kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5 #include "sched.h"
   6
   7 DEFINE_MUTEX(sched_domains_mutex);
   8
   9 /* Protected by sched_domains_mutex: */
  10 static cpumask_var_t sched_domains_tmpmask;
  11 static cpumask_var_t sched_domains_tmpmask2;
  12
  13 #ifdef CONFIG_SCHED_DEBUG
  14
  15 static int __init sched_debug_setup(char *str)
  16 {
  17         sched_debug_enabled = true;
  18
  19         return 0;
  20 }
  21 early_param("sched_debug", sched_debug_setup);
  22
  23 static inline bool sched_debug(void)
  24 {
  25         return sched_debug_enabled;
  26 }
  27
  28 static 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
 115 static 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)
 142 static inline bool sched_debug(void)
 143 {
 144         return false;
 145 }
 146 #endif /* CONFIG_SCHED_DEBUG */
 147
 148 static 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
 173 static int
 174 sd_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)
 205 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
 206 unsigned int sysctl_sched_energy_aware = 1;
 207 DEFINE_MUTEX(sched_energy_mutex);
 208 bool sched_energy_update;
 209
 210 #ifdef CONFIG_PROC_SYSCTL
 211 int sched_energy_aware_handler(struct ctl_table *table, int write,
 212                 void *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
 235 static 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
 246 static 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
 257 static 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
 276 static 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
 295 static 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
 303 static 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. no SMT is detected.
 321  *    4. the EM complexity is low enough to keep scheduling overheads low;
 322  *    5. schedutil is driving the frequency of all CPUs of the rd;
 323  *
 324  * The complexity of the Energy Model is defined as:
 325  *
 326  *              C = nr_pd * (nr_cpus + nr_cs)
 327  *
 328  * with parameters defined as:
 329  *  - nr_pd:    the number of performance domains
 330  *  - nr_cpus:  the number of CPUs
 331  *  - nr_cs:    the sum of the number of capacity states of all performance
 332  *              domains (for example, on a system with 2 performance domains,
 333  *              with 10 capacity states each, nr_cs = 2 * 10 = 20).
 334  *
 335  * It is generally not a good idea to use such a model in the wake-up path on
 336  * very complex platforms because of the associated scheduling overheads. The
 337  * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
 338  * with per-CPU DVFS and less than 8 capacity states each, for example.
 339  */
 340 #define EM_MAX_COMPLEXITY 2048
 341
 342 extern struct cpufreq_governor schedutil_gov;
 343 static bool build_perf_domains(const struct cpumask *cpu_map)
 344 {
 345         int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
 346         struct perf_domain *pd = NULL, *tmp;
 347         int cpu = cpumask_first(cpu_map);
 348         struct root_domain *rd = cpu_rq(cpu)->rd;
 349         struct cpufreq_policy *policy;
 350         struct cpufreq_governor *gov;
 351
 352         if (!sysctl_sched_energy_aware)
 353                 goto free;
 354
 355         /* EAS is enabled for asymmetric CPU capacity topologies. */
 356         if (!per_cpu(sd_asym_cpucapacity, cpu)) {
 357                 if (sched_debug()) {
 358                         pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
 359                                         cpumask_pr_args(cpu_map));
 360                 }
 361                 goto free;
 362         }
 363
 364         /* EAS definitely does *not* handle SMT */
 365         if (sched_smt_active()) {
 366                 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
 367                         cpumask_pr_args(cpu_map));
 368                 goto free;
 369         }
 370
 371         for_each_cpu(i, cpu_map) {
 372                 /* Skip already covered CPUs. */
 373                 if (find_pd(pd, i))
 374                         continue;
 375
 376                 /* Do not attempt EAS if schedutil is not being used. */
 377                 policy = cpufreq_cpu_get(i);
 378                 if (!policy)
 379                         goto free;
 380                 gov = policy->governor;
 381                 cpufreq_cpu_put(policy);
 382                 if (gov != &schedutil_gov) {
 383                         if (rd->pd)
 384                                 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
 385                                                 cpumask_pr_args(cpu_map));
 386                         goto free;
 387                 }
 388
 389                 /* Create the new pd and add it to the local list. */
 390                 tmp = pd_init(i);
 391                 if (!tmp)
 392                         goto free;
 393                 tmp->next = pd;
 394                 pd = tmp;
 395
 396                 /*
 397                  * Count performance domains and capacity states for the
 398                  * complexity check.
 399                  */
 400                 nr_pd++;
 401                 nr_cs += em_pd_nr_cap_states(pd->em_pd);
 402         }
 403
 404         /* Bail out if the Energy Model complexity is too high. */
 405         if (nr_pd * (nr_cs + nr_cpus) > EM_MAX_COMPLEXITY) {
 406                 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
 407                                                 cpumask_pr_args(cpu_map));
 408                 goto free;
 409         }
 410
 411         perf_domain_debug(cpu_map, pd);
 412
 413         /* Attach the new list of performance domains to the root domain. */
 414         tmp = rd->pd;
 415         rcu_assign_pointer(rd->pd, pd);
 416         if (tmp)
 417                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 418
 419         return !!pd;
 420
 421 free:
 422         free_pd(pd);
 423         tmp = rd->pd;
 424         rcu_assign_pointer(rd->pd, NULL);
 425         if (tmp)
 426                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 427
 428         return false;
 429 }
 430 #else
 431 static void free_pd(struct perf_domain *pd) { }
 432 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
 433
 434 static void free_rootdomain(struct rcu_head *rcu)
 435 {
 436         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 437
 438         cpupri_cleanup(&rd->cpupri);
 439         cpudl_cleanup(&rd->cpudl);
 440         free_cpumask_var(rd->dlo_mask);
 441         free_cpumask_var(rd->rto_mask);
 442         free_cpumask_var(rd->online);
 443         free_cpumask_var(rd->span);
 444         free_pd(rd->pd);
 445         kfree(rd);
 446 }
 447
 448 void rq_attach_root(struct rq *rq, struct root_domain *rd)
 449 {
 450         struct root_domain *old_rd = NULL;
 451         unsigned long flags;
 452
 453         raw_spin_lock_irqsave(&rq->lock, flags);
 454
 455         if (rq->rd) {
 456                 old_rd = rq->rd;
 457
 458                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
 459                         set_rq_offline(rq);
 460
 461                 cpumask_clear_cpu(rq->cpu, old_rd->span);
 462
 463                 /*
 464                  * If we dont want to free the old_rd yet then
 465                  * set old_rd to NULL to skip the freeing later
 466                  * in this function:
 467                  */
 468                 if (!atomic_dec_and_test(&old_rd->refcount))
 469                         old_rd = NULL;
 470         }
 471
 472         atomic_inc(&rd->refcount);
 473         rq->rd = rd;
 474
 475         cpumask_set_cpu(rq->cpu, rd->span);
 476         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 477                 set_rq_online(rq);
 478
 479         raw_spin_unlock_irqrestore(&rq->lock, flags);
 480
 481         if (old_rd)
 482                 call_rcu(&old_rd->rcu, free_rootdomain);
 483 }
 484
 485 void sched_get_rd(struct root_domain *rd)
 486 {
 487         atomic_inc(&rd->refcount);
 488 }
 489
 490 void sched_put_rd(struct root_domain *rd)
 491 {
 492         if (!atomic_dec_and_test(&rd->refcount))
 493                 return;
 494
 495         call_rcu(&rd->rcu, free_rootdomain);
 496 }
 497
 498 static int init_rootdomain(struct root_domain *rd)
 499 {
 500         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 501                 goto out;
 502         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 503                 goto free_span;
 504         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 505                 goto free_online;
 506         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 507                 goto free_dlo_mask;
 508
 509 #ifdef HAVE_RT_PUSH_IPI
 510         rd->rto_cpu = -1;
 511         raw_spin_lock_init(&rd->rto_lock);
 512         init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
 513 #endif
 514
 515         init_dl_bw(&rd->dl_bw);
 516         if (cpudl_init(&rd->cpudl) != 0)
 517                 goto free_rto_mask;
 518
 519         if (cpupri_init(&rd->cpupri) != 0)
 520                 goto free_cpudl;
 521         return 0;
 522
 523 free_cpudl:
 524         cpudl_cleanup(&rd->cpudl);
 525 free_rto_mask:
 526         free_cpumask_var(rd->rto_mask);
 527 free_dlo_mask:
 528         free_cpumask_var(rd->dlo_mask);
 529 free_online:
 530         free_cpumask_var(rd->online);
 531 free_span:
 532         free_cpumask_var(rd->span);
 533 out:
 534         return -ENOMEM;
 535 }
 536
 537 /*
 538  * By default the system creates a single root-domain with all CPUs as
 539  * members (mimicking the global state we have today).
 540  */
 541 struct root_domain def_root_domain;
 542
 543 void init_defrootdomain(void)
 544 {
 545         init_rootdomain(&def_root_domain);
 546
 547         atomic_set(&def_root_domain.refcount, 1);
 548 }
 549
 550 static struct root_domain *alloc_rootdomain(void)
 551 {
 552         struct root_domain *rd;
 553
 554         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 555         if (!rd)
 556                 return NULL;
 557
 558         if (init_rootdomain(rd) != 0) {
 559                 kfree(rd);
 560                 return NULL;
 561         }
 562
 563         return rd;
 564 }
 565
 566 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 567 {
 568         struct sched_group *tmp, *first;
 569
 570         if (!sg)
 571                 return;
 572
 573         first = sg;
 574         do {
 575                 tmp = sg->next;
 576
 577                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 578                         kfree(sg->sgc);
 579
 580                 if (atomic_dec_and_test(&sg->ref))
 581                         kfree(sg);
 582                 sg = tmp;
 583         } while (sg != first);
 584 }
 585
 586 static void destroy_sched_domain(struct sched_domain *sd)
 587 {
 588         /*
 589          * A normal sched domain may have multiple group references, an
 590          * overlapping domain, having private groups, only one.  Iterate,
 591          * dropping group/capacity references, freeing where none remain.
 592          */
 593         free_sched_groups(sd->groups, 1);
 594
 595         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 596                 kfree(sd->shared);
 597         kfree(sd);
 598 }
 599
 600 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 601 {
 602         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 603
 604         while (sd) {
 605                 struct sched_domain *parent = sd->parent;
 606                 destroy_sched_domain(sd);
 607                 sd = parent;
 608         }
 609 }
 610
 611 static void destroy_sched_domains(struct sched_domain *sd)
 612 {
 613         if (sd)
 614                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 615 }
 616
 617 /*
 618  * Keep a special pointer to the highest sched_domain that has
 619  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
 620  * allows us to avoid some pointer chasing select_idle_sibling().
 621  *
 622  * Also keep a unique ID per domain (we use the first CPU number in
 623  * the cpumask of the domain), this allows us to quickly tell if
 624  * two CPUs are in the same cache domain, see cpus_share_cache().
 625  */
 626 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 627 DEFINE_PER_CPU(int, sd_llc_size);
 628 DEFINE_PER_CPU(int, sd_llc_id);
 629 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 630 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 631 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 632 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 633 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 634
 635 static void update_top_cache_domain(int cpu)
 636 {
 637         struct sched_domain_shared *sds = NULL;
 638         struct sched_domain *sd;
 639         int id = cpu;
 640         int size = 1;
 641
 642         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 643         if (sd) {
 644                 id = cpumask_first(sched_domain_span(sd));
 645                 size = cpumask_weight(sched_domain_span(sd));
 646                 sds = sd->shared;
 647         }
 648
 649         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 650         per_cpu(sd_llc_size, cpu) = size;
 651         per_cpu(sd_llc_id, cpu) = id;
 652         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 653
 654         sd = lowest_flag_domain(cpu, SD_NUMA);
 655         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 656
 657         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 658         rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 659
 660         sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
 661         rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 662 }
 663
 664 /*
 665  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 666  * hold the hotplug lock.
 667  */
 668 static void
 669 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 670 {
 671         struct rq *rq = cpu_rq(cpu);
 672         struct sched_domain *tmp;
 673
 674         /* Remove the sched domains which do not contribute to scheduling. */
 675         for (tmp = sd; tmp; ) {
 676                 struct sched_domain *parent = tmp->parent;
 677                 if (!parent)
 678                         break;
 679
 680                 if (sd_parent_degenerate(tmp, parent)) {
 681                         tmp->parent = parent->parent;
 682                         if (parent->parent)
 683                                 parent->parent->child = tmp;
 684                         /*
 685                          * Transfer SD_PREFER_SIBLING down in case of a
 686                          * degenerate parent; the spans match for this
 687                          * so the property transfers.
 688                          */
 689                         if (parent->flags & SD_PREFER_SIBLING)
 690                                 tmp->flags |= SD_PREFER_SIBLING;
 691                         destroy_sched_domain(parent);
 692                 } else
 693                         tmp = tmp->parent;
 694         }
 695
 696         if (sd && sd_degenerate(sd)) {
 697                 tmp = sd;
 698                 sd = sd->parent;
 699                 destroy_sched_domain(tmp);
 700                 if (sd)
 701                         sd->child = NULL;
 702         }
 703
 704         sched_domain_debug(sd, cpu);
 705
 706         rq_attach_root(rq, rd);
 707         tmp = rq->sd;
 708         rcu_assign_pointer(rq->sd, sd);
 709         dirty_sched_domain_sysctl(cpu);
 710         destroy_sched_domains(tmp);
 711
 712         update_top_cache_domain(cpu);
 713 }
 714
 715 struct s_data {
 716         struct sched_domain * __percpu *sd;
 717         struct root_domain      *rd;
 718 };
 719
 720 enum s_alloc {
 721         sa_rootdomain,
 722         sa_sd,
 723         sa_sd_storage,
 724         sa_none,
 725 };
 726
 727 /*
 728  * Return the canonical balance CPU for this group, this is the first CPU
 729  * of this group that's also in the balance mask.
 730  *
 731  * The balance mask are all those CPUs that could actually end up at this
 732  * group. See build_balance_mask().
 733  *
 734  * Also see should_we_balance().
 735  */
 736 int group_balance_cpu(struct sched_group *sg)
 737 {
 738         return cpumask_first(group_balance_mask(sg));
 739 }
 740
 741
 742 /*
 743  * NUMA topology (first read the regular topology blurb below)
 744  *
 745  * Given a node-distance table, for example:
 746  *
 747  *   node   0   1   2   3
 748  *     0:  10  20  30  20
 749  *     1:  20  10  20  30
 750  *     2:  30  20  10  20
 751  *     3:  20  30  20  10
 752  *
 753  * which represents a 4 node ring topology like:
 754  *
 755  *   0 ----- 1
 756  *   |       |
 757  *   |       |
 758  *   |       |
 759  *   3 ----- 2
 760  *
 761  * We want to construct domains and groups to represent this. The way we go
 762  * about doing this is to build the domains on 'hops'. For each NUMA level we
 763  * construct the mask of all nodes reachable in @level hops.
 764  *
 765  * For the above NUMA topology that gives 3 levels:
 766  *
 767  * NUMA-2       0-3             0-3             0-3             0-3
 768  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 769  *
 770  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 771  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 772  *
 773  * NUMA-0       0               1               2               3
 774  *
 775  *
 776  * As can be seen; things don't nicely line up as with the regular topology.
 777  * When we iterate a domain in child domain chunks some nodes can be
 778  * represented multiple times -- hence the "overlap" naming for this part of
 779  * the topology.
 780  *
 781  * In order to minimize this overlap, we only build enough groups to cover the
 782  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 783  *
 784  * Because:
 785  *
 786  *  - the first group of each domain is its child domain; this
 787  *    gets us the first 0-1,3
 788  *  - the only uncovered node is 2, who's child domain is 1-3.
 789  *
 790  * However, because of the overlap, computing a unique CPU for each group is
 791  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 792  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 793  * end up at those groups (they would end up in group: 0-1,3).
 794  *
 795  * To correct this we have to introduce the group balance mask. This mask
 796  * will contain those CPUs in the group that can reach this group given the
 797  * (child) domain tree.
 798  *
 799  * With this we can once again compute balance_cpu and sched_group_capacity
 800  * relations.
 801  *
 802  * XXX include words on how balance_cpu is unique and therefore can be
 803  * used for sched_group_capacity links.
 804  *
 805  *
 806  * Another 'interesting' topology is:
 807  *
 808  *   node   0   1   2   3
 809  *     0:  10  20  20  30
 810  *     1:  20  10  20  20
 811  *     2:  20  20  10  20
 812  *     3:  30  20  20  10
 813  *
 814  * Which looks a little like:
 815  *
 816  *   0 ----- 1
 817  *   |     / |
 818  *   |   /   |
 819  *   | /     |
 820  *   2 ----- 3
 821  *
 822  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 823  * are not.
 824  *
 825  * This leads to a few particularly weird cases where the sched_domain's are
 826  * not of the same number for each CPU. Consider:
 827  *
 828  * NUMA-2       0-3                                             0-3
 829  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 830  *
 831  * NUMA-1       0-2             0-3             0-3             1-3
 832  *
 833  * NUMA-0       0               1               2               3
 834  *
 835  */
 836
 837
 838 /*
 839  * Build the balance mask; it contains only those CPUs that can arrive at this
 840  * group and should be considered to continue balancing.
 841  *
 842  * We do this during the group creation pass, therefore the group information
 843  * isn't complete yet, however since each group represents a (child) domain we
 844  * can fully construct this using the sched_domain bits (which are already
 845  * complete).
 846  */
 847 static void
 848 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 849 {
 850         const struct cpumask *sg_span = sched_group_span(sg);
 851         struct sd_data *sdd = sd->private;
 852         struct sched_domain *sibling;
 853         int i;
 854
 855         cpumask_clear(mask);
 856
 857         for_each_cpu(i, sg_span) {
 858                 sibling = *per_cpu_ptr(sdd->sd, i);
 859
 860                 /*
 861                  * Can happen in the asymmetric case, where these siblings are
 862                  * unused. The mask will not be empty because those CPUs that
 863                  * do have the top domain _should_ span the domain.
 864                  */
 865                 if (!sibling->child)
 866                         continue;
 867
 868                 /* If we would not end up here, we can't continue from here */
 869                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 870                         continue;
 871
 872                 cpumask_set_cpu(i, mask);
 873         }
 874
 875         /* We must not have empty masks here */
 876         WARN_ON_ONCE(cpumask_empty(mask));
 877 }
 878
 879 /*
 880  * XXX: This creates per-node group entries; since the load-balancer will
 881  * immediately access remote memory to construct this group's load-balance
 882  * statistics having the groups node local is of dubious benefit.
 883  */
 884 static struct sched_group *
 885 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 886 {
 887         struct sched_group *sg;
 888         struct cpumask *sg_span;
 889
 890         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 891                         GFP_KERNEL, cpu_to_node(cpu));
 892
 893         if (!sg)
 894                 return NULL;
 895
 896         sg_span = sched_group_span(sg);
 897         if (sd->child)
 898                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 899         else
 900                 cpumask_copy(sg_span, sched_domain_span(sd));
 901
 902         atomic_inc(&sg->ref);
 903         return sg;
 904 }
 905
 906 static void init_overlap_sched_group(struct sched_domain *sd,
 907                                      struct sched_group *sg)
 908 {
 909         struct cpumask *mask = sched_domains_tmpmask2;
 910         struct sd_data *sdd = sd->private;
 911         struct cpumask *sg_span;
 912         int cpu;
 913
 914         build_balance_mask(sd, sg, mask);
 915         cpu = cpumask_first_and(sched_group_span(sg), mask);
 916
 917         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 918         if (atomic_inc_return(&sg->sgc->ref) == 1)
 919                 cpumask_copy(group_balance_mask(sg), mask);
 920         else
 921                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 922
 923         /*
 924          * Initialize sgc->capacity such that even if we mess up the
 925          * domains and no possible iteration will get us here, we won't
 926          * die on a /0 trap.
 927          */
 928         sg_span = sched_group_span(sg);
 929         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 930         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 931         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 932 }
 933
 934 static int
 935 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
 936 {
 937         struct sched_group *first = NULL, *last = NULL, *sg;
 938         const struct cpumask *span = sched_domain_span(sd);
 939         struct cpumask *covered = sched_domains_tmpmask;
 940         struct sd_data *sdd = sd->private;
 941         struct sched_domain *sibling;
 942         int i;
 943
 944         cpumask_clear(covered);
 945
 946         for_each_cpu_wrap(i, span, cpu) {
 947                 struct cpumask *sg_span;
 948
 949                 if (cpumask_test_cpu(i, covered))
 950                         continue;
 951
 952                 sibling = *per_cpu_ptr(sdd->sd, i);
 953
 954                 /*
 955                  * Asymmetric node setups can result in situations where the
 956                  * domain tree is of unequal depth, make sure to skip domains
 957                  * that already cover the entire range.
 958                  *
 959                  * In that case build_sched_domains() will have terminated the
 960                  * iteration early and our sibling sd spans will be empty.
 961                  * Domains should always include the CPU they're built on, so
 962                  * check that.
 963                  */
 964                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
 965                         continue;
 966
 967                 sg = build_group_from_child_sched_domain(sibling, cpu);
 968                 if (!sg)
 969                         goto fail;
 970
 971                 sg_span = sched_group_span(sg);
 972                 cpumask_or(covered, covered, sg_span);
 973
 974                 init_overlap_sched_group(sd, sg);
 975
 976                 if (!first)
 977                         first = sg;
 978                 if (last)
 979                         last->next = sg;
 980                 last = sg;
 981                 last->next = first;
 982         }
 983         sd->groups = first;
 984
 985         return 0;
 986
 987 fail:
 988         free_sched_groups(first, 0);
 989
 990         return -ENOMEM;
 991 }
 992
 993
 994 /*
 995  * Package topology (also see the load-balance blurb in fair.c)
 996  *
 997  * The scheduler builds a tree structure to represent a number of important
 998  * topology features. By default (default_topology[]) these include:
 999  *
1000  *  - Simultaneous multithreading (SMT)
1001  *  - Multi-Core Cache (MC)
1002  *  - Package (DIE)
1003  *
1004  * Where the last one more or less denotes everything up to a NUMA node.
1005  *
1006  * The tree consists of 3 primary data structures:
1007  *
1008  *      sched_domain -> sched_group -> sched_group_capacity
1009  *          ^ ^             ^ ^
1010  *          `-'             `-'
1011  *
1012  * The sched_domains are per-CPU and have a two way link (parent & child) and
1013  * denote the ever growing mask of CPUs belonging to that level of topology.
1014  *
1015  * Each sched_domain has a circular (double) linked list of sched_group's, each
1016  * denoting the domains of the level below (or individual CPUs in case of the
1017  * first domain level). The sched_group linked by a sched_domain includes the
1018  * CPU of that sched_domain [*].
1019  *
1020  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1021  *
1022  * CPU   0   1   2   3   4   5   6   7
1023  *
1024  * DIE  [                             ]
1025  * MC   [             ] [             ]
1026  * SMT  [     ] [     ] [     ] [     ]
1027  *
1028  *  - or -
1029  *
1030  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1031  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1032  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1033  *
1034  * CPU   0   1   2   3   4   5   6   7
1035  *
1036  * One way to think about it is: sched_domain moves you up and down among these
1037  * topology levels, while sched_group moves you sideways through it, at child
1038  * domain granularity.
1039  *
1040  * sched_group_capacity ensures each unique sched_group has shared storage.
1041  *
1042  * There are two related construction problems, both require a CPU that
1043  * uniquely identify each group (for a given domain):
1044  *
1045  *  - The first is the balance_cpu (see should_we_balance() and the
1046  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1047  *    continue balancing at a higher domain.
1048  *
1049  *  - The second is the sched_group_capacity; we want all identical groups
1050  *    to share a single sched_group_capacity.
1051  *
1052  * Since these topologies are exclusive by construction. That is, its
1053  * impossible for an SMT thread to belong to multiple cores, and cores to
1054  * be part of multiple caches. There is a very clear and unique location
1055  * for each CPU in the hierarchy.
1056  *
1057  * Therefore computing a unique CPU for each group is trivial (the iteration
1058  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1059  * group), we can simply pick the first CPU in each group.
1060  *
1061  *
1062  * [*] in other words, the first group of each domain is its child domain.
1063  */
1064
1065 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1066 {
1067         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1068         struct sched_domain *child = sd->child;
1069         struct sched_group *sg;
1070         bool already_visited;
1071
1072         if (child)
1073                 cpu = cpumask_first(sched_domain_span(child));
1074
1075         sg = *per_cpu_ptr(sdd->sg, cpu);
1076         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1077
1078         /* Increase refcounts for claim_allocations: */
1079         already_visited = atomic_inc_return(&sg->ref) > 1;
1080         /* sgc visits should follow a similar trend as sg */
1081         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1082
1083         /* If we have already visited that group, it's already initialized. */
1084         if (already_visited)
1085                 return sg;
1086
1087         if (child) {
1088                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1089                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1090         } else {
1091                 cpumask_set_cpu(cpu, sched_group_span(sg));
1092                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1093         }
1094
1095         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1096         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1097         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1098
1099         return sg;
1100 }
1101
1102 /*
1103  * build_sched_groups will build a circular linked list of the groups
1104  * covered by the given span, will set each group's ->cpumask correctly,
1105  * and will initialize their ->sgc.
1106  *
1107  * Assumes the sched_domain tree is fully constructed
1108  */
1109 static int
1110 build_sched_groups(struct sched_domain *sd, int cpu)
1111 {
1112         struct sched_group *first = NULL, *last = NULL;
1113         struct sd_data *sdd = sd->private;
1114         const struct cpumask *span = sched_domain_span(sd);
1115         struct cpumask *covered;
1116         int i;
1117
1118         lockdep_assert_held(&sched_domains_mutex);
1119         covered = sched_domains_tmpmask;
1120
1121         cpumask_clear(covered);
1122
1123         for_each_cpu_wrap(i, span, cpu) {
1124                 struct sched_group *sg;
1125
1126                 if (cpumask_test_cpu(i, covered))
1127                         continue;
1128
1129                 sg = get_group(i, sdd);
1130
1131                 cpumask_or(covered, covered, sched_group_span(sg));
1132
1133                 if (!first)
1134                         first = sg;
1135                 if (last)
1136                         last->next = sg;
1137                 last = sg;
1138         }
1139         last->next = first;
1140         sd->groups = first;
1141
1142         return 0;
1143 }
1144
1145 /*
1146  * Initialize sched groups cpu_capacity.
1147  *
1148  * cpu_capacity indicates the capacity of sched group, which is used while
1149  * distributing the load between different sched groups in a sched domain.
1150  * Typically cpu_capacity for all the groups in a sched domain will be same
1151  * unless there are asymmetries in the topology. If there are asymmetries,
1152  * group having more cpu_capacity will pickup more load compared to the
1153  * group having less cpu_capacity.
1154  */
1155 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1156 {
1157         struct sched_group *sg = sd->groups;
1158
1159         WARN_ON(!sg);
1160
1161         do {
1162                 int cpu, max_cpu = -1;
1163
1164                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1165
1166                 if (!(sd->flags & SD_ASYM_PACKING))
1167                         goto next;
1168
1169                 for_each_cpu(cpu, sched_group_span(sg)) {
1170                         if (max_cpu < 0)
1171                                 max_cpu = cpu;
1172                         else if (sched_asym_prefer(cpu, max_cpu))
1173                                 max_cpu = cpu;
1174                 }
1175                 sg->asym_prefer_cpu = max_cpu;
1176
1177 next:
1178                 sg = sg->next;
1179         } while (sg != sd->groups);
1180
1181         if (cpu != group_balance_cpu(sg))
1182                 return;
1183
1184         update_group_capacity(sd, cpu);
1185 }
1186
1187 /*
1188  * Initializers for schedule domains
1189  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1190  */
1191
1192 static int default_relax_domain_level = -1;
1193 int sched_domain_level_max;
1194
1195 static int __init setup_relax_domain_level(char *str)
1196 {
1197         if (kstrtoint(str, 0, &default_relax_domain_level))
1198                 pr_warn("Unable to set relax_domain_level\n");
1199
1200         return 1;
1201 }
1202 __setup("relax_domain_level=", setup_relax_domain_level);
1203
1204 static void set_domain_attribute(struct sched_domain *sd,
1205                                  struct sched_domain_attr *attr)
1206 {
1207         int request;
1208
1209         if (!attr || attr->relax_domain_level < 0) {
1210                 if (default_relax_domain_level < 0)
1211                         return;
1212                 request = default_relax_domain_level;
1213         } else
1214                 request = attr->relax_domain_level;
1215
1216         if (sd->level > request) {
1217                 /* Turn off idle balance on this domain: */
1218                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1219         }
1220 }
1221
1222 static void __sdt_free(const struct cpumask *cpu_map);
1223 static int __sdt_alloc(const struct cpumask *cpu_map);
1224
1225 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1226                                  const struct cpumask *cpu_map)
1227 {
1228         switch (what) {
1229         case sa_rootdomain:
1230                 if (!atomic_read(&d->rd->refcount))
1231                         free_rootdomain(&d->rd->rcu);
1232                 /* Fall through */
1233         case sa_sd:
1234                 free_percpu(d->sd);
1235                 /* Fall through */
1236         case sa_sd_storage:
1237                 __sdt_free(cpu_map);
1238                 /* Fall through */
1239         case sa_none:
1240                 break;
1241         }
1242 }
1243
1244 static enum s_alloc
1245 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1246 {
1247         memset(d, 0, sizeof(*d));
1248
1249         if (__sdt_alloc(cpu_map))
1250                 return sa_sd_storage;
1251         d->sd = alloc_percpu(struct sched_domain *);
1252         if (!d->sd)
1253                 return sa_sd_storage;
1254         d->rd = alloc_rootdomain();
1255         if (!d->rd)
1256                 return sa_sd;
1257
1258         return sa_rootdomain;
1259 }
1260
1261 /*
1262  * NULL the sd_data elements we've used to build the sched_domain and
1263  * sched_group structure so that the subsequent __free_domain_allocs()
1264  * will not free the data we're using.
1265  */
1266 static void claim_allocations(int cpu, struct sched_domain *sd)
1267 {
1268         struct sd_data *sdd = sd->private;
1269
1270         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1271         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1272
1273         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1274                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1275
1276         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1277                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1278
1279         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1280                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1281 }
1282
1283 #ifdef CONFIG_NUMA
1284 enum numa_topology_type sched_numa_topology_type;
1285
1286 static int                      sched_domains_numa_levels;
1287 static int                      sched_domains_curr_level;
1288
1289 int                             sched_max_numa_distance;
1290 static int                      *sched_domains_numa_distance;
1291 static struct cpumask           ***sched_domains_numa_masks;
1292 int __read_mostly               node_reclaim_distance = RECLAIM_DISTANCE;
1293 #endif
1294
1295 /*
1296  * SD_flags allowed in topology descriptions.
1297  *
1298  * These flags are purely descriptive of the topology and do not prescribe
1299  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1300  * function:
1301  *
1302  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1303  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1304  *   SD_NUMA                - describes NUMA topologies
1305  *   SD_SHARE_POWERDOMAIN   - describes shared power domain
1306  *
1307  * Odd one out, which beside describing the topology has a quirk also
1308  * prescribes the desired behaviour that goes along with it:
1309  *
1310  *   SD_ASYM_PACKING        - describes SMT quirks
1311  */
1312 #define TOPOLOGY_SD_FLAGS               \
1313         (SD_SHARE_CPUCAPACITY   |       \
1314          SD_SHARE_PKG_RESOURCES |       \
1315          SD_NUMA                |       \
1316          SD_ASYM_PACKING        |       \
1317          SD_SHARE_POWERDOMAIN)
1318
1319 static struct sched_domain *
1320 sd_init(struct sched_domain_topology_level *tl,
1321         const struct cpumask *cpu_map,
1322         struct sched_domain *child, int dflags, int cpu)
1323 {
1324         struct sd_data *sdd = &tl->data;
1325         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1326         int sd_id, sd_weight, sd_flags = 0;
1327
1328 #ifdef CONFIG_NUMA
1329         /*
1330          * Ugly hack to pass state to sd_numa_mask()...
1331          */
1332         sched_domains_curr_level = tl->numa_level;
1333 #endif
1334
1335         sd_weight = cpumask_weight(tl->mask(cpu));
1336
1337         if (tl->sd_flags)
1338                 sd_flags = (*tl->sd_flags)();
1339         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1340                         "wrong sd_flags in topology description\n"))
1341                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1342
1343         /* Apply detected topology flags */
1344         sd_flags |= dflags;
1345
1346         *sd = (struct sched_domain){
1347                 .min_interval           = sd_weight,
1348                 .max_interval           = 2*sd_weight,
1349                 .busy_factor            = 32,
1350                 .imbalance_pct          = 125,
1351
1352                 .cache_nice_tries       = 0,
1353
1354                 .flags                  = 1*SD_LOAD_BALANCE
1355                                         | 1*SD_BALANCE_NEWIDLE
1356                                         | 1*SD_BALANCE_EXEC
1357                                         | 1*SD_BALANCE_FORK
1358                                         | 0*SD_BALANCE_WAKE
1359                                         | 1*SD_WAKE_AFFINE
1360                                         | 0*SD_SHARE_CPUCAPACITY
1361                                         | 0*SD_SHARE_PKG_RESOURCES
1362                                         | 0*SD_SERIALIZE
1363                                         | 1*SD_PREFER_SIBLING
1364                                         | 0*SD_NUMA
1365                                         | sd_flags
1366                                         ,
1367
1368                 .last_balance           = jiffies,
1369                 .balance_interval       = sd_weight,
1370                 .max_newidle_lb_cost    = 0,
1371                 .next_decay_max_lb_cost = jiffies,
1372                 .child                  = child,
1373 #ifdef CONFIG_SCHED_DEBUG
1374                 .name                   = tl->name,
1375 #endif
1376         };
1377
1378         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1379         sd_id = cpumask_first(sched_domain_span(sd));
1380
1381         /*
1382          * Convert topological properties into behaviour.
1383          */
1384
1385         /* Don't attempt to spread across CPUs of different capacities. */
1386         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1387                 sd->child->flags &= ~SD_PREFER_SIBLING;
1388
1389         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1390                 sd->imbalance_pct = 110;
1391
1392         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1393                 sd->imbalance_pct = 117;
1394                 sd->cache_nice_tries = 1;
1395
1396 #ifdef CONFIG_NUMA
1397         } else if (sd->flags & SD_NUMA) {
1398                 sd->cache_nice_tries = 2;
1399
1400                 sd->flags &= ~SD_PREFER_SIBLING;
1401                 sd->flags |= SD_SERIALIZE;
1402                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1403                         sd->flags &= ~(SD_BALANCE_EXEC |
1404                                        SD_BALANCE_FORK |
1405                                        SD_WAKE_AFFINE);
1406                 }
1407
1408 #endif
1409         } else {
1410                 sd->cache_nice_tries = 1;
1411         }
1412
1413         /*
1414          * For all levels sharing cache; connect a sched_domain_shared
1415          * instance.
1416          */
1417         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1418                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1419                 atomic_inc(&sd->shared->ref);
1420                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1421         }
1422
1423         sd->private = sdd;
1424
1425         return sd;
1426 }
1427
1428 /*
1429  * Topology list, bottom-up.
1430  */
1431 static struct sched_domain_topology_level default_topology[] = {
1432 #ifdef CONFIG_SCHED_SMT
1433         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1434 #endif
1435 #ifdef CONFIG_SCHED_MC
1436         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1437 #endif
1438         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1439         { NULL, },
1440 };
1441
1442 static struct sched_domain_topology_level *sched_domain_topology =
1443         default_topology;
1444
1445 #define for_each_sd_topology(tl)                        \
1446         for (tl = sched_domain_topology; tl->mask; tl++)
1447
1448 void set_sched_topology(struct sched_domain_topology_level *tl)
1449 {
1450         if (WARN_ON_ONCE(sched_smp_initialized))
1451                 return;
1452
1453         sched_domain_topology = tl;
1454 }
1455
1456 #ifdef CONFIG_NUMA
1457
1458 static const struct cpumask *sd_numa_mask(int cpu)
1459 {
1460         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1461 }
1462
1463 static void sched_numa_warn(const char *str)
1464 {
1465         static int done = false;
1466         int i,j;
1467
1468         if (done)
1469                 return;
1470
1471         done = true;
1472
1473         printk(KERN_WARNING "ERROR: %s\n\n", str);
1474
1475         for (i = 0; i < nr_node_ids; i++) {
1476                 printk(KERN_WARNING "  ");
1477                 for (j = 0; j < nr_node_ids; j++)
1478                         printk(KERN_CONT "%02d ", node_distance(i,j));
1479                 printk(KERN_CONT "\n");
1480         }
1481         printk(KERN_WARNING "\n");
1482 }
1483
1484 bool find_numa_distance(int distance)
1485 {
1486         int i;
1487
1488         if (distance == node_distance(0, 0))
1489                 return true;
1490
1491         for (i = 0; i < sched_domains_numa_levels; i++) {
1492                 if (sched_domains_numa_distance[i] == distance)
1493                         return true;
1494         }
1495
1496         return false;
1497 }
1498
1499 /*
1500  * A system can have three types of NUMA topology:
1501  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1502  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1503  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1504  *
1505  * The difference between a glueless mesh topology and a backplane
1506  * topology lies in whether communication between not directly
1507  * connected nodes goes through intermediary nodes (where programs
1508  * could run), or through backplane controllers. This affects
1509  * placement of programs.
1510  *
1511  * The type of topology can be discerned with the following tests:
1512  * - If the maximum distance between any nodes is 1 hop, the system
1513  *   is directly connected.
1514  * - If for two nodes A and B, located N > 1 hops away from each other,
1515  *   there is an intermediary node C, which is < N hops away from both
1516  *   nodes A and B, the system is a glueless mesh.
1517  */
1518 static void init_numa_topology_type(void)
1519 {
1520         int a, b, c, n;
1521
1522         n = sched_max_numa_distance;
1523
1524         if (sched_domains_numa_levels <= 2) {
1525                 sched_numa_topology_type = NUMA_DIRECT;
1526                 return;
1527         }
1528
1529         for_each_online_node(a) {
1530                 for_each_online_node(b) {
1531                         /* Find two nodes furthest removed from each other. */
1532                         if (node_distance(a, b) < n)
1533                                 continue;
1534
1535                         /* Is there an intermediary node between a and b? */
1536                         for_each_online_node(c) {
1537                                 if (node_distance(a, c) < n &&
1538                                     node_distance(b, c) < n) {
1539                                         sched_numa_topology_type =
1540                                                         NUMA_GLUELESS_MESH;
1541                                         return;
1542                                 }
1543                         }
1544
1545                         sched_numa_topology_type = NUMA_BACKPLANE;
1546                         return;
1547                 }
1548         }
1549 }
1550
1551 void sched_init_numa(void)
1552 {
1553         int next_distance, curr_distance = node_distance(0, 0);
1554         struct sched_domain_topology_level *tl;
1555         int level = 0;
1556         int i, j, k;
1557
1558         sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1559         if (!sched_domains_numa_distance)
1560                 return;
1561
1562         /* Includes NUMA identity node at level 0. */
1563         sched_domains_numa_distance[level++] = curr_distance;
1564         sched_domains_numa_levels = level;
1565
1566         /*
1567          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1568          * unique distances in the node_distance() table.
1569          *
1570          * Assumes node_distance(0,j) includes all distances in
1571          * node_distance(i,j) in order to avoid cubic time.
1572          */
1573         next_distance = curr_distance;
1574         for (i = 0; i < nr_node_ids; i++) {
1575                 for (j = 0; j < nr_node_ids; j++) {
1576                         for (k = 0; k < nr_node_ids; k++) {
1577                                 int distance = node_distance(i, k);
1578
1579                                 if (distance > curr_distance &&
1580                                     (distance < next_distance ||
1581                                      next_distance == curr_distance))
1582                                         next_distance = distance;
1583
1584                                 /*
1585                                  * While not a strong assumption it would be nice to know
1586                                  * about cases where if node A is connected to B, B is not
1587                                  * equally connected to A.
1588                                  */
1589                                 if (sched_debug() && node_distance(k, i) != distance)
1590                                         sched_numa_warn("Node-distance not symmetric");
1591
1592                                 if (sched_debug() && i && !find_numa_distance(distance))
1593                                         sched_numa_warn("Node-0 not representative");
1594                         }
1595                         if (next_distance != curr_distance) {
1596                                 sched_domains_numa_distance[level++] = next_distance;
1597                                 sched_domains_numa_levels = level;
1598                                 curr_distance = next_distance;
1599                         } else break;
1600                 }
1601
1602                 /*
1603                  * In case of sched_debug() we verify the above assumption.
1604                  */
1605                 if (!sched_debug())
1606                         break;
1607         }
1608
1609         /*
1610          * 'level' contains the number of unique distances
1611          *
1612          * The sched_domains_numa_distance[] array includes the actual distance
1613          * numbers.
1614          */
1615
1616         /*
1617          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1618          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1619          * the array will contain less then 'level' members. This could be
1620          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1621          * in other functions.
1622          *
1623          * We reset it to 'level' at the end of this function.
1624          */
1625         sched_domains_numa_levels = 0;
1626
1627         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1628         if (!sched_domains_numa_masks)
1629                 return;
1630
1631         /*
1632          * Now for each level, construct a mask per node which contains all
1633          * CPUs of nodes that are that many hops away from us.
1634          */
1635         for (i = 0; i < level; i++) {
1636                 sched_domains_numa_masks[i] =
1637                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1638                 if (!sched_domains_numa_masks[i])
1639                         return;
1640
1641                 for (j = 0; j < nr_node_ids; j++) {
1642                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1643                         if (!mask)
1644                                 return;
1645
1646                         sched_domains_numa_masks[i][j] = mask;
1647
1648                         for_each_node(k) {
1649                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1650                                         continue;
1651
1652                                 cpumask_or(mask, mask, cpumask_of_node(k));
1653                         }
1654                 }
1655         }
1656
1657         /* Compute default topology size */
1658         for (i = 0; sched_domain_topology[i].mask; i++);
1659
1660         tl = kzalloc((i + level + 1) *
1661                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1662         if (!tl)
1663                 return;
1664
1665         /*
1666          * Copy the default topology bits..
1667          */
1668         for (i = 0; sched_domain_topology[i].mask; i++)
1669                 tl[i] = sched_domain_topology[i];
1670
1671         /*
1672          * Add the NUMA identity distance, aka single NODE.
1673          */
1674         tl[i++] = (struct sched_domain_topology_level){
1675                 .mask = sd_numa_mask,
1676                 .numa_level = 0,
1677                 SD_INIT_NAME(NODE)
1678         };
1679
1680         /*
1681          * .. and append 'j' levels of NUMA goodness.
1682          */
1683         for (j = 1; j < level; i++, j++) {
1684                 tl[i] = (struct sched_domain_topology_level){
1685                         .mask = sd_numa_mask,
1686                         .sd_flags = cpu_numa_flags,
1687                         .flags = SDTL_OVERLAP,
1688                         .numa_level = j,
1689                         SD_INIT_NAME(NUMA)
1690                 };
1691         }
1692
1693         sched_domain_topology = tl;
1694
1695         sched_domains_numa_levels = level;
1696         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1697
1698         init_numa_topology_type();
1699 }
1700
1701 void sched_domains_numa_masks_set(unsigned int cpu)
1702 {
1703         int node = cpu_to_node(cpu);
1704         int i, j;
1705
1706         for (i = 0; i < sched_domains_numa_levels; i++) {
1707                 for (j = 0; j < nr_node_ids; j++) {
1708                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
1709                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1710                 }
1711         }
1712 }
1713
1714 void sched_domains_numa_masks_clear(unsigned int cpu)
1715 {
1716         int i, j;
1717
1718         for (i = 0; i < sched_domains_numa_levels; i++) {
1719                 for (j = 0; j < nr_node_ids; j++)
1720                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1721         }
1722 }
1723
1724 /*
1725  * sched_numa_find_closest() - given the NUMA topology, find the cpu
1726  *                             closest to @cpu from @cpumask.
1727  * cpumask: cpumask to find a cpu from
1728  * cpu: cpu to be close to
1729  *
1730  * returns: cpu, or nr_cpu_ids when nothing found.
1731  */
1732 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1733 {
1734         int i, j = cpu_to_node(cpu);
1735
1736         for (i = 0; i < sched_domains_numa_levels; i++) {
1737                 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1738                 if (cpu < nr_cpu_ids)
1739                         return cpu;
1740         }
1741         return nr_cpu_ids;
1742 }
1743
1744 #endif /* CONFIG_NUMA */
1745
1746 static int __sdt_alloc(const struct cpumask *cpu_map)
1747 {
1748         struct sched_domain_topology_level *tl;
1749         int j;
1750
1751         for_each_sd_topology(tl) {
1752                 struct sd_data *sdd = &tl->data;
1753
1754                 sdd->sd = alloc_percpu(struct sched_domain *);
1755                 if (!sdd->sd)
1756                         return -ENOMEM;
1757
1758                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1759                 if (!sdd->sds)
1760                         return -ENOMEM;
1761
1762                 sdd->sg = alloc_percpu(struct sched_group *);
1763                 if (!sdd->sg)
1764                         return -ENOMEM;
1765
1766                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1767                 if (!sdd->sgc)
1768                         return -ENOMEM;
1769
1770                 for_each_cpu(j, cpu_map) {
1771                         struct sched_domain *sd;
1772                         struct sched_domain_shared *sds;
1773                         struct sched_group *sg;
1774                         struct sched_group_capacity *sgc;
1775
1776                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1777                                         GFP_KERNEL, cpu_to_node(j));
1778                         if (!sd)
1779                                 return -ENOMEM;
1780
1781                         *per_cpu_ptr(sdd->sd, j) = sd;
1782
1783                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
1784                                         GFP_KERNEL, cpu_to_node(j));
1785                         if (!sds)
1786                                 return -ENOMEM;
1787
1788                         *per_cpu_ptr(sdd->sds, j) = sds;
1789
1790                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1791                                         GFP_KERNEL, cpu_to_node(j));
1792                         if (!sg)
1793                                 return -ENOMEM;
1794
1795                         sg->next = sg;
1796
1797                         *per_cpu_ptr(sdd->sg, j) = sg;
1798
1799                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1800                                         GFP_KERNEL, cpu_to_node(j));
1801                         if (!sgc)
1802                                 return -ENOMEM;
1803
1804 #ifdef CONFIG_SCHED_DEBUG
1805                         sgc->id = j;
1806 #endif
1807
1808                         *per_cpu_ptr(sdd->sgc, j) = sgc;
1809                 }
1810         }
1811
1812         return 0;
1813 }
1814
1815 static void __sdt_free(const struct cpumask *cpu_map)
1816 {
1817         struct sched_domain_topology_level *tl;
1818         int j;
1819
1820         for_each_sd_topology(tl) {
1821                 struct sd_data *sdd = &tl->data;
1822
1823                 for_each_cpu(j, cpu_map) {
1824                         struct sched_domain *sd;
1825
1826                         if (sdd->sd) {
1827                                 sd = *per_cpu_ptr(sdd->sd, j);
1828                                 if (sd && (sd->flags & SD_OVERLAP))
1829                                         free_sched_groups(sd->groups, 0);
1830                                 kfree(*per_cpu_ptr(sdd->sd, j));
1831                         }
1832
1833                         if (sdd->sds)
1834                                 kfree(*per_cpu_ptr(sdd->sds, j));
1835                         if (sdd->sg)
1836                                 kfree(*per_cpu_ptr(sdd->sg, j));
1837                         if (sdd->sgc)
1838                                 kfree(*per_cpu_ptr(sdd->sgc, j));
1839                 }
1840                 free_percpu(sdd->sd);
1841                 sdd->sd = NULL;
1842                 free_percpu(sdd->sds);
1843                 sdd->sds = NULL;
1844                 free_percpu(sdd->sg);
1845                 sdd->sg = NULL;
1846                 free_percpu(sdd->sgc);
1847                 sdd->sgc = NULL;
1848         }
1849 }
1850
1851 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1852                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1853                 struct sched_domain *child, int dflags, int cpu)
1854 {
1855         struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1856
1857         if (child) {
1858                 sd->level = child->level + 1;
1859                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1860                 child->parent = sd;
1861
1862                 if (!cpumask_subset(sched_domain_span(child),
1863                                     sched_domain_span(sd))) {
1864                         pr_err("BUG: arch topology borken\n");
1865 #ifdef CONFIG_SCHED_DEBUG
1866                         pr_err("     the %s domain not a subset of the %s domain\n",
1867                                         child->name, sd->name);
1868 #endif
1869                         /* Fixup, ensure @sd has at least @child CPUs. */
1870                         cpumask_or(sched_domain_span(sd),
1871                                    sched_domain_span(sd),
1872                                    sched_domain_span(child));
1873                 }
1874
1875         }
1876         set_domain_attribute(sd, attr);
1877
1878         return sd;
1879 }
1880
1881 /*
1882  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
1883  * any two given CPUs at this (non-NUMA) topology level.
1884  */
1885 static bool topology_span_sane(struct sched_domain_topology_level *tl,
1886                               const struct cpumask *cpu_map, int cpu)
1887 {
1888         int i;
1889
1890         /* NUMA levels are allowed to overlap */
1891         if (tl->flags & SDTL_OVERLAP)
1892                 return true;
1893
1894         /*
1895          * Non-NUMA levels cannot partially overlap - they must be either
1896          * completely equal or completely disjoint. Otherwise we can end up
1897          * breaking the sched_group lists - i.e. a later get_group() pass
1898          * breaks the linking done for an earlier span.
1899          */
1900         for_each_cpu(i, cpu_map) {
1901                 if (i == cpu)
1902                         continue;
1903                 /*
1904                  * We should 'and' all those masks with 'cpu_map' to exactly
1905                  * match the topology we're about to build, but that can only
1906                  * remove CPUs, which only lessens our ability to detect
1907                  * overlaps
1908                  */
1909                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
1910                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
1911                         return false;
1912         }
1913
1914         return true;
1915 }
1916
1917 /*
1918  * Find the sched_domain_topology_level where all CPU capacities are visible
1919  * for all CPUs.
1920  */
1921 static struct sched_domain_topology_level
1922 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1923 {
1924         int i, j, asym_level = 0;
1925         bool asym = false;
1926         struct sched_domain_topology_level *tl, *asym_tl = NULL;
1927         unsigned long cap;
1928
1929         /* Is there any asymmetry? */
1930         cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1931
1932         for_each_cpu(i, cpu_map) {
1933                 if (arch_scale_cpu_capacity(i) != cap) {
1934                         asym = true;
1935                         break;
1936                 }
1937         }
1938
1939         if (!asym)
1940                 return NULL;
1941
1942         /*
1943          * Examine topology from all CPU's point of views to detect the lowest
1944          * sched_domain_topology_level where a highest capacity CPU is visible
1945          * to everyone.
1946          */
1947         for_each_cpu(i, cpu_map) {
1948                 unsigned long max_capacity = arch_scale_cpu_capacity(i);
1949                 int tl_id = 0;
1950
1951                 for_each_sd_topology(tl) {
1952                         if (tl_id < asym_level)
1953                                 goto next_level;
1954
1955                         for_each_cpu_and(j, tl->mask(i), cpu_map) {
1956                                 unsigned long capacity;
1957
1958                                 capacity = arch_scale_cpu_capacity(j);
1959
1960                                 if (capacity <= max_capacity)
1961                                         continue;
1962
1963                                 max_capacity = capacity;
1964                                 asym_level = tl_id;
1965                                 asym_tl = tl;
1966                         }
1967 next_level:
1968                         tl_id++;
1969                 }
1970         }
1971
1972         return asym_tl;
1973 }
1974
1975
1976 /*
1977  * Build sched domains for a given set of CPUs and attach the sched domains
1978  * to the individual CPUs
1979  */
1980 static int
1981 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1982 {
1983         enum s_alloc alloc_state = sa_none;
1984         struct sched_domain *sd;
1985         struct s_data d;
1986         struct rq *rq = NULL;
1987         int i, ret = -ENOMEM;
1988         struct sched_domain_topology_level *tl_asym;
1989         bool has_asym = false;
1990
1991         if (WARN_ON(cpumask_empty(cpu_map)))
1992                 goto error;
1993
1994         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1995         if (alloc_state != sa_rootdomain)
1996                 goto error;
1997
1998         tl_asym = asym_cpu_capacity_level(cpu_map);
1999
2000         /* Set up domains for CPUs specified by the cpu_map: */
2001         for_each_cpu(i, cpu_map) {
2002                 struct sched_domain_topology_level *tl;
2003
2004                 sd = NULL;
2005                 for_each_sd_topology(tl) {
2006                         int dflags = 0;
2007
2008                         if (tl == tl_asym) {
2009                                 dflags |= SD_ASYM_CPUCAPACITY;
2010                                 has_asym = true;
2011                         }
2012
2013                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2014                                 goto error;
2015
2016                         sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2017
2018                         if (tl == sched_domain_topology)
2019                                 *per_cpu_ptr(d.sd, i) = sd;
2020                         if (tl->flags & SDTL_OVERLAP)
2021                                 sd->flags |= SD_OVERLAP;
2022                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2023                                 break;
2024                 }
2025         }
2026
2027         /* Build the groups for the domains */
2028         for_each_cpu(i, cpu_map) {
2029                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2030                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2031                         if (sd->flags & SD_OVERLAP) {
2032                                 if (build_overlap_sched_groups(sd, i))
2033                                         goto error;
2034                         } else {
2035                                 if (build_sched_groups(sd, i))
2036                                         goto error;
2037                         }
2038                 }
2039         }
2040
2041         /* Calculate CPU capacity for physical packages and nodes */
2042         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2043                 if (!cpumask_test_cpu(i, cpu_map))
2044                         continue;
2045
2046                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2047                         claim_allocations(i, sd);
2048                         init_sched_groups_capacity(i, sd);
2049                 }
2050         }
2051
2052         /* Attach the domains */
2053         rcu_read_lock();
2054         for_each_cpu(i, cpu_map) {
2055                 rq = cpu_rq(i);
2056                 sd = *per_cpu_ptr(d.sd, i);
2057
2058                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2059                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2060                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2061
2062                 cpu_attach_domain(sd, d.rd, i);
2063         }
2064         rcu_read_unlock();
2065
2066         if (has_asym)
2067                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2068
2069         if (rq && sched_debug_enabled) {
2070                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2071                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2072         }
2073
2074         ret = 0;
2075 error:
2076         __free_domain_allocs(&d, alloc_state, cpu_map);
2077
2078         return ret;
2079 }
2080
2081 /* Current sched domains: */
2082 static cpumask_var_t                    *doms_cur;
2083
2084 /* Number of sched domains in 'doms_cur': */
2085 static int                              ndoms_cur;
2086
2087 /* Attribues of custom domains in 'doms_cur' */
2088 static struct sched_domain_attr         *dattr_cur;
2089
2090 /*
2091  * Special case: If a kmalloc() of a doms_cur partition (array of
2092  * cpumask) fails, then fallback to a single sched domain,
2093  * as determined by the single cpumask fallback_doms.
2094  */
2095 static cpumask_var_t                    fallback_doms;
2096
2097 /*
2098  * arch_update_cpu_topology lets virtualized architectures update the
2099  * CPU core maps. It is supposed to return 1 if the topology changed
2100  * or 0 if it stayed the same.
2101  */
2102 int __weak arch_update_cpu_topology(void)
2103 {
2104         return 0;
2105 }
2106
2107 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2108 {
2109         int i;
2110         cpumask_var_t *doms;
2111
2112         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2113         if (!doms)
2114                 return NULL;
2115         for (i = 0; i < ndoms; i++) {
2116                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2117                         free_sched_domains(doms, i);
2118                         return NULL;
2119                 }
2120         }
2121         return doms;
2122 }
2123
2124 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2125 {
2126         unsigned int i;
2127         for (i = 0; i < ndoms; i++)
2128                 free_cpumask_var(doms[i]);
2129         kfree(doms);
2130 }
2131
2132 /*
2133  * Set up scheduler domains and groups.  For now this just excludes isolated
2134  * CPUs, but could be used to exclude other special cases in the future.
2135  */
2136 int sched_init_domains(const struct cpumask *cpu_map)
2137 {
2138         int err;
2139
2140         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2141         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2142         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2143
2144         arch_update_cpu_topology();
2145         ndoms_cur = 1;
2146         doms_cur = alloc_sched_domains(ndoms_cur);
2147         if (!doms_cur)
2148                 doms_cur = &fallback_doms;
2149         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2150         err = build_sched_domains(doms_cur[0], NULL);
2151         register_sched_domain_sysctl();
2152
2153         return err;
2154 }
2155
2156 /*
2157  * Detach sched domains from a group of CPUs specified in cpu_map
2158  * These CPUs will now be attached to the NULL domain
2159  */
2160 static void detach_destroy_domains(const struct cpumask *cpu_map)
2161 {
2162         unsigned int cpu = cpumask_any(cpu_map);
2163         int i;
2164
2165         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2166                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2167
2168         rcu_read_lock();
2169         for_each_cpu(i, cpu_map)
2170                 cpu_attach_domain(NULL, &def_root_domain, i);
2171         rcu_read_unlock();
2172 }
2173
2174 /* handle null as "default" */
2175 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2176                         struct sched_domain_attr *new, int idx_new)
2177 {
2178         struct sched_domain_attr tmp;
2179
2180         /* Fast path: */
2181         if (!new && !cur)
2182                 return 1;
2183
2184         tmp = SD_ATTR_INIT;
2185
2186         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2187                         new ? (new + idx_new) : &tmp,
2188                         sizeof(struct sched_domain_attr));
2189 }
2190
2191 /*
2192  * Partition sched domains as specified by the 'ndoms_new'
2193  * cpumasks in the array doms_new[] of cpumasks. This compares
2194  * doms_new[] to the current sched domain partitioning, doms_cur[].
2195  * It destroys each deleted domain and builds each new domain.
2196  *
2197  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2198  * The masks don't intersect (don't overlap.) We should setup one
2199  * sched domain for each mask. CPUs not in any of the cpumasks will
2200  * not be load balanced. If the same cpumask appears both in the
2201  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2202  * it as it is.
2203  *
2204  * The passed in 'doms_new' should be allocated using
2205  * alloc_sched_domains.  This routine takes ownership of it and will
2206  * free_sched_domains it when done with it. If the caller failed the
2207  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2208  * and partition_sched_domains() will fallback to the single partition
2209  * 'fallback_doms', it also forces the domains to be rebuilt.
2210  *
2211  * If doms_new == NULL it will be replaced with cpu_online_mask.
2212  * ndoms_new == 0 is a special case for destroying existing domains,
2213  * and it will not create the default domain.
2214  *
2215  * Call with hotplug lock and sched_domains_mutex held
2216  */
2217 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2218                                     struct sched_domain_attr *dattr_new)
2219 {
2220         bool __maybe_unused has_eas = false;
2221         int i, j, n;
2222         int new_topology;
2223
2224         lockdep_assert_held(&sched_domains_mutex);
2225
2226         /* Always unregister in case we don't destroy any domains: */
2227         unregister_sched_domain_sysctl();
2228
2229         /* Let the architecture update CPU core mappings: */
2230         new_topology = arch_update_cpu_topology();
2231
2232         if (!doms_new) {
2233                 WARN_ON_ONCE(dattr_new);
2234                 n = 0;
2235                 doms_new = alloc_sched_domains(1);
2236                 if (doms_new) {
2237                         n = 1;
2238                         cpumask_and(doms_new[0], cpu_active_mask,
2239                                     housekeeping_cpumask(HK_FLAG_DOMAIN));
2240                 }
2241         } else {
2242                 n = ndoms_new;
2243         }
2244
2245         /* Destroy deleted domains: */
2246         for (i = 0; i < ndoms_cur; i++) {
2247                 for (j = 0; j < n && !new_topology; j++) {
2248                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2249                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2250                                 struct root_domain *rd;
2251
2252                                 /*
2253                                  * This domain won't be destroyed and as such
2254                                  * its dl_bw->total_bw needs to be cleared.  It
2255                                  * will be recomputed in function
2256                                  * update_tasks_root_domain().
2257                                  */
2258                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2259                                 dl_clear_root_domain(rd);
2260                                 goto match1;
2261                         }
2262                 }
2263                 /* No match - a current sched domain not in new doms_new[] */
2264                 detach_destroy_domains(doms_cur[i]);
2265 match1:
2266                 ;
2267         }
2268
2269         n = ndoms_cur;
2270         if (!doms_new) {
2271                 n = 0;
2272                 doms_new = &fallback_doms;
2273                 cpumask_and(doms_new[0], cpu_active_mask,
2274                             housekeeping_cpumask(HK_FLAG_DOMAIN));
2275         }
2276
2277         /* Build new domains: */
2278         for (i = 0; i < ndoms_new; i++) {
2279                 for (j = 0; j < n && !new_topology; j++) {
2280                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2281                             dattrs_equal(dattr_new, i, dattr_cur, j))
2282                                 goto match2;
2283                 }
2284                 /* No match - add a new doms_new */
2285                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2286 match2:
2287                 ;
2288         }
2289
2290 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2291         /* Build perf. domains: */
2292         for (i = 0; i < ndoms_new; i++) {
2293                 for (j = 0; j < n && !sched_energy_update; j++) {
2294                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2295                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2296                                 has_eas = true;
2297                                 goto match3;
2298                         }
2299                 }
2300                 /* No match - add perf. domains for a new rd */
2301                 has_eas |= build_perf_domains(doms_new[i]);
2302 match3:
2303                 ;
2304         }
2305         sched_energy_set(has_eas);
2306 #endif
2307
2308         /* Remember the new sched domains: */
2309         if (doms_cur != &fallback_doms)
2310                 free_sched_domains(doms_cur, ndoms_cur);
2311
2312         kfree(dattr_cur);
2313         doms_cur = doms_new;
2314         dattr_cur = dattr_new;
2315         ndoms_cur = ndoms_new;
2316
2317         register_sched_domain_sysctl();
2318 }
2319
2320 /*
2321  * Call with hotplug lock held
2322  */
2323 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2324                              struct sched_domain_attr *dattr_new)
2325 {
2326         mutex_lock(&sched_domains_mutex);
2327         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2328         mutex_unlock(&sched_domains_mutex);
2329 }