1 // SPDX-License-Identifier: GPL-2.0 <<
2 /* 1 /*
3 * Deadline Scheduling Class (SCHED_DEADLINE) 2 * Deadline Scheduling Class (SCHED_DEADLINE)
4 * 3 *
5 * Earliest Deadline First (EDF) + Constant Ba 4 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 * 5 *
7 * Tasks that periodically executes their inst 6 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines. 7 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or 8 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowe 9 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't af 10 * miss some of their deadlines), and won't affect any other task.
12 * 11 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin 12 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@g 13 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <micha 14 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fcheccon 15 * Fabio Checconi <fchecconi@gmail.com>
17 */ 16 */
>> 17 #include "sched.h"
18 18
19 #include <linux/cpuset.h> !! 19 #include <linux/slab.h>
20 20
21 /* !! 21 struct dl_bandwidth def_dl_bandwidth;
22 * Default limits for DL period; on the top en <<
23 * tasks still getting ridiculously long effec <<
24 * guard against timer DoS. <<
25 */ <<
26 static unsigned int sysctl_sched_dl_period_max <<
27 static unsigned int sysctl_sched_dl_period_min <<
28 #ifdef CONFIG_SYSCTL <<
29 static struct ctl_table sched_dl_sysctls[] = { <<
30 { <<
31 .procname = "sched_deadl <<
32 .data = &sysctl_sche <<
33 .maxlen = sizeof(unsig <<
34 .mode = 0644, <<
35 .proc_handler = proc_douintv <<
36 .extra1 = (void *)&sys <<
37 }, <<
38 { <<
39 .procname = "sched_deadl <<
40 .data = &sysctl_sche <<
41 .maxlen = sizeof(unsig <<
42 .mode = 0644, <<
43 .proc_handler = proc_douintv <<
44 .extra2 = (void *)&sys <<
45 }, <<
46 }; <<
47 <<
48 static int __init sched_dl_sysctl_init(void) <<
49 { <<
50 register_sysctl_init("kernel", sched_d <<
51 return 0; <<
52 } <<
53 late_initcall(sched_dl_sysctl_init); <<
54 #endif <<
55 <<
56 static bool dl_server(struct sched_dl_entity * <<
57 { <<
58 return dl_se->dl_server; <<
59 } <<
60 22
61 static inline struct task_struct *dl_task_of(s 23 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
62 { 24 {
63 BUG_ON(dl_server(dl_se)); <<
64 return container_of(dl_se, struct task 25 return container_of(dl_se, struct task_struct, dl);
65 } 26 }
66 27
67 static inline struct rq *rq_of_dl_rq(struct dl 28 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
68 { 29 {
69 return container_of(dl_rq, struct rq, 30 return container_of(dl_rq, struct rq, dl);
70 } 31 }
71 32
72 static inline struct rq *rq_of_dl_se(struct sc <<
73 { <<
74 struct rq *rq = dl_se->rq; <<
75 <<
76 if (!dl_server(dl_se)) <<
77 rq = task_rq(dl_task_of(dl_se) <<
78 <<
79 return rq; <<
80 } <<
81 <<
82 static inline struct dl_rq *dl_rq_of_se(struct 33 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
83 { 34 {
84 return &rq_of_dl_se(dl_se)->dl; !! 35 struct task_struct *p = dl_task_of(dl_se);
>> 36 struct rq *rq = task_rq(p);
>> 37
>> 38 return &rq->dl;
85 } 39 }
86 40
87 static inline int on_dl_rq(struct sched_dl_ent 41 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
88 { 42 {
89 return !RB_EMPTY_NODE(&dl_se->rb_node) 43 return !RB_EMPTY_NODE(&dl_se->rb_node);
90 } 44 }
91 45
92 #ifdef CONFIG_RT_MUTEXES !! 46 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
93 static inline struct sched_dl_entity *pi_of(st <<
94 { <<
95 return dl_se->pi_se; <<
96 } <<
97 <<
98 static inline bool is_dl_boosted(struct sched_ <<
99 { <<
100 return pi_of(dl_se) != dl_se; <<
101 } <<
102 #else <<
103 static inline struct sched_dl_entity *pi_of(st <<
104 { <<
105 return dl_se; <<
106 } <<
107 <<
108 static inline bool is_dl_boosted(struct sched_ <<
109 { <<
110 return false; <<
111 } <<
112 #endif <<
113 <<
114 #ifdef CONFIG_SMP <<
115 static inline struct dl_bw *dl_bw_of(int i) <<
116 { <<
117 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ <<
118 "sched RCU must be he <<
119 return &cpu_rq(i)->rd->dl_bw; <<
120 } <<
121 <<
122 static inline int dl_bw_cpus(int i) <<
123 { <<
124 struct root_domain *rd = cpu_rq(i)->rd <<
125 int cpus; <<
126 <<
127 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ <<
128 "sched RCU must be he <<
129 <<
130 if (cpumask_subset(rd->span, cpu_activ <<
131 return cpumask_weight(rd->span <<
132 <<
133 cpus = 0; <<
134 <<
135 for_each_cpu_and(i, rd->span, cpu_acti <<
136 cpus++; <<
137 <<
138 return cpus; <<
139 } <<
140 <<
141 static inline unsigned long __dl_bw_capacity(c <<
142 { 47 {
143 unsigned long cap = 0; !! 48 struct sched_dl_entity *dl_se = &p->dl;
144 int i; <<
145 <<
146 for_each_cpu_and(i, mask, cpu_active_m <<
147 cap += arch_scale_cpu_capacity <<
148 <<
149 return cap; <<
150 } <<
151 <<
152 /* <<
153 * XXX Fix: If 'rq->rd == def_root_domain' per <<
154 * of the CPU the task is running on rather rd <<
155 */ <<
156 static inline unsigned long dl_bw_capacity(int <<
157 { <<
158 if (!sched_asym_cpucap_active() && <<
159 arch_scale_cpu_capacity(i) == SCHE <<
160 return dl_bw_cpus(i) << SCHED_ <<
161 } else { <<
162 RCU_LOCKDEP_WARN(!rcu_read_loc <<
163 "sched RCU mu <<
164 <<
165 return __dl_bw_capacity(cpu_rq <<
166 } <<
167 } <<
168 <<
169 static inline bool dl_bw_visited(int cpu, u64 <<
170 { <<
171 struct root_domain *rd = cpu_rq(cpu)-> <<
172 <<
173 if (rd->visit_gen == gen) <<
174 return true; <<
175 <<
176 rd->visit_gen = gen; <<
177 return false; <<
178 } <<
179 <<
180 static inline <<
181 void __dl_update(struct dl_bw *dl_b, s64 bw) <<
182 { <<
183 struct root_domain *rd = container_of( <<
184 int i; <<
185 <<
186 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ <<
187 "sched RCU must be he <<
188 for_each_cpu_and(i, rd->span, cpu_acti <<
189 struct rq *rq = cpu_rq(i); <<
190 <<
191 rq->dl.extra_bw += bw; <<
192 } <<
193 } <<
194 #else <<
195 static inline struct dl_bw *dl_bw_of(int i) <<
196 { <<
197 return &cpu_rq(i)->dl.dl_bw; <<
198 } <<
199 <<
200 static inline int dl_bw_cpus(int i) <<
201 { <<
202 return 1; <<
203 } <<
204 <<
205 static inline unsigned long dl_bw_capacity(int <<
206 { <<
207 return SCHED_CAPACITY_SCALE; <<
208 } <<
209 <<
210 static inline bool dl_bw_visited(int cpu, u64 <<
211 { <<
212 return false; <<
213 } <<
214 <<
215 static inline <<
216 void __dl_update(struct dl_bw *dl_b, s64 bw) <<
217 { <<
218 struct dl_rq *dl = container_of(dl_b, <<
219 <<
220 dl->extra_bw += bw; <<
221 } <<
222 #endif <<
223 <<
224 static inline <<
225 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, <<
226 { <<
227 dl_b->total_bw -= tsk_bw; <<
228 __dl_update(dl_b, (s32)tsk_bw / cpus); <<
229 } <<
230 <<
231 static inline <<
232 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, <<
233 { <<
234 dl_b->total_bw += tsk_bw; <<
235 __dl_update(dl_b, -((s32)tsk_bw / cpus <<
236 } <<
237 <<
238 static inline bool <<
239 __dl_overflow(struct dl_bw *dl_b, unsigned lon <<
240 { <<
241 return dl_b->bw != -1 && <<
242 cap_scale(dl_b->bw, cap) < dl_b <<
243 } <<
244 <<
245 static inline <<
246 void __add_running_bw(u64 dl_bw, struct dl_rq <<
247 { <<
248 u64 old = dl_rq->running_bw; <<
249 <<
250 lockdep_assert_rq_held(rq_of_dl_rq(dl_ <<
251 dl_rq->running_bw += dl_bw; <<
252 SCHED_WARN_ON(dl_rq->running_bw < old) <<
253 SCHED_WARN_ON(dl_rq->running_bw > dl_r <<
254 /* kick cpufreq (see the comment in ke <<
255 cpufreq_update_util(rq_of_dl_rq(dl_rq) <<
256 } <<
257 <<
258 static inline <<
259 void __sub_running_bw(u64 dl_bw, struct dl_rq <<
260 { <<
261 u64 old = dl_rq->running_bw; <<
262 <<
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_ <<
264 dl_rq->running_bw -= dl_bw; <<
265 SCHED_WARN_ON(dl_rq->running_bw > old) <<
266 if (dl_rq->running_bw > old) <<
267 dl_rq->running_bw = 0; <<
268 /* kick cpufreq (see the comment in ke <<
269 cpufreq_update_util(rq_of_dl_rq(dl_rq) <<
270 } <<
271 <<
272 static inline <<
273 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_r <<
274 { <<
275 u64 old = dl_rq->this_bw; <<
276 <<
277 lockdep_assert_rq_held(rq_of_dl_rq(dl_ <<
278 dl_rq->this_bw += dl_bw; <<
279 SCHED_WARN_ON(dl_rq->this_bw < old); / <<
280 } <<
281 <<
282 static inline <<
283 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_r <<
284 { <<
285 u64 old = dl_rq->this_bw; <<
286 <<
287 lockdep_assert_rq_held(rq_of_dl_rq(dl_ <<
288 dl_rq->this_bw -= dl_bw; <<
289 SCHED_WARN_ON(dl_rq->this_bw > old); / <<
290 if (dl_rq->this_bw > old) <<
291 dl_rq->this_bw = 0; <<
292 SCHED_WARN_ON(dl_rq->running_bw > dl_r <<
293 } <<
294 <<
295 static inline <<
296 void add_rq_bw(struct sched_dl_entity *dl_se, <<
297 { <<
298 if (!dl_entity_is_special(dl_se)) <<
299 __add_rq_bw(dl_se->dl_bw, dl_r <<
300 } <<
301 <<
302 static inline <<
303 void sub_rq_bw(struct sched_dl_entity *dl_se, <<
304 { <<
305 if (!dl_entity_is_special(dl_se)) <<
306 __sub_rq_bw(dl_se->dl_bw, dl_r <<
307 } <<
308 <<
309 static inline <<
310 void add_running_bw(struct sched_dl_entity *dl <<
311 { <<
312 if (!dl_entity_is_special(dl_se)) <<
313 __add_running_bw(dl_se->dl_bw, <<
314 } <<
315 <<
316 static inline <<
317 void sub_running_bw(struct sched_dl_entity *dl <<
318 { <<
319 if (!dl_entity_is_special(dl_se)) <<
320 __sub_running_bw(dl_se->dl_bw, <<
321 } <<
322 <<
323 static void dl_rq_change_utilization(struct rq <<
324 { <<
325 if (dl_se->dl_non_contending) { <<
326 sub_running_bw(dl_se, &rq->dl) <<
327 dl_se->dl_non_contending = 0; <<
328 <<
329 /* <<
330 * If the timer handler is cur <<
331 * timer cannot be canceled, i <<
332 * will see that dl_not_conten <<
333 * will not touch the rq's act <<
334 * so we are still safe. <<
335 */ <<
336 if (hrtimer_try_to_cancel(&dl_ <<
337 if (!dl_server(dl_se)) <<
338 put_task_struc <<
339 } <<
340 } <<
341 __sub_rq_bw(dl_se->dl_bw, &rq->dl); <<
342 __add_rq_bw(new_bw, &rq->dl); <<
343 } <<
344 <<
345 static void dl_change_utilization(struct task_ <<
346 { <<
347 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_ <<
348 <<
349 if (task_on_rq_queued(p)) <<
350 return; <<
351 <<
352 dl_rq_change_utilization(task_rq(p), & <<
353 } <<
354 <<
355 static void __dl_clear_params(struct sched_dl_ <<
356 <<
357 /* <<
358 * The utilization of a task cannot be immedia <<
359 * the rq active utilization (running_bw) when <<
360 * Instead, we have to wait for the so called <<
361 * <<
362 * If a task blocks before the "0-lag time", a <<
363 * timer) is armed, and running_bw is decrease <<
364 * fires. <<
365 * <<
366 * If the task wakes up again before the inact <<
367 * the timer is canceled, whereas if the task <<
368 * inactive timer fired (and running_bw has be <<
369 * task's utilization has to be added to runni <<
370 * A flag in the deadline scheduling entity (d <<
371 * is used to avoid race conditions between th <<
372 * and task wakeups. <<
373 * <<
374 * The following diagram shows how running_bw <<
375 * "ACTIVE" when its utilization contributes t <<
376 * "ACTIVE contending" task is in the TASK_RUN <<
377 * "ACTIVE non contending" task is a blocked t <<
378 * has not passed yet. An "INACTIVE" task is a <<
379 * time already passed, which does not contrib <<
380 * +------------- <<
381 * wakeup | ACTIVE <<
382 * +------------------>+ contending <<
383 * | add_running_bw | <<
384 * | +----+------+- <<
385 * | | ^ <<
386 * | dequeue | | <<
387 * +--------+-------+ | | <<
388 * | | t >= 0-lag | | <<
389 * | INACTIVE |<---------------+ | <<
390 * | | sub_running_bw | | <<
391 * +--------+-------+ | | <<
392 * ^ | | <<
393 * | t < 0-lag | | <<
394 * | | | <<
395 * | V | <<
396 * | +----+------+- <<
397 * | sub_running_bw | ACTIVE <<
398 * +-------------------+ <<
399 * inactive timer | non contend <<
400 * fired +------------- <<
401 * <<
402 * The task_non_contending() function is invok <<
403 * blocks, and checks if the 0-lag time alread <<
404 * not (in the first case, it directly updates <<
405 * in the second case, it arms the inactive ti <<
406 * <<
407 * The task_contending() function is invoked w <<
408 * up, and checks if the task is still in the <<
409 * state or not (in the second case, it update <<
410 */ <<
411 static void task_non_contending(struct sched_d <<
412 { <<
413 struct hrtimer *timer = &dl_se->inacti <<
414 struct rq *rq = rq_of_dl_se(dl_se); <<
415 struct dl_rq *dl_rq = &rq->dl; <<
416 s64 zerolag_time; <<
417 <<
418 /* <<
419 * If this is a non-deadline task that <<
420 * do nothing <<
421 */ <<
422 if (dl_se->dl_runtime == 0) <<
423 return; <<
424 <<
425 if (dl_entity_is_special(dl_se)) <<
426 return; <<
427 <<
428 WARN_ON(dl_se->dl_non_contending); <<
429 <<
430 zerolag_time = dl_se->deadline - <<
431 div64_long((dl_se->runtime * <<
432 dl_se->dl_runtime); <<
433 <<
434 /* <<
435 * Using relative times instead of the <<
436 * allows to simplify the code <<
437 */ <<
438 zerolag_time -= rq_clock(rq); <<
439 <<
440 /* <<
441 * If the "0-lag time" already passed, <<
442 * utilization now, instead of startin <<
443 */ <<
444 if ((zerolag_time < 0) || hrtimer_acti <<
445 if (dl_server(dl_se)) { <<
446 sub_running_bw(dl_se, <<
447 } else { <<
448 struct task_struct *p <<
449 <<
450 if (dl_task(p)) <<
451 sub_running_bw <<
452 <<
453 if (!dl_task(p) || REA <<
454 struct dl_bw * <<
455 <<
456 if (READ_ONCE( <<
457 sub_rq <<
458 raw_spin_lock( <<
459 __dl_sub(dl_b, <<
460 raw_spin_unloc <<
461 __dl_clear_par <<
462 } <<
463 } <<
464 <<
465 return; <<
466 } <<
467 <<
468 dl_se->dl_non_contending = 1; <<
469 if (!dl_server(dl_se)) <<
470 get_task_struct(dl_task_of(dl_ <<
471 <<
472 hrtimer_start(timer, ns_to_ktime(zerol <<
473 } <<
474 <<
475 static void task_contending(struct sched_dl_en <<
476 { <<
477 struct dl_rq *dl_rq = dl_rq_of_se(dl_s <<
478 <<
479 /* <<
480 * If this is a non-deadline task that <<
481 * do nothing <<
482 */ <<
483 if (dl_se->dl_runtime == 0) <<
484 return; <<
485 <<
486 if (flags & ENQUEUE_MIGRATED) <<
487 add_rq_bw(dl_se, dl_rq); <<
488 49
489 if (dl_se->dl_non_contending) { !! 50 return dl_rq->rb_leftmost == &dl_se->rb_node;
490 dl_se->dl_non_contending = 0; <<
491 /* <<
492 * If the timer handler is cur <<
493 * timer cannot be canceled, i <<
494 * will see that dl_not_conten <<
495 * will not touch the rq's act <<
496 * so we are still safe. <<
497 */ <<
498 if (hrtimer_try_to_cancel(&dl_ <<
499 if (!dl_server(dl_se)) <<
500 put_task_struc <<
501 } <<
502 } else { <<
503 /* <<
504 * Since "dl_non_contending" i <<
505 * task's utilization has alre <<
506 * active utilization (either <<
507 * when the "inactive timer" f <<
508 * So, add it back. <<
509 */ <<
510 add_running_bw(dl_se, dl_rq); <<
511 } <<
512 } 51 }
513 52
514 static inline int is_leftmost(struct sched_dl_ !! 53 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
515 { 54 {
516 return rb_first_cached(&dl_rq->root) = !! 55 raw_spin_lock_init(&dl_b->dl_runtime_lock);
>> 56 dl_b->dl_period = period;
>> 57 dl_b->dl_runtime = runtime;
517 } 58 }
518 59
519 static void init_dl_rq_bw_ratio(struct dl_rq * <<
520 <<
521 void init_dl_bw(struct dl_bw *dl_b) 60 void init_dl_bw(struct dl_bw *dl_b)
522 { 61 {
523 raw_spin_lock_init(&dl_b->lock); 62 raw_spin_lock_init(&dl_b->lock);
>> 63 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
524 if (global_rt_runtime() == RUNTIME_INF 64 if (global_rt_runtime() == RUNTIME_INF)
525 dl_b->bw = -1; 65 dl_b->bw = -1;
526 else 66 else
527 dl_b->bw = to_ratio(global_rt_ 67 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
>> 68 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
528 dl_b->total_bw = 0; 69 dl_b->total_bw = 0;
529 } 70 }
530 71
531 void init_dl_rq(struct dl_rq *dl_rq) 72 void init_dl_rq(struct dl_rq *dl_rq)
532 { 73 {
533 dl_rq->root = RB_ROOT_CACHED; !! 74 dl_rq->rb_root = RB_ROOT;
534 75
535 #ifdef CONFIG_SMP 76 #ifdef CONFIG_SMP
536 /* zero means no -deadline tasks */ 77 /* zero means no -deadline tasks */
537 dl_rq->earliest_dl.curr = dl_rq->earli 78 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
538 79
>> 80 dl_rq->dl_nr_migratory = 0;
539 dl_rq->overloaded = 0; 81 dl_rq->overloaded = 0;
540 dl_rq->pushable_dl_tasks_root = RB_ROO !! 82 dl_rq->pushable_dl_tasks_root = RB_ROOT;
541 #else 83 #else
542 init_dl_bw(&dl_rq->dl_bw); 84 init_dl_bw(&dl_rq->dl_bw);
543 #endif 85 #endif
544 <<
545 dl_rq->running_bw = 0; <<
546 dl_rq->this_bw = 0; <<
547 init_dl_rq_bw_ratio(dl_rq); <<
548 } 86 }
549 87
550 #ifdef CONFIG_SMP 88 #ifdef CONFIG_SMP
551 89
552 static inline int dl_overloaded(struct rq *rq) 90 static inline int dl_overloaded(struct rq *rq)
553 { 91 {
554 return atomic_read(&rq->rd->dlo_count) 92 return atomic_read(&rq->rd->dlo_count);
555 } 93 }
556 94
557 static inline void dl_set_overload(struct rq * 95 static inline void dl_set_overload(struct rq *rq)
558 { 96 {
559 if (!rq->online) 97 if (!rq->online)
560 return; 98 return;
561 99
562 cpumask_set_cpu(rq->cpu, rq->rd->dlo_m 100 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
563 /* 101 /*
564 * Must be visible before the overload 102 * Must be visible before the overload count is
565 * set (as in sched_rt.c). 103 * set (as in sched_rt.c).
566 * 104 *
567 * Matched by the barrier in pull_dl_t 105 * Matched by the barrier in pull_dl_task().
568 */ 106 */
569 smp_wmb(); 107 smp_wmb();
570 atomic_inc(&rq->rd->dlo_count); 108 atomic_inc(&rq->rd->dlo_count);
571 } 109 }
572 110
573 static inline void dl_clear_overload(struct rq 111 static inline void dl_clear_overload(struct rq *rq)
574 { 112 {
575 if (!rq->online) 113 if (!rq->online)
576 return; 114 return;
577 115
578 atomic_dec(&rq->rd->dlo_count); 116 atomic_dec(&rq->rd->dlo_count);
579 cpumask_clear_cpu(rq->cpu, rq->rd->dlo 117 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
580 } 118 }
581 119
582 #define __node_2_pdl(node) \ !! 120 static void update_dl_migration(struct dl_rq *dl_rq)
583 rb_entry((node), struct task_struct, p !! 121 {
>> 122 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
>> 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 }
584 132
585 static inline bool __pushable_less(struct rb_n !! 133 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
586 { 134 {
587 return dl_entity_preempt(&__node_2_pdl !! 135 struct task_struct *p = dl_task_of(dl_se);
>> 136
>> 137 if (p->nr_cpus_allowed > 1)
>> 138 dl_rq->dl_nr_migratory++;
>> 139
>> 140 update_dl_migration(dl_rq);
588 } 141 }
589 142
590 static inline int has_pushable_dl_tasks(struct !! 143 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
591 { 144 {
592 return !RB_EMPTY_ROOT(&rq->dl.pushable !! 145 struct task_struct *p = dl_task_of(dl_se);
>> 146
>> 147 if (p->nr_cpus_allowed > 1)
>> 148 dl_rq->dl_nr_migratory--;
>> 149
>> 150 update_dl_migration(dl_rq);
593 } 151 }
594 152
595 /* 153 /*
596 * The list of pushable -deadline task is not 154 * The list of pushable -deadline task is not a plist, like in
597 * sched_rt.c, it is an rb-tree with tasks ord 155 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
598 */ 156 */
599 static void enqueue_pushable_dl_task(struct rq 157 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
600 { 158 {
601 struct rb_node *leftmost; !! 159 struct dl_rq *dl_rq = &rq->dl;
602 !! 160 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
603 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushab !! 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 }
604 178
605 leftmost = rb_add_cached(&p->pushable_ <<
606 &rq->dl.pusha <<
607 __pushable_le <<
608 if (leftmost) 179 if (leftmost)
609 rq->dl.earliest_dl.next = p->d !! 180 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
610 181
611 if (!rq->dl.overloaded) { !! 182 rb_link_node(&p->pushable_dl_tasks, parent, link);
612 dl_set_overload(rq); !! 183 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
613 rq->dl.overloaded = 1; <<
614 } <<
615 } 184 }
616 185
617 static void dequeue_pushable_dl_task(struct rq 186 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
618 { 187 {
619 struct dl_rq *dl_rq = &rq->dl; 188 struct dl_rq *dl_rq = &rq->dl;
620 struct rb_root_cached *root = &dl_rq-> <<
621 struct rb_node *leftmost; <<
622 189
623 if (RB_EMPTY_NODE(&p->pushable_dl_task 190 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
624 return; 191 return;
625 192
626 leftmost = rb_erase_cached(&p->pushabl !! 193 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
627 if (leftmost) !! 194 struct rb_node *next_node;
628 dl_rq->earliest_dl.next = __no !! 195
>> 196 next_node = rb_next(&p->pushable_dl_tasks);
>> 197 dl_rq->pushable_dl_tasks_leftmost = next_node;
>> 198 }
629 199
>> 200 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
630 RB_CLEAR_NODE(&p->pushable_dl_tasks); 201 RB_CLEAR_NODE(&p->pushable_dl_tasks);
>> 202 }
631 203
632 if (!has_pushable_dl_tasks(rq) && rq-> !! 204 static inline int has_pushable_dl_tasks(struct rq *rq)
633 dl_clear_overload(rq); !! 205 {
634 rq->dl.overloaded = 0; !! 206 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
635 } <<
636 } 207 }
637 208
638 static int push_dl_task(struct rq *rq); 209 static int push_dl_task(struct rq *rq);
639 210
640 static inline bool need_pull_dl_task(struct rq 211 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
641 { 212 {
642 return rq->online && dl_task(prev); !! 213 return dl_task(prev);
643 } 214 }
644 215
645 static DEFINE_PER_CPU(struct balance_callback, !! 216 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
646 static DEFINE_PER_CPU(struct balance_callback, !! 217 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
647 218
648 static void push_dl_tasks(struct rq *); 219 static void push_dl_tasks(struct rq *);
649 static void pull_dl_task(struct rq *); 220 static void pull_dl_task(struct rq *);
650 221
651 static inline void deadline_queue_push_tasks(s !! 222 static inline void queue_push_tasks(struct rq *rq)
652 { 223 {
653 if (!has_pushable_dl_tasks(rq)) 224 if (!has_pushable_dl_tasks(rq))
654 return; 225 return;
655 226
656 queue_balance_callback(rq, &per_cpu(dl 227 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
657 } 228 }
658 229
659 static inline void deadline_queue_pull_task(st !! 230 static inline void queue_pull_task(struct rq *rq)
660 { 231 {
661 queue_balance_callback(rq, &per_cpu(dl 232 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
662 } 233 }
663 234
664 static struct rq *find_lock_later_rq(struct ta 235 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
665 236
666 static struct rq *dl_task_offline_migration(st 237 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
667 { 238 {
668 struct rq *later_rq = NULL; 239 struct rq *later_rq = NULL;
669 struct dl_bw *dl_b; !! 240 bool fallback = false;
670 241
671 later_rq = find_lock_later_rq(p, rq); 242 later_rq = find_lock_later_rq(p, rq);
>> 243
672 if (!later_rq) { 244 if (!later_rq) {
673 int cpu; 245 int cpu;
674 246
675 /* 247 /*
676 * If we cannot preempt any rq 248 * If we cannot preempt any rq, fall back to pick any
677 * online CPU: !! 249 * online cpu.
678 */ 250 */
679 cpu = cpumask_any_and(cpu_acti !! 251 fallback = true;
>> 252 cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p));
680 if (cpu >= nr_cpu_ids) { 253 if (cpu >= nr_cpu_ids) {
681 /* 254 /*
682 * Failed to find any !! 255 * Fail to find any suitable cpu.
683 * The task will never 256 * The task will never come back!
684 */ 257 */
685 WARN_ON_ONCE(dl_bandwi !! 258 BUG_ON(dl_bandwidth_enabled());
686 259
687 /* 260 /*
688 * If admission contro 261 * If admission control is disabled we
689 * try a little harder 262 * try a little harder to let the task
690 * run. 263 * run.
691 */ 264 */
692 cpu = cpumask_any(cpu_ 265 cpu = cpumask_any(cpu_active_mask);
693 } 266 }
694 later_rq = cpu_rq(cpu); 267 later_rq = cpu_rq(cpu);
695 double_lock_balance(rq, later_ 268 double_lock_balance(rq, later_rq);
696 } 269 }
697 270
698 if (p->dl.dl_non_contending || p->dl.d <<
699 /* <<
700 * Inactive timer is armed (or <<
701 * waiting for us to release r <<
702 * will fire (or continue), it <<
703 * task migrated to later_rq ( <<
704 */ <<
705 sub_running_bw(&p->dl, &rq->dl <<
706 sub_rq_bw(&p->dl, &rq->dl); <<
707 <<
708 add_rq_bw(&p->dl, &later_rq->d <<
709 add_running_bw(&p->dl, &later_ <<
710 } else { <<
711 sub_rq_bw(&p->dl, &rq->dl); <<
712 add_rq_bw(&p->dl, &later_rq->d <<
713 } <<
714 <<
715 /* 271 /*
716 * And we finally need to fix up root_ !! 272 * By now the task is replenished and enqueued; migrate it.
717 * since p is still hanging out in the <<
718 * domain. <<
719 */ 273 */
720 dl_b = &rq->rd->dl_bw; !! 274 deactivate_task(rq, p, 0);
721 raw_spin_lock(&dl_b->lock); !! 275 set_task_cpu(p, later_rq->cpu);
722 __dl_sub(dl_b, p->dl.dl_bw, cpumask_we !! 276 activate_task(later_rq, p, 0);
723 raw_spin_unlock(&dl_b->lock); <<
724 277
725 dl_b = &later_rq->rd->dl_bw; !! 278 if (!fallback)
726 raw_spin_lock(&dl_b->lock); !! 279 resched_curr(later_rq);
727 __dl_add(dl_b, p->dl.dl_bw, cpumask_we <<
728 raw_spin_unlock(&dl_b->lock); <<
729 280
730 set_task_cpu(p, later_rq->cpu); <<
731 double_unlock_balance(later_rq, rq); 281 double_unlock_balance(later_rq, rq);
732 282
733 return later_rq; 283 return later_rq;
734 } 284 }
735 285
736 #else 286 #else
737 287
738 static inline 288 static inline
739 void enqueue_pushable_dl_task(struct rq *rq, s 289 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
740 { 290 {
741 } 291 }
742 292
743 static inline 293 static inline
744 void dequeue_pushable_dl_task(struct rq *rq, s 294 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
745 { 295 {
746 } 296 }
747 297
748 static inline 298 static inline
749 void inc_dl_migration(struct sched_dl_entity * 299 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
750 { 300 {
751 } 301 }
752 302
753 static inline 303 static inline
754 void dec_dl_migration(struct sched_dl_entity * 304 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
755 { 305 {
756 } 306 }
757 307
758 static inline void deadline_queue_push_tasks(s !! 308 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
759 { 309 {
>> 310 return false;
760 } 311 }
761 312
762 static inline void deadline_queue_pull_task(st !! 313 static inline void pull_dl_task(struct rq *rq)
763 { 314 {
764 } 315 }
765 #endif /* CONFIG_SMP */ <<
766 <<
767 static void <<
768 enqueue_dl_entity(struct sched_dl_entity *dl_s <<
769 static void enqueue_task_dl(struct rq *rq, str <<
770 static void dequeue_dl_entity(struct sched_dl_ <<
771 static void wakeup_preempt_dl(struct rq *rq, s <<
772 316
773 static inline void replenish_dl_new_period(str !! 317 static inline void queue_push_tasks(struct rq *rq)
774 st <<
775 { 318 {
776 /* for non-boosted task, pi_of(dl_se) !! 319 }
777 dl_se->deadline = rq_clock(rq) + pi_of <<
778 dl_se->runtime = pi_of(dl_se)->dl_runt <<
779 320
780 /* !! 321 static inline void queue_pull_task(struct rq *rq)
781 * If it is a deferred reservation, an !! 322 {
782 * is not handling an starvation case, <<
783 */ <<
784 if (dl_se->dl_defer & !dl_se->dl_defer <<
785 dl_se->dl_throttled = 1; <<
786 dl_se->dl_defer_armed = 1; <<
787 } <<
788 } 323 }
>> 324 #endif /* CONFIG_SMP */
>> 325
>> 326 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
>> 327 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
>> 328 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
>> 329 int flags);
789 330
790 /* 331 /*
791 * We are being explicitly informed that a new 332 * We are being explicitly informed that a new instance is starting,
792 * and this means that: 333 * and this means that:
793 * - the absolute deadline of the entity has 334 * - the absolute deadline of the entity has to be placed at
794 * current time + relative deadline; 335 * current time + relative deadline;
795 * - the runtime of the entity has to be set 336 * - the runtime of the entity has to be set to the maximum value.
796 * 337 *
797 * The capability of specifying such event is 338 * The capability of specifying such event is useful whenever a -deadline
798 * entity wants to (try to!) synchronize its b 339 * entity wants to (try to!) synchronize its behaviour with the scheduler's
799 * one, and to (try to!) reconcile itself with 340 * one, and to (try to!) reconcile itself with its own scheduling
800 * parameters. 341 * parameters.
801 */ 342 */
802 static inline void setup_new_dl_entity(struct !! 343 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
>> 344 struct sched_dl_entity *pi_se)
803 { 345 {
804 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 346 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
805 struct rq *rq = rq_of_dl_rq(dl_rq); 347 struct rq *rq = rq_of_dl_rq(dl_rq);
806 348
807 WARN_ON(is_dl_boosted(dl_se)); !! 349 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
808 WARN_ON(dl_time_before(rq_clock(rq), d <<
809 <<
810 /* <<
811 * We are racing with the deadline tim <<
812 * the deadline timer handler will tak <<
813 * the runtime and postponing the dead <<
814 */ <<
815 if (dl_se->dl_throttled) <<
816 return; <<
817 350
818 /* 351 /*
819 * We use the regular wall clock time 352 * We use the regular wall clock time to set deadlines in the
820 * future; in fact, we must consider e 353 * future; in fact, we must consider execution overheads (time
821 * spent on hardirq context, etc.). 354 * spent on hardirq context, etc.).
822 */ 355 */
823 replenish_dl_new_period(dl_se, rq); !! 356 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
>> 357 dl_se->runtime = pi_se->dl_runtime;
>> 358 dl_se->dl_new = 0;
824 } 359 }
825 360
826 static int start_dl_timer(struct sched_dl_enti <<
827 static bool dl_entity_overflow(struct sched_dl <<
828 <<
829 /* 361 /*
830 * Pure Earliest Deadline First (EDF) scheduli 362 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
831 * possibility of a entity lasting more than w 363 * possibility of a entity lasting more than what it declared, and thus
832 * exhausting its runtime. 364 * exhausting its runtime.
833 * 365 *
834 * Here we are interested in making runtime ov 366 * Here we are interested in making runtime overrun possible, but we do
835 * not want a entity which is misbehaving to a 367 * not want a entity which is misbehaving to affect the scheduling of all
836 * other entities. 368 * other entities.
837 * Therefore, a budgeting strategy called Cons 369 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
838 * is used, in order to confine each entity wi 370 * is used, in order to confine each entity within its own bandwidth.
839 * 371 *
840 * This function deals exactly with that, and 372 * This function deals exactly with that, and ensures that when the runtime
841 * of a entity is replenished, its deadline is 373 * of a entity is replenished, its deadline is also postponed. That ensures
842 * the overrunning entity can't interfere with 374 * the overrunning entity can't interfere with other entity in the system and
843 * can't make them miss their deadlines. Reaso 375 * can't make them miss their deadlines. Reasons why this kind of overruns
844 * could happen are, typically, a entity volun 376 * could happen are, typically, a entity voluntarily trying to overcome its
845 * runtime, or it just underestimated it durin 377 * runtime, or it just underestimated it during sched_setattr().
846 */ 378 */
847 static void replenish_dl_entity(struct sched_d !! 379 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
>> 380 struct sched_dl_entity *pi_se)
848 { 381 {
849 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 382 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
850 struct rq *rq = rq_of_dl_rq(dl_rq); 383 struct rq *rq = rq_of_dl_rq(dl_rq);
851 384
852 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime !! 385 BUG_ON(pi_se->dl_runtime <= 0);
853 386
854 /* 387 /*
855 * This could be the case for a !-dl t 388 * This could be the case for a !-dl task that is boosted.
856 * Just go with full inherited paramet 389 * Just go with full inherited parameters.
857 * <<
858 * Or, it could be the case of a defer <<
859 * was not able to consume its runtime <<
860 * reached this point with current u > <<
861 * <<
862 * In both cases, set a new period. <<
863 */ 390 */
864 if (dl_se->dl_deadline == 0 || !! 391 if (dl_se->dl_deadline == 0) {
865 (dl_se->dl_defer_armed && dl_entit !! 392 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
866 dl_se->deadline = rq_clock(rq) !! 393 dl_se->runtime = pi_se->dl_runtime;
867 dl_se->runtime = pi_of(dl_se)- <<
868 } 394 }
869 395
870 if (dl_se->dl_yielded && dl_se->runtim <<
871 dl_se->runtime = 0; <<
872 <<
873 /* 396 /*
874 * We keep moving the deadline away un 397 * We keep moving the deadline away until we get some
875 * available runtime for the entity. T 398 * available runtime for the entity. This ensures correct
876 * handling of situations where the ru 399 * handling of situations where the runtime overrun is
877 * arbitrary large. 400 * arbitrary large.
878 */ 401 */
879 while (dl_se->runtime <= 0) { 402 while (dl_se->runtime <= 0) {
880 dl_se->deadline += pi_of(dl_se !! 403 dl_se->deadline += pi_se->dl_period;
881 dl_se->runtime += pi_of(dl_se) !! 404 dl_se->runtime += pi_se->dl_runtime;
882 } 405 }
883 406
884 /* 407 /*
885 * At this point, the deadline really 408 * At this point, the deadline really should be "in
886 * the future" with respect to rq->clo 409 * the future" with respect to rq->clock. If it's
887 * not, we are, for some reason, laggi 410 * not, we are, for some reason, lagging too much!
888 * Anyway, after having warn userspace 411 * Anyway, after having warn userspace abut that,
889 * we still try to keep the things run 412 * we still try to keep the things running by
890 * resetting the deadline and the budg 413 * resetting the deadline and the budget of the
891 * entity. 414 * entity.
892 */ 415 */
893 if (dl_time_before(dl_se->deadline, rq 416 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
894 printk_deferred_once("sched: D !! 417 printk_deferred_once("sched: DL replenish lagged to much\n");
895 replenish_dl_new_period(dl_se, !! 418 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
>> 419 dl_se->runtime = pi_se->dl_runtime;
896 } 420 }
897 421
898 if (dl_se->dl_yielded) 422 if (dl_se->dl_yielded)
899 dl_se->dl_yielded = 0; 423 dl_se->dl_yielded = 0;
900 if (dl_se->dl_throttled) 424 if (dl_se->dl_throttled)
901 dl_se->dl_throttled = 0; 425 dl_se->dl_throttled = 0;
902 <<
903 /* <<
904 * If this is the replenishment of a d <<
905 * clear the flag and return. <<
906 */ <<
907 if (dl_se->dl_defer_armed) { <<
908 dl_se->dl_defer_armed = 0; <<
909 return; <<
910 } <<
911 <<
912 /* <<
913 * A this point, if the deferred serve <<
914 * is in the future, if it is not runn <<
915 * and arm the defer timer. <<
916 */ <<
917 if (dl_se->dl_defer && !dl_se->dl_defe <<
918 dl_time_before(rq_clock(dl_se->rq) <<
919 if (!is_dl_boosted(dl_se) && d <<
920 <<
921 /* <<
922 * Set dl_se->dl_defer <<
923 * inform the start_dl <<
924 * activation. <<
925 */ <<
926 dl_se->dl_defer_armed <<
927 dl_se->dl_throttled = <<
928 if (!start_dl_timer(dl <<
929 /* <<
930 * If for what <<
931 * queued but <<
932 * deferrable <<
933 */ <<
934 hrtimer_try_to <<
935 dl_se->dl_defe <<
936 dl_se->dl_thro <<
937 } <<
938 } <<
939 } <<
940 } 426 }
941 427
942 /* 428 /*
943 * Here we check if --at time t-- an entity (w 429 * Here we check if --at time t-- an entity (which is probably being
944 * [re]activated or, in general, enqueued) can 430 * [re]activated or, in general, enqueued) can use its remaining runtime
945 * and its current deadline _without_ exceedin 431 * and its current deadline _without_ exceeding the bandwidth it is
946 * assigned (function returns true if it can't 432 * assigned (function returns true if it can't). We are in fact applying
947 * one of the CBS rules: when a task wakes up, 433 * one of the CBS rules: when a task wakes up, if the residual runtime
948 * over residual deadline fits within the allo 434 * over residual deadline fits within the allocated bandwidth, then we
949 * can keep the current (absolute) deadline an 435 * can keep the current (absolute) deadline and residual budget without
950 * disrupting the schedulability of the system 436 * disrupting the schedulability of the system. Otherwise, we should
951 * refill the runtime and set the deadline a p 437 * refill the runtime and set the deadline a period in the future,
952 * because keeping the current (absolute) dead 438 * because keeping the current (absolute) deadline of the task would
953 * result in breaking guarantees promised to o 439 * result in breaking guarantees promised to other tasks (refer to
954 * Documentation/scheduler/sched-deadline.rst !! 440 * Documentation/scheduler/sched-deadline.txt for more informations).
955 * 441 *
956 * This function returns true if: 442 * This function returns true if:
957 * 443 *
958 * runtime / (deadline - t) > dl_runtime / d 444 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
959 * 445 *
960 * IOW we can't recycle current parameters. 446 * IOW we can't recycle current parameters.
961 * 447 *
962 * Notice that the bandwidth check is done aga 448 * Notice that the bandwidth check is done against the deadline. For
963 * task with deadline equal to period this is 449 * task with deadline equal to period this is the same of using
964 * dl_period instead of dl_deadline in the equ 450 * dl_period instead of dl_deadline in the equation above.
965 */ 451 */
966 static bool dl_entity_overflow(struct sched_dl !! 452 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
>> 453 struct sched_dl_entity *pi_se, u64 t)
967 { 454 {
968 u64 left, right; 455 u64 left, right;
969 456
970 /* 457 /*
971 * left and right are the two sides of 458 * left and right are the two sides of the equation above,
972 * after a bit of shuffling to use mul 459 * after a bit of shuffling to use multiplications instead
973 * of divisions. 460 * of divisions.
974 * 461 *
975 * Note that none of the time values i 462 * Note that none of the time values involved in the two
976 * multiplications are absolute: dl_de 463 * multiplications are absolute: dl_deadline and dl_runtime
977 * are the relative deadline and the m 464 * are the relative deadline and the maximum runtime of each
978 * instance, runtime is the runtime le 465 * instance, runtime is the runtime left for the last instance
979 * and (deadline - t), since t is rq-> 466 * and (deadline - t), since t is rq->clock, is the time left
980 * to the (absolute) deadline. Even if 467 * to the (absolute) deadline. Even if overflowing the u64 type
981 * is very unlikely to occur in both c 468 * is very unlikely to occur in both cases, here we scale down
982 * as we want to avoid that risk at al 469 * as we want to avoid that risk at all. Scaling down by 10
983 * means that we reduce granularity to 470 * means that we reduce granularity to 1us. We are fine with it,
984 * since this is only a true/false che 471 * since this is only a true/false check and, anyway, thinking
985 * of anything below microseconds reso 472 * of anything below microseconds resolution is actually fiction
986 * (but still we want to give the user 473 * (but still we want to give the user that illusion >;).
987 */ 474 */
988 left = (pi_of(dl_se)->dl_deadline >> D !! 475 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
989 right = ((dl_se->deadline - t) >> DL_S 476 right = ((dl_se->deadline - t) >> DL_SCALE) *
990 (pi_of(dl_se)->dl_runtime >> D !! 477 (pi_se->dl_runtime >> DL_SCALE);
991 478
992 return dl_time_before(right, left); 479 return dl_time_before(right, left);
993 } 480 }
994 481
995 /* 482 /*
996 * Revised wakeup rule [1]: For self-suspendin 483 * Revised wakeup rule [1]: For self-suspending tasks, rather then
997 * re-initializing task's runtime and deadline 484 * re-initializing task's runtime and deadline, the revised wakeup
998 * rule adjusts the task's runtime to avoid th 485 * rule adjusts the task's runtime to avoid the task to overrun its
999 * density. 486 * density.
1000 * 487 *
1001 * Reasoning: a task may overrun the density 488 * Reasoning: a task may overrun the density if:
1002 * runtime / (deadline - t) > dl_runtime / 489 * runtime / (deadline - t) > dl_runtime / dl_deadline
1003 * 490 *
1004 * Therefore, runtime can be adjusted to: 491 * Therefore, runtime can be adjusted to:
1005 * runtime = (dl_runtime / dl_deadline) * 492 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
1006 * 493 *
1007 * In such way that runtime will be equal to 494 * In such way that runtime will be equal to the maximum density
1008 * the task can use without breaking any rule 495 * the task can use without breaking any rule.
1009 * 496 *
1010 * [1] Luca Abeni, Giuseppe Lipari, and Juri 497 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
1011 * bandwidth server revisited. SIGBED Rev. 11 498 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
1012 */ 499 */
1013 static void 500 static void
1014 update_dl_revised_wakeup(struct sched_dl_enti 501 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
1015 { 502 {
1016 u64 laxity = dl_se->deadline - rq_clo 503 u64 laxity = dl_se->deadline - rq_clock(rq);
1017 504
1018 /* 505 /*
1019 * If the task has deadline < period, 506 * If the task has deadline < period, and the deadline is in the past,
1020 * it should already be throttled bef 507 * it should already be throttled before this check.
1021 * 508 *
1022 * See update_dl_entity() comments fo 509 * See update_dl_entity() comments for further details.
1023 */ 510 */
1024 WARN_ON(dl_time_before(dl_se->deadlin 511 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
1025 512
1026 dl_se->runtime = (dl_se->dl_density * !! 513 dl_se->runtime = (dl_se->dl_density * laxity) >> 20;
1027 } 514 }
1028 515
1029 /* 516 /*
1030 * Regarding the deadline, a task with implic 517 * Regarding the deadline, a task with implicit deadline has a relative
1031 * deadline == relative period. A task with c 518 * deadline == relative period. A task with constrained deadline has a
1032 * relative deadline <= relative period. 519 * relative deadline <= relative period.
1033 * 520 *
1034 * We support constrained deadline tasks. How 521 * We support constrained deadline tasks. However, there are some restrictions
1035 * applied only for tasks which do not have a 522 * applied only for tasks which do not have an implicit deadline. See
1036 * update_dl_entity() to know more about such 523 * update_dl_entity() to know more about such restrictions.
1037 * 524 *
1038 * The dl_is_implicit() returns true if the t 525 * The dl_is_implicit() returns true if the task has an implicit deadline.
1039 */ 526 */
1040 static inline bool dl_is_implicit(struct sche 527 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1041 { 528 {
1042 return dl_se->dl_deadline == dl_se->d 529 return dl_se->dl_deadline == dl_se->dl_period;
1043 } 530 }
1044 531
1045 /* 532 /*
1046 * When a deadline entity is placed in the ru 533 * When a deadline entity is placed in the runqueue, its runtime and deadline
1047 * might need to be updated. This is done by 534 * might need to be updated. This is done by a CBS wake up rule. There are two
1048 * different rules: 1) the original CBS; and 535 * different rules: 1) the original CBS; and 2) the Revisited CBS.
1049 * 536 *
1050 * When the task is starting a new period, th 537 * When the task is starting a new period, the Original CBS is used. In this
1051 * case, the runtime is replenished and a new 538 * case, the runtime is replenished and a new absolute deadline is set.
1052 * 539 *
1053 * When a task is queued before the begin of 540 * When a task is queued before the begin of the next period, using the
1054 * remaining runtime and deadline could make 541 * remaining runtime and deadline could make the entity to overflow, see
1055 * dl_entity_overflow() to find more about ru 542 * dl_entity_overflow() to find more about runtime overflow. When such case
1056 * is detected, the runtime and deadline need 543 * is detected, the runtime and deadline need to be updated.
1057 * 544 *
1058 * If the task has an implicit deadline, i.e. 545 * If the task has an implicit deadline, i.e., deadline == period, the Original
1059 * CBS is applied. The runtime is replenished !! 546 * CBS is applied. the runtime is replenished and a new absolute deadline is
1060 * set, as in the previous cases. 547 * set, as in the previous cases.
1061 * 548 *
1062 * However, the Original CBS does not work pr 549 * However, the Original CBS does not work properly for tasks with
1063 * deadline < period, which are said to have 550 * deadline < period, which are said to have a constrained deadline. By
1064 * applying the Original CBS, a constrained d 551 * applying the Original CBS, a constrained deadline task would be able to run
1065 * runtime/deadline in a period. With deadlin 552 * runtime/deadline in a period. With deadline < period, the task would
1066 * overrun the runtime/period allowed bandwid 553 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1067 * 554 *
1068 * In order to prevent this misbehave, the Re 555 * In order to prevent this misbehave, the Revisited CBS is used for
1069 * constrained deadline tasks when a runtime 556 * constrained deadline tasks when a runtime overflow is detected. In the
1070 * Revisited CBS, rather than replenishing & 557 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1071 * the remaining runtime of the task is reduc 558 * the remaining runtime of the task is reduced to avoid runtime overflow.
1072 * Please refer to the comments update_dl_rev 559 * Please refer to the comments update_dl_revised_wakeup() function to find
1073 * more about the Revised CBS rule. 560 * more about the Revised CBS rule.
1074 */ 561 */
1075 static void update_dl_entity(struct sched_dl_ !! 562 static void update_dl_entity(struct sched_dl_entity *dl_se,
>> 563 struct sched_dl_entity *pi_se)
1076 { 564 {
1077 struct rq *rq = rq_of_dl_se(dl_se); !! 565 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
>> 566 struct rq *rq = rq_of_dl_rq(dl_rq);
>> 567
>> 568 /*
>> 569 * The arrival of a new instance needs special treatment, i.e.,
>> 570 * the actual scheduling parameters have to be "renewed".
>> 571 */
>> 572 if (dl_se->dl_new) {
>> 573 setup_new_dl_entity(dl_se, pi_se);
>> 574 return;
>> 575 }
1078 576
1079 if (dl_time_before(dl_se->deadline, r 577 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1080 dl_entity_overflow(dl_se, rq_cloc !! 578 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
1081 579
1082 if (unlikely(!dl_is_implicit( 580 if (unlikely(!dl_is_implicit(dl_se) &&
1083 !dl_time_before( 581 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 !is_dl_boosted(d !! 582 !dl_se->dl_boosted)){
1085 update_dl_revised_wak 583 update_dl_revised_wakeup(dl_se, rq);
1086 return; 584 return;
1087 } 585 }
1088 586
1089 replenish_dl_new_period(dl_se !! 587 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
1090 } else if (dl_server(dl_se) && dl_se- !! 588 dl_se->runtime = pi_se->dl_runtime;
1091 /* <<
1092 * The server can still use i <<
1093 * it left the dl_defer_runni <<
1094 */ <<
1095 if (!dl_se->dl_defer_running) <<
1096 dl_se->dl_defer_armed <<
1097 dl_se->dl_throttled = <<
1098 } <<
1099 } 589 }
1100 } 590 }
1101 591
1102 static inline u64 dl_next_period(struct sched 592 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1103 { 593 {
1104 return dl_se->deadline - dl_se->dl_de 594 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1105 } 595 }
1106 596
1107 /* 597 /*
1108 * If the entity depleted all its runtime, an 598 * If the entity depleted all its runtime, and if we want it to sleep
1109 * while waiting for some new execution time 599 * while waiting for some new execution time to become available, we
1110 * set the bandwidth replenishment timer to t 600 * set the bandwidth replenishment timer to the replenishment instant
1111 * and try to activate it. 601 * and try to activate it.
1112 * 602 *
1113 * Notice that it is important for the caller 603 * Notice that it is important for the caller to know if the timer
1114 * actually started or not (i.e., the repleni 604 * actually started or not (i.e., the replenishment instant is in
1115 * the future or in the past). 605 * the future or in the past).
1116 */ 606 */
1117 static int start_dl_timer(struct sched_dl_ent !! 607 static int start_dl_timer(struct task_struct *p)
1118 { 608 {
>> 609 struct sched_dl_entity *dl_se = &p->dl;
1119 struct hrtimer *timer = &dl_se->dl_ti 610 struct hrtimer *timer = &dl_se->dl_timer;
1120 struct dl_rq *dl_rq = dl_rq_of_se(dl_ !! 611 struct rq *rq = task_rq(p);
1121 struct rq *rq = rq_of_dl_rq(dl_rq); <<
1122 ktime_t now, act; 612 ktime_t now, act;
1123 s64 delta; 613 s64 delta;
1124 614
1125 lockdep_assert_rq_held(rq); !! 615 lockdep_assert_held(&rq->lock);
1126 616
1127 /* 617 /*
1128 * We want the timer to fire at the d 618 * We want the timer to fire at the deadline, but considering
1129 * that it is actually coming from rq 619 * that it is actually coming from rq->clock and not from
1130 * hrtimer's time base reading. 620 * hrtimer's time base reading.
1131 * !! 621 */
1132 * The deferred reservation will have !! 622 act = ns_to_ktime(dl_next_period(dl_se));
1133 * (deadline - runtime). At that poin <<
1134 * if the current deadline can be use <<
1135 * required to avoid add too much pre <<
1136 * (current u > U). <<
1137 */ <<
1138 if (dl_se->dl_defer_armed) { <<
1139 WARN_ON_ONCE(!dl_se->dl_throt <<
1140 act = ns_to_ktime(dl_se->dead <<
1141 } else { <<
1142 /* act = deadline - rel-deadl <<
1143 act = ns_to_ktime(dl_next_per <<
1144 } <<
1145 <<
1146 now = hrtimer_cb_get_time(timer); 623 now = hrtimer_cb_get_time(timer);
1147 delta = ktime_to_ns(now) - rq_clock(r 624 delta = ktime_to_ns(now) - rq_clock(rq);
1148 act = ktime_add_ns(act, delta); 625 act = ktime_add_ns(act, delta);
1149 626
1150 /* 627 /*
1151 * If the expiry time already passed, 628 * If the expiry time already passed, e.g., because the value
1152 * chosen as the deadline is too smal 629 * chosen as the deadline is too small, don't even try to
1153 * start the timer in the past! 630 * start the timer in the past!
1154 */ 631 */
1155 if (ktime_us_delta(act, now) < 0) 632 if (ktime_us_delta(act, now) < 0)
1156 return 0; 633 return 0;
1157 634
1158 /* 635 /*
1159 * !enqueued will guarantee another c 636 * !enqueued will guarantee another callback; even if one is already in
1160 * progress. This ensures a balanced 637 * progress. This ensures a balanced {get,put}_task_struct().
1161 * 638 *
1162 * The race against __run_timer() cle 639 * The race against __run_timer() clearing the enqueued state is
1163 * harmless because we're holding tas 640 * harmless because we're holding task_rq()->lock, therefore the timer
1164 * expiring after we've done the chec 641 * expiring after we've done the check will wait on its task_rq_lock()
1165 * and observe our state. 642 * and observe our state.
1166 */ 643 */
1167 if (!hrtimer_is_queued(timer)) { 644 if (!hrtimer_is_queued(timer)) {
1168 if (!dl_server(dl_se)) !! 645 get_task_struct(p);
1169 get_task_struct(dl_ta !! 646 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
1170 hrtimer_start(timer, act, HRT <<
1171 } 647 }
1172 648
1173 return 1; 649 return 1;
1174 } 650 }
1175 651
1176 static void __push_dl_task(struct rq *rq, str <<
1177 { <<
1178 #ifdef CONFIG_SMP <<
1179 /* <<
1180 * Queueing this task back might have <<
1181 * to kick someone away. <<
1182 */ <<
1183 if (has_pushable_dl_tasks(rq)) { <<
1184 /* <<
1185 * Nothing relies on rq->lock <<
1186 * rq->lock. <<
1187 */ <<
1188 rq_unpin_lock(rq, rf); <<
1189 push_dl_task(rq); <<
1190 rq_repin_lock(rq, rf); <<
1191 } <<
1192 #endif <<
1193 } <<
1194 <<
1195 /* a defer timer will not be reset if the run <<
1196 static const u64 dl_server_min_res = 1 * NSEC <<
1197 <<
1198 static enum hrtimer_restart dl_server_timer(s <<
1199 { <<
1200 struct rq *rq = rq_of_dl_se(dl_se); <<
1201 u64 fw; <<
1202 <<
1203 scoped_guard (rq_lock, rq) { <<
1204 struct rq_flags *rf = &scope. <<
1205 <<
1206 if (!dl_se->dl_throttled || ! <<
1207 return HRTIMER_NOREST <<
1208 <<
1209 sched_clock_tick(); <<
1210 update_rq_clock(rq); <<
1211 <<
1212 if (!dl_se->dl_runtime) <<
1213 return HRTIMER_NOREST <<
1214 <<
1215 if (!dl_se->server_has_tasks( <<
1216 replenish_dl_entity(d <<
1217 return HRTIMER_NOREST <<
1218 } <<
1219 <<
1220 if (dl_se->dl_defer_armed) { <<
1221 /* <<
1222 * First check if the <<
1223 * If so, it is possi <<
1224 * of time. The dl_se <<
1225 * forwarding the tim <<
1226 */ <<
1227 if (dl_time_before(rq <<
1228 (d <<
1229 <<
1230 /* reset the <<
1231 fw = dl_se->d <<
1232 <<
1233 hrtimer_forwa <<
1234 return HRTIME <<
1235 } <<
1236 <<
1237 dl_se->dl_defer_runni <<
1238 } <<
1239 <<
1240 enqueue_dl_entity(dl_se, ENQU <<
1241 <<
1242 if (!dl_task(dl_se->rq->curr) <<
1243 resched_curr(rq); <<
1244 <<
1245 __push_dl_task(rq, rf); <<
1246 } <<
1247 <<
1248 return HRTIMER_NORESTART; <<
1249 } <<
1250 <<
1251 /* 652 /*
1252 * This is the bandwidth enforcement timer ca 653 * This is the bandwidth enforcement timer callback. If here, we know
1253 * a task is not on its dl_rq, since the fact 654 * a task is not on its dl_rq, since the fact that the timer was running
1254 * means the task is throttled and needs a ru 655 * means the task is throttled and needs a runtime replenishment.
1255 * 656 *
1256 * However, what we actually do depends on th 657 * However, what we actually do depends on the fact the task is active,
1257 * (it is on its rq) or has been removed from 658 * (it is on its rq) or has been removed from there by a call to
1258 * dequeue_task_dl(). In the former case we m 659 * dequeue_task_dl(). In the former case we must issue the runtime
1259 * replenishment and add the task back to the 660 * replenishment and add the task back to the dl_rq; in the latter, we just
1260 * do nothing but clearing dl_throttled, so t 661 * do nothing but clearing dl_throttled, so that runtime and deadline
1261 * updating (and the queueing back to dl_rq) 662 * updating (and the queueing back to dl_rq) will be done by the
1262 * next call to enqueue_task_dl(). 663 * next call to enqueue_task_dl().
1263 */ 664 */
1264 static enum hrtimer_restart dl_task_timer(str 665 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1265 { 666 {
1266 struct sched_dl_entity *dl_se = conta 667 struct sched_dl_entity *dl_se = container_of(timer,
1267 668 struct sched_dl_entity,
1268 669 dl_timer);
1269 struct task_struct *p; !! 670 struct task_struct *p = dl_task_of(dl_se);
1270 struct rq_flags rf; !! 671 unsigned long flags;
1271 struct rq *rq; 672 struct rq *rq;
1272 673
1273 if (dl_server(dl_se)) !! 674 rq = task_rq_lock(p, &flags);
1274 return dl_server_timer(timer, <<
1275 <<
1276 p = dl_task_of(dl_se); <<
1277 rq = task_rq_lock(p, &rf); <<
1278 675
1279 /* 676 /*
1280 * The task might have changed its sc 677 * The task might have changed its scheduling policy to something
1281 * different than SCHED_DEADLINE (thr !! 678 * different than SCHED_DEADLINE (through switched_fromd_dl()).
1282 */ 679 */
1283 if (!dl_task(p)) !! 680 if (!dl_task(p)) {
>> 681 __dl_clear_params(p);
>> 682 goto unlock;
>> 683 }
>> 684
>> 685 /*
>> 686 * This is possible if switched_from_dl() raced against a running
>> 687 * callback that took the above !dl_task() path and we've since then
>> 688 * switched back into SCHED_DEADLINE.
>> 689 *
>> 690 * There's nothing to do except drop our task reference.
>> 691 */
>> 692 if (dl_se->dl_new)
1284 goto unlock; 693 goto unlock;
1285 694
1286 /* 695 /*
1287 * The task might have been boosted b 696 * The task might have been boosted by someone else and might be in the
1288 * boosting/deboosting path, its not 697 * boosting/deboosting path, its not throttled.
1289 */ 698 */
1290 if (is_dl_boosted(dl_se)) !! 699 if (dl_se->dl_boosted)
1291 goto unlock; 700 goto unlock;
1292 701
1293 /* 702 /*
1294 * Spurious timer due to start_dl_tim 703 * Spurious timer due to start_dl_timer() race; or we already received
1295 * a replenishment from rt_mutex_setp 704 * a replenishment from rt_mutex_setprio().
1296 */ 705 */
1297 if (!dl_se->dl_throttled) 706 if (!dl_se->dl_throttled)
1298 goto unlock; 707 goto unlock;
1299 708
1300 sched_clock_tick(); 709 sched_clock_tick();
1301 update_rq_clock(rq); 710 update_rq_clock(rq);
1302 711
1303 /* 712 /*
1304 * If the throttle happened during sc 713 * If the throttle happened during sched-out; like:
1305 * 714 *
1306 * schedule() 715 * schedule()
1307 * deactivate_task() 716 * deactivate_task()
1308 * dequeue_task_dl() 717 * dequeue_task_dl()
1309 * update_curr_dl() 718 * update_curr_dl()
1310 * start_dl_timer() 719 * start_dl_timer()
1311 * __dequeue_task_dl() 720 * __dequeue_task_dl()
1312 * prev->on_rq = 0; 721 * prev->on_rq = 0;
1313 * 722 *
1314 * We can be both throttled and !queu 723 * We can be both throttled and !queued. Replenish the counter
1315 * but do not enqueue -- wait for our 724 * but do not enqueue -- wait for our wakeup to do that.
1316 */ 725 */
1317 if (!task_on_rq_queued(p)) { 726 if (!task_on_rq_queued(p)) {
1318 replenish_dl_entity(dl_se); !! 727 replenish_dl_entity(dl_se, dl_se);
1319 goto unlock; 728 goto unlock;
1320 } 729 }
1321 730
>> 731 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
>> 732 if (dl_task(rq->curr))
>> 733 check_preempt_curr_dl(rq, p, 0);
>> 734 else
>> 735 resched_curr(rq);
>> 736
1322 #ifdef CONFIG_SMP 737 #ifdef CONFIG_SMP
1323 if (unlikely(!rq->online)) { !! 738 /*
1324 /* !! 739 * Perform balancing operations here; after the replenishments. We
1325 * If the runqueue is no long !! 740 * cannot drop rq->lock before this, otherwise the assertion in
1326 * task elsewhere. This neces !! 741 * start_dl_timer() about not missing updates is not true.
1327 */ !! 742 *
1328 lockdep_unpin_lock(__rq_lockp !! 743 * If we find that the rq the task was on is no longer available, we
>> 744 * need to select a new rq.
>> 745 *
>> 746 * XXX figure out if select_task_rq_dl() deals with offline cpus.
>> 747 */
>> 748 if (unlikely(!rq->online))
1329 rq = dl_task_offline_migratio 749 rq = dl_task_offline_migration(rq, p);
1330 rf.cookie = lockdep_pin_lock( <<
1331 update_rq_clock(rq); <<
1332 750
>> 751 /*
>> 752 * Queueing this task back might have overloaded rq, check if we need
>> 753 * to kick someone away.
>> 754 */
>> 755 if (has_pushable_dl_tasks(rq)) {
1333 /* 756 /*
1334 * Now that the task has been !! 757 * Nothing relies on rq->lock after this, so its safe to drop
1335 * have that locked, proceed !! 758 * rq->lock.
1336 * there. <<
1337 */ 759 */
>> 760 lockdep_unpin_lock(&rq->lock);
>> 761 push_dl_task(rq);
>> 762 lockdep_pin_lock(&rq->lock);
1338 } 763 }
1339 #endif 764 #endif
1340 765
1341 enqueue_task_dl(rq, p, ENQUEUE_REPLEN <<
1342 if (dl_task(rq->curr)) <<
1343 wakeup_preempt_dl(rq, p, 0); <<
1344 else <<
1345 resched_curr(rq); <<
1346 <<
1347 __push_dl_task(rq, &rf); <<
1348 <<
1349 unlock: 766 unlock:
1350 task_rq_unlock(rq, p, &rf); !! 767 task_rq_unlock(rq, p, &flags);
1351 768
1352 /* 769 /*
1353 * This can free the task_struct, inc 770 * This can free the task_struct, including this hrtimer, do not touch
1354 * anything related to that after thi 771 * anything related to that after this.
1355 */ 772 */
1356 put_task_struct(p); 773 put_task_struct(p);
1357 774
1358 return HRTIMER_NORESTART; 775 return HRTIMER_NORESTART;
1359 } 776 }
1360 777
1361 static void init_dl_task_timer(struct sched_d !! 778 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1362 { 779 {
1363 struct hrtimer *timer = &dl_se->dl_ti 780 struct hrtimer *timer = &dl_se->dl_timer;
1364 781
1365 hrtimer_init(timer, CLOCK_MONOTONIC, !! 782 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1366 timer->function = dl_task_timer; 783 timer->function = dl_task_timer;
1367 } 784 }
1368 785
1369 /* 786 /*
1370 * During the activation, CBS checks if it ca 787 * During the activation, CBS checks if it can reuse the current task's
1371 * runtime and period. If the deadline of the 788 * runtime and period. If the deadline of the task is in the past, CBS
1372 * cannot use the runtime, and so it replenis 789 * cannot use the runtime, and so it replenishes the task. This rule
1373 * works fine for implicit deadline tasks (de 790 * works fine for implicit deadline tasks (deadline == period), and the
1374 * CBS was designed for implicit deadline tas 791 * CBS was designed for implicit deadline tasks. However, a task with
1375 * constrained deadline (deadline < period) m !! 792 * constrained deadline (deadine < period) might be awakened after the
1376 * deadline, but before the next period. In t 793 * deadline, but before the next period. In this case, replenishing the
1377 * task would allow it to run for runtime / d 794 * task would allow it to run for runtime / deadline. As in this case
1378 * deadline < period, CBS enables a task to r 795 * deadline < period, CBS enables a task to run for more than the
1379 * runtime / period. In a very loaded system, 796 * runtime / period. In a very loaded system, this can cause a domino
1380 * effect, making other tasks miss their dead 797 * effect, making other tasks miss their deadlines.
1381 * 798 *
1382 * To avoid this problem, in the activation o 799 * To avoid this problem, in the activation of a constrained deadline
1383 * task after the deadline but before the nex 800 * task after the deadline but before the next period, throttle the
1384 * task and set the replenishing timer to the 801 * task and set the replenishing timer to the begin of the next period,
1385 * unless it is boosted. 802 * unless it is boosted.
1386 */ 803 */
1387 static inline void dl_check_constrained_dl(st 804 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1388 { 805 {
1389 struct rq *rq = rq_of_dl_se(dl_se); !! 806 struct task_struct *p = dl_task_of(dl_se);
>> 807 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1390 808
1391 if (dl_time_before(dl_se->deadline, r 809 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1392 dl_time_before(rq_clock(rq), dl_n 810 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1393 if (unlikely(is_dl_boosted(dl !! 811 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1394 return; 812 return;
1395 dl_se->dl_throttled = 1; 813 dl_se->dl_throttled = 1;
1396 if (dl_se->runtime > 0) 814 if (dl_se->runtime > 0)
1397 dl_se->runtime = 0; 815 dl_se->runtime = 0;
1398 } 816 }
1399 } 817 }
1400 818
1401 static 819 static
1402 int dl_runtime_exceeded(struct sched_dl_entit 820 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1403 { 821 {
1404 return (dl_se->runtime <= 0); 822 return (dl_se->runtime <= 0);
1405 } 823 }
1406 824
>> 825 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
>> 826
1407 /* 827 /*
1408 * This function implements the GRUB accounti !! 828 * Update the current task's runtime statistics (provided it is still
1409 * GRUB reclaiming algorithm, the runtime is !! 829 * a -deadline task and has not been removed from the dl_rq).
1410 * but as "dq = -(max{u, (Umax - Uinact - Uex <<
1411 * where u is the utilization of the task, Um <<
1412 * utilization, Uinact is the (per-runqueue) <<
1413 * as the difference between the "total runqu <<
1414 * "runqueue active utilization", and Uextra <<
1415 * reclaimable utilization. <<
1416 * Since rq->dl.running_bw and rq->dl.this_bw <<
1417 * by 2^BW_SHIFT, the result has to be shifte <<
1418 * Since rq->dl.bw_ratio contains 1 / Umax mu <<
1419 * is multiplied by rq->dl.bw_ratio and shift <<
1420 * Since delta is a 64 bit variable, to have <<
1421 * larger than 2^(64 - 20 - 8), which is more <<
1422 * not an issue here. <<
1423 */ 830 */
1424 static u64 grub_reclaim(u64 delta, struct rq !! 831 static void update_curr_dl(struct rq *rq)
1425 { <<
1426 u64 u_act; <<
1427 u64 u_inact = rq->dl.this_bw - rq->dl <<
1428 <<
1429 /* <<
1430 * Instead of computing max{u, (u_max <<
1431 * compare u_inact + u_extra with u_m <<
1432 * can be larger than u_max. So, u_ma <<
1433 * negative leading to wrong results. <<
1434 */ <<
1435 if (u_inact + rq->dl.extra_bw > rq->d <<
1436 u_act = dl_se->dl_bw; <<
1437 else <<
1438 u_act = rq->dl.max_bw - u_ina <<
1439 <<
1440 u_act = (u_act * rq->dl.bw_ratio) >> <<
1441 return (delta * u_act) >> BW_SHIFT; <<
1442 } <<
1443 <<
1444 s64 dl_scaled_delta_exec(struct rq *rq, struc <<
1445 { <<
1446 s64 scaled_delta_exec; <<
1447 <<
1448 /* <<
1449 * For tasks that participate in GRUB <<
1450 * spare reclaimed bandwidth is used <<
1451 * <<
1452 * For the others, we still need to s <<
1453 * according to current frequency and <<
1454 */ <<
1455 if (unlikely(dl_se->flags & SCHED_FLA <<
1456 scaled_delta_exec = grub_recl <<
1457 } else { <<
1458 int cpu = cpu_of(rq); <<
1459 unsigned long scale_freq = ar <<
1460 unsigned long scale_cpu = arc <<
1461 <<
1462 scaled_delta_exec = cap_scale <<
1463 scaled_delta_exec = cap_scale <<
1464 } <<
1465 <<
1466 return scaled_delta_exec; <<
1467 } <<
1468 <<
1469 static inline void <<
1470 update_stats_dequeue_dl(struct dl_rq *dl_rq, <<
1471 int flags); <<
1472 static void update_curr_dl_se(struct rq *rq, <<
1473 { 832 {
1474 s64 scaled_delta_exec; !! 833 struct task_struct *curr = rq->curr;
1475 !! 834 struct sched_dl_entity *dl_se = &curr->dl;
1476 if (unlikely(delta_exec <= 0)) { !! 835 u64 delta_exec;
1477 if (unlikely(dl_se->dl_yielde <<
1478 goto throttle; <<
1479 return; <<
1480 } <<
1481 <<
1482 if (dl_server(dl_se) && dl_se->dl_thr <<
1483 return; <<
1484 836
1485 if (dl_entity_is_special(dl_se)) !! 837 if (!dl_task(curr) || !on_dl_rq(dl_se))
1486 return; 838 return;
1487 839
1488 scaled_delta_exec = dl_scaled_delta_e <<
1489 <<
1490 dl_se->runtime -= scaled_delta_exec; <<
1491 <<
1492 /* 840 /*
1493 * The fair server can consume its ru !! 841 * Consumed budget is computed considering the time as
1494 * running as regular CFS). !! 842 * observed by schedulable tasks (excluding time spent
1495 * !! 843 * in hardirq context, etc.). Deadlines are instead
1496 * If the server consumes its entire !! 844 * computed using hard walltime. This seems to be the more
1497 * is not required for the current pe !! 845 * natural solution, but the full ramifications of this
1498 * starting a new period, pushing the !! 846 * approach need further study.
1499 */ 847 */
1500 if (dl_se->dl_defer && dl_se->dl_thro !! 848 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
1501 /* !! 849 if (unlikely((s64)delta_exec <= 0))
1502 * If the server was previous !! 850 return;
1503 * took place, it this point <<
1504 * was able to get runtime in <<
1505 * state. <<
1506 */ <<
1507 dl_se->dl_defer_running = 0; <<
1508 851
1509 hrtimer_try_to_cancel(&dl_se- !! 852 schedstat_set(curr->se.statistics.exec_max,
>> 853 max(curr->se.statistics.exec_max, delta_exec));
1510 854
1511 replenish_dl_new_period(dl_se !! 855 curr->se.sum_exec_runtime += delta_exec;
>> 856 account_group_exec_runtime(curr, delta_exec);
1512 857
1513 /* !! 858 curr->se.exec_start = rq_clock_task(rq);
1514 * Not being able to start th !! 859 cpuacct_charge(curr, delta_exec);
1515 * be started for whatever re <<
1516 * and queue right away. Othe <<
1517 * to what enqueue_dl_entity( <<
1518 */ <<
1519 WARN_ON_ONCE(!start_dl_timer( <<
1520 860
1521 return; !! 861 sched_rt_avg_update(rq, delta_exec);
1522 } <<
1523 862
1524 throttle: !! 863 dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec;
1525 if (dl_runtime_exceeded(dl_se) || dl_ !! 864 if (dl_runtime_exceeded(dl_se)) {
1526 dl_se->dl_throttled = 1; 865 dl_se->dl_throttled = 1;
>> 866 __dequeue_task_dl(rq, curr, 0);
>> 867 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
>> 868 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1527 869
1528 /* If requested, inform the u !! 870 if (!is_leftmost(curr, &rq->dl))
1529 if (dl_runtime_exceeded(dl_se <<
1530 (dl_se->flags & SCHED_FLA <<
1531 dl_se->dl_overrun = 1 <<
1532 <<
1533 dequeue_dl_entity(dl_se, 0); <<
1534 if (!dl_server(dl_se)) { <<
1535 update_stats_dequeue_ <<
1536 dequeue_pushable_dl_t <<
1537 } <<
1538 <<
1539 if (unlikely(is_dl_boosted(dl <<
1540 if (dl_server(dl_se)) <<
1541 enqueue_dl_en <<
1542 else <<
1543 enqueue_task_ <<
1544 } <<
1545 <<
1546 if (!is_leftmost(dl_se, &rq-> <<
1547 resched_curr(rq); 871 resched_curr(rq);
1548 } 872 }
1549 873
1550 /* 874 /*
1551 * The fair server (sole dl_server) d <<
1552 * workload because it is running fai <<
1553 */ <<
1554 if (dl_se == &rq->fair_server) <<
1555 return; <<
1556 <<
1557 #ifdef CONFIG_RT_GROUP_SCHED <<
1558 /* <<
1559 * Because -- for now -- we share the 875 * Because -- for now -- we share the rt bandwidth, we need to
1560 * account our runtime there too, oth 876 * account our runtime there too, otherwise actual rt tasks
1561 * would be able to exceed the shared 877 * would be able to exceed the shared quota.
1562 * 878 *
1563 * Account to the root rt group for n 879 * Account to the root rt group for now.
1564 * 880 *
1565 * The solution we're working towards 881 * The solution we're working towards is having the RT groups scheduled
1566 * using deadline servers -- however 882 * using deadline servers -- however there's a few nasties to figure
1567 * out before that can happen. 883 * out before that can happen.
1568 */ 884 */
1569 if (rt_bandwidth_enabled()) { 885 if (rt_bandwidth_enabled()) {
1570 struct rt_rq *rt_rq = &rq->rt 886 struct rt_rq *rt_rq = &rq->rt;
1571 887
1572 raw_spin_lock(&rt_rq->rt_runt 888 raw_spin_lock(&rt_rq->rt_runtime_lock);
1573 /* 889 /*
1574 * We'll let actual RT tasks 890 * We'll let actual RT tasks worry about the overflow here, we
1575 * have our own CBS to keep u 891 * have our own CBS to keep us inline; only account when RT
1576 * bandwidth is relevant. 892 * bandwidth is relevant.
1577 */ 893 */
1578 if (sched_rt_bandwidth_accoun 894 if (sched_rt_bandwidth_account(rt_rq))
1579 rt_rq->rt_time += del 895 rt_rq->rt_time += delta_exec;
1580 raw_spin_unlock(&rt_rq->rt_ru 896 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1581 } 897 }
1582 #endif <<
1583 } <<
1584 <<
1585 /* <<
1586 * In the non-defer mode, the idle time is no <<
1587 * server provides a guarantee. <<
1588 * <<
1589 * If the dl_server is in defer mode, the idl <<
1590 * as time available for the fair server, avo <<
1591 * rt scheduler that did not consumed that ti <<
1592 */ <<
1593 void dl_server_update_idle_time(struct rq *rq <<
1594 { <<
1595 s64 delta_exec, scaled_delta_exec; <<
1596 <<
1597 if (!rq->fair_server.dl_defer) <<
1598 return; <<
1599 <<
1600 /* no need to discount more */ <<
1601 if (rq->fair_server.runtime < 0) <<
1602 return; <<
1603 <<
1604 delta_exec = rq_clock_task(rq) - p->s <<
1605 if (delta_exec < 0) <<
1606 return; <<
1607 <<
1608 scaled_delta_exec = dl_scaled_delta_e <<
1609 <<
1610 rq->fair_server.runtime -= scaled_del <<
1611 <<
1612 if (rq->fair_server.runtime < 0) { <<
1613 rq->fair_server.dl_defer_runn <<
1614 rq->fair_server.runtime = 0; <<
1615 } <<
1616 <<
1617 p->se.exec_start = rq_clock_task(rq); <<
1618 } 898 }
1619 899
1620 void dl_server_update(struct sched_dl_entity !! 900 #ifdef CONFIG_SMP
1621 { <<
1622 /* 0 runtime = fair server disabled * <<
1623 if (dl_se->dl_runtime) <<
1624 update_curr_dl_se(dl_se->rq, <<
1625 } <<
1626 <<
1627 void dl_server_start(struct sched_dl_entity * <<
1628 { <<
1629 struct rq *rq = dl_se->rq; <<
1630 <<
1631 /* <<
1632 * XXX: the apply do not work fine at <<
1633 * fair server because things are not <<
1634 * this before getting generic. <<
1635 */ <<
1636 if (!dl_server(dl_se)) { <<
1637 u64 runtime = 50 * NSEC_PER_ <<
1638 u64 period = 1000 * NSEC_PER_ <<
1639 <<
1640 dl_server_apply_params(dl_se, <<
1641 <<
1642 dl_se->dl_server = 1; <<
1643 dl_se->dl_defer = 1; <<
1644 setup_new_dl_entity(dl_se); <<
1645 } <<
1646 <<
1647 if (!dl_se->dl_runtime) <<
1648 return; <<
1649 <<
1650 enqueue_dl_entity(dl_se, ENQUEUE_WAKE <<
1651 if (!dl_task(dl_se->rq->curr) || dl_e <<
1652 resched_curr(dl_se->rq); <<
1653 } <<
1654 <<
1655 void dl_server_stop(struct sched_dl_entity *d <<
1656 { <<
1657 if (!dl_se->dl_runtime) <<
1658 return; <<
1659 <<
1660 dequeue_dl_entity(dl_se, DEQUEUE_SLEE <<
1661 hrtimer_try_to_cancel(&dl_se->dl_time <<
1662 dl_se->dl_defer_armed = 0; <<
1663 dl_se->dl_throttled = 0; <<
1664 } <<
1665 <<
1666 void dl_server_init(struct sched_dl_entity *d <<
1667 dl_server_has_tasks_f has <<
1668 dl_server_pick_f pick_tas <<
1669 { <<
1670 dl_se->rq = rq; <<
1671 dl_se->server_has_tasks = has_tasks; <<
1672 dl_se->server_pick_task = pick_task; <<
1673 } <<
1674 <<
1675 void __dl_server_attach_root(struct sched_dl_ <<
1676 { <<
1677 u64 new_bw = dl_se->dl_bw; <<
1678 int cpu = cpu_of(rq); <<
1679 struct dl_bw *dl_b; <<
1680 <<
1681 dl_b = dl_bw_of(cpu_of(rq)); <<
1682 guard(raw_spinlock)(&dl_b->lock); <<
1683 <<
1684 if (!dl_bw_cpus(cpu)) <<
1685 return; <<
1686 <<
1687 __dl_add(dl_b, new_bw, dl_bw_cpus(cpu <<
1688 } <<
1689 <<
1690 int dl_server_apply_params(struct sched_dl_en <<
1691 { <<
1692 u64 old_bw = init ? 0 : to_ratio(dl_s <<
1693 u64 new_bw = to_ratio(period, runtime <<
1694 struct rq *rq = dl_se->rq; <<
1695 int cpu = cpu_of(rq); <<
1696 struct dl_bw *dl_b; <<
1697 unsigned long cap; <<
1698 int retval = 0; <<
1699 int cpus; <<
1700 <<
1701 dl_b = dl_bw_of(cpu); <<
1702 guard(raw_spinlock)(&dl_b->lock); <<
1703 <<
1704 cpus = dl_bw_cpus(cpu); <<
1705 cap = dl_bw_capacity(cpu); <<
1706 <<
1707 if (__dl_overflow(dl_b, cap, old_bw, <<
1708 return -EBUSY; <<
1709 <<
1710 if (init) { <<
1711 __add_rq_bw(new_bw, &rq->dl); <<
1712 __dl_add(dl_b, new_bw, cpus); <<
1713 } else { <<
1714 __dl_sub(dl_b, dl_se->dl_bw, <<
1715 __dl_add(dl_b, new_bw, cpus); <<
1716 <<
1717 dl_rq_change_utilization(rq, <<
1718 } <<
1719 <<
1720 dl_se->dl_runtime = runtime; <<
1721 dl_se->dl_deadline = period; <<
1722 dl_se->dl_period = period; <<
1723 <<
1724 dl_se->runtime = 0; <<
1725 dl_se->deadline = 0; <<
1726 <<
1727 dl_se->dl_bw = to_ratio(dl_se->dl_per <<
1728 dl_se->dl_density = to_ratio(dl_se->d <<
1729 <<
1730 return retval; <<
1731 } <<
1732 <<
1733 /* <<
1734 * Update the current task's runtime statisti <<
1735 * a -deadline task and has not been removed <<
1736 */ <<
1737 static void update_curr_dl(struct rq *rq) <<
1738 { <<
1739 struct task_struct *curr = rq->curr; <<
1740 struct sched_dl_entity *dl_se = &curr <<
1741 s64 delta_exec; <<
1742 <<
1743 if (!dl_task(curr) || !on_dl_rq(dl_se <<
1744 return; <<
1745 <<
1746 /* <<
1747 * Consumed budget is computed consid <<
1748 * observed by schedulable tasks (exc <<
1749 * in hardirq context, etc.). Deadlin <<
1750 * computed using hard walltime. This <<
1751 * natural solution, but the full ram <<
1752 * approach need further study. <<
1753 */ <<
1754 delta_exec = update_curr_common(rq); <<
1755 update_curr_dl_se(rq, dl_se, delta_ex <<
1756 } <<
1757 <<
1758 static enum hrtimer_restart inactive_task_tim <<
1759 { <<
1760 struct sched_dl_entity *dl_se = conta <<
1761 <<
1762 <<
1763 struct task_struct *p = NULL; <<
1764 struct rq_flags rf; <<
1765 struct rq *rq; <<
1766 <<
1767 if (!dl_server(dl_se)) { <<
1768 p = dl_task_of(dl_se); <<
1769 rq = task_rq_lock(p, &rf); <<
1770 } else { <<
1771 rq = dl_se->rq; <<
1772 rq_lock(rq, &rf); <<
1773 } <<
1774 <<
1775 sched_clock_tick(); <<
1776 update_rq_clock(rq); <<
1777 <<
1778 if (dl_server(dl_se)) <<
1779 goto no_task; <<
1780 <<
1781 if (!dl_task(p) || READ_ONCE(p->__sta <<
1782 struct dl_bw *dl_b = dl_bw_of <<
1783 <<
1784 if (READ_ONCE(p->__state) == <<
1785 sub_running_bw(&p->dl <<
1786 sub_rq_bw(&p->dl, dl_ <<
1787 dl_se->dl_non_contend <<
1788 } <<
1789 <<
1790 raw_spin_lock(&dl_b->lock); <<
1791 __dl_sub(dl_b, p->dl.dl_bw, d <<
1792 raw_spin_unlock(&dl_b->lock); <<
1793 __dl_clear_params(dl_se); <<
1794 <<
1795 goto unlock; <<
1796 } <<
1797 <<
1798 no_task: <<
1799 if (dl_se->dl_non_contending == 0) <<
1800 goto unlock; <<
1801 <<
1802 sub_running_bw(dl_se, &rq->dl); <<
1803 dl_se->dl_non_contending = 0; <<
1804 unlock: <<
1805 <<
1806 if (!dl_server(dl_se)) { <<
1807 task_rq_unlock(rq, p, &rf); <<
1808 put_task_struct(p); <<
1809 } else { <<
1810 rq_unlock(rq, &rf); <<
1811 } <<
1812 901
1813 return HRTIMER_NORESTART; !! 902 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
1814 } <<
1815 903
1816 static void init_dl_inactive_task_timer(struc !! 904 static inline u64 next_deadline(struct rq *rq)
1817 { 905 {
1818 struct hrtimer *timer = &dl_se->inact !! 906 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
1819 907
1820 hrtimer_init(timer, CLOCK_MONOTONIC, !! 908 if (next && dl_prio(next->prio))
1821 timer->function = inactive_task_timer !! 909 return next->dl.deadline;
>> 910 else
>> 911 return 0;
1822 } 912 }
1823 913
1824 #define __node_2_dle(node) \ <<
1825 rb_entry((node), struct sched_dl_enti <<
1826 <<
1827 #ifdef CONFIG_SMP <<
1828 <<
1829 static void inc_dl_deadline(struct dl_rq *dl_ 914 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1830 { 915 {
1831 struct rq *rq = rq_of_dl_rq(dl_rq); 916 struct rq *rq = rq_of_dl_rq(dl_rq);
1832 917
1833 if (dl_rq->earliest_dl.curr == 0 || 918 if (dl_rq->earliest_dl.curr == 0 ||
1834 dl_time_before(deadline, dl_rq->e 919 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1835 if (dl_rq->earliest_dl.curr = !! 920 /*
1836 cpupri_set(&rq->rd->c !! 921 * If the dl_rq had no -deadline tasks, or if the new task
>> 922 * has shorter deadline than the current one on dl_rq, we
>> 923 * know that the previous earliest becomes our next earliest,
>> 924 * as the new task becomes the earliest itself.
>> 925 */
>> 926 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
1837 dl_rq->earliest_dl.curr = dea 927 dl_rq->earliest_dl.curr = deadline;
1838 cpudl_set(&rq->rd->cpudl, rq- !! 928 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
>> 929 } else if (dl_rq->earliest_dl.next == 0 ||
>> 930 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
>> 931 /*
>> 932 * On the other hand, if the new -deadline task has a
>> 933 * a later deadline than the earliest one on dl_rq, but
>> 934 * it is earlier than the next (if any), we must
>> 935 * recompute the next-earliest.
>> 936 */
>> 937 dl_rq->earliest_dl.next = next_deadline(rq);
1839 } 938 }
1840 } 939 }
1841 940
1842 static void dec_dl_deadline(struct dl_rq *dl_ 941 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1843 { 942 {
1844 struct rq *rq = rq_of_dl_rq(dl_rq); 943 struct rq *rq = rq_of_dl_rq(dl_rq);
1845 944
1846 /* 945 /*
1847 * Since we may have removed our earl 946 * Since we may have removed our earliest (and/or next earliest)
1848 * task we must recompute them. 947 * task we must recompute them.
1849 */ 948 */
1850 if (!dl_rq->dl_nr_running) { 949 if (!dl_rq->dl_nr_running) {
1851 dl_rq->earliest_dl.curr = 0; 950 dl_rq->earliest_dl.curr = 0;
1852 dl_rq->earliest_dl.next = 0; 951 dl_rq->earliest_dl.next = 0;
1853 cpudl_clear(&rq->rd->cpudl, r !! 952 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1854 cpupri_set(&rq->rd->cpupri, r <<
1855 } else { 953 } else {
1856 struct rb_node *leftmost = rb !! 954 struct rb_node *leftmost = dl_rq->rb_leftmost;
1857 struct sched_dl_entity *entry !! 955 struct sched_dl_entity *entry;
1858 956
>> 957 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1859 dl_rq->earliest_dl.curr = ent 958 dl_rq->earliest_dl.curr = entry->deadline;
1860 cpudl_set(&rq->rd->cpudl, rq- !! 959 dl_rq->earliest_dl.next = next_deadline(rq);
>> 960 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
1861 } 961 }
1862 } 962 }
1863 963
1864 #else 964 #else
1865 965
1866 static inline void inc_dl_deadline(struct dl_ 966 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1867 static inline void dec_dl_deadline(struct dl_ 967 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1868 968
1869 #endif /* CONFIG_SMP */ 969 #endif /* CONFIG_SMP */
1870 970
1871 static inline 971 static inline
1872 void inc_dl_tasks(struct sched_dl_entity *dl_ 972 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1873 { 973 {
>> 974 int prio = dl_task_of(dl_se)->prio;
1874 u64 deadline = dl_se->deadline; 975 u64 deadline = dl_se->deadline;
1875 976
>> 977 WARN_ON(!dl_prio(prio));
1876 dl_rq->dl_nr_running++; 978 dl_rq->dl_nr_running++;
1877 add_nr_running(rq_of_dl_rq(dl_rq), 1) 979 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1878 980
1879 inc_dl_deadline(dl_rq, deadline); 981 inc_dl_deadline(dl_rq, deadline);
>> 982 inc_dl_migration(dl_se, dl_rq);
1880 } 983 }
1881 984
1882 static inline 985 static inline
1883 void dec_dl_tasks(struct sched_dl_entity *dl_ 986 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1884 { 987 {
>> 988 int prio = dl_task_of(dl_se)->prio;
>> 989
>> 990 WARN_ON(!dl_prio(prio));
1885 WARN_ON(!dl_rq->dl_nr_running); 991 WARN_ON(!dl_rq->dl_nr_running);
1886 dl_rq->dl_nr_running--; 992 dl_rq->dl_nr_running--;
1887 sub_nr_running(rq_of_dl_rq(dl_rq), 1) 993 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1888 994
1889 dec_dl_deadline(dl_rq, dl_se->deadlin 995 dec_dl_deadline(dl_rq, dl_se->deadline);
1890 } !! 996 dec_dl_migration(dl_se, dl_rq);
1891 <<
1892 static inline bool __dl_less(struct rb_node * <<
1893 { <<
1894 return dl_time_before(__node_2_dle(a) <<
1895 } <<
1896 <<
1897 static __always_inline struct sched_statistic <<
1898 __schedstats_from_dl_se(struct sched_dl_entit <<
1899 { <<
1900 if (!schedstat_enabled()) <<
1901 return NULL; <<
1902 <<
1903 if (dl_server(dl_se)) <<
1904 return NULL; <<
1905 <<
1906 return &dl_task_of(dl_se)->stats; <<
1907 } <<
1908 <<
1909 static inline void <<
1910 update_stats_wait_start_dl(struct dl_rq *dl_r <<
1911 { <<
1912 struct sched_statistics *stats = __sc <<
1913 if (stats) <<
1914 __update_stats_wait_start(rq_ <<
1915 } <<
1916 <<
1917 static inline void <<
1918 update_stats_wait_end_dl(struct dl_rq *dl_rq, <<
1919 { <<
1920 struct sched_statistics *stats = __sc <<
1921 if (stats) <<
1922 __update_stats_wait_end(rq_of <<
1923 } <<
1924 <<
1925 static inline void <<
1926 update_stats_enqueue_sleeper_dl(struct dl_rq <<
1927 { <<
1928 struct sched_statistics *stats = __sc <<
1929 if (stats) <<
1930 __update_stats_enqueue_sleepe <<
1931 } <<
1932 <<
1933 static inline void <<
1934 update_stats_enqueue_dl(struct dl_rq *dl_rq, <<
1935 int flags) <<
1936 { <<
1937 if (!schedstat_enabled()) <<
1938 return; <<
1939 <<
1940 if (flags & ENQUEUE_WAKEUP) <<
1941 update_stats_enqueue_sleeper_ <<
1942 } <<
1943 <<
1944 static inline void <<
1945 update_stats_dequeue_dl(struct dl_rq *dl_rq, <<
1946 int flags) <<
1947 { <<
1948 struct task_struct *p = dl_task_of(dl <<
1949 <<
1950 if (!schedstat_enabled()) <<
1951 return; <<
1952 <<
1953 if ((flags & DEQUEUE_SLEEP)) { <<
1954 unsigned int state; <<
1955 <<
1956 state = READ_ONCE(p->__state) <<
1957 if (state & TASK_INTERRUPTIBL <<
1958 __schedstat_set(p->st <<
1959 rq_cl <<
1960 <<
1961 if (state & TASK_UNINTERRUPTI <<
1962 __schedstat_set(p->st <<
1963 rq_cl <<
1964 } <<
1965 } 997 }
1966 998
1967 static void __enqueue_dl_entity(struct sched_ 999 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1968 { 1000 {
1969 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1001 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
>> 1002 struct rb_node **link = &dl_rq->rb_root.rb_node;
>> 1003 struct rb_node *parent = NULL;
>> 1004 struct sched_dl_entity *entry;
>> 1005 int leftmost = 1;
>> 1006
>> 1007 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
>> 1008
>> 1009 while (*link) {
>> 1010 parent = *link;
>> 1011 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
>> 1012 if (dl_time_before(dl_se->deadline, entry->deadline))
>> 1013 link = &parent->rb_left;
>> 1014 else {
>> 1015 link = &parent->rb_right;
>> 1016 leftmost = 0;
>> 1017 }
>> 1018 }
1970 1019
1971 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->r !! 1020 if (leftmost)
>> 1021 dl_rq->rb_leftmost = &dl_se->rb_node;
1972 1022
1973 rb_add_cached(&dl_se->rb_node, &dl_rq !! 1023 rb_link_node(&dl_se->rb_node, parent, link);
>> 1024 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
1974 1025
1975 inc_dl_tasks(dl_se, dl_rq); 1026 inc_dl_tasks(dl_se, dl_rq);
1976 } 1027 }
1977 1028
1978 static void __dequeue_dl_entity(struct sched_ 1029 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1979 { 1030 {
1980 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1031 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1981 1032
1982 if (RB_EMPTY_NODE(&dl_se->rb_node)) 1033 if (RB_EMPTY_NODE(&dl_se->rb_node))
1983 return; 1034 return;
1984 1035
1985 rb_erase_cached(&dl_se->rb_node, &dl_ !! 1036 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
>> 1037 struct rb_node *next_node;
1986 1038
>> 1039 next_node = rb_next(&dl_se->rb_node);
>> 1040 dl_rq->rb_leftmost = next_node;
>> 1041 }
>> 1042
>> 1043 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
1987 RB_CLEAR_NODE(&dl_se->rb_node); 1044 RB_CLEAR_NODE(&dl_se->rb_node);
1988 1045
1989 dec_dl_tasks(dl_se, dl_rq); 1046 dec_dl_tasks(dl_se, dl_rq);
1990 } 1047 }
1991 1048
1992 static void 1049 static void
1993 enqueue_dl_entity(struct sched_dl_entity *dl_ !! 1050 enqueue_dl_entity(struct sched_dl_entity *dl_se,
>> 1051 struct sched_dl_entity *pi_se, int flags)
1994 { 1052 {
1995 WARN_ON_ONCE(on_dl_rq(dl_se)); !! 1053 BUG_ON(on_dl_rq(dl_se));
1996 <<
1997 update_stats_enqueue_dl(dl_rq_of_se(d <<
1998 <<
1999 /* <<
2000 * Check if a constrained deadline ta <<
2001 * after the deadline but before the <<
2002 * If that is the case, the task will <<
2003 * the replenishment timer will be se <<
2004 */ <<
2005 if (!dl_se->dl_throttled && !dl_is_im <<
2006 dl_check_constrained_dl(dl_se <<
2007 <<
2008 if (flags & (ENQUEUE_RESTORE|ENQUEUE_ <<
2009 struct dl_rq *dl_rq = dl_rq_o <<
2010 <<
2011 add_rq_bw(dl_se, dl_rq); <<
2012 add_running_bw(dl_se, dl_rq); <<
2013 } <<
2014 <<
2015 /* <<
2016 * If p is throttled, we do not enque <<
2017 * its budget it needs a replenishmen <<
2018 * its rq, the bandwidth timer callba <<
2019 * run yet) will take care of this. <<
2020 * However, the active utilization do <<
2021 * that the task is on the runqueue o <<
2022 * task's state - in GRUB parlance, " <<
2023 * In other words, even if a task is <<
2024 * be counted in the active utilizati <<
2025 * add_running_bw(). <<
2026 */ <<
2027 if (!dl_se->dl_defer && dl_se->dl_thr <<
2028 if (flags & ENQUEUE_WAKEUP) <<
2029 task_contending(dl_se <<
2030 <<
2031 return; <<
2032 } <<
2033 1054
2034 /* 1055 /*
2035 * If this is a wakeup or a new insta 1056 * If this is a wakeup or a new instance, the scheduling
2036 * parameters of the task might need 1057 * parameters of the task might need updating. Otherwise,
2037 * we want a replenishment of its run 1058 * we want a replenishment of its runtime.
2038 */ 1059 */
2039 if (flags & ENQUEUE_WAKEUP) { !! 1060 if (dl_se->dl_new || flags & ENQUEUE_WAKEUP)
2040 task_contending(dl_se, flags) !! 1061 update_dl_entity(dl_se, pi_se);
2041 update_dl_entity(dl_se); !! 1062 else if (flags & ENQUEUE_REPLENISH)
2042 } else if (flags & ENQUEUE_REPLENISH) !! 1063 replenish_dl_entity(dl_se, pi_se);
2043 replenish_dl_entity(dl_se); <<
2044 } else if ((flags & ENQUEUE_RESTORE) <<
2045 dl_time_before(dl_se->dead <<
2046 setup_new_dl_entity(dl_se); <<
2047 } <<
2048 <<
2049 /* <<
2050 * If the reservation is still thrott <<
2051 * deferred task and still got to wai <<
2052 */ <<
2053 if (dl_se->dl_throttled && start_dl_t <<
2054 return; <<
2055 <<
2056 /* <<
2057 * We're about to enqueue, make sure <<
2058 * In case the timer was not started, <<
2059 * has passed, mark as not throttled <<
2060 * Also cancel earlier timers, since <<
2061 */ <<
2062 if (dl_se->dl_throttled) { <<
2063 hrtimer_try_to_cancel(&dl_se- <<
2064 dl_se->dl_defer_armed = 0; <<
2065 dl_se->dl_throttled = 0; <<
2066 } <<
2067 1064
2068 __enqueue_dl_entity(dl_se); 1065 __enqueue_dl_entity(dl_se);
2069 } 1066 }
2070 1067
2071 static void dequeue_dl_entity(struct sched_dl !! 1068 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
2072 { 1069 {
2073 __dequeue_dl_entity(dl_se); 1070 __dequeue_dl_entity(dl_se);
2074 <<
2075 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIG <<
2076 struct dl_rq *dl_rq = dl_rq_o <<
2077 <<
2078 sub_running_bw(dl_se, dl_rq); <<
2079 sub_rq_bw(dl_se, dl_rq); <<
2080 } <<
2081 <<
2082 /* <<
2083 * This check allows to start the ina <<
2084 * decrease the active utilization, i <<
2085 * when the task blocks and when it i <<
2086 * (p->state == TASK_DEAD). We can ha <<
2087 * way, because from GRUB's point of <<
2088 * (the task moves from "active conte <<
2089 * or "inactive") <<
2090 */ <<
2091 if (flags & DEQUEUE_SLEEP) <<
2092 task_non_contending(dl_se); <<
2093 } 1071 }
2094 1072
2095 static void enqueue_task_dl(struct rq *rq, st 1073 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2096 { 1074 {
2097 if (is_dl_boosted(&p->dl)) { !! 1075 struct task_struct *pi_task = rt_mutex_get_top_task(p);
2098 /* !! 1076 struct sched_dl_entity *pi_se = &p->dl;
2099 * Because of delays in the d !! 1077
2100 * thread's runtime, it might !! 1078 /*
2101 * goes to sleep in a rt mute !! 1079 * Use the scheduling parameters of the top pi-waiter
2102 * a consequence, the thread !! 1080 * task if we have one and its (absolute) deadline is
2103 * !! 1081 * smaller than our one... OTW we keep our runtime and
2104 * While waiting for the mute !! 1082 * deadline.
2105 * boosted via PI, resulting !! 1083 */
2106 * and boosted at the same ti !! 1084 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) {
2107 * !! 1085 pi_se = &pi_task->dl;
2108 * In this case, the boost ov <<
2109 */ <<
2110 if (p->dl.dl_throttled) { <<
2111 /* <<
2112 * The replenish time <<
2113 * problem if it fire <<
2114 * are ignored in dl_ <<
2115 * <<
2116 * If the timer callb <<
2117 * it will eventually <<
2118 */ <<
2119 if (hrtimer_try_to_ca <<
2120 !dl_server(&p->dl <<
2121 put_task_stru <<
2122 p->dl.dl_throttled = <<
2123 } <<
2124 } else if (!dl_prio(p->normal_prio)) 1086 } else if (!dl_prio(p->normal_prio)) {
2125 /* 1087 /*
2126 * Special case in which we h !! 1088 * Special case in which we have a !SCHED_DEADLINE task
2127 * to be deboosted, but excee !! 1089 * that is going to be deboosted, but exceedes its
2128 * replenishing it, as it's g !! 1090 * runtime while doing so. No point in replenishing
2129 * scheduling class after thi !! 1091 * it, as it's going to return back to its original
2130 * clear the flag, otherwise !! 1092 * scheduling class after this.
2131 * being boosted again with n <<
2132 * the throttle. <<
2133 */ 1093 */
2134 p->dl.dl_throttled = 0; !! 1094 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
2135 if (!(flags & ENQUEUE_REPLENI <<
2136 printk_deferred_once( <<
2137 <<
2138 <<
2139 return; 1095 return;
2140 } 1096 }
2141 1097
2142 check_schedstat_required(); !! 1098 /*
2143 update_stats_wait_start_dl(dl_rq_of_s !! 1099 * Check if a constrained deadline task was activated
2144 !! 1100 * after the deadline but before the next period.
2145 if (p->on_rq == TASK_ON_RQ_MIGRATING) !! 1101 * If that is the case, the task will be throttled and
2146 flags |= ENQUEUE_MIGRATING; !! 1102 * the replenishment timer will be set to the next period.
2147 !! 1103 */
2148 enqueue_dl_entity(&p->dl, flags); !! 1104 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
>> 1105 dl_check_constrained_dl(&p->dl);
2149 1106
2150 if (dl_server(&p->dl)) !! 1107 /*
>> 1108 * If p is throttled, we do nothing. In fact, if it exhausted
>> 1109 * its budget it needs a replenishment and, since it now is on
>> 1110 * its rq, the bandwidth timer callback (which clearly has not
>> 1111 * run yet) will take care of this.
>> 1112 */
>> 1113 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH))
2151 return; 1114 return;
2152 1115
2153 if (!task_current(rq, p) && !p->dl.dl !! 1116 enqueue_dl_entity(&p->dl, pi_se, flags);
>> 1117
>> 1118 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
2154 enqueue_pushable_dl_task(rq, 1119 enqueue_pushable_dl_task(rq, p);
2155 } 1120 }
2156 1121
2157 static bool dequeue_task_dl(struct rq *rq, st !! 1122 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2158 { 1123 {
2159 update_curr_dl(rq); !! 1124 dequeue_dl_entity(&p->dl);
2160 !! 1125 dequeue_pushable_dl_task(rq, p);
2161 if (p->on_rq == TASK_ON_RQ_MIGRATING) !! 1126 }
2162 flags |= DEQUEUE_MIGRATING; <<
2163 <<
2164 dequeue_dl_entity(&p->dl, flags); <<
2165 if (!p->dl.dl_throttled && !dl_server <<
2166 dequeue_pushable_dl_task(rq, <<
2167 1127
2168 return true; !! 1128 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
>> 1129 {
>> 1130 update_curr_dl(rq);
>> 1131 __dequeue_task_dl(rq, p, flags);
2169 } 1132 }
2170 1133
2171 /* 1134 /*
2172 * Yield task semantic for -deadline tasks is 1135 * Yield task semantic for -deadline tasks is:
2173 * 1136 *
2174 * get off from the CPU until our next inst 1137 * get off from the CPU until our next instance, with
2175 * a new runtime. This is of little use now 1138 * a new runtime. This is of little use now, since we
2176 * don't have a bandwidth reclaiming mechan 1139 * don't have a bandwidth reclaiming mechanism. Anyway,
2177 * bandwidth reclaiming is planned for the 1140 * bandwidth reclaiming is planned for the future, and
2178 * yield_task_dl will indicate that some sp 1141 * yield_task_dl will indicate that some spare budget
2179 * is available for other task instances to 1142 * is available for other task instances to use it.
2180 */ 1143 */
2181 static void yield_task_dl(struct rq *rq) 1144 static void yield_task_dl(struct rq *rq)
2182 { 1145 {
>> 1146 struct task_struct *p = rq->curr;
>> 1147
2183 /* 1148 /*
2184 * We make the task go to sleep until 1149 * We make the task go to sleep until its current deadline by
2185 * forcing its runtime to zero. This 1150 * forcing its runtime to zero. This way, update_curr_dl() stops
2186 * it and the bandwidth timer will wa 1151 * it and the bandwidth timer will wake it up and will give it
2187 * new scheduling parameters (thanks 1152 * new scheduling parameters (thanks to dl_yielded=1).
2188 */ 1153 */
2189 rq->curr->dl.dl_yielded = 1; !! 1154 if (p->dl.runtime > 0) {
2190 !! 1155 rq->curr->dl.dl_yielded = 1;
>> 1156 p->dl.runtime = 0;
>> 1157 }
2191 update_rq_clock(rq); 1158 update_rq_clock(rq);
2192 update_curr_dl(rq); 1159 update_curr_dl(rq);
2193 /* 1160 /*
2194 * Tell update_rq_clock() that we've 1161 * Tell update_rq_clock() that we've just updated,
2195 * so we don't do microscopic update 1162 * so we don't do microscopic update in schedule()
2196 * and double the fastpath cost. 1163 * and double the fastpath cost.
2197 */ 1164 */
2198 rq_clock_skip_update(rq); !! 1165 rq_clock_skip_update(rq, true);
2199 } 1166 }
2200 1167
2201 #ifdef CONFIG_SMP 1168 #ifdef CONFIG_SMP
2202 1169
2203 static inline bool dl_task_is_earliest_deadli <<
2204 <<
2205 { <<
2206 return (!rq->dl.dl_nr_running || <<
2207 dl_time_before(p->dl.deadline <<
2208 rq->dl.earlies <<
2209 } <<
2210 <<
2211 static int find_later_rq(struct task_struct * 1170 static int find_later_rq(struct task_struct *task);
2212 1171
2213 static int 1172 static int
2214 select_task_rq_dl(struct task_struct *p, int !! 1173 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
2215 { 1174 {
2216 struct task_struct *curr; 1175 struct task_struct *curr;
2217 bool select_rq; <<
2218 struct rq *rq; 1176 struct rq *rq;
2219 1177
2220 if (!(flags & WF_TTWU)) !! 1178 if (sd_flag != SD_BALANCE_WAKE)
2221 goto out; 1179 goto out;
2222 1180
2223 rq = cpu_rq(cpu); 1181 rq = cpu_rq(cpu);
2224 1182
2225 rcu_read_lock(); 1183 rcu_read_lock();
2226 curr = READ_ONCE(rq->curr); /* unlock 1184 curr = READ_ONCE(rq->curr); /* unlocked access */
2227 1185
2228 /* 1186 /*
2229 * If we are dealing with a -deadline 1187 * If we are dealing with a -deadline task, we must
2230 * decide where to wake it up. 1188 * decide where to wake it up.
2231 * If it has a later deadline and the 1189 * If it has a later deadline and the current task
2232 * on this rq can't move (provided th 1190 * on this rq can't move (provided the waking task
2233 * can!) we prefer to send it somewhe 1191 * can!) we prefer to send it somewhere else. On the
2234 * other hand, if it has a shorter de 1192 * other hand, if it has a shorter deadline, we
2235 * try to make it stay here, it might 1193 * try to make it stay here, it might be important.
2236 */ 1194 */
2237 select_rq = unlikely(dl_task(curr)) & !! 1195 if (unlikely(dl_task(curr)) &&
2238 (curr->nr_cpus_allowed < !! 1196 (curr->nr_cpus_allowed < 2 ||
2239 !dl_entity_preempt(&p->d !! 1197 !dl_entity_preempt(&p->dl, &curr->dl)) &&
2240 p->nr_cpus_allowed > 1; !! 1198 (p->nr_cpus_allowed > 1)) {
2241 <<
2242 /* <<
2243 * Take the capacity of the CPU into <<
2244 * ensure it fits the requirement of <<
2245 */ <<
2246 if (sched_asym_cpucap_active()) <<
2247 select_rq |= !dl_task_fits_ca <<
2248 <<
2249 if (select_rq) { <<
2250 int target = find_later_rq(p) 1199 int target = find_later_rq(p);
2251 1200
2252 if (target != -1 && 1201 if (target != -1 &&
2253 dl_task_is_earliest_deadl !! 1202 (dl_time_before(p->dl.deadline,
>> 1203 cpu_rq(target)->dl.earliest_dl.curr) ||
>> 1204 (cpu_rq(target)->dl.dl_nr_running == 0)))
2254 cpu = target; 1205 cpu = target;
2255 } 1206 }
2256 rcu_read_unlock(); 1207 rcu_read_unlock();
2257 1208
2258 out: 1209 out:
2259 return cpu; 1210 return cpu;
2260 } 1211 }
2261 1212
2262 static void migrate_task_rq_dl(struct task_st <<
2263 { <<
2264 struct rq_flags rf; <<
2265 struct rq *rq; <<
2266 <<
2267 if (READ_ONCE(p->__state) != TASK_WAK <<
2268 return; <<
2269 <<
2270 rq = task_rq(p); <<
2271 /* <<
2272 * Since p->state == TASK_WAKING, set <<
2273 * from try_to_wake_up(). Hence, p->p <<
2274 * rq->lock is not... So, lock it <<
2275 */ <<
2276 rq_lock(rq, &rf); <<
2277 if (p->dl.dl_non_contending) { <<
2278 update_rq_clock(rq); <<
2279 sub_running_bw(&p->dl, &rq->d <<
2280 p->dl.dl_non_contending = 0; <<
2281 /* <<
2282 * If the timer handler is cu <<
2283 * timer cannot be canceled, <<
2284 * will see that dl_not_conte <<
2285 * will not touch the rq's ac <<
2286 * so we are still safe. <<
2287 */ <<
2288 if (hrtimer_try_to_cancel(&p- <<
2289 put_task_struct(p); <<
2290 } <<
2291 sub_rq_bw(&p->dl, &rq->dl); <<
2292 rq_unlock(rq, &rf); <<
2293 } <<
2294 <<
2295 static void check_preempt_equal_dl(struct rq 1213 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
2296 { 1214 {
2297 /* 1215 /*
2298 * Current can't be migrated, useless 1216 * Current can't be migrated, useless to reschedule,
2299 * let's hope p can move out. 1217 * let's hope p can move out.
2300 */ 1218 */
2301 if (rq->curr->nr_cpus_allowed == 1 || 1219 if (rq->curr->nr_cpus_allowed == 1 ||
2302 !cpudl_find(&rq->rd->cpudl, rq->c !! 1220 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
2303 return; 1221 return;
2304 1222
2305 /* 1223 /*
2306 * p is migratable, so let's not sche 1224 * p is migratable, so let's not schedule it and
2307 * see if it is pushed or pulled some 1225 * see if it is pushed or pulled somewhere else.
2308 */ 1226 */
2309 if (p->nr_cpus_allowed != 1 && 1227 if (p->nr_cpus_allowed != 1 &&
2310 cpudl_find(&rq->rd->cpudl, p, NUL !! 1228 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
2311 return; 1229 return;
2312 1230
2313 resched_curr(rq); 1231 resched_curr(rq);
2314 } 1232 }
2315 1233
2316 static int balance_dl(struct rq *rq, struct t <<
2317 { <<
2318 if (!on_dl_rq(&p->dl) && need_pull_dl <<
2319 /* <<
2320 * This is OK, because curren <<
2321 * picked for load-balance an <<
2322 * disabled avoiding further <<
2323 * not yet started the pickin <<
2324 */ <<
2325 rq_unpin_lock(rq, rf); <<
2326 pull_dl_task(rq); <<
2327 rq_repin_lock(rq, rf); <<
2328 } <<
2329 <<
2330 return sched_stop_runnable(rq) || sch <<
2331 } <<
2332 #endif /* CONFIG_SMP */ 1234 #endif /* CONFIG_SMP */
2333 1235
2334 /* 1236 /*
2335 * Only called when both the current and waki 1237 * Only called when both the current and waking task are -deadline
2336 * tasks. 1238 * tasks.
2337 */ 1239 */
2338 static void wakeup_preempt_dl(struct rq *rq, !! 1240 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
2339 int flags) 1241 int flags)
2340 { 1242 {
2341 if (dl_entity_preempt(&p->dl, &rq->cu 1243 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
2342 resched_curr(rq); 1244 resched_curr(rq);
2343 return; 1245 return;
2344 } 1246 }
2345 1247
2346 #ifdef CONFIG_SMP 1248 #ifdef CONFIG_SMP
2347 /* 1249 /*
2348 * In the unlikely case current and p 1250 * In the unlikely case current and p have the same deadline
2349 * let us try to decide what's the be 1251 * let us try to decide what's the best thing to do...
2350 */ 1252 */
2351 if ((p->dl.deadline == rq->curr->dl.d 1253 if ((p->dl.deadline == rq->curr->dl.deadline) &&
2352 !test_tsk_need_resched(rq->curr)) 1254 !test_tsk_need_resched(rq->curr))
2353 check_preempt_equal_dl(rq, p) 1255 check_preempt_equal_dl(rq, p);
2354 #endif /* CONFIG_SMP */ 1256 #endif /* CONFIG_SMP */
2355 } 1257 }
2356 1258
2357 #ifdef CONFIG_SCHED_HRTICK 1259 #ifdef CONFIG_SCHED_HRTICK
2358 static void start_hrtick_dl(struct rq *rq, st !! 1260 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2359 { 1261 {
2360 hrtick_start(rq, dl_se->runtime); !! 1262 hrtick_start(rq, p->dl.runtime);
2361 } 1263 }
2362 #else /* !CONFIG_SCHED_HRTICK */ 1264 #else /* !CONFIG_SCHED_HRTICK */
2363 static void start_hrtick_dl(struct rq *rq, st !! 1265 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2364 { 1266 {
2365 } 1267 }
2366 #endif 1268 #endif
2367 1269
2368 static void set_next_task_dl(struct rq *rq, s !! 1270 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
>> 1271 struct dl_rq *dl_rq)
2369 { 1272 {
2370 struct sched_dl_entity *dl_se = &p->d !! 1273 struct rb_node *left = dl_rq->rb_leftmost;
2371 struct dl_rq *dl_rq = &rq->dl; <<
2372 <<
2373 p->se.exec_start = rq_clock_task(rq); <<
2374 if (on_dl_rq(&p->dl)) <<
2375 update_stats_wait_end_dl(dl_r <<
2376 <<
2377 /* You can't push away the running ta <<
2378 dequeue_pushable_dl_task(rq, p); <<
2379 <<
2380 if (!first) <<
2381 return; <<
2382 <<
2383 if (rq->curr->sched_class != &dl_sche <<
2384 update_dl_rq_load_avg(rq_cloc <<
2385 <<
2386 deadline_queue_push_tasks(rq); <<
2387 <<
2388 if (hrtick_enabled_dl(rq)) <<
2389 start_hrtick_dl(rq, &p->dl); <<
2390 } <<
2391 <<
2392 static struct sched_dl_entity *pick_next_dl_e <<
2393 { <<
2394 struct rb_node *left = rb_first_cache <<
2395 1274
2396 if (!left) 1275 if (!left)
2397 return NULL; 1276 return NULL;
2398 1277
2399 return __node_2_dle(left); !! 1278 return rb_entry(left, struct sched_dl_entity, rb_node);
2400 } 1279 }
2401 1280
2402 /* !! 1281 struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev)
2403 * __pick_next_task_dl - Helper to pick the n <<
2404 * @rq: The runqueue to pick the next task fr <<
2405 */ <<
2406 static struct task_struct *__pick_task_dl(str <<
2407 { 1282 {
2408 struct sched_dl_entity *dl_se; 1283 struct sched_dl_entity *dl_se;
2409 struct dl_rq *dl_rq = &rq->dl; <<
2410 struct task_struct *p; 1284 struct task_struct *p;
>> 1285 struct dl_rq *dl_rq;
2411 1286
2412 again: !! 1287 dl_rq = &rq->dl;
2413 if (!sched_dl_runnable(rq)) !! 1288
>> 1289 if (need_pull_dl_task(rq, prev)) {
>> 1290 /*
>> 1291 * This is OK, because current is on_cpu, which avoids it being
>> 1292 * picked for load-balance and preemption/IRQs are still
>> 1293 * disabled avoiding further scheduler activity on it and we're
>> 1294 * being very careful to re-start the picking loop.
>> 1295 */
>> 1296 lockdep_unpin_lock(&rq->lock);
>> 1297 pull_dl_task(rq);
>> 1298 lockdep_pin_lock(&rq->lock);
>> 1299 /*
>> 1300 * pull_rt_task() can drop (and re-acquire) rq->lock; this
>> 1301 * means a stop task can slip in, in which case we need to
>> 1302 * re-start task selection.
>> 1303 */
>> 1304 if (rq->stop && task_on_rq_queued(rq->stop))
>> 1305 return RETRY_TASK;
>> 1306 }
>> 1307
>> 1308 /*
>> 1309 * When prev is DL, we may throttle it in put_prev_task().
>> 1310 * So, we update time before we check for dl_nr_running.
>> 1311 */
>> 1312 if (prev->sched_class == &dl_sched_class)
>> 1313 update_curr_dl(rq);
>> 1314
>> 1315 if (unlikely(!dl_rq->dl_nr_running))
2414 return NULL; 1316 return NULL;
2415 1317
2416 dl_se = pick_next_dl_entity(dl_rq); !! 1318 put_prev_task(rq, prev);
2417 WARN_ON_ONCE(!dl_se); <<
2418 1319
2419 if (dl_server(dl_se)) { !! 1320 dl_se = pick_next_dl_entity(rq, dl_rq);
2420 p = dl_se->server_pick_task(d !! 1321 BUG_ON(!dl_se);
2421 if (!p) { <<
2422 dl_se->dl_yielded = 1 <<
2423 update_curr_dl_se(rq, <<
2424 goto again; <<
2425 } <<
2426 rq->dl_server = dl_se; <<
2427 } else { <<
2428 p = dl_task_of(dl_se); <<
2429 } <<
2430 1322
2431 return p; !! 1323 p = dl_task_of(dl_se);
2432 } !! 1324 p->se.exec_start = rq_clock_task(rq);
2433 1325
2434 static struct task_struct *pick_task_dl(struc !! 1326 /* Running task will never be pushed. */
2435 { !! 1327 dequeue_pushable_dl_task(rq, p);
2436 return __pick_task_dl(rq); <<
2437 } <<
2438 1328
2439 static void put_prev_task_dl(struct rq *rq, s !! 1329 if (hrtick_enabled(rq))
2440 { !! 1330 start_hrtick_dl(rq, p);
2441 struct sched_dl_entity *dl_se = &p->d <<
2442 struct dl_rq *dl_rq = &rq->dl; <<
2443 1331
2444 if (on_dl_rq(&p->dl)) !! 1332 queue_push_tasks(rq);
2445 update_stats_wait_start_dl(dl <<
2446 1333
>> 1334 return p;
>> 1335 }
>> 1336
>> 1337 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
>> 1338 {
2447 update_curr_dl(rq); 1339 update_curr_dl(rq);
2448 1340
2449 update_dl_rq_load_avg(rq_clock_pelt(r <<
2450 if (on_dl_rq(&p->dl) && p->nr_cpus_al 1341 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2451 enqueue_pushable_dl_task(rq, 1342 enqueue_pushable_dl_task(rq, p);
2452 } 1343 }
2453 1344
2454 /* <<
2455 * scheduler tick hitting a task of our sched <<
2456 * <<
2457 * NOTE: This function can be called remotely <<
2458 * goes along full dynticks. Therefore no loc <<
2459 * and everything must be accessed through th <<
2460 * parameters. <<
2461 */ <<
2462 static void task_tick_dl(struct rq *rq, struc 1345 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2463 { 1346 {
2464 update_curr_dl(rq); 1347 update_curr_dl(rq);
2465 1348
2466 update_dl_rq_load_avg(rq_clock_pelt(r <<
2467 /* 1349 /*
2468 * Even when we have runtime, update_ 1350 * Even when we have runtime, update_curr_dl() might have resulted in us
2469 * not being the leftmost task anymor 1351 * not being the leftmost task anymore. In that case NEED_RESCHED will
2470 * be set and schedule() will start a 1352 * be set and schedule() will start a new hrtick for the next task.
2471 */ 1353 */
2472 if (hrtick_enabled_dl(rq) && queued & !! 1354 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
2473 is_leftmost(&p->dl, &rq->dl)) !! 1355 is_leftmost(p, &rq->dl))
2474 start_hrtick_dl(rq, &p->dl); !! 1356 start_hrtick_dl(rq, p);
2475 } 1357 }
2476 1358
2477 static void task_fork_dl(struct task_struct * 1359 static void task_fork_dl(struct task_struct *p)
2478 { 1360 {
2479 /* 1361 /*
2480 * SCHED_DEADLINE tasks cannot fork a 1362 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2481 * sched_fork() 1363 * sched_fork()
2482 */ 1364 */
2483 } 1365 }
2484 1366
>> 1367 static void task_dead_dl(struct task_struct *p)
>> 1368 {
>> 1369 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
>> 1370
>> 1371 /*
>> 1372 * Since we are TASK_DEAD we won't slip out of the domain!
>> 1373 */
>> 1374 raw_spin_lock_irq(&dl_b->lock);
>> 1375 /* XXX we should retain the bw until 0-lag */
>> 1376 dl_b->total_bw -= p->dl.dl_bw;
>> 1377 raw_spin_unlock_irq(&dl_b->lock);
>> 1378 }
>> 1379
>> 1380 static void set_curr_task_dl(struct rq *rq)
>> 1381 {
>> 1382 struct task_struct *p = rq->curr;
>> 1383
>> 1384 p->se.exec_start = rq_clock_task(rq);
>> 1385
>> 1386 /* You can't push away the running task */
>> 1387 dequeue_pushable_dl_task(rq, p);
>> 1388 }
>> 1389
2485 #ifdef CONFIG_SMP 1390 #ifdef CONFIG_SMP
2486 1391
2487 /* Only try algorithms three times */ 1392 /* Only try algorithms three times */
2488 #define DL_MAX_TRIES 3 1393 #define DL_MAX_TRIES 3
2489 1394
2490 static int pick_dl_task(struct rq *rq, struct 1395 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2491 { 1396 {
2492 if (!task_on_cpu(rq, p) && !! 1397 if (!task_running(rq, p) &&
2493 cpumask_test_cpu(cpu, &p->cpus_ma !! 1398 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2494 return 1; 1399 return 1;
2495 return 0; 1400 return 0;
2496 } 1401 }
2497 1402
>> 1403 /* Returns the second earliest -deadline task, NULL otherwise */
>> 1404 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
>> 1405 {
>> 1406 struct rb_node *next_node = rq->dl.rb_leftmost;
>> 1407 struct sched_dl_entity *dl_se;
>> 1408 struct task_struct *p = NULL;
>> 1409
>> 1410 next_node:
>> 1411 next_node = rb_next(next_node);
>> 1412 if (next_node) {
>> 1413 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
>> 1414 p = dl_task_of(dl_se);
>> 1415
>> 1416 if (pick_dl_task(rq, p, cpu))
>> 1417 return p;
>> 1418
>> 1419 goto next_node;
>> 1420 }
>> 1421
>> 1422 return NULL;
>> 1423 }
>> 1424
2498 /* 1425 /*
2499 * Return the earliest pushable rq's task, wh 1426 * Return the earliest pushable rq's task, which is suitable to be executed
2500 * on the CPU, NULL otherwise: 1427 * on the CPU, NULL otherwise:
2501 */ 1428 */
2502 static struct task_struct *pick_earliest_push 1429 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2503 { 1430 {
>> 1431 struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost;
2504 struct task_struct *p = NULL; 1432 struct task_struct *p = NULL;
2505 struct rb_node *next_node; <<
2506 1433
2507 if (!has_pushable_dl_tasks(rq)) 1434 if (!has_pushable_dl_tasks(rq))
2508 return NULL; 1435 return NULL;
2509 1436
2510 next_node = rb_first_cached(&rq->dl.p <<
2511 <<
2512 next_node: 1437 next_node:
2513 if (next_node) { 1438 if (next_node) {
2514 p = __node_2_pdl(next_node); !! 1439 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
2515 1440
2516 if (pick_dl_task(rq, p, cpu)) 1441 if (pick_dl_task(rq, p, cpu))
2517 return p; 1442 return p;
2518 1443
2519 next_node = rb_next(next_node 1444 next_node = rb_next(next_node);
2520 goto next_node; 1445 goto next_node;
2521 } 1446 }
2522 1447
2523 return NULL; 1448 return NULL;
2524 } 1449 }
2525 1450
2526 static DEFINE_PER_CPU(cpumask_var_t, local_cp 1451 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2527 1452
2528 static int find_later_rq(struct task_struct * 1453 static int find_later_rq(struct task_struct *task)
2529 { 1454 {
2530 struct sched_domain *sd; 1455 struct sched_domain *sd;
2531 struct cpumask *later_mask = this_cpu 1456 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2532 int this_cpu = smp_processor_id(); 1457 int this_cpu = smp_processor_id();
2533 int cpu = task_cpu(task); !! 1458 int best_cpu, cpu = task_cpu(task);
2534 1459
2535 /* Make sure the mask is initialized 1460 /* Make sure the mask is initialized first */
2536 if (unlikely(!later_mask)) 1461 if (unlikely(!later_mask))
2537 return -1; 1462 return -1;
2538 1463
2539 if (task->nr_cpus_allowed == 1) 1464 if (task->nr_cpus_allowed == 1)
2540 return -1; 1465 return -1;
2541 1466
2542 /* 1467 /*
2543 * We have to consider system topolog 1468 * We have to consider system topology and task affinity
2544 * first, then we can look for a suit !! 1469 * first, then we can look for a suitable cpu.
2545 */ 1470 */
2546 if (!cpudl_find(&task_rq(task)->rd->c !! 1471 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
>> 1472 task, later_mask);
>> 1473 if (best_cpu == -1)
2547 return -1; 1474 return -1;
2548 1475
2549 /* 1476 /*
2550 * If we are here, some targets have !! 1477 * If we are here, some target has been found,
2551 * the most suitable which is, among !! 1478 * the most suitable of which is cached in best_cpu.
2552 * current tasks have later deadlines !! 1479 * This is, among the runqueues where the current tasks
2553 * rq with the latest possible one. !! 1480 * have later deadlines than the task's one, the rq
>> 1481 * with the latest possible one.
2554 * 1482 *
2555 * Now we check how well this matches 1483 * Now we check how well this matches with task's
2556 * affinity and system topology. 1484 * affinity and system topology.
2557 * 1485 *
2558 * The last CPU where the task run is !! 1486 * The last cpu where the task run is our first
2559 * guess, since it is most likely cac 1487 * guess, since it is most likely cache-hot there.
2560 */ 1488 */
2561 if (cpumask_test_cpu(cpu, later_mask) 1489 if (cpumask_test_cpu(cpu, later_mask))
2562 return cpu; 1490 return cpu;
2563 /* 1491 /*
2564 * Check if this_cpu is to be skipped 1492 * Check if this_cpu is to be skipped (i.e., it is
2565 * not in the mask) or not. 1493 * not in the mask) or not.
2566 */ 1494 */
2567 if (!cpumask_test_cpu(this_cpu, later 1495 if (!cpumask_test_cpu(this_cpu, later_mask))
2568 this_cpu = -1; 1496 this_cpu = -1;
2569 1497
2570 rcu_read_lock(); 1498 rcu_read_lock();
2571 for_each_domain(cpu, sd) { 1499 for_each_domain(cpu, sd) {
2572 if (sd->flags & SD_WAKE_AFFIN 1500 if (sd->flags & SD_WAKE_AFFINE) {
2573 int best_cpu; <<
2574 1501
2575 /* 1502 /*
2576 * If possible, preem 1503 * If possible, preempting this_cpu is
2577 * cheaper than migra 1504 * cheaper than migrating.
2578 */ 1505 */
2579 if (this_cpu != -1 && 1506 if (this_cpu != -1 &&
2580 cpumask_test_cpu( 1507 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2581 rcu_read_unlo 1508 rcu_read_unlock();
2582 return this_c 1509 return this_cpu;
2583 } 1510 }
2584 1511
2585 best_cpu = cpumask_an <<
2586 <<
2587 /* 1512 /*
2588 * Last chance: if a !! 1513 * Last chance: if best_cpu is valid and is
2589 * and current sd spa !! 1514 * in the mask, that becomes our choice.
2590 * choice. Of course, <<
2591 * already under cons <<
2592 */ 1515 */
2593 if (best_cpu < nr_cpu !! 1516 if (best_cpu < nr_cpu_ids &&
>> 1517 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
2594 rcu_read_unlo 1518 rcu_read_unlock();
2595 return best_c 1519 return best_cpu;
2596 } 1520 }
2597 } 1521 }
2598 } 1522 }
2599 rcu_read_unlock(); 1523 rcu_read_unlock();
2600 1524
2601 /* 1525 /*
2602 * At this point, all our guesses fai 1526 * At this point, all our guesses failed, we just return
2603 * 'something', and let the caller so 1527 * 'something', and let the caller sort the things out.
2604 */ 1528 */
2605 if (this_cpu != -1) 1529 if (this_cpu != -1)
2606 return this_cpu; 1530 return this_cpu;
2607 1531
2608 cpu = cpumask_any_distribute(later_ma !! 1532 cpu = cpumask_any(later_mask);
2609 if (cpu < nr_cpu_ids) 1533 if (cpu < nr_cpu_ids)
2610 return cpu; 1534 return cpu;
2611 1535
2612 return -1; 1536 return -1;
2613 } 1537 }
2614 1538
2615 /* Locks the rq it finds */ 1539 /* Locks the rq it finds */
2616 static struct rq *find_lock_later_rq(struct t 1540 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2617 { 1541 {
2618 struct rq *later_rq = NULL; 1542 struct rq *later_rq = NULL;
2619 int tries; 1543 int tries;
2620 int cpu; 1544 int cpu;
2621 1545
2622 for (tries = 0; tries < DL_MAX_TRIES; 1546 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2623 cpu = find_later_rq(task); 1547 cpu = find_later_rq(task);
2624 1548
2625 if ((cpu == -1) || (cpu == rq 1549 if ((cpu == -1) || (cpu == rq->cpu))
2626 break; 1550 break;
2627 1551
2628 later_rq = cpu_rq(cpu); 1552 later_rq = cpu_rq(cpu);
2629 1553
2630 if (!dl_task_is_earliest_dead !! 1554 if (later_rq->dl.dl_nr_running &&
>> 1555 !dl_time_before(task->dl.deadline,
>> 1556 later_rq->dl.earliest_dl.curr)) {
2631 /* 1557 /*
2632 * Target rq has task 1558 * Target rq has tasks of equal or earlier deadline,
2633 * retrying does not 1559 * retrying does not release any lock and is unlikely
2634 * to yield a differe 1560 * to yield a different result.
2635 */ 1561 */
2636 later_rq = NULL; 1562 later_rq = NULL;
2637 break; 1563 break;
2638 } 1564 }
2639 1565
2640 /* Retry if something changed 1566 /* Retry if something changed. */
2641 if (double_lock_balance(rq, l 1567 if (double_lock_balance(rq, later_rq)) {
2642 if (unlikely(task_rq( 1568 if (unlikely(task_rq(task) != rq ||
2643 !cpumask !! 1569 !cpumask_test_cpu(later_rq->cpu,
2644 task_on_ !! 1570 &task->cpus_allowed) ||
2645 !dl_task !! 1571 task_running(rq, task) ||
2646 is_migra <<
2647 !task_on 1572 !task_on_rq_queued(task))) {
2648 double_unlock 1573 double_unlock_balance(rq, later_rq);
2649 later_rq = NU 1574 later_rq = NULL;
2650 break; 1575 break;
2651 } 1576 }
2652 } 1577 }
2653 1578
2654 /* 1579 /*
2655 * If the rq we found has no 1580 * If the rq we found has no -deadline task, or
2656 * its earliest one has a lat 1581 * its earliest one has a later deadline than our
2657 * task, the rq is a good one 1582 * task, the rq is a good one.
2658 */ 1583 */
2659 if (dl_task_is_earliest_deadl !! 1584 if (!later_rq->dl.dl_nr_running ||
>> 1585 dl_time_before(task->dl.deadline,
>> 1586 later_rq->dl.earliest_dl.curr))
2660 break; 1587 break;
2661 1588
2662 /* Otherwise we try again. */ 1589 /* Otherwise we try again. */
2663 double_unlock_balance(rq, lat 1590 double_unlock_balance(rq, later_rq);
2664 later_rq = NULL; 1591 later_rq = NULL;
2665 } 1592 }
2666 1593
2667 return later_rq; 1594 return later_rq;
2668 } 1595 }
2669 1596
2670 static struct task_struct *pick_next_pushable 1597 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2671 { 1598 {
2672 struct task_struct *p; 1599 struct task_struct *p;
2673 1600
2674 if (!has_pushable_dl_tasks(rq)) 1601 if (!has_pushable_dl_tasks(rq))
2675 return NULL; 1602 return NULL;
2676 1603
2677 p = __node_2_pdl(rb_first_cached(&rq- !! 1604 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
>> 1605 struct task_struct, pushable_dl_tasks);
2678 1606
2679 WARN_ON_ONCE(rq->cpu != task_cpu(p)); !! 1607 BUG_ON(rq->cpu != task_cpu(p));
2680 WARN_ON_ONCE(task_current(rq, p)); !! 1608 BUG_ON(task_current(rq, p));
2681 WARN_ON_ONCE(p->nr_cpus_allowed <= 1) !! 1609 BUG_ON(p->nr_cpus_allowed <= 1);
2682 1610
2683 WARN_ON_ONCE(!task_on_rq_queued(p)); !! 1611 BUG_ON(!task_on_rq_queued(p));
2684 WARN_ON_ONCE(!dl_task(p)); !! 1612 BUG_ON(!dl_task(p));
2685 1613
2686 return p; 1614 return p;
2687 } 1615 }
2688 1616
2689 /* 1617 /*
2690 * See if the non running -deadline tasks on 1618 * See if the non running -deadline tasks on this rq
2691 * can be sent to some other CPU where they c 1619 * can be sent to some other CPU where they can preempt
2692 * and start executing. 1620 * and start executing.
2693 */ 1621 */
2694 static int push_dl_task(struct rq *rq) 1622 static int push_dl_task(struct rq *rq)
2695 { 1623 {
2696 struct task_struct *next_task; 1624 struct task_struct *next_task;
2697 struct rq *later_rq; 1625 struct rq *later_rq;
2698 int ret = 0; 1626 int ret = 0;
2699 1627
>> 1628 if (!rq->dl.overloaded)
>> 1629 return 0;
>> 1630
2700 next_task = pick_next_pushable_dl_tas 1631 next_task = pick_next_pushable_dl_task(rq);
2701 if (!next_task) 1632 if (!next_task)
2702 return 0; 1633 return 0;
2703 1634
2704 retry: 1635 retry:
>> 1636 if (unlikely(next_task == rq->curr)) {
>> 1637 WARN_ON(1);
>> 1638 return 0;
>> 1639 }
>> 1640
2705 /* 1641 /*
2706 * If next_task preempts rq->curr, an 1642 * If next_task preempts rq->curr, and rq->curr
2707 * can move away, it makes sense to j 1643 * can move away, it makes sense to just reschedule
2708 * without going further in pushing n 1644 * without going further in pushing next_task.
2709 */ 1645 */
2710 if (dl_task(rq->curr) && 1646 if (dl_task(rq->curr) &&
2711 dl_time_before(next_task->dl.dead 1647 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2712 rq->curr->nr_cpus_allowed > 1) { 1648 rq->curr->nr_cpus_allowed > 1) {
2713 resched_curr(rq); 1649 resched_curr(rq);
2714 return 0; 1650 return 0;
2715 } 1651 }
2716 1652
2717 if (is_migration_disabled(next_task)) <<
2718 return 0; <<
2719 <<
2720 if (WARN_ON(next_task == rq->curr)) <<
2721 return 0; <<
2722 <<
2723 /* We might release rq lock */ 1653 /* We might release rq lock */
2724 get_task_struct(next_task); 1654 get_task_struct(next_task);
2725 1655
2726 /* Will lock the rq it'll find */ 1656 /* Will lock the rq it'll find */
2727 later_rq = find_lock_later_rq(next_ta 1657 later_rq = find_lock_later_rq(next_task, rq);
2728 if (!later_rq) { 1658 if (!later_rq) {
2729 struct task_struct *task; 1659 struct task_struct *task;
2730 1660
2731 /* 1661 /*
2732 * We must check all this aga 1662 * We must check all this again, since
2733 * find_lock_later_rq release 1663 * find_lock_later_rq releases rq->lock and it is
2734 * then possible that next_ta 1664 * then possible that next_task has migrated.
2735 */ 1665 */
2736 task = pick_next_pushable_dl_ 1666 task = pick_next_pushable_dl_task(rq);
2737 if (task == next_task) { !! 1667 if (task_cpu(next_task) == rq->cpu && task == next_task) {
2738 /* 1668 /*
2739 * The task is still 1669 * The task is still there. We don't try
2740 * again, some other !! 1670 * again, some other cpu will pull it when ready.
2741 */ 1671 */
2742 goto out; 1672 goto out;
2743 } 1673 }
2744 1674
2745 if (!task) 1675 if (!task)
2746 /* No more tasks */ 1676 /* No more tasks */
2747 goto out; 1677 goto out;
2748 1678
2749 put_task_struct(next_task); 1679 put_task_struct(next_task);
2750 next_task = task; 1680 next_task = task;
2751 goto retry; 1681 goto retry;
2752 } 1682 }
2753 1683
2754 deactivate_task(rq, next_task, 0); 1684 deactivate_task(rq, next_task, 0);
2755 set_task_cpu(next_task, later_rq->cpu 1685 set_task_cpu(next_task, later_rq->cpu);
2756 activate_task(later_rq, next_task, 0) 1686 activate_task(later_rq, next_task, 0);
2757 ret = 1; 1687 ret = 1;
2758 1688
2759 resched_curr(later_rq); 1689 resched_curr(later_rq);
2760 1690
2761 double_unlock_balance(rq, later_rq); 1691 double_unlock_balance(rq, later_rq);
2762 1692
2763 out: 1693 out:
2764 put_task_struct(next_task); 1694 put_task_struct(next_task);
2765 1695
2766 return ret; 1696 return ret;
2767 } 1697 }
2768 1698
2769 static void push_dl_tasks(struct rq *rq) 1699 static void push_dl_tasks(struct rq *rq)
2770 { 1700 {
2771 /* push_dl_task() will return true if 1701 /* push_dl_task() will return true if it moved a -deadline task */
2772 while (push_dl_task(rq)) 1702 while (push_dl_task(rq))
2773 ; 1703 ;
2774 } 1704 }
2775 1705
2776 static void pull_dl_task(struct rq *this_rq) 1706 static void pull_dl_task(struct rq *this_rq)
2777 { 1707 {
2778 int this_cpu = this_rq->cpu, cpu; 1708 int this_cpu = this_rq->cpu, cpu;
2779 struct task_struct *p, *push_task; !! 1709 struct task_struct *p;
2780 bool resched = false; 1710 bool resched = false;
2781 struct rq *src_rq; 1711 struct rq *src_rq;
2782 u64 dmin = LONG_MAX; 1712 u64 dmin = LONG_MAX;
2783 1713
2784 if (likely(!dl_overloaded(this_rq))) 1714 if (likely(!dl_overloaded(this_rq)))
2785 return; 1715 return;
2786 1716
2787 /* 1717 /*
2788 * Match the barrier from dl_set_over 1718 * Match the barrier from dl_set_overloaded; this guarantees that if we
2789 * see overloaded we must also see th 1719 * see overloaded we must also see the dlo_mask bit.
2790 */ 1720 */
2791 smp_rmb(); 1721 smp_rmb();
2792 1722
2793 for_each_cpu(cpu, this_rq->rd->dlo_ma 1723 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2794 if (this_cpu == cpu) 1724 if (this_cpu == cpu)
2795 continue; 1725 continue;
2796 1726
2797 src_rq = cpu_rq(cpu); 1727 src_rq = cpu_rq(cpu);
2798 1728
2799 /* 1729 /*
2800 * It looks racy, and it is! !! 1730 * It looks racy, abd it is! However, as in sched_rt.c,
2801 * we are fine with this. 1731 * we are fine with this.
2802 */ 1732 */
2803 if (this_rq->dl.dl_nr_running 1733 if (this_rq->dl.dl_nr_running &&
2804 dl_time_before(this_rq->d 1734 dl_time_before(this_rq->dl.earliest_dl.curr,
2805 src_rq->dl 1735 src_rq->dl.earliest_dl.next))
2806 continue; 1736 continue;
2807 1737
2808 /* Might drop this_rq->lock * 1738 /* Might drop this_rq->lock */
2809 push_task = NULL; <<
2810 double_lock_balance(this_rq, 1739 double_lock_balance(this_rq, src_rq);
2811 1740
2812 /* 1741 /*
2813 * If there are no more pulla 1742 * If there are no more pullable tasks on the
2814 * rq, we're done with it. 1743 * rq, we're done with it.
2815 */ 1744 */
2816 if (src_rq->dl.dl_nr_running 1745 if (src_rq->dl.dl_nr_running <= 1)
2817 goto skip; 1746 goto skip;
2818 1747
2819 p = pick_earliest_pushable_dl 1748 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2820 1749
2821 /* 1750 /*
2822 * We found a task to be pull 1751 * We found a task to be pulled if:
2823 * - it preempts our current 1752 * - it preempts our current (if there's one),
2824 * - it will preempt the las 1753 * - it will preempt the last one we pulled (if any).
2825 */ 1754 */
2826 if (p && dl_time_before(p->dl 1755 if (p && dl_time_before(p->dl.deadline, dmin) &&
2827 dl_task_is_earliest_deadl !! 1756 (!this_rq->dl.dl_nr_running ||
>> 1757 dl_time_before(p->dl.deadline,
>> 1758 this_rq->dl.earliest_dl.curr))) {
2828 WARN_ON(p == src_rq-> 1759 WARN_ON(p == src_rq->curr);
2829 WARN_ON(!task_on_rq_q 1760 WARN_ON(!task_on_rq_queued(p));
2830 1761
2831 /* 1762 /*
2832 * Then we pull iff p 1763 * Then we pull iff p has actually an earlier
2833 * deadline than the 1764 * deadline than the current task of its runqueue.
2834 */ 1765 */
2835 if (dl_time_before(p- 1766 if (dl_time_before(p->dl.deadline,
2836 sr 1767 src_rq->curr->dl.deadline))
2837 goto skip; 1768 goto skip;
2838 1769
2839 if (is_migration_disa !! 1770 resched = true;
2840 push_task = g !! 1771
2841 } else { !! 1772 deactivate_task(src_rq, p, 0);
2842 deactivate_ta !! 1773 set_task_cpu(p, this_cpu);
2843 set_task_cpu( !! 1774 activate_task(this_rq, p, 0);
2844 activate_task !! 1775 dmin = p->dl.deadline;
2845 dmin = p->dl. <<
2846 resched = tru <<
2847 } <<
2848 1776
2849 /* Is there any other 1777 /* Is there any other task even earlier? */
2850 } 1778 }
2851 skip: 1779 skip:
2852 double_unlock_balance(this_rq 1780 double_unlock_balance(this_rq, src_rq);
2853 <<
2854 if (push_task) { <<
2855 preempt_disable(); <<
2856 raw_spin_rq_unlock(th <<
2857 stop_one_cpu_nowait(s <<
2858 p <<
2859 preempt_enable(); <<
2860 raw_spin_rq_lock(this <<
2861 } <<
2862 } 1781 }
2863 1782
2864 if (resched) 1783 if (resched)
2865 resched_curr(this_rq); 1784 resched_curr(this_rq);
2866 } 1785 }
2867 1786
2868 /* 1787 /*
2869 * Since the task is not running and a resche 1788 * Since the task is not running and a reschedule is not going to happen
2870 * anytime soon on its runqueue, we try pushi 1789 * anytime soon on its runqueue, we try pushing it away now.
2871 */ 1790 */
2872 static void task_woken_dl(struct rq *rq, stru 1791 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2873 { 1792 {
2874 if (!task_on_cpu(rq, p) && !! 1793 if (!task_running(rq, p) &&
2875 !test_tsk_need_resched(rq->curr) 1794 !test_tsk_need_resched(rq->curr) &&
2876 p->nr_cpus_allowed > 1 && 1795 p->nr_cpus_allowed > 1 &&
2877 dl_task(rq->curr) && 1796 dl_task(rq->curr) &&
2878 (rq->curr->nr_cpus_allowed < 2 || 1797 (rq->curr->nr_cpus_allowed < 2 ||
2879 !dl_entity_preempt(&p->dl, &rq-> 1798 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2880 push_dl_tasks(rq); 1799 push_dl_tasks(rq);
2881 } 1800 }
2882 } 1801 }
2883 1802
2884 static void set_cpus_allowed_dl(struct task_s 1803 static void set_cpus_allowed_dl(struct task_struct *p,
2885 struct affini !! 1804 const struct cpumask *new_mask)
2886 { 1805 {
2887 struct root_domain *src_rd; 1806 struct root_domain *src_rd;
2888 struct rq *rq; 1807 struct rq *rq;
2889 1808
2890 WARN_ON_ONCE(!dl_task(p)); !! 1809 BUG_ON(!dl_task(p));
2891 1810
2892 rq = task_rq(p); 1811 rq = task_rq(p);
2893 src_rd = rq->rd; 1812 src_rd = rq->rd;
2894 /* 1813 /*
2895 * Migrating a SCHED_DEADLINE task be 1814 * Migrating a SCHED_DEADLINE task between exclusive
2896 * cpusets (different root_domains) e 1815 * cpusets (different root_domains) entails a bandwidth
2897 * update. We already made space for 1816 * update. We already made space for us in the destination
2898 * domain (see cpuset_can_attach()). 1817 * domain (see cpuset_can_attach()).
2899 */ 1818 */
2900 if (!cpumask_intersects(src_rd->span, !! 1819 if (!cpumask_intersects(src_rd->span, new_mask)) {
2901 struct dl_bw *src_dl_b; 1820 struct dl_bw *src_dl_b;
2902 1821
2903 src_dl_b = dl_bw_of(cpu_of(rq 1822 src_dl_b = dl_bw_of(cpu_of(rq));
2904 /* 1823 /*
2905 * We now free resources of t 1824 * We now free resources of the root_domain we are migrating
2906 * off. In the worst case, sc 1825 * off. In the worst case, sched_setattr() may temporary fail
2907 * until we complete the upda 1826 * until we complete the update.
2908 */ 1827 */
2909 raw_spin_lock(&src_dl_b->lock 1828 raw_spin_lock(&src_dl_b->lock);
2910 __dl_sub(src_dl_b, p->dl.dl_b !! 1829 __dl_clear(src_dl_b, p->dl.dl_bw);
2911 raw_spin_unlock(&src_dl_b->lo 1830 raw_spin_unlock(&src_dl_b->lock);
2912 } 1831 }
2913 1832
2914 set_cpus_allowed_common(p, ctx); !! 1833 set_cpus_allowed_common(p, new_mask);
2915 } 1834 }
2916 1835
2917 /* Assumes rq->lock is held */ 1836 /* Assumes rq->lock is held */
2918 static void rq_online_dl(struct rq *rq) 1837 static void rq_online_dl(struct rq *rq)
2919 { 1838 {
2920 if (rq->dl.overloaded) 1839 if (rq->dl.overloaded)
2921 dl_set_overload(rq); 1840 dl_set_overload(rq);
2922 1841
2923 cpudl_set_freecpu(&rq->rd->cpudl, rq- 1842 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2924 if (rq->dl.dl_nr_running > 0) 1843 if (rq->dl.dl_nr_running > 0)
2925 cpudl_set(&rq->rd->cpudl, rq- !! 1844 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
2926 } 1845 }
2927 1846
2928 /* Assumes rq->lock is held */ 1847 /* Assumes rq->lock is held */
2929 static void rq_offline_dl(struct rq *rq) 1848 static void rq_offline_dl(struct rq *rq)
2930 { 1849 {
2931 if (rq->dl.overloaded) 1850 if (rq->dl.overloaded)
2932 dl_clear_overload(rq); 1851 dl_clear_overload(rq);
2933 1852
2934 cpudl_clear(&rq->rd->cpudl, rq->cpu); !! 1853 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
2935 cpudl_clear_freecpu(&rq->rd->cpudl, r 1854 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2936 } 1855 }
2937 1856
2938 void __init init_sched_dl_class(void) 1857 void __init init_sched_dl_class(void)
2939 { 1858 {
2940 unsigned int i; 1859 unsigned int i;
2941 1860
2942 for_each_possible_cpu(i) 1861 for_each_possible_cpu(i)
2943 zalloc_cpumask_var_node(&per_ 1862 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2944 GFP_K 1863 GFP_KERNEL, cpu_to_node(i));
2945 } 1864 }
2946 1865
2947 void dl_add_task_root_domain(struct task_stru <<
2948 { <<
2949 struct rq_flags rf; <<
2950 struct rq *rq; <<
2951 struct dl_bw *dl_b; <<
2952 <<
2953 raw_spin_lock_irqsave(&p->pi_lock, rf <<
2954 if (!dl_task(p)) { <<
2955 raw_spin_unlock_irqrestore(&p <<
2956 return; <<
2957 } <<
2958 <<
2959 rq = __task_rq_lock(p, &rf); <<
2960 <<
2961 dl_b = &rq->rd->dl_bw; <<
2962 raw_spin_lock(&dl_b->lock); <<
2963 <<
2964 __dl_add(dl_b, p->dl.dl_bw, cpumask_w <<
2965 <<
2966 raw_spin_unlock(&dl_b->lock); <<
2967 <<
2968 task_rq_unlock(rq, p, &rf); <<
2969 } <<
2970 <<
2971 void dl_clear_root_domain(struct root_domain <<
2972 { <<
2973 unsigned long flags; <<
2974 <<
2975 raw_spin_lock_irqsave(&rd->dl_bw.lock <<
2976 rd->dl_bw.total_bw = 0; <<
2977 raw_spin_unlock_irqrestore(&rd->dl_bw <<
2978 } <<
2979 <<
2980 #endif /* CONFIG_SMP */ 1866 #endif /* CONFIG_SMP */
2981 1867
2982 static void switched_from_dl(struct rq *rq, s 1868 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2983 { 1869 {
2984 /* 1870 /*
2985 * task_non_contending() can start th !! 1871 * Start the deadline timer; if we switch back to dl before this we'll
2986 * time is in the future). If the tas !! 1872 * continue consuming our current CBS slice. If we stay outside of
2987 * the "inactive timer" fires, it can !! 1873 * SCHED_DEADLINE until the deadline passes, the timer will reset the
2988 * runtime using its current deadline !! 1874 * task.
2989 * SCHED_DEADLINE until the 0-lag tim <<
2990 * will reset the task parameters. <<
2991 */ <<
2992 if (task_on_rq_queued(p) && p->dl.dl_ <<
2993 task_non_contending(&p->dl); <<
2994 <<
2995 /* <<
2996 * In case a task is setscheduled out <<
2997 * keep track of that on its cpuset ( <<
2998 */ <<
2999 dec_dl_tasks_cs(p); <<
3000 <<
3001 if (!task_on_rq_queued(p)) { <<
3002 /* <<
3003 * Inactive timer is armed. H <<
3004 * might migrate away from th <<
3005 * some other class. We need <<
3006 * this rq running_bw now, or <<
3007 */ <<
3008 if (p->dl.dl_non_contending) <<
3009 sub_running_bw(&p->dl <<
3010 sub_rq_bw(&p->dl, &rq->dl); <<
3011 } <<
3012 <<
3013 /* <<
3014 * We cannot use inactive_task_timer( <<
3015 * at the 0-lag time, because the tas <<
3016 * while SCHED_OTHER in the meanwhile <<
3017 */ 1875 */
3018 if (p->dl.dl_non_contending) !! 1876 if (!start_dl_timer(p))
3019 p->dl.dl_non_contending = 0; !! 1877 __dl_clear_params(p);
3020 1878
3021 /* 1879 /*
3022 * Since this might be the only -dead 1880 * Since this might be the only -deadline task on the rq,
3023 * this is the right place to try to 1881 * this is the right place to try to pull some other one
3024 * from an overloaded CPU, if any. !! 1882 * from an overloaded cpu, if any.
3025 */ 1883 */
3026 if (!task_on_rq_queued(p) || rq->dl.d 1884 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
3027 return; 1885 return;
3028 1886
3029 deadline_queue_pull_task(rq); !! 1887 queue_pull_task(rq);
3030 } 1888 }
3031 1889
3032 /* 1890 /*
3033 * When switching to -deadline, we may overlo 1891 * When switching to -deadline, we may overload the rq, then
3034 * we try to push someone off, if possible. 1892 * we try to push someone off, if possible.
3035 */ 1893 */
3036 static void switched_to_dl(struct rq *rq, str 1894 static void switched_to_dl(struct rq *rq, struct task_struct *p)
3037 { 1895 {
3038 if (hrtimer_try_to_cancel(&p->dl.inac !! 1896 if (task_on_rq_queued(p) && rq->curr != p) {
3039 put_task_struct(p); <<
3040 <<
3041 /* <<
3042 * In case a task is setscheduled to <<
3043 * track of that on its cpuset (for c <<
3044 */ <<
3045 inc_dl_tasks_cs(p); <<
3046 <<
3047 /* If p is not queued we will update <<
3048 if (!task_on_rq_queued(p)) { <<
3049 add_rq_bw(&p->dl, &rq->dl); <<
3050 <<
3051 return; <<
3052 } <<
3053 <<
3054 if (rq->curr != p) { <<
3055 #ifdef CONFIG_SMP 1897 #ifdef CONFIG_SMP
3056 if (p->nr_cpus_allowed > 1 && 1898 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
3057 deadline_queue_push_t !! 1899 queue_push_tasks(rq);
3058 #endif 1900 #endif
3059 if (dl_task(rq->curr)) 1901 if (dl_task(rq->curr))
3060 wakeup_preempt_dl(rq, !! 1902 check_preempt_curr_dl(rq, p, 0);
3061 else 1903 else
3062 resched_curr(rq); 1904 resched_curr(rq);
3063 } else { <<
3064 update_dl_rq_load_avg(rq_cloc <<
3065 } 1905 }
3066 } 1906 }
3067 1907
3068 /* 1908 /*
3069 * If the scheduling parameters of a -deadlin 1909 * If the scheduling parameters of a -deadline task changed,
3070 * a push or pull operation might be needed. 1910 * a push or pull operation might be needed.
3071 */ 1911 */
3072 static void prio_changed_dl(struct rq *rq, st 1912 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
3073 int oldprio) 1913 int oldprio)
3074 { 1914 {
3075 if (!task_on_rq_queued(p)) !! 1915 if (task_on_rq_queued(p) || rq->curr == p) {
3076 return; <<
3077 <<
3078 #ifdef CONFIG_SMP 1916 #ifdef CONFIG_SMP
3079 /* !! 1917 /*
3080 * This might be too much, but unfort !! 1918 * This might be too much, but unfortunately
3081 * we don't have the old deadline val !! 1919 * we don't have the old deadline value, and
3082 * we can't argue if the task is incr !! 1920 * we can't argue if the task is increasing
3083 * or lowering its prio, so... !! 1921 * or lowering its prio, so...
3084 */ !! 1922 */
3085 if (!rq->dl.overloaded) !! 1923 if (!rq->dl.overloaded)
3086 deadline_queue_pull_task(rq); !! 1924 queue_pull_task(rq);
3087 1925
3088 if (task_current(rq, p)) { <<
3089 /* 1926 /*
3090 * If we now have a earlier d 1927 * If we now have a earlier deadline task than p,
3091 * then reschedule, provided 1928 * then reschedule, provided p is still on this
3092 * runqueue. 1929 * runqueue.
3093 */ 1930 */
3094 if (dl_time_before(rq->dl.ear 1931 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
3095 resched_curr(rq); 1932 resched_curr(rq);
3096 } else { !! 1933 #else
3097 /* 1934 /*
3098 * Current may not be deadlin !! 1935 * Again, we don't know if p has a earlier
3099 * have just replenished it ( !! 1936 * or later deadline, so let's blindly set a
3100 * !! 1937 * (maybe not needed) rescheduling point.
3101 * Otherwise, if p was given <<
3102 */ 1938 */
3103 if (!dl_task(rq->curr) || !! 1939 resched_curr(rq);
3104 dl_time_before(p->dl.dead !! 1940 #endif /* CONFIG_SMP */
3105 resched_curr(rq); !! 1941 } else
3106 } !! 1942 switched_to_dl(rq, p);
3107 #else <<
3108 /* <<
3109 * We don't know if p has a earlier o <<
3110 * set a (maybe not needed) reschedul <<
3111 */ <<
3112 resched_curr(rq); <<
3113 #endif <<
3114 } <<
3115 <<
3116 #ifdef CONFIG_SCHED_CORE <<
3117 static int task_is_throttled_dl(struct task_s <<
3118 { <<
3119 return p->dl.dl_throttled; <<
3120 } 1943 }
3121 #endif <<
3122 <<
3123 DEFINE_SCHED_CLASS(dl) = { <<
3124 1944
>> 1945 const struct sched_class dl_sched_class = {
>> 1946 .next = &rt_sched_class,
3125 .enqueue_task = enqueue_tas 1947 .enqueue_task = enqueue_task_dl,
3126 .dequeue_task = dequeue_tas 1948 .dequeue_task = dequeue_task_dl,
3127 .yield_task = yield_task_ 1949 .yield_task = yield_task_dl,
3128 1950
3129 .wakeup_preempt = wakeup_pree !! 1951 .check_preempt_curr = check_preempt_curr_dl,
3130 1952
3131 .pick_task = pick_task_d !! 1953 .pick_next_task = pick_next_task_dl,
3132 .put_prev_task = put_prev_ta 1954 .put_prev_task = put_prev_task_dl,
3133 .set_next_task = set_next_ta <<
3134 1955
3135 #ifdef CONFIG_SMP 1956 #ifdef CONFIG_SMP
3136 .balance = balance_dl, <<
3137 .select_task_rq = select_task 1957 .select_task_rq = select_task_rq_dl,
3138 .migrate_task_rq = migrate_tas <<
3139 .set_cpus_allowed = set_cpus_al 1958 .set_cpus_allowed = set_cpus_allowed_dl,
3140 .rq_online = rq_online_d 1959 .rq_online = rq_online_dl,
3141 .rq_offline = rq_offline_ 1960 .rq_offline = rq_offline_dl,
3142 .task_woken = task_woken_ 1961 .task_woken = task_woken_dl,
3143 .find_lock_rq = find_lock_l <<
3144 #endif 1962 #endif
3145 1963
>> 1964 .set_curr_task = set_curr_task_dl,
3146 .task_tick = task_tick_d 1965 .task_tick = task_tick_dl,
3147 .task_fork = task_fork_d 1966 .task_fork = task_fork_dl,
>> 1967 .task_dead = task_dead_dl,
3148 1968
3149 .prio_changed = prio_change 1969 .prio_changed = prio_changed_dl,
3150 .switched_from = switched_fr 1970 .switched_from = switched_from_dl,
3151 .switched_to = switched_to 1971 .switched_to = switched_to_dl,
3152 1972
3153 .update_curr = update_curr 1973 .update_curr = update_curr_dl,
3154 #ifdef CONFIG_SCHED_CORE <<
3155 .task_is_throttled = task_is_thr <<
3156 #endif <<
3157 }; <<
3158 <<
3159 /* Used for dl_bw check and update, used unde <<
3160 static u64 dl_generation; <<
3161 <<
3162 int sched_dl_global_validate(void) <<
3163 { <<
3164 u64 runtime = global_rt_runtime(); <<
3165 u64 period = global_rt_period(); <<
3166 u64 new_bw = to_ratio(period, runtime <<
3167 u64 gen = ++dl_generation; <<
3168 struct dl_bw *dl_b; <<
3169 int cpu, cpus, ret = 0; <<
3170 unsigned long flags; <<
3171 <<
3172 /* <<
3173 * Here we want to check the bandwidt <<
3174 * value smaller than the currently a <<
3175 * any of the root_domains. <<
3176 */ <<
3177 for_each_possible_cpu(cpu) { <<
3178 rcu_read_lock_sched(); <<
3179 <<
3180 if (dl_bw_visited(cpu, gen)) <<
3181 goto next; <<
3182 <<
3183 dl_b = dl_bw_of(cpu); <<
3184 cpus = dl_bw_cpus(cpu); <<
3185 <<
3186 raw_spin_lock_irqsave(&dl_b-> <<
3187 if (new_bw * cpus < dl_b->tot <<
3188 ret = -EBUSY; <<
3189 raw_spin_unlock_irqrestore(&d <<
3190 <<
3191 next: <<
3192 rcu_read_unlock_sched(); <<
3193 <<
3194 if (ret) <<
3195 break; <<
3196 } <<
3197 <<
3198 return ret; <<
3199 } <<
3200 <<
3201 static void init_dl_rq_bw_ratio(struct dl_rq <<
3202 { <<
3203 if (global_rt_runtime() == RUNTIME_IN <<
3204 dl_rq->bw_ratio = 1 << RATIO_ <<
3205 dl_rq->max_bw = dl_rq->extra_ <<
3206 } else { <<
3207 dl_rq->bw_ratio = to_ratio(gl <<
3208 global_rt_period()) <<
3209 dl_rq->max_bw = dl_rq->extra_ <<
3210 to_ratio(global_rt_pe <<
3211 } <<
3212 } <<
3213 <<
3214 void sched_dl_do_global(void) <<
3215 { <<
3216 u64 new_bw = -1; <<
3217 u64 gen = ++dl_generation; <<
3218 struct dl_bw *dl_b; <<
3219 int cpu; <<
3220 unsigned long flags; <<
3221 <<
3222 if (global_rt_runtime() != RUNTIME_IN <<
3223 new_bw = to_ratio(global_rt_p <<
3224 <<
3225 for_each_possible_cpu(cpu) { <<
3226 rcu_read_lock_sched(); <<
3227 <<
3228 if (dl_bw_visited(cpu, gen)) <<
3229 rcu_read_unlock_sched <<
3230 continue; <<
3231 } <<
3232 <<
3233 dl_b = dl_bw_of(cpu); <<
3234 <<
3235 raw_spin_lock_irqsave(&dl_b-> <<
3236 dl_b->bw = new_bw; <<
3237 raw_spin_unlock_irqrestore(&d <<
3238 <<
3239 rcu_read_unlock_sched(); <<
3240 init_dl_rq_bw_ratio(&cpu_rq(c <<
3241 } <<
3242 } <<
3243 <<
3244 /* <<
3245 * We must be sure that accepting a new task <<
3246 * parameters of an existing one) is consiste <<
3247 * constraints. If yes, this function also ac <<
3248 * allocated bandwidth to reflect the new sit <<
3249 * <<
3250 * This function is called while holding p's <<
3251 */ <<
3252 int sched_dl_overflow(struct task_struct *p, <<
3253 const struct sched_attr <<
3254 { <<
3255 u64 period = attr->sched_period ?: at <<
3256 u64 runtime = attr->sched_runtime; <<
3257 u64 new_bw = dl_policy(policy) ? to_r <<
3258 int cpus, err = -1, cpu = task_cpu(p) <<
3259 struct dl_bw *dl_b = dl_bw_of(cpu); <<
3260 unsigned long cap; <<
3261 <<
3262 if (attr->sched_flags & SCHED_FLAG_SU <<
3263 return 0; <<
3264 <<
3265 /* !deadline task may carry old deadl <<
3266 if (new_bw == p->dl.dl_bw && task_has <<
3267 return 0; <<
3268 <<
3269 /* <<
3270 * Either if a task, enters, leave, o <<
3271 * its parameters, we may need to upd <<
3272 * allocated bandwidth of the contain <<
3273 */ <<
3274 raw_spin_lock(&dl_b->lock); <<
3275 cpus = dl_bw_cpus(cpu); <<
3276 cap = dl_bw_capacity(cpu); <<
3277 <<
3278 if (dl_policy(policy) && !task_has_dl <<
3279 !__dl_overflow(dl_b, cap, 0, new_ <<
3280 if (hrtimer_active(&p->dl.ina <<
3281 __dl_sub(dl_b, p->dl. <<
3282 __dl_add(dl_b, new_bw, cpus); <<
3283 err = 0; <<
3284 } else if (dl_policy(policy) && task_ <<
3285 !__dl_overflow(dl_b, cap, <<
3286 /* <<
3287 * XXX this is slightly incor <<
3288 * utilization decreases, we <<
3289 * utilization change until t <<
3290 * But this would require to <<
3291 * timer" when the task is no <<
3292 */ <<
3293 __dl_sub(dl_b, p->dl.dl_bw, c <<
3294 __dl_add(dl_b, new_bw, cpus); <<
3295 dl_change_utilization(p, new_ <<
3296 err = 0; <<
3297 } else if (!dl_policy(policy) && task <<
3298 /* <<
3299 * Do not decrease the total <<
3300 * switched_from_dl() will ta <<
3301 * (0-lag) time. <<
3302 */ <<
3303 err = 0; <<
3304 } <<
3305 raw_spin_unlock(&dl_b->lock); <<
3306 <<
3307 return err; <<
3308 } <<
3309 <<
3310 /* <<
3311 * This function initializes the sched_dl_ent <<
3312 * SCHED_DEADLINE task. <<
3313 * <<
3314 * Only the static values are considered here <<
3315 * absolute deadline will be properly calcula <<
3316 * for the first time with its new policy. <<
3317 */ <<
3318 void __setparam_dl(struct task_struct *p, con <<
3319 { <<
3320 struct sched_dl_entity *dl_se = &p->d <<
3321 <<
3322 dl_se->dl_runtime = attr->sched_runti <<
3323 dl_se->dl_deadline = attr->sched_dead <<
3324 dl_se->dl_period = attr->sched_period <<
3325 dl_se->flags = attr->sched_flags & SC <<
3326 dl_se->dl_bw = to_ratio(dl_se->dl_per <<
3327 dl_se->dl_density = to_ratio(dl_se->d <<
3328 } <<
3329 <<
3330 void __getparam_dl(struct task_struct *p, str <<
3331 { <<
3332 struct sched_dl_entity *dl_se = &p->d <<
3333 <<
3334 attr->sched_priority = p->rt_priority <<
3335 attr->sched_runtime = dl_se->dl_runti <<
3336 attr->sched_deadline = dl_se->dl_dead <<
3337 attr->sched_period = dl_se->dl_period <<
3338 attr->sched_flags &= ~SCHED_DL_FLAGS; <<
3339 attr->sched_flags |= dl_se->flags; <<
3340 } <<
3341 <<
3342 /* <<
3343 * This function validates the new parameters <<
3344 * We ask for the deadline not being zero, an <<
3345 * than the runtime, as well as the period of <<
3346 * greater than deadline. Furthermore, we hav <<
3347 * user parameters are above the internal res <<
3348 * check sched_runtime only since it is alway <<
3349 * below 2^63 ns (we have to check both sched <<
3350 * sched_period, as the latter can be zero). <<
3351 */ <<
3352 bool __checkparam_dl(const struct sched_attr <<
3353 { <<
3354 u64 period, max, min; <<
3355 <<
3356 /* special dl tasks don't actually us <<
3357 if (attr->sched_flags & SCHED_FLAG_SU <<
3358 return true; <<
3359 <<
3360 /* deadline != 0 */ <<
3361 if (attr->sched_deadline == 0) <<
3362 return false; <<
3363 <<
3364 /* <<
3365 * Since we truncate DL_SCALE bits, m <<
3366 * that big. <<
3367 */ <<
3368 if (attr->sched_runtime < (1ULL << DL <<
3369 return false; <<
3370 <<
3371 /* <<
3372 * Since we use the MSB for wrap-arou <<
3373 * sure it's not set (mind that perio <<
3374 */ <<
3375 if (attr->sched_deadline & (1ULL << 6 <<
3376 attr->sched_period & (1ULL << 63) <<
3377 return false; <<
3378 <<
3379 period = attr->sched_period; <<
3380 if (!period) <<
3381 period = attr->sched_deadline <<
3382 <<
3383 /* runtime <= deadline <= period (if <<
3384 if (period < attr->sched_deadline || <<
3385 attr->sched_deadline < attr->sche <<
3386 return false; <<
3387 <<
3388 max = (u64)READ_ONCE(sysctl_sched_dl_ <<
3389 min = (u64)READ_ONCE(sysctl_sched_dl_ <<
3390 <<
3391 if (period < min || period > max) <<
3392 return false; <<
3393 <<
3394 return true; <<
3395 } <<
3396 <<
3397 /* <<
3398 * This function clears the sched_dl_entity s <<
3399 */ <<
3400 static void __dl_clear_params(struct sched_dl <<
3401 { <<
3402 dl_se->dl_runtime = 0; <<
3403 dl_se->dl_deadline = 0; <<
3404 dl_se->dl_period = 0; <<
3405 dl_se->flags = 0; <<
3406 dl_se->dl_bw = 0; <<
3407 dl_se->dl_density = 0; <<
3408 <<
3409 dl_se->dl_throttled = 0; <<
3410 dl_se->dl_yielded = 0; <<
3411 dl_se->dl_non_contending = 0; <<
3412 dl_se->dl_overrun = 0; <<
3413 dl_se->dl_server = 0; <<
3414 <<
3415 #ifdef CONFIG_RT_MUTEXES <<
3416 dl_se->pi_se = dl_ <<
3417 #endif <<
3418 } <<
3419 <<
3420 void init_dl_entity(struct sched_dl_entity *d <<
3421 { <<
3422 RB_CLEAR_NODE(&dl_se->rb_node); <<
3423 init_dl_task_timer(dl_se); <<
3424 init_dl_inactive_task_timer(dl_se); <<
3425 __dl_clear_params(dl_se); <<
3426 } <<
3427 <<
3428 bool dl_param_changed(struct task_struct *p, <<
3429 { <<
3430 struct sched_dl_entity *dl_se = &p->d <<
3431 <<
3432 if (dl_se->dl_runtime != attr->sched_ <<
3433 dl_se->dl_deadline != attr->sched <<
3434 dl_se->dl_period != attr->sched_p <<
3435 dl_se->flags != (attr->sched_flag <<
3436 return true; <<
3437 <<
3438 return false; <<
3439 } <<
3440 <<
3441 #ifdef CONFIG_SMP <<
3442 int dl_cpuset_cpumask_can_shrink(const struct <<
3443 const struct <<
3444 { <<
3445 unsigned long flags, cap; <<
3446 struct dl_bw *cur_dl_b; <<
3447 int ret = 1; <<
3448 <<
3449 rcu_read_lock_sched(); <<
3450 cur_dl_b = dl_bw_of(cpumask_any(cur)) <<
3451 cap = __dl_bw_capacity(trial); <<
3452 raw_spin_lock_irqsave(&cur_dl_b->lock <<
3453 if (__dl_overflow(cur_dl_b, cap, 0, 0 <<
3454 ret = 0; <<
3455 raw_spin_unlock_irqrestore(&cur_dl_b- <<
3456 rcu_read_unlock_sched(); <<
3457 <<
3458 return ret; <<
3459 } <<
3460 <<
3461 enum dl_bw_request { <<
3462 dl_bw_req_check_overflow = 0, <<
3463 dl_bw_req_alloc, <<
3464 dl_bw_req_free <<
3465 }; 1974 };
3466 1975
3467 static int dl_bw_manage(enum dl_bw_request re <<
3468 { <<
3469 unsigned long flags; <<
3470 struct dl_bw *dl_b; <<
3471 bool overflow = 0; <<
3472 <<
3473 rcu_read_lock_sched(); <<
3474 dl_b = dl_bw_of(cpu); <<
3475 raw_spin_lock_irqsave(&dl_b->lock, fl <<
3476 <<
3477 if (req == dl_bw_req_free) { <<
3478 __dl_sub(dl_b, dl_bw, dl_bw_c <<
3479 } else { <<
3480 unsigned long cap = dl_bw_cap <<
3481 <<
3482 overflow = __dl_overflow(dl_b <<
3483 <<
3484 if (req == dl_bw_req_alloc && <<
3485 /* <<
3486 * We reserve space i <<
3487 * root_domain, as we <<
3488 * We will free resou <<
3489 * later on (see set_ <<
3490 */ <<
3491 __dl_add(dl_b, dl_bw, <<
3492 } <<
3493 } <<
3494 <<
3495 raw_spin_unlock_irqrestore(&dl_b->loc <<
3496 rcu_read_unlock_sched(); <<
3497 <<
3498 return overflow ? -EBUSY : 0; <<
3499 } <<
3500 <<
3501 int dl_bw_check_overflow(int cpu) <<
3502 { <<
3503 return dl_bw_manage(dl_bw_req_check_o <<
3504 } <<
3505 <<
3506 int dl_bw_alloc(int cpu, u64 dl_bw) <<
3507 { <<
3508 return dl_bw_manage(dl_bw_req_alloc, <<
3509 } <<
3510 <<
3511 void dl_bw_free(int cpu, u64 dl_bw) <<
3512 { <<
3513 dl_bw_manage(dl_bw_req_free, cpu, dl_ <<
3514 } <<
3515 #endif <<
3516 <<
3517 #ifdef CONFIG_SCHED_DEBUG 1976 #ifdef CONFIG_SCHED_DEBUG
>> 1977 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
>> 1978
3518 void print_dl_stats(struct seq_file *m, int c 1979 void print_dl_stats(struct seq_file *m, int cpu)
3519 { 1980 {
3520 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl) 1981 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3521 } 1982 }
3522 #endif /* CONFIG_SCHED_DEBUG */ 1983 #endif /* CONFIG_SCHED_DEBUG */
3523 1984
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.