kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5 #include "sched.h"
   6
   7 DEFINE_MUTEX(sched_domains_mutex);
   8
   9 /* Protected by sched_domains_mutex: */
  10 static cpumask_var_t sched_domains_tmpmask;
  11 static cpumask_var_t sched_domains_tmpmask2;
  12
  13 #ifdef CONFIG_SCHED_DEBUG
  14
  15 static int __init sched_debug_setup(char *str)
  16 {
  17         sched_debug_enabled = true;
  18
  19         return 0;
  20 }
  21 early_param("sched_debug", sched_debug_setup);
  22
  23 static inline bool sched_debug(void)
  24 {
  25         return sched_debug_enabled;
  26 }
  27
  28 #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_enabled)
 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_enabled 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_lock_irqsave(&rq->lock, 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_unlock_irqrestore(&rq->lock, 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);
 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         /*
 727          * FIXME: Diameter >=3 is misrepresented.
 728          *
 729          * Smallest diameter=3 topology is:
 730          *
 731          *   node   0   1   2   3
 732          *     0:  10  20  30  40
 733          *     1:  20  10  20  30
 734          *     2:  30  20  10  20
 735          *     3:  40  30  20  10
 736          *
 737          *   0 --- 1 --- 2 --- 3
 738          *
 739          * NUMA-3       0-3             N/A             N/A             0-3
 740          *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 741          *
 742          * NUMA-2       0-2             0-3             0-3             1-3
 743          *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
 744          *
 745          * NUMA-1       0-1             0-2             1-3             2-3
 746          *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
 747          *
 748          * NUMA-0       0               1               2               3
 749          *
 750          * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
 751          * group span isn't a subset of the domain span.
 752          */
 753         WARN_ONCE(numa_distance > 2, "Shortest NUMA path spans too many nodes\n");
 754
 755         sched_domain_debug(sd, cpu);
 756
 757         rq_attach_root(rq, rd);
 758         tmp = rq->sd;
 759         rcu_assign_pointer(rq->sd, sd);
 760         dirty_sched_domain_sysctl(cpu);
 761         destroy_sched_domains(tmp);
 762
 763         update_top_cache_domain(cpu);
 764 }
 765
 766 struct s_data {
 767         struct sched_domain * __percpu *sd;
 768         struct root_domain      *rd;
 769 };
 770
 771 enum s_alloc {
 772         sa_rootdomain,
 773         sa_sd,
 774         sa_sd_storage,
 775         sa_none,
 776 };
 777
 778 /*
 779  * Return the canonical balance CPU for this group, this is the first CPU
 780  * of this group that's also in the balance mask.
 781  *
 782  * The balance mask are all those CPUs that could actually end up at this
 783  * group. See build_balance_mask().
 784  *
 785  * Also see should_we_balance().
 786  */
 787 int group_balance_cpu(struct sched_group *sg)
 788 {
 789         return cpumask_first(group_balance_mask(sg));
 790 }
 791
 792
 793 /*
 794  * NUMA topology (first read the regular topology blurb below)
 795  *
 796  * Given a node-distance table, for example:
 797  *
 798  *   node   0   1   2   3
 799  *     0:  10  20  30  20
 800  *     1:  20  10  20  30
 801  *     2:  30  20  10  20
 802  *     3:  20  30  20  10
 803  *
 804  * which represents a 4 node ring topology like:
 805  *
 806  *   0 ----- 1
 807  *   |       |
 808  *   |       |
 809  *   |       |
 810  *   3 ----- 2
 811  *
 812  * We want to construct domains and groups to represent this. The way we go
 813  * about doing this is to build the domains on 'hops'. For each NUMA level we
 814  * construct the mask of all nodes reachable in @level hops.
 815  *
 816  * For the above NUMA topology that gives 3 levels:
 817  *
 818  * NUMA-2       0-3             0-3             0-3             0-3
 819  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 820  *
 821  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 822  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 823  *
 824  * NUMA-0       0               1               2               3
 825  *
 826  *
 827  * As can be seen; things don't nicely line up as with the regular topology.
 828  * When we iterate a domain in child domain chunks some nodes can be
 829  * represented multiple times -- hence the "overlap" naming for this part of
 830  * the topology.
 831  *
 832  * In order to minimize this overlap, we only build enough groups to cover the
 833  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 834  *
 835  * Because:
 836  *
 837  *  - the first group of each domain is its child domain; this
 838  *    gets us the first 0-1,3
 839  *  - the only uncovered node is 2, who's child domain is 1-3.
 840  *
 841  * However, because of the overlap, computing a unique CPU for each group is
 842  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 843  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 844  * end up at those groups (they would end up in group: 0-1,3).
 845  *
 846  * To correct this we have to introduce the group balance mask. This mask
 847  * will contain those CPUs in the group that can reach this group given the
 848  * (child) domain tree.
 849  *
 850  * With this we can once again compute balance_cpu and sched_group_capacity
 851  * relations.
 852  *
 853  * XXX include words on how balance_cpu is unique and therefore can be
 854  * used for sched_group_capacity links.
 855  *
 856  *
 857  * Another 'interesting' topology is:
 858  *
 859  *   node   0   1   2   3
 860  *     0:  10  20  20  30
 861  *     1:  20  10  20  20
 862  *     2:  20  20  10  20
 863  *     3:  30  20  20  10
 864  *
 865  * Which looks a little like:
 866  *
 867  *   0 ----- 1
 868  *   |     / |
 869  *   |   /   |
 870  *   | /     |
 871  *   2 ----- 3
 872  *
 873  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 874  * are not.
 875  *
 876  * This leads to a few particularly weird cases where the sched_domain's are
 877  * not of the same number for each CPU. Consider:
 878  *
 879  * NUMA-2       0-3                                             0-3
 880  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 881  *
 882  * NUMA-1       0-2             0-3             0-3             1-3
 883  *
 884  * NUMA-0       0               1               2               3
 885  *
 886  */
 887
 888
 889 /*
 890  * Build the balance mask; it contains only those CPUs that can arrive at this
 891  * group and should be considered to continue balancing.
 892  *
 893  * We do this during the group creation pass, therefore the group information
 894  * isn't complete yet, however since each group represents a (child) domain we
 895  * can fully construct this using the sched_domain bits (which are already
 896  * complete).
 897  */
 898 static void
 899 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 900 {
 901         const struct cpumask *sg_span = sched_group_span(sg);
 902         struct sd_data *sdd = sd->private;
 903         struct sched_domain *sibling;
 904         int i;
 905
 906         cpumask_clear(mask);
 907
 908         for_each_cpu(i, sg_span) {
 909                 sibling = *per_cpu_ptr(sdd->sd, i);
 910
 911                 /*
 912                  * Can happen in the asymmetric case, where these siblings are
 913                  * unused. The mask will not be empty because those CPUs that
 914                  * do have the top domain _should_ span the domain.
 915                  */
 916                 if (!sibling->child)
 917                         continue;
 918
 919                 /* If we would not end up here, we can't continue from here */
 920                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 921                         continue;
 922
 923                 cpumask_set_cpu(i, mask);
 924         }
 925
 926         /* We must not have empty masks here */
 927         WARN_ON_ONCE(cpumask_empty(mask));
 928 }
 929
 930 /*
 931  * XXX: This creates per-node group entries; since the load-balancer will
 932  * immediately access remote memory to construct this group's load-balance
 933  * statistics having the groups node local is of dubious benefit.
 934  */
 935 static struct sched_group *
 936 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 937 {
 938         struct sched_group *sg;
 939         struct cpumask *sg_span;
 940
 941         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 942                         GFP_KERNEL, cpu_to_node(cpu));
 943
 944         if (!sg)
 945                 return NULL;
 946
 947         sg_span = sched_group_span(sg);
 948         if (sd->child)
 949                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 950         else
 951                 cpumask_copy(sg_span, sched_domain_span(sd));
 952
 953         atomic_inc(&sg->ref);
 954         return sg;
 955 }
 956
 957 static void init_overlap_sched_group(struct sched_domain *sd,
 958                                      struct sched_group *sg)
 959 {
 960         struct cpumask *mask = sched_domains_tmpmask2;
 961         struct sd_data *sdd = sd->private;
 962         struct cpumask *sg_span;
 963         int cpu;
 964
 965         build_balance_mask(sd, sg, mask);
 966         cpu = cpumask_first_and(sched_group_span(sg), mask);
 967
 968         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 969         if (atomic_inc_return(&sg->sgc->ref) == 1)
 970                 cpumask_copy(group_balance_mask(sg), mask);
 971         else
 972                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 973
 974         /*
 975          * Initialize sgc->capacity such that even if we mess up the
 976          * domains and no possible iteration will get us here, we won't
 977          * die on a /0 trap.
 978          */
 979         sg_span = sched_group_span(sg);
 980         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 981         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 982         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 983 }
 984
 985 static int
 986 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
 987 {
 988         struct sched_group *first = NULL, *last = NULL, *sg;
 989         const struct cpumask *span = sched_domain_span(sd);
 990         struct cpumask *covered = sched_domains_tmpmask;
 991         struct sd_data *sdd = sd->private;
 992         struct sched_domain *sibling;
 993         int i;
 994
 995         cpumask_clear(covered);
 996
 997         for_each_cpu_wrap(i, span, cpu) {
 998                 struct cpumask *sg_span;
 999
1000                 if (cpumask_test_cpu(i, covered))
1001                         continue;
1002
1003                 sibling = *per_cpu_ptr(sdd->sd, i);
1004
1005                 /*
1006                  * Asymmetric node setups can result in situations where the
1007                  * domain tree is of unequal depth, make sure to skip domains
1008                  * that already cover the entire range.
1009                  *
1010                  * In that case build_sched_domains() will have terminated the
1011                  * iteration early and our sibling sd spans will be empty.
1012                  * Domains should always include the CPU they're built on, so
1013                  * check that.
1014                  */
1015                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1016                         continue;
1017
1018                 sg = build_group_from_child_sched_domain(sibling, cpu);
1019                 if (!sg)
1020                         goto fail;
1021
1022                 sg_span = sched_group_span(sg);
1023                 cpumask_or(covered, covered, sg_span);
1024
1025                 init_overlap_sched_group(sd, sg);
1026
1027                 if (!first)
1028                         first = sg;
1029                 if (last)
1030                         last->next = sg;
1031                 last = sg;
1032                 last->next = first;
1033         }
1034         sd->groups = first;
1035
1036         return 0;
1037
1038 fail:
1039         free_sched_groups(first, 0);
1040
1041         return -ENOMEM;
1042 }
1043
1044
1045 /*
1046  * Package topology (also see the load-balance blurb in fair.c)
1047  *
1048  * The scheduler builds a tree structure to represent a number of important
1049  * topology features. By default (default_topology[]) these include:
1050  *
1051  *  - Simultaneous multithreading (SMT)
1052  *  - Multi-Core Cache (MC)
1053  *  - Package (DIE)
1054  *
1055  * Where the last one more or less denotes everything up to a NUMA node.
1056  *
1057  * The tree consists of 3 primary data structures:
1058  *
1059  *      sched_domain -> sched_group -> sched_group_capacity
1060  *          ^ ^             ^ ^
1061  *          `-'             `-'
1062  *
1063  * The sched_domains are per-CPU and have a two way link (parent & child) and
1064  * denote the ever growing mask of CPUs belonging to that level of topology.
1065  *
1066  * Each sched_domain has a circular (double) linked list of sched_group's, each
1067  * denoting the domains of the level below (or individual CPUs in case of the
1068  * first domain level). The sched_group linked by a sched_domain includes the
1069  * CPU of that sched_domain [*].
1070  *
1071  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1072  *
1073  * CPU   0   1   2   3   4   5   6   7
1074  *
1075  * DIE  [                             ]
1076  * MC   [             ] [             ]
1077  * SMT  [     ] [     ] [     ] [     ]
1078  *
1079  *  - or -
1080  *
1081  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1082  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1083  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1084  *
1085  * CPU   0   1   2   3   4   5   6   7
1086  *
1087  * One way to think about it is: sched_domain moves you up and down among these
1088  * topology levels, while sched_group moves you sideways through it, at child
1089  * domain granularity.
1090  *
1091  * sched_group_capacity ensures each unique sched_group has shared storage.
1092  *
1093  * There are two related construction problems, both require a CPU that
1094  * uniquely identify each group (for a given domain):
1095  *
1096  *  - The first is the balance_cpu (see should_we_balance() and the
1097  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1098  *    continue balancing at a higher domain.
1099  *
1100  *  - The second is the sched_group_capacity; we want all identical groups
1101  *    to share a single sched_group_capacity.
1102  *
1103  * Since these topologies are exclusive by construction. That is, its
1104  * impossible for an SMT thread to belong to multiple cores, and cores to
1105  * be part of multiple caches. There is a very clear and unique location
1106  * for each CPU in the hierarchy.
1107  *
1108  * Therefore computing a unique CPU for each group is trivial (the iteration
1109  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1110  * group), we can simply pick the first CPU in each group.
1111  *
1112  *
1113  * [*] in other words, the first group of each domain is its child domain.
1114  */
1115
1116 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1117 {
1118         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1119         struct sched_domain *child = sd->child;
1120         struct sched_group *sg;
1121         bool already_visited;
1122
1123         if (child)
1124                 cpu = cpumask_first(sched_domain_span(child));
1125
1126         sg = *per_cpu_ptr(sdd->sg, cpu);
1127         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1128
1129         /* Increase refcounts for claim_allocations: */
1130         already_visited = atomic_inc_return(&sg->ref) > 1;
1131         /* sgc visits should follow a similar trend as sg */
1132         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1133
1134         /* If we have already visited that group, it's already initialized. */
1135         if (already_visited)
1136                 return sg;
1137
1138         if (child) {
1139                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1140                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1141         } else {
1142                 cpumask_set_cpu(cpu, sched_group_span(sg));
1143                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1144         }
1145
1146         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1147         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1148         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1149
1150         return sg;
1151 }
1152
1153 /*
1154  * build_sched_groups will build a circular linked list of the groups
1155  * covered by the given span, will set each group's ->cpumask correctly,
1156  * and will initialize their ->sgc.
1157  *
1158  * Assumes the sched_domain tree is fully constructed
1159  */
1160 static int
1161 build_sched_groups(struct sched_domain *sd, int cpu)
1162 {
1163         struct sched_group *first = NULL, *last = NULL;
1164         struct sd_data *sdd = sd->private;
1165         const struct cpumask *span = sched_domain_span(sd);
1166         struct cpumask *covered;
1167         int i;
1168
1169         lockdep_assert_held(&sched_domains_mutex);
1170         covered = sched_domains_tmpmask;
1171
1172         cpumask_clear(covered);
1173
1174         for_each_cpu_wrap(i, span, cpu) {
1175                 struct sched_group *sg;
1176
1177                 if (cpumask_test_cpu(i, covered))
1178                         continue;
1179
1180                 sg = get_group(i, sdd);
1181
1182                 cpumask_or(covered, covered, sched_group_span(sg));
1183
1184                 if (!first)
1185                         first = sg;
1186                 if (last)
1187                         last->next = sg;
1188                 last = sg;
1189         }
1190         last->next = first;
1191         sd->groups = first;
1192
1193         return 0;
1194 }
1195
1196 /*
1197  * Initialize sched groups cpu_capacity.
1198  *
1199  * cpu_capacity indicates the capacity of sched group, which is used while
1200  * distributing the load between different sched groups in a sched domain.
1201  * Typically cpu_capacity for all the groups in a sched domain will be same
1202  * unless there are asymmetries in the topology. If there are asymmetries,
1203  * group having more cpu_capacity will pickup more load compared to the
1204  * group having less cpu_capacity.
1205  */
1206 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1207 {
1208         struct sched_group *sg = sd->groups;
1209
1210         WARN_ON(!sg);
1211
1212         do {
1213                 int cpu, max_cpu = -1;
1214
1215                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1216
1217                 if (!(sd->flags & SD_ASYM_PACKING))
1218                         goto next;
1219
1220                 for_each_cpu(cpu, sched_group_span(sg)) {
1221                         if (max_cpu < 0)
1222                                 max_cpu = cpu;
1223                         else if (sched_asym_prefer(cpu, max_cpu))
1224                                 max_cpu = cpu;
1225                 }
1226                 sg->asym_prefer_cpu = max_cpu;
1227
1228 next:
1229                 sg = sg->next;
1230         } while (sg != sd->groups);
1231
1232         if (cpu != group_balance_cpu(sg))
1233                 return;
1234
1235         update_group_capacity(sd, cpu);
1236 }
1237
1238 /*
1239  * Initializers for schedule domains
1240  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1241  */
1242
1243 static int default_relax_domain_level = -1;
1244 int sched_domain_level_max;
1245
1246 static int __init setup_relax_domain_level(char *str)
1247 {
1248         if (kstrtoint(str, 0, &default_relax_domain_level))
1249                 pr_warn("Unable to set relax_domain_level\n");
1250
1251         return 1;
1252 }
1253 __setup("relax_domain_level=", setup_relax_domain_level);
1254
1255 static void set_domain_attribute(struct sched_domain *sd,
1256                                  struct sched_domain_attr *attr)
1257 {
1258         int request;
1259
1260         if (!attr || attr->relax_domain_level < 0) {
1261                 if (default_relax_domain_level < 0)
1262                         return;
1263                 request = default_relax_domain_level;
1264         } else
1265                 request = attr->relax_domain_level;
1266
1267         if (sd->level > request) {
1268                 /* Turn off idle balance on this domain: */
1269                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1270         }
1271 }
1272
1273 static void __sdt_free(const struct cpumask *cpu_map);
1274 static int __sdt_alloc(const struct cpumask *cpu_map);
1275
1276 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1277                                  const struct cpumask *cpu_map)
1278 {
1279         switch (what) {
1280         case sa_rootdomain:
1281                 if (!atomic_read(&d->rd->refcount))
1282                         free_rootdomain(&d->rd->rcu);
1283                 fallthrough;
1284         case sa_sd:
1285                 free_percpu(d->sd);
1286                 fallthrough;
1287         case sa_sd_storage:
1288                 __sdt_free(cpu_map);
1289                 fallthrough;
1290         case sa_none:
1291                 break;
1292         }
1293 }
1294
1295 static enum s_alloc
1296 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1297 {
1298         memset(d, 0, sizeof(*d));
1299
1300         if (__sdt_alloc(cpu_map))
1301                 return sa_sd_storage;
1302         d->sd = alloc_percpu(struct sched_domain *);
1303         if (!d->sd)
1304                 return sa_sd_storage;
1305         d->rd = alloc_rootdomain();
1306         if (!d->rd)
1307                 return sa_sd;
1308
1309         return sa_rootdomain;
1310 }
1311
1312 /*
1313  * NULL the sd_data elements we've used to build the sched_domain and
1314  * sched_group structure so that the subsequent __free_domain_allocs()
1315  * will not free the data we're using.
1316  */
1317 static void claim_allocations(int cpu, struct sched_domain *sd)
1318 {
1319         struct sd_data *sdd = sd->private;
1320
1321         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1322         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1323
1324         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1325                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1326
1327         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1328                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1329
1330         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1331                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1332 }
1333
1334 #ifdef CONFIG_NUMA
1335 enum numa_topology_type sched_numa_topology_type;
1336
1337 static int                      sched_domains_numa_levels;
1338 static int                      sched_domains_curr_level;
1339
1340 int                             sched_max_numa_distance;
1341 static int                      *sched_domains_numa_distance;
1342 static struct cpumask           ***sched_domains_numa_masks;
1343 int __read_mostly               node_reclaim_distance = RECLAIM_DISTANCE;
1344 #endif
1345
1346 /*
1347  * SD_flags allowed in topology descriptions.
1348  *
1349  * These flags are purely descriptive of the topology and do not prescribe
1350  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1351  * function:
1352  *
1353  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1354  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1355  *   SD_NUMA                - describes NUMA topologies
1356  *
1357  * Odd one out, which beside describing the topology has a quirk also
1358  * prescribes the desired behaviour that goes along with it:
1359  *
1360  *   SD_ASYM_PACKING        - describes SMT quirks
1361  */
1362 #define TOPOLOGY_SD_FLAGS               \
1363         (SD_SHARE_CPUCAPACITY   |       \
1364          SD_SHARE_PKG_RESOURCES |       \
1365          SD_NUMA                |       \
1366          SD_ASYM_PACKING)
1367
1368 static struct sched_domain *
1369 sd_init(struct sched_domain_topology_level *tl,
1370         const struct cpumask *cpu_map,
1371         struct sched_domain *child, int dflags, int cpu)
1372 {
1373         struct sd_data *sdd = &tl->data;
1374         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1375         int sd_id, sd_weight, sd_flags = 0;
1376
1377 #ifdef CONFIG_NUMA
1378         /*
1379          * Ugly hack to pass state to sd_numa_mask()...
1380          */
1381         sched_domains_curr_level = tl->numa_level;
1382 #endif
1383
1384         sd_weight = cpumask_weight(tl->mask(cpu));
1385
1386         if (tl->sd_flags)
1387                 sd_flags = (*tl->sd_flags)();
1388         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1389                         "wrong sd_flags in topology description\n"))
1390                 sd_flags &= TOPOLOGY_SD_FLAGS;
1391
1392         /* Apply detected topology flags */
1393         sd_flags |= dflags;
1394
1395         *sd = (struct sched_domain){
1396                 .min_interval           = sd_weight,
1397                 .max_interval           = 2*sd_weight,
1398                 .busy_factor            = 16,
1399                 .imbalance_pct          = 117,
1400
1401                 .cache_nice_tries       = 0,
1402
1403                 .flags                  = 1*SD_BALANCE_NEWIDLE
1404                                         | 1*SD_BALANCE_EXEC
1405                                         | 1*SD_BALANCE_FORK
1406                                         | 0*SD_BALANCE_WAKE
1407                                         | 1*SD_WAKE_AFFINE
1408                                         | 0*SD_SHARE_CPUCAPACITY
1409                                         | 0*SD_SHARE_PKG_RESOURCES
1410                                         | 0*SD_SERIALIZE
1411                                         | 1*SD_PREFER_SIBLING
1412                                         | 0*SD_NUMA
1413                                         | sd_flags
1414                                         ,
1415
1416                 .last_balance           = jiffies,
1417                 .balance_interval       = sd_weight,
1418                 .max_newidle_lb_cost    = 0,
1419                 .next_decay_max_lb_cost = jiffies,
1420                 .child                  = child,
1421 #ifdef CONFIG_SCHED_DEBUG
1422                 .name                   = tl->name,
1423 #endif
1424         };
1425
1426         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1427         sd_id = cpumask_first(sched_domain_span(sd));
1428
1429         /*
1430          * Convert topological properties into behaviour.
1431          */
1432
1433         /* Don't attempt to spread across CPUs of different capacities. */
1434         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1435                 sd->child->flags &= ~SD_PREFER_SIBLING;
1436
1437         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1438                 sd->imbalance_pct = 110;
1439
1440         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1441                 sd->imbalance_pct = 117;
1442                 sd->cache_nice_tries = 1;
1443
1444 #ifdef CONFIG_NUMA
1445         } else if (sd->flags & SD_NUMA) {
1446                 sd->cache_nice_tries = 2;
1447
1448                 sd->flags &= ~SD_PREFER_SIBLING;
1449                 sd->flags |= SD_SERIALIZE;
1450                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1451                         sd->flags &= ~(SD_BALANCE_EXEC |
1452                                        SD_BALANCE_FORK |
1453                                        SD_WAKE_AFFINE);
1454                 }
1455
1456 #endif
1457         } else {
1458                 sd->cache_nice_tries = 1;
1459         }
1460
1461         /*
1462          * For all levels sharing cache; connect a sched_domain_shared
1463          * instance.
1464          */
1465         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1466                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1467                 atomic_inc(&sd->shared->ref);
1468                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1469         }
1470
1471         sd->private = sdd;
1472
1473         return sd;
1474 }
1475
1476 /*
1477  * Topology list, bottom-up.
1478  */
1479 static struct sched_domain_topology_level default_topology[] = {
1480 #ifdef CONFIG_SCHED_SMT
1481         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1482 #endif
1483 #ifdef CONFIG_SCHED_MC
1484         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1485 #endif
1486         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1487         { NULL, },
1488 };
1489
1490 static struct sched_domain_topology_level *sched_domain_topology =
1491         default_topology;
1492
1493 #define for_each_sd_topology(tl)                        \
1494         for (tl = sched_domain_topology; tl->mask; tl++)
1495
1496 void set_sched_topology(struct sched_domain_topology_level *tl)
1497 {
1498         if (WARN_ON_ONCE(sched_smp_initialized))
1499                 return;
1500
1501         sched_domain_topology = tl;
1502 }
1503
1504 #ifdef CONFIG_NUMA
1505
1506 static const struct cpumask *sd_numa_mask(int cpu)
1507 {
1508         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1509 }
1510
1511 static void sched_numa_warn(const char *str)
1512 {
1513         static int done = false;
1514         int i,j;
1515
1516         if (done)
1517                 return;
1518
1519         done = true;
1520
1521         printk(KERN_WARNING "ERROR: %s\n\n", str);
1522
1523         for (i = 0; i < nr_node_ids; i++) {
1524                 printk(KERN_WARNING "  ");
1525                 for (j = 0; j < nr_node_ids; j++)
1526                         printk(KERN_CONT "%02d ", node_distance(i,j));
1527                 printk(KERN_CONT "\n");
1528         }
1529         printk(KERN_WARNING "\n");
1530 }
1531
1532 bool find_numa_distance(int distance)
1533 {
1534         int i;
1535
1536         if (distance == node_distance(0, 0))
1537                 return true;
1538
1539         for (i = 0; i < sched_domains_numa_levels; i++) {
1540                 if (sched_domains_numa_distance[i] == distance)
1541                         return true;
1542         }
1543
1544         return false;
1545 }
1546
1547 /*
1548  * A system can have three types of NUMA topology:
1549  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1550  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1551  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1552  *
1553  * The difference between a glueless mesh topology and a backplane
1554  * topology lies in whether communication between not directly
1555  * connected nodes goes through intermediary nodes (where programs
1556  * could run), or through backplane controllers. This affects
1557  * placement of programs.
1558  *
1559  * The type of topology can be discerned with the following tests:
1560  * - If the maximum distance between any nodes is 1 hop, the system
1561  *   is directly connected.
1562  * - If for two nodes A and B, located N > 1 hops away from each other,
1563  *   there is an intermediary node C, which is < N hops away from both
1564  *   nodes A and B, the system is a glueless mesh.
1565  */
1566 static void init_numa_topology_type(void)
1567 {
1568         int a, b, c, n;
1569
1570         n = sched_max_numa_distance;
1571
1572         if (sched_domains_numa_levels <= 2) {
1573                 sched_numa_topology_type = NUMA_DIRECT;
1574                 return;
1575         }
1576
1577         for_each_online_node(a) {
1578                 for_each_online_node(b) {
1579                         /* Find two nodes furthest removed from each other. */
1580                         if (node_distance(a, b) < n)
1581                                 continue;
1582
1583                         /* Is there an intermediary node between a and b? */
1584                         for_each_online_node(c) {
1585                                 if (node_distance(a, c) < n &&
1586                                     node_distance(b, c) < n) {
1587                                         sched_numa_topology_type =
1588                                                         NUMA_GLUELESS_MESH;
1589                                         return;
1590                                 }
1591                         }
1592
1593                         sched_numa_topology_type = NUMA_BACKPLANE;
1594                         return;
1595                 }
1596         }
1597 }
1598
1599 void sched_init_numa(void)
1600 {
1601         int next_distance, curr_distance = node_distance(0, 0);
1602         struct sched_domain_topology_level *tl;
1603         int level = 0;
1604         int i, j, k;
1605
1606         sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1607         if (!sched_domains_numa_distance)
1608                 return;
1609
1610         /* Includes NUMA identity node at level 0. */
1611         sched_domains_numa_distance[level++] = curr_distance;
1612         sched_domains_numa_levels = level;
1613
1614         /*
1615          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1616          * unique distances in the node_distance() table.
1617          *
1618          * Assumes node_distance(0,j) includes all distances in
1619          * node_distance(i,j) in order to avoid cubic time.
1620          */
1621         next_distance = curr_distance;
1622         for (i = 0; i < nr_node_ids; i++) {
1623                 for (j = 0; j < nr_node_ids; j++) {
1624                         for (k = 0; k < nr_node_ids; k++) {
1625                                 int distance = node_distance(i, k);
1626
1627                                 if (distance > curr_distance &&
1628                                     (distance < next_distance ||
1629                                      next_distance == curr_distance))
1630                                         next_distance = distance;
1631
1632                                 /*
1633                                  * While not a strong assumption it would be nice to know
1634                                  * about cases where if node A is connected to B, B is not
1635                                  * equally connected to A.
1636                                  */
1637                                 if (sched_debug() && node_distance(k, i) != distance)
1638                                         sched_numa_warn("Node-distance not symmetric");
1639
1640                                 if (sched_debug() && i && !find_numa_distance(distance))
1641                                         sched_numa_warn("Node-0 not representative");
1642                         }
1643                         if (next_distance != curr_distance) {
1644                                 sched_domains_numa_distance[level++] = next_distance;
1645                                 sched_domains_numa_levels = level;
1646                                 curr_distance = next_distance;
1647                         } else break;
1648                 }
1649
1650                 /*
1651                  * In case of sched_debug() we verify the above assumption.
1652                  */
1653                 if (!sched_debug())
1654                         break;
1655         }
1656
1657         /*
1658          * 'level' contains the number of unique distances
1659          *
1660          * The sched_domains_numa_distance[] array includes the actual distance
1661          * numbers.
1662          */
1663
1664         /*
1665          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1666          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1667          * the array will contain less then 'level' members. This could be
1668          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1669          * in other functions.
1670          *
1671          * We reset it to 'level' at the end of this function.
1672          */
1673         sched_domains_numa_levels = 0;
1674
1675         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1676         if (!sched_domains_numa_masks)
1677                 return;
1678
1679         /*
1680          * Now for each level, construct a mask per node which contains all
1681          * CPUs of nodes that are that many hops away from us.
1682          */
1683         for (i = 0; i < level; i++) {
1684                 sched_domains_numa_masks[i] =
1685                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1686                 if (!sched_domains_numa_masks[i])
1687                         return;
1688
1689                 for (j = 0; j < nr_node_ids; j++) {
1690                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1691                         if (!mask)
1692                                 return;
1693
1694                         sched_domains_numa_masks[i][j] = mask;
1695
1696                         for_each_node(k) {
1697                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1698                                         continue;
1699
1700                                 cpumask_or(mask, mask, cpumask_of_node(k));
1701                         }
1702                 }
1703         }
1704
1705         /* Compute default topology size */
1706         for (i = 0; sched_domain_topology[i].mask; i++);
1707
1708         tl = kzalloc((i + level + 1) *
1709                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1710         if (!tl)
1711                 return;
1712
1713         /*
1714          * Copy the default topology bits..
1715          */
1716         for (i = 0; sched_domain_topology[i].mask; i++)
1717                 tl[i] = sched_domain_topology[i];
1718
1719         /*
1720          * Add the NUMA identity distance, aka single NODE.
1721          */
1722         tl[i++] = (struct sched_domain_topology_level){
1723                 .mask = sd_numa_mask,
1724                 .numa_level = 0,
1725                 SD_INIT_NAME(NODE)
1726         };
1727
1728         /*
1729          * .. and append 'j' levels of NUMA goodness.
1730          */
1731         for (j = 1; j < level; i++, j++) {
1732                 tl[i] = (struct sched_domain_topology_level){
1733                         .mask = sd_numa_mask,
1734                         .sd_flags = cpu_numa_flags,
1735                         .flags = SDTL_OVERLAP,
1736                         .numa_level = j,
1737                         SD_INIT_NAME(NUMA)
1738                 };
1739         }
1740
1741         sched_domain_topology = tl;
1742
1743         sched_domains_numa_levels = level;
1744         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1745
1746         init_numa_topology_type();
1747 }
1748
1749 void sched_domains_numa_masks_set(unsigned int cpu)
1750 {
1751         int node = cpu_to_node(cpu);
1752         int i, j;
1753
1754         for (i = 0; i < sched_domains_numa_levels; i++) {
1755                 for (j = 0; j < nr_node_ids; j++) {
1756                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
1757                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1758                 }
1759         }
1760 }
1761
1762 void sched_domains_numa_masks_clear(unsigned int cpu)
1763 {
1764         int i, j;
1765
1766         for (i = 0; i < sched_domains_numa_levels; i++) {
1767                 for (j = 0; j < nr_node_ids; j++)
1768                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1769         }
1770 }
1771
1772 /*
1773  * sched_numa_find_closest() - given the NUMA topology, find the cpu
1774  *                             closest to @cpu from @cpumask.
1775  * cpumask: cpumask to find a cpu from
1776  * cpu: cpu to be close to
1777  *
1778  * returns: cpu, or nr_cpu_ids when nothing found.
1779  */
1780 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1781 {
1782         int i, j = cpu_to_node(cpu);
1783
1784         for (i = 0; i < sched_domains_numa_levels; i++) {
1785                 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1786                 if (cpu < nr_cpu_ids)
1787                         return cpu;
1788         }
1789         return nr_cpu_ids;
1790 }
1791
1792 #endif /* CONFIG_NUMA */
1793
1794 static int __sdt_alloc(const struct cpumask *cpu_map)
1795 {
1796         struct sched_domain_topology_level *tl;
1797         int j;
1798
1799         for_each_sd_topology(tl) {
1800                 struct sd_data *sdd = &tl->data;
1801
1802                 sdd->sd = alloc_percpu(struct sched_domain *);
1803                 if (!sdd->sd)
1804                         return -ENOMEM;
1805
1806                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1807                 if (!sdd->sds)
1808                         return -ENOMEM;
1809
1810                 sdd->sg = alloc_percpu(struct sched_group *);
1811                 if (!sdd->sg)
1812                         return -ENOMEM;
1813
1814                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1815                 if (!sdd->sgc)
1816                         return -ENOMEM;
1817
1818                 for_each_cpu(j, cpu_map) {
1819                         struct sched_domain *sd;
1820                         struct sched_domain_shared *sds;
1821                         struct sched_group *sg;
1822                         struct sched_group_capacity *sgc;
1823
1824                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1825                                         GFP_KERNEL, cpu_to_node(j));
1826                         if (!sd)
1827                                 return -ENOMEM;
1828
1829                         *per_cpu_ptr(sdd->sd, j) = sd;
1830
1831                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
1832                                         GFP_KERNEL, cpu_to_node(j));
1833                         if (!sds)
1834                                 return -ENOMEM;
1835
1836                         *per_cpu_ptr(sdd->sds, j) = sds;
1837
1838                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1839                                         GFP_KERNEL, cpu_to_node(j));
1840                         if (!sg)
1841                                 return -ENOMEM;
1842
1843                         sg->next = sg;
1844
1845                         *per_cpu_ptr(sdd->sg, j) = sg;
1846
1847                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1848                                         GFP_KERNEL, cpu_to_node(j));
1849                         if (!sgc)
1850                                 return -ENOMEM;
1851
1852 #ifdef CONFIG_SCHED_DEBUG
1853                         sgc->id = j;
1854 #endif
1855
1856                         *per_cpu_ptr(sdd->sgc, j) = sgc;
1857                 }
1858         }
1859
1860         return 0;
1861 }
1862
1863 static void __sdt_free(const struct cpumask *cpu_map)
1864 {
1865         struct sched_domain_topology_level *tl;
1866         int j;
1867
1868         for_each_sd_topology(tl) {
1869                 struct sd_data *sdd = &tl->data;
1870
1871                 for_each_cpu(j, cpu_map) {
1872                         struct sched_domain *sd;
1873
1874                         if (sdd->sd) {
1875                                 sd = *per_cpu_ptr(sdd->sd, j);
1876                                 if (sd && (sd->flags & SD_OVERLAP))
1877                                         free_sched_groups(sd->groups, 0);
1878                                 kfree(*per_cpu_ptr(sdd->sd, j));
1879                         }
1880
1881                         if (sdd->sds)
1882                                 kfree(*per_cpu_ptr(sdd->sds, j));
1883                         if (sdd->sg)
1884                                 kfree(*per_cpu_ptr(sdd->sg, j));
1885                         if (sdd->sgc)
1886                                 kfree(*per_cpu_ptr(sdd->sgc, j));
1887                 }
1888                 free_percpu(sdd->sd);
1889                 sdd->sd = NULL;
1890                 free_percpu(sdd->sds);
1891                 sdd->sds = NULL;
1892                 free_percpu(sdd->sg);
1893                 sdd->sg = NULL;
1894                 free_percpu(sdd->sgc);
1895                 sdd->sgc = NULL;
1896         }
1897 }
1898
1899 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1900                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1901                 struct sched_domain *child, int dflags, int cpu)
1902 {
1903         struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1904
1905         if (child) {
1906                 sd->level = child->level + 1;
1907                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1908                 child->parent = sd;
1909
1910                 if (!cpumask_subset(sched_domain_span(child),
1911                                     sched_domain_span(sd))) {
1912                         pr_err("BUG: arch topology borken\n");
1913 #ifdef CONFIG_SCHED_DEBUG
1914                         pr_err("     the %s domain not a subset of the %s domain\n",
1915                                         child->name, sd->name);
1916 #endif
1917                         /* Fixup, ensure @sd has at least @child CPUs. */
1918                         cpumask_or(sched_domain_span(sd),
1919                                    sched_domain_span(sd),
1920                                    sched_domain_span(child));
1921                 }
1922
1923         }
1924         set_domain_attribute(sd, attr);
1925
1926         return sd;
1927 }
1928
1929 /*
1930  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
1931  * any two given CPUs at this (non-NUMA) topology level.
1932  */
1933 static bool topology_span_sane(struct sched_domain_topology_level *tl,
1934                               const struct cpumask *cpu_map, int cpu)
1935 {
1936         int i;
1937
1938         /* NUMA levels are allowed to overlap */
1939         if (tl->flags & SDTL_OVERLAP)
1940                 return true;
1941
1942         /*
1943          * Non-NUMA levels cannot partially overlap - they must be either
1944          * completely equal or completely disjoint. Otherwise we can end up
1945          * breaking the sched_group lists - i.e. a later get_group() pass
1946          * breaks the linking done for an earlier span.
1947          */
1948         for_each_cpu(i, cpu_map) {
1949                 if (i == cpu)
1950                         continue;
1951                 /*
1952                  * We should 'and' all those masks with 'cpu_map' to exactly
1953                  * match the topology we're about to build, but that can only
1954                  * remove CPUs, which only lessens our ability to detect
1955                  * overlaps
1956                  */
1957                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
1958                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
1959                         return false;
1960         }
1961
1962         return true;
1963 }
1964
1965 /*
1966  * Find the sched_domain_topology_level where all CPU capacities are visible
1967  * for all CPUs.
1968  */
1969 static struct sched_domain_topology_level
1970 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1971 {
1972         int i, j, asym_level = 0;
1973         bool asym = false;
1974         struct sched_domain_topology_level *tl, *asym_tl = NULL;
1975         unsigned long cap;
1976
1977         /* Is there any asymmetry? */
1978         cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1979
1980         for_each_cpu(i, cpu_map) {
1981                 if (arch_scale_cpu_capacity(i) != cap) {
1982                         asym = true;
1983                         break;
1984                 }
1985         }
1986
1987         if (!asym)
1988                 return NULL;
1989
1990         /*
1991          * Examine topology from all CPU's point of views to detect the lowest
1992          * sched_domain_topology_level where a highest capacity CPU is visible
1993          * to everyone.
1994          */
1995         for_each_cpu(i, cpu_map) {
1996                 unsigned long max_capacity = arch_scale_cpu_capacity(i);
1997                 int tl_id = 0;
1998
1999                 for_each_sd_topology(tl) {
2000                         if (tl_id < asym_level)
2001                                 goto next_level;
2002
2003                         for_each_cpu_and(j, tl->mask(i), cpu_map) {
2004                                 unsigned long capacity;
2005
2006                                 capacity = arch_scale_cpu_capacity(j);
2007
2008                                 if (capacity <= max_capacity)
2009                                         continue;
2010
2011                                 max_capacity = capacity;
2012                                 asym_level = tl_id;
2013                                 asym_tl = tl;
2014                         }
2015 next_level:
2016                         tl_id++;
2017                 }
2018         }
2019
2020         return asym_tl;
2021 }
2022
2023
2024 /*
2025  * Build sched domains for a given set of CPUs and attach the sched domains
2026  * to the individual CPUs
2027  */
2028 static int
2029 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2030 {
2031         enum s_alloc alloc_state = sa_none;
2032         struct sched_domain *sd;
2033         struct s_data d;
2034         struct rq *rq = NULL;
2035         int i, ret = -ENOMEM;
2036         struct sched_domain_topology_level *tl_asym;
2037         bool has_asym = false;
2038
2039         if (WARN_ON(cpumask_empty(cpu_map)))
2040                 goto error;
2041
2042         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2043         if (alloc_state != sa_rootdomain)
2044                 goto error;
2045
2046         tl_asym = asym_cpu_capacity_level(cpu_map);
2047
2048         /* Set up domains for CPUs specified by the cpu_map: */
2049         for_each_cpu(i, cpu_map) {
2050                 struct sched_domain_topology_level *tl;
2051                 int dflags = 0;
2052
2053                 sd = NULL;
2054                 for_each_sd_topology(tl) {
2055                         if (tl == tl_asym) {
2056                                 dflags |= SD_ASYM_CPUCAPACITY;
2057                                 has_asym = true;
2058                         }
2059
2060                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2061                                 goto error;
2062
2063                         sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2064
2065                         if (tl == sched_domain_topology)
2066                                 *per_cpu_ptr(d.sd, i) = sd;
2067                         if (tl->flags & SDTL_OVERLAP)
2068                                 sd->flags |= SD_OVERLAP;
2069                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2070                                 break;
2071                 }
2072         }
2073
2074         /* Build the groups for the domains */
2075         for_each_cpu(i, cpu_map) {
2076                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2077                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2078                         if (sd->flags & SD_OVERLAP) {
2079                                 if (build_overlap_sched_groups(sd, i))
2080                                         goto error;
2081                         } else {
2082                                 if (build_sched_groups(sd, i))
2083                                         goto error;
2084                         }
2085                 }
2086         }
2087
2088         /* Calculate CPU capacity for physical packages and nodes */
2089         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2090                 if (!cpumask_test_cpu(i, cpu_map))
2091                         continue;
2092
2093                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2094                         claim_allocations(i, sd);
2095                         init_sched_groups_capacity(i, sd);
2096                 }
2097         }
2098
2099         /* Attach the domains */
2100         rcu_read_lock();
2101         for_each_cpu(i, cpu_map) {
2102                 rq = cpu_rq(i);
2103                 sd = *per_cpu_ptr(d.sd, i);
2104
2105                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2106                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2107                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2108
2109                 cpu_attach_domain(sd, d.rd, i);
2110         }
2111         rcu_read_unlock();
2112
2113         if (has_asym)
2114                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2115
2116         if (rq && sched_debug_enabled) {
2117                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2118                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2119         }
2120
2121         ret = 0;
2122 error:
2123         __free_domain_allocs(&d, alloc_state, cpu_map);
2124
2125         return ret;
2126 }
2127
2128 /* Current sched domains: */
2129 static cpumask_var_t                    *doms_cur;
2130
2131 /* Number of sched domains in 'doms_cur': */
2132 static int                              ndoms_cur;
2133
2134 /* Attribues of custom domains in 'doms_cur' */
2135 static struct sched_domain_attr         *dattr_cur;
2136
2137 /*
2138  * Special case: If a kmalloc() of a doms_cur partition (array of
2139  * cpumask) fails, then fallback to a single sched domain,
2140  * as determined by the single cpumask fallback_doms.
2141  */
2142 static cpumask_var_t                    fallback_doms;
2143
2144 /*
2145  * arch_update_cpu_topology lets virtualized architectures update the
2146  * CPU core maps. It is supposed to return 1 if the topology changed
2147  * or 0 if it stayed the same.
2148  */
2149 int __weak arch_update_cpu_topology(void)
2150 {
2151         return 0;
2152 }
2153
2154 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2155 {
2156         int i;
2157         cpumask_var_t *doms;
2158
2159         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2160         if (!doms)
2161                 return NULL;
2162         for (i = 0; i < ndoms; i++) {
2163                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2164                         free_sched_domains(doms, i);
2165                         return NULL;
2166                 }
2167         }
2168         return doms;
2169 }
2170
2171 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2172 {
2173         unsigned int i;
2174         for (i = 0; i < ndoms; i++)
2175                 free_cpumask_var(doms[i]);
2176         kfree(doms);
2177 }
2178
2179 /*
2180  * Set up scheduler domains and groups.  For now this just excludes isolated
2181  * CPUs, but could be used to exclude other special cases in the future.
2182  */
2183 int sched_init_domains(const struct cpumask *cpu_map)
2184 {
2185         int err;
2186
2187         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2188         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2189         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2190
2191         arch_update_cpu_topology();
2192         ndoms_cur = 1;
2193         doms_cur = alloc_sched_domains(ndoms_cur);
2194         if (!doms_cur)
2195                 doms_cur = &fallback_doms;
2196         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2197         err = build_sched_domains(doms_cur[0], NULL);
2198         register_sched_domain_sysctl();
2199
2200         return err;
2201 }
2202
2203 /*
2204  * Detach sched domains from a group of CPUs specified in cpu_map
2205  * These CPUs will now be attached to the NULL domain
2206  */
2207 static void detach_destroy_domains(const struct cpumask *cpu_map)
2208 {
2209         unsigned int cpu = cpumask_any(cpu_map);
2210         int i;
2211
2212         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2213                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2214
2215         rcu_read_lock();
2216         for_each_cpu(i, cpu_map)
2217                 cpu_attach_domain(NULL, &def_root_domain, i);
2218         rcu_read_unlock();
2219 }
2220
2221 /* handle null as "default" */
2222 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2223                         struct sched_domain_attr *new, int idx_new)
2224 {
2225         struct sched_domain_attr tmp;
2226
2227         /* Fast path: */
2228         if (!new && !cur)
2229                 return 1;
2230
2231         tmp = SD_ATTR_INIT;
2232
2233         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2234                         new ? (new + idx_new) : &tmp,
2235                         sizeof(struct sched_domain_attr));
2236 }
2237
2238 /*
2239  * Partition sched domains as specified by the 'ndoms_new'
2240  * cpumasks in the array doms_new[] of cpumasks. This compares
2241  * doms_new[] to the current sched domain partitioning, doms_cur[].
2242  * It destroys each deleted domain and builds each new domain.
2243  *
2244  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2245  * The masks don't intersect (don't overlap.) We should setup one
2246  * sched domain for each mask. CPUs not in any of the cpumasks will
2247  * not be load balanced. If the same cpumask appears both in the
2248  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2249  * it as it is.
2250  *
2251  * The passed in 'doms_new' should be allocated using
2252  * alloc_sched_domains.  This routine takes ownership of it and will
2253  * free_sched_domains it when done with it. If the caller failed the
2254  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2255  * and partition_sched_domains() will fallback to the single partition
2256  * 'fallback_doms', it also forces the domains to be rebuilt.
2257  *
2258  * If doms_new == NULL it will be replaced with cpu_online_mask.
2259  * ndoms_new == 0 is a special case for destroying existing domains,
2260  * and it will not create the default domain.
2261  *
2262  * Call with hotplug lock and sched_domains_mutex held
2263  */
2264 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2265                                     struct sched_domain_attr *dattr_new)
2266 {
2267         bool __maybe_unused has_eas = false;
2268         int i, j, n;
2269         int new_topology;
2270
2271         lockdep_assert_held(&sched_domains_mutex);
2272
2273         /* Always unregister in case we don't destroy any domains: */
2274         unregister_sched_domain_sysctl();
2275
2276         /* Let the architecture update CPU core mappings: */
2277         new_topology = arch_update_cpu_topology();
2278
2279         if (!doms_new) {
2280                 WARN_ON_ONCE(dattr_new);
2281                 n = 0;
2282                 doms_new = alloc_sched_domains(1);
2283                 if (doms_new) {
2284                         n = 1;
2285                         cpumask_and(doms_new[0], cpu_active_mask,
2286                                     housekeeping_cpumask(HK_FLAG_DOMAIN));
2287                 }
2288         } else {
2289                 n = ndoms_new;
2290         }
2291
2292         /* Destroy deleted domains: */
2293         for (i = 0; i < ndoms_cur; i++) {
2294                 for (j = 0; j < n && !new_topology; j++) {
2295                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2296                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2297                                 struct root_domain *rd;
2298
2299                                 /*
2300                                  * This domain won't be destroyed and as such
2301                                  * its dl_bw->total_bw needs to be cleared.  It
2302                                  * will be recomputed in function
2303                                  * update_tasks_root_domain().
2304                                  */
2305                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2306                                 dl_clear_root_domain(rd);
2307                                 goto match1;
2308                         }
2309                 }
2310                 /* No match - a current sched domain not in new doms_new[] */
2311                 detach_destroy_domains(doms_cur[i]);
2312 match1:
2313                 ;
2314         }
2315
2316         n = ndoms_cur;
2317         if (!doms_new) {
2318                 n = 0;
2319                 doms_new = &fallback_doms;
2320                 cpumask_and(doms_new[0], cpu_active_mask,
2321                             housekeeping_cpumask(HK_FLAG_DOMAIN));
2322         }
2323
2324         /* Build new domains: */
2325         for (i = 0; i < ndoms_new; i++) {
2326                 for (j = 0; j < n && !new_topology; j++) {
2327                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2328                             dattrs_equal(dattr_new, i, dattr_cur, j))
2329                                 goto match2;
2330                 }
2331                 /* No match - add a new doms_new */
2332                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2333 match2:
2334                 ;
2335         }
2336
2337 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2338         /* Build perf. domains: */
2339         for (i = 0; i < ndoms_new; i++) {
2340                 for (j = 0; j < n && !sched_energy_update; j++) {
2341                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2342                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2343                                 has_eas = true;
2344                                 goto match3;
2345                         }
2346                 }
2347                 /* No match - add perf. domains for a new rd */
2348                 has_eas |= build_perf_domains(doms_new[i]);
2349 match3:
2350                 ;
2351         }
2352         sched_energy_set(has_eas);
2353 #endif
2354
2355         /* Remember the new sched domains: */
2356         if (doms_cur != &fallback_doms)
2357                 free_sched_domains(doms_cur, ndoms_cur);
2358
2359         kfree(dattr_cur);
2360         doms_cur = doms_new;
2361         dattr_cur = dattr_new;
2362         ndoms_cur = ndoms_new;
2363
2364         register_sched_domain_sysctl();
2365 }
2366
2367 /*
2368  * Call with hotplug lock held
2369  */
2370 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2371                              struct sched_domain_attr *dattr_new)
2372 {
2373         mutex_lock(&sched_domains_mutex);
2374         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2375         mutex_unlock(&sched_domains_mutex);
2376 }