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(const 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 const 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 don't 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         /*
 520          * Because the rq is not a task, dl_add_task_root_domain() did not
 521          * move the fair server bw to the rd if it already started.
 522          * Add it now.
 523          */
 524         if (rq->fair_server.dl_server)
 525                 __dl_server_attach_root(&rq->fair_server, rq);
 526
 527         rq_unlock_irqrestore(rq, &rf);
 528
 529         if (old_rd)
 530                 call_rcu(&old_rd->rcu, free_rootdomain);
 531 }
 532
 533 void sched_get_rd(struct root_domain *rd)
 534 {
 535         atomic_inc(&rd->refcount);
 536 }
 537
 538 void sched_put_rd(struct root_domain *rd)
 539 {
 540         if (!atomic_dec_and_test(&rd->refcount))
 541                 return;
 542
 543         call_rcu(&rd->rcu, free_rootdomain);
 544 }
 545
 546 static int init_rootdomain(struct root_domain *rd)
 547 {
 548         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 549                 goto out;
 550         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 551                 goto free_span;
 552         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 553                 goto free_online;
 554         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 555                 goto free_dlo_mask;
 556
 557 #ifdef HAVE_RT_PUSH_IPI
 558         rd->rto_cpu = -1;
 559         raw_spin_lock_init(&rd->rto_lock);
 560         rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
 561 #endif
 562
 563         rd->visit_gen = 0;
 564         init_dl_bw(&rd->dl_bw);
 565         if (cpudl_init(&rd->cpudl) != 0)
 566                 goto free_rto_mask;
 567
 568         if (cpupri_init(&rd->cpupri) != 0)
 569                 goto free_cpudl;
 570         return 0;
 571
 572 free_cpudl:
 573         cpudl_cleanup(&rd->cpudl);
 574 free_rto_mask:
 575         free_cpumask_var(rd->rto_mask);
 576 free_dlo_mask:
 577         free_cpumask_var(rd->dlo_mask);
 578 free_online:
 579         free_cpumask_var(rd->online);
 580 free_span:
 581         free_cpumask_var(rd->span);
 582 out:
 583         return -ENOMEM;
 584 }
 585
 586 /*
 587  * By default the system creates a single root-domain with all CPUs as
 588  * members (mimicking the global state we have today).
 589  */
 590 struct root_domain def_root_domain;
 591
 592 void __init init_defrootdomain(void)
 593 {
 594         init_rootdomain(&def_root_domain);
 595
 596         atomic_set(&def_root_domain.refcount, 1);
 597 }
 598
 599 static struct root_domain *alloc_rootdomain(void)
 600 {
 601         struct root_domain *rd;
 602
 603         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 604         if (!rd)
 605                 return NULL;
 606
 607         if (init_rootdomain(rd) != 0) {
 608                 kfree(rd);
 609                 return NULL;
 610         }
 611
 612         return rd;
 613 }
 614
 615 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 616 {
 617         struct sched_group *tmp, *first;
 618
 619         if (!sg)
 620                 return;
 621
 622         first = sg;
 623         do {
 624                 tmp = sg->next;
 625
 626                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 627                         kfree(sg->sgc);
 628
 629                 if (atomic_dec_and_test(&sg->ref))
 630                         kfree(sg);
 631                 sg = tmp;
 632         } while (sg != first);
 633 }
 634
 635 static void destroy_sched_domain(struct sched_domain *sd)
 636 {
 637         /*
 638          * A normal sched domain may have multiple group references, an
 639          * overlapping domain, having private groups, only one.  Iterate,
 640          * dropping group/capacity references, freeing where none remain.
 641          */
 642         free_sched_groups(sd->groups, 1);
 643
 644         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 645                 kfree(sd->shared);
 646         kfree(sd);
 647 }
 648
 649 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 650 {
 651         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 652
 653         while (sd) {
 654                 struct sched_domain *parent = sd->parent;
 655                 destroy_sched_domain(sd);
 656                 sd = parent;
 657         }
 658 }
 659
 660 static void destroy_sched_domains(struct sched_domain *sd)
 661 {
 662         if (sd)
 663                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 664 }
 665
 666 /*
 667  * Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
 668  * (Last Level Cache Domain) for this allows us to avoid some pointer chasing
 669  * select_idle_sibling().
 670  *
 671  * Also keep a unique ID per domain (we use the first CPU number in the cpumask
 672  * of the domain), this allows us to quickly tell if two CPUs are in the same
 673  * cache domain, see cpus_share_cache().
 674  */
 675 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 676 DEFINE_PER_CPU(int, sd_llc_size);
 677 DEFINE_PER_CPU(int, sd_llc_id);
 678 DEFINE_PER_CPU(int, sd_share_id);
 679 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 680 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 681 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 682 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 683
 684 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 685 DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
 686
 687 static void update_top_cache_domain(int cpu)
 688 {
 689         struct sched_domain_shared *sds = NULL;
 690         struct sched_domain *sd;
 691         int id = cpu;
 692         int size = 1;
 693
 694         sd = highest_flag_domain(cpu, SD_SHARE_LLC);
 695         if (sd) {
 696                 id = cpumask_first(sched_domain_span(sd));
 697                 size = cpumask_weight(sched_domain_span(sd));
 698                 sds = sd->shared;
 699         }
 700
 701         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 702         per_cpu(sd_llc_size, cpu) = size;
 703         per_cpu(sd_llc_id, cpu) = id;
 704         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 705
 706         sd = lowest_flag_domain(cpu, SD_CLUSTER);
 707         if (sd)
 708                 id = cpumask_first(sched_domain_span(sd));
 709
 710         /*
 711          * This assignment should be placed after the sd_llc_id as
 712          * we want this id equals to cluster id on cluster machines
 713          * but equals to LLC id on non-Cluster machines.
 714          */
 715         per_cpu(sd_share_id, cpu) = id;
 716
 717         sd = lowest_flag_domain(cpu, SD_NUMA);
 718         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 719
 720         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 721         rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 722
 723         sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
 724         rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 725 }
 726
 727 /*
 728  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 729  * hold the hotplug lock.
 730  */
 731 static void
 732 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 733 {
 734         struct rq *rq = cpu_rq(cpu);
 735         struct sched_domain *tmp;
 736
 737         /* Remove the sched domains which do not contribute to scheduling. */
 738         for (tmp = sd; tmp; ) {
 739                 struct sched_domain *parent = tmp->parent;
 740                 if (!parent)
 741                         break;
 742
 743                 if (sd_parent_degenerate(tmp, parent)) {
 744                         tmp->parent = parent->parent;
 745
 746                         if (parent->parent) {
 747                                 parent->parent->child = tmp;
 748                                 parent->parent->groups->flags = tmp->flags;
 749                         }
 750
 751                         /*
 752                          * Transfer SD_PREFER_SIBLING down in case of a
 753                          * degenerate parent; the spans match for this
 754                          * so the property transfers.
 755                          */
 756                         if (parent->flags & SD_PREFER_SIBLING)
 757                                 tmp->flags |= SD_PREFER_SIBLING;
 758                         destroy_sched_domain(parent);
 759                 } else
 760                         tmp = tmp->parent;
 761         }
 762
 763         if (sd && sd_degenerate(sd)) {
 764                 tmp = sd;
 765                 sd = sd->parent;
 766                 destroy_sched_domain(tmp);
 767                 if (sd) {
 768                         struct sched_group *sg = sd->groups;
 769
 770                         /*
 771                          * sched groups hold the flags of the child sched
 772                          * domain for convenience. Clear such flags since
 773                          * the child is being destroyed.
 774                          */
 775                         do {
 776                                 sg->flags = 0;
 777                         } while (sg != sd->groups);
 778
 779                         sd->child = NULL;
 780                 }
 781         }
 782
 783         sched_domain_debug(sd, cpu);
 784
 785         rq_attach_root(rq, rd);
 786         tmp = rq->sd;
 787         rcu_assign_pointer(rq->sd, sd);
 788         dirty_sched_domain_sysctl(cpu);
 789         destroy_sched_domains(tmp);
 790
 791         update_top_cache_domain(cpu);
 792 }
 793
 794 struct s_data {
 795         struct sched_domain * __percpu *sd;
 796         struct root_domain      *rd;
 797 };
 798
 799 enum s_alloc {
 800         sa_rootdomain,
 801         sa_sd,
 802         sa_sd_storage,
 803         sa_none,
 804 };
 805
 806 /*
 807  * Return the canonical balance CPU for this group, this is the first CPU
 808  * of this group that's also in the balance mask.
 809  *
 810  * The balance mask are all those CPUs that could actually end up at this
 811  * group. See build_balance_mask().
 812  *
 813  * Also see should_we_balance().
 814  */
 815 int group_balance_cpu(struct sched_group *sg)
 816 {
 817         return cpumask_first(group_balance_mask(sg));
 818 }
 819
 820
 821 /*
 822  * NUMA topology (first read the regular topology blurb below)
 823  *
 824  * Given a node-distance table, for example:
 825  *
 826  *   node   0   1   2   3
 827  *     0:  10  20  30  20
 828  *     1:  20  10  20  30
 829  *     2:  30  20  10  20
 830  *     3:  20  30  20  10
 831  *
 832  * which represents a 4 node ring topology like:
 833  *
 834  *   0 ----- 1
 835  *   |       |
 836  *   |       |
 837  *   |       |
 838  *   3 ----- 2
 839  *
 840  * We want to construct domains and groups to represent this. The way we go
 841  * about doing this is to build the domains on 'hops'. For each NUMA level we
 842  * construct the mask of all nodes reachable in @level hops.
 843  *
 844  * For the above NUMA topology that gives 3 levels:
 845  *
 846  * NUMA-2       0-3             0-3             0-3             0-3
 847  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 848  *
 849  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 850  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 851  *
 852  * NUMA-0       0               1               2               3
 853  *
 854  *
 855  * As can be seen; things don't nicely line up as with the regular topology.
 856  * When we iterate a domain in child domain chunks some nodes can be
 857  * represented multiple times -- hence the "overlap" naming for this part of
 858  * the topology.
 859  *
 860  * In order to minimize this overlap, we only build enough groups to cover the
 861  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 862  *
 863  * Because:
 864  *
 865  *  - the first group of each domain is its child domain; this
 866  *    gets us the first 0-1,3
 867  *  - the only uncovered node is 2, who's child domain is 1-3.
 868  *
 869  * However, because of the overlap, computing a unique CPU for each group is
 870  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 871  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 872  * end up at those groups (they would end up in group: 0-1,3).
 873  *
 874  * To correct this we have to introduce the group balance mask. This mask
 875  * will contain those CPUs in the group that can reach this group given the
 876  * (child) domain tree.
 877  *
 878  * With this we can once again compute balance_cpu and sched_group_capacity
 879  * relations.
 880  *
 881  * XXX include words on how balance_cpu is unique and therefore can be
 882  * used for sched_group_capacity links.
 883  *
 884  *
 885  * Another 'interesting' topology is:
 886  *
 887  *   node   0   1   2   3
 888  *     0:  10  20  20  30
 889  *     1:  20  10  20  20
 890  *     2:  20  20  10  20
 891  *     3:  30  20  20  10
 892  *
 893  * Which looks a little like:
 894  *
 895  *   0 ----- 1
 896  *   |     / |
 897  *   |   /   |
 898  *   | /     |
 899  *   2 ----- 3
 900  *
 901  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 902  * are not.
 903  *
 904  * This leads to a few particularly weird cases where the sched_domain's are
 905  * not of the same number for each CPU. Consider:
 906  *
 907  * NUMA-2       0-3                                             0-3
 908  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 909  *
 910  * NUMA-1       0-2             0-3             0-3             1-3
 911  *
 912  * NUMA-0       0               1               2               3
 913  *
 914  */
 915
 916
 917 /*
 918  * Build the balance mask; it contains only those CPUs that can arrive at this
 919  * group and should be considered to continue balancing.
 920  *
 921  * We do this during the group creation pass, therefore the group information
 922  * isn't complete yet, however since each group represents a (child) domain we
 923  * can fully construct this using the sched_domain bits (which are already
 924  * complete).
 925  */
 926 static void
 927 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 928 {
 929         const struct cpumask *sg_span = sched_group_span(sg);
 930         struct sd_data *sdd = sd->private;
 931         struct sched_domain *sibling;
 932         int i;
 933
 934         cpumask_clear(mask);
 935
 936         for_each_cpu(i, sg_span) {
 937                 sibling = *per_cpu_ptr(sdd->sd, i);
 938
 939                 /*
 940                  * Can happen in the asymmetric case, where these siblings are
 941                  * unused. The mask will not be empty because those CPUs that
 942                  * do have the top domain _should_ span the domain.
 943                  */
 944                 if (!sibling->child)
 945                         continue;
 946
 947                 /* If we would not end up here, we can't continue from here */
 948                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 949                         continue;
 950
 951                 cpumask_set_cpu(i, mask);
 952         }
 953
 954         /* We must not have empty masks here */
 955         WARN_ON_ONCE(cpumask_empty(mask));
 956 }
 957
 958 /*
 959  * XXX: This creates per-node group entries; since the load-balancer will
 960  * immediately access remote memory to construct this group's load-balance
 961  * statistics having the groups node local is of dubious benefit.
 962  */
 963 static struct sched_group *
 964 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 965 {
 966         struct sched_group *sg;
 967         struct cpumask *sg_span;
 968
 969         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 970                         GFP_KERNEL, cpu_to_node(cpu));
 971
 972         if (!sg)
 973                 return NULL;
 974
 975         sg_span = sched_group_span(sg);
 976         if (sd->child) {
 977                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 978                 sg->flags = sd->child->flags;
 979         } else {
 980                 cpumask_copy(sg_span, sched_domain_span(sd));
 981         }
 982
 983         atomic_inc(&sg->ref);
 984         return sg;
 985 }
 986
 987 static void init_overlap_sched_group(struct sched_domain *sd,
 988                                      struct sched_group *sg)
 989 {
 990         struct cpumask *mask = sched_domains_tmpmask2;
 991         struct sd_data *sdd = sd->private;
 992         struct cpumask *sg_span;
 993         int cpu;
 994
 995         build_balance_mask(sd, sg, mask);
 996         cpu = cpumask_first(mask);
 997
 998         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 999         if (atomic_inc_return(&sg->sgc->ref) == 1)
1000                 cpumask_copy(group_balance_mask(sg), mask);
1001         else
1002                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
1003
1004         /*
1005          * Initialize sgc->capacity such that even if we mess up the
1006          * domains and no possible iteration will get us here, we won't
1007          * die on a /0 trap.
1008          */
1009         sg_span = sched_group_span(sg);
1010         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
1011         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1012         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1013 }
1014
1015 static struct sched_domain *
1016 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1017 {
1018         /*
1019          * The proper descendant would be the one whose child won't span out
1020          * of sd
1021          */
1022         while (sibling->child &&
1023                !cpumask_subset(sched_domain_span(sibling->child),
1024                                sched_domain_span(sd)))
1025                 sibling = sibling->child;
1026
1027         /*
1028          * As we are referencing sgc across different topology level, we need
1029          * to go down to skip those sched_domains which don't contribute to
1030          * scheduling because they will be degenerated in cpu_attach_domain
1031          */
1032         while (sibling->child &&
1033                cpumask_equal(sched_domain_span(sibling->child),
1034                              sched_domain_span(sibling)))
1035                 sibling = sibling->child;
1036
1037         return sibling;
1038 }
1039
1040 static int
1041 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1042 {
1043         struct sched_group *first = NULL, *last = NULL, *sg;
1044         const struct cpumask *span = sched_domain_span(sd);
1045         struct cpumask *covered = sched_domains_tmpmask;
1046         struct sd_data *sdd = sd->private;
1047         struct sched_domain *sibling;
1048         int i;
1049
1050         cpumask_clear(covered);
1051
1052         for_each_cpu_wrap(i, span, cpu) {
1053                 struct cpumask *sg_span;
1054
1055                 if (cpumask_test_cpu(i, covered))
1056                         continue;
1057
1058                 sibling = *per_cpu_ptr(sdd->sd, i);
1059
1060                 /*
1061                  * Asymmetric node setups can result in situations where the
1062                  * domain tree is of unequal depth, make sure to skip domains
1063                  * that already cover the entire range.
1064                  *
1065                  * In that case build_sched_domains() will have terminated the
1066                  * iteration early and our sibling sd spans will be empty.
1067                  * Domains should always include the CPU they're built on, so
1068                  * check that.
1069                  */
1070                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1071                         continue;
1072
1073                 /*
1074                  * Usually we build sched_group by sibling's child sched_domain
1075                  * But for machines whose NUMA diameter are 3 or above, we move
1076                  * to build sched_group by sibling's proper descendant's child
1077                  * domain because sibling's child sched_domain will span out of
1078                  * the sched_domain being built as below.
1079                  *
1080                  * Smallest diameter=3 topology is:
1081                  *
1082                  *   node   0   1   2   3
1083                  *     0:  10  20  30  40
1084                  *     1:  20  10  20  30
1085                  *     2:  30  20  10  20
1086                  *     3:  40  30  20  10
1087                  *
1088                  *   0 --- 1 --- 2 --- 3
1089                  *
1090                  * NUMA-3       0-3             N/A             N/A             0-3
1091                  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1092                  *
1093                  * NUMA-2       0-2             0-3             0-3             1-3
1094                  *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1095                  *
1096                  * NUMA-1       0-1             0-2             1-3             2-3
1097                  *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1098                  *
1099                  * NUMA-0       0               1               2               3
1100                  *
1101                  * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1102                  * group span isn't a subset of the domain span.
1103                  */
1104                 if (sibling->child &&
1105                     !cpumask_subset(sched_domain_span(sibling->child), span))
1106                         sibling = find_descended_sibling(sd, sibling);
1107
1108                 sg = build_group_from_child_sched_domain(sibling, cpu);
1109                 if (!sg)
1110                         goto fail;
1111
1112                 sg_span = sched_group_span(sg);
1113                 cpumask_or(covered, covered, sg_span);
1114
1115                 init_overlap_sched_group(sibling, sg);
1116
1117                 if (!first)
1118                         first = sg;
1119                 if (last)
1120                         last->next = sg;
1121                 last = sg;
1122                 last->next = first;
1123         }
1124         sd->groups = first;
1125
1126         return 0;
1127
1128 fail:
1129         free_sched_groups(first, 0);
1130
1131         return -ENOMEM;
1132 }
1133
1134
1135 /*
1136  * Package topology (also see the load-balance blurb in fair.c)
1137  *
1138  * The scheduler builds a tree structure to represent a number of important
1139  * topology features. By default (default_topology[]) these include:
1140  *
1141  *  - Simultaneous multithreading (SMT)
1142  *  - Multi-Core Cache (MC)
1143  *  - Package (PKG)
1144  *
1145  * Where the last one more or less denotes everything up to a NUMA node.
1146  *
1147  * The tree consists of 3 primary data structures:
1148  *
1149  *      sched_domain -> sched_group -> sched_group_capacity
1150  *          ^ ^             ^ ^
1151  *          `-'             `-'
1152  *
1153  * The sched_domains are per-CPU and have a two way link (parent & child) and
1154  * denote the ever growing mask of CPUs belonging to that level of topology.
1155  *
1156  * Each sched_domain has a circular (double) linked list of sched_group's, each
1157  * denoting the domains of the level below (or individual CPUs in case of the
1158  * first domain level). The sched_group linked by a sched_domain includes the
1159  * CPU of that sched_domain [*].
1160  *
1161  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1162  *
1163  * CPU   0   1   2   3   4   5   6   7
1164  *
1165  * PKG  [                             ]
1166  * MC   [             ] [             ]
1167  * SMT  [     ] [     ] [     ] [     ]
1168  *
1169  *  - or -
1170  *
1171  * PKG  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1172  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1173  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1174  *
1175  * CPU   0   1   2   3   4   5   6   7
1176  *
1177  * One way to think about it is: sched_domain moves you up and down among these
1178  * topology levels, while sched_group moves you sideways through it, at child
1179  * domain granularity.
1180  *
1181  * sched_group_capacity ensures each unique sched_group has shared storage.
1182  *
1183  * There are two related construction problems, both require a CPU that
1184  * uniquely identify each group (for a given domain):
1185  *
1186  *  - The first is the balance_cpu (see should_we_balance() and the
1187  *    load-balance blurb in fair.c); for each group we only want 1 CPU to
1188  *    continue balancing at a higher domain.
1189  *
1190  *  - The second is the sched_group_capacity; we want all identical groups
1191  *    to share a single sched_group_capacity.
1192  *
1193  * Since these topologies are exclusive by construction. That is, its
1194  * impossible for an SMT thread to belong to multiple cores, and cores to
1195  * be part of multiple caches. There is a very clear and unique location
1196  * for each CPU in the hierarchy.
1197  *
1198  * Therefore computing a unique CPU for each group is trivial (the iteration
1199  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1200  * group), we can simply pick the first CPU in each group.
1201  *
1202  *
1203  * [*] in other words, the first group of each domain is its child domain.
1204  */
1205
1206 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1207 {
1208         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1209         struct sched_domain *child = sd->child;
1210         struct sched_group *sg;
1211         bool already_visited;
1212
1213         if (child)
1214                 cpu = cpumask_first(sched_domain_span(child));
1215
1216         sg = *per_cpu_ptr(sdd->sg, cpu);
1217         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1218
1219         /* Increase refcounts for claim_allocations: */
1220         already_visited = atomic_inc_return(&sg->ref) > 1;
1221         /* sgc visits should follow a similar trend as sg */
1222         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1223
1224         /* If we have already visited that group, it's already initialized. */
1225         if (already_visited)
1226                 return sg;
1227
1228         if (child) {
1229                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1230                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1231                 sg->flags = child->flags;
1232         } else {
1233                 cpumask_set_cpu(cpu, sched_group_span(sg));
1234                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1235         }
1236
1237         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1238         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1239         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1240
1241         return sg;
1242 }
1243
1244 /*
1245  * build_sched_groups will build a circular linked list of the groups
1246  * covered by the given span, will set each group's ->cpumask correctly,
1247  * and will initialize their ->sgc.
1248  *
1249  * Assumes the sched_domain tree is fully constructed
1250  */
1251 static int
1252 build_sched_groups(struct sched_domain *sd, int cpu)
1253 {
1254         struct sched_group *first = NULL, *last = NULL;
1255         struct sd_data *sdd = sd->private;
1256         const struct cpumask *span = sched_domain_span(sd);
1257         struct cpumask *covered;
1258         int i;
1259
1260         lockdep_assert_held(&sched_domains_mutex);
1261         covered = sched_domains_tmpmask;
1262
1263         cpumask_clear(covered);
1264
1265         for_each_cpu_wrap(i, span, cpu) {
1266                 struct sched_group *sg;
1267
1268                 if (cpumask_test_cpu(i, covered))
1269                         continue;
1270
1271                 sg = get_group(i, sdd);
1272
1273                 cpumask_or(covered, covered, sched_group_span(sg));
1274
1275                 if (!first)
1276                         first = sg;
1277                 if (last)
1278                         last->next = sg;
1279                 last = sg;
1280         }
1281         last->next = first;
1282         sd->groups = first;
1283
1284         return 0;
1285 }
1286
1287 /*
1288  * Initialize sched groups cpu_capacity.
1289  *
1290  * cpu_capacity indicates the capacity of sched group, which is used while
1291  * distributing the load between different sched groups in a sched domain.
1292  * Typically cpu_capacity for all the groups in a sched domain will be same
1293  * unless there are asymmetries in the topology. If there are asymmetries,
1294  * group having more cpu_capacity will pickup more load compared to the
1295  * group having less cpu_capacity.
1296  */
1297 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1298 {
1299         struct sched_group *sg = sd->groups;
1300         struct cpumask *mask = sched_domains_tmpmask2;
1301
1302         WARN_ON(!sg);
1303
1304         do {
1305                 int cpu, cores = 0, max_cpu = -1;
1306
1307                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1308
1309                 cpumask_copy(mask, sched_group_span(sg));
1310                 for_each_cpu(cpu, mask) {
1311                         cores++;
1312 #ifdef CONFIG_SCHED_SMT
1313                         cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1314 #endif
1315                 }
1316                 sg->cores = cores;
1317
1318                 if (!(sd->flags & SD_ASYM_PACKING))
1319                         goto next;
1320
1321                 for_each_cpu(cpu, sched_group_span(sg)) {
1322                         if (max_cpu < 0)
1323                                 max_cpu = cpu;
1324                         else if (sched_asym_prefer(cpu, max_cpu))
1325                                 max_cpu = cpu;
1326                 }
1327                 sg->asym_prefer_cpu = max_cpu;
1328
1329 next:
1330                 sg = sg->next;
1331         } while (sg != sd->groups);
1332
1333         if (cpu != group_balance_cpu(sg))
1334                 return;
1335
1336         update_group_capacity(sd, cpu);
1337 }
1338
1339 /*
1340  * Set of available CPUs grouped by their corresponding capacities
1341  * Each list entry contains a CPU mask reflecting CPUs that share the same
1342  * capacity.
1343  * The lifespan of data is unlimited.
1344  */
1345 LIST_HEAD(asym_cap_list);
1346
1347 /*
1348  * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1349  * Provides sd_flags reflecting the asymmetry scope.
1350  */
1351 static inline int
1352 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1353                            const struct cpumask *cpu_map)
1354 {
1355         struct asym_cap_data *entry;
1356         int count = 0, miss = 0;
1357
1358         /*
1359          * Count how many unique CPU capacities this domain spans across
1360          * (compare sched_domain CPUs mask with ones representing  available
1361          * CPUs capacities). Take into account CPUs that might be offline:
1362          * skip those.
1363          */
1364         list_for_each_entry(entry, &asym_cap_list, link) {
1365                 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1366                         ++count;
1367                 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1368                         ++miss;
1369         }
1370
1371         WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1372
1373         /* No asymmetry detected */
1374         if (count < 2)
1375                 return 0;
1376         /* Some of the available CPU capacity values have not been detected */
1377         if (miss)
1378                 return SD_ASYM_CPUCAPACITY;
1379
1380         /* Full asymmetry */
1381         return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1382
1383 }
1384
1385 static void free_asym_cap_entry(struct rcu_head *head)
1386 {
1387         struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
1388         kfree(entry);
1389 }
1390
1391 static inline void asym_cpu_capacity_update_data(int cpu)
1392 {
1393         unsigned long capacity = arch_scale_cpu_capacity(cpu);
1394         struct asym_cap_data *insert_entry = NULL;
1395         struct asym_cap_data *entry;
1396
1397         /*
1398          * Search if capacity already exits. If not, track which the entry
1399          * where we should insert to keep the list ordered descending.
1400          */
1401         list_for_each_entry(entry, &asym_cap_list, link) {
1402                 if (capacity == entry->capacity)
1403                         goto done;
1404                 else if (!insert_entry && capacity > entry->capacity)
1405                         insert_entry = list_prev_entry(entry, link);
1406         }
1407
1408         entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1409         if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1410                 return;
1411         entry->capacity = capacity;
1412
1413         /* If NULL then the new capacity is the smallest, add last. */
1414         if (!insert_entry)
1415                 list_add_tail_rcu(&entry->link, &asym_cap_list);
1416         else
1417                 list_add_rcu(&entry->link, &insert_entry->link);
1418 done:
1419         __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1420 }
1421
1422 /*
1423  * Build-up/update list of CPUs grouped by their capacities
1424  * An update requires explicit request to rebuild sched domains
1425  * with state indicating CPU topology changes.
1426  */
1427 static void asym_cpu_capacity_scan(void)
1428 {
1429         struct asym_cap_data *entry, *next;
1430         int cpu;
1431
1432         list_for_each_entry(entry, &asym_cap_list, link)
1433                 cpumask_clear(cpu_capacity_span(entry));
1434
1435         for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1436                 asym_cpu_capacity_update_data(cpu);
1437
1438         list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1439                 if (cpumask_empty(cpu_capacity_span(entry))) {
1440                         list_del_rcu(&entry->link);
1441                         call_rcu(&entry->rcu, free_asym_cap_entry);
1442                 }
1443         }
1444
1445         /*
1446          * Only one capacity value has been detected i.e. this system is symmetric.
1447          * No need to keep this data around.
1448          */
1449         if (list_is_singular(&asym_cap_list)) {
1450                 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1451                 list_del_rcu(&entry->link);
1452                 call_rcu(&entry->rcu, free_asym_cap_entry);
1453         }
1454 }
1455
1456 /*
1457  * Initializers for schedule domains
1458  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1459  */
1460
1461 static int default_relax_domain_level = -1;
1462 int sched_domain_level_max;
1463
1464 static int __init setup_relax_domain_level(char *str)
1465 {
1466         if (kstrtoint(str, 0, &default_relax_domain_level))
1467                 pr_warn("Unable to set relax_domain_level\n");
1468
1469         return 1;
1470 }
1471 __setup("relax_domain_level=", setup_relax_domain_level);
1472
1473 static void set_domain_attribute(struct sched_domain *sd,
1474                                  struct sched_domain_attr *attr)
1475 {
1476         int request;
1477
1478         if (!attr || attr->relax_domain_level < 0) {
1479                 if (default_relax_domain_level < 0)
1480                         return;
1481                 request = default_relax_domain_level;
1482         } else
1483                 request = attr->relax_domain_level;
1484
1485         if (sd->level >= request) {
1486                 /* Turn off idle balance on this domain: */
1487                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1488         }
1489 }
1490
1491 static void __sdt_free(const struct cpumask *cpu_map);
1492 static int __sdt_alloc(const struct cpumask *cpu_map);
1493
1494 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1495                                  const struct cpumask *cpu_map)
1496 {
1497         switch (what) {
1498         case sa_rootdomain:
1499                 if (!atomic_read(&d->rd->refcount))
1500                         free_rootdomain(&d->rd->rcu);
1501                 fallthrough;
1502         case sa_sd:
1503                 free_percpu(d->sd);
1504                 fallthrough;
1505         case sa_sd_storage:
1506                 __sdt_free(cpu_map);
1507                 fallthrough;
1508         case sa_none:
1509                 break;
1510         }
1511 }
1512
1513 static enum s_alloc
1514 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1515 {
1516         memset(d, 0, sizeof(*d));
1517
1518         if (__sdt_alloc(cpu_map))
1519                 return sa_sd_storage;
1520         d->sd = alloc_percpu(struct sched_domain *);
1521         if (!d->sd)
1522                 return sa_sd_storage;
1523         d->rd = alloc_rootdomain();
1524         if (!d->rd)
1525                 return sa_sd;
1526
1527         return sa_rootdomain;
1528 }
1529
1530 /*
1531  * NULL the sd_data elements we've used to build the sched_domain and
1532  * sched_group structure so that the subsequent __free_domain_allocs()
1533  * will not free the data we're using.
1534  */
1535 static void claim_allocations(int cpu, struct sched_domain *sd)
1536 {
1537         struct sd_data *sdd = sd->private;
1538
1539         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1540         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1541
1542         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1543                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1544
1545         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1546                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1547
1548         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1549                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1550 }
1551
1552 #ifdef CONFIG_NUMA
1553 enum numa_topology_type sched_numa_topology_type;
1554
1555 static int                      sched_domains_numa_levels;
1556 static int                      sched_domains_curr_level;
1557
1558 int                             sched_max_numa_distance;
1559 static int                      *sched_domains_numa_distance;
1560 static struct cpumask           ***sched_domains_numa_masks;
1561 #endif
1562
1563 /*
1564  * SD_flags allowed in topology descriptions.
1565  *
1566  * These flags are purely descriptive of the topology and do not prescribe
1567  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1568  * function. For details, see include/linux/sched/sd_flags.h.
1569  *
1570  *   SD_SHARE_CPUCAPACITY
1571  *   SD_SHARE_LLC
1572  *   SD_CLUSTER
1573  *   SD_NUMA
1574  *
1575  * Odd one out, which beside describing the topology has a quirk also
1576  * prescribes the desired behaviour that goes along with it:
1577  *
1578  *   SD_ASYM_PACKING        - describes SMT quirks
1579  */
1580 #define TOPOLOGY_SD_FLAGS               \
1581         (SD_SHARE_CPUCAPACITY   |       \
1582          SD_CLUSTER             |       \
1583          SD_SHARE_LLC           |       \
1584          SD_NUMA                |       \
1585          SD_ASYM_PACKING)
1586
1587 static struct sched_domain *
1588 sd_init(struct sched_domain_topology_level *tl,
1589         const struct cpumask *cpu_map,
1590         struct sched_domain *child, int cpu)
1591 {
1592         struct sd_data *sdd = &tl->data;
1593         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1594         int sd_id, sd_weight, sd_flags = 0;
1595         struct cpumask *sd_span;
1596
1597 #ifdef CONFIG_NUMA
1598         /*
1599          * Ugly hack to pass state to sd_numa_mask()...
1600          */
1601         sched_domains_curr_level = tl->numa_level;
1602 #endif
1603
1604         sd_weight = cpumask_weight(tl->mask(cpu));
1605
1606         if (tl->sd_flags)
1607                 sd_flags = (*tl->sd_flags)();
1608         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1609                         "wrong sd_flags in topology description\n"))
1610                 sd_flags &= TOPOLOGY_SD_FLAGS;
1611
1612         *sd = (struct sched_domain){
1613                 .min_interval           = sd_weight,
1614                 .max_interval           = 2*sd_weight,
1615                 .busy_factor            = 16,
1616                 .imbalance_pct          = 117,
1617
1618                 .cache_nice_tries       = 0,
1619
1620                 .flags                  = 1*SD_BALANCE_NEWIDLE
1621                                         | 1*SD_BALANCE_EXEC
1622                                         | 1*SD_BALANCE_FORK
1623                                         | 0*SD_BALANCE_WAKE
1624                                         | 1*SD_WAKE_AFFINE
1625                                         | 0*SD_SHARE_CPUCAPACITY
1626                                         | 0*SD_SHARE_LLC
1627                                         | 0*SD_SERIALIZE
1628                                         | 1*SD_PREFER_SIBLING
1629                                         | 0*SD_NUMA
1630                                         | sd_flags
1631                                         ,
1632
1633                 .last_balance           = jiffies,
1634                 .balance_interval       = sd_weight,
1635                 .max_newidle_lb_cost    = 0,
1636                 .last_decay_max_lb_cost = jiffies,
1637                 .child                  = child,
1638                 .name                   = tl->name,
1639         };
1640
1641         sd_span = sched_domain_span(sd);
1642         cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1643         sd_id = cpumask_first(sd_span);
1644
1645         sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1646
1647         WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1648                   (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1649                   "CPU capacity asymmetry not supported on SMT\n");
1650
1651         /*
1652          * Convert topological properties into behaviour.
1653          */
1654         /* Don't attempt to spread across CPUs of different capacities. */
1655         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1656                 sd->child->flags &= ~SD_PREFER_SIBLING;
1657
1658         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1659                 sd->imbalance_pct = 110;
1660
1661         } else if (sd->flags & SD_SHARE_LLC) {
1662                 sd->imbalance_pct = 117;
1663                 sd->cache_nice_tries = 1;
1664
1665 #ifdef CONFIG_NUMA
1666         } else if (sd->flags & SD_NUMA) {
1667                 sd->cache_nice_tries = 2;
1668
1669                 sd->flags &= ~SD_PREFER_SIBLING;
1670                 sd->flags |= SD_SERIALIZE;
1671                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1672                         sd->flags &= ~(SD_BALANCE_EXEC |
1673                                        SD_BALANCE_FORK |
1674                                        SD_WAKE_AFFINE);
1675                 }
1676
1677 #endif
1678         } else {
1679                 sd->cache_nice_tries = 1;
1680         }
1681
1682         /*
1683          * For all levels sharing cache; connect a sched_domain_shared
1684          * instance.
1685          */
1686         if (sd->flags & SD_SHARE_LLC) {
1687                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1688                 atomic_inc(&sd->shared->ref);
1689                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1690         }
1691
1692         sd->private = sdd;
1693
1694         return sd;
1695 }
1696
1697 /*
1698  * Topology list, bottom-up.
1699  */
1700 static struct sched_domain_topology_level default_topology[] = {
1701 #ifdef CONFIG_SCHED_SMT
1702         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1703 #endif
1704
1705 #ifdef CONFIG_SCHED_CLUSTER
1706         { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1707 #endif
1708
1709 #ifdef CONFIG_SCHED_MC
1710         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1711 #endif
1712         { cpu_cpu_mask, SD_INIT_NAME(PKG) },
1713         { NULL, },
1714 };
1715
1716 static struct sched_domain_topology_level *sched_domain_topology =
1717         default_topology;
1718 static struct sched_domain_topology_level *sched_domain_topology_saved;
1719
1720 #define for_each_sd_topology(tl)                        \
1721         for (tl = sched_domain_topology; tl->mask; tl++)
1722
1723 void __init set_sched_topology(struct sched_domain_topology_level *tl)
1724 {
1725         if (WARN_ON_ONCE(sched_smp_initialized))
1726                 return;
1727
1728         sched_domain_topology = tl;
1729         sched_domain_topology_saved = NULL;
1730 }
1731
1732 #ifdef CONFIG_NUMA
1733
1734 static const struct cpumask *sd_numa_mask(int cpu)
1735 {
1736         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1737 }
1738
1739 static void sched_numa_warn(const char *str)
1740 {
1741         static int done = false;
1742         int i,j;
1743
1744         if (done)
1745                 return;
1746
1747         done = true;
1748
1749         printk(KERN_WARNING "ERROR: %s\n\n", str);
1750
1751         for (i = 0; i < nr_node_ids; i++) {
1752                 printk(KERN_WARNING "  ");
1753                 for (j = 0; j < nr_node_ids; j++) {
1754                         if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1755                                 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1756                         else
1757                                 printk(KERN_CONT " %02d  ", node_distance(i,j));
1758                 }
1759                 printk(KERN_CONT "\n");
1760         }
1761         printk(KERN_WARNING "\n");
1762 }
1763
1764 bool find_numa_distance(int distance)
1765 {
1766         bool found = false;
1767         int i, *distances;
1768
1769         if (distance == node_distance(0, 0))
1770                 return true;
1771
1772         rcu_read_lock();
1773         distances = rcu_dereference(sched_domains_numa_distance);
1774         if (!distances)
1775                 goto unlock;
1776         for (i = 0; i < sched_domains_numa_levels; i++) {
1777                 if (distances[i] == distance) {
1778                         found = true;
1779                         break;
1780                 }
1781         }
1782 unlock:
1783         rcu_read_unlock();
1784
1785         return found;
1786 }
1787
1788 #define for_each_cpu_node_but(n, nbut)          \
1789         for_each_node_state(n, N_CPU)           \
1790                 if (n == nbut)                  \
1791                         continue;               \
1792                 else
1793
1794 /*
1795  * A system can have three types of NUMA topology:
1796  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1797  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1798  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1799  *
1800  * The difference between a glueless mesh topology and a backplane
1801  * topology lies in whether communication between not directly
1802  * connected nodes goes through intermediary nodes (where programs
1803  * could run), or through backplane controllers. This affects
1804  * placement of programs.
1805  *
1806  * The type of topology can be discerned with the following tests:
1807  * - If the maximum distance between any nodes is 1 hop, the system
1808  *   is directly connected.
1809  * - If for two nodes A and B, located N > 1 hops away from each other,
1810  *   there is an intermediary node C, which is < N hops away from both
1811  *   nodes A and B, the system is a glueless mesh.
1812  */
1813 static void init_numa_topology_type(int offline_node)
1814 {
1815         int a, b, c, n;
1816
1817         n = sched_max_numa_distance;
1818
1819         if (sched_domains_numa_levels <= 2) {
1820                 sched_numa_topology_type = NUMA_DIRECT;
1821                 return;
1822         }
1823
1824         for_each_cpu_node_but(a, offline_node) {
1825                 for_each_cpu_node_but(b, offline_node) {
1826                         /* Find two nodes furthest removed from each other. */
1827                         if (node_distance(a, b) < n)
1828                                 continue;
1829
1830                         /* Is there an intermediary node between a and b? */
1831                         for_each_cpu_node_but(c, offline_node) {
1832                                 if (node_distance(a, c) < n &&
1833                                     node_distance(b, c) < n) {
1834                                         sched_numa_topology_type =
1835                                                         NUMA_GLUELESS_MESH;
1836                                         return;
1837                                 }
1838                         }
1839
1840                         sched_numa_topology_type = NUMA_BACKPLANE;
1841                         return;
1842                 }
1843         }
1844
1845         pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1846         sched_numa_topology_type = NUMA_DIRECT;
1847 }
1848
1849
1850 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1851
1852 void sched_init_numa(int offline_node)
1853 {
1854         struct sched_domain_topology_level *tl;
1855         unsigned long *distance_map;
1856         int nr_levels = 0;
1857         int i, j;
1858         int *distances;
1859         struct cpumask ***masks;
1860
1861         /*
1862          * O(nr_nodes^2) de-duplicating selection sort -- in order to find the
1863          * unique distances in the node_distance() table.
1864          */
1865         distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1866         if (!distance_map)
1867                 return;
1868
1869         bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1870         for_each_cpu_node_but(i, offline_node) {
1871                 for_each_cpu_node_but(j, offline_node) {
1872                         int distance = node_distance(i, j);
1873
1874                         if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1875                                 sched_numa_warn("Invalid distance value range");
1876                                 bitmap_free(distance_map);
1877                                 return;
1878                         }
1879
1880                         bitmap_set(distance_map, distance, 1);
1881                 }
1882         }
1883         /*
1884          * We can now figure out how many unique distance values there are and
1885          * allocate memory accordingly.
1886          */
1887         nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1888
1889         distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1890         if (!distances) {
1891                 bitmap_free(distance_map);
1892                 return;
1893         }
1894
1895         for (i = 0, j = 0; i < nr_levels; i++, j++) {
1896                 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1897                 distances[i] = j;
1898         }
1899         rcu_assign_pointer(sched_domains_numa_distance, distances);
1900
1901         bitmap_free(distance_map);
1902
1903         /*
1904          * 'nr_levels' contains the number of unique distances
1905          *
1906          * The sched_domains_numa_distance[] array includes the actual distance
1907          * numbers.
1908          */
1909
1910         /*
1911          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1912          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1913          * the array will contain less then 'nr_levels' members. This could be
1914          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1915          * in other functions.
1916          *
1917          * We reset it to 'nr_levels' at the end of this function.
1918          */
1919         sched_domains_numa_levels = 0;
1920
1921         masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1922         if (!masks)
1923                 return;
1924
1925         /*
1926          * Now for each level, construct a mask per node which contains all
1927          * CPUs of nodes that are that many hops away from us.
1928          */
1929         for (i = 0; i < nr_levels; i++) {
1930                 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1931                 if (!masks[i])
1932                         return;
1933
1934                 for_each_cpu_node_but(j, offline_node) {
1935                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1936                         int k;
1937
1938                         if (!mask)
1939                                 return;
1940
1941                         masks[i][j] = mask;
1942
1943                         for_each_cpu_node_but(k, offline_node) {
1944                                 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1945                                         sched_numa_warn("Node-distance not symmetric");
1946
1947                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1948                                         continue;
1949
1950                                 cpumask_or(mask, mask, cpumask_of_node(k));
1951                         }
1952                 }
1953         }
1954         rcu_assign_pointer(sched_domains_numa_masks, masks);
1955
1956         /* Compute default topology size */
1957         for (i = 0; sched_domain_topology[i].mask; i++);
1958
1959         tl = kzalloc((i + nr_levels + 1) *
1960                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1961         if (!tl)
1962                 return;
1963
1964         /*
1965          * Copy the default topology bits..
1966          */
1967         for (i = 0; sched_domain_topology[i].mask; i++)
1968                 tl[i] = sched_domain_topology[i];
1969
1970         /*
1971          * Add the NUMA identity distance, aka single NODE.
1972          */
1973         tl[i++] = (struct sched_domain_topology_level){
1974                 .mask = sd_numa_mask,
1975                 .numa_level = 0,
1976                 SD_INIT_NAME(NODE)
1977         };
1978
1979         /*
1980          * .. and append 'j' levels of NUMA goodness.
1981          */
1982         for (j = 1; j < nr_levels; i++, j++) {
1983                 tl[i] = (struct sched_domain_topology_level){
1984                         .mask = sd_numa_mask,
1985                         .sd_flags = cpu_numa_flags,
1986                         .flags = SDTL_OVERLAP,
1987                         .numa_level = j,
1988                         SD_INIT_NAME(NUMA)
1989                 };
1990         }
1991
1992         sched_domain_topology_saved = sched_domain_topology;
1993         sched_domain_topology = tl;
1994
1995         sched_domains_numa_levels = nr_levels;
1996         WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1997
1998         init_numa_topology_type(offline_node);
1999 }
2000
2001
2002 static void sched_reset_numa(void)
2003 {
2004         int nr_levels, *distances;
2005         struct cpumask ***masks;
2006
2007         nr_levels = sched_domains_numa_levels;
2008         sched_domains_numa_levels = 0;
2009         sched_max_numa_distance = 0;
2010         sched_numa_topology_type = NUMA_DIRECT;
2011         distances = sched_domains_numa_distance;
2012         rcu_assign_pointer(sched_domains_numa_distance, NULL);
2013         masks = sched_domains_numa_masks;
2014         rcu_assign_pointer(sched_domains_numa_masks, NULL);
2015         if (distances || masks) {
2016                 int i, j;
2017
2018                 synchronize_rcu();
2019                 kfree(distances);
2020                 for (i = 0; i < nr_levels && masks; i++) {
2021                         if (!masks[i])
2022                                 continue;
2023                         for_each_node(j)
2024                                 kfree(masks[i][j]);
2025                         kfree(masks[i]);
2026                 }
2027                 kfree(masks);
2028         }
2029         if (sched_domain_topology_saved) {
2030                 kfree(sched_domain_topology);
2031                 sched_domain_topology = sched_domain_topology_saved;
2032                 sched_domain_topology_saved = NULL;
2033         }
2034 }
2035
2036 /*
2037  * Call with hotplug lock held
2038  */
2039 void sched_update_numa(int cpu, bool online)
2040 {
2041         int node;
2042
2043         node = cpu_to_node(cpu);
2044         /*
2045          * Scheduler NUMA topology is updated when the first CPU of a
2046          * node is onlined or the last CPU of a node is offlined.
2047          */
2048         if (cpumask_weight(cpumask_of_node(node)) != 1)
2049                 return;
2050
2051         sched_reset_numa();
2052         sched_init_numa(online ? NUMA_NO_NODE : node);
2053 }
2054
2055 void sched_domains_numa_masks_set(unsigned int cpu)
2056 {
2057         int node = cpu_to_node(cpu);
2058         int i, j;
2059
2060         for (i = 0; i < sched_domains_numa_levels; i++) {
2061                 for (j = 0; j < nr_node_ids; j++) {
2062                         if (!node_state(j, N_CPU))
2063                                 continue;
2064
2065                         /* Set ourselves in the remote node's masks */
2066                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
2067                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2068                 }
2069         }
2070 }
2071
2072 void sched_domains_numa_masks_clear(unsigned int cpu)
2073 {
2074         int i, j;
2075
2076         for (i = 0; i < sched_domains_numa_levels; i++) {
2077                 for (j = 0; j < nr_node_ids; j++) {
2078                         if (sched_domains_numa_masks[i][j])
2079                                 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2080                 }
2081         }
2082 }
2083
2084 /*
2085  * sched_numa_find_closest() - given the NUMA topology, find the cpu
2086  *                             closest to @cpu from @cpumask.
2087  * cpumask: cpumask to find a cpu from
2088  * cpu: cpu to be close to
2089  *
2090  * returns: cpu, or nr_cpu_ids when nothing found.
2091  */
2092 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2093 {
2094         int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2095         struct cpumask ***masks;
2096
2097         rcu_read_lock();
2098         masks = rcu_dereference(sched_domains_numa_masks);
2099         if (!masks)
2100                 goto unlock;
2101         for (i = 0; i < sched_domains_numa_levels; i++) {
2102                 if (!masks[i][j])
2103                         break;
2104                 cpu = cpumask_any_and(cpus, masks[i][j]);
2105                 if (cpu < nr_cpu_ids) {
2106                         found = cpu;
2107                         break;
2108                 }
2109         }
2110 unlock:
2111         rcu_read_unlock();
2112
2113         return found;
2114 }
2115
2116 struct __cmp_key {
2117         const struct cpumask *cpus;
2118         struct cpumask ***masks;
2119         int node;
2120         int cpu;
2121         int w;
2122 };
2123
2124 static int hop_cmp(const void *a, const void *b)
2125 {
2126         struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2127         struct __cmp_key *k = (struct __cmp_key *)a;
2128
2129         if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2130                 return 1;
2131
2132         if (b == k->masks) {
2133                 k->w = 0;
2134                 return 0;
2135         }
2136
2137         prev_hop = *((struct cpumask ***)b - 1);
2138         k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2139         if (k->w <= k->cpu)
2140                 return 0;
2141
2142         return -1;
2143 }
2144
2145 /**
2146  * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2147  *                             from @cpus to @cpu, taking into account distance
2148  *                             from a given @node.
2149  * @cpus: cpumask to find a cpu from
2150  * @cpu: CPU to start searching
2151  * @node: NUMA node to order CPUs by distance
2152  *
2153  * Return: cpu, or nr_cpu_ids when nothing found.
2154  */
2155 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2156 {
2157         struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2158         struct cpumask ***hop_masks;
2159         int hop, ret = nr_cpu_ids;
2160
2161         if (node == NUMA_NO_NODE)
2162                 return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2163
2164         rcu_read_lock();
2165
2166         /* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2167         node = numa_nearest_node(node, N_CPU);
2168         k.node = node;
2169
2170         k.masks = rcu_dereference(sched_domains_numa_masks);
2171         if (!k.masks)
2172                 goto unlock;
2173
2174         hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2175         hop = hop_masks - k.masks;
2176
2177         ret = hop ?
2178                 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2179                 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2180 unlock:
2181         rcu_read_unlock();
2182         return ret;
2183 }
2184 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2185
2186 /**
2187  * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2188  *                         @node
2189  * @node: The node to count hops from.
2190  * @hops: Include CPUs up to that many hops away. 0 means local node.
2191  *
2192  * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2193  * @node, an error value otherwise.
2194  *
2195  * Requires rcu_lock to be held. Returned cpumask is only valid within that
2196  * read-side section, copy it if required beyond that.
2197  *
2198  * Note that not all hops are equal in distance; see sched_init_numa() for how
2199  * distances and masks are handled.
2200  * Also note that this is a reflection of sched_domains_numa_masks, which may change
2201  * during the lifetime of the system (offline nodes are taken out of the masks).
2202  */
2203 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2204 {
2205         struct cpumask ***masks;
2206
2207         if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2208                 return ERR_PTR(-EINVAL);
2209
2210         masks = rcu_dereference(sched_domains_numa_masks);
2211         if (!masks)
2212                 return ERR_PTR(-EBUSY);
2213
2214         return masks[hops][node];
2215 }
2216 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2217
2218 #endif /* CONFIG_NUMA */
2219
2220 static int __sdt_alloc(const struct cpumask *cpu_map)
2221 {
2222         struct sched_domain_topology_level *tl;
2223         int j;
2224
2225         for_each_sd_topology(tl) {
2226                 struct sd_data *sdd = &tl->data;
2227
2228                 sdd->sd = alloc_percpu(struct sched_domain *);
2229                 if (!sdd->sd)
2230                         return -ENOMEM;
2231
2232                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2233                 if (!sdd->sds)
2234                         return -ENOMEM;
2235
2236                 sdd->sg = alloc_percpu(struct sched_group *);
2237                 if (!sdd->sg)
2238                         return -ENOMEM;
2239
2240                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2241                 if (!sdd->sgc)
2242                         return -ENOMEM;
2243
2244                 for_each_cpu(j, cpu_map) {
2245                         struct sched_domain *sd;
2246                         struct sched_domain_shared *sds;
2247                         struct sched_group *sg;
2248                         struct sched_group_capacity *sgc;
2249
2250                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2251                                         GFP_KERNEL, cpu_to_node(j));
2252                         if (!sd)
2253                                 return -ENOMEM;
2254
2255                         *per_cpu_ptr(sdd->sd, j) = sd;
2256
2257                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
2258                                         GFP_KERNEL, cpu_to_node(j));
2259                         if (!sds)
2260                                 return -ENOMEM;
2261
2262                         *per_cpu_ptr(sdd->sds, j) = sds;
2263
2264                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2265                                         GFP_KERNEL, cpu_to_node(j));
2266                         if (!sg)
2267                                 return -ENOMEM;
2268
2269                         sg->next = sg;
2270
2271                         *per_cpu_ptr(sdd->sg, j) = sg;
2272
2273                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2274                                         GFP_KERNEL, cpu_to_node(j));
2275                         if (!sgc)
2276                                 return -ENOMEM;
2277
2278 #ifdef CONFIG_SCHED_DEBUG
2279                         sgc->id = j;
2280 #endif
2281
2282                         *per_cpu_ptr(sdd->sgc, j) = sgc;
2283                 }
2284         }
2285
2286         return 0;
2287 }
2288
2289 static void __sdt_free(const struct cpumask *cpu_map)
2290 {
2291         struct sched_domain_topology_level *tl;
2292         int j;
2293
2294         for_each_sd_topology(tl) {
2295                 struct sd_data *sdd = &tl->data;
2296
2297                 for_each_cpu(j, cpu_map) {
2298                         struct sched_domain *sd;
2299
2300                         if (sdd->sd) {
2301                                 sd = *per_cpu_ptr(sdd->sd, j);
2302                                 if (sd && (sd->flags & SD_OVERLAP))
2303                                         free_sched_groups(sd->groups, 0);
2304                                 kfree(*per_cpu_ptr(sdd->sd, j));
2305                         }
2306
2307                         if (sdd->sds)
2308                                 kfree(*per_cpu_ptr(sdd->sds, j));
2309                         if (sdd->sg)
2310                                 kfree(*per_cpu_ptr(sdd->sg, j));
2311                         if (sdd->sgc)
2312                                 kfree(*per_cpu_ptr(sdd->sgc, j));
2313                 }
2314                 free_percpu(sdd->sd);
2315                 sdd->sd = NULL;
2316                 free_percpu(sdd->sds);
2317                 sdd->sds = NULL;
2318                 free_percpu(sdd->sg);
2319                 sdd->sg = NULL;
2320                 free_percpu(sdd->sgc);
2321                 sdd->sgc = NULL;
2322         }
2323 }
2324
2325 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2326                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2327                 struct sched_domain *child, int cpu)
2328 {
2329         struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2330
2331         if (child) {
2332                 sd->level = child->level + 1;
2333                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2334                 child->parent = sd;
2335
2336                 if (!cpumask_subset(sched_domain_span(child),
2337                                     sched_domain_span(sd))) {
2338                         pr_err("BUG: arch topology borken\n");
2339                         pr_err("     the %s domain not a subset of the %s domain\n",
2340                                         child->name, sd->name);
2341                         /* Fixup, ensure @sd has at least @child CPUs. */
2342                         cpumask_or(sched_domain_span(sd),
2343                                    sched_domain_span(sd),
2344                                    sched_domain_span(child));
2345                 }
2346
2347         }
2348         set_domain_attribute(sd, attr);
2349
2350         return sd;
2351 }
2352
2353 /*
2354  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2355  * any two given CPUs at this (non-NUMA) topology level.
2356  */
2357 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2358                               const struct cpumask *cpu_map, int cpu)
2359 {
2360         int i = cpu + 1;
2361
2362         /* NUMA levels are allowed to overlap */
2363         if (tl->flags & SDTL_OVERLAP)
2364                 return true;
2365
2366         /*
2367          * Non-NUMA levels cannot partially overlap - they must be either
2368          * completely equal or completely disjoint. Otherwise we can end up
2369          * breaking the sched_group lists - i.e. a later get_group() pass
2370          * breaks the linking done for an earlier span.
2371          */
2372         for_each_cpu_from(i, cpu_map) {
2373                 /*
2374                  * We should 'and' all those masks with 'cpu_map' to exactly
2375                  * match the topology we're about to build, but that can only
2376                  * remove CPUs, which only lessens our ability to detect
2377                  * overlaps
2378                  */
2379                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2380                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2381                         return false;
2382         }
2383
2384         return true;
2385 }
2386
2387 /*
2388  * Build sched domains for a given set of CPUs and attach the sched domains
2389  * to the individual CPUs
2390  */
2391 static int
2392 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2393 {
2394         enum s_alloc alloc_state = sa_none;
2395         struct sched_domain *sd;
2396         struct s_data d;
2397         struct rq *rq = NULL;
2398         int i, ret = -ENOMEM;
2399         bool has_asym = false;
2400         bool has_cluster = false;
2401
2402         if (WARN_ON(cpumask_empty(cpu_map)))
2403                 goto error;
2404
2405         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2406         if (alloc_state != sa_rootdomain)
2407                 goto error;
2408
2409         /* Set up domains for CPUs specified by the cpu_map: */
2410         for_each_cpu(i, cpu_map) {
2411                 struct sched_domain_topology_level *tl;
2412
2413                 sd = NULL;
2414                 for_each_sd_topology(tl) {
2415
2416                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2417                                 goto error;
2418
2419                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2420
2421                         has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2422
2423                         if (tl == sched_domain_topology)
2424                                 *per_cpu_ptr(d.sd, i) = sd;
2425                         if (tl->flags & SDTL_OVERLAP)
2426                                 sd->flags |= SD_OVERLAP;
2427                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2428                                 break;
2429                 }
2430         }
2431
2432         /* Build the groups for the domains */
2433         for_each_cpu(i, cpu_map) {
2434                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2435                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2436                         if (sd->flags & SD_OVERLAP) {
2437                                 if (build_overlap_sched_groups(sd, i))
2438                                         goto error;
2439                         } else {
2440                                 if (build_sched_groups(sd, i))
2441                                         goto error;
2442                         }
2443                 }
2444         }
2445
2446         /*
2447          * Calculate an allowed NUMA imbalance such that LLCs do not get
2448          * imbalanced.
2449          */
2450         for_each_cpu(i, cpu_map) {
2451                 unsigned int imb = 0;
2452                 unsigned int imb_span = 1;
2453
2454                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2455                         struct sched_domain *child = sd->child;
2456
2457                         if (!(sd->flags & SD_SHARE_LLC) && child &&
2458                             (child->flags & SD_SHARE_LLC)) {
2459                                 struct sched_domain __rcu *top_p;
2460                                 unsigned int nr_llcs;
2461
2462                                 /*
2463                                  * For a single LLC per node, allow an
2464                                  * imbalance up to 12.5% of the node. This is
2465                                  * arbitrary cutoff based two factors -- SMT and
2466                                  * memory channels. For SMT-2, the intent is to
2467                                  * avoid premature sharing of HT resources but
2468                                  * SMT-4 or SMT-8 *may* benefit from a different
2469                                  * cutoff. For memory channels, this is a very
2470                                  * rough estimate of how many channels may be
2471                                  * active and is based on recent CPUs with
2472                                  * many cores.
2473                                  *
2474                                  * For multiple LLCs, allow an imbalance
2475                                  * until multiple tasks would share an LLC
2476                                  * on one node while LLCs on another node
2477                                  * remain idle. This assumes that there are
2478                                  * enough logical CPUs per LLC to avoid SMT
2479                                  * factors and that there is a correlation
2480                                  * between LLCs and memory channels.
2481                                  */
2482                                 nr_llcs = sd->span_weight / child->span_weight;
2483                                 if (nr_llcs == 1)
2484                                         imb = sd->span_weight >> 3;
2485                                 else
2486                                         imb = nr_llcs;
2487                                 imb = max(1U, imb);
2488                                 sd->imb_numa_nr = imb;
2489
2490                                 /* Set span based on the first NUMA domain. */
2491                                 top_p = sd->parent;
2492                                 while (top_p && !(top_p->flags & SD_NUMA)) {
2493                                         top_p = top_p->parent;
2494                                 }
2495                                 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2496                         } else {
2497                                 int factor = max(1U, (sd->span_weight / imb_span));
2498
2499                                 sd->imb_numa_nr = imb * factor;
2500                         }
2501                 }
2502         }
2503
2504         /* Calculate CPU capacity for physical packages and nodes */
2505         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2506                 if (!cpumask_test_cpu(i, cpu_map))
2507                         continue;
2508
2509                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2510                         claim_allocations(i, sd);
2511                         init_sched_groups_capacity(i, sd);
2512                 }
2513         }
2514
2515         /* Attach the domains */
2516         rcu_read_lock();
2517         for_each_cpu(i, cpu_map) {
2518                 rq = cpu_rq(i);
2519                 sd = *per_cpu_ptr(d.sd, i);
2520
2521                 cpu_attach_domain(sd, d.rd, i);
2522
2523                 if (lowest_flag_domain(i, SD_CLUSTER))
2524                         has_cluster = true;
2525         }
2526         rcu_read_unlock();
2527
2528         if (has_asym)
2529                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2530
2531         if (has_cluster)
2532                 static_branch_inc_cpuslocked(&sched_cluster_active);
2533
2534         if (rq && sched_debug_verbose)
2535                 pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
2536
2537         ret = 0;
2538 error:
2539         __free_domain_allocs(&d, alloc_state, cpu_map);
2540
2541         return ret;
2542 }
2543
2544 /* Current sched domains: */
2545 static cpumask_var_t                    *doms_cur;
2546
2547 /* Number of sched domains in 'doms_cur': */
2548 static int                              ndoms_cur;
2549
2550 /* Attributes of custom domains in 'doms_cur' */
2551 static struct sched_domain_attr         *dattr_cur;
2552
2553 /*
2554  * Special case: If a kmalloc() of a doms_cur partition (array of
2555  * cpumask) fails, then fallback to a single sched domain,
2556  * as determined by the single cpumask fallback_doms.
2557  */
2558 static cpumask_var_t                    fallback_doms;
2559
2560 /*
2561  * arch_update_cpu_topology lets virtualized architectures update the
2562  * CPU core maps. It is supposed to return 1 if the topology changed
2563  * or 0 if it stayed the same.
2564  */
2565 int __weak arch_update_cpu_topology(void)
2566 {
2567         return 0;
2568 }
2569
2570 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2571 {
2572         int i;
2573         cpumask_var_t *doms;
2574
2575         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2576         if (!doms)
2577                 return NULL;
2578         for (i = 0; i < ndoms; i++) {
2579                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2580                         free_sched_domains(doms, i);
2581                         return NULL;
2582                 }
2583         }
2584         return doms;
2585 }
2586
2587 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2588 {
2589         unsigned int i;
2590         for (i = 0; i < ndoms; i++)
2591                 free_cpumask_var(doms[i]);
2592         kfree(doms);
2593 }
2594
2595 /*
2596  * Set up scheduler domains and groups.  For now this just excludes isolated
2597  * CPUs, but could be used to exclude other special cases in the future.
2598  */
2599 int __init sched_init_domains(const struct cpumask *cpu_map)
2600 {
2601         int err;
2602
2603         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2604         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2605         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2606
2607         arch_update_cpu_topology();
2608         asym_cpu_capacity_scan();
2609         ndoms_cur = 1;
2610         doms_cur = alloc_sched_domains(ndoms_cur);
2611         if (!doms_cur)
2612                 doms_cur = &fallback_doms;
2613         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2614         err = build_sched_domains(doms_cur[0], NULL);
2615
2616         return err;
2617 }
2618
2619 /*
2620  * Detach sched domains from a group of CPUs specified in cpu_map
2621  * These CPUs will now be attached to the NULL domain
2622  */
2623 static void detach_destroy_domains(const struct cpumask *cpu_map)
2624 {
2625         unsigned int cpu = cpumask_any(cpu_map);
2626         int i;
2627
2628         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2629                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2630
2631         if (static_branch_unlikely(&sched_cluster_active))
2632                 static_branch_dec_cpuslocked(&sched_cluster_active);
2633
2634         rcu_read_lock();
2635         for_each_cpu(i, cpu_map)
2636                 cpu_attach_domain(NULL, &def_root_domain, i);
2637         rcu_read_unlock();
2638 }
2639
2640 /* handle null as "default" */
2641 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2642                         struct sched_domain_attr *new, int idx_new)
2643 {
2644         struct sched_domain_attr tmp;
2645
2646         /* Fast path: */
2647         if (!new && !cur)
2648                 return 1;
2649
2650         tmp = SD_ATTR_INIT;
2651
2652         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2653                         new ? (new + idx_new) : &tmp,
2654                         sizeof(struct sched_domain_attr));
2655 }
2656
2657 /*
2658  * Partition sched domains as specified by the 'ndoms_new'
2659  * cpumasks in the array doms_new[] of cpumasks. This compares
2660  * doms_new[] to the current sched domain partitioning, doms_cur[].
2661  * It destroys each deleted domain and builds each new domain.
2662  *
2663  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2664  * The masks don't intersect (don't overlap.) We should setup one
2665  * sched domain for each mask. CPUs not in any of the cpumasks will
2666  * not be load balanced. If the same cpumask appears both in the
2667  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2668  * it as it is.
2669  *
2670  * The passed in 'doms_new' should be allocated using
2671  * alloc_sched_domains.  This routine takes ownership of it and will
2672  * free_sched_domains it when done with it. If the caller failed the
2673  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2674  * and partition_sched_domains() will fallback to the single partition
2675  * 'fallback_doms', it also forces the domains to be rebuilt.
2676  *
2677  * If doms_new == NULL it will be replaced with cpu_online_mask.
2678  * ndoms_new == 0 is a special case for destroying existing domains,
2679  * and it will not create the default domain.
2680  *
2681  * Call with hotplug lock and sched_domains_mutex held
2682  */
2683 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2684                                     struct sched_domain_attr *dattr_new)
2685 {
2686         bool __maybe_unused has_eas = false;
2687         int i, j, n;
2688         int new_topology;
2689
2690         lockdep_assert_held(&sched_domains_mutex);
2691
2692         /* Let the architecture update CPU core mappings: */
2693         new_topology = arch_update_cpu_topology();
2694         /* Trigger rebuilding CPU capacity asymmetry data */
2695         if (new_topology)
2696                 asym_cpu_capacity_scan();
2697
2698         if (!doms_new) {
2699                 WARN_ON_ONCE(dattr_new);
2700                 n = 0;
2701                 doms_new = alloc_sched_domains(1);
2702                 if (doms_new) {
2703                         n = 1;
2704                         cpumask_and(doms_new[0], cpu_active_mask,
2705                                     housekeeping_cpumask(HK_TYPE_DOMAIN));
2706                 }
2707         } else {
2708                 n = ndoms_new;
2709         }
2710
2711         /* Destroy deleted domains: */
2712         for (i = 0; i < ndoms_cur; i++) {
2713                 for (j = 0; j < n && !new_topology; j++) {
2714                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2715                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2716                                 struct root_domain *rd;
2717
2718                                 /*
2719                                  * This domain won't be destroyed and as such
2720                                  * its dl_bw->total_bw needs to be cleared.
2721                                  * Tasks contribution will be then recomputed
2722                                  * in function dl_update_tasks_root_domain(),
2723                                  * dl_servers contribution in function
2724                                  * dl_restore_server_root_domain().
2725                                  */
2726                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2727                                 dl_clear_root_domain(rd);
2728                                 goto match1;
2729                         }
2730                 }
2731                 /* No match - a current sched domain not in new doms_new[] */
2732                 detach_destroy_domains(doms_cur[i]);
2733 match1:
2734                 ;
2735         }
2736
2737         n = ndoms_cur;
2738         if (!doms_new) {
2739                 n = 0;
2740                 doms_new = &fallback_doms;
2741                 cpumask_and(doms_new[0], cpu_active_mask,
2742                             housekeeping_cpumask(HK_TYPE_DOMAIN));
2743         }
2744
2745         /* Build new domains: */
2746         for (i = 0; i < ndoms_new; i++) {
2747                 for (j = 0; j < n && !new_topology; j++) {
2748                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2749                             dattrs_equal(dattr_new, i, dattr_cur, j))
2750                                 goto match2;
2751                 }
2752                 /* No match - add a new doms_new */
2753                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2754 match2:
2755                 ;
2756         }
2757
2758 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2759         /* Build perf domains: */
2760         for (i = 0; i < ndoms_new; i++) {
2761                 for (j = 0; j < n && !sched_energy_update; j++) {
2762                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2763                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2764                                 has_eas = true;
2765                                 goto match3;
2766                         }
2767                 }
2768                 /* No match - add perf domains for a new rd */
2769                 has_eas |= build_perf_domains(doms_new[i]);
2770 match3:
2771                 ;
2772         }
2773         sched_energy_set(has_eas);
2774 #endif
2775
2776         /* Remember the new sched domains: */
2777         if (doms_cur != &fallback_doms)
2778                 free_sched_domains(doms_cur, ndoms_cur);
2779
2780         kfree(dattr_cur);
2781         doms_cur = doms_new;
2782         dattr_cur = dattr_new;
2783         ndoms_cur = ndoms_new;
2784
2785         update_sched_domain_debugfs();
2786 }
2787
2788 /*
2789  * Call with hotplug lock held
2790  */
2791 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2792                              struct sched_domain_attr *dattr_new)
2793 {
2794         mutex_lock(&sched_domains_mutex);
2795         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2796         mutex_unlock(&sched_domains_mutex);
2797 }