1/*
   2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
   3 * policies)
   4 */
   5
   6#include "sched.h"
   7
   8#include <linux/slab.h>
   9#include <linux/irq_work.h>
  10
  11int sched_rr_timeslice = RR_TIMESLICE;
  12
  13static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
  14
  15struct rt_bandwidth def_rt_bandwidth;
  16
  17static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
  18{
  19	struct rt_bandwidth *rt_b =
  20		container_of(timer, struct rt_bandwidth, rt_period_timer);
  21	int idle = 0;
  22	int overrun;
  23
  24	raw_spin_lock(&rt_b->rt_runtime_lock);
  25	for (;;) {
  26		overrun = hrtimer_forward_now(timer, rt_b->rt_period);
  27		if (!overrun)
  28			break;
  29
  30		raw_spin_unlock(&rt_b->rt_runtime_lock);
  31		idle = do_sched_rt_period_timer(rt_b, overrun);
  32		raw_spin_lock(&rt_b->rt_runtime_lock);
  33	}
  34	if (idle)
  35		rt_b->rt_period_active = 0;
  36	raw_spin_unlock(&rt_b->rt_runtime_lock);
  37
  38	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
  39}
  40
  41void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
  42{
  43	rt_b->rt_period = ns_to_ktime(period);
  44	rt_b->rt_runtime = runtime;
  45
  46	raw_spin_lock_init(&rt_b->rt_runtime_lock);
  47
  48	hrtimer_init(&rt_b->rt_period_timer,
  49			CLOCK_MONOTONIC, HRTIMER_MODE_REL);
  50	rt_b->rt_period_timer.function = sched_rt_period_timer;
  51}
  52
  53static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
  54{
  55	if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
  56		return;
  57
  58	raw_spin_lock(&rt_b->rt_runtime_lock);
  59	if (!rt_b->rt_period_active) {
  60		rt_b->rt_period_active = 1;
  61		/*
  62		 * SCHED_DEADLINE updates the bandwidth, as a run away
  63		 * RT task with a DL task could hog a CPU. But DL does
  64		 * not reset the period. If a deadline task was running
  65		 * without an RT task running, it can cause RT tasks to
  66		 * throttle when they start up. Kick the timer right away
  67		 * to update the period.
  68		 */
  69		hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
  70		hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
  71	}
  72	raw_spin_unlock(&rt_b->rt_runtime_lock);
  73}
  74
  75#if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
  76static void push_irq_work_func(struct irq_work *work);
  77#endif
  78
  79void init_rt_rq(struct rt_rq *rt_rq)
  80{
  81	struct rt_prio_array *array;
  82	int i;
  83
  84	array = &rt_rq->active;
  85	for (i = 0; i < MAX_RT_PRIO; i++) {
  86		INIT_LIST_HEAD(array->queue + i);
  87		__clear_bit(i, array->bitmap);
  88	}
  89	/* delimiter for bitsearch: */
  90	__set_bit(MAX_RT_PRIO, array->bitmap);
  91
  92#if defined CONFIG_SMP
  93	rt_rq->highest_prio.curr = MAX_RT_PRIO;
  94	rt_rq->highest_prio.next = MAX_RT_PRIO;
  95	rt_rq->rt_nr_migratory = 0;
  96	rt_rq->overloaded = 0;
  97	plist_head_init(&rt_rq->pushable_tasks);
  98
  99#ifdef HAVE_RT_PUSH_IPI
 100	rt_rq->push_flags = 0;
 101	rt_rq->push_cpu = nr_cpu_ids;
 102	raw_spin_lock_init(&rt_rq->push_lock);
 103	init_irq_work(&rt_rq->push_work, push_irq_work_func);
 104#endif
 105#endif /* CONFIG_SMP */
 106	/* We start is dequeued state, because no RT tasks are queued */
 107	rt_rq->rt_queued = 0;
 108
 109	rt_rq->rt_time = 0;
 110	rt_rq->rt_throttled = 0;
 111	rt_rq->rt_runtime = 0;
 112	raw_spin_lock_init(&rt_rq->rt_runtime_lock);
 113}
 114
 115#ifdef CONFIG_RT_GROUP_SCHED
 116static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 117{
 118	hrtimer_cancel(&rt_b->rt_period_timer);
 119}
 120
 121#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
 122
 123static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
 124{
 125#ifdef CONFIG_SCHED_DEBUG
 126	WARN_ON_ONCE(!rt_entity_is_task(rt_se));
 127#endif
 128	return container_of(rt_se, struct task_struct, rt);
 129}
 130
 131static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 132{
 133	return rt_rq->rq;
 134}
 135
 136static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 137{
 138	return rt_se->rt_rq;
 139}
 140
 141static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
 142{
 143	struct rt_rq *rt_rq = rt_se->rt_rq;
 144
 145	return rt_rq->rq;
 146}
 147
 148void free_rt_sched_group(struct task_group *tg)
 149{
 150	int i;
 151
 152	if (tg->rt_se)
 153		destroy_rt_bandwidth(&tg->rt_bandwidth);
 154
 155	for_each_possible_cpu(i) {
 156		if (tg->rt_rq)
 157			kfree(tg->rt_rq[i]);
 158		if (tg->rt_se)
 159			kfree(tg->rt_se[i]);
 160	}
 161
 162	kfree(tg->rt_rq);
 163	kfree(tg->rt_se);
 164}
 165
 166void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
 167		struct sched_rt_entity *rt_se, int cpu,
 168		struct sched_rt_entity *parent)
 169{
 170	struct rq *rq = cpu_rq(cpu);
 171
 172	rt_rq->highest_prio.curr = MAX_RT_PRIO;
 173	rt_rq->rt_nr_boosted = 0;
 174	rt_rq->rq = rq;
 175	rt_rq->tg = tg;
 176
 177	tg->rt_rq[cpu] = rt_rq;
 178	tg->rt_se[cpu] = rt_se;
 179
 180	if (!rt_se)
 181		return;
 182
 183	if (!parent)
 184		rt_se->rt_rq = &rq->rt;
 185	else
 186		rt_se->rt_rq = parent->my_q;
 187
 188	rt_se->my_q = rt_rq;
 189	rt_se->parent = parent;
 190	INIT_LIST_HEAD(&rt_se->run_list);
 191}
 192
 193int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
 194{
 195	struct rt_rq *rt_rq;
 196	struct sched_rt_entity *rt_se;
 197	int i;
 198
 199	tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
 200	if (!tg->rt_rq)
 201		goto err;
 202	tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
 203	if (!tg->rt_se)
 204		goto err;
 205
 206	init_rt_bandwidth(&tg->rt_bandwidth,
 207			ktime_to_ns(def_rt_bandwidth.rt_period), 0);
 208
 209	for_each_possible_cpu(i) {
 210		rt_rq = kzalloc_node(sizeof(struct rt_rq),
 211				     GFP_KERNEL, cpu_to_node(i));
 212		if (!rt_rq)
 213			goto err;
 214
 215		rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
 216				     GFP_KERNEL, cpu_to_node(i));
 217		if (!rt_se)
 218			goto err_free_rq;
 219
 220		init_rt_rq(rt_rq);
 221		rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
 222		init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
 223	}
 224
 225	return 1;
 226
 227err_free_rq:
 228	kfree(rt_rq);
 229err:
 230	return 0;
 231}
 232
 233#else /* CONFIG_RT_GROUP_SCHED */
 234
 235#define rt_entity_is_task(rt_se) (1)
 236
 237static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
 238{
 239	return container_of(rt_se, struct task_struct, rt);
 240}
 241
 242static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 243{
 244	return container_of(rt_rq, struct rq, rt);
 245}
 246
 247static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
 248{
 249	struct task_struct *p = rt_task_of(rt_se);
 250
 251	return task_rq(p);
 252}
 253
 254static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 255{
 256	struct rq *rq = rq_of_rt_se(rt_se);
 257
 258	return &rq->rt;
 259}
 260
 261void free_rt_sched_group(struct task_group *tg) { }
 262
 263int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
 264{
 265	return 1;
 266}
 267#endif /* CONFIG_RT_GROUP_SCHED */
 268
 269#ifdef CONFIG_SMP
 270
 271static void pull_rt_task(struct rq *this_rq);
 272
 273static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
 274{
 275	/* Try to pull RT tasks here if we lower this rq's prio */
 276	return rq->rt.highest_prio.curr > prev->prio;
 277}
 278
 279static inline int rt_overloaded(struct rq *rq)
 280{
 281	return atomic_read(&rq->rd->rto_count);
 282}
 283
 284static inline void rt_set_overload(struct rq *rq)
 285{
 286	if (!rq->online)
 287		return;
 288
 289	cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
 290	/*
 291	 * Make sure the mask is visible before we set
 292	 * the overload count. That is checked to determine
 293	 * if we should look at the mask. It would be a shame
 294	 * if we looked at the mask, but the mask was not
 295	 * updated yet.
 296	 *
 297	 * Matched by the barrier in pull_rt_task().
 298	 */
 299	smp_wmb();
 300	atomic_inc(&rq->rd->rto_count);
 301}
 302
 303static inline void rt_clear_overload(struct rq *rq)
 304{
 305	if (!rq->online)
 306		return;
 307
 308	/* the order here really doesn't matter */
 309	atomic_dec(&rq->rd->rto_count);
 310	cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
 311}
 312
 313static void update_rt_migration(struct rt_rq *rt_rq)
 314{
 315	if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
 316		if (!rt_rq->overloaded) {
 317			rt_set_overload(rq_of_rt_rq(rt_rq));
 318			rt_rq->overloaded = 1;
 319		}
 320	} else if (rt_rq->overloaded) {
 321		rt_clear_overload(rq_of_rt_rq(rt_rq));
 322		rt_rq->overloaded = 0;
 323	}
 324}
 325
 326static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 327{
 328	struct task_struct *p;
 329
 330	if (!rt_entity_is_task(rt_se))
 331		return;
 332
 333	p = rt_task_of(rt_se);
 334	rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 335
 336	rt_rq->rt_nr_total++;
 337	if (tsk_nr_cpus_allowed(p) > 1)
 338		rt_rq->rt_nr_migratory++;
 339
 340	update_rt_migration(rt_rq);
 341}
 342
 343static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 344{
 345	struct task_struct *p;
 346
 347	if (!rt_entity_is_task(rt_se))
 348		return;
 349
 350	p = rt_task_of(rt_se);
 351	rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 352
 353	rt_rq->rt_nr_total--;
 354	if (tsk_nr_cpus_allowed(p) > 1)
 355		rt_rq->rt_nr_migratory--;
 356
 357	update_rt_migration(rt_rq);
 358}
 359
 360static inline int has_pushable_tasks(struct rq *rq)
 361{
 362	return !plist_head_empty(&rq->rt.pushable_tasks);
 363}
 364
 365static DEFINE_PER_CPU(struct callback_head, rt_push_head);
 366static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
 367
 368static void push_rt_tasks(struct rq *);
 369static void pull_rt_task(struct rq *);
 370
 371static inline void queue_push_tasks(struct rq *rq)
 372{
 373	if (!has_pushable_tasks(rq))
 374		return;
 375
 376	queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
 377}
 378
 379static inline void queue_pull_task(struct rq *rq)
 380{
 381	queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
 382}
 383
 384static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 385{
 386	plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 387	plist_node_init(&p->pushable_tasks, p->prio);
 388	plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
 389
 390	/* Update the highest prio pushable task */
 391	if (p->prio < rq->rt.highest_prio.next)
 392		rq->rt.highest_prio.next = p->prio;
 393}
 394
 395static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 396{
 397	plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 398
 399	/* Update the new highest prio pushable task */
 400	if (has_pushable_tasks(rq)) {
 401		p = plist_first_entry(&rq->rt.pushable_tasks,
 402				      struct task_struct, pushable_tasks);
 403		rq->rt.highest_prio.next = p->prio;
 404	} else
 405		rq->rt.highest_prio.next = MAX_RT_PRIO;
 406}
 407
 408#else
 409
 410static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 411{
 412}
 413
 414static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 415{
 416}
 417
 418static inline
 419void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 420{
 421}
 422
 423static inline
 424void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 425{
 426}
 427
 428static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
 429{
 430	return false;
 431}
 432
 433static inline void pull_rt_task(struct rq *this_rq)
 434{
 435}
 436
 437static inline void queue_push_tasks(struct rq *rq)
 438{
 439}
 440#endif /* CONFIG_SMP */
 441
 442static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
 443static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
 444
 445static inline int on_rt_rq(struct sched_rt_entity *rt_se)
 446{
 447	return rt_se->on_rq;
 448}
 449
 450#ifdef CONFIG_RT_GROUP_SCHED
 451
 452static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 453{
 454	if (!rt_rq->tg)
 455		return RUNTIME_INF;
 456
 457	return rt_rq->rt_runtime;
 458}
 459
 460static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 461{
 462	return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
 463}
 464
 465typedef struct task_group *rt_rq_iter_t;
 466
 467static inline struct task_group *next_task_group(struct task_group *tg)
 468{
 469	do {
 470		tg = list_entry_rcu(tg->list.next,
 471			typeof(struct task_group), list);
 472	} while (&tg->list != &task_groups && task_group_is_autogroup(tg));
 473
 474	if (&tg->list == &task_groups)
 475		tg = NULL;
 476
 477	return tg;
 478}
 479
 480#define for_each_rt_rq(rt_rq, iter, rq)					\
 481	for (iter = container_of(&task_groups, typeof(*iter), list);	\
 482		(iter = next_task_group(iter)) &&			\
 483		(rt_rq = iter->rt_rq[cpu_of(rq)]);)
 484
 485#define for_each_sched_rt_entity(rt_se) \
 486	for (; rt_se; rt_se = rt_se->parent)
 487
 488static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 489{
 490	return rt_se->my_q;
 491}
 492
 493static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
 494static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
 495
 496static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 497{
 498	struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 499	struct rq *rq = rq_of_rt_rq(rt_rq);
 500	struct sched_rt_entity *rt_se;
 501
 502	int cpu = cpu_of(rq);
 503
 504	rt_se = rt_rq->tg->rt_se[cpu];
 505
 506	if (rt_rq->rt_nr_running) {
 507		if (!rt_se)
 508			enqueue_top_rt_rq(rt_rq);
 509		else if (!on_rt_rq(rt_se))
 510			enqueue_rt_entity(rt_se, 0);
 511
 512		if (rt_rq->highest_prio.curr < curr->prio)
 513			resched_curr(rq);
 514	}
 515}
 516
 517static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 518{
 519	struct sched_rt_entity *rt_se;
 520	int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 521
 522	rt_se = rt_rq->tg->rt_se[cpu];
 523
 524	if (!rt_se)
 525		dequeue_top_rt_rq(rt_rq);
 526	else if (on_rt_rq(rt_se))
 527		dequeue_rt_entity(rt_se, 0);
 528}
 529
 530static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 531{
 532	return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 533}
 534
 535static int rt_se_boosted(struct sched_rt_entity *rt_se)
 536{
 537	struct rt_rq *rt_rq = group_rt_rq(rt_se);
 538	struct task_struct *p;
 539
 540	if (rt_rq)
 541		return !!rt_rq->rt_nr_boosted;
 542
 543	p = rt_task_of(rt_se);
 544	return p->prio != p->normal_prio;
 545}
 546
 547#ifdef CONFIG_SMP
 548static inline const struct cpumask *sched_rt_period_mask(void)
 549{
 550	return this_rq()->rd->span;
 551}
 552#else
 553static inline const struct cpumask *sched_rt_period_mask(void)
 554{
 555	return cpu_online_mask;
 556}
 557#endif
 558
 559static inline
 560struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 561{
 562	return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 563}
 564
 565static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 566{
 567	return &rt_rq->tg->rt_bandwidth;
 568}
 569
 570#else /* !CONFIG_RT_GROUP_SCHED */
 571
 572static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 573{
 574	return rt_rq->rt_runtime;
 575}
 576
 577static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 578{
 579	return ktime_to_ns(def_rt_bandwidth.rt_period);
 580}
 581
 582typedef struct rt_rq *rt_rq_iter_t;
 583
 584#define for_each_rt_rq(rt_rq, iter, rq) \
 585	for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 586
 587#define for_each_sched_rt_entity(rt_se) \
 588	for (; rt_se; rt_se = NULL)
 589
 590static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 591{
 592	return NULL;
 593}
 594
 595static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 596{
 597	struct rq *rq = rq_of_rt_rq(rt_rq);
 598
 599	if (!rt_rq->rt_nr_running)
 600		return;
 601
 602	enqueue_top_rt_rq(rt_rq);
 603	resched_curr(rq);
 604}
 605
 606static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 607{
 608	dequeue_top_rt_rq(rt_rq);
 609}
 610
 611static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 612{
 613	return rt_rq->rt_throttled;
 614}
 615
 616static inline const struct cpumask *sched_rt_period_mask(void)
 617{
 618	return cpu_online_mask;
 619}
 620
 621static inline
 622struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 623{
 624	return &cpu_rq(cpu)->rt;
 625}
 626
 627static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 628{
 629	return &def_rt_bandwidth;
 630}
 631
 632#endif /* CONFIG_RT_GROUP_SCHED */
 633
 634bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
 635{
 636	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 637
 638	return (hrtimer_active(&rt_b->rt_period_timer) ||
 639		rt_rq->rt_time < rt_b->rt_runtime);
 640}
 641
 642#ifdef CONFIG_SMP
 643/*
 644 * We ran out of runtime, see if we can borrow some from our neighbours.
 645 */
 646static void do_balance_runtime(struct rt_rq *rt_rq)
 647{
 648	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 649	struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
 650	int i, weight;
 651	u64 rt_period;
 652
 653	weight = cpumask_weight(rd->span);
 654
 655	raw_spin_lock(&rt_b->rt_runtime_lock);
 656	rt_period = ktime_to_ns(rt_b->rt_period);
 657	for_each_cpu(i, rd->span) {
 658		struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 659		s64 diff;
 660
 661		if (iter == rt_rq)
 662			continue;
 663
 664		raw_spin_lock(&iter->rt_runtime_lock);
 665		/*
 666		 * Either all rqs have inf runtime and there's nothing to steal
 667		 * or __disable_runtime() below sets a specific rq to inf to
 668		 * indicate its been disabled and disalow stealing.
 669		 */
 670		if (iter->rt_runtime == RUNTIME_INF)
 671			goto next;
 672
 673		/*
 674		 * From runqueues with spare time, take 1/n part of their
 675		 * spare time, but no more than our period.
 676		 */
 677		diff = iter->rt_runtime - iter->rt_time;
 678		if (diff > 0) {
 679			diff = div_u64((u64)diff, weight);
 680			if (rt_rq->rt_runtime + diff > rt_period)
 681				diff = rt_period - rt_rq->rt_runtime;
 682			iter->rt_runtime -= diff;
 683			rt_rq->rt_runtime += diff;
 684			if (rt_rq->rt_runtime == rt_period) {
 685				raw_spin_unlock(&iter->rt_runtime_lock);
 686				break;
 687			}
 688		}
 689next:
 690		raw_spin_unlock(&iter->rt_runtime_lock);
 691	}
 692	raw_spin_unlock(&rt_b->rt_runtime_lock);
 693}
 694
 695/*
 696 * Ensure this RQ takes back all the runtime it lend to its neighbours.
 697 */
 698static void __disable_runtime(struct rq *rq)
 699{
 700	struct root_domain *rd = rq->rd;
 701	rt_rq_iter_t iter;
 702	struct rt_rq *rt_rq;
 703
 704	if (unlikely(!scheduler_running))
 705		return;
 706
 707	for_each_rt_rq(rt_rq, iter, rq) {
 708		struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 709		s64 want;
 710		int i;
 711
 712		raw_spin_lock(&rt_b->rt_runtime_lock);
 713		raw_spin_lock(&rt_rq->rt_runtime_lock);
 714		/*
 715		 * Either we're all inf and nobody needs to borrow, or we're
 716		 * already disabled and thus have nothing to do, or we have
 717		 * exactly the right amount of runtime to take out.
 718		 */
 719		if (rt_rq->rt_runtime == RUNTIME_INF ||
 720				rt_rq->rt_runtime == rt_b->rt_runtime)
 721			goto balanced;
 722		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 723
 724		/*
 725		 * Calculate the difference between what we started out with
 726		 * and what we current have, that's the amount of runtime
 727		 * we lend and now have to reclaim.
 728		 */
 729		want = rt_b->rt_runtime - rt_rq->rt_runtime;
 730
 731		/*
 732		 * Greedy reclaim, take back as much as we can.
 733		 */
 734		for_each_cpu(i, rd->span) {
 735			struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 736			s64 diff;
 737
 738			/*
 739			 * Can't reclaim from ourselves or disabled runqueues.
 740			 */
 741			if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 742				continue;
 743
 744			raw_spin_lock(&iter->rt_runtime_lock);
 745			if (want > 0) {
 746				diff = min_t(s64, iter->rt_runtime, want);
 747				iter->rt_runtime -= diff;
 748				want -= diff;
 749			} else {
 750				iter->rt_runtime -= want;
 751				want -= want;
 752			}
 753			raw_spin_unlock(&iter->rt_runtime_lock);
 754
 755			if (!want)
 756				break;
 757		}
 758
 759		raw_spin_lock(&rt_rq->rt_runtime_lock);
 760		/*
 761		 * We cannot be left wanting - that would mean some runtime
 762		 * leaked out of the system.
 763		 */
 764		BUG_ON(want);
 765balanced:
 766		/*
 767		 * Disable all the borrow logic by pretending we have inf
 768		 * runtime - in which case borrowing doesn't make sense.
 769		 */
 770		rt_rq->rt_runtime = RUNTIME_INF;
 771		rt_rq->rt_throttled = 0;
 772		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 773		raw_spin_unlock(&rt_b->rt_runtime_lock);
 774
 775		/* Make rt_rq available for pick_next_task() */
 776		sched_rt_rq_enqueue(rt_rq);
 777	}
 778}
 779
 780static void __enable_runtime(struct rq *rq)
 781{
 782	rt_rq_iter_t iter;
 783	struct rt_rq *rt_rq;
 784
 785	if (unlikely(!scheduler_running))
 786		return;
 787
 788	/*
 789	 * Reset each runqueue's bandwidth settings
 790	 */
 791	for_each_rt_rq(rt_rq, iter, rq) {
 792		struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 793
 794		raw_spin_lock(&rt_b->rt_runtime_lock);
 795		raw_spin_lock(&rt_rq->rt_runtime_lock);
 796		rt_rq->rt_runtime = rt_b->rt_runtime;
 797		rt_rq->rt_time = 0;
 798		rt_rq->rt_throttled = 0;
 799		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 800		raw_spin_unlock(&rt_b->rt_runtime_lock);
 801	}
 802}
 803
 804static void balance_runtime(struct rt_rq *rt_rq)
 805{
 806	if (!sched_feat(RT_RUNTIME_SHARE))
 807		return;
 808
 809	if (rt_rq->rt_time > rt_rq->rt_runtime) {
 810		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 811		do_balance_runtime(rt_rq);
 812		raw_spin_lock(&rt_rq->rt_runtime_lock);
 813	}
 814}
 815#else /* !CONFIG_SMP */
 816static inline void balance_runtime(struct rt_rq *rt_rq) {}
 817#endif /* CONFIG_SMP */
 818
 819static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 820{
 821	int i, idle = 1, throttled = 0;
 822	const struct cpumask *span;
 823
 824	span = sched_rt_period_mask();
 825#ifdef CONFIG_RT_GROUP_SCHED
 826	/*
 827	 * FIXME: isolated CPUs should really leave the root task group,
 828	 * whether they are isolcpus or were isolated via cpusets, lest
 829	 * the timer run on a CPU which does not service all runqueues,
 830	 * potentially leaving other CPUs indefinitely throttled.  If
 831	 * isolation is really required, the user will turn the throttle
 832	 * off to kill the perturbations it causes anyway.  Meanwhile,
 833	 * this maintains functionality for boot and/or troubleshooting.
 834	 */
 835	if (rt_b == &root_task_group.rt_bandwidth)
 836		span = cpu_online_mask;
 837#endif
 838	for_each_cpu(i, span) {
 839		int enqueue = 0;
 840		struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 841		struct rq *rq = rq_of_rt_rq(rt_rq);
 842
 843		raw_spin_lock(&rq->lock);
 844		if (rt_rq->rt_time) {
 845			u64 runtime;
 846
 847			raw_spin_lock(&rt_rq->rt_runtime_lock);
 848			if (rt_rq->rt_throttled)
 849				balance_runtime(rt_rq);
 850			runtime = rt_rq->rt_runtime;
 851			rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 852			if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 853				rt_rq->rt_throttled = 0;
 854				enqueue = 1;
 855
 856				/*
 857				 * When we're idle and a woken (rt) task is
 858				 * throttled check_preempt_curr() will set
 859				 * skip_update and the time between the wakeup
 860				 * and this unthrottle will get accounted as
 861				 * 'runtime'.
 862				 */
 863				if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 864					rq_clock_skip_update(rq, false);
 865			}
 866			if (rt_rq->rt_time || rt_rq->rt_nr_running)
 867				idle = 0;
 868			raw_spin_unlock(&rt_rq->rt_runtime_lock);
 869		} else if (rt_rq->rt_nr_running) {
 870			idle = 0;
 871			if (!rt_rq_throttled(rt_rq))
 872				enqueue = 1;
 873		}
 874		if (rt_rq->rt_throttled)
 875			throttled = 1;
 876
 877		if (enqueue)
 878			sched_rt_rq_enqueue(rt_rq);
 879		raw_spin_unlock(&rq->lock);
 880	}
 881
 882	if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
 883		return 1;
 884
 885	return idle;
 886}
 887
 888static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 889{
 890#ifdef CONFIG_RT_GROUP_SCHED
 891	struct rt_rq *rt_rq = group_rt_rq(rt_se);
 892
 893	if (rt_rq)
 894		return rt_rq->highest_prio.curr;
 895#endif
 896
 897	return rt_task_of(rt_se)->prio;
 898}
 899
 900static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 901{
 902	u64 runtime = sched_rt_runtime(rt_rq);
 903
 904	if (rt_rq->rt_throttled)
 905		return rt_rq_throttled(rt_rq);
 906
 907	if (runtime >= sched_rt_period(rt_rq))
 908		return 0;
 909
 910	balance_runtime(rt_rq);
 911	runtime = sched_rt_runtime(rt_rq);
 912	if (runtime == RUNTIME_INF)
 913		return 0;
 914
 915	if (rt_rq->rt_time > runtime) {
 916		struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 917
 918		/*
 919		 * Don't actually throttle groups that have no runtime assigned
 920		 * but accrue some time due to boosting.
 921		 */
 922		if (likely(rt_b->rt_runtime)) {
 923			rt_rq->rt_throttled = 1;
 924			printk_deferred_once("sched: RT throttling activated\n");
 925		} else {
 926			/*
 927			 * In case we did anyway, make it go away,
 928			 * replenishment is a joke, since it will replenish us
 929			 * with exactly 0 ns.
 930			 */
 931			rt_rq->rt_time = 0;
 932		}
 933
 934		if (rt_rq_throttled(rt_rq)) {
 935			sched_rt_rq_dequeue(rt_rq);
 936			return 1;
 937		}
 938	}
 939
 940	return 0;
 941}
 942
 943/*
 944 * Update the current task's runtime statistics. Skip current tasks that
 945 * are not in our scheduling class.
 946 */
 947static void update_curr_rt(struct rq *rq)
 948{
 949	struct task_struct *curr = rq->curr;
 950	struct sched_rt_entity *rt_se = &curr->rt;
 951	u64 delta_exec;
 952
 953	if (curr->sched_class != &rt_sched_class)
 954		return;
 955
 
 
 
 
 956	delta_exec = rq_clock_task(rq) - curr->se.exec_start;
 957	if (unlikely((s64)delta_exec <= 0))
 958		return;
 959
 960	/* Kick cpufreq (see the comment in kernel/sched/sched.h). */
 961	cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_RT);
 962
 963	schedstat_set(curr->se.statistics.exec_max,
 964		      max(curr->se.statistics.exec_max, delta_exec));
 965
 966	curr->se.sum_exec_runtime += delta_exec;
 967	account_group_exec_runtime(curr, delta_exec);
 968
 969	curr->se.exec_start = rq_clock_task(rq);
 970	cpuacct_charge(curr, delta_exec);
 971
 972	sched_rt_avg_update(rq, delta_exec);
 973
 974	if (!rt_bandwidth_enabled())
 975		return;
 976
 977	for_each_sched_rt_entity(rt_se) {
 978		struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 979
 980		if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 981			raw_spin_lock(&rt_rq->rt_runtime_lock);
 982			rt_rq->rt_time += delta_exec;
 983			if (sched_rt_runtime_exceeded(rt_rq))
 984				resched_curr(rq);
 985			raw_spin_unlock(&rt_rq->rt_runtime_lock);
 986		}
 987	}
 988}
 989
 990static void
 991dequeue_top_rt_rq(struct rt_rq *rt_rq)
 992{
 993	struct rq *rq = rq_of_rt_rq(rt_rq);
 994
 995	BUG_ON(&rq->rt != rt_rq);
 996
 997	if (!rt_rq->rt_queued)
 998		return;
 999
1000	BUG_ON(!rq->nr_running);
1001
1002	sub_nr_running(rq, rt_rq->rt_nr_running);
1003	rt_rq->rt_queued = 0;
1004}
1005
1006static void
1007enqueue_top_rt_rq(struct rt_rq *rt_rq)
1008{
1009	struct rq *rq = rq_of_rt_rq(rt_rq);
1010
1011	BUG_ON(&rq->rt != rt_rq);
1012
1013	if (rt_rq->rt_queued)
1014		return;
1015	if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1016		return;
1017
1018	add_nr_running(rq, rt_rq->rt_nr_running);
1019	rt_rq->rt_queued = 1;
1020}
1021
1022#if defined CONFIG_SMP
1023
1024static void
1025inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1026{
1027	struct rq *rq = rq_of_rt_rq(rt_rq);
1028
1029#ifdef CONFIG_RT_GROUP_SCHED
1030	/*
1031	 * Change rq's cpupri only if rt_rq is the top queue.
1032	 */
1033	if (&rq->rt != rt_rq)
1034		return;
1035#endif
1036	if (rq->online && prio < prev_prio)
1037		cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1038}
1039
1040static void
1041dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1042{
1043	struct rq *rq = rq_of_rt_rq(rt_rq);
1044
1045#ifdef CONFIG_RT_GROUP_SCHED
1046	/*
1047	 * Change rq's cpupri only if rt_rq is the top queue.
1048	 */
1049	if (&rq->rt != rt_rq)
1050		return;
1051#endif
1052	if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1053		cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1054}
1055
1056#else /* CONFIG_SMP */
1057
1058static inline
1059void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1060static inline
1061void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1062
1063#endif /* CONFIG_SMP */
1064
1065#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1066static void
1067inc_rt_prio(struct rt_rq *rt_rq, int prio)
1068{
1069	int prev_prio = rt_rq->highest_prio.curr;
1070
1071	if (prio < prev_prio)
1072		rt_rq->highest_prio.curr = prio;
1073
1074	inc_rt_prio_smp(rt_rq, prio, prev_prio);
1075}
1076
1077static void
1078dec_rt_prio(struct rt_rq *rt_rq, int prio)
1079{
1080	int prev_prio = rt_rq->highest_prio.curr;
1081
1082	if (rt_rq->rt_nr_running) {
1083
1084		WARN_ON(prio < prev_prio);
1085
1086		/*
1087		 * This may have been our highest task, and therefore
1088		 * we may have some recomputation to do
1089		 */
1090		if (prio == prev_prio) {
1091			struct rt_prio_array *array = &rt_rq->active;
1092
1093			rt_rq->highest_prio.curr =
1094				sched_find_first_bit(array->bitmap);
1095		}
1096
1097	} else
1098		rt_rq->highest_prio.curr = MAX_RT_PRIO;
1099
1100	dec_rt_prio_smp(rt_rq, prio, prev_prio);
1101}
1102
1103#else
1104
1105static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1106static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1107
1108#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1109
1110#ifdef CONFIG_RT_GROUP_SCHED
1111
1112static void
1113inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1114{
1115	if (rt_se_boosted(rt_se))
1116		rt_rq->rt_nr_boosted++;
1117
1118	if (rt_rq->tg)
1119		start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1120}
1121
1122static void
1123dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1124{
1125	if (rt_se_boosted(rt_se))
1126		rt_rq->rt_nr_boosted--;
1127
1128	WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1129}
1130
1131#else /* CONFIG_RT_GROUP_SCHED */
1132
1133static void
1134inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1135{
1136	start_rt_bandwidth(&def_rt_bandwidth);
1137}
1138
1139static inline
1140void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1141
1142#endif /* CONFIG_RT_GROUP_SCHED */
1143
1144static inline
1145unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1146{
1147	struct rt_rq *group_rq = group_rt_rq(rt_se);
1148
1149	if (group_rq)
1150		return group_rq->rt_nr_running;
1151	else
1152		return 1;
1153}
1154
1155static inline
1156unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1157{
1158	struct rt_rq *group_rq = group_rt_rq(rt_se);
1159	struct task_struct *tsk;
1160
1161	if (group_rq)
1162		return group_rq->rr_nr_running;
1163
1164	tsk = rt_task_of(rt_se);
1165
1166	return (tsk->policy == SCHED_RR) ? 1 : 0;
1167}
1168
1169static inline
1170void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1171{
1172	int prio = rt_se_prio(rt_se);
1173
1174	WARN_ON(!rt_prio(prio));
1175	rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1176	rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1177
1178	inc_rt_prio(rt_rq, prio);
1179	inc_rt_migration(rt_se, rt_rq);
1180	inc_rt_group(rt_se, rt_rq);
1181}
1182
1183static inline
1184void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1185{
1186	WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1187	WARN_ON(!rt_rq->rt_nr_running);
1188	rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1189	rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1190
1191	dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1192	dec_rt_migration(rt_se, rt_rq);
1193	dec_rt_group(rt_se, rt_rq);
1194}
1195
1196/*
1197 * Change rt_se->run_list location unless SAVE && !MOVE
1198 *
1199 * assumes ENQUEUE/DEQUEUE flags match
1200 */
1201static inline bool move_entity(unsigned int flags)
1202{
1203	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1204		return false;
1205
1206	return true;
1207}
1208
1209static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1210{
1211	list_del_init(&rt_se->run_list);
1212
1213	if (list_empty(array->queue + rt_se_prio(rt_se)))
1214		__clear_bit(rt_se_prio(rt_se), array->bitmap);
1215
1216	rt_se->on_list = 0;
1217}
1218
1219static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1220{
1221	struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1222	struct rt_prio_array *array = &rt_rq->active;
1223	struct rt_rq *group_rq = group_rt_rq(rt_se);
1224	struct list_head *queue = array->queue + rt_se_prio(rt_se);
1225
1226	/*
1227	 * Don't enqueue the group if its throttled, or when empty.
1228	 * The latter is a consequence of the former when a child group
1229	 * get throttled and the current group doesn't have any other
1230	 * active members.
1231	 */
1232	if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1233		if (rt_se->on_list)
1234			__delist_rt_entity(rt_se, array);
1235		return;
1236	}
1237
1238	if (move_entity(flags)) {
1239		WARN_ON_ONCE(rt_se->on_list);
1240		if (flags & ENQUEUE_HEAD)
1241			list_add(&rt_se->run_list, queue);
1242		else
1243			list_add_tail(&rt_se->run_list, queue);
1244
1245		__set_bit(rt_se_prio(rt_se), array->bitmap);
1246		rt_se->on_list = 1;
1247	}
1248	rt_se->on_rq = 1;
1249
1250	inc_rt_tasks(rt_se, rt_rq);
1251}
1252
1253static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1254{
1255	struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1256	struct rt_prio_array *array = &rt_rq->active;
1257
1258	if (move_entity(flags)) {
1259		WARN_ON_ONCE(!rt_se->on_list);
1260		__delist_rt_entity(rt_se, array);
1261	}
1262	rt_se->on_rq = 0;
1263
1264	dec_rt_tasks(rt_se, rt_rq);
1265}
1266
1267/*
1268 * Because the prio of an upper entry depends on the lower
1269 * entries, we must remove entries top - down.
1270 */
1271static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1272{
1273	struct sched_rt_entity *back = NULL;
1274
1275	for_each_sched_rt_entity(rt_se) {
1276		rt_se->back = back;
1277		back = rt_se;
1278	}
1279
1280	dequeue_top_rt_rq(rt_rq_of_se(back));
1281
1282	for (rt_se = back; rt_se; rt_se = rt_se->back) {
1283		if (on_rt_rq(rt_se))
1284			__dequeue_rt_entity(rt_se, flags);
1285	}
1286}
1287
1288static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1289{
1290	struct rq *rq = rq_of_rt_se(rt_se);
1291
1292	dequeue_rt_stack(rt_se, flags);
1293	for_each_sched_rt_entity(rt_se)
1294		__enqueue_rt_entity(rt_se, flags);
1295	enqueue_top_rt_rq(&rq->rt);
1296}
1297
1298static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1299{
1300	struct rq *rq = rq_of_rt_se(rt_se);
1301
1302	dequeue_rt_stack(rt_se, flags);
1303
1304	for_each_sched_rt_entity(rt_se) {
1305		struct rt_rq *rt_rq = group_rt_rq(rt_se);
1306
1307		if (rt_rq && rt_rq->rt_nr_running)
1308			__enqueue_rt_entity(rt_se, flags);
1309	}
1310	enqueue_top_rt_rq(&rq->rt);
1311}
1312
1313/*
1314 * Adding/removing a task to/from a priority array:
1315 */
1316static void
1317enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1318{
1319	struct sched_rt_entity *rt_se = &p->rt;
1320
1321	if (flags & ENQUEUE_WAKEUP)
1322		rt_se->timeout = 0;
1323
1324	enqueue_rt_entity(rt_se, flags);
1325
1326	if (!task_current(rq, p) && tsk_nr_cpus_allowed(p) > 1)
1327		enqueue_pushable_task(rq, p);
1328}
1329
1330static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1331{
1332	struct sched_rt_entity *rt_se = &p->rt;
1333
1334	update_curr_rt(rq);
1335	dequeue_rt_entity(rt_se, flags);
1336
1337	dequeue_pushable_task(rq, p);
1338}
1339
1340/*
1341 * Put task to the head or the end of the run list without the overhead of
1342 * dequeue followed by enqueue.
1343 */
1344static void
1345requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1346{
1347	if (on_rt_rq(rt_se)) {
1348		struct rt_prio_array *array = &rt_rq->active;
1349		struct list_head *queue = array->queue + rt_se_prio(rt_se);
1350
1351		if (head)
1352			list_move(&rt_se->run_list, queue);
1353		else
1354			list_move_tail(&rt_se->run_list, queue);
1355	}
1356}
1357
1358static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1359{
1360	struct sched_rt_entity *rt_se = &p->rt;
1361	struct rt_rq *rt_rq;
1362
1363	for_each_sched_rt_entity(rt_se) {
1364		rt_rq = rt_rq_of_se(rt_se);
1365		requeue_rt_entity(rt_rq, rt_se, head);
1366	}
1367}
1368
1369static void yield_task_rt(struct rq *rq)
1370{
1371	requeue_task_rt(rq, rq->curr, 0);
1372}
1373
1374#ifdef CONFIG_SMP
1375static int find_lowest_rq(struct task_struct *task);
1376
1377static int
1378select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1379{
1380	struct task_struct *curr;
1381	struct rq *rq;
1382
1383	/* For anything but wake ups, just return the task_cpu */
1384	if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1385		goto out;
1386
1387	rq = cpu_rq(cpu);
1388
1389	rcu_read_lock();
1390	curr = READ_ONCE(rq->curr); /* unlocked access */
1391
1392	/*
1393	 * If the current task on @p's runqueue is an RT task, then
1394	 * try to see if we can wake this RT task up on another
1395	 * runqueue. Otherwise simply start this RT task
1396	 * on its current runqueue.
1397	 *
1398	 * We want to avoid overloading runqueues. If the woken
1399	 * task is a higher priority, then it will stay on this CPU
1400	 * and the lower prio task should be moved to another CPU.
1401	 * Even though this will probably make the lower prio task
1402	 * lose its cache, we do not want to bounce a higher task
1403	 * around just because it gave up its CPU, perhaps for a
1404	 * lock?
1405	 *
1406	 * For equal prio tasks, we just let the scheduler sort it out.
1407	 *
1408	 * Otherwise, just let it ride on the affined RQ and the
1409	 * post-schedule router will push the preempted task away
1410	 *
1411	 * This test is optimistic, if we get it wrong the load-balancer
1412	 * will have to sort it out.
1413	 */
1414	if (curr && unlikely(rt_task(curr)) &&
1415	    (tsk_nr_cpus_allowed(curr) < 2 ||
1416	     curr->prio <= p->prio)) {
1417		int target = find_lowest_rq(p);
1418
1419		/*
1420		 * Don't bother moving it if the destination CPU is
1421		 * not running a lower priority task.
1422		 */
1423		if (target != -1 &&
1424		    p->prio < cpu_rq(target)->rt.highest_prio.curr)
1425			cpu = target;
1426	}
1427	rcu_read_unlock();
1428
1429out:
1430	return cpu;
1431}
1432
1433static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1434{
1435	/*
1436	 * Current can't be migrated, useless to reschedule,
1437	 * let's hope p can move out.
1438	 */
1439	if (tsk_nr_cpus_allowed(rq->curr) == 1 ||
1440	    !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1441		return;
1442
1443	/*
1444	 * p is migratable, so let's not schedule it and
1445	 * see if it is pushed or pulled somewhere else.
1446	 */
1447	if (tsk_nr_cpus_allowed(p) != 1
1448	    && cpupri_find(&rq->rd->cpupri, p, NULL))
1449		return;
1450
1451	/*
1452	 * There appears to be other cpus that can accept
1453	 * current and none to run 'p', so lets reschedule
1454	 * to try and push current away:
1455	 */
1456	requeue_task_rt(rq, p, 1);
1457	resched_curr(rq);
1458}
1459
1460#endif /* CONFIG_SMP */
1461
1462/*
1463 * Preempt the current task with a newly woken task if needed:
1464 */
1465static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1466{
1467	if (p->prio < rq->curr->prio) {
1468		resched_curr(rq);
1469		return;
1470	}
1471
1472#ifdef CONFIG_SMP
1473	/*
1474	 * If:
1475	 *
1476	 * - the newly woken task is of equal priority to the current task
1477	 * - the newly woken task is non-migratable while current is migratable
1478	 * - current will be preempted on the next reschedule
1479	 *
1480	 * we should check to see if current can readily move to a different
1481	 * cpu.  If so, we will reschedule to allow the push logic to try
1482	 * to move current somewhere else, making room for our non-migratable
1483	 * task.
1484	 */
1485	if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1486		check_preempt_equal_prio(rq, p);
1487#endif
1488}
1489
1490static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1491						   struct rt_rq *rt_rq)
1492{
1493	struct rt_prio_array *array = &rt_rq->active;
1494	struct sched_rt_entity *next = NULL;
1495	struct list_head *queue;
1496	int idx;
1497
1498	idx = sched_find_first_bit(array->bitmap);
1499	BUG_ON(idx >= MAX_RT_PRIO);
1500
1501	queue = array->queue + idx;
1502	next = list_entry(queue->next, struct sched_rt_entity, run_list);
1503
1504	return next;
1505}
1506
1507static struct task_struct *_pick_next_task_rt(struct rq *rq)
1508{
1509	struct sched_rt_entity *rt_se;
1510	struct task_struct *p;
1511	struct rt_rq *rt_rq  = &rq->rt;
1512
1513	do {
1514		rt_se = pick_next_rt_entity(rq, rt_rq);
1515		BUG_ON(!rt_se);
1516		rt_rq = group_rt_rq(rt_se);
1517	} while (rt_rq);
1518
1519	p = rt_task_of(rt_se);
1520	p->se.exec_start = rq_clock_task(rq);
1521
1522	return p;
1523}
1524
1525static struct task_struct *
1526pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
1527{
1528	struct task_struct *p;
1529	struct rt_rq *rt_rq = &rq->rt;
1530
1531	if (need_pull_rt_task(rq, prev)) {
1532		/*
1533		 * This is OK, because current is on_cpu, which avoids it being
1534		 * picked for load-balance and preemption/IRQs are still
1535		 * disabled avoiding further scheduler activity on it and we're
1536		 * being very careful to re-start the picking loop.
1537		 */
1538		lockdep_unpin_lock(&rq->lock, cookie);
1539		pull_rt_task(rq);
1540		lockdep_repin_lock(&rq->lock, cookie);
1541		/*
1542		 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1543		 * means a dl or stop task can slip in, in which case we need
1544		 * to re-start task selection.
1545		 */
1546		if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1547			     rq->dl.dl_nr_running))
1548			return RETRY_TASK;
1549	}
1550
1551	/*
1552	 * We may dequeue prev's rt_rq in put_prev_task().
1553	 * So, we update time before rt_nr_running check.
1554	 */
1555	if (prev->sched_class == &rt_sched_class)
1556		update_curr_rt(rq);
1557
1558	if (!rt_rq->rt_queued)
1559		return NULL;
1560
1561	put_prev_task(rq, prev);
1562
1563	p = _pick_next_task_rt(rq);
1564
1565	/* The running task is never eligible for pushing */
1566	dequeue_pushable_task(rq, p);
1567
1568	queue_push_tasks(rq);
1569
1570	return p;
1571}
1572
1573static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1574{
1575	update_curr_rt(rq);
1576
1577	/*
1578	 * The previous task needs to be made eligible for pushing
1579	 * if it is still active
1580	 */
1581	if (on_rt_rq(&p->rt) && tsk_nr_cpus_allowed(p) > 1)
1582		enqueue_pushable_task(rq, p);
1583}
1584
1585#ifdef CONFIG_SMP
1586
1587/* Only try algorithms three times */
1588#define RT_MAX_TRIES 3
1589
1590static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1591{
1592	if (!task_running(rq, p) &&
1593	    cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1594		return 1;
1595	return 0;
1596}
1597
1598/*
1599 * Return the highest pushable rq's task, which is suitable to be executed
1600 * on the cpu, NULL otherwise
1601 */
1602static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1603{
1604	struct plist_head *head = &rq->rt.pushable_tasks;
1605	struct task_struct *p;
1606
1607	if (!has_pushable_tasks(rq))
1608		return NULL;
1609
1610	plist_for_each_entry(p, head, pushable_tasks) {
1611		if (pick_rt_task(rq, p, cpu))
1612			return p;
1613	}
1614
1615	return NULL;
1616}
1617
1618static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1619
1620static int find_lowest_rq(struct task_struct *task)
1621{
1622	struct sched_domain *sd;
1623	struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1624	int this_cpu = smp_processor_id();
1625	int cpu      = task_cpu(task);
1626
1627	/* Make sure the mask is initialized first */
1628	if (unlikely(!lowest_mask))
1629		return -1;
1630
1631	if (tsk_nr_cpus_allowed(task) == 1)
1632		return -1; /* No other targets possible */
1633
1634	if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1635		return -1; /* No targets found */
1636
1637	/*
1638	 * At this point we have built a mask of cpus representing the
1639	 * lowest priority tasks in the system.  Now we want to elect
1640	 * the best one based on our affinity and topology.
1641	 *
1642	 * We prioritize the last cpu that the task executed on since
1643	 * it is most likely cache-hot in that location.
1644	 */
1645	if (cpumask_test_cpu(cpu, lowest_mask))
1646		return cpu;
1647
1648	/*
1649	 * Otherwise, we consult the sched_domains span maps to figure
1650	 * out which cpu is logically closest to our hot cache data.
1651	 */
1652	if (!cpumask_test_cpu(this_cpu, lowest_mask))
1653		this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1654
1655	rcu_read_lock();
1656	for_each_domain(cpu, sd) {
1657		if (sd->flags & SD_WAKE_AFFINE) {
1658			int best_cpu;
1659
1660			/*
1661			 * "this_cpu" is cheaper to preempt than a
1662			 * remote processor.
1663			 */
1664			if (this_cpu != -1 &&
1665			    cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1666				rcu_read_unlock();
1667				return this_cpu;
1668			}
1669
1670			best_cpu = cpumask_first_and(lowest_mask,
1671						     sched_domain_span(sd));
1672			if (best_cpu < nr_cpu_ids) {
1673				rcu_read_unlock();
1674				return best_cpu;
1675			}
1676		}
1677	}
1678	rcu_read_unlock();
1679
1680	/*
1681	 * And finally, if there were no matches within the domains
1682	 * just give the caller *something* to work with from the compatible
1683	 * locations.
1684	 */
1685	if (this_cpu != -1)
1686		return this_cpu;
1687
1688	cpu = cpumask_any(lowest_mask);
1689	if (cpu < nr_cpu_ids)
1690		return cpu;
1691	return -1;
1692}
1693
1694/* Will lock the rq it finds */
1695static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1696{
1697	struct rq *lowest_rq = NULL;
1698	int tries;
1699	int cpu;
1700
1701	for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1702		cpu = find_lowest_rq(task);
1703
1704		if ((cpu == -1) || (cpu == rq->cpu))
1705			break;
1706
1707		lowest_rq = cpu_rq(cpu);
1708
1709		if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1710			/*
1711			 * Target rq has tasks of equal or higher priority,
1712			 * retrying does not release any lock and is unlikely
1713			 * to yield a different result.
1714			 */
1715			lowest_rq = NULL;
1716			break;
1717		}
1718
1719		/* if the prio of this runqueue changed, try again */
1720		if (double_lock_balance(rq, lowest_rq)) {
1721			/*
1722			 * We had to unlock the run queue. In
1723			 * the mean time, task could have
1724			 * migrated already or had its affinity changed.
1725			 * Also make sure that it wasn't scheduled on its rq.
1726			 */
1727			if (unlikely(task_rq(task) != rq ||
1728				     !cpumask_test_cpu(lowest_rq->cpu,
1729						       tsk_cpus_allowed(task)) ||
1730				     task_running(rq, task) ||
1731				     !rt_task(task) ||
1732				     !task_on_rq_queued(task))) {
1733
1734				double_unlock_balance(rq, lowest_rq);
1735				lowest_rq = NULL;
1736				break;
1737			}
1738		}
1739
1740		/* If this rq is still suitable use it. */
1741		if (lowest_rq->rt.highest_prio.curr > task->prio)
1742			break;
1743
1744		/* try again */
1745		double_unlock_balance(rq, lowest_rq);
1746		lowest_rq = NULL;
1747	}
1748
1749	return lowest_rq;
1750}
1751
1752static struct task_struct *pick_next_pushable_task(struct rq *rq)
1753{
1754	struct task_struct *p;
1755
1756	if (!has_pushable_tasks(rq))
1757		return NULL;
1758
1759	p = plist_first_entry(&rq->rt.pushable_tasks,
1760			      struct task_struct, pushable_tasks);
1761
1762	BUG_ON(rq->cpu != task_cpu(p));
1763	BUG_ON(task_current(rq, p));
1764	BUG_ON(tsk_nr_cpus_allowed(p) <= 1);
1765
1766	BUG_ON(!task_on_rq_queued(p));
1767	BUG_ON(!rt_task(p));
1768
1769	return p;
1770}
1771
1772/*
1773 * If the current CPU has more than one RT task, see if the non
1774 * running task can migrate over to a CPU that is running a task
1775 * of lesser priority.
1776 */
1777static int push_rt_task(struct rq *rq)
1778{
1779	struct task_struct *next_task;
1780	struct rq *lowest_rq;
1781	int ret = 0;
1782
1783	if (!rq->rt.overloaded)
1784		return 0;
1785
1786	next_task = pick_next_pushable_task(rq);
1787	if (!next_task)
1788		return 0;
1789
1790retry:
1791	if (unlikely(next_task == rq->curr)) {
1792		WARN_ON(1);
1793		return 0;
1794	}
1795
1796	/*
1797	 * It's possible that the next_task slipped in of
1798	 * higher priority than current. If that's the case
1799	 * just reschedule current.
1800	 */
1801	if (unlikely(next_task->prio < rq->curr->prio)) {
1802		resched_curr(rq);
1803		return 0;
1804	}
1805
1806	/* We might release rq lock */
1807	get_task_struct(next_task);
1808
1809	/* find_lock_lowest_rq locks the rq if found */
1810	lowest_rq = find_lock_lowest_rq(next_task, rq);
1811	if (!lowest_rq) {
1812		struct task_struct *task;
1813		/*
1814		 * find_lock_lowest_rq releases rq->lock
1815		 * so it is possible that next_task has migrated.
1816		 *
1817		 * We need to make sure that the task is still on the same
1818		 * run-queue and is also still the next task eligible for
1819		 * pushing.
1820		 */
1821		task = pick_next_pushable_task(rq);
1822		if (task_cpu(next_task) == rq->cpu && task == next_task) {
1823			/*
1824			 * The task hasn't migrated, and is still the next
1825			 * eligible task, but we failed to find a run-queue
1826			 * to push it to.  Do not retry in this case, since
1827			 * other cpus will pull from us when ready.
1828			 */
1829			goto out;
1830		}
1831
1832		if (!task)
1833			/* No more tasks, just exit */
1834			goto out;
1835
1836		/*
1837		 * Something has shifted, try again.
1838		 */
1839		put_task_struct(next_task);
1840		next_task = task;
1841		goto retry;
1842	}
1843
1844	deactivate_task(rq, next_task, 0);
1845	set_task_cpu(next_task, lowest_rq->cpu);
1846	activate_task(lowest_rq, next_task, 0);
1847	ret = 1;
1848
1849	resched_curr(lowest_rq);
1850
1851	double_unlock_balance(rq, lowest_rq);
1852
1853out:
1854	put_task_struct(next_task);
1855
1856	return ret;
1857}
1858
1859static void push_rt_tasks(struct rq *rq)
1860{
1861	/* push_rt_task will return true if it moved an RT */
1862	while (push_rt_task(rq))
1863		;
1864}
1865
1866#ifdef HAVE_RT_PUSH_IPI
1867/*
1868 * The search for the next cpu always starts at rq->cpu and ends
1869 * when we reach rq->cpu again. It will never return rq->cpu.
1870 * This returns the next cpu to check, or nr_cpu_ids if the loop
1871 * is complete.
1872 *
1873 * rq->rt.push_cpu holds the last cpu returned by this function,
1874 * or if this is the first instance, it must hold rq->cpu.
1875 */
1876static int rto_next_cpu(struct rq *rq)
1877{
1878	int prev_cpu = rq->rt.push_cpu;
1879	int cpu;
1880
1881	cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1882
1883	/*
1884	 * If the previous cpu is less than the rq's CPU, then it already
1885	 * passed the end of the mask, and has started from the beginning.
1886	 * We end if the next CPU is greater or equal to rq's CPU.
1887	 */
1888	if (prev_cpu < rq->cpu) {
1889		if (cpu >= rq->cpu)
1890			return nr_cpu_ids;
1891
1892	} else if (cpu >= nr_cpu_ids) {
1893		/*
1894		 * We passed the end of the mask, start at the beginning.
1895		 * If the result is greater or equal to the rq's CPU, then
1896		 * the loop is finished.
1897		 */
1898		cpu = cpumask_first(rq->rd->rto_mask);
1899		if (cpu >= rq->cpu)
1900			return nr_cpu_ids;
1901	}
1902	rq->rt.push_cpu = cpu;
1903
1904	/* Return cpu to let the caller know if the loop is finished or not */
1905	return cpu;
1906}
1907
1908static int find_next_push_cpu(struct rq *rq)
1909{
1910	struct rq *next_rq;
1911	int cpu;
1912
1913	while (1) {
1914		cpu = rto_next_cpu(rq);
1915		if (cpu >= nr_cpu_ids)
1916			break;
1917		next_rq = cpu_rq(cpu);
1918
1919		/* Make sure the next rq can push to this rq */
1920		if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1921			break;
1922	}
1923
1924	return cpu;
1925}
1926
1927#define RT_PUSH_IPI_EXECUTING		1
1928#define RT_PUSH_IPI_RESTART		2
1929
1930static void tell_cpu_to_push(struct rq *rq)
1931{
1932	int cpu;
1933
1934	if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1935		raw_spin_lock(&rq->rt.push_lock);
1936		/* Make sure it's still executing */
1937		if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1938			/*
1939			 * Tell the IPI to restart the loop as things have
1940			 * changed since it started.
1941			 */
1942			rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1943			raw_spin_unlock(&rq->rt.push_lock);
1944			return;
1945		}
1946		raw_spin_unlock(&rq->rt.push_lock);
1947	}
1948
1949	/* When here, there's no IPI going around */
1950
1951	rq->rt.push_cpu = rq->cpu;
1952	cpu = find_next_push_cpu(rq);
1953	if (cpu >= nr_cpu_ids)
1954		return;
1955
1956	rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1957
1958	irq_work_queue_on(&rq->rt.push_work, cpu);
1959}
1960
1961/* Called from hardirq context */
1962static void try_to_push_tasks(void *arg)
1963{
1964	struct rt_rq *rt_rq = arg;
1965	struct rq *rq, *src_rq;
1966	int this_cpu;
1967	int cpu;
1968
1969	this_cpu = rt_rq->push_cpu;
1970
1971	/* Paranoid check */
1972	BUG_ON(this_cpu != smp_processor_id());
1973
1974	rq = cpu_rq(this_cpu);
1975	src_rq = rq_of_rt_rq(rt_rq);
1976
1977again:
1978	if (has_pushable_tasks(rq)) {
1979		raw_spin_lock(&rq->lock);
1980		push_rt_task(rq);
1981		raw_spin_unlock(&rq->lock);
1982	}
1983
1984	/* Pass the IPI to the next rt overloaded queue */
1985	raw_spin_lock(&rt_rq->push_lock);
1986	/*
1987	 * If the source queue changed since the IPI went out,
1988	 * we need to restart the search from that CPU again.
1989	 */
1990	if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1991		rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1992		rt_rq->push_cpu = src_rq->cpu;
1993	}
1994
1995	cpu = find_next_push_cpu(src_rq);
1996
1997	if (cpu >= nr_cpu_ids)
1998		rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
1999	raw_spin_unlock(&rt_rq->push_lock);
2000
2001	if (cpu >= nr_cpu_ids)
2002		return;
2003
2004	/*
2005	 * It is possible that a restart caused this CPU to be
2006	 * chosen again. Don't bother with an IPI, just see if we
2007	 * have more to push.
2008	 */
2009	if (unlikely(cpu == rq->cpu))
2010		goto again;
2011
2012	/* Try the next RT overloaded CPU */
2013	irq_work_queue_on(&rt_rq->push_work, cpu);
2014}
2015
2016static void push_irq_work_func(struct irq_work *work)
2017{
2018	struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2019
2020	try_to_push_tasks(rt_rq);
2021}
2022#endif /* HAVE_RT_PUSH_IPI */
2023
2024static void pull_rt_task(struct rq *this_rq)
2025{
2026	int this_cpu = this_rq->cpu, cpu;
2027	bool resched = false;
2028	struct task_struct *p;
2029	struct rq *src_rq;
2030
2031	if (likely(!rt_overloaded(this_rq)))
2032		return;
2033
2034	/*
2035	 * Match the barrier from rt_set_overloaded; this guarantees that if we
2036	 * see overloaded we must also see the rto_mask bit.
2037	 */
2038	smp_rmb();
2039
2040#ifdef HAVE_RT_PUSH_IPI
2041	if (sched_feat(RT_PUSH_IPI)) {
2042		tell_cpu_to_push(this_rq);
2043		return;
2044	}
2045#endif
2046
2047	for_each_cpu(cpu, this_rq->rd->rto_mask) {
2048		if (this_cpu == cpu)
2049			continue;
2050
2051		src_rq = cpu_rq(cpu);
2052
2053		/*
2054		 * Don't bother taking the src_rq->lock if the next highest
2055		 * task is known to be lower-priority than our current task.
2056		 * This may look racy, but if this value is about to go
2057		 * logically higher, the src_rq will push this task away.
2058		 * And if its going logically lower, we do not care
2059		 */
2060		if (src_rq->rt.highest_prio.next >=
2061		    this_rq->rt.highest_prio.curr)
2062			continue;
2063
2064		/*
2065		 * We can potentially drop this_rq's lock in
2066		 * double_lock_balance, and another CPU could
2067		 * alter this_rq
2068		 */
2069		double_lock_balance(this_rq, src_rq);
2070
2071		/*
2072		 * We can pull only a task, which is pushable
2073		 * on its rq, and no others.
2074		 */
2075		p = pick_highest_pushable_task(src_rq, this_cpu);
2076
2077		/*
2078		 * Do we have an RT task that preempts
2079		 * the to-be-scheduled task?
2080		 */
2081		if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2082			WARN_ON(p == src_rq->curr);
2083			WARN_ON(!task_on_rq_queued(p));
2084
2085			/*
2086			 * There's a chance that p is higher in priority
2087			 * than what's currently running on its cpu.
2088			 * This is just that p is wakeing up and hasn't
2089			 * had a chance to schedule. We only pull
2090			 * p if it is lower in priority than the
2091			 * current task on the run queue
2092			 */
2093			if (p->prio < src_rq->curr->prio)
2094				goto skip;
2095
2096			resched = true;
2097
2098			deactivate_task(src_rq, p, 0);
2099			set_task_cpu(p, this_cpu);
2100			activate_task(this_rq, p, 0);
2101			/*
2102			 * We continue with the search, just in
2103			 * case there's an even higher prio task
2104			 * in another runqueue. (low likelihood
2105			 * but possible)
2106			 */
2107		}
2108skip:
2109		double_unlock_balance(this_rq, src_rq);
2110	}
2111
2112	if (resched)
2113		resched_curr(this_rq);
2114}
2115
2116/*
2117 * If we are not running and we are not going to reschedule soon, we should
2118 * try to push tasks away now
2119 */
2120static void task_woken_rt(struct rq *rq, struct task_struct *p)
2121{
2122	if (!task_running(rq, p) &&
2123	    !test_tsk_need_resched(rq->curr) &&
2124	    tsk_nr_cpus_allowed(p) > 1 &&
2125	    (dl_task(rq->curr) || rt_task(rq->curr)) &&
2126	    (tsk_nr_cpus_allowed(rq->curr) < 2 ||
2127	     rq->curr->prio <= p->prio))
2128		push_rt_tasks(rq);
2129}
2130
2131/* Assumes rq->lock is held */
2132static void rq_online_rt(struct rq *rq)
2133{
2134	if (rq->rt.overloaded)
2135		rt_set_overload(rq);
2136
2137	__enable_runtime(rq);
2138
2139	cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2140}
2141
2142/* Assumes rq->lock is held */
2143static void rq_offline_rt(struct rq *rq)
2144{
2145	if (rq->rt.overloaded)
2146		rt_clear_overload(rq);
2147
2148	__disable_runtime(rq);
2149
2150	cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2151}
2152
2153/*
2154 * When switch from the rt queue, we bring ourselves to a position
2155 * that we might want to pull RT tasks from other runqueues.
2156 */
2157static void switched_from_rt(struct rq *rq, struct task_struct *p)
2158{
2159	/*
2160	 * If there are other RT tasks then we will reschedule
2161	 * and the scheduling of the other RT tasks will handle
2162	 * the balancing. But if we are the last RT task
2163	 * we may need to handle the pulling of RT tasks
2164	 * now.
2165	 */
2166	if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2167		return;
2168
2169	queue_pull_task(rq);
2170}
2171
2172void __init init_sched_rt_class(void)
2173{
2174	unsigned int i;
2175
2176	for_each_possible_cpu(i) {
2177		zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2178					GFP_KERNEL, cpu_to_node(i));
2179	}
2180}
2181#endif /* CONFIG_SMP */
2182
2183/*
2184 * When switching a task to RT, we may overload the runqueue
2185 * with RT tasks. In this case we try to push them off to
2186 * other runqueues.
2187 */
2188static void switched_to_rt(struct rq *rq, struct task_struct *p)
2189{
2190	/*
2191	 * If we are already running, then there's nothing
2192	 * that needs to be done. But if we are not running
2193	 * we may need to preempt the current running task.
2194	 * If that current running task is also an RT task
2195	 * then see if we can move to another run queue.
2196	 */
2197	if (task_on_rq_queued(p) && rq->curr != p) {
2198#ifdef CONFIG_SMP
2199		if (tsk_nr_cpus_allowed(p) > 1 && rq->rt.overloaded)
2200			queue_push_tasks(rq);
2201#endif /* CONFIG_SMP */
2202		if (p->prio < rq->curr->prio)
2203			resched_curr(rq);
 
2204	}
2205}
2206
2207/*
2208 * Priority of the task has changed. This may cause
2209 * us to initiate a push or pull.
2210 */
2211static void
2212prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2213{
2214	if (!task_on_rq_queued(p))
2215		return;
2216
2217	if (rq->curr == p) {
2218#ifdef CONFIG_SMP
2219		/*
2220		 * If our priority decreases while running, we
2221		 * may need to pull tasks to this runqueue.
2222		 */
2223		if (oldprio < p->prio)
2224			queue_pull_task(rq);
2225
2226		/*
2227		 * If there's a higher priority task waiting to run
2228		 * then reschedule.
2229		 */
2230		if (p->prio > rq->rt.highest_prio.curr)
2231			resched_curr(rq);
2232#else
2233		/* For UP simply resched on drop of prio */
2234		if (oldprio < p->prio)
2235			resched_curr(rq);
2236#endif /* CONFIG_SMP */
2237	} else {
2238		/*
2239		 * This task is not running, but if it is
2240		 * greater than the current running task
2241		 * then reschedule.
2242		 */
2243		if (p->prio < rq->curr->prio)
2244			resched_curr(rq);
2245	}
2246}
2247
2248static void watchdog(struct rq *rq, struct task_struct *p)
2249{
2250	unsigned long soft, hard;
2251
2252	/* max may change after cur was read, this will be fixed next tick */
2253	soft = task_rlimit(p, RLIMIT_RTTIME);
2254	hard = task_rlimit_max(p, RLIMIT_RTTIME);
2255
2256	if (soft != RLIM_INFINITY) {
2257		unsigned long next;
2258
2259		if (p->rt.watchdog_stamp != jiffies) {
2260			p->rt.timeout++;
2261			p->rt.watchdog_stamp = jiffies;
2262		}
2263
2264		next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2265		if (p->rt.timeout > next)
2266			p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2267	}
2268}
2269
2270static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2271{
2272	struct sched_rt_entity *rt_se = &p->rt;
2273
2274	update_curr_rt(rq);
2275
2276	watchdog(rq, p);
2277
2278	/*
2279	 * RR tasks need a special form of timeslice management.
2280	 * FIFO tasks have no timeslices.
2281	 */
2282	if (p->policy != SCHED_RR)
2283		return;
2284
2285	if (--p->rt.time_slice)
2286		return;
2287
2288	p->rt.time_slice = sched_rr_timeslice;
2289
2290	/*
2291	 * Requeue to the end of queue if we (and all of our ancestors) are not
2292	 * the only element on the queue
2293	 */
2294	for_each_sched_rt_entity(rt_se) {
2295		if (rt_se->run_list.prev != rt_se->run_list.next) {
2296			requeue_task_rt(rq, p, 0);
2297			resched_curr(rq);
2298			return;
2299		}
2300	}
2301}
2302
2303static void set_curr_task_rt(struct rq *rq)
2304{
2305	struct task_struct *p = rq->curr;
2306
2307	p->se.exec_start = rq_clock_task(rq);
2308
2309	/* The running task is never eligible for pushing */
2310	dequeue_pushable_task(rq, p);
2311}
2312
2313static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2314{
2315	/*
2316	 * Time slice is 0 for SCHED_FIFO tasks
2317	 */
2318	if (task->policy == SCHED_RR)
2319		return sched_rr_timeslice;
2320	else
2321		return 0;
2322}
2323
2324const struct sched_class rt_sched_class = {
2325	.next			= &fair_sched_class,
2326	.enqueue_task		= enqueue_task_rt,
2327	.dequeue_task		= dequeue_task_rt,
2328	.yield_task		= yield_task_rt,
2329
2330	.check_preempt_curr	= check_preempt_curr_rt,
2331
2332	.pick_next_task		= pick_next_task_rt,
2333	.put_prev_task		= put_prev_task_rt,
2334
2335#ifdef CONFIG_SMP
2336	.select_task_rq		= select_task_rq_rt,
2337
2338	.set_cpus_allowed       = set_cpus_allowed_common,
2339	.rq_online              = rq_online_rt,
2340	.rq_offline             = rq_offline_rt,
2341	.task_woken		= task_woken_rt,
2342	.switched_from		= switched_from_rt,
2343#endif
2344
2345	.set_curr_task          = set_curr_task_rt,
2346	.task_tick		= task_tick_rt,
2347
2348	.get_rr_interval	= get_rr_interval_rt,
2349
2350	.prio_changed		= prio_changed_rt,
2351	.switched_to		= switched_to_rt,
2352
2353	.update_curr		= update_curr_rt,
2354};
2355
2356#ifdef CONFIG_SCHED_DEBUG
2357extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2358
2359void print_rt_stats(struct seq_file *m, int cpu)
2360{
2361	rt_rq_iter_t iter;
2362	struct rt_rq *rt_rq;
2363
2364	rcu_read_lock();
2365	for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2366		print_rt_rq(m, cpu, rt_rq);
2367	rcu_read_unlock();
2368}
2369#endif /* CONFIG_SCHED_DEBUG */