]>
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 | ||
60 | extern unsigned long to_ratio(u64 period, u64 runtime); | |
61 | ||
62 | void init_dl_bw(struct dl_bw *dl_b) | |
63 | { | |
64 | raw_spin_lock_init(&dl_b->lock); | |
65 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | |
1724813d | 66 | if (global_rt_runtime() == RUNTIME_INF) |
332ac17e DF |
67 | dl_b->bw = -1; |
68 | else | |
1724813d | 69 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
332ac17e DF |
70 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
71 | dl_b->total_bw = 0; | |
72 | } | |
73 | ||
aab03e05 DF |
74 | void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq) |
75 | { | |
76 | dl_rq->rb_root = RB_ROOT; | |
1baca4ce JL |
77 | |
78 | #ifdef CONFIG_SMP | |
79 | /* zero means no -deadline tasks */ | |
80 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | |
81 | ||
82 | dl_rq->dl_nr_migratory = 0; | |
83 | dl_rq->overloaded = 0; | |
84 | dl_rq->pushable_dl_tasks_root = RB_ROOT; | |
332ac17e DF |
85 | #else |
86 | init_dl_bw(&dl_rq->dl_bw); | |
1baca4ce JL |
87 | #endif |
88 | } | |
89 | ||
90 | #ifdef CONFIG_SMP | |
91 | ||
92 | static inline int dl_overloaded(struct rq *rq) | |
93 | { | |
94 | return atomic_read(&rq->rd->dlo_count); | |
95 | } | |
96 | ||
97 | static inline void dl_set_overload(struct rq *rq) | |
98 | { | |
99 | if (!rq->online) | |
100 | return; | |
101 | ||
102 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | |
103 | /* | |
104 | * Must be visible before the overload count is | |
105 | * set (as in sched_rt.c). | |
106 | * | |
107 | * Matched by the barrier in pull_dl_task(). | |
108 | */ | |
109 | smp_wmb(); | |
110 | atomic_inc(&rq->rd->dlo_count); | |
111 | } | |
112 | ||
113 | static inline void dl_clear_overload(struct rq *rq) | |
114 | { | |
115 | if (!rq->online) | |
116 | return; | |
117 | ||
118 | atomic_dec(&rq->rd->dlo_count); | |
119 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | |
120 | } | |
121 | ||
122 | static void update_dl_migration(struct dl_rq *dl_rq) | |
123 | { | |
124 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_total > 1) { | |
125 | if (!dl_rq->overloaded) { | |
126 | dl_set_overload(rq_of_dl_rq(dl_rq)); | |
127 | dl_rq->overloaded = 1; | |
128 | } | |
129 | } else if (dl_rq->overloaded) { | |
130 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | |
131 | dl_rq->overloaded = 0; | |
132 | } | |
133 | } | |
134 | ||
135 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
136 | { | |
137 | struct task_struct *p = dl_task_of(dl_se); | |
138 | dl_rq = &rq_of_dl_rq(dl_rq)->dl; | |
139 | ||
140 | dl_rq->dl_nr_total++; | |
141 | if (p->nr_cpus_allowed > 1) | |
142 | dl_rq->dl_nr_migratory++; | |
143 | ||
144 | update_dl_migration(dl_rq); | |
145 | } | |
146 | ||
147 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
148 | { | |
149 | struct task_struct *p = dl_task_of(dl_se); | |
150 | dl_rq = &rq_of_dl_rq(dl_rq)->dl; | |
151 | ||
152 | dl_rq->dl_nr_total--; | |
153 | if (p->nr_cpus_allowed > 1) | |
154 | dl_rq->dl_nr_migratory--; | |
155 | ||
156 | update_dl_migration(dl_rq); | |
157 | } | |
158 | ||
159 | /* | |
160 | * The list of pushable -deadline task is not a plist, like in | |
161 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | |
162 | */ | |
163 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
164 | { | |
165 | struct dl_rq *dl_rq = &rq->dl; | |
166 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; | |
167 | struct rb_node *parent = NULL; | |
168 | struct task_struct *entry; | |
169 | int leftmost = 1; | |
170 | ||
171 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | |
172 | ||
173 | while (*link) { | |
174 | parent = *link; | |
175 | entry = rb_entry(parent, struct task_struct, | |
176 | pushable_dl_tasks); | |
177 | if (dl_entity_preempt(&p->dl, &entry->dl)) | |
178 | link = &parent->rb_left; | |
179 | else { | |
180 | link = &parent->rb_right; | |
181 | leftmost = 0; | |
182 | } | |
183 | } | |
184 | ||
185 | if (leftmost) | |
186 | dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; | |
187 | ||
188 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
189 | rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | |
aab03e05 DF |
190 | } |
191 | ||
1baca4ce JL |
192 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
193 | { | |
194 | struct dl_rq *dl_rq = &rq->dl; | |
195 | ||
196 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
197 | return; | |
198 | ||
199 | if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { | |
200 | struct rb_node *next_node; | |
201 | ||
202 | next_node = rb_next(&p->pushable_dl_tasks); | |
203 | dl_rq->pushable_dl_tasks_leftmost = next_node; | |
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 | ||
217 | #else | |
218 | ||
219 | static inline | |
220 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
221 | { | |
222 | } | |
223 | ||
224 | static inline | |
225 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
226 | { | |
227 | } | |
228 | ||
229 | static inline | |
230 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
231 | { | |
232 | } | |
233 | ||
234 | static inline | |
235 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
236 | { | |
237 | } | |
238 | ||
239 | #endif /* CONFIG_SMP */ | |
240 | ||
aab03e05 DF |
241 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
242 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
243 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
244 | int flags); | |
245 | ||
246 | /* | |
247 | * We are being explicitly informed that a new instance is starting, | |
248 | * and this means that: | |
249 | * - the absolute deadline of the entity has to be placed at | |
250 | * current time + relative deadline; | |
251 | * - the runtime of the entity has to be set to the maximum value. | |
252 | * | |
253 | * The capability of specifying such event is useful whenever a -deadline | |
254 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
255 | * one, and to (try to!) reconcile itself with its own scheduling | |
256 | * parameters. | |
257 | */ | |
2d3d891d DF |
258 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, |
259 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
260 | { |
261 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
262 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
263 | ||
264 | WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); | |
265 | ||
266 | /* | |
267 | * We use the regular wall clock time to set deadlines in the | |
268 | * future; in fact, we must consider execution overheads (time | |
269 | * spent on hardirq context, etc.). | |
270 | */ | |
2d3d891d DF |
271 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
272 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
273 | dl_se->dl_new = 0; |
274 | } | |
275 | ||
276 | /* | |
277 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
278 | * possibility of a entity lasting more than what it declared, and thus | |
279 | * exhausting its runtime. | |
280 | * | |
281 | * Here we are interested in making runtime overrun possible, but we do | |
282 | * not want a entity which is misbehaving to affect the scheduling of all | |
283 | * other entities. | |
284 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
285 | * is used, in order to confine each entity within its own bandwidth. | |
286 | * | |
287 | * This function deals exactly with that, and ensures that when the runtime | |
288 | * of a entity is replenished, its deadline is also postponed. That ensures | |
289 | * the overrunning entity can't interfere with other entity in the system and | |
290 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
291 | * could happen are, typically, a entity voluntarily trying to overcome its | |
292 | * runtime, or it just underestimated it during sched_setscheduler_ex(). | |
293 | */ | |
2d3d891d DF |
294 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
295 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
296 | { |
297 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
298 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
299 | ||
2d3d891d DF |
300 | BUG_ON(pi_se->dl_runtime <= 0); |
301 | ||
302 | /* | |
303 | * This could be the case for a !-dl task that is boosted. | |
304 | * Just go with full inherited parameters. | |
305 | */ | |
306 | if (dl_se->dl_deadline == 0) { | |
307 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
308 | dl_se->runtime = pi_se->dl_runtime; | |
309 | } | |
310 | ||
aab03e05 DF |
311 | /* |
312 | * We keep moving the deadline away until we get some | |
313 | * available runtime for the entity. This ensures correct | |
314 | * handling of situations where the runtime overrun is | |
315 | * arbitrary large. | |
316 | */ | |
317 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
318 | dl_se->deadline += pi_se->dl_period; |
319 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
320 | } |
321 | ||
322 | /* | |
323 | * At this point, the deadline really should be "in | |
324 | * the future" with respect to rq->clock. If it's | |
325 | * not, we are, for some reason, lagging too much! | |
326 | * Anyway, after having warn userspace abut that, | |
327 | * we still try to keep the things running by | |
328 | * resetting the deadline and the budget of the | |
329 | * entity. | |
330 | */ | |
331 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
332 | static bool lag_once = false; | |
333 | ||
334 | if (!lag_once) { | |
335 | lag_once = true; | |
336 | printk_sched("sched: DL replenish lagged to much\n"); | |
337 | } | |
2d3d891d DF |
338 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
339 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
340 | } |
341 | } | |
342 | ||
343 | /* | |
344 | * Here we check if --at time t-- an entity (which is probably being | |
345 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
346 | * and its current deadline _without_ exceeding the bandwidth it is | |
347 | * assigned (function returns true if it can't). We are in fact applying | |
348 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
349 | * over residual deadline fits within the allocated bandwidth, then we | |
350 | * can keep the current (absolute) deadline and residual budget without | |
351 | * disrupting the schedulability of the system. Otherwise, we should | |
352 | * refill the runtime and set the deadline a period in the future, | |
353 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 DF |
354 | * result in breaking guarantees promised to other tasks (refer to |
355 | * Documentation/scheduler/sched-deadline.txt for more informations). | |
aab03e05 DF |
356 | * |
357 | * This function returns true if: | |
358 | * | |
755378a4 | 359 | * runtime / (deadline - t) > dl_runtime / dl_period , |
aab03e05 DF |
360 | * |
361 | * IOW we can't recycle current parameters. | |
755378a4 HG |
362 | * |
363 | * Notice that the bandwidth check is done against the period. For | |
364 | * task with deadline equal to period this is the same of using | |
365 | * dl_deadline instead of dl_period in the equation above. | |
aab03e05 | 366 | */ |
2d3d891d DF |
367 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
368 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
369 | { |
370 | u64 left, right; | |
371 | ||
372 | /* | |
373 | * left and right are the two sides of the equation above, | |
374 | * after a bit of shuffling to use multiplications instead | |
375 | * of divisions. | |
376 | * | |
377 | * Note that none of the time values involved in the two | |
378 | * multiplications are absolute: dl_deadline and dl_runtime | |
379 | * are the relative deadline and the maximum runtime of each | |
380 | * instance, runtime is the runtime left for the last instance | |
381 | * and (deadline - t), since t is rq->clock, is the time left | |
382 | * to the (absolute) deadline. Even if overflowing the u64 type | |
383 | * is very unlikely to occur in both cases, here we scale down | |
384 | * as we want to avoid that risk at all. Scaling down by 10 | |
385 | * means that we reduce granularity to 1us. We are fine with it, | |
386 | * since this is only a true/false check and, anyway, thinking | |
387 | * of anything below microseconds resolution is actually fiction | |
388 | * (but still we want to give the user that illusion >;). | |
389 | */ | |
332ac17e DF |
390 | left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
391 | right = ((dl_se->deadline - t) >> DL_SCALE) * | |
392 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
393 | |
394 | return dl_time_before(right, left); | |
395 | } | |
396 | ||
397 | /* | |
398 | * When a -deadline entity is queued back on the runqueue, its runtime and | |
399 | * deadline might need updating. | |
400 | * | |
401 | * The policy here is that we update the deadline of the entity only if: | |
402 | * - the current deadline is in the past, | |
403 | * - using the remaining runtime with the current deadline would make | |
404 | * the entity exceed its bandwidth. | |
405 | */ | |
2d3d891d DF |
406 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
407 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
408 | { |
409 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
410 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
411 | ||
412 | /* | |
413 | * The arrival of a new instance needs special treatment, i.e., | |
414 | * the actual scheduling parameters have to be "renewed". | |
415 | */ | |
416 | if (dl_se->dl_new) { | |
2d3d891d | 417 | setup_new_dl_entity(dl_se, pi_se); |
aab03e05 DF |
418 | return; |
419 | } | |
420 | ||
421 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || | |
2d3d891d DF |
422 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
423 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
424 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
425 | } |
426 | } | |
427 | ||
428 | /* | |
429 | * If the entity depleted all its runtime, and if we want it to sleep | |
430 | * while waiting for some new execution time to become available, we | |
431 | * set the bandwidth enforcement timer to the replenishment instant | |
432 | * and try to activate it. | |
433 | * | |
434 | * Notice that it is important for the caller to know if the timer | |
435 | * actually started or not (i.e., the replenishment instant is in | |
436 | * the future or in the past). | |
437 | */ | |
2d3d891d | 438 | static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted) |
aab03e05 DF |
439 | { |
440 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
441 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
442 | ktime_t now, act; | |
443 | ktime_t soft, hard; | |
444 | unsigned long range; | |
445 | s64 delta; | |
446 | ||
2d3d891d DF |
447 | if (boosted) |
448 | return 0; | |
aab03e05 DF |
449 | /* |
450 | * We want the timer to fire at the deadline, but considering | |
451 | * that it is actually coming from rq->clock and not from | |
452 | * hrtimer's time base reading. | |
453 | */ | |
454 | act = ns_to_ktime(dl_se->deadline); | |
455 | now = hrtimer_cb_get_time(&dl_se->dl_timer); | |
456 | delta = ktime_to_ns(now) - rq_clock(rq); | |
457 | act = ktime_add_ns(act, delta); | |
458 | ||
459 | /* | |
460 | * If the expiry time already passed, e.g., because the value | |
461 | * chosen as the deadline is too small, don't even try to | |
462 | * start the timer in the past! | |
463 | */ | |
464 | if (ktime_us_delta(act, now) < 0) | |
465 | return 0; | |
466 | ||
467 | hrtimer_set_expires(&dl_se->dl_timer, act); | |
468 | ||
469 | soft = hrtimer_get_softexpires(&dl_se->dl_timer); | |
470 | hard = hrtimer_get_expires(&dl_se->dl_timer); | |
471 | range = ktime_to_ns(ktime_sub(hard, soft)); | |
472 | __hrtimer_start_range_ns(&dl_se->dl_timer, soft, | |
473 | range, HRTIMER_MODE_ABS, 0); | |
474 | ||
475 | return hrtimer_active(&dl_se->dl_timer); | |
476 | } | |
477 | ||
478 | /* | |
479 | * This is the bandwidth enforcement timer callback. If here, we know | |
480 | * a task is not on its dl_rq, since the fact that the timer was running | |
481 | * means the task is throttled and needs a runtime replenishment. | |
482 | * | |
483 | * However, what we actually do depends on the fact the task is active, | |
484 | * (it is on its rq) or has been removed from there by a call to | |
485 | * dequeue_task_dl(). In the former case we must issue the runtime | |
486 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
487 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
488 | * updating (and the queueing back to dl_rq) will be done by the | |
489 | * next call to enqueue_task_dl(). | |
490 | */ | |
491 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
492 | { | |
493 | struct sched_dl_entity *dl_se = container_of(timer, | |
494 | struct sched_dl_entity, | |
495 | dl_timer); | |
496 | struct task_struct *p = dl_task_of(dl_se); | |
497 | struct rq *rq = task_rq(p); | |
498 | raw_spin_lock(&rq->lock); | |
499 | ||
500 | /* | |
501 | * We need to take care of a possible races here. In fact, the | |
502 | * task might have changed its scheduling policy to something | |
503 | * different from SCHED_DEADLINE or changed its reservation | |
504 | * parameters (through sched_setscheduler()). | |
505 | */ | |
506 | if (!dl_task(p) || dl_se->dl_new) | |
507 | goto unlock; | |
508 | ||
509 | sched_clock_tick(); | |
510 | update_rq_clock(rq); | |
511 | dl_se->dl_throttled = 0; | |
512 | if (p->on_rq) { | |
513 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); | |
514 | if (task_has_dl_policy(rq->curr)) | |
515 | check_preempt_curr_dl(rq, p, 0); | |
516 | else | |
517 | resched_task(rq->curr); | |
1baca4ce JL |
518 | #ifdef CONFIG_SMP |
519 | /* | |
520 | * Queueing this task back might have overloaded rq, | |
521 | * check if we need to kick someone away. | |
522 | */ | |
523 | if (has_pushable_dl_tasks(rq)) | |
524 | push_dl_task(rq); | |
525 | #endif | |
aab03e05 DF |
526 | } |
527 | unlock: | |
528 | raw_spin_unlock(&rq->lock); | |
529 | ||
530 | return HRTIMER_NORESTART; | |
531 | } | |
532 | ||
533 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
534 | { | |
535 | struct hrtimer *timer = &dl_se->dl_timer; | |
536 | ||
537 | if (hrtimer_active(timer)) { | |
538 | hrtimer_try_to_cancel(timer); | |
539 | return; | |
540 | } | |
541 | ||
542 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
543 | timer->function = dl_task_timer; | |
544 | } | |
545 | ||
546 | static | |
547 | int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) | |
548 | { | |
549 | int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq)); | |
550 | int rorun = dl_se->runtime <= 0; | |
551 | ||
552 | if (!rorun && !dmiss) | |
553 | return 0; | |
554 | ||
555 | /* | |
556 | * If we are beyond our current deadline and we are still | |
557 | * executing, then we have already used some of the runtime of | |
558 | * the next instance. Thus, if we do not account that, we are | |
559 | * stealing bandwidth from the system at each deadline miss! | |
560 | */ | |
561 | if (dmiss) { | |
562 | dl_se->runtime = rorun ? dl_se->runtime : 0; | |
563 | dl_se->runtime -= rq_clock(rq) - dl_se->deadline; | |
564 | } | |
565 | ||
566 | return 1; | |
567 | } | |
568 | ||
569 | /* | |
570 | * Update the current task's runtime statistics (provided it is still | |
571 | * a -deadline task and has not been removed from the dl_rq). | |
572 | */ | |
573 | static void update_curr_dl(struct rq *rq) | |
574 | { | |
575 | struct task_struct *curr = rq->curr; | |
576 | struct sched_dl_entity *dl_se = &curr->dl; | |
577 | u64 delta_exec; | |
578 | ||
579 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
580 | return; | |
581 | ||
582 | /* | |
583 | * Consumed budget is computed considering the time as | |
584 | * observed by schedulable tasks (excluding time spent | |
585 | * in hardirq context, etc.). Deadlines are instead | |
586 | * computed using hard walltime. This seems to be the more | |
587 | * natural solution, but the full ramifications of this | |
588 | * approach need further study. | |
589 | */ | |
590 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; | |
591 | if (unlikely((s64)delta_exec < 0)) | |
592 | delta_exec = 0; | |
593 | ||
594 | schedstat_set(curr->se.statistics.exec_max, | |
595 | max(curr->se.statistics.exec_max, delta_exec)); | |
596 | ||
597 | curr->se.sum_exec_runtime += delta_exec; | |
598 | account_group_exec_runtime(curr, delta_exec); | |
599 | ||
600 | curr->se.exec_start = rq_clock_task(rq); | |
601 | cpuacct_charge(curr, delta_exec); | |
602 | ||
239be4a9 DF |
603 | sched_rt_avg_update(rq, delta_exec); |
604 | ||
aab03e05 DF |
605 | dl_se->runtime -= delta_exec; |
606 | if (dl_runtime_exceeded(rq, dl_se)) { | |
607 | __dequeue_task_dl(rq, curr, 0); | |
2d3d891d | 608 | if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted))) |
aab03e05 DF |
609 | dl_se->dl_throttled = 1; |
610 | else | |
611 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); | |
612 | ||
613 | if (!is_leftmost(curr, &rq->dl)) | |
614 | resched_task(curr); | |
615 | } | |
1724813d PZ |
616 | |
617 | /* | |
618 | * Because -- for now -- we share the rt bandwidth, we need to | |
619 | * account our runtime there too, otherwise actual rt tasks | |
620 | * would be able to exceed the shared quota. | |
621 | * | |
622 | * Account to the root rt group for now. | |
623 | * | |
624 | * The solution we're working towards is having the RT groups scheduled | |
625 | * using deadline servers -- however there's a few nasties to figure | |
626 | * out before that can happen. | |
627 | */ | |
628 | if (rt_bandwidth_enabled()) { | |
629 | struct rt_rq *rt_rq = &rq->rt; | |
630 | ||
631 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
632 | rt_rq->rt_time += delta_exec; | |
633 | /* | |
634 | * We'll let actual RT tasks worry about the overflow here, we | |
635 | * have our own CBS to keep us inline -- see above. | |
636 | */ | |
637 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
638 | } | |
aab03e05 DF |
639 | } |
640 | ||
1baca4ce JL |
641 | #ifdef CONFIG_SMP |
642 | ||
643 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu); | |
644 | ||
645 | static inline u64 next_deadline(struct rq *rq) | |
646 | { | |
647 | struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu); | |
648 | ||
649 | if (next && dl_prio(next->prio)) | |
650 | return next->dl.deadline; | |
651 | else | |
652 | return 0; | |
653 | } | |
654 | ||
655 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
656 | { | |
657 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
658 | ||
659 | if (dl_rq->earliest_dl.curr == 0 || | |
660 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
661 | /* | |
662 | * If the dl_rq had no -deadline tasks, or if the new task | |
663 | * has shorter deadline than the current one on dl_rq, we | |
664 | * know that the previous earliest becomes our next earliest, | |
665 | * as the new task becomes the earliest itself. | |
666 | */ | |
667 | dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr; | |
668 | dl_rq->earliest_dl.curr = deadline; | |
6bfd6d72 | 669 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); |
1baca4ce JL |
670 | } else if (dl_rq->earliest_dl.next == 0 || |
671 | dl_time_before(deadline, dl_rq->earliest_dl.next)) { | |
672 | /* | |
673 | * On the other hand, if the new -deadline task has a | |
674 | * a later deadline than the earliest one on dl_rq, but | |
675 | * it is earlier than the next (if any), we must | |
676 | * recompute the next-earliest. | |
677 | */ | |
678 | dl_rq->earliest_dl.next = next_deadline(rq); | |
679 | } | |
680 | } | |
681 | ||
682 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
683 | { | |
684 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
685 | ||
686 | /* | |
687 | * Since we may have removed our earliest (and/or next earliest) | |
688 | * task we must recompute them. | |
689 | */ | |
690 | if (!dl_rq->dl_nr_running) { | |
691 | dl_rq->earliest_dl.curr = 0; | |
692 | dl_rq->earliest_dl.next = 0; | |
6bfd6d72 | 693 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); |
1baca4ce JL |
694 | } else { |
695 | struct rb_node *leftmost = dl_rq->rb_leftmost; | |
696 | struct sched_dl_entity *entry; | |
697 | ||
698 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
699 | dl_rq->earliest_dl.curr = entry->deadline; | |
700 | dl_rq->earliest_dl.next = next_deadline(rq); | |
6bfd6d72 | 701 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); |
1baca4ce JL |
702 | } |
703 | } | |
704 | ||
705 | #else | |
706 | ||
707 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
708 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
709 | ||
710 | #endif /* CONFIG_SMP */ | |
711 | ||
712 | static inline | |
713 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
714 | { | |
715 | int prio = dl_task_of(dl_se)->prio; | |
716 | u64 deadline = dl_se->deadline; | |
717 | ||
718 | WARN_ON(!dl_prio(prio)); | |
719 | dl_rq->dl_nr_running++; | |
720 | ||
721 | inc_dl_deadline(dl_rq, deadline); | |
722 | inc_dl_migration(dl_se, dl_rq); | |
723 | } | |
724 | ||
725 | static inline | |
726 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
727 | { | |
728 | int prio = dl_task_of(dl_se)->prio; | |
729 | ||
730 | WARN_ON(!dl_prio(prio)); | |
731 | WARN_ON(!dl_rq->dl_nr_running); | |
732 | dl_rq->dl_nr_running--; | |
733 | ||
734 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
735 | dec_dl_migration(dl_se, dl_rq); | |
736 | } | |
737 | ||
aab03e05 DF |
738 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
739 | { | |
740 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
741 | struct rb_node **link = &dl_rq->rb_root.rb_node; | |
742 | struct rb_node *parent = NULL; | |
743 | struct sched_dl_entity *entry; | |
744 | int leftmost = 1; | |
745 | ||
746 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
747 | ||
748 | while (*link) { | |
749 | parent = *link; | |
750 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
751 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
752 | link = &parent->rb_left; | |
753 | else { | |
754 | link = &parent->rb_right; | |
755 | leftmost = 0; | |
756 | } | |
757 | } | |
758 | ||
759 | if (leftmost) | |
760 | dl_rq->rb_leftmost = &dl_se->rb_node; | |
761 | ||
762 | rb_link_node(&dl_se->rb_node, parent, link); | |
763 | rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); | |
764 | ||
1baca4ce | 765 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
766 | } |
767 | ||
768 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
769 | { | |
770 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
771 | ||
772 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
773 | return; | |
774 | ||
775 | if (dl_rq->rb_leftmost == &dl_se->rb_node) { | |
776 | struct rb_node *next_node; | |
777 | ||
778 | next_node = rb_next(&dl_se->rb_node); | |
779 | dl_rq->rb_leftmost = next_node; | |
780 | } | |
781 | ||
782 | rb_erase(&dl_se->rb_node, &dl_rq->rb_root); | |
783 | RB_CLEAR_NODE(&dl_se->rb_node); | |
784 | ||
1baca4ce | 785 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
786 | } |
787 | ||
788 | static void | |
2d3d891d DF |
789 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
790 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
791 | { |
792 | BUG_ON(on_dl_rq(dl_se)); | |
793 | ||
794 | /* | |
795 | * If this is a wakeup or a new instance, the scheduling | |
796 | * parameters of the task might need updating. Otherwise, | |
797 | * we want a replenishment of its runtime. | |
798 | */ | |
799 | if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH) | |
2d3d891d | 800 | replenish_dl_entity(dl_se, pi_se); |
aab03e05 | 801 | else |
2d3d891d | 802 | update_dl_entity(dl_se, pi_se); |
aab03e05 DF |
803 | |
804 | __enqueue_dl_entity(dl_se); | |
805 | } | |
806 | ||
807 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
808 | { | |
809 | __dequeue_dl_entity(dl_se); | |
810 | } | |
811 | ||
812 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
813 | { | |
2d3d891d DF |
814 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
815 | struct sched_dl_entity *pi_se = &p->dl; | |
816 | ||
817 | /* | |
818 | * Use the scheduling parameters of the top pi-waiter | |
819 | * task if we have one and its (relative) deadline is | |
820 | * smaller than our one... OTW we keep our runtime and | |
821 | * deadline. | |
822 | */ | |
823 | if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) | |
824 | pi_se = &pi_task->dl; | |
825 | ||
aab03e05 DF |
826 | /* |
827 | * If p is throttled, we do nothing. In fact, if it exhausted | |
828 | * its budget it needs a replenishment and, since it now is on | |
829 | * its rq, the bandwidth timer callback (which clearly has not | |
830 | * run yet) will take care of this. | |
831 | */ | |
832 | if (p->dl.dl_throttled) | |
833 | return; | |
834 | ||
2d3d891d | 835 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce JL |
836 | |
837 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) | |
838 | enqueue_pushable_dl_task(rq, p); | |
839 | ||
aab03e05 DF |
840 | inc_nr_running(rq); |
841 | } | |
842 | ||
843 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
844 | { | |
845 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 846 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
847 | } |
848 | ||
849 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
850 | { | |
851 | update_curr_dl(rq); | |
852 | __dequeue_task_dl(rq, p, flags); | |
853 | ||
854 | dec_nr_running(rq); | |
855 | } | |
856 | ||
857 | /* | |
858 | * Yield task semantic for -deadline tasks is: | |
859 | * | |
860 | * get off from the CPU until our next instance, with | |
861 | * a new runtime. This is of little use now, since we | |
862 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
863 | * bandwidth reclaiming is planned for the future, and | |
864 | * yield_task_dl will indicate that some spare budget | |
865 | * is available for other task instances to use it. | |
866 | */ | |
867 | static void yield_task_dl(struct rq *rq) | |
868 | { | |
869 | struct task_struct *p = rq->curr; | |
870 | ||
871 | /* | |
872 | * We make the task go to sleep until its current deadline by | |
873 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
874 | * it and the bandwidth timer will wake it up and will give it | |
875 | * new scheduling parameters (thanks to dl_new=1). | |
876 | */ | |
877 | if (p->dl.runtime > 0) { | |
878 | rq->curr->dl.dl_new = 1; | |
879 | p->dl.runtime = 0; | |
880 | } | |
881 | update_curr_dl(rq); | |
882 | } | |
883 | ||
1baca4ce JL |
884 | #ifdef CONFIG_SMP |
885 | ||
886 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
887 | |
888 | static int | |
889 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
890 | { | |
891 | struct task_struct *curr; | |
892 | struct rq *rq; | |
893 | ||
894 | if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) | |
895 | goto out; | |
896 | ||
897 | rq = cpu_rq(cpu); | |
898 | ||
899 | rcu_read_lock(); | |
900 | curr = ACCESS_ONCE(rq->curr); /* unlocked access */ | |
901 | ||
902 | /* | |
903 | * If we are dealing with a -deadline task, we must | |
904 | * decide where to wake it up. | |
905 | * If it has a later deadline and the current task | |
906 | * on this rq can't move (provided the waking task | |
907 | * can!) we prefer to send it somewhere else. On the | |
908 | * other hand, if it has a shorter deadline, we | |
909 | * try to make it stay here, it might be important. | |
910 | */ | |
911 | if (unlikely(dl_task(curr)) && | |
912 | (curr->nr_cpus_allowed < 2 || | |
913 | !dl_entity_preempt(&p->dl, &curr->dl)) && | |
914 | (p->nr_cpus_allowed > 1)) { | |
915 | int target = find_later_rq(p); | |
916 | ||
917 | if (target != -1) | |
918 | cpu = target; | |
919 | } | |
920 | rcu_read_unlock(); | |
921 | ||
922 | out: | |
923 | return cpu; | |
924 | } | |
925 | ||
926 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) | |
927 | { | |
928 | /* | |
929 | * Current can't be migrated, useless to reschedule, | |
930 | * let's hope p can move out. | |
931 | */ | |
932 | if (rq->curr->nr_cpus_allowed == 1 || | |
6bfd6d72 | 933 | cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) |
1baca4ce JL |
934 | return; |
935 | ||
936 | /* | |
937 | * p is migratable, so let's not schedule it and | |
938 | * see if it is pushed or pulled somewhere else. | |
939 | */ | |
940 | if (p->nr_cpus_allowed != 1 && | |
6bfd6d72 | 941 | cpudl_find(&rq->rd->cpudl, p, NULL) != -1) |
1baca4ce JL |
942 | return; |
943 | ||
944 | resched_task(rq->curr); | |
945 | } | |
946 | ||
947 | #endif /* CONFIG_SMP */ | |
948 | ||
aab03e05 DF |
949 | /* |
950 | * Only called when both the current and waking task are -deadline | |
951 | * tasks. | |
952 | */ | |
953 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
954 | int flags) | |
955 | { | |
1baca4ce | 956 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
aab03e05 | 957 | resched_task(rq->curr); |
1baca4ce JL |
958 | return; |
959 | } | |
960 | ||
961 | #ifdef CONFIG_SMP | |
962 | /* | |
963 | * In the unlikely case current and p have the same deadline | |
964 | * let us try to decide what's the best thing to do... | |
965 | */ | |
332ac17e DF |
966 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
967 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
968 | check_preempt_equal_dl(rq, p); |
969 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
970 | } |
971 | ||
972 | #ifdef CONFIG_SCHED_HRTICK | |
973 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
974 | { | |
975 | s64 delta = p->dl.dl_runtime - p->dl.runtime; | |
976 | ||
977 | if (delta > 10000) | |
978 | hrtick_start(rq, p->dl.runtime); | |
979 | } | |
980 | #endif | |
981 | ||
982 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, | |
983 | struct dl_rq *dl_rq) | |
984 | { | |
985 | struct rb_node *left = dl_rq->rb_leftmost; | |
986 | ||
987 | if (!left) | |
988 | return NULL; | |
989 | ||
990 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
991 | } | |
992 | ||
993 | struct task_struct *pick_next_task_dl(struct rq *rq) | |
994 | { | |
995 | struct sched_dl_entity *dl_se; | |
996 | struct task_struct *p; | |
997 | struct dl_rq *dl_rq; | |
998 | ||
999 | dl_rq = &rq->dl; | |
1000 | ||
1001 | if (unlikely(!dl_rq->dl_nr_running)) | |
1002 | return NULL; | |
1003 | ||
1004 | dl_se = pick_next_dl_entity(rq, dl_rq); | |
1005 | BUG_ON(!dl_se); | |
1006 | ||
1007 | p = dl_task_of(dl_se); | |
1008 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1009 | |
1010 | /* Running task will never be pushed. */ | |
71362650 | 1011 | dequeue_pushable_dl_task(rq, p); |
1baca4ce | 1012 | |
aab03e05 DF |
1013 | #ifdef CONFIG_SCHED_HRTICK |
1014 | if (hrtick_enabled(rq)) | |
1015 | start_hrtick_dl(rq, p); | |
1016 | #endif | |
1baca4ce JL |
1017 | |
1018 | #ifdef CONFIG_SMP | |
1019 | rq->post_schedule = has_pushable_dl_tasks(rq); | |
1020 | #endif /* CONFIG_SMP */ | |
1021 | ||
aab03e05 DF |
1022 | return p; |
1023 | } | |
1024 | ||
1025 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | |
1026 | { | |
1027 | update_curr_dl(rq); | |
1baca4ce JL |
1028 | |
1029 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) | |
1030 | enqueue_pushable_dl_task(rq, p); | |
aab03e05 DF |
1031 | } |
1032 | ||
1033 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) | |
1034 | { | |
1035 | update_curr_dl(rq); | |
1036 | ||
1037 | #ifdef CONFIG_SCHED_HRTICK | |
1038 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0) | |
1039 | start_hrtick_dl(rq, p); | |
1040 | #endif | |
1041 | } | |
1042 | ||
1043 | static void task_fork_dl(struct task_struct *p) | |
1044 | { | |
1045 | /* | |
1046 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1047 | * sched_fork() | |
1048 | */ | |
1049 | } | |
1050 | ||
1051 | static void task_dead_dl(struct task_struct *p) | |
1052 | { | |
1053 | struct hrtimer *timer = &p->dl.dl_timer; | |
332ac17e DF |
1054 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1055 | ||
1056 | /* | |
1057 | * Since we are TASK_DEAD we won't slip out of the domain! | |
1058 | */ | |
1059 | raw_spin_lock_irq(&dl_b->lock); | |
1060 | dl_b->total_bw -= p->dl.dl_bw; | |
1061 | raw_spin_unlock_irq(&dl_b->lock); | |
aab03e05 | 1062 | |
2d3d891d | 1063 | hrtimer_cancel(timer); |
aab03e05 DF |
1064 | } |
1065 | ||
1066 | static void set_curr_task_dl(struct rq *rq) | |
1067 | { | |
1068 | struct task_struct *p = rq->curr; | |
1069 | ||
1070 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1071 | |
1072 | /* You can't push away the running task */ | |
1073 | dequeue_pushable_dl_task(rq, p); | |
1074 | } | |
1075 | ||
1076 | #ifdef CONFIG_SMP | |
1077 | ||
1078 | /* Only try algorithms three times */ | |
1079 | #define DL_MAX_TRIES 3 | |
1080 | ||
1081 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1082 | { | |
1083 | if (!task_running(rq, p) && | |
1084 | (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) && | |
1085 | (p->nr_cpus_allowed > 1)) | |
1086 | return 1; | |
1087 | ||
1088 | return 0; | |
1089 | } | |
1090 | ||
1091 | /* Returns the second earliest -deadline task, NULL otherwise */ | |
1092 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu) | |
1093 | { | |
1094 | struct rb_node *next_node = rq->dl.rb_leftmost; | |
1095 | struct sched_dl_entity *dl_se; | |
1096 | struct task_struct *p = NULL; | |
1097 | ||
1098 | next_node: | |
1099 | next_node = rb_next(next_node); | |
1100 | if (next_node) { | |
1101 | dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node); | |
1102 | p = dl_task_of(dl_se); | |
1103 | ||
1104 | if (pick_dl_task(rq, p, cpu)) | |
1105 | return p; | |
1106 | ||
1107 | goto next_node; | |
1108 | } | |
1109 | ||
1110 | return NULL; | |
1111 | } | |
1112 | ||
1baca4ce JL |
1113 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1114 | ||
1115 | static int find_later_rq(struct task_struct *task) | |
1116 | { | |
1117 | struct sched_domain *sd; | |
1118 | struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl); | |
1119 | int this_cpu = smp_processor_id(); | |
1120 | int best_cpu, cpu = task_cpu(task); | |
1121 | ||
1122 | /* Make sure the mask is initialized first */ | |
1123 | if (unlikely(!later_mask)) | |
1124 | return -1; | |
1125 | ||
1126 | if (task->nr_cpus_allowed == 1) | |
1127 | return -1; | |
1128 | ||
6bfd6d72 JL |
1129 | best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, |
1130 | task, later_mask); | |
1baca4ce JL |
1131 | if (best_cpu == -1) |
1132 | return -1; | |
1133 | ||
1134 | /* | |
1135 | * If we are here, some target has been found, | |
1136 | * the most suitable of which is cached in best_cpu. | |
1137 | * This is, among the runqueues where the current tasks | |
1138 | * have later deadlines than the task's one, the rq | |
1139 | * with the latest possible one. | |
1140 | * | |
1141 | * Now we check how well this matches with task's | |
1142 | * affinity and system topology. | |
1143 | * | |
1144 | * The last cpu where the task run is our first | |
1145 | * guess, since it is most likely cache-hot there. | |
1146 | */ | |
1147 | if (cpumask_test_cpu(cpu, later_mask)) | |
1148 | return cpu; | |
1149 | /* | |
1150 | * Check if this_cpu is to be skipped (i.e., it is | |
1151 | * not in the mask) or not. | |
1152 | */ | |
1153 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1154 | this_cpu = -1; | |
1155 | ||
1156 | rcu_read_lock(); | |
1157 | for_each_domain(cpu, sd) { | |
1158 | if (sd->flags & SD_WAKE_AFFINE) { | |
1159 | ||
1160 | /* | |
1161 | * If possible, preempting this_cpu is | |
1162 | * cheaper than migrating. | |
1163 | */ | |
1164 | if (this_cpu != -1 && | |
1165 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1166 | rcu_read_unlock(); | |
1167 | return this_cpu; | |
1168 | } | |
1169 | ||
1170 | /* | |
1171 | * Last chance: if best_cpu is valid and is | |
1172 | * in the mask, that becomes our choice. | |
1173 | */ | |
1174 | if (best_cpu < nr_cpu_ids && | |
1175 | cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { | |
1176 | rcu_read_unlock(); | |
1177 | return best_cpu; | |
1178 | } | |
1179 | } | |
1180 | } | |
1181 | rcu_read_unlock(); | |
1182 | ||
1183 | /* | |
1184 | * At this point, all our guesses failed, we just return | |
1185 | * 'something', and let the caller sort the things out. | |
1186 | */ | |
1187 | if (this_cpu != -1) | |
1188 | return this_cpu; | |
1189 | ||
1190 | cpu = cpumask_any(later_mask); | |
1191 | if (cpu < nr_cpu_ids) | |
1192 | return cpu; | |
1193 | ||
1194 | return -1; | |
1195 | } | |
1196 | ||
1197 | /* Locks the rq it finds */ | |
1198 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
1199 | { | |
1200 | struct rq *later_rq = NULL; | |
1201 | int tries; | |
1202 | int cpu; | |
1203 | ||
1204 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
1205 | cpu = find_later_rq(task); | |
1206 | ||
1207 | if ((cpu == -1) || (cpu == rq->cpu)) | |
1208 | break; | |
1209 | ||
1210 | later_rq = cpu_rq(cpu); | |
1211 | ||
1212 | /* Retry if something changed. */ | |
1213 | if (double_lock_balance(rq, later_rq)) { | |
1214 | if (unlikely(task_rq(task) != rq || | |
1215 | !cpumask_test_cpu(later_rq->cpu, | |
1216 | &task->cpus_allowed) || | |
1217 | task_running(rq, task) || !task->on_rq)) { | |
1218 | double_unlock_balance(rq, later_rq); | |
1219 | later_rq = NULL; | |
1220 | break; | |
1221 | } | |
1222 | } | |
1223 | ||
1224 | /* | |
1225 | * If the rq we found has no -deadline task, or | |
1226 | * its earliest one has a later deadline than our | |
1227 | * task, the rq is a good one. | |
1228 | */ | |
1229 | if (!later_rq->dl.dl_nr_running || | |
1230 | dl_time_before(task->dl.deadline, | |
1231 | later_rq->dl.earliest_dl.curr)) | |
1232 | break; | |
1233 | ||
1234 | /* Otherwise we try again. */ | |
1235 | double_unlock_balance(rq, later_rq); | |
1236 | later_rq = NULL; | |
1237 | } | |
1238 | ||
1239 | return later_rq; | |
1240 | } | |
1241 | ||
1242 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
1243 | { | |
1244 | struct task_struct *p; | |
1245 | ||
1246 | if (!has_pushable_dl_tasks(rq)) | |
1247 | return NULL; | |
1248 | ||
1249 | p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, | |
1250 | struct task_struct, pushable_dl_tasks); | |
1251 | ||
1252 | BUG_ON(rq->cpu != task_cpu(p)); | |
1253 | BUG_ON(task_current(rq, p)); | |
1254 | BUG_ON(p->nr_cpus_allowed <= 1); | |
1255 | ||
332ac17e | 1256 | BUG_ON(!p->on_rq); |
1baca4ce JL |
1257 | BUG_ON(!dl_task(p)); |
1258 | ||
1259 | return p; | |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * See if the non running -deadline tasks on this rq | |
1264 | * can be sent to some other CPU where they can preempt | |
1265 | * and start executing. | |
1266 | */ | |
1267 | static int push_dl_task(struct rq *rq) | |
1268 | { | |
1269 | struct task_struct *next_task; | |
1270 | struct rq *later_rq; | |
1271 | ||
1272 | if (!rq->dl.overloaded) | |
1273 | return 0; | |
1274 | ||
1275 | next_task = pick_next_pushable_dl_task(rq); | |
1276 | if (!next_task) | |
1277 | return 0; | |
1278 | ||
1279 | retry: | |
1280 | if (unlikely(next_task == rq->curr)) { | |
1281 | WARN_ON(1); | |
1282 | return 0; | |
1283 | } | |
1284 | ||
1285 | /* | |
1286 | * If next_task preempts rq->curr, and rq->curr | |
1287 | * can move away, it makes sense to just reschedule | |
1288 | * without going further in pushing next_task. | |
1289 | */ | |
1290 | if (dl_task(rq->curr) && | |
1291 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
1292 | rq->curr->nr_cpus_allowed > 1) { | |
1293 | resched_task(rq->curr); | |
1294 | return 0; | |
1295 | } | |
1296 | ||
1297 | /* We might release rq lock */ | |
1298 | get_task_struct(next_task); | |
1299 | ||
1300 | /* Will lock the rq it'll find */ | |
1301 | later_rq = find_lock_later_rq(next_task, rq); | |
1302 | if (!later_rq) { | |
1303 | struct task_struct *task; | |
1304 | ||
1305 | /* | |
1306 | * We must check all this again, since | |
1307 | * find_lock_later_rq releases rq->lock and it is | |
1308 | * then possible that next_task has migrated. | |
1309 | */ | |
1310 | task = pick_next_pushable_dl_task(rq); | |
1311 | if (task_cpu(next_task) == rq->cpu && task == next_task) { | |
1312 | /* | |
1313 | * The task is still there. We don't try | |
1314 | * again, some other cpu will pull it when ready. | |
1315 | */ | |
1316 | dequeue_pushable_dl_task(rq, next_task); | |
1317 | goto out; | |
1318 | } | |
1319 | ||
1320 | if (!task) | |
1321 | /* No more tasks */ | |
1322 | goto out; | |
1323 | ||
1324 | put_task_struct(next_task); | |
1325 | next_task = task; | |
1326 | goto retry; | |
1327 | } | |
1328 | ||
1329 | deactivate_task(rq, next_task, 0); | |
1330 | set_task_cpu(next_task, later_rq->cpu); | |
1331 | activate_task(later_rq, next_task, 0); | |
1332 | ||
1333 | resched_task(later_rq->curr); | |
1334 | ||
1335 | double_unlock_balance(rq, later_rq); | |
1336 | ||
1337 | out: | |
1338 | put_task_struct(next_task); | |
1339 | ||
1340 | return 1; | |
1341 | } | |
1342 | ||
1343 | static void push_dl_tasks(struct rq *rq) | |
1344 | { | |
1345 | /* Terminates as it moves a -deadline task */ | |
1346 | while (push_dl_task(rq)) | |
1347 | ; | |
aab03e05 DF |
1348 | } |
1349 | ||
1baca4ce JL |
1350 | static int pull_dl_task(struct rq *this_rq) |
1351 | { | |
1352 | int this_cpu = this_rq->cpu, ret = 0, cpu; | |
1353 | struct task_struct *p; | |
1354 | struct rq *src_rq; | |
1355 | u64 dmin = LONG_MAX; | |
1356 | ||
1357 | if (likely(!dl_overloaded(this_rq))) | |
1358 | return 0; | |
1359 | ||
1360 | /* | |
1361 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
1362 | * see overloaded we must also see the dlo_mask bit. | |
1363 | */ | |
1364 | smp_rmb(); | |
1365 | ||
1366 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
1367 | if (this_cpu == cpu) | |
1368 | continue; | |
1369 | ||
1370 | src_rq = cpu_rq(cpu); | |
1371 | ||
1372 | /* | |
1373 | * It looks racy, abd it is! However, as in sched_rt.c, | |
1374 | * we are fine with this. | |
1375 | */ | |
1376 | if (this_rq->dl.dl_nr_running && | |
1377 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
1378 | src_rq->dl.earliest_dl.next)) | |
1379 | continue; | |
1380 | ||
1381 | /* Might drop this_rq->lock */ | |
1382 | double_lock_balance(this_rq, src_rq); | |
1383 | ||
1384 | /* | |
1385 | * If there are no more pullable tasks on the | |
1386 | * rq, we're done with it. | |
1387 | */ | |
1388 | if (src_rq->dl.dl_nr_running <= 1) | |
1389 | goto skip; | |
1390 | ||
1391 | p = pick_next_earliest_dl_task(src_rq, this_cpu); | |
1392 | ||
1393 | /* | |
1394 | * We found a task to be pulled if: | |
1395 | * - it preempts our current (if there's one), | |
1396 | * - it will preempt the last one we pulled (if any). | |
1397 | */ | |
1398 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
1399 | (!this_rq->dl.dl_nr_running || | |
1400 | dl_time_before(p->dl.deadline, | |
1401 | this_rq->dl.earliest_dl.curr))) { | |
1402 | WARN_ON(p == src_rq->curr); | |
332ac17e | 1403 | WARN_ON(!p->on_rq); |
1baca4ce JL |
1404 | |
1405 | /* | |
1406 | * Then we pull iff p has actually an earlier | |
1407 | * deadline than the current task of its runqueue. | |
1408 | */ | |
1409 | if (dl_time_before(p->dl.deadline, | |
1410 | src_rq->curr->dl.deadline)) | |
1411 | goto skip; | |
1412 | ||
1413 | ret = 1; | |
1414 | ||
1415 | deactivate_task(src_rq, p, 0); | |
1416 | set_task_cpu(p, this_cpu); | |
1417 | activate_task(this_rq, p, 0); | |
1418 | dmin = p->dl.deadline; | |
1419 | ||
1420 | /* Is there any other task even earlier? */ | |
1421 | } | |
1422 | skip: | |
1423 | double_unlock_balance(this_rq, src_rq); | |
1424 | } | |
1425 | ||
1426 | return ret; | |
1427 | } | |
1428 | ||
1429 | static void pre_schedule_dl(struct rq *rq, struct task_struct *prev) | |
1430 | { | |
1431 | /* Try to pull other tasks here */ | |
1432 | if (dl_task(prev)) | |
1433 | pull_dl_task(rq); | |
1434 | } | |
1435 | ||
1436 | static void post_schedule_dl(struct rq *rq) | |
1437 | { | |
1438 | push_dl_tasks(rq); | |
1439 | } | |
1440 | ||
1441 | /* | |
1442 | * Since the task is not running and a reschedule is not going to happen | |
1443 | * anytime soon on its runqueue, we try pushing it away now. | |
1444 | */ | |
1445 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
1446 | { | |
1447 | if (!task_running(rq, p) && | |
1448 | !test_tsk_need_resched(rq->curr) && | |
1449 | has_pushable_dl_tasks(rq) && | |
1450 | p->nr_cpus_allowed > 1 && | |
1451 | dl_task(rq->curr) && | |
1452 | (rq->curr->nr_cpus_allowed < 2 || | |
1453 | dl_entity_preempt(&rq->curr->dl, &p->dl))) { | |
1454 | push_dl_tasks(rq); | |
1455 | } | |
1456 | } | |
1457 | ||
1458 | static void set_cpus_allowed_dl(struct task_struct *p, | |
1459 | const struct cpumask *new_mask) | |
1460 | { | |
1461 | struct rq *rq; | |
1462 | int weight; | |
1463 | ||
1464 | BUG_ON(!dl_task(p)); | |
1465 | ||
1466 | /* | |
1467 | * Update only if the task is actually running (i.e., | |
1468 | * it is on the rq AND it is not throttled). | |
1469 | */ | |
1470 | if (!on_dl_rq(&p->dl)) | |
1471 | return; | |
1472 | ||
1473 | weight = cpumask_weight(new_mask); | |
1474 | ||
1475 | /* | |
1476 | * Only update if the process changes its state from whether it | |
1477 | * can migrate or not. | |
1478 | */ | |
1479 | if ((p->nr_cpus_allowed > 1) == (weight > 1)) | |
1480 | return; | |
1481 | ||
1482 | rq = task_rq(p); | |
1483 | ||
1484 | /* | |
1485 | * The process used to be able to migrate OR it can now migrate | |
1486 | */ | |
1487 | if (weight <= 1) { | |
1488 | if (!task_current(rq, p)) | |
1489 | dequeue_pushable_dl_task(rq, p); | |
1490 | BUG_ON(!rq->dl.dl_nr_migratory); | |
1491 | rq->dl.dl_nr_migratory--; | |
1492 | } else { | |
1493 | if (!task_current(rq, p)) | |
1494 | enqueue_pushable_dl_task(rq, p); | |
1495 | rq->dl.dl_nr_migratory++; | |
1496 | } | |
1497 | ||
1498 | update_dl_migration(&rq->dl); | |
1499 | } | |
1500 | ||
1501 | /* Assumes rq->lock is held */ | |
1502 | static void rq_online_dl(struct rq *rq) | |
1503 | { | |
1504 | if (rq->dl.overloaded) | |
1505 | dl_set_overload(rq); | |
6bfd6d72 JL |
1506 | |
1507 | if (rq->dl.dl_nr_running > 0) | |
1508 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); | |
1baca4ce JL |
1509 | } |
1510 | ||
1511 | /* Assumes rq->lock is held */ | |
1512 | static void rq_offline_dl(struct rq *rq) | |
1513 | { | |
1514 | if (rq->dl.overloaded) | |
1515 | dl_clear_overload(rq); | |
6bfd6d72 JL |
1516 | |
1517 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); | |
1baca4ce JL |
1518 | } |
1519 | ||
1520 | void init_sched_dl_class(void) | |
1521 | { | |
1522 | unsigned int i; | |
1523 | ||
1524 | for_each_possible_cpu(i) | |
1525 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
1526 | GFP_KERNEL, cpu_to_node(i)); | |
1527 | } | |
1528 | ||
1529 | #endif /* CONFIG_SMP */ | |
1530 | ||
aab03e05 DF |
1531 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
1532 | { | |
1baca4ce | 1533 | if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy)) |
aab03e05 | 1534 | hrtimer_try_to_cancel(&p->dl.dl_timer); |
1baca4ce JL |
1535 | |
1536 | #ifdef CONFIG_SMP | |
1537 | /* | |
1538 | * Since this might be the only -deadline task on the rq, | |
1539 | * this is the right place to try to pull some other one | |
1540 | * from an overloaded cpu, if any. | |
1541 | */ | |
1542 | if (!rq->dl.dl_nr_running) | |
1543 | pull_dl_task(rq); | |
1544 | #endif | |
aab03e05 DF |
1545 | } |
1546 | ||
1baca4ce JL |
1547 | /* |
1548 | * When switching to -deadline, we may overload the rq, then | |
1549 | * we try to push someone off, if possible. | |
1550 | */ | |
aab03e05 DF |
1551 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
1552 | { | |
1baca4ce JL |
1553 | int check_resched = 1; |
1554 | ||
aab03e05 DF |
1555 | /* |
1556 | * If p is throttled, don't consider the possibility | |
1557 | * of preempting rq->curr, the check will be done right | |
1558 | * after its runtime will get replenished. | |
1559 | */ | |
1560 | if (unlikely(p->dl.dl_throttled)) | |
1561 | return; | |
1562 | ||
1563 | if (p->on_rq || rq->curr != p) { | |
1baca4ce JL |
1564 | #ifdef CONFIG_SMP |
1565 | if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p)) | |
1566 | /* Only reschedule if pushing failed */ | |
1567 | check_resched = 0; | |
1568 | #endif /* CONFIG_SMP */ | |
1569 | if (check_resched && task_has_dl_policy(rq->curr)) | |
aab03e05 | 1570 | check_preempt_curr_dl(rq, p, 0); |
aab03e05 DF |
1571 | } |
1572 | } | |
1573 | ||
1baca4ce JL |
1574 | /* |
1575 | * If the scheduling parameters of a -deadline task changed, | |
1576 | * a push or pull operation might be needed. | |
1577 | */ | |
aab03e05 DF |
1578 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
1579 | int oldprio) | |
1580 | { | |
1baca4ce | 1581 | if (p->on_rq || rq->curr == p) { |
aab03e05 | 1582 | #ifdef CONFIG_SMP |
1baca4ce JL |
1583 | /* |
1584 | * This might be too much, but unfortunately | |
1585 | * we don't have the old deadline value, and | |
1586 | * we can't argue if the task is increasing | |
1587 | * or lowering its prio, so... | |
1588 | */ | |
1589 | if (!rq->dl.overloaded) | |
1590 | pull_dl_task(rq); | |
1591 | ||
1592 | /* | |
1593 | * If we now have a earlier deadline task than p, | |
1594 | * then reschedule, provided p is still on this | |
1595 | * runqueue. | |
1596 | */ | |
1597 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) && | |
1598 | rq->curr == p) | |
1599 | resched_task(p); | |
1600 | #else | |
1601 | /* | |
1602 | * Again, we don't know if p has a earlier | |
1603 | * or later deadline, so let's blindly set a | |
1604 | * (maybe not needed) rescheduling point. | |
1605 | */ | |
1606 | resched_task(p); | |
1607 | #endif /* CONFIG_SMP */ | |
1608 | } else | |
1609 | switched_to_dl(rq, p); | |
aab03e05 | 1610 | } |
aab03e05 DF |
1611 | |
1612 | const struct sched_class dl_sched_class = { | |
1613 | .next = &rt_sched_class, | |
1614 | .enqueue_task = enqueue_task_dl, | |
1615 | .dequeue_task = dequeue_task_dl, | |
1616 | .yield_task = yield_task_dl, | |
1617 | ||
1618 | .check_preempt_curr = check_preempt_curr_dl, | |
1619 | ||
1620 | .pick_next_task = pick_next_task_dl, | |
1621 | .put_prev_task = put_prev_task_dl, | |
1622 | ||
1623 | #ifdef CONFIG_SMP | |
1624 | .select_task_rq = select_task_rq_dl, | |
1baca4ce JL |
1625 | .set_cpus_allowed = set_cpus_allowed_dl, |
1626 | .rq_online = rq_online_dl, | |
1627 | .rq_offline = rq_offline_dl, | |
1628 | .pre_schedule = pre_schedule_dl, | |
1629 | .post_schedule = post_schedule_dl, | |
1630 | .task_woken = task_woken_dl, | |
aab03e05 DF |
1631 | #endif |
1632 | ||
1633 | .set_curr_task = set_curr_task_dl, | |
1634 | .task_tick = task_tick_dl, | |
1635 | .task_fork = task_fork_dl, | |
1636 | .task_dead = task_dead_dl, | |
1637 | ||
1638 | .prio_changed = prio_changed_dl, | |
1639 | .switched_from = switched_from_dl, | |
1640 | .switched_to = switched_to_dl, | |
1641 | }; |