Commit | Line | Data |
---|---|---|
aab03e05 DF |
1 | /* |
2 | * Deadline Scheduling Class (SCHED_DEADLINE) | |
3 | * | |
4 | * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS). | |
5 | * | |
6 | * Tasks that periodically executes their instances for less than their | |
7 | * runtime won't miss any of their deadlines. | |
8 | * Tasks that are not periodic or sporadic or that tries to execute more | |
9 | * than their reserved bandwidth will be slowed down (and may potentially | |
10 | * miss some of their deadlines), and won't affect any other task. | |
11 | * | |
12 | * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>, | |
1baca4ce | 13 | * Juri Lelli <juri.lelli@gmail.com>, |
aab03e05 DF |
14 | * Michael Trimarchi <michael@amarulasolutions.com>, |
15 | * Fabio Checconi <fchecconi@gmail.com> | |
16 | */ | |
17 | #include "sched.h" | |
18 | ||
6bfd6d72 JL |
19 | #include <linux/slab.h> |
20 | ||
332ac17e DF |
21 | struct dl_bandwidth def_dl_bandwidth; |
22 | ||
aab03e05 DF |
23 | static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se) |
24 | { | |
25 | return container_of(dl_se, struct task_struct, dl); | |
26 | } | |
27 | ||
28 | static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) | |
29 | { | |
30 | return container_of(dl_rq, struct rq, dl); | |
31 | } | |
32 | ||
33 | static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se) | |
34 | { | |
35 | struct task_struct *p = dl_task_of(dl_se); | |
36 | struct rq *rq = task_rq(p); | |
37 | ||
38 | return &rq->dl; | |
39 | } | |
40 | ||
41 | static inline int on_dl_rq(struct sched_dl_entity *dl_se) | |
42 | { | |
43 | return !RB_EMPTY_NODE(&dl_se->rb_node); | |
44 | } | |
45 | ||
46 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) | |
47 | { | |
48 | struct sched_dl_entity *dl_se = &p->dl; | |
49 | ||
50 | return dl_rq->rb_leftmost == &dl_se->rb_node; | |
51 | } | |
52 | ||
332ac17e DF |
53 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) |
54 | { | |
55 | raw_spin_lock_init(&dl_b->dl_runtime_lock); | |
56 | dl_b->dl_period = period; | |
57 | dl_b->dl_runtime = runtime; | |
58 | } | |
59 | ||
332ac17e DF |
60 | void init_dl_bw(struct dl_bw *dl_b) |
61 | { | |
62 | raw_spin_lock_init(&dl_b->lock); | |
63 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | |
1724813d | 64 | if (global_rt_runtime() == RUNTIME_INF) |
332ac17e DF |
65 | dl_b->bw = -1; |
66 | else | |
1724813d | 67 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
332ac17e DF |
68 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
69 | dl_b->total_bw = 0; | |
70 | } | |
71 | ||
07c54f7a | 72 | void init_dl_rq(struct dl_rq *dl_rq) |
aab03e05 DF |
73 | { |
74 | dl_rq->rb_root = RB_ROOT; | |
1baca4ce JL |
75 | |
76 | #ifdef CONFIG_SMP | |
77 | /* zero means no -deadline tasks */ | |
78 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | |
79 | ||
80 | dl_rq->dl_nr_migratory = 0; | |
81 | dl_rq->overloaded = 0; | |
82 | dl_rq->pushable_dl_tasks_root = RB_ROOT; | |
332ac17e DF |
83 | #else |
84 | init_dl_bw(&dl_rq->dl_bw); | |
1baca4ce JL |
85 | #endif |
86 | } | |
87 | ||
88 | #ifdef CONFIG_SMP | |
89 | ||
90 | static inline int dl_overloaded(struct rq *rq) | |
91 | { | |
92 | return atomic_read(&rq->rd->dlo_count); | |
93 | } | |
94 | ||
95 | static inline void dl_set_overload(struct rq *rq) | |
96 | { | |
97 | if (!rq->online) | |
98 | return; | |
99 | ||
100 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | |
101 | /* | |
102 | * Must be visible before the overload count is | |
103 | * set (as in sched_rt.c). | |
104 | * | |
105 | * Matched by the barrier in pull_dl_task(). | |
106 | */ | |
107 | smp_wmb(); | |
108 | atomic_inc(&rq->rd->dlo_count); | |
109 | } | |
110 | ||
111 | static inline void dl_clear_overload(struct rq *rq) | |
112 | { | |
113 | if (!rq->online) | |
114 | return; | |
115 | ||
116 | atomic_dec(&rq->rd->dlo_count); | |
117 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | |
118 | } | |
119 | ||
120 | static void update_dl_migration(struct dl_rq *dl_rq) | |
121 | { | |
995b9ea4 | 122 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { |
1baca4ce JL |
123 | if (!dl_rq->overloaded) { |
124 | dl_set_overload(rq_of_dl_rq(dl_rq)); | |
125 | dl_rq->overloaded = 1; | |
126 | } | |
127 | } else if (dl_rq->overloaded) { | |
128 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | |
129 | dl_rq->overloaded = 0; | |
130 | } | |
131 | } | |
132 | ||
133 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
134 | { | |
135 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 136 | |
1baca4ce JL |
137 | if (p->nr_cpus_allowed > 1) |
138 | dl_rq->dl_nr_migratory++; | |
139 | ||
140 | update_dl_migration(dl_rq); | |
141 | } | |
142 | ||
143 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
144 | { | |
145 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 146 | |
1baca4ce JL |
147 | if (p->nr_cpus_allowed > 1) |
148 | dl_rq->dl_nr_migratory--; | |
149 | ||
150 | update_dl_migration(dl_rq); | |
151 | } | |
152 | ||
153 | /* | |
154 | * The list of pushable -deadline task is not a plist, like in | |
155 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | |
156 | */ | |
157 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
158 | { | |
159 | struct dl_rq *dl_rq = &rq->dl; | |
160 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; | |
161 | struct rb_node *parent = NULL; | |
162 | struct task_struct *entry; | |
163 | int leftmost = 1; | |
164 | ||
165 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | |
166 | ||
167 | while (*link) { | |
168 | parent = *link; | |
169 | entry = rb_entry(parent, struct task_struct, | |
170 | pushable_dl_tasks); | |
171 | if (dl_entity_preempt(&p->dl, &entry->dl)) | |
172 | link = &parent->rb_left; | |
173 | else { | |
174 | link = &parent->rb_right; | |
175 | leftmost = 0; | |
176 | } | |
177 | } | |
178 | ||
179 | if (leftmost) | |
180 | dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; | |
181 | ||
182 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
183 | rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | |
aab03e05 DF |
184 | } |
185 | ||
1baca4ce JL |
186 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
187 | { | |
188 | struct dl_rq *dl_rq = &rq->dl; | |
189 | ||
190 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
191 | return; | |
192 | ||
193 | if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { | |
194 | struct rb_node *next_node; | |
195 | ||
196 | next_node = rb_next(&p->pushable_dl_tasks); | |
197 | dl_rq->pushable_dl_tasks_leftmost = next_node; | |
198 | } | |
199 | ||
200 | rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | |
201 | RB_CLEAR_NODE(&p->pushable_dl_tasks); | |
202 | } | |
203 | ||
204 | static inline int has_pushable_dl_tasks(struct rq *rq) | |
205 | { | |
206 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); | |
207 | } | |
208 | ||
209 | static int push_dl_task(struct rq *rq); | |
210 | ||
dc877341 PZ |
211 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
212 | { | |
213 | return dl_task(prev); | |
214 | } | |
215 | ||
216 | static inline void set_post_schedule(struct rq *rq) | |
217 | { | |
218 | rq->post_schedule = has_pushable_dl_tasks(rq); | |
219 | } | |
220 | ||
fa9c9d10 WL |
221 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); |
222 | ||
223 | static void dl_task_offline_migration(struct rq *rq, struct task_struct *p) | |
224 | { | |
225 | struct rq *later_rq = NULL; | |
226 | bool fallback = false; | |
227 | ||
228 | later_rq = find_lock_later_rq(p, rq); | |
229 | ||
230 | if (!later_rq) { | |
231 | int cpu; | |
232 | ||
233 | /* | |
234 | * If we cannot preempt any rq, fall back to pick any | |
235 | * online cpu. | |
236 | */ | |
237 | fallback = true; | |
238 | cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p)); | |
239 | if (cpu >= nr_cpu_ids) { | |
240 | /* | |
241 | * Fail to find any suitable cpu. | |
242 | * The task will never come back! | |
243 | */ | |
244 | BUG_ON(dl_bandwidth_enabled()); | |
245 | ||
246 | /* | |
247 | * If admission control is disabled we | |
248 | * try a little harder to let the task | |
249 | * run. | |
250 | */ | |
251 | cpu = cpumask_any(cpu_active_mask); | |
252 | } | |
253 | later_rq = cpu_rq(cpu); | |
254 | double_lock_balance(rq, later_rq); | |
255 | } | |
256 | ||
257 | deactivate_task(rq, p, 0); | |
258 | set_task_cpu(p, later_rq->cpu); | |
259 | activate_task(later_rq, p, ENQUEUE_REPLENISH); | |
260 | ||
261 | if (!fallback) | |
262 | resched_curr(later_rq); | |
263 | ||
264 | double_unlock_balance(rq, later_rq); | |
265 | } | |
266 | ||
1baca4ce JL |
267 | #else |
268 | ||
269 | static inline | |
270 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
271 | { | |
272 | } | |
273 | ||
274 | static inline | |
275 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
276 | { | |
277 | } | |
278 | ||
279 | static inline | |
280 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
281 | { | |
282 | } | |
283 | ||
284 | static inline | |
285 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
286 | { | |
287 | } | |
288 | ||
dc877341 PZ |
289 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
290 | { | |
291 | return false; | |
292 | } | |
293 | ||
294 | static inline int pull_dl_task(struct rq *rq) | |
295 | { | |
296 | return 0; | |
297 | } | |
298 | ||
299 | static inline void set_post_schedule(struct rq *rq) | |
300 | { | |
301 | } | |
1baca4ce JL |
302 | #endif /* CONFIG_SMP */ |
303 | ||
aab03e05 DF |
304 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
305 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
306 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
307 | int flags); | |
308 | ||
309 | /* | |
310 | * We are being explicitly informed that a new instance is starting, | |
311 | * and this means that: | |
312 | * - the absolute deadline of the entity has to be placed at | |
313 | * current time + relative deadline; | |
314 | * - the runtime of the entity has to be set to the maximum value. | |
315 | * | |
316 | * The capability of specifying such event is useful whenever a -deadline | |
317 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
318 | * one, and to (try to!) reconcile itself with its own scheduling | |
319 | * parameters. | |
320 | */ | |
2d3d891d DF |
321 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, |
322 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
323 | { |
324 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
325 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
326 | ||
327 | WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); | |
328 | ||
329 | /* | |
330 | * We use the regular wall clock time to set deadlines in the | |
331 | * future; in fact, we must consider execution overheads (time | |
332 | * spent on hardirq context, etc.). | |
333 | */ | |
2d3d891d DF |
334 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
335 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
336 | dl_se->dl_new = 0; |
337 | } | |
338 | ||
339 | /* | |
340 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
341 | * possibility of a entity lasting more than what it declared, and thus | |
342 | * exhausting its runtime. | |
343 | * | |
344 | * Here we are interested in making runtime overrun possible, but we do | |
345 | * not want a entity which is misbehaving to affect the scheduling of all | |
346 | * other entities. | |
347 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
348 | * is used, in order to confine each entity within its own bandwidth. | |
349 | * | |
350 | * This function deals exactly with that, and ensures that when the runtime | |
351 | * of a entity is replenished, its deadline is also postponed. That ensures | |
352 | * the overrunning entity can't interfere with other entity in the system and | |
353 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
354 | * could happen are, typically, a entity voluntarily trying to overcome its | |
1b09d29b | 355 | * runtime, or it just underestimated it during sched_setattr(). |
aab03e05 | 356 | */ |
2d3d891d DF |
357 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
358 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
359 | { |
360 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
361 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
362 | ||
2d3d891d DF |
363 | BUG_ON(pi_se->dl_runtime <= 0); |
364 | ||
365 | /* | |
366 | * This could be the case for a !-dl task that is boosted. | |
367 | * Just go with full inherited parameters. | |
368 | */ | |
369 | if (dl_se->dl_deadline == 0) { | |
370 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
371 | dl_se->runtime = pi_se->dl_runtime; | |
372 | } | |
373 | ||
aab03e05 DF |
374 | /* |
375 | * We keep moving the deadline away until we get some | |
376 | * available runtime for the entity. This ensures correct | |
377 | * handling of situations where the runtime overrun is | |
378 | * arbitrary large. | |
379 | */ | |
380 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
381 | dl_se->deadline += pi_se->dl_period; |
382 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
383 | } |
384 | ||
385 | /* | |
386 | * At this point, the deadline really should be "in | |
387 | * the future" with respect to rq->clock. If it's | |
388 | * not, we are, for some reason, lagging too much! | |
389 | * Anyway, after having warn userspace abut that, | |
390 | * we still try to keep the things running by | |
391 | * resetting the deadline and the budget of the | |
392 | * entity. | |
393 | */ | |
394 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
c224815d | 395 | printk_deferred_once("sched: DL replenish lagged to much\n"); |
2d3d891d DF |
396 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
397 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 | 398 | } |
1019a359 PZ |
399 | |
400 | if (dl_se->dl_yielded) | |
401 | dl_se->dl_yielded = 0; | |
402 | if (dl_se->dl_throttled) | |
403 | dl_se->dl_throttled = 0; | |
aab03e05 DF |
404 | } |
405 | ||
406 | /* | |
407 | * Here we check if --at time t-- an entity (which is probably being | |
408 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
409 | * and its current deadline _without_ exceeding the bandwidth it is | |
410 | * assigned (function returns true if it can't). We are in fact applying | |
411 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
412 | * over residual deadline fits within the allocated bandwidth, then we | |
413 | * can keep the current (absolute) deadline and residual budget without | |
414 | * disrupting the schedulability of the system. Otherwise, we should | |
415 | * refill the runtime and set the deadline a period in the future, | |
416 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 DF |
417 | * result in breaking guarantees promised to other tasks (refer to |
418 | * Documentation/scheduler/sched-deadline.txt for more informations). | |
aab03e05 DF |
419 | * |
420 | * This function returns true if: | |
421 | * | |
755378a4 | 422 | * runtime / (deadline - t) > dl_runtime / dl_period , |
aab03e05 DF |
423 | * |
424 | * IOW we can't recycle current parameters. | |
755378a4 HG |
425 | * |
426 | * Notice that the bandwidth check is done against the period. For | |
427 | * task with deadline equal to period this is the same of using | |
428 | * dl_deadline instead of dl_period in the equation above. | |
aab03e05 | 429 | */ |
2d3d891d DF |
430 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
431 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
432 | { |
433 | u64 left, right; | |
434 | ||
435 | /* | |
436 | * left and right are the two sides of the equation above, | |
437 | * after a bit of shuffling to use multiplications instead | |
438 | * of divisions. | |
439 | * | |
440 | * Note that none of the time values involved in the two | |
441 | * multiplications are absolute: dl_deadline and dl_runtime | |
442 | * are the relative deadline and the maximum runtime of each | |
443 | * instance, runtime is the runtime left for the last instance | |
444 | * and (deadline - t), since t is rq->clock, is the time left | |
445 | * to the (absolute) deadline. Even if overflowing the u64 type | |
446 | * is very unlikely to occur in both cases, here we scale down | |
447 | * as we want to avoid that risk at all. Scaling down by 10 | |
448 | * means that we reduce granularity to 1us. We are fine with it, | |
449 | * since this is only a true/false check and, anyway, thinking | |
450 | * of anything below microseconds resolution is actually fiction | |
451 | * (but still we want to give the user that illusion >;). | |
452 | */ | |
332ac17e DF |
453 | left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
454 | right = ((dl_se->deadline - t) >> DL_SCALE) * | |
455 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
456 | |
457 | return dl_time_before(right, left); | |
458 | } | |
459 | ||
460 | /* | |
461 | * When a -deadline entity is queued back on the runqueue, its runtime and | |
462 | * deadline might need updating. | |
463 | * | |
464 | * The policy here is that we update the deadline of the entity only if: | |
465 | * - the current deadline is in the past, | |
466 | * - using the remaining runtime with the current deadline would make | |
467 | * the entity exceed its bandwidth. | |
468 | */ | |
2d3d891d DF |
469 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
470 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
471 | { |
472 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
473 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
474 | ||
475 | /* | |
476 | * The arrival of a new instance needs special treatment, i.e., | |
477 | * the actual scheduling parameters have to be "renewed". | |
478 | */ | |
479 | if (dl_se->dl_new) { | |
2d3d891d | 480 | setup_new_dl_entity(dl_se, pi_se); |
aab03e05 DF |
481 | return; |
482 | } | |
483 | ||
484 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || | |
2d3d891d DF |
485 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
486 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
487 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
488 | } |
489 | } | |
490 | ||
491 | /* | |
492 | * If the entity depleted all its runtime, and if we want it to sleep | |
493 | * while waiting for some new execution time to become available, we | |
494 | * set the bandwidth enforcement timer to the replenishment instant | |
495 | * and try to activate it. | |
496 | * | |
497 | * Notice that it is important for the caller to know if the timer | |
498 | * actually started or not (i.e., the replenishment instant is in | |
499 | * the future or in the past). | |
500 | */ | |
2d3d891d | 501 | static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted) |
aab03e05 DF |
502 | { |
503 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
504 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
505 | ktime_t now, act; | |
506 | ktime_t soft, hard; | |
507 | unsigned long range; | |
508 | s64 delta; | |
509 | ||
2d3d891d DF |
510 | if (boosted) |
511 | return 0; | |
aab03e05 DF |
512 | /* |
513 | * We want the timer to fire at the deadline, but considering | |
514 | * that it is actually coming from rq->clock and not from | |
515 | * hrtimer's time base reading. | |
516 | */ | |
517 | act = ns_to_ktime(dl_se->deadline); | |
518 | now = hrtimer_cb_get_time(&dl_se->dl_timer); | |
519 | delta = ktime_to_ns(now) - rq_clock(rq); | |
520 | act = ktime_add_ns(act, delta); | |
521 | ||
522 | /* | |
523 | * If the expiry time already passed, e.g., because the value | |
524 | * chosen as the deadline is too small, don't even try to | |
525 | * start the timer in the past! | |
526 | */ | |
527 | if (ktime_us_delta(act, now) < 0) | |
528 | return 0; | |
529 | ||
530 | hrtimer_set_expires(&dl_se->dl_timer, act); | |
531 | ||
532 | soft = hrtimer_get_softexpires(&dl_se->dl_timer); | |
533 | hard = hrtimer_get_expires(&dl_se->dl_timer); | |
534 | range = ktime_to_ns(ktime_sub(hard, soft)); | |
535 | __hrtimer_start_range_ns(&dl_se->dl_timer, soft, | |
536 | range, HRTIMER_MODE_ABS, 0); | |
537 | ||
538 | return hrtimer_active(&dl_se->dl_timer); | |
539 | } | |
540 | ||
541 | /* | |
542 | * This is the bandwidth enforcement timer callback. If here, we know | |
543 | * a task is not on its dl_rq, since the fact that the timer was running | |
544 | * means the task is throttled and needs a runtime replenishment. | |
545 | * | |
546 | * However, what we actually do depends on the fact the task is active, | |
547 | * (it is on its rq) or has been removed from there by a call to | |
548 | * dequeue_task_dl(). In the former case we must issue the runtime | |
549 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
550 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
551 | * updating (and the queueing back to dl_rq) will be done by the | |
552 | * next call to enqueue_task_dl(). | |
553 | */ | |
554 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
555 | { | |
556 | struct sched_dl_entity *dl_se = container_of(timer, | |
557 | struct sched_dl_entity, | |
558 | dl_timer); | |
559 | struct task_struct *p = dl_task_of(dl_se); | |
3960c8c0 | 560 | unsigned long flags; |
0f397f2c | 561 | struct rq *rq; |
3960c8c0 | 562 | |
4cd57f97 | 563 | rq = task_rq_lock(p, &flags); |
0f397f2c | 564 | |
aab03e05 | 565 | /* |
aee38ea9 JL |
566 | * We need to take care of several possible races here: |
567 | * | |
568 | * - the task might have changed its scheduling policy | |
569 | * to something different than SCHED_DEADLINE | |
570 | * - the task might have changed its reservation parameters | |
571 | * (through sched_setattr()) | |
572 | * - the task might have been boosted by someone else and | |
573 | * might be in the boosting/deboosting path | |
574 | * | |
575 | * In all this cases we bail out, as the task is already | |
576 | * in the runqueue or is going to be enqueued back anyway. | |
aab03e05 | 577 | */ |
aee38ea9 JL |
578 | if (!dl_task(p) || dl_se->dl_new || |
579 | dl_se->dl_boosted || !dl_se->dl_throttled) | |
aab03e05 DF |
580 | goto unlock; |
581 | ||
582 | sched_clock_tick(); | |
583 | update_rq_clock(rq); | |
a79ec89f | 584 | |
fa9c9d10 WL |
585 | #ifdef CONFIG_SMP |
586 | /* | |
587 | * If we find that the rq the task was on is no longer | |
588 | * available, we need to select a new rq. | |
589 | */ | |
590 | if (unlikely(!rq->online)) { | |
591 | dl_task_offline_migration(rq, p); | |
592 | goto unlock; | |
593 | } | |
594 | #endif | |
595 | ||
a79ec89f KT |
596 | /* |
597 | * If the throttle happened during sched-out; like: | |
598 | * | |
599 | * schedule() | |
600 | * deactivate_task() | |
601 | * dequeue_task_dl() | |
602 | * update_curr_dl() | |
603 | * start_dl_timer() | |
604 | * __dequeue_task_dl() | |
605 | * prev->on_rq = 0; | |
606 | * | |
607 | * We can be both throttled and !queued. Replenish the counter | |
608 | * but do not enqueue -- wait for our wakeup to do that. | |
609 | */ | |
610 | if (!task_on_rq_queued(p)) { | |
611 | replenish_dl_entity(dl_se, dl_se); | |
612 | goto unlock; | |
613 | } | |
614 | ||
1019a359 PZ |
615 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
616 | if (dl_task(rq->curr)) | |
617 | check_preempt_curr_dl(rq, p, 0); | |
618 | else | |
619 | resched_curr(rq); | |
1baca4ce | 620 | #ifdef CONFIG_SMP |
1019a359 PZ |
621 | /* |
622 | * Queueing this task back might have overloaded rq, | |
623 | * check if we need to kick someone away. | |
624 | */ | |
625 | if (has_pushable_dl_tasks(rq)) | |
626 | push_dl_task(rq); | |
1baca4ce | 627 | #endif |
aab03e05 | 628 | unlock: |
4cd57f97 | 629 | task_rq_unlock(rq, p, &flags); |
aab03e05 DF |
630 | |
631 | return HRTIMER_NORESTART; | |
632 | } | |
633 | ||
634 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
635 | { | |
636 | struct hrtimer *timer = &dl_se->dl_timer; | |
637 | ||
aab03e05 DF |
638 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
639 | timer->function = dl_task_timer; | |
640 | } | |
641 | ||
642 | static | |
643 | int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) | |
644 | { | |
269ad801 | 645 | return (dl_se->runtime <= 0); |
aab03e05 DF |
646 | } |
647 | ||
faa59937 JL |
648 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
649 | ||
aab03e05 DF |
650 | /* |
651 | * Update the current task's runtime statistics (provided it is still | |
652 | * a -deadline task and has not been removed from the dl_rq). | |
653 | */ | |
654 | static void update_curr_dl(struct rq *rq) | |
655 | { | |
656 | struct task_struct *curr = rq->curr; | |
657 | struct sched_dl_entity *dl_se = &curr->dl; | |
658 | u64 delta_exec; | |
659 | ||
660 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
661 | return; | |
662 | ||
663 | /* | |
664 | * Consumed budget is computed considering the time as | |
665 | * observed by schedulable tasks (excluding time spent | |
666 | * in hardirq context, etc.). Deadlines are instead | |
667 | * computed using hard walltime. This seems to be the more | |
668 | * natural solution, but the full ramifications of this | |
669 | * approach need further study. | |
670 | */ | |
671 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; | |
734ff2a7 KT |
672 | if (unlikely((s64)delta_exec <= 0)) |
673 | return; | |
aab03e05 DF |
674 | |
675 | schedstat_set(curr->se.statistics.exec_max, | |
676 | max(curr->se.statistics.exec_max, delta_exec)); | |
677 | ||
678 | curr->se.sum_exec_runtime += delta_exec; | |
679 | account_group_exec_runtime(curr, delta_exec); | |
680 | ||
681 | curr->se.exec_start = rq_clock_task(rq); | |
682 | cpuacct_charge(curr, delta_exec); | |
683 | ||
239be4a9 DF |
684 | sched_rt_avg_update(rq, delta_exec); |
685 | ||
80496880 | 686 | dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec; |
aab03e05 | 687 | if (dl_runtime_exceeded(rq, dl_se)) { |
1019a359 | 688 | dl_se->dl_throttled = 1; |
aab03e05 | 689 | __dequeue_task_dl(rq, curr, 0); |
1019a359 | 690 | if (unlikely(!start_dl_timer(dl_se, curr->dl.dl_boosted))) |
aab03e05 DF |
691 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
692 | ||
693 | if (!is_leftmost(curr, &rq->dl)) | |
8875125e | 694 | resched_curr(rq); |
aab03e05 | 695 | } |
1724813d PZ |
696 | |
697 | /* | |
698 | * Because -- for now -- we share the rt bandwidth, we need to | |
699 | * account our runtime there too, otherwise actual rt tasks | |
700 | * would be able to exceed the shared quota. | |
701 | * | |
702 | * Account to the root rt group for now. | |
703 | * | |
704 | * The solution we're working towards is having the RT groups scheduled | |
705 | * using deadline servers -- however there's a few nasties to figure | |
706 | * out before that can happen. | |
707 | */ | |
708 | if (rt_bandwidth_enabled()) { | |
709 | struct rt_rq *rt_rq = &rq->rt; | |
710 | ||
711 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
1724813d PZ |
712 | /* |
713 | * We'll let actual RT tasks worry about the overflow here, we | |
faa59937 JL |
714 | * have our own CBS to keep us inline; only account when RT |
715 | * bandwidth is relevant. | |
1724813d | 716 | */ |
faa59937 JL |
717 | if (sched_rt_bandwidth_account(rt_rq)) |
718 | rt_rq->rt_time += delta_exec; | |
1724813d PZ |
719 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
720 | } | |
aab03e05 DF |
721 | } |
722 | ||
1baca4ce JL |
723 | #ifdef CONFIG_SMP |
724 | ||
725 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu); | |
726 | ||
727 | static inline u64 next_deadline(struct rq *rq) | |
728 | { | |
729 | struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu); | |
730 | ||
731 | if (next && dl_prio(next->prio)) | |
732 | return next->dl.deadline; | |
733 | else | |
734 | return 0; | |
735 | } | |
736 | ||
737 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
738 | { | |
739 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
740 | ||
741 | if (dl_rq->earliest_dl.curr == 0 || | |
742 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
743 | /* | |
744 | * If the dl_rq had no -deadline tasks, or if the new task | |
745 | * has shorter deadline than the current one on dl_rq, we | |
746 | * know that the previous earliest becomes our next earliest, | |
747 | * as the new task becomes the earliest itself. | |
748 | */ | |
749 | dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr; | |
750 | dl_rq->earliest_dl.curr = deadline; | |
6bfd6d72 | 751 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); |
1baca4ce JL |
752 | } else if (dl_rq->earliest_dl.next == 0 || |
753 | dl_time_before(deadline, dl_rq->earliest_dl.next)) { | |
754 | /* | |
755 | * On the other hand, if the new -deadline task has a | |
756 | * a later deadline than the earliest one on dl_rq, but | |
757 | * it is earlier than the next (if any), we must | |
758 | * recompute the next-earliest. | |
759 | */ | |
760 | dl_rq->earliest_dl.next = next_deadline(rq); | |
761 | } | |
762 | } | |
763 | ||
764 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
765 | { | |
766 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
767 | ||
768 | /* | |
769 | * Since we may have removed our earliest (and/or next earliest) | |
770 | * task we must recompute them. | |
771 | */ | |
772 | if (!dl_rq->dl_nr_running) { | |
773 | dl_rq->earliest_dl.curr = 0; | |
774 | dl_rq->earliest_dl.next = 0; | |
6bfd6d72 | 775 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); |
1baca4ce JL |
776 | } else { |
777 | struct rb_node *leftmost = dl_rq->rb_leftmost; | |
778 | struct sched_dl_entity *entry; | |
779 | ||
780 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
781 | dl_rq->earliest_dl.curr = entry->deadline; | |
782 | dl_rq->earliest_dl.next = next_deadline(rq); | |
6bfd6d72 | 783 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); |
1baca4ce JL |
784 | } |
785 | } | |
786 | ||
787 | #else | |
788 | ||
789 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
790 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
791 | ||
792 | #endif /* CONFIG_SMP */ | |
793 | ||
794 | static inline | |
795 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
796 | { | |
797 | int prio = dl_task_of(dl_se)->prio; | |
798 | u64 deadline = dl_se->deadline; | |
799 | ||
800 | WARN_ON(!dl_prio(prio)); | |
801 | dl_rq->dl_nr_running++; | |
72465447 | 802 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
803 | |
804 | inc_dl_deadline(dl_rq, deadline); | |
805 | inc_dl_migration(dl_se, dl_rq); | |
806 | } | |
807 | ||
808 | static inline | |
809 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
810 | { | |
811 | int prio = dl_task_of(dl_se)->prio; | |
812 | ||
813 | WARN_ON(!dl_prio(prio)); | |
814 | WARN_ON(!dl_rq->dl_nr_running); | |
815 | dl_rq->dl_nr_running--; | |
72465447 | 816 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
817 | |
818 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
819 | dec_dl_migration(dl_se, dl_rq); | |
820 | } | |
821 | ||
aab03e05 DF |
822 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
823 | { | |
824 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
825 | struct rb_node **link = &dl_rq->rb_root.rb_node; | |
826 | struct rb_node *parent = NULL; | |
827 | struct sched_dl_entity *entry; | |
828 | int leftmost = 1; | |
829 | ||
830 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
831 | ||
832 | while (*link) { | |
833 | parent = *link; | |
834 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
835 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
836 | link = &parent->rb_left; | |
837 | else { | |
838 | link = &parent->rb_right; | |
839 | leftmost = 0; | |
840 | } | |
841 | } | |
842 | ||
843 | if (leftmost) | |
844 | dl_rq->rb_leftmost = &dl_se->rb_node; | |
845 | ||
846 | rb_link_node(&dl_se->rb_node, parent, link); | |
847 | rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); | |
848 | ||
1baca4ce | 849 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
850 | } |
851 | ||
852 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
853 | { | |
854 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
855 | ||
856 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
857 | return; | |
858 | ||
859 | if (dl_rq->rb_leftmost == &dl_se->rb_node) { | |
860 | struct rb_node *next_node; | |
861 | ||
862 | next_node = rb_next(&dl_se->rb_node); | |
863 | dl_rq->rb_leftmost = next_node; | |
864 | } | |
865 | ||
866 | rb_erase(&dl_se->rb_node, &dl_rq->rb_root); | |
867 | RB_CLEAR_NODE(&dl_se->rb_node); | |
868 | ||
1baca4ce | 869 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
870 | } |
871 | ||
872 | static void | |
2d3d891d DF |
873 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
874 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
875 | { |
876 | BUG_ON(on_dl_rq(dl_se)); | |
877 | ||
878 | /* | |
879 | * If this is a wakeup or a new instance, the scheduling | |
880 | * parameters of the task might need updating. Otherwise, | |
881 | * we want a replenishment of its runtime. | |
882 | */ | |
6a503c3b | 883 | if (dl_se->dl_new || flags & ENQUEUE_WAKEUP) |
2d3d891d | 884 | update_dl_entity(dl_se, pi_se); |
6a503c3b LA |
885 | else if (flags & ENQUEUE_REPLENISH) |
886 | replenish_dl_entity(dl_se, pi_se); | |
aab03e05 DF |
887 | |
888 | __enqueue_dl_entity(dl_se); | |
889 | } | |
890 | ||
891 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
892 | { | |
893 | __dequeue_dl_entity(dl_se); | |
894 | } | |
895 | ||
896 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
897 | { | |
2d3d891d DF |
898 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
899 | struct sched_dl_entity *pi_se = &p->dl; | |
900 | ||
901 | /* | |
902 | * Use the scheduling parameters of the top pi-waiter | |
903 | * task if we have one and its (relative) deadline is | |
904 | * smaller than our one... OTW we keep our runtime and | |
905 | * deadline. | |
906 | */ | |
64be6f1f | 907 | if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) { |
2d3d891d | 908 | pi_se = &pi_task->dl; |
64be6f1f JL |
909 | } else if (!dl_prio(p->normal_prio)) { |
910 | /* | |
911 | * Special case in which we have a !SCHED_DEADLINE task | |
912 | * that is going to be deboosted, but exceedes its | |
913 | * runtime while doing so. No point in replenishing | |
914 | * it, as it's going to return back to its original | |
915 | * scheduling class after this. | |
916 | */ | |
917 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | |
918 | return; | |
919 | } | |
2d3d891d | 920 | |
aab03e05 DF |
921 | /* |
922 | * If p is throttled, we do nothing. In fact, if it exhausted | |
923 | * its budget it needs a replenishment and, since it now is on | |
924 | * its rq, the bandwidth timer callback (which clearly has not | |
925 | * run yet) will take care of this. | |
926 | */ | |
1019a359 | 927 | if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) |
aab03e05 DF |
928 | return; |
929 | ||
2d3d891d | 930 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce JL |
931 | |
932 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) | |
933 | enqueue_pushable_dl_task(rq, p); | |
aab03e05 DF |
934 | } |
935 | ||
936 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
937 | { | |
938 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 939 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
940 | } |
941 | ||
942 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
943 | { | |
944 | update_curr_dl(rq); | |
945 | __dequeue_task_dl(rq, p, flags); | |
aab03e05 DF |
946 | } |
947 | ||
948 | /* | |
949 | * Yield task semantic for -deadline tasks is: | |
950 | * | |
951 | * get off from the CPU until our next instance, with | |
952 | * a new runtime. This is of little use now, since we | |
953 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
954 | * bandwidth reclaiming is planned for the future, and | |
955 | * yield_task_dl will indicate that some spare budget | |
956 | * is available for other task instances to use it. | |
957 | */ | |
958 | static void yield_task_dl(struct rq *rq) | |
959 | { | |
960 | struct task_struct *p = rq->curr; | |
961 | ||
962 | /* | |
963 | * We make the task go to sleep until its current deadline by | |
964 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
965 | * it and the bandwidth timer will wake it up and will give it | |
5bfd126e | 966 | * new scheduling parameters (thanks to dl_yielded=1). |
aab03e05 DF |
967 | */ |
968 | if (p->dl.runtime > 0) { | |
5bfd126e | 969 | rq->curr->dl.dl_yielded = 1; |
aab03e05 DF |
970 | p->dl.runtime = 0; |
971 | } | |
6f1607f1 | 972 | update_rq_clock(rq); |
aab03e05 | 973 | update_curr_dl(rq); |
44fb085b WL |
974 | /* |
975 | * Tell update_rq_clock() that we've just updated, | |
976 | * so we don't do microscopic update in schedule() | |
977 | * and double the fastpath cost. | |
978 | */ | |
979 | rq_clock_skip_update(rq, true); | |
aab03e05 DF |
980 | } |
981 | ||
1baca4ce JL |
982 | #ifdef CONFIG_SMP |
983 | ||
984 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
985 | |
986 | static int | |
987 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
988 | { | |
989 | struct task_struct *curr; | |
990 | struct rq *rq; | |
991 | ||
1d7e974c | 992 | if (sd_flag != SD_BALANCE_WAKE) |
1baca4ce JL |
993 | goto out; |
994 | ||
995 | rq = cpu_rq(cpu); | |
996 | ||
997 | rcu_read_lock(); | |
998 | curr = ACCESS_ONCE(rq->curr); /* unlocked access */ | |
999 | ||
1000 | /* | |
1001 | * If we are dealing with a -deadline task, we must | |
1002 | * decide where to wake it up. | |
1003 | * If it has a later deadline and the current task | |
1004 | * on this rq can't move (provided the waking task | |
1005 | * can!) we prefer to send it somewhere else. On the | |
1006 | * other hand, if it has a shorter deadline, we | |
1007 | * try to make it stay here, it might be important. | |
1008 | */ | |
1009 | if (unlikely(dl_task(curr)) && | |
1010 | (curr->nr_cpus_allowed < 2 || | |
1011 | !dl_entity_preempt(&p->dl, &curr->dl)) && | |
1012 | (p->nr_cpus_allowed > 1)) { | |
1013 | int target = find_later_rq(p); | |
1014 | ||
1015 | if (target != -1) | |
1016 | cpu = target; | |
1017 | } | |
1018 | rcu_read_unlock(); | |
1019 | ||
1020 | out: | |
1021 | return cpu; | |
1022 | } | |
1023 | ||
1024 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) | |
1025 | { | |
1026 | /* | |
1027 | * Current can't be migrated, useless to reschedule, | |
1028 | * let's hope p can move out. | |
1029 | */ | |
1030 | if (rq->curr->nr_cpus_allowed == 1 || | |
6bfd6d72 | 1031 | cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) |
1baca4ce JL |
1032 | return; |
1033 | ||
1034 | /* | |
1035 | * p is migratable, so let's not schedule it and | |
1036 | * see if it is pushed or pulled somewhere else. | |
1037 | */ | |
1038 | if (p->nr_cpus_allowed != 1 && | |
6bfd6d72 | 1039 | cpudl_find(&rq->rd->cpudl, p, NULL) != -1) |
1baca4ce JL |
1040 | return; |
1041 | ||
8875125e | 1042 | resched_curr(rq); |
1baca4ce JL |
1043 | } |
1044 | ||
38033c37 PZ |
1045 | static int pull_dl_task(struct rq *this_rq); |
1046 | ||
1baca4ce JL |
1047 | #endif /* CONFIG_SMP */ |
1048 | ||
aab03e05 DF |
1049 | /* |
1050 | * Only called when both the current and waking task are -deadline | |
1051 | * tasks. | |
1052 | */ | |
1053 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
1054 | int flags) | |
1055 | { | |
1baca4ce | 1056 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
8875125e | 1057 | resched_curr(rq); |
1baca4ce JL |
1058 | return; |
1059 | } | |
1060 | ||
1061 | #ifdef CONFIG_SMP | |
1062 | /* | |
1063 | * In the unlikely case current and p have the same deadline | |
1064 | * let us try to decide what's the best thing to do... | |
1065 | */ | |
332ac17e DF |
1066 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1067 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
1068 | check_preempt_equal_dl(rq, p); |
1069 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
1070 | } |
1071 | ||
1072 | #ifdef CONFIG_SCHED_HRTICK | |
1073 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1074 | { | |
177ef2a6 | 1075 | hrtick_start(rq, p->dl.runtime); |
aab03e05 | 1076 | } |
36ce9881 WL |
1077 | #else /* !CONFIG_SCHED_HRTICK */ |
1078 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1079 | { | |
1080 | } | |
aab03e05 DF |
1081 | #endif |
1082 | ||
1083 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, | |
1084 | struct dl_rq *dl_rq) | |
1085 | { | |
1086 | struct rb_node *left = dl_rq->rb_leftmost; | |
1087 | ||
1088 | if (!left) | |
1089 | return NULL; | |
1090 | ||
1091 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
1092 | } | |
1093 | ||
606dba2e | 1094 | struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev) |
aab03e05 DF |
1095 | { |
1096 | struct sched_dl_entity *dl_se; | |
1097 | struct task_struct *p; | |
1098 | struct dl_rq *dl_rq; | |
1099 | ||
1100 | dl_rq = &rq->dl; | |
1101 | ||
a1d9a323 | 1102 | if (need_pull_dl_task(rq, prev)) { |
38033c37 | 1103 | pull_dl_task(rq); |
a1d9a323 KT |
1104 | /* |
1105 | * pull_rt_task() can drop (and re-acquire) rq->lock; this | |
1106 | * means a stop task can slip in, in which case we need to | |
1107 | * re-start task selection. | |
1108 | */ | |
da0c1e65 | 1109 | if (rq->stop && task_on_rq_queued(rq->stop)) |
a1d9a323 KT |
1110 | return RETRY_TASK; |
1111 | } | |
1112 | ||
734ff2a7 KT |
1113 | /* |
1114 | * When prev is DL, we may throttle it in put_prev_task(). | |
1115 | * So, we update time before we check for dl_nr_running. | |
1116 | */ | |
1117 | if (prev->sched_class == &dl_sched_class) | |
1118 | update_curr_dl(rq); | |
38033c37 | 1119 | |
aab03e05 DF |
1120 | if (unlikely(!dl_rq->dl_nr_running)) |
1121 | return NULL; | |
1122 | ||
3f1d2a31 | 1123 | put_prev_task(rq, prev); |
606dba2e | 1124 | |
aab03e05 DF |
1125 | dl_se = pick_next_dl_entity(rq, dl_rq); |
1126 | BUG_ON(!dl_se); | |
1127 | ||
1128 | p = dl_task_of(dl_se); | |
1129 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1130 | |
1131 | /* Running task will never be pushed. */ | |
71362650 | 1132 | dequeue_pushable_dl_task(rq, p); |
1baca4ce | 1133 | |
aab03e05 DF |
1134 | if (hrtick_enabled(rq)) |
1135 | start_hrtick_dl(rq, p); | |
1baca4ce | 1136 | |
dc877341 | 1137 | set_post_schedule(rq); |
1baca4ce | 1138 | |
aab03e05 DF |
1139 | return p; |
1140 | } | |
1141 | ||
1142 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | |
1143 | { | |
1144 | update_curr_dl(rq); | |
1baca4ce JL |
1145 | |
1146 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) | |
1147 | enqueue_pushable_dl_task(rq, p); | |
aab03e05 DF |
1148 | } |
1149 | ||
1150 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) | |
1151 | { | |
1152 | update_curr_dl(rq); | |
1153 | ||
a7bebf48 WL |
1154 | /* |
1155 | * Even when we have runtime, update_curr_dl() might have resulted in us | |
1156 | * not being the leftmost task anymore. In that case NEED_RESCHED will | |
1157 | * be set and schedule() will start a new hrtick for the next task. | |
1158 | */ | |
1159 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && | |
1160 | is_leftmost(p, &rq->dl)) | |
aab03e05 | 1161 | start_hrtick_dl(rq, p); |
aab03e05 DF |
1162 | } |
1163 | ||
1164 | static void task_fork_dl(struct task_struct *p) | |
1165 | { | |
1166 | /* | |
1167 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1168 | * sched_fork() | |
1169 | */ | |
1170 | } | |
1171 | ||
1172 | static void task_dead_dl(struct task_struct *p) | |
1173 | { | |
1174 | struct hrtimer *timer = &p->dl.dl_timer; | |
332ac17e DF |
1175 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1176 | ||
1177 | /* | |
1178 | * Since we are TASK_DEAD we won't slip out of the domain! | |
1179 | */ | |
1180 | raw_spin_lock_irq(&dl_b->lock); | |
40767b0d | 1181 | /* XXX we should retain the bw until 0-lag */ |
332ac17e DF |
1182 | dl_b->total_bw -= p->dl.dl_bw; |
1183 | raw_spin_unlock_irq(&dl_b->lock); | |
aab03e05 | 1184 | |
2d3d891d | 1185 | hrtimer_cancel(timer); |
aab03e05 DF |
1186 | } |
1187 | ||
1188 | static void set_curr_task_dl(struct rq *rq) | |
1189 | { | |
1190 | struct task_struct *p = rq->curr; | |
1191 | ||
1192 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1193 | |
1194 | /* You can't push away the running task */ | |
1195 | dequeue_pushable_dl_task(rq, p); | |
1196 | } | |
1197 | ||
1198 | #ifdef CONFIG_SMP | |
1199 | ||
1200 | /* Only try algorithms three times */ | |
1201 | #define DL_MAX_TRIES 3 | |
1202 | ||
1203 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1204 | { | |
1205 | if (!task_running(rq, p) && | |
1ba93d42 | 1206 | cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
1baca4ce | 1207 | return 1; |
1baca4ce JL |
1208 | return 0; |
1209 | } | |
1210 | ||
1211 | /* Returns the second earliest -deadline task, NULL otherwise */ | |
1212 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu) | |
1213 | { | |
1214 | struct rb_node *next_node = rq->dl.rb_leftmost; | |
1215 | struct sched_dl_entity *dl_se; | |
1216 | struct task_struct *p = NULL; | |
1217 | ||
1218 | next_node: | |
1219 | next_node = rb_next(next_node); | |
1220 | if (next_node) { | |
1221 | dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node); | |
1222 | p = dl_task_of(dl_se); | |
1223 | ||
1224 | if (pick_dl_task(rq, p, cpu)) | |
1225 | return p; | |
1226 | ||
1227 | goto next_node; | |
1228 | } | |
1229 | ||
1230 | return NULL; | |
1231 | } | |
1232 | ||
1baca4ce JL |
1233 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1234 | ||
1235 | static int find_later_rq(struct task_struct *task) | |
1236 | { | |
1237 | struct sched_domain *sd; | |
4ba29684 | 1238 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1baca4ce JL |
1239 | int this_cpu = smp_processor_id(); |
1240 | int best_cpu, cpu = task_cpu(task); | |
1241 | ||
1242 | /* Make sure the mask is initialized first */ | |
1243 | if (unlikely(!later_mask)) | |
1244 | return -1; | |
1245 | ||
1246 | if (task->nr_cpus_allowed == 1) | |
1247 | return -1; | |
1248 | ||
91ec6778 JL |
1249 | /* |
1250 | * We have to consider system topology and task affinity | |
1251 | * first, then we can look for a suitable cpu. | |
1252 | */ | |
6bfd6d72 JL |
1253 | best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, |
1254 | task, later_mask); | |
1baca4ce JL |
1255 | if (best_cpu == -1) |
1256 | return -1; | |
1257 | ||
1258 | /* | |
1259 | * If we are here, some target has been found, | |
1260 | * the most suitable of which is cached in best_cpu. | |
1261 | * This is, among the runqueues where the current tasks | |
1262 | * have later deadlines than the task's one, the rq | |
1263 | * with the latest possible one. | |
1264 | * | |
1265 | * Now we check how well this matches with task's | |
1266 | * affinity and system topology. | |
1267 | * | |
1268 | * The last cpu where the task run is our first | |
1269 | * guess, since it is most likely cache-hot there. | |
1270 | */ | |
1271 | if (cpumask_test_cpu(cpu, later_mask)) | |
1272 | return cpu; | |
1273 | /* | |
1274 | * Check if this_cpu is to be skipped (i.e., it is | |
1275 | * not in the mask) or not. | |
1276 | */ | |
1277 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1278 | this_cpu = -1; | |
1279 | ||
1280 | rcu_read_lock(); | |
1281 | for_each_domain(cpu, sd) { | |
1282 | if (sd->flags & SD_WAKE_AFFINE) { | |
1283 | ||
1284 | /* | |
1285 | * If possible, preempting this_cpu is | |
1286 | * cheaper than migrating. | |
1287 | */ | |
1288 | if (this_cpu != -1 && | |
1289 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1290 | rcu_read_unlock(); | |
1291 | return this_cpu; | |
1292 | } | |
1293 | ||
1294 | /* | |
1295 | * Last chance: if best_cpu is valid and is | |
1296 | * in the mask, that becomes our choice. | |
1297 | */ | |
1298 | if (best_cpu < nr_cpu_ids && | |
1299 | cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { | |
1300 | rcu_read_unlock(); | |
1301 | return best_cpu; | |
1302 | } | |
1303 | } | |
1304 | } | |
1305 | rcu_read_unlock(); | |
1306 | ||
1307 | /* | |
1308 | * At this point, all our guesses failed, we just return | |
1309 | * 'something', and let the caller sort the things out. | |
1310 | */ | |
1311 | if (this_cpu != -1) | |
1312 | return this_cpu; | |
1313 | ||
1314 | cpu = cpumask_any(later_mask); | |
1315 | if (cpu < nr_cpu_ids) | |
1316 | return cpu; | |
1317 | ||
1318 | return -1; | |
1319 | } | |
1320 | ||
1321 | /* Locks the rq it finds */ | |
1322 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
1323 | { | |
1324 | struct rq *later_rq = NULL; | |
1325 | int tries; | |
1326 | int cpu; | |
1327 | ||
1328 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
1329 | cpu = find_later_rq(task); | |
1330 | ||
1331 | if ((cpu == -1) || (cpu == rq->cpu)) | |
1332 | break; | |
1333 | ||
1334 | later_rq = cpu_rq(cpu); | |
1335 | ||
1336 | /* Retry if something changed. */ | |
1337 | if (double_lock_balance(rq, later_rq)) { | |
1338 | if (unlikely(task_rq(task) != rq || | |
1339 | !cpumask_test_cpu(later_rq->cpu, | |
1340 | &task->cpus_allowed) || | |
da0c1e65 KT |
1341 | task_running(rq, task) || |
1342 | !task_on_rq_queued(task))) { | |
1baca4ce JL |
1343 | double_unlock_balance(rq, later_rq); |
1344 | later_rq = NULL; | |
1345 | break; | |
1346 | } | |
1347 | } | |
1348 | ||
1349 | /* | |
1350 | * If the rq we found has no -deadline task, or | |
1351 | * its earliest one has a later deadline than our | |
1352 | * task, the rq is a good one. | |
1353 | */ | |
1354 | if (!later_rq->dl.dl_nr_running || | |
1355 | dl_time_before(task->dl.deadline, | |
1356 | later_rq->dl.earliest_dl.curr)) | |
1357 | break; | |
1358 | ||
1359 | /* Otherwise we try again. */ | |
1360 | double_unlock_balance(rq, later_rq); | |
1361 | later_rq = NULL; | |
1362 | } | |
1363 | ||
1364 | return later_rq; | |
1365 | } | |
1366 | ||
1367 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
1368 | { | |
1369 | struct task_struct *p; | |
1370 | ||
1371 | if (!has_pushable_dl_tasks(rq)) | |
1372 | return NULL; | |
1373 | ||
1374 | p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, | |
1375 | struct task_struct, pushable_dl_tasks); | |
1376 | ||
1377 | BUG_ON(rq->cpu != task_cpu(p)); | |
1378 | BUG_ON(task_current(rq, p)); | |
1379 | BUG_ON(p->nr_cpus_allowed <= 1); | |
1380 | ||
da0c1e65 | 1381 | BUG_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
1382 | BUG_ON(!dl_task(p)); |
1383 | ||
1384 | return p; | |
1385 | } | |
1386 | ||
1387 | /* | |
1388 | * See if the non running -deadline tasks on this rq | |
1389 | * can be sent to some other CPU where they can preempt | |
1390 | * and start executing. | |
1391 | */ | |
1392 | static int push_dl_task(struct rq *rq) | |
1393 | { | |
1394 | struct task_struct *next_task; | |
1395 | struct rq *later_rq; | |
c51b8ab5 | 1396 | int ret = 0; |
1baca4ce JL |
1397 | |
1398 | if (!rq->dl.overloaded) | |
1399 | return 0; | |
1400 | ||
1401 | next_task = pick_next_pushable_dl_task(rq); | |
1402 | if (!next_task) | |
1403 | return 0; | |
1404 | ||
1405 | retry: | |
1406 | if (unlikely(next_task == rq->curr)) { | |
1407 | WARN_ON(1); | |
1408 | return 0; | |
1409 | } | |
1410 | ||
1411 | /* | |
1412 | * If next_task preempts rq->curr, and rq->curr | |
1413 | * can move away, it makes sense to just reschedule | |
1414 | * without going further in pushing next_task. | |
1415 | */ | |
1416 | if (dl_task(rq->curr) && | |
1417 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
1418 | rq->curr->nr_cpus_allowed > 1) { | |
8875125e | 1419 | resched_curr(rq); |
1baca4ce JL |
1420 | return 0; |
1421 | } | |
1422 | ||
1423 | /* We might release rq lock */ | |
1424 | get_task_struct(next_task); | |
1425 | ||
1426 | /* Will lock the rq it'll find */ | |
1427 | later_rq = find_lock_later_rq(next_task, rq); | |
1428 | if (!later_rq) { | |
1429 | struct task_struct *task; | |
1430 | ||
1431 | /* | |
1432 | * We must check all this again, since | |
1433 | * find_lock_later_rq releases rq->lock and it is | |
1434 | * then possible that next_task has migrated. | |
1435 | */ | |
1436 | task = pick_next_pushable_dl_task(rq); | |
1437 | if (task_cpu(next_task) == rq->cpu && task == next_task) { | |
1438 | /* | |
1439 | * The task is still there. We don't try | |
1440 | * again, some other cpu will pull it when ready. | |
1441 | */ | |
1baca4ce JL |
1442 | goto out; |
1443 | } | |
1444 | ||
1445 | if (!task) | |
1446 | /* No more tasks */ | |
1447 | goto out; | |
1448 | ||
1449 | put_task_struct(next_task); | |
1450 | next_task = task; | |
1451 | goto retry; | |
1452 | } | |
1453 | ||
1454 | deactivate_task(rq, next_task, 0); | |
1455 | set_task_cpu(next_task, later_rq->cpu); | |
1456 | activate_task(later_rq, next_task, 0); | |
c51b8ab5 | 1457 | ret = 1; |
1baca4ce | 1458 | |
8875125e | 1459 | resched_curr(later_rq); |
1baca4ce JL |
1460 | |
1461 | double_unlock_balance(rq, later_rq); | |
1462 | ||
1463 | out: | |
1464 | put_task_struct(next_task); | |
1465 | ||
c51b8ab5 | 1466 | return ret; |
1baca4ce JL |
1467 | } |
1468 | ||
1469 | static void push_dl_tasks(struct rq *rq) | |
1470 | { | |
1471 | /* Terminates as it moves a -deadline task */ | |
1472 | while (push_dl_task(rq)) | |
1473 | ; | |
aab03e05 DF |
1474 | } |
1475 | ||
1baca4ce JL |
1476 | static int pull_dl_task(struct rq *this_rq) |
1477 | { | |
1478 | int this_cpu = this_rq->cpu, ret = 0, cpu; | |
1479 | struct task_struct *p; | |
1480 | struct rq *src_rq; | |
1481 | u64 dmin = LONG_MAX; | |
1482 | ||
1483 | if (likely(!dl_overloaded(this_rq))) | |
1484 | return 0; | |
1485 | ||
1486 | /* | |
1487 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
1488 | * see overloaded we must also see the dlo_mask bit. | |
1489 | */ | |
1490 | smp_rmb(); | |
1491 | ||
1492 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
1493 | if (this_cpu == cpu) | |
1494 | continue; | |
1495 | ||
1496 | src_rq = cpu_rq(cpu); | |
1497 | ||
1498 | /* | |
1499 | * It looks racy, abd it is! However, as in sched_rt.c, | |
1500 | * we are fine with this. | |
1501 | */ | |
1502 | if (this_rq->dl.dl_nr_running && | |
1503 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
1504 | src_rq->dl.earliest_dl.next)) | |
1505 | continue; | |
1506 | ||
1507 | /* Might drop this_rq->lock */ | |
1508 | double_lock_balance(this_rq, src_rq); | |
1509 | ||
1510 | /* | |
1511 | * If there are no more pullable tasks on the | |
1512 | * rq, we're done with it. | |
1513 | */ | |
1514 | if (src_rq->dl.dl_nr_running <= 1) | |
1515 | goto skip; | |
1516 | ||
1517 | p = pick_next_earliest_dl_task(src_rq, this_cpu); | |
1518 | ||
1519 | /* | |
1520 | * We found a task to be pulled if: | |
1521 | * - it preempts our current (if there's one), | |
1522 | * - it will preempt the last one we pulled (if any). | |
1523 | */ | |
1524 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
1525 | (!this_rq->dl.dl_nr_running || | |
1526 | dl_time_before(p->dl.deadline, | |
1527 | this_rq->dl.earliest_dl.curr))) { | |
1528 | WARN_ON(p == src_rq->curr); | |
da0c1e65 | 1529 | WARN_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
1530 | |
1531 | /* | |
1532 | * Then we pull iff p has actually an earlier | |
1533 | * deadline than the current task of its runqueue. | |
1534 | */ | |
1535 | if (dl_time_before(p->dl.deadline, | |
1536 | src_rq->curr->dl.deadline)) | |
1537 | goto skip; | |
1538 | ||
1539 | ret = 1; | |
1540 | ||
1541 | deactivate_task(src_rq, p, 0); | |
1542 | set_task_cpu(p, this_cpu); | |
1543 | activate_task(this_rq, p, 0); | |
1544 | dmin = p->dl.deadline; | |
1545 | ||
1546 | /* Is there any other task even earlier? */ | |
1547 | } | |
1548 | skip: | |
1549 | double_unlock_balance(this_rq, src_rq); | |
1550 | } | |
1551 | ||
1552 | return ret; | |
1553 | } | |
1554 | ||
1baca4ce JL |
1555 | static void post_schedule_dl(struct rq *rq) |
1556 | { | |
1557 | push_dl_tasks(rq); | |
1558 | } | |
1559 | ||
1560 | /* | |
1561 | * Since the task is not running and a reschedule is not going to happen | |
1562 | * anytime soon on its runqueue, we try pushing it away now. | |
1563 | */ | |
1564 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
1565 | { | |
1566 | if (!task_running(rq, p) && | |
1567 | !test_tsk_need_resched(rq->curr) && | |
1568 | has_pushable_dl_tasks(rq) && | |
1569 | p->nr_cpus_allowed > 1 && | |
1570 | dl_task(rq->curr) && | |
1571 | (rq->curr->nr_cpus_allowed < 2 || | |
6b0a563f | 1572 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1baca4ce JL |
1573 | push_dl_tasks(rq); |
1574 | } | |
1575 | } | |
1576 | ||
1577 | static void set_cpus_allowed_dl(struct task_struct *p, | |
1578 | const struct cpumask *new_mask) | |
1579 | { | |
1580 | struct rq *rq; | |
7f51412a | 1581 | struct root_domain *src_rd; |
1baca4ce JL |
1582 | int weight; |
1583 | ||
1584 | BUG_ON(!dl_task(p)); | |
1585 | ||
7f51412a JL |
1586 | rq = task_rq(p); |
1587 | src_rd = rq->rd; | |
1588 | /* | |
1589 | * Migrating a SCHED_DEADLINE task between exclusive | |
1590 | * cpusets (different root_domains) entails a bandwidth | |
1591 | * update. We already made space for us in the destination | |
1592 | * domain (see cpuset_can_attach()). | |
1593 | */ | |
1594 | if (!cpumask_intersects(src_rd->span, new_mask)) { | |
1595 | struct dl_bw *src_dl_b; | |
1596 | ||
1597 | src_dl_b = dl_bw_of(cpu_of(rq)); | |
1598 | /* | |
1599 | * We now free resources of the root_domain we are migrating | |
1600 | * off. In the worst case, sched_setattr() may temporary fail | |
1601 | * until we complete the update. | |
1602 | */ | |
1603 | raw_spin_lock(&src_dl_b->lock); | |
1604 | __dl_clear(src_dl_b, p->dl.dl_bw); | |
1605 | raw_spin_unlock(&src_dl_b->lock); | |
1606 | } | |
1607 | ||
1baca4ce JL |
1608 | /* |
1609 | * Update only if the task is actually running (i.e., | |
1610 | * it is on the rq AND it is not throttled). | |
1611 | */ | |
1612 | if (!on_dl_rq(&p->dl)) | |
1613 | return; | |
1614 | ||
1615 | weight = cpumask_weight(new_mask); | |
1616 | ||
1617 | /* | |
1618 | * Only update if the process changes its state from whether it | |
1619 | * can migrate or not. | |
1620 | */ | |
1621 | if ((p->nr_cpus_allowed > 1) == (weight > 1)) | |
1622 | return; | |
1623 | ||
1baca4ce JL |
1624 | /* |
1625 | * The process used to be able to migrate OR it can now migrate | |
1626 | */ | |
1627 | if (weight <= 1) { | |
1628 | if (!task_current(rq, p)) | |
1629 | dequeue_pushable_dl_task(rq, p); | |
1630 | BUG_ON(!rq->dl.dl_nr_migratory); | |
1631 | rq->dl.dl_nr_migratory--; | |
1632 | } else { | |
1633 | if (!task_current(rq, p)) | |
1634 | enqueue_pushable_dl_task(rq, p); | |
1635 | rq->dl.dl_nr_migratory++; | |
1636 | } | |
1637 | ||
1638 | update_dl_migration(&rq->dl); | |
1639 | } | |
1640 | ||
1641 | /* Assumes rq->lock is held */ | |
1642 | static void rq_online_dl(struct rq *rq) | |
1643 | { | |
1644 | if (rq->dl.overloaded) | |
1645 | dl_set_overload(rq); | |
6bfd6d72 | 1646 | |
16b26943 | 1647 | cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); |
6bfd6d72 JL |
1648 | if (rq->dl.dl_nr_running > 0) |
1649 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); | |
1baca4ce JL |
1650 | } |
1651 | ||
1652 | /* Assumes rq->lock is held */ | |
1653 | static void rq_offline_dl(struct rq *rq) | |
1654 | { | |
1655 | if (rq->dl.overloaded) | |
1656 | dl_clear_overload(rq); | |
6bfd6d72 JL |
1657 | |
1658 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); | |
16b26943 | 1659 | cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); |
1baca4ce JL |
1660 | } |
1661 | ||
1662 | void init_sched_dl_class(void) | |
1663 | { | |
1664 | unsigned int i; | |
1665 | ||
1666 | for_each_possible_cpu(i) | |
1667 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
1668 | GFP_KERNEL, cpu_to_node(i)); | |
1669 | } | |
1670 | ||
1671 | #endif /* CONFIG_SMP */ | |
1672 | ||
67dfa1b7 KT |
1673 | /* |
1674 | * Ensure p's dl_timer is cancelled. May drop rq->lock for a while. | |
1675 | */ | |
1676 | static void cancel_dl_timer(struct rq *rq, struct task_struct *p) | |
1677 | { | |
1678 | struct hrtimer *dl_timer = &p->dl.dl_timer; | |
1679 | ||
1680 | /* Nobody will change task's class if pi_lock is held */ | |
1681 | lockdep_assert_held(&p->pi_lock); | |
1682 | ||
1683 | if (hrtimer_active(dl_timer)) { | |
1684 | int ret = hrtimer_try_to_cancel(dl_timer); | |
1685 | ||
1686 | if (unlikely(ret == -1)) { | |
1687 | /* | |
1688 | * Note, p may migrate OR new deadline tasks | |
1689 | * may appear in rq when we are unlocking it. | |
1690 | * A caller of us must be fine with that. | |
1691 | */ | |
1692 | raw_spin_unlock(&rq->lock); | |
1693 | hrtimer_cancel(dl_timer); | |
1694 | raw_spin_lock(&rq->lock); | |
1695 | } | |
1696 | } | |
1697 | } | |
1698 | ||
aab03e05 DF |
1699 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
1700 | { | |
40767b0d | 1701 | /* XXX we should retain the bw until 0-lag */ |
67dfa1b7 | 1702 | cancel_dl_timer(rq, p); |
a5e7be3b JL |
1703 | __dl_clear_params(p); |
1704 | ||
1baca4ce JL |
1705 | /* |
1706 | * Since this might be the only -deadline task on the rq, | |
1707 | * this is the right place to try to pull some other one | |
1708 | * from an overloaded cpu, if any. | |
1709 | */ | |
cd660911 WL |
1710 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
1711 | return; | |
1712 | ||
1713 | if (pull_dl_task(rq)) | |
1714 | resched_curr(rq); | |
aab03e05 DF |
1715 | } |
1716 | ||
1baca4ce JL |
1717 | /* |
1718 | * When switching to -deadline, we may overload the rq, then | |
1719 | * we try to push someone off, if possible. | |
1720 | */ | |
aab03e05 DF |
1721 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
1722 | { | |
1baca4ce JL |
1723 | int check_resched = 1; |
1724 | ||
da0c1e65 | 1725 | if (task_on_rq_queued(p) && rq->curr != p) { |
1baca4ce | 1726 | #ifdef CONFIG_SMP |
d9aade7a WL |
1727 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded && |
1728 | push_dl_task(rq) && rq != task_rq(p)) | |
1baca4ce JL |
1729 | /* Only reschedule if pushing failed */ |
1730 | check_resched = 0; | |
1731 | #endif /* CONFIG_SMP */ | |
f3a7e1a9 KT |
1732 | if (check_resched) { |
1733 | if (dl_task(rq->curr)) | |
1734 | check_preempt_curr_dl(rq, p, 0); | |
1735 | else | |
1736 | resched_curr(rq); | |
1737 | } | |
aab03e05 DF |
1738 | } |
1739 | } | |
1740 | ||
1baca4ce JL |
1741 | /* |
1742 | * If the scheduling parameters of a -deadline task changed, | |
1743 | * a push or pull operation might be needed. | |
1744 | */ | |
aab03e05 DF |
1745 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
1746 | int oldprio) | |
1747 | { | |
da0c1e65 | 1748 | if (task_on_rq_queued(p) || rq->curr == p) { |
aab03e05 | 1749 | #ifdef CONFIG_SMP |
1baca4ce JL |
1750 | /* |
1751 | * This might be too much, but unfortunately | |
1752 | * we don't have the old deadline value, and | |
1753 | * we can't argue if the task is increasing | |
1754 | * or lowering its prio, so... | |
1755 | */ | |
1756 | if (!rq->dl.overloaded) | |
1757 | pull_dl_task(rq); | |
1758 | ||
1759 | /* | |
1760 | * If we now have a earlier deadline task than p, | |
1761 | * then reschedule, provided p is still on this | |
1762 | * runqueue. | |
1763 | */ | |
1764 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) && | |
1765 | rq->curr == p) | |
8875125e | 1766 | resched_curr(rq); |
1baca4ce JL |
1767 | #else |
1768 | /* | |
1769 | * Again, we don't know if p has a earlier | |
1770 | * or later deadline, so let's blindly set a | |
1771 | * (maybe not needed) rescheduling point. | |
1772 | */ | |
8875125e | 1773 | resched_curr(rq); |
1baca4ce JL |
1774 | #endif /* CONFIG_SMP */ |
1775 | } else | |
1776 | switched_to_dl(rq, p); | |
aab03e05 | 1777 | } |
aab03e05 DF |
1778 | |
1779 | const struct sched_class dl_sched_class = { | |
1780 | .next = &rt_sched_class, | |
1781 | .enqueue_task = enqueue_task_dl, | |
1782 | .dequeue_task = dequeue_task_dl, | |
1783 | .yield_task = yield_task_dl, | |
1784 | ||
1785 | .check_preempt_curr = check_preempt_curr_dl, | |
1786 | ||
1787 | .pick_next_task = pick_next_task_dl, | |
1788 | .put_prev_task = put_prev_task_dl, | |
1789 | ||
1790 | #ifdef CONFIG_SMP | |
1791 | .select_task_rq = select_task_rq_dl, | |
1baca4ce JL |
1792 | .set_cpus_allowed = set_cpus_allowed_dl, |
1793 | .rq_online = rq_online_dl, | |
1794 | .rq_offline = rq_offline_dl, | |
1baca4ce JL |
1795 | .post_schedule = post_schedule_dl, |
1796 | .task_woken = task_woken_dl, | |
aab03e05 DF |
1797 | #endif |
1798 | ||
1799 | .set_curr_task = set_curr_task_dl, | |
1800 | .task_tick = task_tick_dl, | |
1801 | .task_fork = task_fork_dl, | |
1802 | .task_dead = task_dead_dl, | |
1803 | ||
1804 | .prio_changed = prio_changed_dl, | |
1805 | .switched_from = switched_from_dl, | |
1806 | .switched_to = switched_to_dl, | |
6e998916 SG |
1807 | |
1808 | .update_curr = update_curr_dl, | |
aab03e05 | 1809 | }; |
acb32132 WL |
1810 | |
1811 | #ifdef CONFIG_SCHED_DEBUG | |
1812 | extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); | |
1813 | ||
1814 | void print_dl_stats(struct seq_file *m, int cpu) | |
1815 | { | |
1816 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | |
1817 | } | |
1818 | #endif /* CONFIG_SCHED_DEBUG */ |