1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Scheduler topology setup/handling methods
   4 */
   5#include "sched.h"
 
   6
   7DEFINE_MUTEX(sched_domains_mutex);
   8
   9/* Protected by sched_domains_mutex: */
  10static cpumask_var_t sched_domains_tmpmask;
  11static cpumask_var_t sched_domains_tmpmask2;
  12
  13#ifdef CONFIG_SCHED_DEBUG
  14
  15static int __init sched_debug_setup(char *str)
  16{
  17	sched_debug_enabled = true;
  18
  19	return 0;
  20}
  21early_param("sched_debug", sched_debug_setup);
  22
  23static inline bool sched_debug(void)
  24{
  25	return sched_debug_enabled;
  26}
  27
 
 
 
 
 
 
  28static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
  29				  struct cpumask *groupmask)
  30{
  31	struct sched_group *group = sd->groups;
 
 
  32
  33	cpumask_clear(groupmask);
  34
  35	printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
  36
  37	if (!(sd->flags & SD_LOAD_BALANCE)) {
  38		printk("does not load-balance\n");
  39		if (sd->parent)
  40			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
  41		return -1;
  42	}
  43
  44	printk(KERN_CONT "span=%*pbl level=%s\n",
  45	       cpumask_pr_args(sched_domain_span(sd)), sd->name);
  46
  47	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
  48		printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
  49	}
  50	if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
  51		printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
  52	}
  53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  54	printk(KERN_DEBUG "%*s groups:", level + 1, "");
  55	do {
  56		if (!group) {
  57			printk("\n");
  58			printk(KERN_ERR "ERROR: group is NULL\n");
  59			break;
  60		}
  61
  62		if (!cpumask_weight(sched_group_span(group))) {
  63			printk(KERN_CONT "\n");
  64			printk(KERN_ERR "ERROR: empty group\n");
  65			break;
  66		}
  67
  68		if (!(sd->flags & SD_OVERLAP) &&
  69		    cpumask_intersects(groupmask, sched_group_span(group))) {
  70			printk(KERN_CONT "\n");
  71			printk(KERN_ERR "ERROR: repeated CPUs\n");
  72			break;
  73		}
  74
  75		cpumask_or(groupmask, groupmask, sched_group_span(group));
  76
  77		printk(KERN_CONT " %d:{ span=%*pbl",
  78				group->sgc->id,
  79				cpumask_pr_args(sched_group_span(group)));
  80
  81		if ((sd->flags & SD_OVERLAP) &&
  82		    !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
  83			printk(KERN_CONT " mask=%*pbl",
  84				cpumask_pr_args(group_balance_mask(group)));
  85		}
  86
  87		if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
  88			printk(KERN_CONT " cap=%lu", group->sgc->capacity);
  89
  90		if (group == sd->groups && sd->child &&
  91		    !cpumask_equal(sched_domain_span(sd->child),
  92				   sched_group_span(group))) {
  93			printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
  94		}
  95
  96		printk(KERN_CONT " }");
  97
  98		group = group->next;
  99
 100		if (group != sd->groups)
 101			printk(KERN_CONT ",");
 102
 103	} while (group != sd->groups);
 104	printk(KERN_CONT "\n");
 105
 106	if (!cpumask_equal(sched_domain_span(sd), groupmask))
 107		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 108
 109	if (sd->parent &&
 110	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
 111		printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
 112	return 0;
 113}
 114
 115static void sched_domain_debug(struct sched_domain *sd, int cpu)
 116{
 117	int level = 0;
 118
 119	if (!sched_debug_enabled)
 120		return;
 121
 122	if (!sd) {
 123		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
 124		return;
 125	}
 126
 127	printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
 128
 129	for (;;) {
 130		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
 131			break;
 132		level++;
 133		sd = sd->parent;
 134		if (!sd)
 135			break;
 136	}
 137}
 138#else /* !CONFIG_SCHED_DEBUG */
 139
 140# define sched_debug_enabled 0
 141# define sched_domain_debug(sd, cpu) do { } while (0)
 142static inline bool sched_debug(void)
 143{
 144	return false;
 145}
 146#endif /* CONFIG_SCHED_DEBUG */
 147
 
 
 
 
 
 
 
 148static int sd_degenerate(struct sched_domain *sd)
 149{
 150	if (cpumask_weight(sched_domain_span(sd)) == 1)
 151		return 1;
 152
 153	/* Following flags need at least 2 groups */
 154	if (sd->flags & (SD_LOAD_BALANCE |
 155			 SD_BALANCE_NEWIDLE |
 156			 SD_BALANCE_FORK |
 157			 SD_BALANCE_EXEC |
 158			 SD_SHARE_CPUCAPACITY |
 159			 SD_ASYM_CPUCAPACITY |
 160			 SD_SHARE_PKG_RESOURCES |
 161			 SD_SHARE_POWERDOMAIN)) {
 162		if (sd->groups != sd->groups->next)
 163			return 0;
 164	}
 165
 166	/* Following flags don't use groups */
 167	if (sd->flags & (SD_WAKE_AFFINE))
 168		return 0;
 169
 170	return 1;
 171}
 172
 173static int
 174sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
 175{
 176	unsigned long cflags = sd->flags, pflags = parent->flags;
 177
 178	if (sd_degenerate(parent))
 179		return 1;
 180
 181	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
 182		return 0;
 183
 184	/* Flags needing groups don't count if only 1 group in parent */
 185	if (parent->groups == parent->groups->next) {
 186		pflags &= ~(SD_LOAD_BALANCE |
 187				SD_BALANCE_NEWIDLE |
 188				SD_BALANCE_FORK |
 189				SD_BALANCE_EXEC |
 190				SD_ASYM_CPUCAPACITY |
 191				SD_SHARE_CPUCAPACITY |
 192				SD_SHARE_PKG_RESOURCES |
 193				SD_PREFER_SIBLING |
 194				SD_SHARE_POWERDOMAIN);
 195		if (nr_node_ids == 1)
 196			pflags &= ~SD_SERIALIZE;
 197	}
 198	if (~cflags & pflags)
 199		return 0;
 200
 201	return 1;
 202}
 203
 204#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
 205DEFINE_STATIC_KEY_FALSE(sched_energy_present);
 206unsigned int sysctl_sched_energy_aware = 1;
 207DEFINE_MUTEX(sched_energy_mutex);
 208bool sched_energy_update;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 209
 210#ifdef CONFIG_PROC_SYSCTL
 211int sched_energy_aware_handler(struct ctl_table *table, int write,
 212			 void __user *buffer, size_t *lenp, loff_t *ppos)
 213{
 214	int ret, state;
 215
 216	if (write && !capable(CAP_SYS_ADMIN))
 217		return -EPERM;
 218
 
 
 
 
 
 
 
 
 
 219	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 220	if (!ret && write) {
 221		state = static_branch_unlikely(&sched_energy_present);
 222		if (state != sysctl_sched_energy_aware) {
 223			mutex_lock(&sched_energy_mutex);
 224			sched_energy_update = 1;
 225			rebuild_sched_domains();
 226			sched_energy_update = 0;
 227			mutex_unlock(&sched_energy_mutex);
 228		}
 229	}
 230
 231	return ret;
 232}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 233#endif
 234
 235static void free_pd(struct perf_domain *pd)
 236{
 237	struct perf_domain *tmp;
 238
 239	while (pd) {
 240		tmp = pd->next;
 241		kfree(pd);
 242		pd = tmp;
 243	}
 244}
 245
 246static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
 247{
 248	while (pd) {
 249		if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
 250			return pd;
 251		pd = pd->next;
 252	}
 253
 254	return NULL;
 255}
 256
 257static struct perf_domain *pd_init(int cpu)
 258{
 259	struct em_perf_domain *obj = em_cpu_get(cpu);
 260	struct perf_domain *pd;
 261
 262	if (!obj) {
 263		if (sched_debug())
 264			pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
 265		return NULL;
 266	}
 267
 268	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 269	if (!pd)
 270		return NULL;
 271	pd->em_pd = obj;
 272
 273	return pd;
 274}
 275
 276static void perf_domain_debug(const struct cpumask *cpu_map,
 277						struct perf_domain *pd)
 278{
 279	if (!sched_debug() || !pd)
 280		return;
 281
 282	printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
 283
 284	while (pd) {
 285		printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_cstate=%d }",
 286				cpumask_first(perf_domain_span(pd)),
 287				cpumask_pr_args(perf_domain_span(pd)),
 288				em_pd_nr_cap_states(pd->em_pd));
 289		pd = pd->next;
 290	}
 291
 292	printk(KERN_CONT "\n");
 293}
 294
 295static void destroy_perf_domain_rcu(struct rcu_head *rp)
 296{
 297	struct perf_domain *pd;
 298
 299	pd = container_of(rp, struct perf_domain, rcu);
 300	free_pd(pd);
 301}
 302
 303static void sched_energy_set(bool has_eas)
 304{
 305	if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
 306		if (sched_debug())
 307			pr_info("%s: stopping EAS\n", __func__);
 308		static_branch_disable_cpuslocked(&sched_energy_present);
 309	} else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
 310		if (sched_debug())
 311			pr_info("%s: starting EAS\n", __func__);
 312		static_branch_enable_cpuslocked(&sched_energy_present);
 313	}
 314}
 315
 316/*
 317 * EAS can be used on a root domain if it meets all the following conditions:
 318 *    1. an Energy Model (EM) is available;
 319 *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
 320 *    3. the EM complexity is low enough to keep scheduling overheads low;
 321 *    4. schedutil is driving the frequency of all CPUs of the rd;
 322 *
 323 * The complexity of the Energy Model is defined as:
 324 *
 325 *              C = nr_pd * (nr_cpus + nr_cs)
 326 *
 327 * with parameters defined as:
 328 *  - nr_pd:    the number of performance domains
 329 *  - nr_cpus:  the number of CPUs
 330 *  - nr_cs:    the sum of the number of capacity states of all performance
 331 *              domains (for example, on a system with 2 performance domains,
 332 *              with 10 capacity states each, nr_cs = 2 * 10 = 20).
 333 *
 334 * It is generally not a good idea to use such a model in the wake-up path on
 335 * very complex platforms because of the associated scheduling overheads. The
 336 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
 337 * with per-CPU DVFS and less than 8 capacity states each, for example.
 338 */
 339#define EM_MAX_COMPLEXITY 2048
 340
 341extern struct cpufreq_governor schedutil_gov;
 342static bool build_perf_domains(const struct cpumask *cpu_map)
 343{
 344	int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
 345	struct perf_domain *pd = NULL, *tmp;
 346	int cpu = cpumask_first(cpu_map);
 347	struct root_domain *rd = cpu_rq(cpu)->rd;
 348	struct cpufreq_policy *policy;
 349	struct cpufreq_governor *gov;
 350
 351	if (!sysctl_sched_energy_aware)
 352		goto free;
 353
 354	/* EAS is enabled for asymmetric CPU capacity topologies. */
 355	if (!per_cpu(sd_asym_cpucapacity, cpu)) {
 356		if (sched_debug()) {
 357			pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
 358					cpumask_pr_args(cpu_map));
 359		}
 360		goto free;
 361	}
 362
 363	for_each_cpu(i, cpu_map) {
 364		/* Skip already covered CPUs. */
 365		if (find_pd(pd, i))
 366			continue;
 367
 368		/* Do not attempt EAS if schedutil is not being used. */
 369		policy = cpufreq_cpu_get(i);
 370		if (!policy)
 371			goto free;
 372		gov = policy->governor;
 373		cpufreq_cpu_put(policy);
 374		if (gov != &schedutil_gov) {
 375			if (rd->pd)
 376				pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
 377						cpumask_pr_args(cpu_map));
 378			goto free;
 379		}
 380
 381		/* Create the new pd and add it to the local list. */
 382		tmp = pd_init(i);
 383		if (!tmp)
 384			goto free;
 385		tmp->next = pd;
 386		pd = tmp;
 387
 388		/*
 389		 * Count performance domains and capacity states for the
 390		 * complexity check.
 391		 */
 392		nr_pd++;
 393		nr_cs += em_pd_nr_cap_states(pd->em_pd);
 394	}
 395
 396	/* Bail out if the Energy Model complexity is too high. */
 397	if (nr_pd * (nr_cs + nr_cpus) > EM_MAX_COMPLEXITY) {
 398		WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
 399						cpumask_pr_args(cpu_map));
 400		goto free;
 401	}
 402
 403	perf_domain_debug(cpu_map, pd);
 404
 405	/* Attach the new list of performance domains to the root domain. */
 406	tmp = rd->pd;
 407	rcu_assign_pointer(rd->pd, pd);
 408	if (tmp)
 409		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 410
 411	return !!pd;
 412
 413free:
 414	free_pd(pd);
 415	tmp = rd->pd;
 416	rcu_assign_pointer(rd->pd, NULL);
 417	if (tmp)
 418		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 419
 420	return false;
 421}
 422#else
 423static void free_pd(struct perf_domain *pd) { }
 424#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
 425
 426static void free_rootdomain(struct rcu_head *rcu)
 427{
 428	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 429
 430	cpupri_cleanup(&rd->cpupri);
 431	cpudl_cleanup(&rd->cpudl);
 432	free_cpumask_var(rd->dlo_mask);
 433	free_cpumask_var(rd->rto_mask);
 434	free_cpumask_var(rd->online);
 435	free_cpumask_var(rd->span);
 436	free_pd(rd->pd);
 437	kfree(rd);
 438}
 439
 440void rq_attach_root(struct rq *rq, struct root_domain *rd)
 441{
 442	struct root_domain *old_rd = NULL;
 443	unsigned long flags;
 444
 445	raw_spin_lock_irqsave(&rq->lock, flags);
 446
 447	if (rq->rd) {
 448		old_rd = rq->rd;
 449
 450		if (cpumask_test_cpu(rq->cpu, old_rd->online))
 451			set_rq_offline(rq);
 452
 453		cpumask_clear_cpu(rq->cpu, old_rd->span);
 454
 455		/*
 456		 * If we dont want to free the old_rd yet then
 457		 * set old_rd to NULL to skip the freeing later
 458		 * in this function:
 459		 */
 460		if (!atomic_dec_and_test(&old_rd->refcount))
 461			old_rd = NULL;
 462	}
 463
 464	atomic_inc(&rd->refcount);
 465	rq->rd = rd;
 466
 467	cpumask_set_cpu(rq->cpu, rd->span);
 468	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 469		set_rq_online(rq);
 470
 471	raw_spin_unlock_irqrestore(&rq->lock, flags);
 472
 473	if (old_rd)
 474		call_rcu(&old_rd->rcu, free_rootdomain);
 475}
 476
 477void sched_get_rd(struct root_domain *rd)
 478{
 479	atomic_inc(&rd->refcount);
 480}
 481
 482void sched_put_rd(struct root_domain *rd)
 483{
 484	if (!atomic_dec_and_test(&rd->refcount))
 485		return;
 486
 487	call_rcu(&rd->rcu, free_rootdomain);
 488}
 489
 490static int init_rootdomain(struct root_domain *rd)
 491{
 492	if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 493		goto out;
 494	if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 495		goto free_span;
 496	if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 497		goto free_online;
 498	if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 499		goto free_dlo_mask;
 500
 501#ifdef HAVE_RT_PUSH_IPI
 502	rd->rto_cpu = -1;
 503	raw_spin_lock_init(&rd->rto_lock);
 504	init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
 505#endif
 506
 
 507	init_dl_bw(&rd->dl_bw);
 508	if (cpudl_init(&rd->cpudl) != 0)
 509		goto free_rto_mask;
 510
 511	if (cpupri_init(&rd->cpupri) != 0)
 512		goto free_cpudl;
 513	return 0;
 514
 515free_cpudl:
 516	cpudl_cleanup(&rd->cpudl);
 517free_rto_mask:
 518	free_cpumask_var(rd->rto_mask);
 519free_dlo_mask:
 520	free_cpumask_var(rd->dlo_mask);
 521free_online:
 522	free_cpumask_var(rd->online);
 523free_span:
 524	free_cpumask_var(rd->span);
 525out:
 526	return -ENOMEM;
 527}
 528
 529/*
 530 * By default the system creates a single root-domain with all CPUs as
 531 * members (mimicking the global state we have today).
 532 */
 533struct root_domain def_root_domain;
 534
 535void init_defrootdomain(void)
 536{
 537	init_rootdomain(&def_root_domain);
 538
 539	atomic_set(&def_root_domain.refcount, 1);
 540}
 541
 542static struct root_domain *alloc_rootdomain(void)
 543{
 544	struct root_domain *rd;
 545
 546	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 547	if (!rd)
 548		return NULL;
 549
 550	if (init_rootdomain(rd) != 0) {
 551		kfree(rd);
 552		return NULL;
 553	}
 554
 555	return rd;
 556}
 557
 558static void free_sched_groups(struct sched_group *sg, int free_sgc)
 559{
 560	struct sched_group *tmp, *first;
 561
 562	if (!sg)
 563		return;
 564
 565	first = sg;
 566	do {
 567		tmp = sg->next;
 568
 569		if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 570			kfree(sg->sgc);
 571
 572		if (atomic_dec_and_test(&sg->ref))
 573			kfree(sg);
 574		sg = tmp;
 575	} while (sg != first);
 576}
 577
 578static void destroy_sched_domain(struct sched_domain *sd)
 579{
 580	/*
 581	 * A normal sched domain may have multiple group references, an
 582	 * overlapping domain, having private groups, only one.  Iterate,
 583	 * dropping group/capacity references, freeing where none remain.
 584	 */
 585	free_sched_groups(sd->groups, 1);
 586
 587	if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 588		kfree(sd->shared);
 589	kfree(sd);
 590}
 591
 592static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 593{
 594	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 595
 596	while (sd) {
 597		struct sched_domain *parent = sd->parent;
 598		destroy_sched_domain(sd);
 599		sd = parent;
 600	}
 601}
 602
 603static void destroy_sched_domains(struct sched_domain *sd)
 604{
 605	if (sd)
 606		call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 607}
 608
 609/*
 610 * Keep a special pointer to the highest sched_domain that has
 611 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
 612 * allows us to avoid some pointer chasing select_idle_sibling().
 613 *
 614 * Also keep a unique ID per domain (we use the first CPU number in
 615 * the cpumask of the domain), this allows us to quickly tell if
 616 * two CPUs are in the same cache domain, see cpus_share_cache().
 617 */
 618DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 619DEFINE_PER_CPU(int, sd_llc_size);
 620DEFINE_PER_CPU(int, sd_llc_id);
 
 621DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 622DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 623DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 624DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 
 625DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 
 626
 627static void update_top_cache_domain(int cpu)
 628{
 629	struct sched_domain_shared *sds = NULL;
 630	struct sched_domain *sd;
 631	int id = cpu;
 632	int size = 1;
 633
 634	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 635	if (sd) {
 636		id = cpumask_first(sched_domain_span(sd));
 637		size = cpumask_weight(sched_domain_span(sd));
 638		sds = sd->shared;
 639	}
 640
 641	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 642	per_cpu(sd_llc_size, cpu) = size;
 643	per_cpu(sd_llc_id, cpu) = id;
 644	rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 645
 
 
 
 
 
 
 
 
 
 
 
 646	sd = lowest_flag_domain(cpu, SD_NUMA);
 647	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 648
 649	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 650	rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 651
 652	sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
 653	rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 654}
 655
 656/*
 657 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 658 * hold the hotplug lock.
 659 */
 660static void
 661cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 662{
 663	struct rq *rq = cpu_rq(cpu);
 664	struct sched_domain *tmp;
 665
 666	/* Remove the sched domains which do not contribute to scheduling. */
 667	for (tmp = sd; tmp; ) {
 668		struct sched_domain *parent = tmp->parent;
 669		if (!parent)
 670			break;
 671
 672		if (sd_parent_degenerate(tmp, parent)) {
 673			tmp->parent = parent->parent;
 674			if (parent->parent)
 
 675				parent->parent->child = tmp;
 
 
 
 676			/*
 677			 * Transfer SD_PREFER_SIBLING down in case of a
 678			 * degenerate parent; the spans match for this
 679			 * so the property transfers.
 680			 */
 681			if (parent->flags & SD_PREFER_SIBLING)
 682				tmp->flags |= SD_PREFER_SIBLING;
 683			destroy_sched_domain(parent);
 684		} else
 685			tmp = tmp->parent;
 686	}
 687
 688	if (sd && sd_degenerate(sd)) {
 689		tmp = sd;
 690		sd = sd->parent;
 691		destroy_sched_domain(tmp);
 692		if (sd)
 
 
 
 
 
 
 
 
 
 
 
 693			sd->child = NULL;
 
 694	}
 695
 696	sched_domain_debug(sd, cpu);
 697
 698	rq_attach_root(rq, rd);
 699	tmp = rq->sd;
 700	rcu_assign_pointer(rq->sd, sd);
 701	dirty_sched_domain_sysctl(cpu);
 702	destroy_sched_domains(tmp);
 703
 704	update_top_cache_domain(cpu);
 705}
 706
 707struct s_data {
 708	struct sched_domain * __percpu *sd;
 709	struct root_domain	*rd;
 710};
 711
 712enum s_alloc {
 713	sa_rootdomain,
 714	sa_sd,
 715	sa_sd_storage,
 716	sa_none,
 717};
 718
 719/*
 720 * Return the canonical balance CPU for this group, this is the first CPU
 721 * of this group that's also in the balance mask.
 722 *
 723 * The balance mask are all those CPUs that could actually end up at this
 724 * group. See build_balance_mask().
 725 *
 726 * Also see should_we_balance().
 727 */
 728int group_balance_cpu(struct sched_group *sg)
 729{
 730	return cpumask_first(group_balance_mask(sg));
 731}
 732
 733
 734/*
 735 * NUMA topology (first read the regular topology blurb below)
 736 *
 737 * Given a node-distance table, for example:
 738 *
 739 *   node   0   1   2   3
 740 *     0:  10  20  30  20
 741 *     1:  20  10  20  30
 742 *     2:  30  20  10  20
 743 *     3:  20  30  20  10
 744 *
 745 * which represents a 4 node ring topology like:
 746 *
 747 *   0 ----- 1
 748 *   |       |
 749 *   |       |
 750 *   |       |
 751 *   3 ----- 2
 752 *
 753 * We want to construct domains and groups to represent this. The way we go
 754 * about doing this is to build the domains on 'hops'. For each NUMA level we
 755 * construct the mask of all nodes reachable in @level hops.
 756 *
 757 * For the above NUMA topology that gives 3 levels:
 758 *
 759 * NUMA-2	0-3		0-3		0-3		0-3
 760 *  groups:	{0-1,3},{1-3}	{0-2},{0,2-3}	{1-3},{0-1,3}	{0,2-3},{0-2}
 761 *
 762 * NUMA-1	0-1,3		0-2		1-3		0,2-3
 763 *  groups:	{0},{1},{3}	{0},{1},{2}	{1},{2},{3}	{0},{2},{3}
 764 *
 765 * NUMA-0	0		1		2		3
 766 *
 767 *
 768 * As can be seen; things don't nicely line up as with the regular topology.
 769 * When we iterate a domain in child domain chunks some nodes can be
 770 * represented multiple times -- hence the "overlap" naming for this part of
 771 * the topology.
 772 *
 773 * In order to minimize this overlap, we only build enough groups to cover the
 774 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 775 *
 776 * Because:
 777 *
 778 *  - the first group of each domain is its child domain; this
 779 *    gets us the first 0-1,3
 780 *  - the only uncovered node is 2, who's child domain is 1-3.
 781 *
 782 * However, because of the overlap, computing a unique CPU for each group is
 783 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 784 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 785 * end up at those groups (they would end up in group: 0-1,3).
 786 *
 787 * To correct this we have to introduce the group balance mask. This mask
 788 * will contain those CPUs in the group that can reach this group given the
 789 * (child) domain tree.
 790 *
 791 * With this we can once again compute balance_cpu and sched_group_capacity
 792 * relations.
 793 *
 794 * XXX include words on how balance_cpu is unique and therefore can be
 795 * used for sched_group_capacity links.
 796 *
 797 *
 798 * Another 'interesting' topology is:
 799 *
 800 *   node   0   1   2   3
 801 *     0:  10  20  20  30
 802 *     1:  20  10  20  20
 803 *     2:  20  20  10  20
 804 *     3:  30  20  20  10
 805 *
 806 * Which looks a little like:
 807 *
 808 *   0 ----- 1
 809 *   |     / |
 810 *   |   /   |
 811 *   | /     |
 812 *   2 ----- 3
 813 *
 814 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 815 * are not.
 816 *
 817 * This leads to a few particularly weird cases where the sched_domain's are
 818 * not of the same number for each CPU. Consider:
 819 *
 820 * NUMA-2	0-3						0-3
 821 *  groups:	{0-2},{1-3}					{1-3},{0-2}
 822 *
 823 * NUMA-1	0-2		0-3		0-3		1-3
 824 *
 825 * NUMA-0	0		1		2		3
 826 *
 827 */
 828
 829
 830/*
 831 * Build the balance mask; it contains only those CPUs that can arrive at this
 832 * group and should be considered to continue balancing.
 833 *
 834 * We do this during the group creation pass, therefore the group information
 835 * isn't complete yet, however since each group represents a (child) domain we
 836 * can fully construct this using the sched_domain bits (which are already
 837 * complete).
 838 */
 839static void
 840build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 841{
 842	const struct cpumask *sg_span = sched_group_span(sg);
 843	struct sd_data *sdd = sd->private;
 844	struct sched_domain *sibling;
 845	int i;
 846
 847	cpumask_clear(mask);
 848
 849	for_each_cpu(i, sg_span) {
 850		sibling = *per_cpu_ptr(sdd->sd, i);
 851
 852		/*
 853		 * Can happen in the asymmetric case, where these siblings are
 854		 * unused. The mask will not be empty because those CPUs that
 855		 * do have the top domain _should_ span the domain.
 856		 */
 857		if (!sibling->child)
 858			continue;
 859
 860		/* If we would not end up here, we can't continue from here */
 861		if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 862			continue;
 863
 864		cpumask_set_cpu(i, mask);
 865	}
 866
 867	/* We must not have empty masks here */
 868	WARN_ON_ONCE(cpumask_empty(mask));
 869}
 870
 871/*
 872 * XXX: This creates per-node group entries; since the load-balancer will
 873 * immediately access remote memory to construct this group's load-balance
 874 * statistics having the groups node local is of dubious benefit.
 875 */
 876static struct sched_group *
 877build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 878{
 879	struct sched_group *sg;
 880	struct cpumask *sg_span;
 881
 882	sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 883			GFP_KERNEL, cpu_to_node(cpu));
 884
 885	if (!sg)
 886		return NULL;
 887
 888	sg_span = sched_group_span(sg);
 889	if (sd->child)
 890		cpumask_copy(sg_span, sched_domain_span(sd->child));
 891	else
 
 892		cpumask_copy(sg_span, sched_domain_span(sd));
 
 893
 894	atomic_inc(&sg->ref);
 895	return sg;
 896}
 897
 898static void init_overlap_sched_group(struct sched_domain *sd,
 899				     struct sched_group *sg)
 900{
 901	struct cpumask *mask = sched_domains_tmpmask2;
 902	struct sd_data *sdd = sd->private;
 903	struct cpumask *sg_span;
 904	int cpu;
 905
 906	build_balance_mask(sd, sg, mask);
 907	cpu = cpumask_first_and(sched_group_span(sg), mask);
 908
 909	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 910	if (atomic_inc_return(&sg->sgc->ref) == 1)
 911		cpumask_copy(group_balance_mask(sg), mask);
 912	else
 913		WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 914
 915	/*
 916	 * Initialize sgc->capacity such that even if we mess up the
 917	 * domains and no possible iteration will get us here, we won't
 918	 * die on a /0 trap.
 919	 */
 920	sg_span = sched_group_span(sg);
 921	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 922	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 923	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 924}
 925
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 926static int
 927build_overlap_sched_groups(struct sched_domain *sd, int cpu)
 928{
 929	struct sched_group *first = NULL, *last = NULL, *sg;
 930	const struct cpumask *span = sched_domain_span(sd);
 931	struct cpumask *covered = sched_domains_tmpmask;
 932	struct sd_data *sdd = sd->private;
 933	struct sched_domain *sibling;
 934	int i;
 935
 936	cpumask_clear(covered);
 937
 938	for_each_cpu_wrap(i, span, cpu) {
 939		struct cpumask *sg_span;
 940
 941		if (cpumask_test_cpu(i, covered))
 942			continue;
 943
 944		sibling = *per_cpu_ptr(sdd->sd, i);
 945
 946		/*
 947		 * Asymmetric node setups can result in situations where the
 948		 * domain tree is of unequal depth, make sure to skip domains
 949		 * that already cover the entire range.
 950		 *
 951		 * In that case build_sched_domains() will have terminated the
 952		 * iteration early and our sibling sd spans will be empty.
 953		 * Domains should always include the CPU they're built on, so
 954		 * check that.
 955		 */
 956		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
 957			continue;
 958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 959		sg = build_group_from_child_sched_domain(sibling, cpu);
 960		if (!sg)
 961			goto fail;
 962
 963		sg_span = sched_group_span(sg);
 964		cpumask_or(covered, covered, sg_span);
 965
 966		init_overlap_sched_group(sd, sg);
 967
 968		if (!first)
 969			first = sg;
 970		if (last)
 971			last->next = sg;
 972		last = sg;
 973		last->next = first;
 974	}
 975	sd->groups = first;
 976
 977	return 0;
 978
 979fail:
 980	free_sched_groups(first, 0);
 981
 982	return -ENOMEM;
 983}
 984
 985
 986/*
 987 * Package topology (also see the load-balance blurb in fair.c)
 988 *
 989 * The scheduler builds a tree structure to represent a number of important
 990 * topology features. By default (default_topology[]) these include:
 991 *
 992 *  - Simultaneous multithreading (SMT)
 993 *  - Multi-Core Cache (MC)
 994 *  - Package (DIE)
 995 *
 996 * Where the last one more or less denotes everything up to a NUMA node.
 997 *
 998 * The tree consists of 3 primary data structures:
 999 *
1000 *	sched_domain -> sched_group -> sched_group_capacity
1001 *	    ^ ^             ^ ^
1002 *          `-'             `-'
1003 *
1004 * The sched_domains are per-CPU and have a two way link (parent & child) and
1005 * denote the ever growing mask of CPUs belonging to that level of topology.
1006 *
1007 * Each sched_domain has a circular (double) linked list of sched_group's, each
1008 * denoting the domains of the level below (or individual CPUs in case of the
1009 * first domain level). The sched_group linked by a sched_domain includes the
1010 * CPU of that sched_domain [*].
1011 *
1012 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1013 *
1014 * CPU   0   1   2   3   4   5   6   7
1015 *
1016 * DIE  [                             ]
1017 * MC   [             ] [             ]
1018 * SMT  [     ] [     ] [     ] [     ]
1019 *
1020 *  - or -
1021 *
1022 * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1023 * MC	0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1024 * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1025 *
1026 * CPU   0   1   2   3   4   5   6   7
1027 *
1028 * One way to think about it is: sched_domain moves you up and down among these
1029 * topology levels, while sched_group moves you sideways through it, at child
1030 * domain granularity.
1031 *
1032 * sched_group_capacity ensures each unique sched_group has shared storage.
1033 *
1034 * There are two related construction problems, both require a CPU that
1035 * uniquely identify each group (for a given domain):
1036 *
1037 *  - The first is the balance_cpu (see should_we_balance() and the
1038 *    load-balance blub in fair.c); for each group we only want 1 CPU to
1039 *    continue balancing at a higher domain.
1040 *
1041 *  - The second is the sched_group_capacity; we want all identical groups
1042 *    to share a single sched_group_capacity.
1043 *
1044 * Since these topologies are exclusive by construction. That is, its
1045 * impossible for an SMT thread to belong to multiple cores, and cores to
1046 * be part of multiple caches. There is a very clear and unique location
1047 * for each CPU in the hierarchy.
1048 *
1049 * Therefore computing a unique CPU for each group is trivial (the iteration
1050 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1051 * group), we can simply pick the first CPU in each group.
1052 *
1053 *
1054 * [*] in other words, the first group of each domain is its child domain.
1055 */
1056
1057static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1058{
1059	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1060	struct sched_domain *child = sd->child;
1061	struct sched_group *sg;
1062	bool already_visited;
1063
1064	if (child)
1065		cpu = cpumask_first(sched_domain_span(child));
1066
1067	sg = *per_cpu_ptr(sdd->sg, cpu);
1068	sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1069
1070	/* Increase refcounts for claim_allocations: */
1071	already_visited = atomic_inc_return(&sg->ref) > 1;
1072	/* sgc visits should follow a similar trend as sg */
1073	WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1074
1075	/* If we have already visited that group, it's already initialized. */
1076	if (already_visited)
1077		return sg;
1078
1079	if (child) {
1080		cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1081		cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
 
1082	} else {
1083		cpumask_set_cpu(cpu, sched_group_span(sg));
1084		cpumask_set_cpu(cpu, group_balance_mask(sg));
1085	}
1086
1087	sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1088	sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1089	sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1090
1091	return sg;
1092}
1093
1094/*
1095 * build_sched_groups will build a circular linked list of the groups
1096 * covered by the given span, will set each group's ->cpumask correctly,
1097 * and will initialize their ->sgc.
1098 *
1099 * Assumes the sched_domain tree is fully constructed
1100 */
1101static int
1102build_sched_groups(struct sched_domain *sd, int cpu)
1103{
1104	struct sched_group *first = NULL, *last = NULL;
1105	struct sd_data *sdd = sd->private;
1106	const struct cpumask *span = sched_domain_span(sd);
1107	struct cpumask *covered;
1108	int i;
1109
1110	lockdep_assert_held(&sched_domains_mutex);
1111	covered = sched_domains_tmpmask;
1112
1113	cpumask_clear(covered);
1114
1115	for_each_cpu_wrap(i, span, cpu) {
1116		struct sched_group *sg;
1117
1118		if (cpumask_test_cpu(i, covered))
1119			continue;
1120
1121		sg = get_group(i, sdd);
1122
1123		cpumask_or(covered, covered, sched_group_span(sg));
1124
1125		if (!first)
1126			first = sg;
1127		if (last)
1128			last->next = sg;
1129		last = sg;
1130	}
1131	last->next = first;
1132	sd->groups = first;
1133
1134	return 0;
1135}
1136
1137/*
1138 * Initialize sched groups cpu_capacity.
1139 *
1140 * cpu_capacity indicates the capacity of sched group, which is used while
1141 * distributing the load between different sched groups in a sched domain.
1142 * Typically cpu_capacity for all the groups in a sched domain will be same
1143 * unless there are asymmetries in the topology. If there are asymmetries,
1144 * group having more cpu_capacity will pickup more load compared to the
1145 * group having less cpu_capacity.
1146 */
1147static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1148{
1149	struct sched_group *sg = sd->groups;
 
1150
1151	WARN_ON(!sg);
1152
1153	do {
1154		int cpu, max_cpu = -1;
1155
1156		sg->group_weight = cpumask_weight(sched_group_span(sg));
1157
 
 
 
 
 
 
 
 
 
1158		if (!(sd->flags & SD_ASYM_PACKING))
1159			goto next;
1160
1161		for_each_cpu(cpu, sched_group_span(sg)) {
1162			if (max_cpu < 0)
1163				max_cpu = cpu;
1164			else if (sched_asym_prefer(cpu, max_cpu))
1165				max_cpu = cpu;
1166		}
1167		sg->asym_prefer_cpu = max_cpu;
1168
1169next:
1170		sg = sg->next;
1171	} while (sg != sd->groups);
1172
1173	if (cpu != group_balance_cpu(sg))
1174		return;
1175
1176	update_group_capacity(sd, cpu);
1177}
1178
1179/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1180 * Initializers for schedule domains
1181 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1182 */
1183
1184static int default_relax_domain_level = -1;
1185int sched_domain_level_max;
1186
1187static int __init setup_relax_domain_level(char *str)
1188{
1189	if (kstrtoint(str, 0, &default_relax_domain_level))
1190		pr_warn("Unable to set relax_domain_level\n");
1191
1192	return 1;
1193}
1194__setup("relax_domain_level=", setup_relax_domain_level);
1195
1196static void set_domain_attribute(struct sched_domain *sd,
1197				 struct sched_domain_attr *attr)
1198{
1199	int request;
1200
1201	if (!attr || attr->relax_domain_level < 0) {
1202		if (default_relax_domain_level < 0)
1203			return;
1204		else
1205			request = default_relax_domain_level;
1206	} else
1207		request = attr->relax_domain_level;
1208	if (request < sd->level) {
 
1209		/* Turn off idle balance on this domain: */
1210		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1211	} else {
1212		/* Turn on idle balance on this domain: */
1213		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1214	}
1215}
1216
1217static void __sdt_free(const struct cpumask *cpu_map);
1218static int __sdt_alloc(const struct cpumask *cpu_map);
1219
1220static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1221				 const struct cpumask *cpu_map)
1222{
1223	switch (what) {
1224	case sa_rootdomain:
1225		if (!atomic_read(&d->rd->refcount))
1226			free_rootdomain(&d->rd->rcu);
1227		/* Fall through */
1228	case sa_sd:
1229		free_percpu(d->sd);
1230		/* Fall through */
1231	case sa_sd_storage:
1232		__sdt_free(cpu_map);
1233		/* Fall through */
1234	case sa_none:
1235		break;
1236	}
1237}
1238
1239static enum s_alloc
1240__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1241{
1242	memset(d, 0, sizeof(*d));
1243
1244	if (__sdt_alloc(cpu_map))
1245		return sa_sd_storage;
1246	d->sd = alloc_percpu(struct sched_domain *);
1247	if (!d->sd)
1248		return sa_sd_storage;
1249	d->rd = alloc_rootdomain();
1250	if (!d->rd)
1251		return sa_sd;
1252
1253	return sa_rootdomain;
1254}
1255
1256/*
1257 * NULL the sd_data elements we've used to build the sched_domain and
1258 * sched_group structure so that the subsequent __free_domain_allocs()
1259 * will not free the data we're using.
1260 */
1261static void claim_allocations(int cpu, struct sched_domain *sd)
1262{
1263	struct sd_data *sdd = sd->private;
1264
1265	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1266	*per_cpu_ptr(sdd->sd, cpu) = NULL;
1267
1268	if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1269		*per_cpu_ptr(sdd->sds, cpu) = NULL;
1270
1271	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1272		*per_cpu_ptr(sdd->sg, cpu) = NULL;
1273
1274	if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1275		*per_cpu_ptr(sdd->sgc, cpu) = NULL;
1276}
1277
1278#ifdef CONFIG_NUMA
1279enum numa_topology_type sched_numa_topology_type;
1280
1281static int			sched_domains_numa_levels;
1282static int			sched_domains_curr_level;
1283
1284int				sched_max_numa_distance;
1285static int			*sched_domains_numa_distance;
1286static struct cpumask		***sched_domains_numa_masks;
1287int __read_mostly		node_reclaim_distance = RECLAIM_DISTANCE;
1288#endif
1289
1290/*
1291 * SD_flags allowed in topology descriptions.
1292 *
1293 * These flags are purely descriptive of the topology and do not prescribe
1294 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1295 * function:
1296 *
1297 *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1298 *   SD_SHARE_PKG_RESOURCES - describes shared caches
1299 *   SD_NUMA                - describes NUMA topologies
1300 *   SD_SHARE_POWERDOMAIN   - describes shared power domain
1301 *
1302 * Odd one out, which beside describing the topology has a quirk also
1303 * prescribes the desired behaviour that goes along with it:
1304 *
1305 *   SD_ASYM_PACKING        - describes SMT quirks
1306 */
1307#define TOPOLOGY_SD_FLAGS		\
1308	(SD_SHARE_CPUCAPACITY	|	\
 
1309	 SD_SHARE_PKG_RESOURCES |	\
1310	 SD_NUMA		|	\
1311	 SD_ASYM_PACKING	|	\
1312	 SD_SHARE_POWERDOMAIN)
1313
1314static struct sched_domain *
1315sd_init(struct sched_domain_topology_level *tl,
1316	const struct cpumask *cpu_map,
1317	struct sched_domain *child, int dflags, int cpu)
1318{
1319	struct sd_data *sdd = &tl->data;
1320	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1321	int sd_id, sd_weight, sd_flags = 0;
 
1322
1323#ifdef CONFIG_NUMA
1324	/*
1325	 * Ugly hack to pass state to sd_numa_mask()...
1326	 */
1327	sched_domains_curr_level = tl->numa_level;
1328#endif
1329
1330	sd_weight = cpumask_weight(tl->mask(cpu));
1331
1332	if (tl->sd_flags)
1333		sd_flags = (*tl->sd_flags)();
1334	if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1335			"wrong sd_flags in topology description\n"))
1336		sd_flags &= ~TOPOLOGY_SD_FLAGS;
1337
1338	/* Apply detected topology flags */
1339	sd_flags |= dflags;
1340
1341	*sd = (struct sched_domain){
1342		.min_interval		= sd_weight,
1343		.max_interval		= 2*sd_weight,
1344		.busy_factor		= 32,
1345		.imbalance_pct		= 125,
1346
1347		.cache_nice_tries	= 0,
1348
1349		.flags			= 1*SD_LOAD_BALANCE
1350					| 1*SD_BALANCE_NEWIDLE
1351					| 1*SD_BALANCE_EXEC
1352					| 1*SD_BALANCE_FORK
1353					| 0*SD_BALANCE_WAKE
1354					| 1*SD_WAKE_AFFINE
1355					| 0*SD_SHARE_CPUCAPACITY
1356					| 0*SD_SHARE_PKG_RESOURCES
1357					| 0*SD_SERIALIZE
1358					| 1*SD_PREFER_SIBLING
1359					| 0*SD_NUMA
1360					| sd_flags
1361					,
1362
1363		.last_balance		= jiffies,
1364		.balance_interval	= sd_weight,
1365		.max_newidle_lb_cost	= 0,
1366		.next_decay_max_lb_cost	= jiffies,
1367		.child			= child,
1368#ifdef CONFIG_SCHED_DEBUG
1369		.name			= tl->name,
1370#endif
1371	};
1372
1373	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1374	sd_id = cpumask_first(sched_domain_span(sd));
 
 
 
 
 
 
 
1375
1376	/*
1377	 * Convert topological properties into behaviour.
1378	 */
1379
1380	if (sd->flags & SD_ASYM_CPUCAPACITY) {
1381		struct sched_domain *t = sd;
1382
1383		/*
1384		 * Don't attempt to spread across CPUs of different capacities.
1385		 */
1386		if (sd->child)
1387			sd->child->flags &= ~SD_PREFER_SIBLING;
1388
1389		for_each_lower_domain(t)
1390			t->flags |= SD_BALANCE_WAKE;
1391	}
1392
1393	if (sd->flags & SD_SHARE_CPUCAPACITY) {
1394		sd->imbalance_pct = 110;
1395
1396	} else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1397		sd->imbalance_pct = 117;
1398		sd->cache_nice_tries = 1;
1399
1400#ifdef CONFIG_NUMA
1401	} else if (sd->flags & SD_NUMA) {
1402		sd->cache_nice_tries = 2;
1403
1404		sd->flags &= ~SD_PREFER_SIBLING;
1405		sd->flags |= SD_SERIALIZE;
1406		if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1407			sd->flags &= ~(SD_BALANCE_EXEC |
1408				       SD_BALANCE_FORK |
1409				       SD_WAKE_AFFINE);
1410		}
1411
1412#endif
1413	} else {
1414		sd->cache_nice_tries = 1;
1415	}
1416
1417	/*
1418	 * For all levels sharing cache; connect a sched_domain_shared
1419	 * instance.
1420	 */
1421	if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1422		sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1423		atomic_inc(&sd->shared->ref);
1424		atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1425	}
1426
1427	sd->private = sdd;
1428
1429	return sd;
1430}
1431
1432/*
1433 * Topology list, bottom-up.
1434 */
1435static struct sched_domain_topology_level default_topology[] = {
1436#ifdef CONFIG_SCHED_SMT
1437	{ cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1438#endif
 
 
 
 
 
1439#ifdef CONFIG_SCHED_MC
1440	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1441#endif
1442	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1443	{ NULL, },
1444};
1445
1446static struct sched_domain_topology_level *sched_domain_topology =
1447	default_topology;
 
1448
1449#define for_each_sd_topology(tl)			\
1450	for (tl = sched_domain_topology; tl->mask; tl++)
1451
1452void set_sched_topology(struct sched_domain_topology_level *tl)
1453{
1454	if (WARN_ON_ONCE(sched_smp_initialized))
1455		return;
1456
1457	sched_domain_topology = tl;
 
1458}
1459
1460#ifdef CONFIG_NUMA
1461
1462static const struct cpumask *sd_numa_mask(int cpu)
1463{
1464	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1465}
1466
1467static void sched_numa_warn(const char *str)
1468{
1469	static int done = false;
1470	int i,j;
1471
1472	if (done)
1473		return;
1474
1475	done = true;
1476
1477	printk(KERN_WARNING "ERROR: %s\n\n", str);
1478
1479	for (i = 0; i < nr_node_ids; i++) {
1480		printk(KERN_WARNING "  ");
1481		for (j = 0; j < nr_node_ids; j++)
1482			printk(KERN_CONT "%02d ", node_distance(i,j));
 
 
 
 
1483		printk(KERN_CONT "\n");
1484	}
1485	printk(KERN_WARNING "\n");
1486}
1487
1488bool find_numa_distance(int distance)
1489{
1490	int i;
 
1491
1492	if (distance == node_distance(0, 0))
1493		return true;
1494
 
 
 
 
1495	for (i = 0; i < sched_domains_numa_levels; i++) {
1496		if (sched_domains_numa_distance[i] == distance)
1497			return true;
 
 
1498	}
 
 
1499
1500	return false;
1501}
1502
 
 
 
 
 
 
1503/*
1504 * A system can have three types of NUMA topology:
1505 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1506 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1507 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1508 *
1509 * The difference between a glueless mesh topology and a backplane
1510 * topology lies in whether communication between not directly
1511 * connected nodes goes through intermediary nodes (where programs
1512 * could run), or through backplane controllers. This affects
1513 * placement of programs.
1514 *
1515 * The type of topology can be discerned with the following tests:
1516 * - If the maximum distance between any nodes is 1 hop, the system
1517 *   is directly connected.
1518 * - If for two nodes A and B, located N > 1 hops away from each other,
1519 *   there is an intermediary node C, which is < N hops away from both
1520 *   nodes A and B, the system is a glueless mesh.
1521 */
1522static void init_numa_topology_type(void)
1523{
1524	int a, b, c, n;
1525
1526	n = sched_max_numa_distance;
1527
1528	if (sched_domains_numa_levels <= 2) {
1529		sched_numa_topology_type = NUMA_DIRECT;
1530		return;
1531	}
1532
1533	for_each_online_node(a) {
1534		for_each_online_node(b) {
1535			/* Find two nodes furthest removed from each other. */
1536			if (node_distance(a, b) < n)
1537				continue;
1538
1539			/* Is there an intermediary node between a and b? */
1540			for_each_online_node(c) {
1541				if (node_distance(a, c) < n &&
1542				    node_distance(b, c) < n) {
1543					sched_numa_topology_type =
1544							NUMA_GLUELESS_MESH;
1545					return;
1546				}
1547			}
1548
1549			sched_numa_topology_type = NUMA_BACKPLANE;
1550			return;
1551		}
1552	}
 
 
 
1553}
1554
1555void sched_init_numa(void)
1556{
1557	int next_distance, curr_distance = node_distance(0, 0);
1558	struct sched_domain_topology_level *tl;
1559	int level = 0;
1560	int i, j, k;
1561
1562	sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1563	if (!sched_domains_numa_distance)
1564		return;
1565
1566	/* Includes NUMA identity node at level 0. */
1567	sched_domains_numa_distance[level++] = curr_distance;
1568	sched_domains_numa_levels = level;
 
 
 
 
 
1569
1570	/*
1571	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1572	 * unique distances in the node_distance() table.
1573	 *
1574	 * Assumes node_distance(0,j) includes all distances in
1575	 * node_distance(i,j) in order to avoid cubic time.
1576	 */
1577	next_distance = curr_distance;
1578	for (i = 0; i < nr_node_ids; i++) {
1579		for (j = 0; j < nr_node_ids; j++) {
1580			for (k = 0; k < nr_node_ids; k++) {
1581				int distance = node_distance(i, k);
1582
1583				if (distance > curr_distance &&
1584				    (distance < next_distance ||
1585				     next_distance == curr_distance))
1586					next_distance = distance;
1587
1588				/*
1589				 * While not a strong assumption it would be nice to know
1590				 * about cases where if node A is connected to B, B is not
1591				 * equally connected to A.
1592				 */
1593				if (sched_debug() && node_distance(k, i) != distance)
1594					sched_numa_warn("Node-distance not symmetric");
1595
1596				if (sched_debug() && i && !find_numa_distance(distance))
1597					sched_numa_warn("Node-0 not representative");
 
 
 
 
 
 
 
1598			}
1599			if (next_distance != curr_distance) {
1600				sched_domains_numa_distance[level++] = next_distance;
1601				sched_domains_numa_levels = level;
1602				curr_distance = next_distance;
1603			} else break;
1604		}
 
 
 
 
 
 
1605
1606		/*
1607		 * In case of sched_debug() we verify the above assumption.
1608		 */
1609		if (!sched_debug())
1610			break;
 
 
 
 
1611	}
 
 
 
1612
1613	/*
1614	 * 'level' contains the number of unique distances
1615	 *
1616	 * The sched_domains_numa_distance[] array includes the actual distance
1617	 * numbers.
1618	 */
1619
1620	/*
1621	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1622	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1623	 * the array will contain less then 'level' members. This could be
1624	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1625	 * in other functions.
1626	 *
1627	 * We reset it to 'level' at the end of this function.
1628	 */
1629	sched_domains_numa_levels = 0;
1630
1631	sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1632	if (!sched_domains_numa_masks)
1633		return;
1634
1635	/*
1636	 * Now for each level, construct a mask per node which contains all
1637	 * CPUs of nodes that are that many hops away from us.
1638	 */
1639	for (i = 0; i < level; i++) {
1640		sched_domains_numa_masks[i] =
1641			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1642		if (!sched_domains_numa_masks[i])
1643			return;
1644
1645		for (j = 0; j < nr_node_ids; j++) {
1646			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
 
 
1647			if (!mask)
1648				return;
1649
1650			sched_domains_numa_masks[i][j] = mask;
 
 
 
 
1651
1652			for_each_node(k) {
1653				if (node_distance(j, k) > sched_domains_numa_distance[i])
1654					continue;
1655
1656				cpumask_or(mask, mask, cpumask_of_node(k));
1657			}
1658		}
1659	}
 
1660
1661	/* Compute default topology size */
1662	for (i = 0; sched_domain_topology[i].mask; i++);
1663
1664	tl = kzalloc((i + level + 1) *
1665			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1666	if (!tl)
1667		return;
1668
1669	/*
1670	 * Copy the default topology bits..
1671	 */
1672	for (i = 0; sched_domain_topology[i].mask; i++)
1673		tl[i] = sched_domain_topology[i];
1674
1675	/*
1676	 * Add the NUMA identity distance, aka single NODE.
1677	 */
1678	tl[i++] = (struct sched_domain_topology_level){
1679		.mask = sd_numa_mask,
1680		.numa_level = 0,
1681		SD_INIT_NAME(NODE)
1682	};
1683
1684	/*
1685	 * .. and append 'j' levels of NUMA goodness.
1686	 */
1687	for (j = 1; j < level; i++, j++) {
1688		tl[i] = (struct sched_domain_topology_level){
1689			.mask = sd_numa_mask,
1690			.sd_flags = cpu_numa_flags,
1691			.flags = SDTL_OVERLAP,
1692			.numa_level = j,
1693			SD_INIT_NAME(NUMA)
1694		};
1695	}
1696
 
1697	sched_domain_topology = tl;
1698
1699	sched_domains_numa_levels = level;
1700	sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1701
1702	init_numa_topology_type();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1703}
1704
1705void sched_domains_numa_masks_set(unsigned int cpu)
1706{
1707	int node = cpu_to_node(cpu);
1708	int i, j;
1709
1710	for (i = 0; i < sched_domains_numa_levels; i++) {
1711		for (j = 0; j < nr_node_ids; j++) {
 
 
 
 
1712			if (node_distance(j, node) <= sched_domains_numa_distance[i])
1713				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1714		}
1715	}
1716}
1717
1718void sched_domains_numa_masks_clear(unsigned int cpu)
1719{
1720	int i, j;
1721
1722	for (i = 0; i < sched_domains_numa_levels; i++) {
1723		for (j = 0; j < nr_node_ids; j++)
1724			cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
 
 
1725	}
1726}
1727
1728/*
1729 * sched_numa_find_closest() - given the NUMA topology, find the cpu
1730 *                             closest to @cpu from @cpumask.
1731 * cpumask: cpumask to find a cpu from
1732 * cpu: cpu to be close to
1733 *
1734 * returns: cpu, or nr_cpu_ids when nothing found.
1735 */
1736int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1737{
1738	int i, j = cpu_to_node(cpu);
 
1739
 
 
 
 
1740	for (i = 0; i < sched_domains_numa_levels; i++) {
1741		cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1742		if (cpu < nr_cpu_ids)
1743			return cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1744	}
1745	return nr_cpu_ids;
 
 
 
 
 
 
1746}
1747
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1748#endif /* CONFIG_NUMA */
1749
1750static int __sdt_alloc(const struct cpumask *cpu_map)
1751{
1752	struct sched_domain_topology_level *tl;
1753	int j;
1754
1755	for_each_sd_topology(tl) {
1756		struct sd_data *sdd = &tl->data;
1757
1758		sdd->sd = alloc_percpu(struct sched_domain *);
1759		if (!sdd->sd)
1760			return -ENOMEM;
1761
1762		sdd->sds = alloc_percpu(struct sched_domain_shared *);
1763		if (!sdd->sds)
1764			return -ENOMEM;
1765
1766		sdd->sg = alloc_percpu(struct sched_group *);
1767		if (!sdd->sg)
1768			return -ENOMEM;
1769
1770		sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1771		if (!sdd->sgc)
1772			return -ENOMEM;
1773
1774		for_each_cpu(j, cpu_map) {
1775			struct sched_domain *sd;
1776			struct sched_domain_shared *sds;
1777			struct sched_group *sg;
1778			struct sched_group_capacity *sgc;
1779
1780			sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1781					GFP_KERNEL, cpu_to_node(j));
1782			if (!sd)
1783				return -ENOMEM;
1784
1785			*per_cpu_ptr(sdd->sd, j) = sd;
1786
1787			sds = kzalloc_node(sizeof(struct sched_domain_shared),
1788					GFP_KERNEL, cpu_to_node(j));
1789			if (!sds)
1790				return -ENOMEM;
1791
1792			*per_cpu_ptr(sdd->sds, j) = sds;
1793
1794			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1795					GFP_KERNEL, cpu_to_node(j));
1796			if (!sg)
1797				return -ENOMEM;
1798
1799			sg->next = sg;
1800
1801			*per_cpu_ptr(sdd->sg, j) = sg;
1802
1803			sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1804					GFP_KERNEL, cpu_to_node(j));
1805			if (!sgc)
1806				return -ENOMEM;
1807
1808#ifdef CONFIG_SCHED_DEBUG
1809			sgc->id = j;
1810#endif
1811
1812			*per_cpu_ptr(sdd->sgc, j) = sgc;
1813		}
1814	}
1815
1816	return 0;
1817}
1818
1819static void __sdt_free(const struct cpumask *cpu_map)
1820{
1821	struct sched_domain_topology_level *tl;
1822	int j;
1823
1824	for_each_sd_topology(tl) {
1825		struct sd_data *sdd = &tl->data;
1826
1827		for_each_cpu(j, cpu_map) {
1828			struct sched_domain *sd;
1829
1830			if (sdd->sd) {
1831				sd = *per_cpu_ptr(sdd->sd, j);
1832				if (sd && (sd->flags & SD_OVERLAP))
1833					free_sched_groups(sd->groups, 0);
1834				kfree(*per_cpu_ptr(sdd->sd, j));
1835			}
1836
1837			if (sdd->sds)
1838				kfree(*per_cpu_ptr(sdd->sds, j));
1839			if (sdd->sg)
1840				kfree(*per_cpu_ptr(sdd->sg, j));
1841			if (sdd->sgc)
1842				kfree(*per_cpu_ptr(sdd->sgc, j));
1843		}
1844		free_percpu(sdd->sd);
1845		sdd->sd = NULL;
1846		free_percpu(sdd->sds);
1847		sdd->sds = NULL;
1848		free_percpu(sdd->sg);
1849		sdd->sg = NULL;
1850		free_percpu(sdd->sgc);
1851		sdd->sgc = NULL;
1852	}
1853}
1854
1855static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1856		const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1857		struct sched_domain *child, int dflags, int cpu)
1858{
1859	struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1860
1861	if (child) {
1862		sd->level = child->level + 1;
1863		sched_domain_level_max = max(sched_domain_level_max, sd->level);
1864		child->parent = sd;
1865
1866		if (!cpumask_subset(sched_domain_span(child),
1867				    sched_domain_span(sd))) {
1868			pr_err("BUG: arch topology borken\n");
1869#ifdef CONFIG_SCHED_DEBUG
1870			pr_err("     the %s domain not a subset of the %s domain\n",
1871					child->name, sd->name);
1872#endif
1873			/* Fixup, ensure @sd has at least @child CPUs. */
1874			cpumask_or(sched_domain_span(sd),
1875				   sched_domain_span(sd),
1876				   sched_domain_span(child));
1877		}
1878
1879	}
1880	set_domain_attribute(sd, attr);
1881
1882	return sd;
1883}
1884
1885/*
1886 * Find the sched_domain_topology_level where all CPU capacities are visible
1887 * for all CPUs.
1888 */
1889static struct sched_domain_topology_level
1890*asym_cpu_capacity_level(const struct cpumask *cpu_map)
1891{
1892	int i, j, asym_level = 0;
1893	bool asym = false;
1894	struct sched_domain_topology_level *tl, *asym_tl = NULL;
1895	unsigned long cap;
1896
1897	/* Is there any asymmetry? */
1898	cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1899
1900	for_each_cpu(i, cpu_map) {
1901		if (arch_scale_cpu_capacity(i) != cap) {
1902			asym = true;
1903			break;
1904		}
1905	}
1906
1907	if (!asym)
1908		return NULL;
 
1909
1910	/*
1911	 * Examine topology from all CPU's point of views to detect the lowest
1912	 * sched_domain_topology_level where a highest capacity CPU is visible
1913	 * to everyone.
 
1914	 */
1915	for_each_cpu(i, cpu_map) {
1916		unsigned long max_capacity = arch_scale_cpu_capacity(i);
1917		int tl_id = 0;
1918
1919		for_each_sd_topology(tl) {
1920			if (tl_id < asym_level)
1921				goto next_level;
1922
1923			for_each_cpu_and(j, tl->mask(i), cpu_map) {
1924				unsigned long capacity;
1925
1926				capacity = arch_scale_cpu_capacity(j);
1927
1928				if (capacity <= max_capacity)
1929					continue;
1930
1931				max_capacity = capacity;
1932				asym_level = tl_id;
1933				asym_tl = tl;
1934			}
1935next_level:
1936			tl_id++;
1937		}
1938	}
1939
1940	return asym_tl;
1941}
1942
1943
1944/*
1945 * Build sched domains for a given set of CPUs and attach the sched domains
1946 * to the individual CPUs
1947 */
1948static int
1949build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1950{
1951	enum s_alloc alloc_state = sa_none;
1952	struct sched_domain *sd;
1953	struct s_data d;
1954	struct rq *rq = NULL;
1955	int i, ret = -ENOMEM;
1956	struct sched_domain_topology_level *tl_asym;
1957	bool has_asym = false;
 
1958
1959	if (WARN_ON(cpumask_empty(cpu_map)))
1960		goto error;
1961
1962	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1963	if (alloc_state != sa_rootdomain)
1964		goto error;
1965
1966	tl_asym = asym_cpu_capacity_level(cpu_map);
1967
1968	/* Set up domains for CPUs specified by the cpu_map: */
1969	for_each_cpu(i, cpu_map) {
1970		struct sched_domain_topology_level *tl;
1971
1972		sd = NULL;
1973		for_each_sd_topology(tl) {
1974			int dflags = 0;
1975
1976			if (tl == tl_asym) {
1977				dflags |= SD_ASYM_CPUCAPACITY;
1978				has_asym = true;
1979			}
1980
1981			sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
1982
1983			if (tl == sched_domain_topology)
1984				*per_cpu_ptr(d.sd, i) = sd;
1985			if (tl->flags & SDTL_OVERLAP)
1986				sd->flags |= SD_OVERLAP;
1987			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1988				break;
1989		}
1990	}
1991
1992	/* Build the groups for the domains */
1993	for_each_cpu(i, cpu_map) {
1994		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1995			sd->span_weight = cpumask_weight(sched_domain_span(sd));
1996			if (sd->flags & SD_OVERLAP) {
1997				if (build_overlap_sched_groups(sd, i))
1998					goto error;
1999			} else {
2000				if (build_sched_groups(sd, i))
2001					goto error;
2002			}
2003		}
2004	}
2005
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2006	/* Calculate CPU capacity for physical packages and nodes */
2007	for (i = nr_cpumask_bits-1; i >= 0; i--) {
2008		if (!cpumask_test_cpu(i, cpu_map))
2009			continue;
2010
2011		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2012			claim_allocations(i, sd);
2013			init_sched_groups_capacity(i, sd);
2014		}
2015	}
2016
2017	/* Attach the domains */
2018	rcu_read_lock();
2019	for_each_cpu(i, cpu_map) {
 
 
2020		rq = cpu_rq(i);
2021		sd = *per_cpu_ptr(d.sd, i);
2022
 
2023		/* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2024		if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2025			WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2026
2027		cpu_attach_domain(sd, d.rd, i);
 
 
 
2028	}
2029	rcu_read_unlock();
2030
2031	if (has_asym)
2032		static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2033
2034	if (rq && sched_debug_enabled) {
 
 
 
2035		pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2036			cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2037	}
2038
2039	ret = 0;
2040error:
2041	__free_domain_allocs(&d, alloc_state, cpu_map);
2042
2043	return ret;
2044}
2045
2046/* Current sched domains: */
2047static cpumask_var_t			*doms_cur;
2048
2049/* Number of sched domains in 'doms_cur': */
2050static int				ndoms_cur;
2051
2052/* Attribues of custom domains in 'doms_cur' */
2053static struct sched_domain_attr		*dattr_cur;
2054
2055/*
2056 * Special case: If a kmalloc() of a doms_cur partition (array of
2057 * cpumask) fails, then fallback to a single sched domain,
2058 * as determined by the single cpumask fallback_doms.
2059 */
2060static cpumask_var_t			fallback_doms;
2061
2062/*
2063 * arch_update_cpu_topology lets virtualized architectures update the
2064 * CPU core maps. It is supposed to return 1 if the topology changed
2065 * or 0 if it stayed the same.
2066 */
2067int __weak arch_update_cpu_topology(void)
2068{
2069	return 0;
2070}
2071
2072cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2073{
2074	int i;
2075	cpumask_var_t *doms;
2076
2077	doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2078	if (!doms)
2079		return NULL;
2080	for (i = 0; i < ndoms; i++) {
2081		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2082			free_sched_domains(doms, i);
2083			return NULL;
2084		}
2085	}
2086	return doms;
2087}
2088
2089void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2090{
2091	unsigned int i;
2092	for (i = 0; i < ndoms; i++)
2093		free_cpumask_var(doms[i]);
2094	kfree(doms);
2095}
2096
2097/*
2098 * Set up scheduler domains and groups.  For now this just excludes isolated
2099 * CPUs, but could be used to exclude other special cases in the future.
2100 */
2101int sched_init_domains(const struct cpumask *cpu_map)
2102{
2103	int err;
2104
2105	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2106	zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2107	zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2108
2109	arch_update_cpu_topology();
 
2110	ndoms_cur = 1;
2111	doms_cur = alloc_sched_domains(ndoms_cur);
2112	if (!doms_cur)
2113		doms_cur = &fallback_doms;
2114	cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2115	err = build_sched_domains(doms_cur[0], NULL);
2116	register_sched_domain_sysctl();
2117
2118	return err;
2119}
2120
2121/*
2122 * Detach sched domains from a group of CPUs specified in cpu_map
2123 * These CPUs will now be attached to the NULL domain
2124 */
2125static void detach_destroy_domains(const struct cpumask *cpu_map)
2126{
2127	unsigned int cpu = cpumask_any(cpu_map);
2128	int i;
2129
2130	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2131		static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2132
 
 
 
2133	rcu_read_lock();
2134	for_each_cpu(i, cpu_map)
2135		cpu_attach_domain(NULL, &def_root_domain, i);
2136	rcu_read_unlock();
2137}
2138
2139/* handle null as "default" */
2140static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2141			struct sched_domain_attr *new, int idx_new)
2142{
2143	struct sched_domain_attr tmp;
2144
2145	/* Fast path: */
2146	if (!new && !cur)
2147		return 1;
2148
2149	tmp = SD_ATTR_INIT;
2150
2151	return !memcmp(cur ? (cur + idx_cur) : &tmp,
2152			new ? (new + idx_new) : &tmp,
2153			sizeof(struct sched_domain_attr));
2154}
2155
2156/*
2157 * Partition sched domains as specified by the 'ndoms_new'
2158 * cpumasks in the array doms_new[] of cpumasks. This compares
2159 * doms_new[] to the current sched domain partitioning, doms_cur[].
2160 * It destroys each deleted domain and builds each new domain.
2161 *
2162 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2163 * The masks don't intersect (don't overlap.) We should setup one
2164 * sched domain for each mask. CPUs not in any of the cpumasks will
2165 * not be load balanced. If the same cpumask appears both in the
2166 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2167 * it as it is.
2168 *
2169 * The passed in 'doms_new' should be allocated using
2170 * alloc_sched_domains.  This routine takes ownership of it and will
2171 * free_sched_domains it when done with it. If the caller failed the
2172 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2173 * and partition_sched_domains() will fallback to the single partition
2174 * 'fallback_doms', it also forces the domains to be rebuilt.
2175 *
2176 * If doms_new == NULL it will be replaced with cpu_online_mask.
2177 * ndoms_new == 0 is a special case for destroying existing domains,
2178 * and it will not create the default domain.
2179 *
2180 * Call with hotplug lock and sched_domains_mutex held
2181 */
2182void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2183				    struct sched_domain_attr *dattr_new)
2184{
2185	bool __maybe_unused has_eas = false;
2186	int i, j, n;
2187	int new_topology;
2188
2189	lockdep_assert_held(&sched_domains_mutex);
2190
2191	/* Always unregister in case we don't destroy any domains: */
2192	unregister_sched_domain_sysctl();
2193
2194	/* Let the architecture update CPU core mappings: */
2195	new_topology = arch_update_cpu_topology();
 
 
 
2196
2197	if (!doms_new) {
2198		WARN_ON_ONCE(dattr_new);
2199		n = 0;
2200		doms_new = alloc_sched_domains(1);
2201		if (doms_new) {
2202			n = 1;
2203			cpumask_and(doms_new[0], cpu_active_mask,
2204				    housekeeping_cpumask(HK_FLAG_DOMAIN));
2205		}
2206	} else {
2207		n = ndoms_new;
2208	}
2209
2210	/* Destroy deleted domains: */
2211	for (i = 0; i < ndoms_cur; i++) {
2212		for (j = 0; j < n && !new_topology; j++) {
2213			if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2214			    dattrs_equal(dattr_cur, i, dattr_new, j)) {
2215				struct root_domain *rd;
2216
2217				/*
2218				 * This domain won't be destroyed and as such
2219				 * its dl_bw->total_bw needs to be cleared.  It
2220				 * will be recomputed in function
2221				 * update_tasks_root_domain().
2222				 */
2223				rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2224				dl_clear_root_domain(rd);
2225				goto match1;
2226			}
2227		}
2228		/* No match - a current sched domain not in new doms_new[] */
2229		detach_destroy_domains(doms_cur[i]);
2230match1:
2231		;
2232	}
2233
2234	n = ndoms_cur;
2235	if (!doms_new) {
2236		n = 0;
2237		doms_new = &fallback_doms;
2238		cpumask_and(doms_new[0], cpu_active_mask,
2239			    housekeeping_cpumask(HK_FLAG_DOMAIN));
2240	}
2241
2242	/* Build new domains: */
2243	for (i = 0; i < ndoms_new; i++) {
2244		for (j = 0; j < n && !new_topology; j++) {
2245			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2246			    dattrs_equal(dattr_new, i, dattr_cur, j))
2247				goto match2;
2248		}
2249		/* No match - add a new doms_new */
2250		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2251match2:
2252		;
2253	}
2254
2255#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2256	/* Build perf. domains: */
2257	for (i = 0; i < ndoms_new; i++) {
2258		for (j = 0; j < n && !sched_energy_update; j++) {
2259			if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2260			    cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2261				has_eas = true;
2262				goto match3;
2263			}
2264		}
2265		/* No match - add perf. domains for a new rd */
2266		has_eas |= build_perf_domains(doms_new[i]);
2267match3:
2268		;
2269	}
2270	sched_energy_set(has_eas);
2271#endif
2272
2273	/* Remember the new sched domains: */
2274	if (doms_cur != &fallback_doms)
2275		free_sched_domains(doms_cur, ndoms_cur);
2276
2277	kfree(dattr_cur);
2278	doms_cur = doms_new;
2279	dattr_cur = dattr_new;
2280	ndoms_cur = ndoms_new;
2281
2282	register_sched_domain_sysctl();
2283}
2284
2285/*
2286 * Call with hotplug lock held
2287 */
2288void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2289			     struct sched_domain_attr *dattr_new)
2290{
2291	mutex_lock(&sched_domains_mutex);
2292	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2293	mutex_unlock(&sched_domains_mutex);
2294}