kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5
   6 #include <linux/bsearch.h>
   7
   8 DEFINE_MUTEX(sched_domains_mutex);
   9
  10 /* Protected by sched_domains_mutex: */
  11 static cpumask_var_t sched_domains_tmpmask;
  12 static cpumask_var_t sched_domains_tmpmask2;
  13
  14 #ifdef CONFIG_SCHED_DEBUG
  15
  16 static int __init sched_debug_setup(char *str)
  17 {
  18         sched_debug_verbose = true;
  19
  20         return 0;
  21 }
  22 early_param("sched_verbose", sched_debug_setup);
  23
  24 static inline bool sched_debug(void)
  25 {
  26         return sched_debug_verbose;
  27 }
  28
  29 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
  30 const struct sd_flag_debug sd_flag_debug[] = {
  31 #include <linux/sched/sd_flags.h>
  32 };
  33 #undef SD_FLAG
  34
  35 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
  36                                   struct cpumask *groupmask)
  37 {
  38         struct sched_group *group = sd->groups;
  39         unsigned long flags = sd->flags;
  40         unsigned int idx;
  41
  42         cpumask_clear(groupmask);
  43
  44         printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
  45         printk(KERN_CONT "span=%*pbl level=%s\n",
  46                cpumask_pr_args(sched_domain_span(sd)), sd->name);
  47
  48         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
  49                 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
  50         }
  51         if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
  52                 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
  53         }
  54
  55         for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
  56                 unsigned int flag = BIT(idx);
  57                 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
  58
  59                 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
  60                     !(sd->child->flags & flag))
  61                         printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
  62                                sd_flag_debug[idx].name);
  63
  64                 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
  65                     !(sd->parent->flags & flag))
  66                         printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
  67                                sd_flag_debug[idx].name);
  68         }
  69
  70         printk(KERN_DEBUG "%*s groups:", level + 1, "");
  71         do {
  72                 if (!group) {
  73                         printk("\n");
  74                         printk(KERN_ERR "ERROR: group is NULL\n");
  75                         break;
  76                 }
  77
  78                 if (cpumask_empty(sched_group_span(group))) {
  79                         printk(KERN_CONT "\n");
  80                         printk(KERN_ERR "ERROR: empty group\n");
  81                         break;
  82                 }
  83
  84                 if (!(sd->flags & SD_OVERLAP) &&
  85                     cpumask_intersects(groupmask, sched_group_span(group))) {
  86                         printk(KERN_CONT "\n");
  87                         printk(KERN_ERR "ERROR: repeated CPUs\n");
  88                         break;
  89                 }
  90
  91                 cpumask_or(groupmask, groupmask, sched_group_span(group));
  92
  93                 printk(KERN_CONT " %d:{ span=%*pbl",
  94                                 group->sgc->id,
  95                                 cpumask_pr_args(sched_group_span(group)));
  96
  97                 if ((sd->flags & SD_OVERLAP) &&
  98                     !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
  99                         printk(KERN_CONT " mask=%*pbl",
 100                                 cpumask_pr_args(group_balance_mask(group)));
 101                 }
 102
 103                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
 104                         printk(KERN_CONT " cap=%lu", group->sgc->capacity);
 105
 106                 if (group == sd->groups && sd->child &&
 107                     !cpumask_equal(sched_domain_span(sd->child),
 108                                    sched_group_span(group))) {
 109                         printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
 110                 }
 111
 112                 printk(KERN_CONT " }");
 113
 114                 group = group->next;
 115
 116                 if (group != sd->groups)
 117                         printk(KERN_CONT ",");
 118
 119         } while (group != sd->groups);
 120         printk(KERN_CONT "\n");
 121
 122         if (!cpumask_equal(sched_domain_span(sd), groupmask))
 123                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 124
 125         if (sd->parent &&
 126             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
 127                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
 128         return 0;
 129 }
 130
 131 static void sched_domain_debug(struct sched_domain *sd, int cpu)
 132 {
 133         int level = 0;
 134
 135         if (!sched_debug_verbose)
 136                 return;
 137
 138         if (!sd) {
 139                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
 140                 return;
 141         }
 142
 143         printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
 144
 145         for (;;) {
 146                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
 147                         break;
 148                 level++;
 149                 sd = sd->parent;
 150                 if (!sd)
 151                         break;
 152         }
 153 }
 154 #else /* !CONFIG_SCHED_DEBUG */
 155
 156 # define sched_debug_verbose 0
 157 # define sched_domain_debug(sd, cpu) do { } while (0)
 158 static inline bool sched_debug(void)
 159 {
 160         return false;
 161 }
 162 #endif /* CONFIG_SCHED_DEBUG */
 163
 164 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
 165 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
 166 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
 167 #include <linux/sched/sd_flags.h>
 168 0;
 169 #undef SD_FLAG
 170
 171 static int sd_degenerate(struct sched_domain *sd)
 172 {
 173         if (cpumask_weight(sched_domain_span(sd)) == 1)
 174                 return 1;
 175
 176         /* Following flags need at least 2 groups */
 177         if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
 178             (sd->groups != sd->groups->next))
 179                 return 0;
 180
 181         /* Following flags don't use groups */
 182         if (sd->flags & (SD_WAKE_AFFINE))
 183                 return 0;
 184
 185         return 1;
 186 }
 187
 188 static int
 189 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
 190 {
 191         unsigned long cflags = sd->flags, pflags = parent->flags;
 192
 193         if (sd_degenerate(parent))
 194                 return 1;
 195
 196         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
 197                 return 0;
 198
 199         /* Flags needing groups don't count if only 1 group in parent */
 200         if (parent->groups == parent->groups->next)
 201                 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
 202
 203         if (~cflags & pflags)
 204                 return 0;
 205
 206         return 1;
 207 }
 208
 209 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
 210 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
 211 static unsigned int sysctl_sched_energy_aware = 1;
 212 static DEFINE_MUTEX(sched_energy_mutex);
 213 static bool sched_energy_update;
 214
 215 static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
 216 {
 217         bool any_asym_capacity = false;
 218         struct cpufreq_policy *policy;
 219         struct cpufreq_governor *gov;
 220         int i;
 221
 222         /* EAS is enabled for asymmetric CPU capacity topologies. */
 223         for_each_cpu(i, cpu_mask) {
 224                 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
 225                         any_asym_capacity = true;
 226                         break;
 227                 }
 228         }
 229         if (!any_asym_capacity) {
 230                 if (sched_debug()) {
 231                         pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
 232                                 cpumask_pr_args(cpu_mask));
 233                 }
 234                 return false;
 235         }
 236
 237         /* EAS definitely does *not* handle SMT */
 238         if (sched_smt_active()) {
 239                 if (sched_debug()) {
 240                         pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
 241                                 cpumask_pr_args(cpu_mask));
 242                 }
 243                 return false;
 244         }
 245
 246         if (!arch_scale_freq_invariant()) {
 247                 if (sched_debug()) {
 248                         pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
 249                                 cpumask_pr_args(cpu_mask));
 250                 }
 251                 return false;
 252         }
 253
 254         /* Do not attempt EAS if schedutil is not being used. */
 255         for_each_cpu(i, cpu_mask) {
 256                 policy = cpufreq_cpu_get(i);
 257                 if (!policy) {
 258                         if (sched_debug()) {
 259                                 pr_info("rd %*pbl: Checking EAS, cpufreq policy not set for CPU: %d",
 260                                         cpumask_pr_args(cpu_mask), i);
 261                         }
 262                         return false;
 263                 }
 264                 gov = policy->governor;
 265                 cpufreq_cpu_put(policy);
 266                 if (gov != &schedutil_gov) {
 267                         if (sched_debug()) {
 268                                 pr_info("rd %*pbl: Checking EAS, schedutil is mandatory\n",
 269                                         cpumask_pr_args(cpu_mask));
 270                         }
 271                         return false;
 272                 }
 273         }
 274
 275         return true;
 276 }
 277
 278 void rebuild_sched_domains_energy(void)
 279 {
 280         mutex_lock(&sched_energy_mutex);
 281         sched_energy_update = true;
 282         rebuild_sched_domains();
 283         sched_energy_update = false;
 284         mutex_unlock(&sched_energy_mutex);
 285 }
 286
 287 #ifdef CONFIG_PROC_SYSCTL
 288 static int sched_energy_aware_handler(struct ctl_table *table, int write,
 289                 void *buffer, size_t *lenp, loff_t *ppos)
 290 {
 291         int ret, state;
 292
 293         if (write && !capable(CAP_SYS_ADMIN))
 294                 return -EPERM;
 295
 296         if (!sched_is_eas_possible(cpu_active_mask)) {
 297                 if (write) {
 298                         return -EOPNOTSUPP;
 299                 } else {
 300                         *lenp = 0;
 301                         return 0;
 302                 }
 303         }
 304
 305         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 306         if (!ret && write) {
 307                 state = static_branch_unlikely(&sched_energy_present);
 308                 if (state != sysctl_sched_energy_aware)
 309                         rebuild_sched_domains_energy();
 310         }
 311
 312         return ret;
 313 }
 314
 315 static struct ctl_table sched_energy_aware_sysctls[] = {
 316         {
 317                 .procname       = "sched_energy_aware",
 318                 .data           = &sysctl_sched_energy_aware,
 319                 .maxlen         = sizeof(unsigned int),
 320                 .mode           = 0644,
 321                 .proc_handler   = sched_energy_aware_handler,
 322                 .extra1         = SYSCTL_ZERO,
 323                 .extra2         = SYSCTL_ONE,
 324         },
 325 };
 326
 327 static int __init sched_energy_aware_sysctl_init(void)
 328 {
 329         register_sysctl_init("kernel", sched_energy_aware_sysctls);
 330         return 0;
 331 }
 332
 333 late_initcall(sched_energy_aware_sysctl_init);
 334 #endif
 335
 336 static void free_pd(struct perf_domain *pd)
 337 {
 338         struct perf_domain *tmp;
 339
 340         while (pd) {
 341                 tmp = pd->next;
 342                 kfree(pd);
 343                 pd = tmp;
 344         }
 345 }
 346
 347 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
 348 {
 349         while (pd) {
 350                 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
 351                         return pd;
 352                 pd = pd->next;
 353         }
 354
 355         return NULL;
 356 }
 357
 358 static struct perf_domain *pd_init(int cpu)
 359 {
 360         struct em_perf_domain *obj = em_cpu_get(cpu);
 361         struct perf_domain *pd;
 362
 363         if (!obj) {
 364                 if (sched_debug())
 365                         pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
 366                 return NULL;
 367         }
 368
 369         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 370         if (!pd)
 371                 return NULL;
 372         pd->em_pd = obj;
 373
 374         return pd;
 375 }
 376
 377 static void perf_domain_debug(const struct cpumask *cpu_map,
 378                                                 struct perf_domain *pd)
 379 {
 380         if (!sched_debug() || !pd)
 381                 return;
 382
 383         printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
 384
 385         while (pd) {
 386                 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
 387                                 cpumask_first(perf_domain_span(pd)),
 388                                 cpumask_pr_args(perf_domain_span(pd)),
 389                                 em_pd_nr_perf_states(pd->em_pd));
 390                 pd = pd->next;
 391         }
 392
 393         printk(KERN_CONT "\n");
 394 }
 395
 396 static void destroy_perf_domain_rcu(struct rcu_head *rp)
 397 {
 398         struct perf_domain *pd;
 399
 400         pd = container_of(rp, struct perf_domain, rcu);
 401         free_pd(pd);
 402 }
 403
 404 static void sched_energy_set(bool has_eas)
 405 {
 406         if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
 407                 if (sched_debug())
 408                         pr_info("%s: stopping EAS\n", __func__);
 409                 static_branch_disable_cpuslocked(&sched_energy_present);
 410         } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
 411                 if (sched_debug())
 412                         pr_info("%s: starting EAS\n", __func__);
 413                 static_branch_enable_cpuslocked(&sched_energy_present);
 414         }
 415 }
 416
 417 /*
 418  * EAS can be used on a root domain if it meets all the following conditions:
 419  *    1. an Energy Model (EM) is available;
 420  *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
 421  *    3. no SMT is detected.
 422  *    4. schedutil is driving the frequency of all CPUs of the rd;
 423  *    5. frequency invariance support is present;
 424  */
 425 static bool build_perf_domains(const struct cpumask *cpu_map)
 426 {
 427         int i;
 428         struct perf_domain *pd = NULL, *tmp;
 429         int cpu = cpumask_first(cpu_map);
 430         struct root_domain *rd = cpu_rq(cpu)->rd;
 431
 432         if (!sysctl_sched_energy_aware)
 433                 goto free;
 434
 435         if (!sched_is_eas_possible(cpu_map))
 436                 goto free;
 437
 438         for_each_cpu(i, cpu_map) {
 439                 /* Skip already covered CPUs. */
 440                 if (find_pd(pd, i))
 441                         continue;
 442
 443                 /* Create the new pd and add it to the local list. */
 444                 tmp = pd_init(i);
 445                 if (!tmp)
 446                         goto free;
 447                 tmp->next = pd;
 448                 pd = tmp;
 449         }
 450
 451         perf_domain_debug(cpu_map, pd);
 452
 453         /* Attach the new list of performance domains to the root domain. */
 454         tmp = rd->pd;
 455         rcu_assign_pointer(rd->pd, pd);
 456         if (tmp)
 457                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 458
 459         return !!pd;
 460
 461 free:
 462         free_pd(pd);
 463         tmp = rd->pd;
 464         rcu_assign_pointer(rd->pd, NULL);
 465         if (tmp)
 466                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 467
 468         return false;
 469 }
 470 #else
 471 static void free_pd(struct perf_domain *pd) { }
 472 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
 473
 474 static void free_rootdomain(struct rcu_head *rcu)
 475 {
 476         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 477
 478         cpupri_cleanup(&rd->cpupri);
 479         cpudl_cleanup(&rd->cpudl);
 480         free_cpumask_var(rd->dlo_mask);
 481         free_cpumask_var(rd->rto_mask);
 482         free_cpumask_var(rd->online);
 483         free_cpumask_var(rd->span);
 484         free_pd(rd->pd);
 485         kfree(rd);
 486 }
 487
 488 void rq_attach_root(struct rq *rq, struct root_domain *rd)
 489 {
 490         struct root_domain *old_rd = NULL;
 491         struct rq_flags rf;
 492
 493         rq_lock_irqsave(rq, &rf);
 494
 495         if (rq->rd) {
 496                 old_rd = rq->rd;
 497
 498                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
 499                         set_rq_offline(rq);
 500
 501                 cpumask_clear_cpu(rq->cpu, old_rd->span);
 502
 503                 /*
 504                  * If we dont want to free the old_rd yet then
 505                  * set old_rd to NULL to skip the freeing later
 506                  * in this function:
 507                  */
 508                 if (!atomic_dec_and_test(&old_rd->refcount))
 509                         old_rd = NULL;
 510         }
 511
 512         atomic_inc(&rd->refcount);
 513         rq->rd = rd;
 514
 515         cpumask_set_cpu(rq->cpu, rd->span);
 516         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 517                 set_rq_online(rq);
 518
 519         rq_unlock_irqrestore(rq, &rf);
 520
 521         if (old_rd)
 522                 call_rcu(&old_rd->rcu, free_rootdomain);
 523 }
 524
 525 void sched_get_rd(struct root_domain *rd)
 526 {
 527         atomic_inc(&rd->refcount);
 528 }
 529
 530 void sched_put_rd(struct root_domain *rd)
 531 {
 532         if (!atomic_dec_and_test(&rd->refcount))
 533                 return;
 534
 535         call_rcu(&rd->rcu, free_rootdomain);
 536 }
 537
 538 static int init_rootdomain(struct root_domain *rd)
 539 {
 540         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 541                 goto out;
 542         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 543                 goto free_span;
 544         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 545                 goto free_online;
 546         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 547                 goto free_dlo_mask;
 548
 549 #ifdef HAVE_RT_PUSH_IPI
 550         rd->rto_cpu = -1;
 551         raw_spin_lock_init(&rd->rto_lock);
 552         rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
 553 #endif
 554
 555         rd->visit_gen = 0;
 556         init_dl_bw(&rd->dl_bw);
 557         if (cpudl_init(&rd->cpudl) != 0)
 558                 goto free_rto_mask;
 559
 560         if (cpupri_init(&rd->cpupri) != 0)
 561                 goto free_cpudl;
 562         return 0;
 563
 564 free_cpudl:
 565         cpudl_cleanup(&rd->cpudl);
 566 free_rto_mask:
 567         free_cpumask_var(rd->rto_mask);
 568 free_dlo_mask:
 569         free_cpumask_var(rd->dlo_mask);
 570 free_online:
 571         free_cpumask_var(rd->online);
 572 free_span:
 573         free_cpumask_var(rd->span);
 574 out:
 575         return -ENOMEM;
 576 }
 577
 578 /*
 579  * By default the system creates a single root-domain with all CPUs as
 580  * members (mimicking the global state we have today).
 581  */
 582 struct root_domain def_root_domain;
 583
 584 void __init init_defrootdomain(void)
 585 {
 586         init_rootdomain(&def_root_domain);
 587
 588         atomic_set(&def_root_domain.refcount, 1);
 589 }
 590
 591 static struct root_domain *alloc_rootdomain(void)
 592 {
 593         struct root_domain *rd;
 594
 595         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 596         if (!rd)
 597                 return NULL;
 598
 599         if (init_rootdomain(rd) != 0) {
 600                 kfree(rd);
 601                 return NULL;
 602         }
 603
 604         return rd;
 605 }
 606
 607 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 608 {
 609         struct sched_group *tmp, *first;
 610
 611         if (!sg)
 612                 return;
 613
 614         first = sg;
 615         do {
 616                 tmp = sg->next;
 617
 618                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 619                         kfree(sg->sgc);
 620
 621                 if (atomic_dec_and_test(&sg->ref))
 622                         kfree(sg);
 623                 sg = tmp;
 624         } while (sg != first);
 625 }
 626
 627 static void destroy_sched_domain(struct sched_domain *sd)
 628 {
 629         /*
 630          * A normal sched domain may have multiple group references, an
 631          * overlapping domain, having private groups, only one.  Iterate,
 632          * dropping group/capacity references, freeing where none remain.
 633          */
 634         free_sched_groups(sd->groups, 1);
 635
 636         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 637                 kfree(sd->shared);
 638         kfree(sd);
 639 }
 640
 641 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 642 {
 643         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 644
 645         while (sd) {
 646                 struct sched_domain *parent = sd->parent;
 647                 destroy_sched_domain(sd);
 648                 sd = parent;
 649         }
 650 }
 651
 652 static void destroy_sched_domains(struct sched_domain *sd)
 653 {
 654         if (sd)
 655                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 656 }
 657
 658 /*
 659  * Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
 660  * (Last Level Cache Domain) for this allows us to avoid some pointer chasing
 661  * select_idle_sibling().
 662  *
 663  * Also keep a unique ID per domain (we use the first CPU number in the cpumask
 664  * of the domain), this allows us to quickly tell if two CPUs are in the same
 665  * cache domain, see cpus_share_cache().
 666  */
 667 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 668 DEFINE_PER_CPU(int, sd_llc_size);
 669 DEFINE_PER_CPU(int, sd_llc_id);
 670 DEFINE_PER_CPU(int, sd_share_id);
 671 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 672 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 673 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 674 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 675
 676 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 677 DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
 678
 679 static void update_top_cache_domain(int cpu)
 680 {
 681         struct sched_domain_shared *sds = NULL;
 682         struct sched_domain *sd;
 683         int id = cpu;
 684         int size = 1;
 685
 686         sd = highest_flag_domain(cpu, SD_SHARE_LLC);
 687         if (sd) {
 688                 id = cpumask_first(sched_domain_span(sd));
 689                 size = cpumask_weight(sched_domain_span(sd));
 690                 sds = sd->shared;
 691         }
 692
 693         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 694         per_cpu(sd_llc_size, cpu) = size;
 695         per_cpu(sd_llc_id, cpu) = id;
 696         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 697
 698         sd = lowest_flag_domain(cpu, SD_CLUSTER);
 699         if (sd)
 700                 id = cpumask_first(sched_domain_span(sd));
 701
 702         /*
 703          * This assignment should be placed after the sd_llc_id as
 704          * we want this id equals to cluster id on cluster machines
 705          * but equals to LLC id on non-Cluster machines.
 706          */
 707         per_cpu(sd_share_id, cpu) = id;
 708
 709         sd = lowest_flag_domain(cpu, SD_NUMA);
 710         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 711
 712         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 713         rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 714
 715         sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
 716         rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 717 }
 718
 719 /*
 720  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 721  * hold the hotplug lock.
 722  */
 723 static void
 724 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 725 {
 726         struct rq *rq = cpu_rq(cpu);
 727         struct sched_domain *tmp;
 728
 729         /* Remove the sched domains which do not contribute to scheduling. */
 730         for (tmp = sd; tmp; ) {
 731                 struct sched_domain *parent = tmp->parent;
 732                 if (!parent)
 733                         break;
 734
 735                 if (sd_parent_degenerate(tmp, parent)) {
 736                         tmp->parent = parent->parent;
 737
 738                         if (parent->parent) {
 739                                 parent->parent->child = tmp;
 740                                 parent->parent->groups->flags = tmp->flags;
 741                         }
 742
 743                         /*
 744                          * Transfer SD_PREFER_SIBLING down in case of a
 745                          * degenerate parent; the spans match for this
 746                          * so the property transfers.
 747                          */
 748                         if (parent->flags & SD_PREFER_SIBLING)
 749                                 tmp->flags |= SD_PREFER_SIBLING;
 750                         destroy_sched_domain(parent);
 751                 } else
 752                         tmp = tmp->parent;
 753         }
 754
 755         if (sd && sd_degenerate(sd)) {
 756                 tmp = sd;
 757                 sd = sd->parent;
 758                 destroy_sched_domain(tmp);
 759                 if (sd) {
 760                         struct sched_group *sg = sd->groups;
 761
 762                         /*
 763                          * sched groups hold the flags of the child sched
 764                          * domain for convenience. Clear such flags since
 765                          * the child is being destroyed.
 766                          */
 767                         do {
 768                                 sg->flags = 0;
 769                         } while (sg != sd->groups);
 770
 771                         sd->child = NULL;
 772                 }
 773         }
 774
 775         sched_domain_debug(sd, cpu);
 776
 777         rq_attach_root(rq, rd);
 778         tmp = rq->sd;
 779         rcu_assign_pointer(rq->sd, sd);
 780         dirty_sched_domain_sysctl(cpu);
 781         destroy_sched_domains(tmp);
 782
 783         update_top_cache_domain(cpu);
 784 }
 785
 786 struct s_data {
 787         struct sched_domain * __percpu *sd;
 788         struct root_domain      *rd;
 789 };
 790
 791 enum s_alloc {
 792         sa_rootdomain,
 793         sa_sd,
 794         sa_sd_storage,
 795         sa_none,
 796 };
 797
 798 /*
 799  * Return the canonical balance CPU for this group, this is the first CPU
 800  * of this group that's also in the balance mask.
 801  *
 802  * The balance mask are all those CPUs that could actually end up at this
 803  * group. See build_balance_mask().
 804  *
 805  * Also see should_we_balance().
 806  */
 807 int group_balance_cpu(struct sched_group *sg)
 808 {
 809         return cpumask_first(group_balance_mask(sg));
 810 }
 811
 812
 813 /*
 814  * NUMA topology (first read the regular topology blurb below)
 815  *
 816  * Given a node-distance table, for example:
 817  *
 818  *   node   0   1   2   3
 819  *     0:  10  20  30  20
 820  *     1:  20  10  20  30
 821  *     2:  30  20  10  20
 822  *     3:  20  30  20  10
 823  *
 824  * which represents a 4 node ring topology like:
 825  *
 826  *   0 ----- 1
 827  *   |       |
 828  *   |       |
 829  *   |       |
 830  *   3 ----- 2
 831  *
 832  * We want to construct domains and groups to represent this. The way we go
 833  * about doing this is to build the domains on 'hops'. For each NUMA level we
 834  * construct the mask of all nodes reachable in @level hops.
 835  *
 836  * For the above NUMA topology that gives 3 levels:
 837  *
 838  * NUMA-2       0-3             0-3             0-3             0-3
 839  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 840  *
 841  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 842  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 843  *
 844  * NUMA-0       0               1               2               3
 845  *
 846  *
 847  * As can be seen; things don't nicely line up as with the regular topology.
 848  * When we iterate a domain in child domain chunks some nodes can be
 849  * represented multiple times -- hence the "overlap" naming for this part of
 850  * the topology.
 851  *
 852  * In order to minimize this overlap, we only build enough groups to cover the
 853  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 854  *
 855  * Because:
 856  *
 857  *  - the first group of each domain is its child domain; this
 858  *    gets us the first 0-1,3
 859  *  - the only uncovered node is 2, who's child domain is 1-3.
 860  *
 861  * However, because of the overlap, computing a unique CPU for each group is
 862  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 863  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 864  * end up at those groups (they would end up in group: 0-1,3).
 865  *
 866  * To correct this we have to introduce the group balance mask. This mask
 867  * will contain those CPUs in the group that can reach this group given the
 868  * (child) domain tree.
 869  *
 870  * With this we can once again compute balance_cpu and sched_group_capacity
 871  * relations.
 872  *
 873  * XXX include words on how balance_cpu is unique and therefore can be
 874  * used for sched_group_capacity links.
 875  *
 876  *
 877  * Another 'interesting' topology is:
 878  *
 879  *   node   0   1   2   3
 880  *     0:  10  20  20  30
 881  *     1:  20  10  20  20
 882  *     2:  20  20  10  20
 883  *     3:  30  20  20  10
 884  *
 885  * Which looks a little like:
 886  *
 887  *   0 ----- 1
 888  *   |     / |
 889  *   |   /   |
 890  *   | /     |
 891  *   2 ----- 3
 892  *
 893  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 894  * are not.
 895  *
 896  * This leads to a few particularly weird cases where the sched_domain's are
 897  * not of the same number for each CPU. Consider:
 898  *
 899  * NUMA-2       0-3                                             0-3
 900  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 901  *
 902  * NUMA-1       0-2             0-3             0-3             1-3
 903  *
 904  * NUMA-0       0               1               2               3
 905  *
 906  */
 907
 908
 909 /*
 910  * Build the balance mask; it contains only those CPUs that can arrive at this
 911  * group and should be considered to continue balancing.
 912  *
 913  * We do this during the group creation pass, therefore the group information
 914  * isn't complete yet, however since each group represents a (child) domain we
 915  * can fully construct this using the sched_domain bits (which are already
 916  * complete).
 917  */
 918 static void
 919 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 920 {
 921         const struct cpumask *sg_span = sched_group_span(sg);
 922         struct sd_data *sdd = sd->private;
 923         struct sched_domain *sibling;
 924         int i;
 925
 926         cpumask_clear(mask);
 927
 928         for_each_cpu(i, sg_span) {
 929                 sibling = *per_cpu_ptr(sdd->sd, i);
 930
 931                 /*
 932                  * Can happen in the asymmetric case, where these siblings are
 933                  * unused. The mask will not be empty because those CPUs that
 934                  * do have the top domain _should_ span the domain.
 935                  */
 936                 if (!sibling->child)
 937                         continue;
 938
 939                 /* If we would not end up here, we can't continue from here */
 940                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 941                         continue;
 942
 943                 cpumask_set_cpu(i, mask);
 944         }
 945
 946         /* We must not have empty masks here */
 947         WARN_ON_ONCE(cpumask_empty(mask));
 948 }
 949
 950 /*
 951  * XXX: This creates per-node group entries; since the load-balancer will
 952  * immediately access remote memory to construct this group's load-balance
 953  * statistics having the groups node local is of dubious benefit.
 954  */
 955 static struct sched_group *
 956 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 957 {
 958         struct sched_group *sg;
 959         struct cpumask *sg_span;
 960
 961         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 962                         GFP_KERNEL, cpu_to_node(cpu));
 963
 964         if (!sg)
 965                 return NULL;
 966
 967         sg_span = sched_group_span(sg);
 968         if (sd->child) {
 969                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 970                 sg->flags = sd->child->flags;
 971         } else {
 972                 cpumask_copy(sg_span, sched_domain_span(sd));
 973         }
 974
 975         atomic_inc(&sg->ref);
 976         return sg;
 977 }
 978
 979 static void init_overlap_sched_group(struct sched_domain *sd,
 980                                      struct sched_group *sg)
 981 {
 982         struct cpumask *mask = sched_domains_tmpmask2;
 983         struct sd_data *sdd = sd->private;
 984         struct cpumask *sg_span;
 985         int cpu;
 986
 987         build_balance_mask(sd, sg, mask);
 988         cpu = cpumask_first(mask);
 989
 990         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 991         if (atomic_inc_return(&sg->sgc->ref) == 1)
 992                 cpumask_copy(group_balance_mask(sg), mask);
 993         else
 994                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 995
 996         /*
 997          * Initialize sgc->capacity such that even if we mess up the
 998          * domains and no possible iteration will get us here, we won't
 999          * die on a /0 trap.
1000          */
1001         sg_span = sched_group_span(sg);
1002         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
1003         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1004         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1005 }
1006
1007 static struct sched_domain *
1008 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1009 {
1010         /*
1011          * The proper descendant would be the one whose child won't span out
1012          * of sd
1013          */
1014         while (sibling->child &&
1015                !cpumask_subset(sched_domain_span(sibling->child),
1016                                sched_domain_span(sd)))
1017                 sibling = sibling->child;
1018
1019         /*
1020          * As we are referencing sgc across different topology level, we need
1021          * to go down to skip those sched_domains which don't contribute to
1022          * scheduling because they will be degenerated in cpu_attach_domain
1023          */
1024         while (sibling->child &&
1025                cpumask_equal(sched_domain_span(sibling->child),
1026                              sched_domain_span(sibling)))
1027                 sibling = sibling->child;
1028
1029         return sibling;
1030 }
1031
1032 static int
1033 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1034 {
1035         struct sched_group *first = NULL, *last = NULL, *sg;
1036         const struct cpumask *span = sched_domain_span(sd);
1037         struct cpumask *covered = sched_domains_tmpmask;
1038         struct sd_data *sdd = sd->private;
1039         struct sched_domain *sibling;
1040         int i;
1041
1042         cpumask_clear(covered);
1043
1044         for_each_cpu_wrap(i, span, cpu) {
1045                 struct cpumask *sg_span;
1046
1047                 if (cpumask_test_cpu(i, covered))
1048                         continue;
1049
1050                 sibling = *per_cpu_ptr(sdd->sd, i);
1051
1052                 /*
1053                  * Asymmetric node setups can result in situations where the
1054                  * domain tree is of unequal depth, make sure to skip domains
1055                  * that already cover the entire range.
1056                  *
1057                  * In that case build_sched_domains() will have terminated the
1058                  * iteration early and our sibling sd spans will be empty.
1059                  * Domains should always include the CPU they're built on, so
1060                  * check that.
1061                  */
1062                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1063                         continue;
1064
1065                 /*
1066                  * Usually we build sched_group by sibling's child sched_domain
1067                  * But for machines whose NUMA diameter are 3 or above, we move
1068                  * to build sched_group by sibling's proper descendant's child
1069                  * domain because sibling's child sched_domain will span out of
1070                  * the sched_domain being built as below.
1071                  *
1072                  * Smallest diameter=3 topology is:
1073                  *
1074                  *   node   0   1   2   3
1075                  *     0:  10  20  30  40
1076                  *     1:  20  10  20  30
1077                  *     2:  30  20  10  20
1078                  *     3:  40  30  20  10
1079                  *
1080                  *   0 --- 1 --- 2 --- 3
1081                  *
1082                  * NUMA-3       0-3             N/A             N/A             0-3
1083                  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1084                  *
1085                  * NUMA-2       0-2             0-3             0-3             1-3
1086                  *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1087                  *
1088                  * NUMA-1       0-1             0-2             1-3             2-3
1089                  *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1090                  *
1091                  * NUMA-0       0               1               2               3
1092                  *
1093                  * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1094                  * group span isn't a subset of the domain span.
1095                  */
1096                 if (sibling->child &&
1097                     !cpumask_subset(sched_domain_span(sibling->child), span))
1098                         sibling = find_descended_sibling(sd, sibling);
1099
1100                 sg = build_group_from_child_sched_domain(sibling, cpu);
1101                 if (!sg)
1102                         goto fail;
1103
1104                 sg_span = sched_group_span(sg);
1105                 cpumask_or(covered, covered, sg_span);
1106
1107                 init_overlap_sched_group(sibling, sg);
1108
1109                 if (!first)
1110                         first = sg;
1111                 if (last)
1112                         last->next = sg;
1113                 last = sg;
1114                 last->next = first;
1115         }
1116         sd->groups = first;
1117
1118         return 0;
1119
1120 fail:
1121         free_sched_groups(first, 0);
1122
1123         return -ENOMEM;
1124 }
1125
1126
1127 /*
1128  * Package topology (also see the load-balance blurb in fair.c)
1129  *
1130  * The scheduler builds a tree structure to represent a number of important
1131  * topology features. By default (default_topology[]) these include:
1132  *
1133  *  - Simultaneous multithreading (SMT)
1134  *  - Multi-Core Cache (MC)
1135  *  - Package (PKG)
1136  *
1137  * Where the last one more or less denotes everything up to a NUMA node.
1138  *
1139  * The tree consists of 3 primary data structures:
1140  *
1141  *      sched_domain -> sched_group -> sched_group_capacity
1142  *          ^ ^             ^ ^
1143  *          `-'             `-'
1144  *
1145  * The sched_domains are per-CPU and have a two way link (parent & child) and
1146  * denote the ever growing mask of CPUs belonging to that level of topology.
1147  *
1148  * Each sched_domain has a circular (double) linked list of sched_group's, each
1149  * denoting the domains of the level below (or individual CPUs in case of the
1150  * first domain level). The sched_group linked by a sched_domain includes the
1151  * CPU of that sched_domain [*].
1152  *
1153  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1154  *
1155  * CPU   0   1   2   3   4   5   6   7
1156  *
1157  * PKG  [                             ]
1158  * MC   [             ] [             ]
1159  * SMT  [     ] [     ] [     ] [     ]
1160  *
1161  *  - or -
1162  *
1163  * PKG  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1164  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1165  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1166  *
1167  * CPU   0   1   2   3   4   5   6   7
1168  *
1169  * One way to think about it is: sched_domain moves you up and down among these
1170  * topology levels, while sched_group moves you sideways through it, at child
1171  * domain granularity.
1172  *
1173  * sched_group_capacity ensures each unique sched_group has shared storage.
1174  *
1175  * There are two related construction problems, both require a CPU that
1176  * uniquely identify each group (for a given domain):
1177  *
1178  *  - The first is the balance_cpu (see should_we_balance() and the
1179  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1180  *    continue balancing at a higher domain.
1181  *
1182  *  - The second is the sched_group_capacity; we want all identical groups
1183  *    to share a single sched_group_capacity.
1184  *
1185  * Since these topologies are exclusive by construction. That is, its
1186  * impossible for an SMT thread to belong to multiple cores, and cores to
1187  * be part of multiple caches. There is a very clear and unique location
1188  * for each CPU in the hierarchy.
1189  *
1190  * Therefore computing a unique CPU for each group is trivial (the iteration
1191  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1192  * group), we can simply pick the first CPU in each group.
1193  *
1194  *
1195  * [*] in other words, the first group of each domain is its child domain.
1196  */
1197
1198 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1199 {
1200         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1201         struct sched_domain *child = sd->child;
1202         struct sched_group *sg;
1203         bool already_visited;
1204
1205         if (child)
1206                 cpu = cpumask_first(sched_domain_span(child));
1207
1208         sg = *per_cpu_ptr(sdd->sg, cpu);
1209         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1210
1211         /* Increase refcounts for claim_allocations: */
1212         already_visited = atomic_inc_return(&sg->ref) > 1;
1213         /* sgc visits should follow a similar trend as sg */
1214         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1215
1216         /* If we have already visited that group, it's already initialized. */
1217         if (already_visited)
1218                 return sg;
1219
1220         if (child) {
1221                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1222                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1223                 sg->flags = child->flags;
1224         } else {
1225                 cpumask_set_cpu(cpu, sched_group_span(sg));
1226                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1227         }
1228
1229         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1230         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1231         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1232
1233         return sg;
1234 }
1235
1236 /*
1237  * build_sched_groups will build a circular linked list of the groups
1238  * covered by the given span, will set each group's ->cpumask correctly,
1239  * and will initialize their ->sgc.
1240  *
1241  * Assumes the sched_domain tree is fully constructed
1242  */
1243 static int
1244 build_sched_groups(struct sched_domain *sd, int cpu)
1245 {
1246         struct sched_group *first = NULL, *last = NULL;
1247         struct sd_data *sdd = sd->private;
1248         const struct cpumask *span = sched_domain_span(sd);
1249         struct cpumask *covered;
1250         int i;
1251
1252         lockdep_assert_held(&sched_domains_mutex);
1253         covered = sched_domains_tmpmask;
1254
1255         cpumask_clear(covered);
1256
1257         for_each_cpu_wrap(i, span, cpu) {
1258                 struct sched_group *sg;
1259
1260                 if (cpumask_test_cpu(i, covered))
1261                         continue;
1262
1263                 sg = get_group(i, sdd);
1264
1265                 cpumask_or(covered, covered, sched_group_span(sg));
1266
1267                 if (!first)
1268                         first = sg;
1269                 if (last)
1270                         last->next = sg;
1271                 last = sg;
1272         }
1273         last->next = first;
1274         sd->groups = first;
1275
1276         return 0;
1277 }
1278
1279 /*
1280  * Initialize sched groups cpu_capacity.
1281  *
1282  * cpu_capacity indicates the capacity of sched group, which is used while
1283  * distributing the load between different sched groups in a sched domain.
1284  * Typically cpu_capacity for all the groups in a sched domain will be same
1285  * unless there are asymmetries in the topology. If there are asymmetries,
1286  * group having more cpu_capacity will pickup more load compared to the
1287  * group having less cpu_capacity.
1288  */
1289 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1290 {
1291         struct sched_group *sg = sd->groups;
1292         struct cpumask *mask = sched_domains_tmpmask2;
1293
1294         WARN_ON(!sg);
1295
1296         do {
1297                 int cpu, cores = 0, max_cpu = -1;
1298
1299                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1300
1301                 cpumask_copy(mask, sched_group_span(sg));
1302                 for_each_cpu(cpu, mask) {
1303                         cores++;
1304 #ifdef CONFIG_SCHED_SMT
1305                         cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1306 #endif
1307                 }
1308                 sg->cores = cores;
1309
1310                 if (!(sd->flags & SD_ASYM_PACKING))
1311                         goto next;
1312
1313                 for_each_cpu(cpu, sched_group_span(sg)) {
1314                         if (max_cpu < 0)
1315                                 max_cpu = cpu;
1316                         else if (sched_asym_prefer(cpu, max_cpu))
1317                                 max_cpu = cpu;
1318                 }
1319                 sg->asym_prefer_cpu = max_cpu;
1320
1321 next:
1322                 sg = sg->next;
1323         } while (sg != sd->groups);
1324
1325         if (cpu != group_balance_cpu(sg))
1326                 return;
1327
1328         update_group_capacity(sd, cpu);
1329 }
1330
1331 /*
1332  * Set of available CPUs grouped by their corresponding capacities
1333  * Each list entry contains a CPU mask reflecting CPUs that share the same
1334  * capacity.
1335  * The lifespan of data is unlimited.
1336  */
1337 LIST_HEAD(asym_cap_list);
1338
1339 /*
1340  * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1341  * Provides sd_flags reflecting the asymmetry scope.
1342  */
1343 static inline int
1344 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1345                            const struct cpumask *cpu_map)
1346 {
1347         struct asym_cap_data *entry;
1348         int count = 0, miss = 0;
1349
1350         /*
1351          * Count how many unique CPU capacities this domain spans across
1352          * (compare sched_domain CPUs mask with ones representing  available
1353          * CPUs capacities). Take into account CPUs that might be offline:
1354          * skip those.
1355          */
1356         list_for_each_entry(entry, &asym_cap_list, link) {
1357                 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1358                         ++count;
1359                 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1360                         ++miss;
1361         }
1362
1363         WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1364
1365         /* No asymmetry detected */
1366         if (count < 2)
1367                 return 0;
1368         /* Some of the available CPU capacity values have not been detected */
1369         if (miss)
1370                 return SD_ASYM_CPUCAPACITY;
1371
1372         /* Full asymmetry */
1373         return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1374
1375 }
1376
1377 static void free_asym_cap_entry(struct rcu_head *head)
1378 {
1379         struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
1380         kfree(entry);
1381 }
1382
1383 static inline void asym_cpu_capacity_update_data(int cpu)
1384 {
1385         unsigned long capacity = arch_scale_cpu_capacity(cpu);
1386         struct asym_cap_data *insert_entry = NULL;
1387         struct asym_cap_data *entry;
1388
1389         /*
1390          * Search if capacity already exits. If not, track which the entry
1391          * where we should insert to keep the list ordered descendingly.
1392          */
1393         list_for_each_entry(entry, &asym_cap_list, link) {
1394                 if (capacity == entry->capacity)
1395                         goto done;
1396                 else if (!insert_entry && capacity > entry->capacity)
1397                         insert_entry = list_prev_entry(entry, link);
1398         }
1399
1400         entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1401         if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1402                 return;
1403         entry->capacity = capacity;
1404
1405         /* If NULL then the new capacity is the smallest, add last. */
1406         if (!insert_entry)
1407                 list_add_tail_rcu(&entry->link, &asym_cap_list);
1408         else
1409                 list_add_rcu(&entry->link, &insert_entry->link);
1410 done:
1411         __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1412 }
1413
1414 /*
1415  * Build-up/update list of CPUs grouped by their capacities
1416  * An update requires explicit request to rebuild sched domains
1417  * with state indicating CPU topology changes.
1418  */
1419 static void asym_cpu_capacity_scan(void)
1420 {
1421         struct asym_cap_data *entry, *next;
1422         int cpu;
1423
1424         list_for_each_entry(entry, &asym_cap_list, link)
1425                 cpumask_clear(cpu_capacity_span(entry));
1426
1427         for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1428                 asym_cpu_capacity_update_data(cpu);
1429
1430         list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1431                 if (cpumask_empty(cpu_capacity_span(entry))) {
1432                         list_del_rcu(&entry->link);
1433                         call_rcu(&entry->rcu, free_asym_cap_entry);
1434                 }
1435         }
1436
1437         /*
1438          * Only one capacity value has been detected i.e. this system is symmetric.
1439          * No need to keep this data around.
1440          */
1441         if (list_is_singular(&asym_cap_list)) {
1442                 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1443                 list_del_rcu(&entry->link);
1444                 call_rcu(&entry->rcu, free_asym_cap_entry);
1445         }
1446 }
1447
1448 /*
1449  * Initializers for schedule domains
1450  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1451  */
1452
1453 static int default_relax_domain_level = -1;
1454 int sched_domain_level_max;
1455
1456 static int __init setup_relax_domain_level(char *str)
1457 {
1458         if (kstrtoint(str, 0, &default_relax_domain_level))
1459                 pr_warn("Unable to set relax_domain_level\n");
1460
1461         return 1;
1462 }
1463 __setup("relax_domain_level=", setup_relax_domain_level);
1464
1465 static void set_domain_attribute(struct sched_domain *sd,
1466                                  struct sched_domain_attr *attr)
1467 {
1468         int request;
1469
1470         if (!attr || attr->relax_domain_level < 0) {
1471                 if (default_relax_domain_level < 0)
1472                         return;
1473                 request = default_relax_domain_level;
1474         } else
1475                 request = attr->relax_domain_level;
1476
1477         if (sd->level >= request) {
1478                 /* Turn off idle balance on this domain: */
1479                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1480         }
1481 }
1482
1483 static void __sdt_free(const struct cpumask *cpu_map);
1484 static int __sdt_alloc(const struct cpumask *cpu_map);
1485
1486 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1487                                  const struct cpumask *cpu_map)
1488 {
1489         switch (what) {
1490         case sa_rootdomain:
1491                 if (!atomic_read(&d->rd->refcount))
1492                         free_rootdomain(&d->rd->rcu);
1493                 fallthrough;
1494         case sa_sd:
1495                 free_percpu(d->sd);
1496                 fallthrough;
1497         case sa_sd_storage:
1498                 __sdt_free(cpu_map);
1499                 fallthrough;
1500         case sa_none:
1501                 break;
1502         }
1503 }
1504
1505 static enum s_alloc
1506 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1507 {
1508         memset(d, 0, sizeof(*d));
1509
1510         if (__sdt_alloc(cpu_map))
1511                 return sa_sd_storage;
1512         d->sd = alloc_percpu(struct sched_domain *);
1513         if (!d->sd)
1514                 return sa_sd_storage;
1515         d->rd = alloc_rootdomain();
1516         if (!d->rd)
1517                 return sa_sd;
1518
1519         return sa_rootdomain;
1520 }
1521
1522 /*
1523  * NULL the sd_data elements we've used to build the sched_domain and
1524  * sched_group structure so that the subsequent __free_domain_allocs()
1525  * will not free the data we're using.
1526  */
1527 static void claim_allocations(int cpu, struct sched_domain *sd)
1528 {
1529         struct sd_data *sdd = sd->private;
1530
1531         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1532         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1533
1534         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1535                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1536
1537         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1538                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1539
1540         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1541                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1542 }
1543
1544 #ifdef CONFIG_NUMA
1545 enum numa_topology_type sched_numa_topology_type;
1546
1547 static int                      sched_domains_numa_levels;
1548 static int                      sched_domains_curr_level;
1549
1550 int                             sched_max_numa_distance;
1551 static int                      *sched_domains_numa_distance;
1552 static struct cpumask           ***sched_domains_numa_masks;
1553 #endif
1554
1555 /*
1556  * SD_flags allowed in topology descriptions.
1557  *
1558  * These flags are purely descriptive of the topology and do not prescribe
1559  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1560  * function. For details, see include/linux/sched/sd_flags.h.
1561  *
1562  *   SD_SHARE_CPUCAPACITY
1563  *   SD_SHARE_LLC
1564  *   SD_CLUSTER
1565  *   SD_NUMA
1566  *
1567  * Odd one out, which beside describing the topology has a quirk also
1568  * prescribes the desired behaviour that goes along with it:
1569  *
1570  *   SD_ASYM_PACKING        - describes SMT quirks
1571  */
1572 #define TOPOLOGY_SD_FLAGS               \
1573         (SD_SHARE_CPUCAPACITY   |       \
1574          SD_CLUSTER             |       \
1575          SD_SHARE_LLC           |       \
1576          SD_NUMA                |       \
1577          SD_ASYM_PACKING)
1578
1579 static struct sched_domain *
1580 sd_init(struct sched_domain_topology_level *tl,
1581         const struct cpumask *cpu_map,
1582         struct sched_domain *child, int cpu)
1583 {
1584         struct sd_data *sdd = &tl->data;
1585         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1586         int sd_id, sd_weight, sd_flags = 0;
1587         struct cpumask *sd_span;
1588
1589 #ifdef CONFIG_NUMA
1590         /*
1591          * Ugly hack to pass state to sd_numa_mask()...
1592          */
1593         sched_domains_curr_level = tl->numa_level;
1594 #endif
1595
1596         sd_weight = cpumask_weight(tl->mask(cpu));
1597
1598         if (tl->sd_flags)
1599                 sd_flags = (*tl->sd_flags)();
1600         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1601                         "wrong sd_flags in topology description\n"))
1602                 sd_flags &= TOPOLOGY_SD_FLAGS;
1603
1604         *sd = (struct sched_domain){
1605                 .min_interval           = sd_weight,
1606                 .max_interval           = 2*sd_weight,
1607                 .busy_factor            = 16,
1608                 .imbalance_pct          = 117,
1609
1610                 .cache_nice_tries       = 0,
1611
1612                 .flags                  = 1*SD_BALANCE_NEWIDLE
1613                                         | 1*SD_BALANCE_EXEC
1614                                         | 1*SD_BALANCE_FORK
1615                                         | 0*SD_BALANCE_WAKE
1616                                         | 1*SD_WAKE_AFFINE
1617                                         | 0*SD_SHARE_CPUCAPACITY
1618                                         | 0*SD_SHARE_LLC
1619                                         | 0*SD_SERIALIZE
1620                                         | 1*SD_PREFER_SIBLING
1621                                         | 0*SD_NUMA
1622                                         | sd_flags
1623                                         ,
1624
1625                 .last_balance           = jiffies,
1626                 .balance_interval       = sd_weight,
1627                 .max_newidle_lb_cost    = 0,
1628                 .last_decay_max_lb_cost = jiffies,
1629                 .child                  = child,
1630 #ifdef CONFIG_SCHED_DEBUG
1631                 .name                   = tl->name,
1632 #endif
1633         };
1634
1635         sd_span = sched_domain_span(sd);
1636         cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1637         sd_id = cpumask_first(sd_span);
1638
1639         sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1640
1641         WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1642                   (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1643                   "CPU capacity asymmetry not supported on SMT\n");
1644
1645         /*
1646          * Convert topological properties into behaviour.
1647          */
1648         /* Don't attempt to spread across CPUs of different capacities. */
1649         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1650                 sd->child->flags &= ~SD_PREFER_SIBLING;
1651
1652         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1653                 sd->imbalance_pct = 110;
1654
1655         } else if (sd->flags & SD_SHARE_LLC) {
1656                 sd->imbalance_pct = 117;
1657                 sd->cache_nice_tries = 1;
1658
1659 #ifdef CONFIG_NUMA
1660         } else if (sd->flags & SD_NUMA) {
1661                 sd->cache_nice_tries = 2;
1662
1663                 sd->flags &= ~SD_PREFER_SIBLING;
1664                 sd->flags |= SD_SERIALIZE;
1665                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1666                         sd->flags &= ~(SD_BALANCE_EXEC |
1667                                        SD_BALANCE_FORK |
1668                                        SD_WAKE_AFFINE);
1669                 }
1670
1671 #endif
1672         } else {
1673                 sd->cache_nice_tries = 1;
1674         }
1675
1676         /*
1677          * For all levels sharing cache; connect a sched_domain_shared
1678          * instance.
1679          */
1680         if (sd->flags & SD_SHARE_LLC) {
1681                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1682                 atomic_inc(&sd->shared->ref);
1683                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1684         }
1685
1686         sd->private = sdd;
1687
1688         return sd;
1689 }
1690
1691 /*
1692  * Topology list, bottom-up.
1693  */
1694 static struct sched_domain_topology_level default_topology[] = {
1695 #ifdef CONFIG_SCHED_SMT
1696         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1697 #endif
1698
1699 #ifdef CONFIG_SCHED_CLUSTER
1700         { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1701 #endif
1702
1703 #ifdef CONFIG_SCHED_MC
1704         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1705 #endif
1706         { cpu_cpu_mask, SD_INIT_NAME(PKG) },
1707         { NULL, },
1708 };
1709
1710 static struct sched_domain_topology_level *sched_domain_topology =
1711         default_topology;
1712 static struct sched_domain_topology_level *sched_domain_topology_saved;
1713
1714 #define for_each_sd_topology(tl)                        \
1715         for (tl = sched_domain_topology; tl->mask; tl++)
1716
1717 void __init set_sched_topology(struct sched_domain_topology_level *tl)
1718 {
1719         if (WARN_ON_ONCE(sched_smp_initialized))
1720                 return;
1721
1722         sched_domain_topology = tl;
1723         sched_domain_topology_saved = NULL;
1724 }
1725
1726 #ifdef CONFIG_NUMA
1727
1728 static const struct cpumask *sd_numa_mask(int cpu)
1729 {
1730         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1731 }
1732
1733 static void sched_numa_warn(const char *str)
1734 {
1735         static int done = false;
1736         int i,j;
1737
1738         if (done)
1739                 return;
1740
1741         done = true;
1742
1743         printk(KERN_WARNING "ERROR: %s\n\n", str);
1744
1745         for (i = 0; i < nr_node_ids; i++) {
1746                 printk(KERN_WARNING "  ");
1747                 for (j = 0; j < nr_node_ids; j++) {
1748                         if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1749                                 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1750                         else
1751                                 printk(KERN_CONT " %02d  ", node_distance(i,j));
1752                 }
1753                 printk(KERN_CONT "\n");
1754         }
1755         printk(KERN_WARNING "\n");
1756 }
1757
1758 bool find_numa_distance(int distance)
1759 {
1760         bool found = false;
1761         int i, *distances;
1762
1763         if (distance == node_distance(0, 0))
1764                 return true;
1765
1766         rcu_read_lock();
1767         distances = rcu_dereference(sched_domains_numa_distance);
1768         if (!distances)
1769                 goto unlock;
1770         for (i = 0; i < sched_domains_numa_levels; i++) {
1771                 if (distances[i] == distance) {
1772                         found = true;
1773                         break;
1774                 }
1775         }
1776 unlock:
1777         rcu_read_unlock();
1778
1779         return found;
1780 }
1781
1782 #define for_each_cpu_node_but(n, nbut)          \
1783         for_each_node_state(n, N_CPU)           \
1784                 if (n == nbut)                  \
1785                         continue;               \
1786                 else
1787
1788 /*
1789  * A system can have three types of NUMA topology:
1790  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1791  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1792  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1793  *
1794  * The difference between a glueless mesh topology and a backplane
1795  * topology lies in whether communication between not directly
1796  * connected nodes goes through intermediary nodes (where programs
1797  * could run), or through backplane controllers. This affects
1798  * placement of programs.
1799  *
1800  * The type of topology can be discerned with the following tests:
1801  * - If the maximum distance between any nodes is 1 hop, the system
1802  *   is directly connected.
1803  * - If for two nodes A and B, located N > 1 hops away from each other,
1804  *   there is an intermediary node C, which is < N hops away from both
1805  *   nodes A and B, the system is a glueless mesh.
1806  */
1807 static void init_numa_topology_type(int offline_node)
1808 {
1809         int a, b, c, n;
1810
1811         n = sched_max_numa_distance;
1812
1813         if (sched_domains_numa_levels <= 2) {
1814                 sched_numa_topology_type = NUMA_DIRECT;
1815                 return;
1816         }
1817
1818         for_each_cpu_node_but(a, offline_node) {
1819                 for_each_cpu_node_but(b, offline_node) {
1820                         /* Find two nodes furthest removed from each other. */
1821                         if (node_distance(a, b) < n)
1822                                 continue;
1823
1824                         /* Is there an intermediary node between a and b? */
1825                         for_each_cpu_node_but(c, offline_node) {
1826                                 if (node_distance(a, c) < n &&
1827                                     node_distance(b, c) < n) {
1828                                         sched_numa_topology_type =
1829                                                         NUMA_GLUELESS_MESH;
1830                                         return;
1831                                 }
1832                         }
1833
1834                         sched_numa_topology_type = NUMA_BACKPLANE;
1835                         return;
1836                 }
1837         }
1838
1839         pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1840         sched_numa_topology_type = NUMA_DIRECT;
1841 }
1842
1843
1844 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1845
1846 void sched_init_numa(int offline_node)
1847 {
1848         struct sched_domain_topology_level *tl;
1849         unsigned long *distance_map;
1850         int nr_levels = 0;
1851         int i, j;
1852         int *distances;
1853         struct cpumask ***masks;
1854
1855         /*
1856          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1857          * unique distances in the node_distance() table.
1858          */
1859         distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1860         if (!distance_map)
1861                 return;
1862
1863         bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1864         for_each_cpu_node_but(i, offline_node) {
1865                 for_each_cpu_node_but(j, offline_node) {
1866                         int distance = node_distance(i, j);
1867
1868                         if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1869                                 sched_numa_warn("Invalid distance value range");
1870                                 bitmap_free(distance_map);
1871                                 return;
1872                         }
1873
1874                         bitmap_set(distance_map, distance, 1);
1875                 }
1876         }
1877         /*
1878          * We can now figure out how many unique distance values there are and
1879          * allocate memory accordingly.
1880          */
1881         nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1882
1883         distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1884         if (!distances) {
1885                 bitmap_free(distance_map);
1886                 return;
1887         }
1888
1889         for (i = 0, j = 0; i < nr_levels; i++, j++) {
1890                 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1891                 distances[i] = j;
1892         }
1893         rcu_assign_pointer(sched_domains_numa_distance, distances);
1894
1895         bitmap_free(distance_map);
1896
1897         /*
1898          * 'nr_levels' contains the number of unique distances
1899          *
1900          * The sched_domains_numa_distance[] array includes the actual distance
1901          * numbers.
1902          */
1903
1904         /*
1905          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1906          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1907          * the array will contain less then 'nr_levels' members. This could be
1908          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1909          * in other functions.
1910          *
1911          * We reset it to 'nr_levels' at the end of this function.
1912          */
1913         sched_domains_numa_levels = 0;
1914
1915         masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1916         if (!masks)
1917                 return;
1918
1919         /*
1920          * Now for each level, construct a mask per node which contains all
1921          * CPUs of nodes that are that many hops away from us.
1922          */
1923         for (i = 0; i < nr_levels; i++) {
1924                 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1925                 if (!masks[i])
1926                         return;
1927
1928                 for_each_cpu_node_but(j, offline_node) {
1929                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1930                         int k;
1931
1932                         if (!mask)
1933                                 return;
1934
1935                         masks[i][j] = mask;
1936
1937                         for_each_cpu_node_but(k, offline_node) {
1938                                 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1939                                         sched_numa_warn("Node-distance not symmetric");
1940
1941                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1942                                         continue;
1943
1944                                 cpumask_or(mask, mask, cpumask_of_node(k));
1945                         }
1946                 }
1947         }
1948         rcu_assign_pointer(sched_domains_numa_masks, masks);
1949
1950         /* Compute default topology size */
1951         for (i = 0; sched_domain_topology[i].mask; i++);
1952
1953         tl = kzalloc((i + nr_levels + 1) *
1954                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1955         if (!tl)
1956                 return;
1957
1958         /*
1959          * Copy the default topology bits..
1960          */
1961         for (i = 0; sched_domain_topology[i].mask; i++)
1962                 tl[i] = sched_domain_topology[i];
1963
1964         /*
1965          * Add the NUMA identity distance, aka single NODE.
1966          */
1967         tl[i++] = (struct sched_domain_topology_level){
1968                 .mask = sd_numa_mask,
1969                 .numa_level = 0,
1970                 SD_INIT_NAME(NODE)
1971         };
1972
1973         /*
1974          * .. and append 'j' levels of NUMA goodness.
1975          */
1976         for (j = 1; j < nr_levels; i++, j++) {
1977                 tl[i] = (struct sched_domain_topology_level){
1978                         .mask = sd_numa_mask,
1979                         .sd_flags = cpu_numa_flags,
1980                         .flags = SDTL_OVERLAP,
1981                         .numa_level = j,
1982                         SD_INIT_NAME(NUMA)
1983                 };
1984         }
1985
1986         sched_domain_topology_saved = sched_domain_topology;
1987         sched_domain_topology = tl;
1988
1989         sched_domains_numa_levels = nr_levels;
1990         WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1991
1992         init_numa_topology_type(offline_node);
1993 }
1994
1995
1996 static void sched_reset_numa(void)
1997 {
1998         int nr_levels, *distances;
1999         struct cpumask ***masks;
2000
2001         nr_levels = sched_domains_numa_levels;
2002         sched_domains_numa_levels = 0;
2003         sched_max_numa_distance = 0;
2004         sched_numa_topology_type = NUMA_DIRECT;
2005         distances = sched_domains_numa_distance;
2006         rcu_assign_pointer(sched_domains_numa_distance, NULL);
2007         masks = sched_domains_numa_masks;
2008         rcu_assign_pointer(sched_domains_numa_masks, NULL);
2009         if (distances || masks) {
2010                 int i, j;
2011
2012                 synchronize_rcu();
2013                 kfree(distances);
2014                 for (i = 0; i < nr_levels && masks; i++) {
2015                         if (!masks[i])
2016                                 continue;
2017                         for_each_node(j)
2018                                 kfree(masks[i][j]);
2019                         kfree(masks[i]);
2020                 }
2021                 kfree(masks);
2022         }
2023         if (sched_domain_topology_saved) {
2024                 kfree(sched_domain_topology);
2025                 sched_domain_topology = sched_domain_topology_saved;
2026                 sched_domain_topology_saved = NULL;
2027         }
2028 }
2029
2030 /*
2031  * Call with hotplug lock held
2032  */
2033 void sched_update_numa(int cpu, bool online)
2034 {
2035         int node;
2036
2037         node = cpu_to_node(cpu);
2038         /*
2039          * Scheduler NUMA topology is updated when the first CPU of a
2040          * node is onlined or the last CPU of a node is offlined.
2041          */
2042         if (cpumask_weight(cpumask_of_node(node)) != 1)
2043                 return;
2044
2045         sched_reset_numa();
2046         sched_init_numa(online ? NUMA_NO_NODE : node);
2047 }
2048
2049 void sched_domains_numa_masks_set(unsigned int cpu)
2050 {
2051         int node = cpu_to_node(cpu);
2052         int i, j;
2053
2054         for (i = 0; i < sched_domains_numa_levels; i++) {
2055                 for (j = 0; j < nr_node_ids; j++) {
2056                         if (!node_state(j, N_CPU))
2057                                 continue;
2058
2059                         /* Set ourselves in the remote node's masks */
2060                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
2061                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2062                 }
2063         }
2064 }
2065
2066 void sched_domains_numa_masks_clear(unsigned int cpu)
2067 {
2068         int i, j;
2069
2070         for (i = 0; i < sched_domains_numa_levels; i++) {
2071                 for (j = 0; j < nr_node_ids; j++) {
2072                         if (sched_domains_numa_masks[i][j])
2073                                 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2074                 }
2075         }
2076 }
2077
2078 /*
2079  * sched_numa_find_closest() - given the NUMA topology, find the cpu
2080  *                             closest to @cpu from @cpumask.
2081  * cpumask: cpumask to find a cpu from
2082  * cpu: cpu to be close to
2083  *
2084  * returns: cpu, or nr_cpu_ids when nothing found.
2085  */
2086 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2087 {
2088         int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2089         struct cpumask ***masks;
2090
2091         rcu_read_lock();
2092         masks = rcu_dereference(sched_domains_numa_masks);
2093         if (!masks)
2094                 goto unlock;
2095         for (i = 0; i < sched_domains_numa_levels; i++) {
2096                 if (!masks[i][j])
2097                         break;
2098                 cpu = cpumask_any_and(cpus, masks[i][j]);
2099                 if (cpu < nr_cpu_ids) {
2100                         found = cpu;
2101                         break;
2102                 }
2103         }
2104 unlock:
2105         rcu_read_unlock();
2106
2107         return found;
2108 }
2109
2110 struct __cmp_key {
2111         const struct cpumask *cpus;
2112         struct cpumask ***masks;
2113         int node;
2114         int cpu;
2115         int w;
2116 };
2117
2118 static int hop_cmp(const void *a, const void *b)
2119 {
2120         struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2121         struct __cmp_key *k = (struct __cmp_key *)a;
2122
2123         if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2124                 return 1;
2125
2126         if (b == k->masks) {
2127                 k->w = 0;
2128                 return 0;
2129         }
2130
2131         prev_hop = *((struct cpumask ***)b - 1);
2132         k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2133         if (k->w <= k->cpu)
2134                 return 0;
2135
2136         return -1;
2137 }
2138
2139 /**
2140  * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2141  *                             from @cpus to @cpu, taking into account distance
2142  *                             from a given @node.
2143  * @cpus: cpumask to find a cpu from
2144  * @cpu: CPU to start searching
2145  * @node: NUMA node to order CPUs by distance
2146  *
2147  * Return: cpu, or nr_cpu_ids when nothing found.
2148  */
2149 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2150 {
2151         struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2152         struct cpumask ***hop_masks;
2153         int hop, ret = nr_cpu_ids;
2154
2155         if (node == NUMA_NO_NODE)
2156                 return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2157
2158         rcu_read_lock();
2159
2160         /* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2161         node = numa_nearest_node(node, N_CPU);
2162         k.node = node;
2163
2164         k.masks = rcu_dereference(sched_domains_numa_masks);
2165         if (!k.masks)
2166                 goto unlock;
2167
2168         hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2169         hop = hop_masks - k.masks;
2170
2171         ret = hop ?
2172                 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2173                 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2174 unlock:
2175         rcu_read_unlock();
2176         return ret;
2177 }
2178 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2179
2180 /**
2181  * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2182  *                         @node
2183  * @node: The node to count hops from.
2184  * @hops: Include CPUs up to that many hops away. 0 means local node.
2185  *
2186  * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2187  * @node, an error value otherwise.
2188  *
2189  * Requires rcu_lock to be held. Returned cpumask is only valid within that
2190  * read-side section, copy it if required beyond that.
2191  *
2192  * Note that not all hops are equal in distance; see sched_init_numa() for how
2193  * distances and masks are handled.
2194  * Also note that this is a reflection of sched_domains_numa_masks, which may change
2195  * during the lifetime of the system (offline nodes are taken out of the masks).
2196  */
2197 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2198 {
2199         struct cpumask ***masks;
2200
2201         if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2202                 return ERR_PTR(-EINVAL);
2203
2204         masks = rcu_dereference(sched_domains_numa_masks);
2205         if (!masks)
2206                 return ERR_PTR(-EBUSY);
2207
2208         return masks[hops][node];
2209 }
2210 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2211
2212 #endif /* CONFIG_NUMA */
2213
2214 static int __sdt_alloc(const struct cpumask *cpu_map)
2215 {
2216         struct sched_domain_topology_level *tl;
2217         int j;
2218
2219         for_each_sd_topology(tl) {
2220                 struct sd_data *sdd = &tl->data;
2221
2222                 sdd->sd = alloc_percpu(struct sched_domain *);
2223                 if (!sdd->sd)
2224                         return -ENOMEM;
2225
2226                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2227                 if (!sdd->sds)
2228                         return -ENOMEM;
2229
2230                 sdd->sg = alloc_percpu(struct sched_group *);
2231                 if (!sdd->sg)
2232                         return -ENOMEM;
2233
2234                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2235                 if (!sdd->sgc)
2236                         return -ENOMEM;
2237
2238                 for_each_cpu(j, cpu_map) {
2239                         struct sched_domain *sd;
2240                         struct sched_domain_shared *sds;
2241                         struct sched_group *sg;
2242                         struct sched_group_capacity *sgc;
2243
2244                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2245                                         GFP_KERNEL, cpu_to_node(j));
2246                         if (!sd)
2247                                 return -ENOMEM;
2248
2249                         *per_cpu_ptr(sdd->sd, j) = sd;
2250
2251                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
2252                                         GFP_KERNEL, cpu_to_node(j));
2253                         if (!sds)
2254                                 return -ENOMEM;
2255
2256                         *per_cpu_ptr(sdd->sds, j) = sds;
2257
2258                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2259                                         GFP_KERNEL, cpu_to_node(j));
2260                         if (!sg)
2261                                 return -ENOMEM;
2262
2263                         sg->next = sg;
2264
2265                         *per_cpu_ptr(sdd->sg, j) = sg;
2266
2267                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2268                                         GFP_KERNEL, cpu_to_node(j));
2269                         if (!sgc)
2270                                 return -ENOMEM;
2271
2272 #ifdef CONFIG_SCHED_DEBUG
2273                         sgc->id = j;
2274 #endif
2275
2276                         *per_cpu_ptr(sdd->sgc, j) = sgc;
2277                 }
2278         }
2279
2280         return 0;
2281 }
2282
2283 static void __sdt_free(const struct cpumask *cpu_map)
2284 {
2285         struct sched_domain_topology_level *tl;
2286         int j;
2287
2288         for_each_sd_topology(tl) {
2289                 struct sd_data *sdd = &tl->data;
2290
2291                 for_each_cpu(j, cpu_map) {
2292                         struct sched_domain *sd;
2293
2294                         if (sdd->sd) {
2295                                 sd = *per_cpu_ptr(sdd->sd, j);
2296                                 if (sd && (sd->flags & SD_OVERLAP))
2297                                         free_sched_groups(sd->groups, 0);
2298                                 kfree(*per_cpu_ptr(sdd->sd, j));
2299                         }
2300
2301                         if (sdd->sds)
2302                                 kfree(*per_cpu_ptr(sdd->sds, j));
2303                         if (sdd->sg)
2304                                 kfree(*per_cpu_ptr(sdd->sg, j));
2305                         if (sdd->sgc)
2306                                 kfree(*per_cpu_ptr(sdd->sgc, j));
2307                 }
2308                 free_percpu(sdd->sd);
2309                 sdd->sd = NULL;
2310                 free_percpu(sdd->sds);
2311                 sdd->sds = NULL;
2312                 free_percpu(sdd->sg);
2313                 sdd->sg = NULL;
2314                 free_percpu(sdd->sgc);
2315                 sdd->sgc = NULL;
2316         }
2317 }
2318
2319 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2320                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2321                 struct sched_domain *child, int cpu)
2322 {
2323         struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2324
2325         if (child) {
2326                 sd->level = child->level + 1;
2327                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2328                 child->parent = sd;
2329
2330                 if (!cpumask_subset(sched_domain_span(child),
2331                                     sched_domain_span(sd))) {
2332                         pr_err("BUG: arch topology borken\n");
2333 #ifdef CONFIG_SCHED_DEBUG
2334                         pr_err("     the %s domain not a subset of the %s domain\n",
2335                                         child->name, sd->name);
2336 #endif
2337                         /* Fixup, ensure @sd has at least @child CPUs. */
2338                         cpumask_or(sched_domain_span(sd),
2339                                    sched_domain_span(sd),
2340                                    sched_domain_span(child));
2341                 }
2342
2343         }
2344         set_domain_attribute(sd, attr);
2345
2346         return sd;
2347 }
2348
2349 /*
2350  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2351  * any two given CPUs at this (non-NUMA) topology level.
2352  */
2353 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2354                               const struct cpumask *cpu_map, int cpu)
2355 {
2356         int i = cpu + 1;
2357
2358         /* NUMA levels are allowed to overlap */
2359         if (tl->flags & SDTL_OVERLAP)
2360                 return true;
2361
2362         /*
2363          * Non-NUMA levels cannot partially overlap - they must be either
2364          * completely equal or completely disjoint. Otherwise we can end up
2365          * breaking the sched_group lists - i.e. a later get_group() pass
2366          * breaks the linking done for an earlier span.
2367          */
2368         for_each_cpu_from(i, cpu_map) {
2369                 /*
2370                  * We should 'and' all those masks with 'cpu_map' to exactly
2371                  * match the topology we're about to build, but that can only
2372                  * remove CPUs, which only lessens our ability to detect
2373                  * overlaps
2374                  */
2375                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2376                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2377                         return false;
2378         }
2379
2380         return true;
2381 }
2382
2383 /*
2384  * Build sched domains for a given set of CPUs and attach the sched domains
2385  * to the individual CPUs
2386  */
2387 static int
2388 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2389 {
2390         enum s_alloc alloc_state = sa_none;
2391         struct sched_domain *sd;
2392         struct s_data d;
2393         struct rq *rq = NULL;
2394         int i, ret = -ENOMEM;
2395         bool has_asym = false;
2396         bool has_cluster = false;
2397
2398         if (WARN_ON(cpumask_empty(cpu_map)))
2399                 goto error;
2400
2401         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2402         if (alloc_state != sa_rootdomain)
2403                 goto error;
2404
2405         /* Set up domains for CPUs specified by the cpu_map: */
2406         for_each_cpu(i, cpu_map) {
2407                 struct sched_domain_topology_level *tl;
2408
2409                 sd = NULL;
2410                 for_each_sd_topology(tl) {
2411
2412                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2413                                 goto error;
2414
2415                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2416
2417                         has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2418
2419                         if (tl == sched_domain_topology)
2420                                 *per_cpu_ptr(d.sd, i) = sd;
2421                         if (tl->flags & SDTL_OVERLAP)
2422                                 sd->flags |= SD_OVERLAP;
2423                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2424                                 break;
2425                 }
2426         }
2427
2428         /* Build the groups for the domains */
2429         for_each_cpu(i, cpu_map) {
2430                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2431                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2432                         if (sd->flags & SD_OVERLAP) {
2433                                 if (build_overlap_sched_groups(sd, i))
2434                                         goto error;
2435                         } else {
2436                                 if (build_sched_groups(sd, i))
2437                                         goto error;
2438                         }
2439                 }
2440         }
2441
2442         /*
2443          * Calculate an allowed NUMA imbalance such that LLCs do not get
2444          * imbalanced.
2445          */
2446         for_each_cpu(i, cpu_map) {
2447                 unsigned int imb = 0;
2448                 unsigned int imb_span = 1;
2449
2450                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2451                         struct sched_domain *child = sd->child;
2452
2453                         if (!(sd->flags & SD_SHARE_LLC) && child &&
2454                             (child->flags & SD_SHARE_LLC)) {
2455                                 struct sched_domain __rcu *top_p;
2456                                 unsigned int nr_llcs;
2457
2458                                 /*
2459                                  * For a single LLC per node, allow an
2460                                  * imbalance up to 12.5% of the node. This is
2461                                  * arbitrary cutoff based two factors -- SMT and
2462                                  * memory channels. For SMT-2, the intent is to
2463                                  * avoid premature sharing of HT resources but
2464                                  * SMT-4 or SMT-8 *may* benefit from a different
2465                                  * cutoff. For memory channels, this is a very
2466                                  * rough estimate of how many channels may be
2467                                  * active and is based on recent CPUs with
2468                                  * many cores.
2469                                  *
2470                                  * For multiple LLCs, allow an imbalance
2471                                  * until multiple tasks would share an LLC
2472                                  * on one node while LLCs on another node
2473                                  * remain idle. This assumes that there are
2474                                  * enough logical CPUs per LLC to avoid SMT
2475                                  * factors and that there is a correlation
2476                                  * between LLCs and memory channels.
2477                                  */
2478                                 nr_llcs = sd->span_weight / child->span_weight;
2479                                 if (nr_llcs == 1)
2480                                         imb = sd->span_weight >> 3;
2481                                 else
2482                                         imb = nr_llcs;
2483                                 imb = max(1U, imb);
2484                                 sd->imb_numa_nr = imb;
2485
2486                                 /* Set span based on the first NUMA domain. */
2487                                 top_p = sd->parent;
2488                                 while (top_p && !(top_p->flags & SD_NUMA)) {
2489                                         top_p = top_p->parent;
2490                                 }
2491                                 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2492                         } else {
2493                                 int factor = max(1U, (sd->span_weight / imb_span));
2494
2495                                 sd->imb_numa_nr = imb * factor;
2496                         }
2497                 }
2498         }
2499
2500         /* Calculate CPU capacity for physical packages and nodes */
2501         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2502                 if (!cpumask_test_cpu(i, cpu_map))
2503                         continue;
2504
2505                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2506                         claim_allocations(i, sd);
2507                         init_sched_groups_capacity(i, sd);
2508                 }
2509         }
2510
2511         /* Attach the domains */
2512         rcu_read_lock();
2513         for_each_cpu(i, cpu_map) {
2514                 rq = cpu_rq(i);
2515                 sd = *per_cpu_ptr(d.sd, i);
2516
2517                 cpu_attach_domain(sd, d.rd, i);
2518
2519                 if (lowest_flag_domain(i, SD_CLUSTER))
2520                         has_cluster = true;
2521         }
2522         rcu_read_unlock();
2523
2524         if (has_asym)
2525                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2526
2527         if (has_cluster)
2528                 static_branch_inc_cpuslocked(&sched_cluster_active);
2529
2530         if (rq && sched_debug_verbose)
2531                 pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
2532
2533         ret = 0;
2534 error:
2535         __free_domain_allocs(&d, alloc_state, cpu_map);
2536
2537         return ret;
2538 }
2539
2540 /* Current sched domains: */
2541 static cpumask_var_t                    *doms_cur;
2542
2543 /* Number of sched domains in 'doms_cur': */
2544 static int                              ndoms_cur;
2545
2546 /* Attributes of custom domains in 'doms_cur' */
2547 static struct sched_domain_attr         *dattr_cur;
2548
2549 /*
2550  * Special case: If a kmalloc() of a doms_cur partition (array of
2551  * cpumask) fails, then fallback to a single sched domain,
2552  * as determined by the single cpumask fallback_doms.
2553  */
2554 static cpumask_var_t                    fallback_doms;
2555
2556 /*
2557  * arch_update_cpu_topology lets virtualized architectures update the
2558  * CPU core maps. It is supposed to return 1 if the topology changed
2559  * or 0 if it stayed the same.
2560  */
2561 int __weak arch_update_cpu_topology(void)
2562 {
2563         return 0;
2564 }
2565
2566 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2567 {
2568         int i;
2569         cpumask_var_t *doms;
2570
2571         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2572         if (!doms)
2573                 return NULL;
2574         for (i = 0; i < ndoms; i++) {
2575                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2576                         free_sched_domains(doms, i);
2577                         return NULL;
2578                 }
2579         }
2580         return doms;
2581 }
2582
2583 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2584 {
2585         unsigned int i;
2586         for (i = 0; i < ndoms; i++)
2587                 free_cpumask_var(doms[i]);
2588         kfree(doms);
2589 }
2590
2591 /*
2592  * Set up scheduler domains and groups.  For now this just excludes isolated
2593  * CPUs, but could be used to exclude other special cases in the future.
2594  */
2595 int __init sched_init_domains(const struct cpumask *cpu_map)
2596 {
2597         int err;
2598
2599         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2600         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2601         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2602
2603         arch_update_cpu_topology();
2604         asym_cpu_capacity_scan();
2605         ndoms_cur = 1;
2606         doms_cur = alloc_sched_domains(ndoms_cur);
2607         if (!doms_cur)
2608                 doms_cur = &fallback_doms;
2609         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2610         err = build_sched_domains(doms_cur[0], NULL);
2611
2612         return err;
2613 }
2614
2615 /*
2616  * Detach sched domains from a group of CPUs specified in cpu_map
2617  * These CPUs will now be attached to the NULL domain
2618  */
2619 static void detach_destroy_domains(const struct cpumask *cpu_map)
2620 {
2621         unsigned int cpu = cpumask_any(cpu_map);
2622         int i;
2623
2624         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2625                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2626
2627         if (static_branch_unlikely(&sched_cluster_active))
2628                 static_branch_dec_cpuslocked(&sched_cluster_active);
2629
2630         rcu_read_lock();
2631         for_each_cpu(i, cpu_map)
2632                 cpu_attach_domain(NULL, &def_root_domain, i);
2633         rcu_read_unlock();
2634 }
2635
2636 /* handle null as "default" */
2637 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2638                         struct sched_domain_attr *new, int idx_new)
2639 {
2640         struct sched_domain_attr tmp;
2641
2642         /* Fast path: */
2643         if (!new && !cur)
2644                 return 1;
2645
2646         tmp = SD_ATTR_INIT;
2647
2648         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2649                         new ? (new + idx_new) : &tmp,
2650                         sizeof(struct sched_domain_attr));
2651 }
2652
2653 /*
2654  * Partition sched domains as specified by the 'ndoms_new'
2655  * cpumasks in the array doms_new[] of cpumasks. This compares
2656  * doms_new[] to the current sched domain partitioning, doms_cur[].
2657  * It destroys each deleted domain and builds each new domain.
2658  *
2659  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2660  * The masks don't intersect (don't overlap.) We should setup one
2661  * sched domain for each mask. CPUs not in any of the cpumasks will
2662  * not be load balanced. If the same cpumask appears both in the
2663  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2664  * it as it is.
2665  *
2666  * The passed in 'doms_new' should be allocated using
2667  * alloc_sched_domains.  This routine takes ownership of it and will
2668  * free_sched_domains it when done with it. If the caller failed the
2669  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2670  * and partition_sched_domains() will fallback to the single partition
2671  * 'fallback_doms', it also forces the domains to be rebuilt.
2672  *
2673  * If doms_new == NULL it will be replaced with cpu_online_mask.
2674  * ndoms_new == 0 is a special case for destroying existing domains,
2675  * and it will not create the default domain.
2676  *
2677  * Call with hotplug lock and sched_domains_mutex held
2678  */
2679 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2680                                     struct sched_domain_attr *dattr_new)
2681 {
2682         bool __maybe_unused has_eas = false;
2683         int i, j, n;
2684         int new_topology;
2685
2686         lockdep_assert_held(&sched_domains_mutex);
2687
2688         /* Let the architecture update CPU core mappings: */
2689         new_topology = arch_update_cpu_topology();
2690         /* Trigger rebuilding CPU capacity asymmetry data */
2691         if (new_topology)
2692                 asym_cpu_capacity_scan();
2693
2694         if (!doms_new) {
2695                 WARN_ON_ONCE(dattr_new);
2696                 n = 0;
2697                 doms_new = alloc_sched_domains(1);
2698                 if (doms_new) {
2699                         n = 1;
2700                         cpumask_and(doms_new[0], cpu_active_mask,
2701                                     housekeeping_cpumask(HK_TYPE_DOMAIN));
2702                 }
2703         } else {
2704                 n = ndoms_new;
2705         }
2706
2707         /* Destroy deleted domains: */
2708         for (i = 0; i < ndoms_cur; i++) {
2709                 for (j = 0; j < n && !new_topology; j++) {
2710                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2711                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2712                                 struct root_domain *rd;
2713
2714                                 /*
2715                                  * This domain won't be destroyed and as such
2716                                  * its dl_bw->total_bw needs to be cleared.  It
2717                                  * will be recomputed in function
2718                                  * update_tasks_root_domain().
2719                                  */
2720                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2721                                 dl_clear_root_domain(rd);
2722                                 goto match1;
2723                         }
2724                 }
2725                 /* No match - a current sched domain not in new doms_new[] */
2726                 detach_destroy_domains(doms_cur[i]);
2727 match1:
2728                 ;
2729         }
2730
2731         n = ndoms_cur;
2732         if (!doms_new) {
2733                 n = 0;
2734                 doms_new = &fallback_doms;
2735                 cpumask_and(doms_new[0], cpu_active_mask,
2736                             housekeeping_cpumask(HK_TYPE_DOMAIN));
2737         }
2738
2739         /* Build new domains: */
2740         for (i = 0; i < ndoms_new; i++) {
2741                 for (j = 0; j < n && !new_topology; j++) {
2742                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2743                             dattrs_equal(dattr_new, i, dattr_cur, j))
2744                                 goto match2;
2745                 }
2746                 /* No match - add a new doms_new */
2747                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2748 match2:
2749                 ;
2750         }
2751
2752 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2753         /* Build perf. domains: */
2754         for (i = 0; i < ndoms_new; i++) {
2755                 for (j = 0; j < n && !sched_energy_update; j++) {
2756                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2757                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2758                                 has_eas = true;
2759                                 goto match3;
2760                         }
2761                 }
2762                 /* No match - add perf. domains for a new rd */
2763                 has_eas |= build_perf_domains(doms_new[i]);
2764 match3:
2765                 ;
2766         }
2767         sched_energy_set(has_eas);
2768 #endif
2769
2770         /* Remember the new sched domains: */
2771         if (doms_cur != &fallback_doms)
2772                 free_sched_domains(doms_cur, ndoms_cur);
2773
2774         kfree(dattr_cur);
2775         doms_cur = doms_new;
2776         dattr_cur = dattr_new;
2777         ndoms_cur = ndoms_new;
2778
2779         update_sched_domain_debugfs();
2780 }
2781
2782 /*
2783  * Call with hotplug lock held
2784  */
2785 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2786                              struct sched_domain_attr *dattr_new)
2787 {
2788         mutex_lock(&sched_domains_mutex);
2789         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2790         mutex_unlock(&sched_domains_mutex);
2791 }