]>
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 | ||
7d92de3a | 179 | if (leftmost) { |
1baca4ce | 180 | dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; |
7d92de3a WL |
181 | dl_rq->earliest_dl.next = p->dl.deadline; |
182 | } | |
1baca4ce JL |
183 | |
184 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
185 | rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | |
aab03e05 DF |
186 | } |
187 | ||
1baca4ce JL |
188 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
189 | { | |
190 | struct dl_rq *dl_rq = &rq->dl; | |
191 | ||
192 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
193 | return; | |
194 | ||
195 | if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { | |
196 | struct rb_node *next_node; | |
197 | ||
198 | next_node = rb_next(&p->pushable_dl_tasks); | |
199 | dl_rq->pushable_dl_tasks_leftmost = next_node; | |
7d92de3a WL |
200 | if (next_node) { |
201 | dl_rq->earliest_dl.next = rb_entry(next_node, | |
202 | struct task_struct, pushable_dl_tasks)->dl.deadline; | |
203 | } | |
1baca4ce JL |
204 | } |
205 | ||
206 | rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | |
207 | RB_CLEAR_NODE(&p->pushable_dl_tasks); | |
208 | } | |
209 | ||
210 | static inline int has_pushable_dl_tasks(struct rq *rq) | |
211 | { | |
212 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); | |
213 | } | |
214 | ||
215 | static int push_dl_task(struct rq *rq); | |
216 | ||
dc877341 PZ |
217 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
218 | { | |
219 | return dl_task(prev); | |
220 | } | |
221 | ||
9916e214 PZ |
222 | static DEFINE_PER_CPU(struct callback_head, dl_push_head); |
223 | static DEFINE_PER_CPU(struct callback_head, dl_pull_head); | |
e3fca9e7 PZ |
224 | |
225 | static void push_dl_tasks(struct rq *); | |
9916e214 | 226 | static void pull_dl_task(struct rq *); |
e3fca9e7 PZ |
227 | |
228 | static inline void queue_push_tasks(struct rq *rq) | |
dc877341 | 229 | { |
e3fca9e7 PZ |
230 | if (!has_pushable_dl_tasks(rq)) |
231 | return; | |
232 | ||
9916e214 PZ |
233 | queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks); |
234 | } | |
235 | ||
236 | static inline void queue_pull_task(struct rq *rq) | |
237 | { | |
238 | queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task); | |
dc877341 PZ |
239 | } |
240 | ||
fa9c9d10 WL |
241 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); |
242 | ||
a649f237 | 243 | static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p) |
fa9c9d10 WL |
244 | { |
245 | struct rq *later_rq = NULL; | |
246 | bool fallback = false; | |
247 | ||
248 | later_rq = find_lock_later_rq(p, rq); | |
249 | ||
250 | if (!later_rq) { | |
251 | int cpu; | |
252 | ||
253 | /* | |
254 | * If we cannot preempt any rq, fall back to pick any | |
255 | * online cpu. | |
256 | */ | |
257 | fallback = true; | |
258 | cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p)); | |
259 | if (cpu >= nr_cpu_ids) { | |
260 | /* | |
261 | * Fail to find any suitable cpu. | |
262 | * The task will never come back! | |
263 | */ | |
264 | BUG_ON(dl_bandwidth_enabled()); | |
265 | ||
266 | /* | |
267 | * If admission control is disabled we | |
268 | * try a little harder to let the task | |
269 | * run. | |
270 | */ | |
271 | cpu = cpumask_any(cpu_active_mask); | |
272 | } | |
273 | later_rq = cpu_rq(cpu); | |
274 | double_lock_balance(rq, later_rq); | |
275 | } | |
276 | ||
a649f237 PZ |
277 | /* |
278 | * By now the task is replenished and enqueued; migrate it. | |
279 | */ | |
fa9c9d10 WL |
280 | deactivate_task(rq, p, 0); |
281 | set_task_cpu(p, later_rq->cpu); | |
a649f237 | 282 | activate_task(later_rq, p, 0); |
fa9c9d10 WL |
283 | |
284 | if (!fallback) | |
285 | resched_curr(later_rq); | |
286 | ||
a649f237 PZ |
287 | double_unlock_balance(later_rq, rq); |
288 | ||
289 | return later_rq; | |
fa9c9d10 WL |
290 | } |
291 | ||
1baca4ce JL |
292 | #else |
293 | ||
294 | static inline | |
295 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
296 | { | |
297 | } | |
298 | ||
299 | static inline | |
300 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
301 | { | |
302 | } | |
303 | ||
304 | static inline | |
305 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
306 | { | |
307 | } | |
308 | ||
309 | static inline | |
310 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
311 | { | |
312 | } | |
313 | ||
dc877341 PZ |
314 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
315 | { | |
316 | return false; | |
317 | } | |
318 | ||
0ea60c20 | 319 | static inline void pull_dl_task(struct rq *rq) |
dc877341 | 320 | { |
dc877341 PZ |
321 | } |
322 | ||
e3fca9e7 | 323 | static inline void queue_push_tasks(struct rq *rq) |
dc877341 | 324 | { |
dc877341 PZ |
325 | } |
326 | ||
9916e214 | 327 | static inline void queue_pull_task(struct rq *rq) |
dc877341 PZ |
328 | { |
329 | } | |
1baca4ce JL |
330 | #endif /* CONFIG_SMP */ |
331 | ||
aab03e05 DF |
332 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
333 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
334 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
335 | int flags); | |
336 | ||
337 | /* | |
338 | * We are being explicitly informed that a new instance is starting, | |
339 | * and this means that: | |
340 | * - the absolute deadline of the entity has to be placed at | |
341 | * current time + relative deadline; | |
342 | * - the runtime of the entity has to be set to the maximum value. | |
343 | * | |
344 | * The capability of specifying such event is useful whenever a -deadline | |
345 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
346 | * one, and to (try to!) reconcile itself with its own scheduling | |
347 | * parameters. | |
348 | */ | |
2d3d891d DF |
349 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, |
350 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
351 | { |
352 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
353 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
354 | ||
355 | WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); | |
356 | ||
357 | /* | |
358 | * We use the regular wall clock time to set deadlines in the | |
359 | * future; in fact, we must consider execution overheads (time | |
360 | * spent on hardirq context, etc.). | |
361 | */ | |
2d3d891d DF |
362 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
363 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
364 | dl_se->dl_new = 0; |
365 | } | |
366 | ||
367 | /* | |
368 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
369 | * possibility of a entity lasting more than what it declared, and thus | |
370 | * exhausting its runtime. | |
371 | * | |
372 | * Here we are interested in making runtime overrun possible, but we do | |
373 | * not want a entity which is misbehaving to affect the scheduling of all | |
374 | * other entities. | |
375 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
376 | * is used, in order to confine each entity within its own bandwidth. | |
377 | * | |
378 | * This function deals exactly with that, and ensures that when the runtime | |
379 | * of a entity is replenished, its deadline is also postponed. That ensures | |
380 | * the overrunning entity can't interfere with other entity in the system and | |
381 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
382 | * could happen are, typically, a entity voluntarily trying to overcome its | |
1b09d29b | 383 | * runtime, or it just underestimated it during sched_setattr(). |
aab03e05 | 384 | */ |
2d3d891d DF |
385 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
386 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
387 | { |
388 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
389 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
390 | ||
2d3d891d DF |
391 | BUG_ON(pi_se->dl_runtime <= 0); |
392 | ||
393 | /* | |
394 | * This could be the case for a !-dl task that is boosted. | |
395 | * Just go with full inherited parameters. | |
396 | */ | |
397 | if (dl_se->dl_deadline == 0) { | |
398 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
399 | dl_se->runtime = pi_se->dl_runtime; | |
400 | } | |
401 | ||
48be3a67 PZ |
402 | if (dl_se->dl_yielded && dl_se->runtime > 0) |
403 | dl_se->runtime = 0; | |
404 | ||
aab03e05 DF |
405 | /* |
406 | * We keep moving the deadline away until we get some | |
407 | * available runtime for the entity. This ensures correct | |
408 | * handling of situations where the runtime overrun is | |
409 | * arbitrary large. | |
410 | */ | |
411 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
412 | dl_se->deadline += pi_se->dl_period; |
413 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
414 | } |
415 | ||
416 | /* | |
417 | * At this point, the deadline really should be "in | |
418 | * the future" with respect to rq->clock. If it's | |
419 | * not, we are, for some reason, lagging too much! | |
420 | * Anyway, after having warn userspace abut that, | |
421 | * we still try to keep the things running by | |
422 | * resetting the deadline and the budget of the | |
423 | * entity. | |
424 | */ | |
425 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
c219b7dd | 426 | printk_deferred_once("sched: DL replenish lagged too much\n"); |
2d3d891d DF |
427 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
428 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 | 429 | } |
1019a359 PZ |
430 | |
431 | if (dl_se->dl_yielded) | |
432 | dl_se->dl_yielded = 0; | |
433 | if (dl_se->dl_throttled) | |
434 | dl_se->dl_throttled = 0; | |
aab03e05 DF |
435 | } |
436 | ||
437 | /* | |
438 | * Here we check if --at time t-- an entity (which is probably being | |
439 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
440 | * and its current deadline _without_ exceeding the bandwidth it is | |
441 | * assigned (function returns true if it can't). We are in fact applying | |
442 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
443 | * over residual deadline fits within the allocated bandwidth, then we | |
444 | * can keep the current (absolute) deadline and residual budget without | |
445 | * disrupting the schedulability of the system. Otherwise, we should | |
446 | * refill the runtime and set the deadline a period in the future, | |
447 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 DF |
448 | * result in breaking guarantees promised to other tasks (refer to |
449 | * Documentation/scheduler/sched-deadline.txt for more informations). | |
aab03e05 DF |
450 | * |
451 | * This function returns true if: | |
452 | * | |
755378a4 | 453 | * runtime / (deadline - t) > dl_runtime / dl_period , |
aab03e05 DF |
454 | * |
455 | * IOW we can't recycle current parameters. | |
755378a4 HG |
456 | * |
457 | * Notice that the bandwidth check is done against the period. For | |
458 | * task with deadline equal to period this is the same of using | |
459 | * dl_deadline instead of dl_period in the equation above. | |
aab03e05 | 460 | */ |
2d3d891d DF |
461 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
462 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
463 | { |
464 | u64 left, right; | |
465 | ||
466 | /* | |
467 | * left and right are the two sides of the equation above, | |
468 | * after a bit of shuffling to use multiplications instead | |
469 | * of divisions. | |
470 | * | |
471 | * Note that none of the time values involved in the two | |
472 | * multiplications are absolute: dl_deadline and dl_runtime | |
473 | * are the relative deadline and the maximum runtime of each | |
474 | * instance, runtime is the runtime left for the last instance | |
475 | * and (deadline - t), since t is rq->clock, is the time left | |
476 | * to the (absolute) deadline. Even if overflowing the u64 type | |
477 | * is very unlikely to occur in both cases, here we scale down | |
478 | * as we want to avoid that risk at all. Scaling down by 10 | |
479 | * means that we reduce granularity to 1us. We are fine with it, | |
480 | * since this is only a true/false check and, anyway, thinking | |
481 | * of anything below microseconds resolution is actually fiction | |
482 | * (but still we want to give the user that illusion >;). | |
483 | */ | |
332ac17e DF |
484 | left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
485 | right = ((dl_se->deadline - t) >> DL_SCALE) * | |
486 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
487 | |
488 | return dl_time_before(right, left); | |
489 | } | |
490 | ||
491 | /* | |
492 | * When a -deadline entity is queued back on the runqueue, its runtime and | |
493 | * deadline might need updating. | |
494 | * | |
495 | * The policy here is that we update the deadline of the entity only if: | |
496 | * - the current deadline is in the past, | |
497 | * - using the remaining runtime with the current deadline would make | |
498 | * the entity exceed its bandwidth. | |
499 | */ | |
2d3d891d DF |
500 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
501 | struct sched_dl_entity *pi_se) | |
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 | ||
506 | /* | |
507 | * The arrival of a new instance needs special treatment, i.e., | |
508 | * the actual scheduling parameters have to be "renewed". | |
509 | */ | |
510 | if (dl_se->dl_new) { | |
2d3d891d | 511 | setup_new_dl_entity(dl_se, pi_se); |
aab03e05 DF |
512 | return; |
513 | } | |
514 | ||
515 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || | |
2d3d891d DF |
516 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
517 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
518 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
519 | } |
520 | } | |
521 | ||
522 | /* | |
523 | * If the entity depleted all its runtime, and if we want it to sleep | |
524 | * while waiting for some new execution time to become available, we | |
525 | * set the bandwidth enforcement timer to the replenishment instant | |
526 | * and try to activate it. | |
527 | * | |
528 | * Notice that it is important for the caller to know if the timer | |
529 | * actually started or not (i.e., the replenishment instant is in | |
530 | * the future or in the past). | |
531 | */ | |
a649f237 | 532 | static int start_dl_timer(struct task_struct *p) |
aab03e05 | 533 | { |
a649f237 PZ |
534 | struct sched_dl_entity *dl_se = &p->dl; |
535 | struct hrtimer *timer = &dl_se->dl_timer; | |
536 | struct rq *rq = task_rq(p); | |
aab03e05 | 537 | ktime_t now, act; |
aab03e05 DF |
538 | s64 delta; |
539 | ||
a649f237 PZ |
540 | lockdep_assert_held(&rq->lock); |
541 | ||
aab03e05 DF |
542 | /* |
543 | * We want the timer to fire at the deadline, but considering | |
544 | * that it is actually coming from rq->clock and not from | |
545 | * hrtimer's time base reading. | |
546 | */ | |
547 | act = ns_to_ktime(dl_se->deadline); | |
a649f237 | 548 | now = hrtimer_cb_get_time(timer); |
aab03e05 DF |
549 | delta = ktime_to_ns(now) - rq_clock(rq); |
550 | act = ktime_add_ns(act, delta); | |
551 | ||
552 | /* | |
553 | * If the expiry time already passed, e.g., because the value | |
554 | * chosen as the deadline is too small, don't even try to | |
555 | * start the timer in the past! | |
556 | */ | |
557 | if (ktime_us_delta(act, now) < 0) | |
558 | return 0; | |
559 | ||
a649f237 PZ |
560 | /* |
561 | * !enqueued will guarantee another callback; even if one is already in | |
562 | * progress. This ensures a balanced {get,put}_task_struct(). | |
563 | * | |
564 | * The race against __run_timer() clearing the enqueued state is | |
565 | * harmless because we're holding task_rq()->lock, therefore the timer | |
566 | * expiring after we've done the check will wait on its task_rq_lock() | |
567 | * and observe our state. | |
568 | */ | |
569 | if (!hrtimer_is_queued(timer)) { | |
570 | get_task_struct(p); | |
571 | hrtimer_start(timer, act, HRTIMER_MODE_ABS); | |
572 | } | |
aab03e05 | 573 | |
cc9684d3 | 574 | return 1; |
aab03e05 DF |
575 | } |
576 | ||
577 | /* | |
578 | * This is the bandwidth enforcement timer callback. If here, we know | |
579 | * a task is not on its dl_rq, since the fact that the timer was running | |
580 | * means the task is throttled and needs a runtime replenishment. | |
581 | * | |
582 | * However, what we actually do depends on the fact the task is active, | |
583 | * (it is on its rq) or has been removed from there by a call to | |
584 | * dequeue_task_dl(). In the former case we must issue the runtime | |
585 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
586 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
587 | * updating (and the queueing back to dl_rq) will be done by the | |
588 | * next call to enqueue_task_dl(). | |
589 | */ | |
590 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
591 | { | |
592 | struct sched_dl_entity *dl_se = container_of(timer, | |
593 | struct sched_dl_entity, | |
594 | dl_timer); | |
595 | struct task_struct *p = dl_task_of(dl_se); | |
3960c8c0 | 596 | unsigned long flags; |
0f397f2c | 597 | struct rq *rq; |
3960c8c0 | 598 | |
4cd57f97 | 599 | rq = task_rq_lock(p, &flags); |
0f397f2c | 600 | |
aab03e05 | 601 | /* |
a649f237 PZ |
602 | * The task might have changed its scheduling policy to something |
603 | * different than SCHED_DEADLINE (through switched_fromd_dl()). | |
604 | */ | |
605 | if (!dl_task(p)) { | |
606 | __dl_clear_params(p); | |
607 | goto unlock; | |
608 | } | |
609 | ||
610 | /* | |
611 | * This is possible if switched_from_dl() raced against a running | |
612 | * callback that took the above !dl_task() path and we've since then | |
613 | * switched back into SCHED_DEADLINE. | |
aee38ea9 | 614 | * |
a649f237 | 615 | * There's nothing to do except drop our task reference. |
aab03e05 | 616 | */ |
a649f237 | 617 | if (dl_se->dl_new) |
aab03e05 DF |
618 | goto unlock; |
619 | ||
a649f237 PZ |
620 | /* |
621 | * The task might have been boosted by someone else and might be in the | |
622 | * boosting/deboosting path, its not throttled. | |
623 | */ | |
624 | if (dl_se->dl_boosted) | |
625 | goto unlock; | |
a79ec89f | 626 | |
fa9c9d10 | 627 | /* |
a649f237 PZ |
628 | * Spurious timer due to start_dl_timer() race; or we already received |
629 | * a replenishment from rt_mutex_setprio(). | |
fa9c9d10 | 630 | */ |
a649f237 | 631 | if (!dl_se->dl_throttled) |
fa9c9d10 | 632 | goto unlock; |
a649f237 PZ |
633 | |
634 | sched_clock_tick(); | |
635 | update_rq_clock(rq); | |
fa9c9d10 | 636 | |
a79ec89f KT |
637 | /* |
638 | * If the throttle happened during sched-out; like: | |
639 | * | |
640 | * schedule() | |
641 | * deactivate_task() | |
642 | * dequeue_task_dl() | |
643 | * update_curr_dl() | |
644 | * start_dl_timer() | |
645 | * __dequeue_task_dl() | |
646 | * prev->on_rq = 0; | |
647 | * | |
648 | * We can be both throttled and !queued. Replenish the counter | |
649 | * but do not enqueue -- wait for our wakeup to do that. | |
650 | */ | |
651 | if (!task_on_rq_queued(p)) { | |
652 | replenish_dl_entity(dl_se, dl_se); | |
653 | goto unlock; | |
654 | } | |
655 | ||
1019a359 PZ |
656 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
657 | if (dl_task(rq->curr)) | |
658 | check_preempt_curr_dl(rq, p, 0); | |
659 | else | |
660 | resched_curr(rq); | |
a649f237 | 661 | |
1baca4ce | 662 | #ifdef CONFIG_SMP |
1019a359 | 663 | /* |
a649f237 PZ |
664 | * Perform balancing operations here; after the replenishments. We |
665 | * cannot drop rq->lock before this, otherwise the assertion in | |
666 | * start_dl_timer() about not missing updates is not true. | |
667 | * | |
668 | * If we find that the rq the task was on is no longer available, we | |
669 | * need to select a new rq. | |
670 | * | |
671 | * XXX figure out if select_task_rq_dl() deals with offline cpus. | |
672 | */ | |
673 | if (unlikely(!rq->online)) | |
674 | rq = dl_task_offline_migration(rq, p); | |
675 | ||
676 | /* | |
677 | * Queueing this task back might have overloaded rq, check if we need | |
678 | * to kick someone away. | |
1019a359 | 679 | */ |
0aaafaab PZ |
680 | if (has_pushable_dl_tasks(rq)) { |
681 | /* | |
682 | * Nothing relies on rq->lock after this, so its safe to drop | |
683 | * rq->lock. | |
684 | */ | |
685 | lockdep_unpin_lock(&rq->lock); | |
1019a359 | 686 | push_dl_task(rq); |
0aaafaab PZ |
687 | lockdep_pin_lock(&rq->lock); |
688 | } | |
1baca4ce | 689 | #endif |
a649f237 | 690 | |
aab03e05 | 691 | unlock: |
4cd57f97 | 692 | task_rq_unlock(rq, p, &flags); |
aab03e05 | 693 | |
a649f237 PZ |
694 | /* |
695 | * This can free the task_struct, including this hrtimer, do not touch | |
696 | * anything related to that after this. | |
697 | */ | |
698 | put_task_struct(p); | |
699 | ||
aab03e05 DF |
700 | return HRTIMER_NORESTART; |
701 | } | |
702 | ||
703 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
704 | { | |
705 | struct hrtimer *timer = &dl_se->dl_timer; | |
706 | ||
aab03e05 DF |
707 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
708 | timer->function = dl_task_timer; | |
709 | } | |
710 | ||
711 | static | |
6fab5410 | 712 | int dl_runtime_exceeded(struct sched_dl_entity *dl_se) |
aab03e05 | 713 | { |
269ad801 | 714 | return (dl_se->runtime <= 0); |
aab03e05 DF |
715 | } |
716 | ||
faa59937 JL |
717 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
718 | ||
aab03e05 DF |
719 | /* |
720 | * Update the current task's runtime statistics (provided it is still | |
721 | * a -deadline task and has not been removed from the dl_rq). | |
722 | */ | |
723 | static void update_curr_dl(struct rq *rq) | |
724 | { | |
725 | struct task_struct *curr = rq->curr; | |
726 | struct sched_dl_entity *dl_se = &curr->dl; | |
727 | u64 delta_exec; | |
728 | ||
729 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
730 | return; | |
731 | ||
732 | /* | |
733 | * Consumed budget is computed considering the time as | |
734 | * observed by schedulable tasks (excluding time spent | |
735 | * in hardirq context, etc.). Deadlines are instead | |
736 | * computed using hard walltime. This seems to be the more | |
737 | * natural solution, but the full ramifications of this | |
738 | * approach need further study. | |
739 | */ | |
740 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; | |
48be3a67 PZ |
741 | if (unlikely((s64)delta_exec <= 0)) { |
742 | if (unlikely(dl_se->dl_yielded)) | |
743 | goto throttle; | |
734ff2a7 | 744 | return; |
48be3a67 | 745 | } |
aab03e05 DF |
746 | |
747 | schedstat_set(curr->se.statistics.exec_max, | |
748 | max(curr->se.statistics.exec_max, delta_exec)); | |
749 | ||
750 | curr->se.sum_exec_runtime += delta_exec; | |
751 | account_group_exec_runtime(curr, delta_exec); | |
752 | ||
753 | curr->se.exec_start = rq_clock_task(rq); | |
754 | cpuacct_charge(curr, delta_exec); | |
755 | ||
239be4a9 DF |
756 | sched_rt_avg_update(rq, delta_exec); |
757 | ||
48be3a67 PZ |
758 | dl_se->runtime -= delta_exec; |
759 | ||
760 | throttle: | |
761 | if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { | |
1019a359 | 762 | dl_se->dl_throttled = 1; |
aab03e05 | 763 | __dequeue_task_dl(rq, curr, 0); |
a649f237 | 764 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr))) |
aab03e05 DF |
765 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
766 | ||
767 | if (!is_leftmost(curr, &rq->dl)) | |
8875125e | 768 | resched_curr(rq); |
aab03e05 | 769 | } |
1724813d PZ |
770 | |
771 | /* | |
772 | * Because -- for now -- we share the rt bandwidth, we need to | |
773 | * account our runtime there too, otherwise actual rt tasks | |
774 | * would be able to exceed the shared quota. | |
775 | * | |
776 | * Account to the root rt group for now. | |
777 | * | |
778 | * The solution we're working towards is having the RT groups scheduled | |
779 | * using deadline servers -- however there's a few nasties to figure | |
780 | * out before that can happen. | |
781 | */ | |
782 | if (rt_bandwidth_enabled()) { | |
783 | struct rt_rq *rt_rq = &rq->rt; | |
784 | ||
785 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
1724813d PZ |
786 | /* |
787 | * We'll let actual RT tasks worry about the overflow here, we | |
faa59937 JL |
788 | * have our own CBS to keep us inline; only account when RT |
789 | * bandwidth is relevant. | |
1724813d | 790 | */ |
faa59937 JL |
791 | if (sched_rt_bandwidth_account(rt_rq)) |
792 | rt_rq->rt_time += delta_exec; | |
1724813d PZ |
793 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
794 | } | |
aab03e05 DF |
795 | } |
796 | ||
1baca4ce JL |
797 | #ifdef CONFIG_SMP |
798 | ||
1baca4ce JL |
799 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
800 | { | |
801 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
802 | ||
803 | if (dl_rq->earliest_dl.curr == 0 || | |
804 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
1baca4ce | 805 | dl_rq->earliest_dl.curr = deadline; |
6bfd6d72 | 806 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); |
1baca4ce JL |
807 | } |
808 | } | |
809 | ||
810 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
811 | { | |
812 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
813 | ||
814 | /* | |
815 | * Since we may have removed our earliest (and/or next earliest) | |
816 | * task we must recompute them. | |
817 | */ | |
818 | if (!dl_rq->dl_nr_running) { | |
819 | dl_rq->earliest_dl.curr = 0; | |
820 | dl_rq->earliest_dl.next = 0; | |
6bfd6d72 | 821 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); |
1baca4ce JL |
822 | } else { |
823 | struct rb_node *leftmost = dl_rq->rb_leftmost; | |
824 | struct sched_dl_entity *entry; | |
825 | ||
826 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
827 | dl_rq->earliest_dl.curr = entry->deadline; | |
6bfd6d72 | 828 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); |
1baca4ce JL |
829 | } |
830 | } | |
831 | ||
832 | #else | |
833 | ||
834 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
835 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
836 | ||
837 | #endif /* CONFIG_SMP */ | |
838 | ||
839 | static inline | |
840 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
841 | { | |
842 | int prio = dl_task_of(dl_se)->prio; | |
843 | u64 deadline = dl_se->deadline; | |
844 | ||
845 | WARN_ON(!dl_prio(prio)); | |
846 | dl_rq->dl_nr_running++; | |
72465447 | 847 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
848 | |
849 | inc_dl_deadline(dl_rq, deadline); | |
850 | inc_dl_migration(dl_se, dl_rq); | |
851 | } | |
852 | ||
853 | static inline | |
854 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
855 | { | |
856 | int prio = dl_task_of(dl_se)->prio; | |
857 | ||
858 | WARN_ON(!dl_prio(prio)); | |
859 | WARN_ON(!dl_rq->dl_nr_running); | |
860 | dl_rq->dl_nr_running--; | |
72465447 | 861 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
862 | |
863 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
864 | dec_dl_migration(dl_se, dl_rq); | |
865 | } | |
866 | ||
aab03e05 DF |
867 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
868 | { | |
869 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
870 | struct rb_node **link = &dl_rq->rb_root.rb_node; | |
871 | struct rb_node *parent = NULL; | |
872 | struct sched_dl_entity *entry; | |
873 | int leftmost = 1; | |
874 | ||
875 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
876 | ||
877 | while (*link) { | |
878 | parent = *link; | |
879 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
880 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
881 | link = &parent->rb_left; | |
882 | else { | |
883 | link = &parent->rb_right; | |
884 | leftmost = 0; | |
885 | } | |
886 | } | |
887 | ||
888 | if (leftmost) | |
889 | dl_rq->rb_leftmost = &dl_se->rb_node; | |
890 | ||
891 | rb_link_node(&dl_se->rb_node, parent, link); | |
892 | rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); | |
893 | ||
1baca4ce | 894 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
895 | } |
896 | ||
897 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
898 | { | |
899 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
900 | ||
901 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
902 | return; | |
903 | ||
904 | if (dl_rq->rb_leftmost == &dl_se->rb_node) { | |
905 | struct rb_node *next_node; | |
906 | ||
907 | next_node = rb_next(&dl_se->rb_node); | |
908 | dl_rq->rb_leftmost = next_node; | |
909 | } | |
910 | ||
911 | rb_erase(&dl_se->rb_node, &dl_rq->rb_root); | |
912 | RB_CLEAR_NODE(&dl_se->rb_node); | |
913 | ||
1baca4ce | 914 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
915 | } |
916 | ||
917 | static void | |
2d3d891d DF |
918 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
919 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
920 | { |
921 | BUG_ON(on_dl_rq(dl_se)); | |
922 | ||
923 | /* | |
924 | * If this is a wakeup or a new instance, the scheduling | |
925 | * parameters of the task might need updating. Otherwise, | |
926 | * we want a replenishment of its runtime. | |
927 | */ | |
6a503c3b | 928 | if (dl_se->dl_new || flags & ENQUEUE_WAKEUP) |
2d3d891d | 929 | update_dl_entity(dl_se, pi_se); |
6a503c3b LA |
930 | else if (flags & ENQUEUE_REPLENISH) |
931 | replenish_dl_entity(dl_se, pi_se); | |
aab03e05 DF |
932 | |
933 | __enqueue_dl_entity(dl_se); | |
934 | } | |
935 | ||
936 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
937 | { | |
938 | __dequeue_dl_entity(dl_se); | |
939 | } | |
940 | ||
941 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
942 | { | |
2d3d891d DF |
943 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
944 | struct sched_dl_entity *pi_se = &p->dl; | |
945 | ||
946 | /* | |
947 | * Use the scheduling parameters of the top pi-waiter | |
ff277d42 | 948 | * task if we have one and its (absolute) deadline is |
2d3d891d DF |
949 | * smaller than our one... OTW we keep our runtime and |
950 | * deadline. | |
951 | */ | |
64be6f1f | 952 | if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) { |
2d3d891d | 953 | pi_se = &pi_task->dl; |
64be6f1f JL |
954 | } else if (!dl_prio(p->normal_prio)) { |
955 | /* | |
956 | * Special case in which we have a !SCHED_DEADLINE task | |
957 | * that is going to be deboosted, but exceedes its | |
958 | * runtime while doing so. No point in replenishing | |
959 | * it, as it's going to return back to its original | |
960 | * scheduling class after this. | |
961 | */ | |
962 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | |
963 | return; | |
964 | } | |
2d3d891d | 965 | |
aab03e05 DF |
966 | /* |
967 | * If p is throttled, we do nothing. In fact, if it exhausted | |
968 | * its budget it needs a replenishment and, since it now is on | |
969 | * its rq, the bandwidth timer callback (which clearly has not | |
970 | * run yet) will take care of this. | |
971 | */ | |
1019a359 | 972 | if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) |
aab03e05 DF |
973 | return; |
974 | ||
2d3d891d | 975 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce JL |
976 | |
977 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) | |
978 | enqueue_pushable_dl_task(rq, p); | |
aab03e05 DF |
979 | } |
980 | ||
981 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
982 | { | |
983 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 984 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
985 | } |
986 | ||
987 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
988 | { | |
989 | update_curr_dl(rq); | |
990 | __dequeue_task_dl(rq, p, flags); | |
aab03e05 DF |
991 | } |
992 | ||
993 | /* | |
994 | * Yield task semantic for -deadline tasks is: | |
995 | * | |
996 | * get off from the CPU until our next instance, with | |
997 | * a new runtime. This is of little use now, since we | |
998 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
999 | * bandwidth reclaiming is planned for the future, and | |
1000 | * yield_task_dl will indicate that some spare budget | |
1001 | * is available for other task instances to use it. | |
1002 | */ | |
1003 | static void yield_task_dl(struct rq *rq) | |
1004 | { | |
aab03e05 DF |
1005 | /* |
1006 | * We make the task go to sleep until its current deadline by | |
1007 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
1008 | * it and the bandwidth timer will wake it up and will give it | |
5bfd126e | 1009 | * new scheduling parameters (thanks to dl_yielded=1). |
aab03e05 | 1010 | */ |
48be3a67 PZ |
1011 | rq->curr->dl.dl_yielded = 1; |
1012 | ||
6f1607f1 | 1013 | update_rq_clock(rq); |
aab03e05 | 1014 | update_curr_dl(rq); |
44fb085b WL |
1015 | /* |
1016 | * Tell update_rq_clock() that we've just updated, | |
1017 | * so we don't do microscopic update in schedule() | |
1018 | * and double the fastpath cost. | |
1019 | */ | |
1020 | rq_clock_skip_update(rq, true); | |
aab03e05 DF |
1021 | } |
1022 | ||
1baca4ce JL |
1023 | #ifdef CONFIG_SMP |
1024 | ||
1025 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
1026 | |
1027 | static int | |
1028 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
1029 | { | |
1030 | struct task_struct *curr; | |
1031 | struct rq *rq; | |
1032 | ||
1d7e974c | 1033 | if (sd_flag != SD_BALANCE_WAKE) |
1baca4ce JL |
1034 | goto out; |
1035 | ||
1036 | rq = cpu_rq(cpu); | |
1037 | ||
1038 | rcu_read_lock(); | |
316c1608 | 1039 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
1baca4ce JL |
1040 | |
1041 | /* | |
1042 | * If we are dealing with a -deadline task, we must | |
1043 | * decide where to wake it up. | |
1044 | * If it has a later deadline and the current task | |
1045 | * on this rq can't move (provided the waking task | |
1046 | * can!) we prefer to send it somewhere else. On the | |
1047 | * other hand, if it has a shorter deadline, we | |
1048 | * try to make it stay here, it might be important. | |
1049 | */ | |
1050 | if (unlikely(dl_task(curr)) && | |
1051 | (curr->nr_cpus_allowed < 2 || | |
1052 | !dl_entity_preempt(&p->dl, &curr->dl)) && | |
1053 | (p->nr_cpus_allowed > 1)) { | |
1054 | int target = find_later_rq(p); | |
1055 | ||
9d514262 | 1056 | if (target != -1 && |
5aa50507 LA |
1057 | (dl_time_before(p->dl.deadline, |
1058 | cpu_rq(target)->dl.earliest_dl.curr) || | |
1059 | (cpu_rq(target)->dl.dl_nr_running == 0))) | |
1baca4ce JL |
1060 | cpu = target; |
1061 | } | |
1062 | rcu_read_unlock(); | |
1063 | ||
1064 | out: | |
1065 | return cpu; | |
1066 | } | |
1067 | ||
1068 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) | |
1069 | { | |
1070 | /* | |
1071 | * Current can't be migrated, useless to reschedule, | |
1072 | * let's hope p can move out. | |
1073 | */ | |
1074 | if (rq->curr->nr_cpus_allowed == 1 || | |
6bfd6d72 | 1075 | cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) |
1baca4ce JL |
1076 | return; |
1077 | ||
1078 | /* | |
1079 | * p is migratable, so let's not schedule it and | |
1080 | * see if it is pushed or pulled somewhere else. | |
1081 | */ | |
1082 | if (p->nr_cpus_allowed != 1 && | |
6bfd6d72 | 1083 | cpudl_find(&rq->rd->cpudl, p, NULL) != -1) |
1baca4ce JL |
1084 | return; |
1085 | ||
8875125e | 1086 | resched_curr(rq); |
1baca4ce JL |
1087 | } |
1088 | ||
1089 | #endif /* CONFIG_SMP */ | |
1090 | ||
aab03e05 DF |
1091 | /* |
1092 | * Only called when both the current and waking task are -deadline | |
1093 | * tasks. | |
1094 | */ | |
1095 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
1096 | int flags) | |
1097 | { | |
1baca4ce | 1098 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
8875125e | 1099 | resched_curr(rq); |
1baca4ce JL |
1100 | return; |
1101 | } | |
1102 | ||
1103 | #ifdef CONFIG_SMP | |
1104 | /* | |
1105 | * In the unlikely case current and p have the same deadline | |
1106 | * let us try to decide what's the best thing to do... | |
1107 | */ | |
332ac17e DF |
1108 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1109 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
1110 | check_preempt_equal_dl(rq, p); |
1111 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
1112 | } |
1113 | ||
1114 | #ifdef CONFIG_SCHED_HRTICK | |
1115 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1116 | { | |
177ef2a6 | 1117 | hrtick_start(rq, p->dl.runtime); |
aab03e05 | 1118 | } |
36ce9881 WL |
1119 | #else /* !CONFIG_SCHED_HRTICK */ |
1120 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1121 | { | |
1122 | } | |
aab03e05 DF |
1123 | #endif |
1124 | ||
1125 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, | |
1126 | struct dl_rq *dl_rq) | |
1127 | { | |
1128 | struct rb_node *left = dl_rq->rb_leftmost; | |
1129 | ||
1130 | if (!left) | |
1131 | return NULL; | |
1132 | ||
1133 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
1134 | } | |
1135 | ||
606dba2e | 1136 | struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev) |
aab03e05 DF |
1137 | { |
1138 | struct sched_dl_entity *dl_se; | |
1139 | struct task_struct *p; | |
1140 | struct dl_rq *dl_rq; | |
1141 | ||
1142 | dl_rq = &rq->dl; | |
1143 | ||
a1d9a323 | 1144 | if (need_pull_dl_task(rq, prev)) { |
cbce1a68 PZ |
1145 | /* |
1146 | * This is OK, because current is on_cpu, which avoids it being | |
1147 | * picked for load-balance and preemption/IRQs are still | |
1148 | * disabled avoiding further scheduler activity on it and we're | |
1149 | * being very careful to re-start the picking loop. | |
1150 | */ | |
1151 | lockdep_unpin_lock(&rq->lock); | |
38033c37 | 1152 | pull_dl_task(rq); |
cbce1a68 | 1153 | lockdep_pin_lock(&rq->lock); |
a1d9a323 KT |
1154 | /* |
1155 | * pull_rt_task() can drop (and re-acquire) rq->lock; this | |
1156 | * means a stop task can slip in, in which case we need to | |
1157 | * re-start task selection. | |
1158 | */ | |
da0c1e65 | 1159 | if (rq->stop && task_on_rq_queued(rq->stop)) |
a1d9a323 KT |
1160 | return RETRY_TASK; |
1161 | } | |
1162 | ||
734ff2a7 KT |
1163 | /* |
1164 | * When prev is DL, we may throttle it in put_prev_task(). | |
1165 | * So, we update time before we check for dl_nr_running. | |
1166 | */ | |
1167 | if (prev->sched_class == &dl_sched_class) | |
1168 | update_curr_dl(rq); | |
38033c37 | 1169 | |
aab03e05 DF |
1170 | if (unlikely(!dl_rq->dl_nr_running)) |
1171 | return NULL; | |
1172 | ||
3f1d2a31 | 1173 | put_prev_task(rq, prev); |
606dba2e | 1174 | |
aab03e05 DF |
1175 | dl_se = pick_next_dl_entity(rq, dl_rq); |
1176 | BUG_ON(!dl_se); | |
1177 | ||
1178 | p = dl_task_of(dl_se); | |
1179 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1180 | |
1181 | /* Running task will never be pushed. */ | |
71362650 | 1182 | dequeue_pushable_dl_task(rq, p); |
1baca4ce | 1183 | |
aab03e05 DF |
1184 | if (hrtick_enabled(rq)) |
1185 | start_hrtick_dl(rq, p); | |
1baca4ce | 1186 | |
e3fca9e7 | 1187 | queue_push_tasks(rq); |
1baca4ce | 1188 | |
aab03e05 DF |
1189 | return p; |
1190 | } | |
1191 | ||
1192 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | |
1193 | { | |
1194 | update_curr_dl(rq); | |
1baca4ce JL |
1195 | |
1196 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) | |
1197 | enqueue_pushable_dl_task(rq, p); | |
aab03e05 DF |
1198 | } |
1199 | ||
1200 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) | |
1201 | { | |
1202 | update_curr_dl(rq); | |
1203 | ||
a7bebf48 WL |
1204 | /* |
1205 | * Even when we have runtime, update_curr_dl() might have resulted in us | |
1206 | * not being the leftmost task anymore. In that case NEED_RESCHED will | |
1207 | * be set and schedule() will start a new hrtick for the next task. | |
1208 | */ | |
1209 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && | |
1210 | is_leftmost(p, &rq->dl)) | |
aab03e05 | 1211 | start_hrtick_dl(rq, p); |
aab03e05 DF |
1212 | } |
1213 | ||
1214 | static void task_fork_dl(struct task_struct *p) | |
1215 | { | |
1216 | /* | |
1217 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1218 | * sched_fork() | |
1219 | */ | |
1220 | } | |
1221 | ||
1222 | static void task_dead_dl(struct task_struct *p) | |
1223 | { | |
332ac17e DF |
1224 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1225 | ||
1226 | /* | |
1227 | * Since we are TASK_DEAD we won't slip out of the domain! | |
1228 | */ | |
1229 | raw_spin_lock_irq(&dl_b->lock); | |
40767b0d | 1230 | /* XXX we should retain the bw until 0-lag */ |
332ac17e DF |
1231 | dl_b->total_bw -= p->dl.dl_bw; |
1232 | raw_spin_unlock_irq(&dl_b->lock); | |
aab03e05 DF |
1233 | } |
1234 | ||
1235 | static void set_curr_task_dl(struct rq *rq) | |
1236 | { | |
1237 | struct task_struct *p = rq->curr; | |
1238 | ||
1239 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1240 | |
1241 | /* You can't push away the running task */ | |
1242 | dequeue_pushable_dl_task(rq, p); | |
1243 | } | |
1244 | ||
1245 | #ifdef CONFIG_SMP | |
1246 | ||
1247 | /* Only try algorithms three times */ | |
1248 | #define DL_MAX_TRIES 3 | |
1249 | ||
1250 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1251 | { | |
1252 | if (!task_running(rq, p) && | |
1ba93d42 | 1253 | cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
1baca4ce | 1254 | return 1; |
1baca4ce JL |
1255 | return 0; |
1256 | } | |
1257 | ||
8b5e770e WL |
1258 | /* |
1259 | * Return the earliest pushable rq's task, which is suitable to be executed | |
1260 | * on the CPU, NULL otherwise: | |
1261 | */ | |
1262 | static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu) | |
1263 | { | |
1264 | struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost; | |
1265 | struct task_struct *p = NULL; | |
1266 | ||
1267 | if (!has_pushable_dl_tasks(rq)) | |
1268 | return NULL; | |
1269 | ||
1270 | next_node: | |
1271 | if (next_node) { | |
1272 | p = rb_entry(next_node, struct task_struct, pushable_dl_tasks); | |
1273 | ||
1274 | if (pick_dl_task(rq, p, cpu)) | |
1275 | return p; | |
1276 | ||
1277 | next_node = rb_next(next_node); | |
1278 | goto next_node; | |
1279 | } | |
1280 | ||
1281 | return NULL; | |
1282 | } | |
1283 | ||
1baca4ce JL |
1284 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1285 | ||
1286 | static int find_later_rq(struct task_struct *task) | |
1287 | { | |
1288 | struct sched_domain *sd; | |
4ba29684 | 1289 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1baca4ce JL |
1290 | int this_cpu = smp_processor_id(); |
1291 | int best_cpu, cpu = task_cpu(task); | |
1292 | ||
1293 | /* Make sure the mask is initialized first */ | |
1294 | if (unlikely(!later_mask)) | |
1295 | return -1; | |
1296 | ||
1297 | if (task->nr_cpus_allowed == 1) | |
1298 | return -1; | |
1299 | ||
91ec6778 JL |
1300 | /* |
1301 | * We have to consider system topology and task affinity | |
1302 | * first, then we can look for a suitable cpu. | |
1303 | */ | |
6bfd6d72 JL |
1304 | best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, |
1305 | task, later_mask); | |
1baca4ce JL |
1306 | if (best_cpu == -1) |
1307 | return -1; | |
1308 | ||
1309 | /* | |
1310 | * If we are here, some target has been found, | |
1311 | * the most suitable of which is cached in best_cpu. | |
1312 | * This is, among the runqueues where the current tasks | |
1313 | * have later deadlines than the task's one, the rq | |
1314 | * with the latest possible one. | |
1315 | * | |
1316 | * Now we check how well this matches with task's | |
1317 | * affinity and system topology. | |
1318 | * | |
1319 | * The last cpu where the task run is our first | |
1320 | * guess, since it is most likely cache-hot there. | |
1321 | */ | |
1322 | if (cpumask_test_cpu(cpu, later_mask)) | |
1323 | return cpu; | |
1324 | /* | |
1325 | * Check if this_cpu is to be skipped (i.e., it is | |
1326 | * not in the mask) or not. | |
1327 | */ | |
1328 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1329 | this_cpu = -1; | |
1330 | ||
1331 | rcu_read_lock(); | |
1332 | for_each_domain(cpu, sd) { | |
1333 | if (sd->flags & SD_WAKE_AFFINE) { | |
1334 | ||
1335 | /* | |
1336 | * If possible, preempting this_cpu is | |
1337 | * cheaper than migrating. | |
1338 | */ | |
1339 | if (this_cpu != -1 && | |
1340 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1341 | rcu_read_unlock(); | |
1342 | return this_cpu; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * Last chance: if best_cpu is valid and is | |
1347 | * in the mask, that becomes our choice. | |
1348 | */ | |
1349 | if (best_cpu < nr_cpu_ids && | |
1350 | cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { | |
1351 | rcu_read_unlock(); | |
1352 | return best_cpu; | |
1353 | } | |
1354 | } | |
1355 | } | |
1356 | rcu_read_unlock(); | |
1357 | ||
1358 | /* | |
1359 | * At this point, all our guesses failed, we just return | |
1360 | * 'something', and let the caller sort the things out. | |
1361 | */ | |
1362 | if (this_cpu != -1) | |
1363 | return this_cpu; | |
1364 | ||
1365 | cpu = cpumask_any(later_mask); | |
1366 | if (cpu < nr_cpu_ids) | |
1367 | return cpu; | |
1368 | ||
1369 | return -1; | |
1370 | } | |
1371 | ||
1372 | /* Locks the rq it finds */ | |
1373 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
1374 | { | |
1375 | struct rq *later_rq = NULL; | |
1376 | int tries; | |
1377 | int cpu; | |
1378 | ||
1379 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
1380 | cpu = find_later_rq(task); | |
1381 | ||
1382 | if ((cpu == -1) || (cpu == rq->cpu)) | |
1383 | break; | |
1384 | ||
1385 | later_rq = cpu_rq(cpu); | |
1386 | ||
5aa50507 LA |
1387 | if (later_rq->dl.dl_nr_running && |
1388 | !dl_time_before(task->dl.deadline, | |
9d514262 WL |
1389 | later_rq->dl.earliest_dl.curr)) { |
1390 | /* | |
1391 | * Target rq has tasks of equal or earlier deadline, | |
1392 | * retrying does not release any lock and is unlikely | |
1393 | * to yield a different result. | |
1394 | */ | |
1395 | later_rq = NULL; | |
1396 | break; | |
1397 | } | |
1398 | ||
1baca4ce JL |
1399 | /* Retry if something changed. */ |
1400 | if (double_lock_balance(rq, later_rq)) { | |
1401 | if (unlikely(task_rq(task) != rq || | |
1402 | !cpumask_test_cpu(later_rq->cpu, | |
1403 | &task->cpus_allowed) || | |
da0c1e65 KT |
1404 | task_running(rq, task) || |
1405 | !task_on_rq_queued(task))) { | |
1baca4ce JL |
1406 | double_unlock_balance(rq, later_rq); |
1407 | later_rq = NULL; | |
1408 | break; | |
1409 | } | |
1410 | } | |
1411 | ||
1412 | /* | |
1413 | * If the rq we found has no -deadline task, or | |
1414 | * its earliest one has a later deadline than our | |
1415 | * task, the rq is a good one. | |
1416 | */ | |
1417 | if (!later_rq->dl.dl_nr_running || | |
1418 | dl_time_before(task->dl.deadline, | |
1419 | later_rq->dl.earliest_dl.curr)) | |
1420 | break; | |
1421 | ||
1422 | /* Otherwise we try again. */ | |
1423 | double_unlock_balance(rq, later_rq); | |
1424 | later_rq = NULL; | |
1425 | } | |
1426 | ||
1427 | return later_rq; | |
1428 | } | |
1429 | ||
1430 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
1431 | { | |
1432 | struct task_struct *p; | |
1433 | ||
1434 | if (!has_pushable_dl_tasks(rq)) | |
1435 | return NULL; | |
1436 | ||
1437 | p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, | |
1438 | struct task_struct, pushable_dl_tasks); | |
1439 | ||
1440 | BUG_ON(rq->cpu != task_cpu(p)); | |
1441 | BUG_ON(task_current(rq, p)); | |
1442 | BUG_ON(p->nr_cpus_allowed <= 1); | |
1443 | ||
da0c1e65 | 1444 | BUG_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
1445 | BUG_ON(!dl_task(p)); |
1446 | ||
1447 | return p; | |
1448 | } | |
1449 | ||
1450 | /* | |
1451 | * See if the non running -deadline tasks on this rq | |
1452 | * can be sent to some other CPU where they can preempt | |
1453 | * and start executing. | |
1454 | */ | |
1455 | static int push_dl_task(struct rq *rq) | |
1456 | { | |
1457 | struct task_struct *next_task; | |
1458 | struct rq *later_rq; | |
c51b8ab5 | 1459 | int ret = 0; |
1baca4ce JL |
1460 | |
1461 | if (!rq->dl.overloaded) | |
1462 | return 0; | |
1463 | ||
1464 | next_task = pick_next_pushable_dl_task(rq); | |
1465 | if (!next_task) | |
1466 | return 0; | |
1467 | ||
1468 | retry: | |
1469 | if (unlikely(next_task == rq->curr)) { | |
1470 | WARN_ON(1); | |
1471 | return 0; | |
1472 | } | |
1473 | ||
1474 | /* | |
1475 | * If next_task preempts rq->curr, and rq->curr | |
1476 | * can move away, it makes sense to just reschedule | |
1477 | * without going further in pushing next_task. | |
1478 | */ | |
1479 | if (dl_task(rq->curr) && | |
1480 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
1481 | rq->curr->nr_cpus_allowed > 1) { | |
8875125e | 1482 | resched_curr(rq); |
1baca4ce JL |
1483 | return 0; |
1484 | } | |
1485 | ||
1486 | /* We might release rq lock */ | |
1487 | get_task_struct(next_task); | |
1488 | ||
1489 | /* Will lock the rq it'll find */ | |
1490 | later_rq = find_lock_later_rq(next_task, rq); | |
1491 | if (!later_rq) { | |
1492 | struct task_struct *task; | |
1493 | ||
1494 | /* | |
1495 | * We must check all this again, since | |
1496 | * find_lock_later_rq releases rq->lock and it is | |
1497 | * then possible that next_task has migrated. | |
1498 | */ | |
1499 | task = pick_next_pushable_dl_task(rq); | |
1500 | if (task_cpu(next_task) == rq->cpu && task == next_task) { | |
1501 | /* | |
1502 | * The task is still there. We don't try | |
1503 | * again, some other cpu will pull it when ready. | |
1504 | */ | |
1baca4ce JL |
1505 | goto out; |
1506 | } | |
1507 | ||
1508 | if (!task) | |
1509 | /* No more tasks */ | |
1510 | goto out; | |
1511 | ||
1512 | put_task_struct(next_task); | |
1513 | next_task = task; | |
1514 | goto retry; | |
1515 | } | |
1516 | ||
1517 | deactivate_task(rq, next_task, 0); | |
1518 | set_task_cpu(next_task, later_rq->cpu); | |
1519 | activate_task(later_rq, next_task, 0); | |
c51b8ab5 | 1520 | ret = 1; |
1baca4ce | 1521 | |
8875125e | 1522 | resched_curr(later_rq); |
1baca4ce JL |
1523 | |
1524 | double_unlock_balance(rq, later_rq); | |
1525 | ||
1526 | out: | |
1527 | put_task_struct(next_task); | |
1528 | ||
c51b8ab5 | 1529 | return ret; |
1baca4ce JL |
1530 | } |
1531 | ||
1532 | static void push_dl_tasks(struct rq *rq) | |
1533 | { | |
4ffa08ed | 1534 | /* push_dl_task() will return true if it moved a -deadline task */ |
1baca4ce JL |
1535 | while (push_dl_task(rq)) |
1536 | ; | |
aab03e05 DF |
1537 | } |
1538 | ||
0ea60c20 | 1539 | static void pull_dl_task(struct rq *this_rq) |
1baca4ce | 1540 | { |
0ea60c20 | 1541 | int this_cpu = this_rq->cpu, cpu; |
1baca4ce | 1542 | struct task_struct *p; |
0ea60c20 | 1543 | bool resched = false; |
1baca4ce JL |
1544 | struct rq *src_rq; |
1545 | u64 dmin = LONG_MAX; | |
1546 | ||
1547 | if (likely(!dl_overloaded(this_rq))) | |
0ea60c20 | 1548 | return; |
1baca4ce JL |
1549 | |
1550 | /* | |
1551 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
1552 | * see overloaded we must also see the dlo_mask bit. | |
1553 | */ | |
1554 | smp_rmb(); | |
1555 | ||
1556 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
1557 | if (this_cpu == cpu) | |
1558 | continue; | |
1559 | ||
1560 | src_rq = cpu_rq(cpu); | |
1561 | ||
1562 | /* | |
1563 | * It looks racy, abd it is! However, as in sched_rt.c, | |
1564 | * we are fine with this. | |
1565 | */ | |
1566 | if (this_rq->dl.dl_nr_running && | |
1567 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
1568 | src_rq->dl.earliest_dl.next)) | |
1569 | continue; | |
1570 | ||
1571 | /* Might drop this_rq->lock */ | |
1572 | double_lock_balance(this_rq, src_rq); | |
1573 | ||
1574 | /* | |
1575 | * If there are no more pullable tasks on the | |
1576 | * rq, we're done with it. | |
1577 | */ | |
1578 | if (src_rq->dl.dl_nr_running <= 1) | |
1579 | goto skip; | |
1580 | ||
8b5e770e | 1581 | p = pick_earliest_pushable_dl_task(src_rq, this_cpu); |
1baca4ce JL |
1582 | |
1583 | /* | |
1584 | * We found a task to be pulled if: | |
1585 | * - it preempts our current (if there's one), | |
1586 | * - it will preempt the last one we pulled (if any). | |
1587 | */ | |
1588 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
1589 | (!this_rq->dl.dl_nr_running || | |
1590 | dl_time_before(p->dl.deadline, | |
1591 | this_rq->dl.earliest_dl.curr))) { | |
1592 | WARN_ON(p == src_rq->curr); | |
da0c1e65 | 1593 | WARN_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
1594 | |
1595 | /* | |
1596 | * Then we pull iff p has actually an earlier | |
1597 | * deadline than the current task of its runqueue. | |
1598 | */ | |
1599 | if (dl_time_before(p->dl.deadline, | |
1600 | src_rq->curr->dl.deadline)) | |
1601 | goto skip; | |
1602 | ||
0ea60c20 | 1603 | resched = true; |
1baca4ce JL |
1604 | |
1605 | deactivate_task(src_rq, p, 0); | |
1606 | set_task_cpu(p, this_cpu); | |
1607 | activate_task(this_rq, p, 0); | |
1608 | dmin = p->dl.deadline; | |
1609 | ||
1610 | /* Is there any other task even earlier? */ | |
1611 | } | |
1612 | skip: | |
1613 | double_unlock_balance(this_rq, src_rq); | |
1614 | } | |
1615 | ||
0ea60c20 PZ |
1616 | if (resched) |
1617 | resched_curr(this_rq); | |
1baca4ce JL |
1618 | } |
1619 | ||
1620 | /* | |
1621 | * Since the task is not running and a reschedule is not going to happen | |
1622 | * anytime soon on its runqueue, we try pushing it away now. | |
1623 | */ | |
1624 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
1625 | { | |
1626 | if (!task_running(rq, p) && | |
1627 | !test_tsk_need_resched(rq->curr) && | |
1baca4ce JL |
1628 | p->nr_cpus_allowed > 1 && |
1629 | dl_task(rq->curr) && | |
1630 | (rq->curr->nr_cpus_allowed < 2 || | |
6b0a563f | 1631 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1baca4ce JL |
1632 | push_dl_tasks(rq); |
1633 | } | |
1634 | } | |
1635 | ||
1636 | static void set_cpus_allowed_dl(struct task_struct *p, | |
1637 | const struct cpumask *new_mask) | |
1638 | { | |
7f51412a | 1639 | struct root_domain *src_rd; |
6c37067e | 1640 | struct rq *rq; |
1baca4ce JL |
1641 | |
1642 | BUG_ON(!dl_task(p)); | |
1643 | ||
7f51412a JL |
1644 | rq = task_rq(p); |
1645 | src_rd = rq->rd; | |
1646 | /* | |
1647 | * Migrating a SCHED_DEADLINE task between exclusive | |
1648 | * cpusets (different root_domains) entails a bandwidth | |
1649 | * update. We already made space for us in the destination | |
1650 | * domain (see cpuset_can_attach()). | |
1651 | */ | |
1652 | if (!cpumask_intersects(src_rd->span, new_mask)) { | |
1653 | struct dl_bw *src_dl_b; | |
1654 | ||
1655 | src_dl_b = dl_bw_of(cpu_of(rq)); | |
1656 | /* | |
1657 | * We now free resources of the root_domain we are migrating | |
1658 | * off. In the worst case, sched_setattr() may temporary fail | |
1659 | * until we complete the update. | |
1660 | */ | |
1661 | raw_spin_lock(&src_dl_b->lock); | |
1662 | __dl_clear(src_dl_b, p->dl.dl_bw); | |
1663 | raw_spin_unlock(&src_dl_b->lock); | |
1664 | } | |
1665 | ||
6c37067e | 1666 | set_cpus_allowed_common(p, new_mask); |
1baca4ce JL |
1667 | } |
1668 | ||
1669 | /* Assumes rq->lock is held */ | |
1670 | static void rq_online_dl(struct rq *rq) | |
1671 | { | |
1672 | if (rq->dl.overloaded) | |
1673 | dl_set_overload(rq); | |
6bfd6d72 | 1674 | |
16b26943 | 1675 | cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); |
6bfd6d72 JL |
1676 | if (rq->dl.dl_nr_running > 0) |
1677 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); | |
1baca4ce JL |
1678 | } |
1679 | ||
1680 | /* Assumes rq->lock is held */ | |
1681 | static void rq_offline_dl(struct rq *rq) | |
1682 | { | |
1683 | if (rq->dl.overloaded) | |
1684 | dl_clear_overload(rq); | |
6bfd6d72 JL |
1685 | |
1686 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); | |
16b26943 | 1687 | cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); |
1baca4ce JL |
1688 | } |
1689 | ||
a6c0e746 | 1690 | void __init init_sched_dl_class(void) |
1baca4ce JL |
1691 | { |
1692 | unsigned int i; | |
1693 | ||
1694 | for_each_possible_cpu(i) | |
1695 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
1696 | GFP_KERNEL, cpu_to_node(i)); | |
1697 | } | |
1698 | ||
1699 | #endif /* CONFIG_SMP */ | |
1700 | ||
aab03e05 DF |
1701 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
1702 | { | |
a649f237 PZ |
1703 | /* |
1704 | * Start the deadline timer; if we switch back to dl before this we'll | |
1705 | * continue consuming our current CBS slice. If we stay outside of | |
1706 | * SCHED_DEADLINE until the deadline passes, the timer will reset the | |
1707 | * task. | |
1708 | */ | |
1709 | if (!start_dl_timer(p)) | |
1710 | __dl_clear_params(p); | |
a5e7be3b | 1711 | |
1baca4ce JL |
1712 | /* |
1713 | * Since this might be the only -deadline task on the rq, | |
1714 | * this is the right place to try to pull some other one | |
1715 | * from an overloaded cpu, if any. | |
1716 | */ | |
cd660911 WL |
1717 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
1718 | return; | |
1719 | ||
9916e214 | 1720 | queue_pull_task(rq); |
aab03e05 DF |
1721 | } |
1722 | ||
1baca4ce JL |
1723 | /* |
1724 | * When switching to -deadline, we may overload the rq, then | |
1725 | * we try to push someone off, if possible. | |
1726 | */ | |
aab03e05 DF |
1727 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
1728 | { | |
da0c1e65 | 1729 | if (task_on_rq_queued(p) && rq->curr != p) { |
1baca4ce | 1730 | #ifdef CONFIG_SMP |
9916e214 PZ |
1731 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded) |
1732 | queue_push_tasks(rq); | |
1733 | #else | |
1734 | if (dl_task(rq->curr)) | |
1735 | check_preempt_curr_dl(rq, p, 0); | |
1736 | else | |
1737 | resched_curr(rq); | |
1738 | #endif | |
aab03e05 DF |
1739 | } |
1740 | } | |
1741 | ||
1baca4ce JL |
1742 | /* |
1743 | * If the scheduling parameters of a -deadline task changed, | |
1744 | * a push or pull operation might be needed. | |
1745 | */ | |
aab03e05 DF |
1746 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
1747 | int oldprio) | |
1748 | { | |
da0c1e65 | 1749 | if (task_on_rq_queued(p) || rq->curr == p) { |
aab03e05 | 1750 | #ifdef CONFIG_SMP |
1baca4ce JL |
1751 | /* |
1752 | * This might be too much, but unfortunately | |
1753 | * we don't have the old deadline value, and | |
1754 | * we can't argue if the task is increasing | |
1755 | * or lowering its prio, so... | |
1756 | */ | |
1757 | if (!rq->dl.overloaded) | |
9916e214 | 1758 | queue_pull_task(rq); |
1baca4ce JL |
1759 | |
1760 | /* | |
1761 | * If we now have a earlier deadline task than p, | |
1762 | * then reschedule, provided p is still on this | |
1763 | * runqueue. | |
1764 | */ | |
9916e214 | 1765 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline)) |
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 | 1795 | .task_woken = task_woken_dl, |
aab03e05 DF |
1796 | #endif |
1797 | ||
1798 | .set_curr_task = set_curr_task_dl, | |
1799 | .task_tick = task_tick_dl, | |
1800 | .task_fork = task_fork_dl, | |
1801 | .task_dead = task_dead_dl, | |
1802 | ||
1803 | .prio_changed = prio_changed_dl, | |
1804 | .switched_from = switched_from_dl, | |
1805 | .switched_to = switched_to_dl, | |
6e998916 SG |
1806 | |
1807 | .update_curr = update_curr_dl, | |
aab03e05 | 1808 | }; |
acb32132 WL |
1809 | |
1810 | #ifdef CONFIG_SCHED_DEBUG | |
1811 | extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); | |
1812 | ||
1813 | void print_dl_stats(struct seq_file *m, int cpu) | |
1814 | { | |
1815 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | |
1816 | } | |
1817 | #endif /* CONFIG_SCHED_DEBUG */ |