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