kernel/sched/topology.c

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