next->prio < src_rq->curr->prio)) | |
80bf3171 | 690 | goto out; |
f65eda4f SR |
691 | |
692 | ret = 1; | |
693 | ||
694 | deactivate_task(src_rq, p, 0); | |
695 | set_task_cpu(p, this_cpu); | |
696 | activate_task(this_rq, p, 0); | |
697 | /* | |
698 | * We continue with the search, just in | |
699 | * case there's an even higher prio task | |
700 | * in another runqueue. (low likelyhood | |
701 | * but possible) | |
80bf3171 | 702 | * |
f65eda4f SR |
703 | * Update next so that we won't pick a task |
704 | * on another cpu with a priority lower (or equal) | |
705 | * than the one we just picked. | |
706 | */ | |
707 | next = p; | |
708 | ||
709 | } | |
80bf3171 | 710 | out: |
f65eda4f SR |
711 | spin_unlock(&src_rq->lock); |
712 | } | |
713 | ||
714 | return ret; | |
715 | } | |
716 | ||
9a897c5a | 717 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev) |
f65eda4f SR |
718 | { |
719 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
7f51f298 | 720 | if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio) |
f65eda4f SR |
721 | pull_rt_task(rq); |
722 | } | |
723 | ||
9a897c5a | 724 | static void post_schedule_rt(struct rq *rq) |
e8fa1362 SR |
725 | { |
726 | /* | |
727 | * If we have more than one rt_task queued, then | |
728 | * see if we can push the other rt_tasks off to other CPUS. | |
729 | * Note we may release the rq lock, and since | |
730 | * the lock was owned by prev, we need to release it | |
731 | * first via finish_lock_switch and then reaquire it here. | |
732 | */ | |
a22d7fc1 | 733 | if (unlikely(rq->rt.overloaded)) { |
e8fa1362 SR |
734 | spin_lock_irq(&rq->lock); |
735 | push_rt_tasks(rq); | |
736 | spin_unlock_irq(&rq->lock); | |
737 | } | |
738 | } | |
739 | ||
4642dafd | 740 | |
9a897c5a | 741 | static void task_wake_up_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 742 | { |
9a897c5a | 743 | if (!task_running(rq, p) && |
a22d7fc1 GH |
744 | (p->prio >= rq->rt.highest_prio) && |
745 | rq->rt.overloaded) | |
4642dafd SR |
746 | push_rt_tasks(rq); |
747 | } | |
748 | ||
43010659 | 749 | static unsigned long |
bb44e5d1 | 750 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f PW |
751 | unsigned long max_load_move, |
752 | struct sched_domain *sd, enum cpu_idle_type idle, | |
753 | int *all_pinned, int *this_best_prio) | |
bb44e5d1 | 754 | { |
c7a1e46a SR |
755 | /* don't touch RT tasks */ |
756 | return 0; | |
e1d1484f PW |
757 | } |
758 | ||
759 | static int | |
760 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
761 | struct sched_domain *sd, enum cpu_idle_type idle) | |
762 | { | |
c7a1e46a SR |
763 | /* don't touch RT tasks */ |
764 | return 0; | |
bb44e5d1 | 765 | } |
deeeccd4 | 766 | |
73fe6aae GH |
767 | static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask) |
768 | { | |
769 | int weight = cpus_weight(*new_mask); | |
770 | ||
771 | BUG_ON(!rt_task(p)); | |
772 | ||
773 | /* | |
774 | * Update the migration status of the RQ if we have an RT task | |
775 | * which is running AND changing its weight value. | |
776 | */ | |
777 | if (p->se.on_rq && (weight != p->nr_cpus_allowed)) { | |
778 | struct rq *rq = task_rq(p); | |
779 | ||
deeeccd4 | 780 | if ((p->nr_cpus_allowed <= 1) && (weight > 1)) { |
73fe6aae | 781 | rq->rt.rt_nr_migratory++; |
deeeccd4 | 782 | } else if ((p->nr_cpus_allowed > 1) && (weight <= 1)) { |
73fe6aae GH |
783 | BUG_ON(!rq->rt.rt_nr_migratory); |
784 | rq->rt.rt_nr_migratory--; | |
785 | } | |
786 | ||
787 | update_rt_migration(rq); | |
788 | } | |
789 | ||
790 | p->cpus_allowed = *new_mask; | |
791 | p->nr_cpus_allowed = weight; | |
792 | } | |
deeeccd4 | 793 | |
bdd7c81b IM |
794 | /* Assumes rq->lock is held */ |
795 | static void join_domain_rt(struct rq *rq) | |
796 | { | |
797 | if (rq->rt.overloaded) | |
798 | rt_set_overload(rq); | |
799 | } | |
800 | ||
801 | /* Assumes rq->lock is held */ | |
802 | static void leave_domain_rt(struct rq *rq) | |
803 | { | |
804 | if (rq->rt.overloaded) | |
805 | rt_clear_overload(rq); | |
806 | } | |
cb469845 SR |
807 | |
808 | /* | |
809 | * When switch from the rt queue, we bring ourselves to a position | |
810 | * that we might want to pull RT tasks from other runqueues. | |
811 | */ | |
812 | static void switched_from_rt(struct rq *rq, struct task_struct *p, | |
813 | int running) | |
814 | { | |
815 | /* | |
816 | * If there are other RT tasks then we will reschedule | |
817 | * and the scheduling of the other RT tasks will handle | |
818 | * the balancing. But if we are the last RT task | |
819 | * we may need to handle the pulling of RT tasks | |
820 | * now. | |
821 | */ | |
822 | if (!rq->rt.rt_nr_running) | |
823 | pull_rt_task(rq); | |
824 | } | |
825 | #endif /* CONFIG_SMP */ | |
826 | ||
827 | /* | |
828 | * When switching a task to RT, we may overload the runqueue | |
829 | * with RT tasks. In this case we try to push them off to | |
830 | * other runqueues. | |
831 | */ | |
832 | static void switched_to_rt(struct rq *rq, struct task_struct *p, | |
833 | int running) | |
834 | { | |
835 | int check_resched = 1; | |
836 | ||
837 | /* | |
838 | * If we are already running, then there's nothing | |
839 | * that needs to be done. But if we are not running | |
840 | * we may need to preempt the current running task. | |
841 | * If that current running task is also an RT task | |
842 | * then see if we can move to another run queue. | |
843 | */ | |
844 | if (!running) { | |
845 | #ifdef CONFIG_SMP | |
846 | if (rq->rt.overloaded && push_rt_task(rq) && | |
847 | /* Don't resched if we changed runqueues */ | |
848 | rq != task_rq(p)) | |
849 | check_resched = 0; | |
850 | #endif /* CONFIG_SMP */ | |
851 | if (check_resched && p->prio < rq->curr->prio) | |
852 | resched_task(rq->curr); | |
853 | } | |
854 | } | |
855 | ||
856 | /* | |
857 | * Priority of the task has changed. This may cause | |
858 | * us to initiate a push or pull. | |
859 | */ | |
860 | static void prio_changed_rt(struct rq *rq, struct task_struct *p, | |
861 | int oldprio, int running) | |
862 | { | |
863 | if (running) { | |
864 | #ifdef CONFIG_SMP | |
865 | /* | |
866 | * If our priority decreases while running, we | |
867 | * may need to pull tasks to this runqueue. | |
868 | */ | |
869 | if (oldprio < p->prio) | |
870 | pull_rt_task(rq); | |
871 | /* | |
872 | * If there's a higher priority task waiting to run | |
873 | * then reschedule. | |
874 | */ | |
875 | if (p->prio > rq->rt.highest_prio) | |
876 | resched_task(p); | |
877 | #else | |
878 | /* For UP simply resched on drop of prio */ | |
879 | if (oldprio < p->prio) | |
880 | resched_task(p); | |
e8fa1362 | 881 | #endif /* CONFIG_SMP */ |
cb469845 SR |
882 | } else { |
883 | /* | |
884 | * This task is not running, but if it is | |
885 | * greater than the current running task | |
886 | * then reschedule. | |
887 | */ | |
888 | if (p->prio < rq->curr->prio) | |
889 | resched_task(rq->curr); | |
890 | } | |
891 | } | |
892 | ||
78f2c7db PZ |
893 | static void watchdog(struct rq *rq, struct task_struct *p) |
894 | { | |
895 | unsigned long soft, hard; | |
896 | ||
897 | if (!p->signal) | |
898 | return; | |
899 | ||
900 | soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur; | |
901 | hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max; | |
902 | ||
903 | if (soft != RLIM_INFINITY) { | |
904 | unsigned long next; | |
905 | ||
906 | p->rt.timeout++; | |
907 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); | |
908 | if (next > p->rt.timeout) { | |
909 | u64 next_time = p->se.sum_exec_runtime; | |
910 | ||
911 | next_time += next * (NSEC_PER_SEC/HZ); | |
912 | if (p->it_sched_expires > next_time) | |
913 | p->it_sched_expires = next_time; | |
914 | } else | |
915 | p->it_sched_expires = p->se.sum_exec_runtime; | |
916 | } | |
917 | } | |
bb44e5d1 | 918 | |
8f4d37ec | 919 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 920 | { |
67e2be02 PZ |
921 | update_curr_rt(rq); |
922 | ||
78f2c7db PZ |
923 | watchdog(rq, p); |
924 | ||
bb44e5d1 IM |
925 | /* |
926 | * RR tasks need a special form of timeslice management. | |
927 | * FIFO tasks have no timeslices. | |
928 | */ | |
929 | if (p->policy != SCHED_RR) | |
930 | return; | |
931 | ||
fa717060 | 932 | if (--p->rt.time_slice) |
bb44e5d1 IM |
933 | return; |
934 | ||
fa717060 | 935 | p->rt.time_slice = DEF_TIMESLICE; |
bb44e5d1 | 936 | |
98fbc798 DA |
937 | /* |
938 | * Requeue to the end of queue if we are not the only element | |
939 | * on the queue: | |
940 | */ | |
fa717060 | 941 | if (p->rt.run_list.prev != p->rt.run_list.next) { |
98fbc798 DA |
942 | requeue_task_rt(rq, p); |
943 | set_tsk_need_resched(p); | |
944 | } | |
bb44e5d1 IM |
945 | } |
946 | ||
83b699ed SV |
947 | static void set_curr_task_rt(struct rq *rq) |
948 | { | |
949 | struct task_struct *p = rq->curr; | |
950 | ||
951 | p->se.exec_start = rq->clock; | |
952 | } | |
953 | ||
5522d5d5 IM |
954 | const struct sched_class rt_sched_class = { |
955 | .next = &fair_sched_class, | |
bb44e5d1 IM |
956 | .enqueue_task = enqueue_task_rt, |
957 | .dequeue_task = dequeue_task_rt, | |
958 | .yield_task = yield_task_rt, | |
e7693a36 GH |
959 | #ifdef CONFIG_SMP |
960 | .select_task_rq = select_task_rq_rt, | |
961 | #endif /* CONFIG_SMP */ | |
bb44e5d1 IM |
962 | |
963 | .check_preempt_curr = check_preempt_curr_rt, | |
964 | ||
965 | .pick_next_task = pick_next_task_rt, | |
966 | .put_prev_task = put_prev_task_rt, | |
967 | ||
681f3e68 | 968 | #ifdef CONFIG_SMP |
bb44e5d1 | 969 | .load_balance = load_balance_rt, |
e1d1484f | 970 | .move_one_task = move_one_task_rt, |
73fe6aae | 971 | .set_cpus_allowed = set_cpus_allowed_rt, |
bdd7c81b IM |
972 | .join_domain = join_domain_rt, |
973 | .leave_domain = leave_domain_rt, | |
9a897c5a SR |
974 | .pre_schedule = pre_schedule_rt, |
975 | .post_schedule = post_schedule_rt, | |
976 | .task_wake_up = task_wake_up_rt, | |
cb469845 | 977 | .switched_from = switched_from_rt, |
681f3e68 | 978 | #endif |
bb44e5d1 | 979 | |
83b699ed | 980 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 981 | .task_tick = task_tick_rt, |
cb469845 SR |
982 | |
983 | .prio_changed = prio_changed_rt, | |
984 | .switched_to = switched_to_rt, | |
bb44e5d1 | 985 | }; |