kernel/sched/topology.c

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