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         bool already_visited;
1063
1064         if (child)
1065                 cpu = cpumask_first(sched_domain_span(child));
1066
1067         sg = *per_cpu_ptr(sdd->sg, cpu);
1068         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1069
1070         /* Increase refcounts for claim_allocations: */
1071         already_visited = atomic_inc_return(&sg->ref) > 1;
1072         /* sgc visits should follow a similar trend as sg */
1073         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1074
1075         /* If we have already visited that group, it's already initialized. */
1076         if (already_visited)
1077                 return sg;
1078
1079         if (child) {
1080                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1081                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1082         } else {
1083                 cpumask_set_cpu(cpu, sched_group_span(sg));
1084                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1085         }
1086
1087         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1088         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1089         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1090
1091         return sg;
1092 }
1093
1094 /*
1095  * build_sched_groups will build a circular linked list of the groups
1096  * covered by the given span, will set each group's ->cpumask correctly,
1097  * and will initialize their ->sgc.
1098  *
1099  * Assumes the sched_domain tree is fully constructed
1100  */
1101 static int
1102 build_sched_groups(struct sched_domain *sd, int cpu)
1103 {
1104         struct sched_group *first = NULL, *last = NULL;
1105         struct sd_data *sdd = sd->private;
1106         const struct cpumask *span = sched_domain_span(sd);
1107         struct cpumask *covered;
1108         int i;
1109
1110         lockdep_assert_held(&sched_domains_mutex);
1111         covered = sched_domains_tmpmask;
1112
1113         cpumask_clear(covered);
1114
1115         for_each_cpu_wrap(i, span, cpu) {
1116                 struct sched_group *sg;
1117
1118                 if (cpumask_test_cpu(i, covered))
1119                         continue;
1120
1121                 sg = get_group(i, sdd);
1122
1123                 cpumask_or(covered, covered, sched_group_span(sg));
1124
1125                 if (!first)
1126                         first = sg;
1127                 if (last)
1128                         last->next = sg;
1129                 last = sg;
1130         }
1131         last->next = first;
1132         sd->groups = first;
1133
1134         return 0;
1135 }
1136
1137 /*
1138  * Initialize sched groups cpu_capacity.
1139  *
1140  * cpu_capacity indicates the capacity of sched group, which is used while
1141  * distributing the load between different sched groups in a sched domain.
1142  * Typically cpu_capacity for all the groups in a sched domain will be same
1143  * unless there are asymmetries in the topology. If there are asymmetries,
1144  * group having more cpu_capacity will pickup more load compared to the
1145  * group having less cpu_capacity.
1146  */
1147 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1148 {
1149         struct sched_group *sg = sd->groups;
1150
1151         WARN_ON(!sg);
1152
1153         do {
1154                 int cpu, max_cpu = -1;
1155
1156                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1157
1158                 if (!(sd->flags & SD_ASYM_PACKING))
1159                         goto next;
1160
1161                 for_each_cpu(cpu, sched_group_span(sg)) {
1162                         if (max_cpu < 0)
1163                                 max_cpu = cpu;
1164                         else if (sched_asym_prefer(cpu, max_cpu))
1165                                 max_cpu = cpu;
1166                 }
1167                 sg->asym_prefer_cpu = max_cpu;
1168
1169 next:
1170                 sg = sg->next;
1171         } while (sg != sd->groups);
1172
1173         if (cpu != group_balance_cpu(sg))
1174                 return;
1175
1176         update_group_capacity(sd, cpu);
1177 }
1178
1179 /*
1180  * Initializers for schedule domains
1181  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1182  */
1183
1184 static int default_relax_domain_level = -1;
1185 int sched_domain_level_max;
1186
1187 static int __init setup_relax_domain_level(char *str)
1188 {
1189         if (kstrtoint(str, 0, &default_relax_domain_level))
1190                 pr_warn("Unable to set relax_domain_level\n");
1191
1192         return 1;
1193 }
1194 __setup("relax_domain_level=", setup_relax_domain_level);
1195
1196 static void set_domain_attribute(struct sched_domain *sd,
1197                                  struct sched_domain_attr *attr)
1198 {
1199         int request;
1200
1201         if (!attr || attr->relax_domain_level < 0) {
1202                 if (default_relax_domain_level < 0)
1203                         return;
1204                 request = default_relax_domain_level;
1205         } else
1206                 request = attr->relax_domain_level;
1207
1208         if (sd->level > request) {
1209                 /* Turn off idle balance on this domain: */
1210                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1211         }
1212 }
1213
1214 static void __sdt_free(const struct cpumask *cpu_map);
1215 static int __sdt_alloc(const struct cpumask *cpu_map);
1216
1217 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1218                                  const struct cpumask *cpu_map)
1219 {
1220         switch (what) {
1221         case sa_rootdomain:
1222                 if (!atomic_read(&d->rd->refcount))
1223                         free_rootdomain(&d->rd->rcu);
1224                 /* Fall through */
1225         case sa_sd:
1226                 free_percpu(d->sd);
1227                 /* Fall through */
1228         case sa_sd_storage:
1229                 __sdt_free(cpu_map);
1230                 /* Fall through */
1231         case sa_none:
1232                 break;
1233         }
1234 }
1235
1236 static enum s_alloc
1237 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1238 {
1239         memset(d, 0, sizeof(*d));
1240
1241         if (__sdt_alloc(cpu_map))
1242                 return sa_sd_storage;
1243         d->sd = alloc_percpu(struct sched_domain *);
1244         if (!d->sd)
1245                 return sa_sd_storage;
1246         d->rd = alloc_rootdomain();
1247         if (!d->rd)
1248                 return sa_sd;
1249
1250         return sa_rootdomain;
1251 }
1252
1253 /*
1254  * NULL the sd_data elements we've used to build the sched_domain and
1255  * sched_group structure so that the subsequent __free_domain_allocs()
1256  * will not free the data we're using.
1257  */
1258 static void claim_allocations(int cpu, struct sched_domain *sd)
1259 {
1260         struct sd_data *sdd = sd->private;
1261
1262         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1263         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1264
1265         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1266                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1267
1268         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1269                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1270
1271         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1272                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1273 }
1274
1275 #ifdef CONFIG_NUMA
1276 enum numa_topology_type sched_numa_topology_type;
1277
1278 static int                      sched_domains_numa_levels;
1279 static int                      sched_domains_curr_level;
1280
1281 int                             sched_max_numa_distance;
1282 static int                      *sched_domains_numa_distance;
1283 static struct cpumask           ***sched_domains_numa_masks;
1284 int __read_mostly               node_reclaim_distance = RECLAIM_DISTANCE;
1285 #endif
1286
1287 /*
1288  * SD_flags allowed in topology descriptions.
1289  *
1290  * These flags are purely descriptive of the topology and do not prescribe
1291  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1292  * function:
1293  *
1294  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1295  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1296  *   SD_NUMA                - describes NUMA topologies
1297  *   SD_SHARE_POWERDOMAIN   - describes shared power domain
1298  *
1299  * Odd one out, which beside describing the topology has a quirk also
1300  * prescribes the desired behaviour that goes along with it:
1301  *
1302  *   SD_ASYM_PACKING        - describes SMT quirks
1303  */
1304 #define TOPOLOGY_SD_FLAGS               \
1305         (SD_SHARE_CPUCAPACITY   |       \
1306          SD_SHARE_PKG_RESOURCES |       \
1307          SD_NUMA                |       \
1308          SD_ASYM_PACKING        |       \
1309          SD_SHARE_POWERDOMAIN)
1310
1311 static struct sched_domain *
1312 sd_init(struct sched_domain_topology_level *tl,
1313         const struct cpumask *cpu_map,
1314         struct sched_domain *child, int dflags, int cpu)
1315 {
1316         struct sd_data *sdd = &tl->data;
1317         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1318         int sd_id, sd_weight, sd_flags = 0;
1319
1320 #ifdef CONFIG_NUMA
1321         /*
1322          * Ugly hack to pass state to sd_numa_mask()...
1323          */
1324         sched_domains_curr_level = tl->numa_level;
1325 #endif
1326
1327         sd_weight = cpumask_weight(tl->mask(cpu));
1328
1329         if (tl->sd_flags)
1330                 sd_flags = (*tl->sd_flags)();
1331         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1332                         "wrong sd_flags in topology description\n"))
1333                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1334
1335         /* Apply detected topology flags */
1336         sd_flags |= dflags;
1337
1338         *sd = (struct sched_domain){
1339                 .min_interval           = sd_weight,
1340                 .max_interval           = 2*sd_weight,
1341                 .busy_factor            = 32,
1342                 .imbalance_pct          = 125,
1343
1344                 .cache_nice_tries       = 0,
1345
1346                 .flags                  = 1*SD_LOAD_BALANCE
1347                                         | 1*SD_BALANCE_NEWIDLE
1348                                         | 1*SD_BALANCE_EXEC
1349                                         | 1*SD_BALANCE_FORK
1350                                         | 0*SD_BALANCE_WAKE
1351                                         | 1*SD_WAKE_AFFINE
1352                                         | 0*SD_SHARE_CPUCAPACITY
1353                                         | 0*SD_SHARE_PKG_RESOURCES
1354                                         | 0*SD_SERIALIZE
1355                                         | 1*SD_PREFER_SIBLING
1356                                         | 0*SD_NUMA
1357                                         | sd_flags
1358                                         ,
1359
1360                 .last_balance           = jiffies,
1361                 .balance_interval       = sd_weight,
1362                 .max_newidle_lb_cost    = 0,
1363                 .next_decay_max_lb_cost = jiffies,
1364                 .child                  = child,
1365 #ifdef CONFIG_SCHED_DEBUG
1366                 .name                   = tl->name,
1367 #endif
1368         };
1369
1370         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1371         sd_id = cpumask_first(sched_domain_span(sd));
1372
1373         /*
1374          * Convert topological properties into behaviour.
1375          */
1376
1377         if (sd->flags & SD_ASYM_CPUCAPACITY) {
1378                 struct sched_domain *t = sd;
1379
1380                 /*
1381                  * Don't attempt to spread across CPUs of different capacities.
1382                  */
1383                 if (sd->child)
1384                         sd->child->flags &= ~SD_PREFER_SIBLING;
1385
1386                 for_each_lower_domain(t)
1387                         t->flags |= SD_BALANCE_WAKE;
1388         }
1389
1390         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1391                 sd->imbalance_pct = 110;
1392
1393         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1394                 sd->imbalance_pct = 117;
1395                 sd->cache_nice_tries = 1;
1396
1397 #ifdef CONFIG_NUMA
1398         } else if (sd->flags & SD_NUMA) {
1399                 sd->cache_nice_tries = 2;
1400
1401                 sd->flags &= ~SD_PREFER_SIBLING;
1402                 sd->flags |= SD_SERIALIZE;
1403                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1404                         sd->flags &= ~(SD_BALANCE_EXEC |
1405                                        SD_BALANCE_FORK |
1406                                        SD_WAKE_AFFINE);
1407                 }
1408
1409 #endif
1410         } else {
1411                 sd->cache_nice_tries = 1;
1412         }
1413
1414         /*
1415          * For all levels sharing cache; connect a sched_domain_shared
1416          * instance.
1417          */
1418         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1419                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1420                 atomic_inc(&sd->shared->ref);
1421                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1422         }
1423
1424         sd->private = sdd;
1425
1426         return sd;
1427 }
1428
1429 /*
1430  * Topology list, bottom-up.
1431  */
1432 static struct sched_domain_topology_level default_topology[] = {
1433 #ifdef CONFIG_SCHED_SMT
1434         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1435 #endif
1436 #ifdef CONFIG_SCHED_MC
1437         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1438 #endif
1439         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1440         { NULL, },
1441 };
1442
1443 static struct sched_domain_topology_level *sched_domain_topology =
1444         default_topology;
1445
1446 #define for_each_sd_topology(tl)                        \
1447         for (tl = sched_domain_topology; tl->mask; tl++)
1448
1449 void set_sched_topology(struct sched_domain_topology_level *tl)
1450 {
1451         if (WARN_ON_ONCE(sched_smp_initialized))
1452                 return;
1453
1454         sched_domain_topology = tl;
1455 }
1456
1457 #ifdef CONFIG_NUMA
1458
1459 static const struct cpumask *sd_numa_mask(int cpu)
1460 {
1461         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1462 }
1463
1464 static void sched_numa_warn(const char *str)
1465 {
1466         static int done = false;
1467         int i,j;
1468
1469         if (done)
1470                 return;
1471
1472         done = true;
1473
1474         printk(KERN_WARNING "ERROR: %s\n\n", str);
1475
1476         for (i = 0; i < nr_node_ids; i++) {
1477                 printk(KERN_WARNING "  ");
1478                 for (j = 0; j < nr_node_ids; j++)
1479                         printk(KERN_CONT "%02d ", node_distance(i,j));
1480                 printk(KERN_CONT "\n");
1481         }
1482         printk(KERN_WARNING "\n");
1483 }
1484
1485 bool find_numa_distance(int distance)
1486 {
1487         int i;
1488
1489         if (distance == node_distance(0, 0))
1490                 return true;
1491
1492         for (i = 0; i < sched_domains_numa_levels; i++) {
1493                 if (sched_domains_numa_distance[i] == distance)
1494                         return true;
1495         }
1496
1497         return false;
1498 }
1499
1500 /*
1501  * A system can have three types of NUMA topology:
1502  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1503  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1504  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1505  *
1506  * The difference between a glueless mesh topology and a backplane
1507  * topology lies in whether communication between not directly
1508  * connected nodes goes through intermediary nodes (where programs
1509  * could run), or through backplane controllers. This affects
1510  * placement of programs.
1511  *
1512  * The type of topology can be discerned with the following tests:
1513  * - If the maximum distance between any nodes is 1 hop, the system
1514  *   is directly connected.
1515  * - If for two nodes A and B, located N > 1 hops away from each other,
1516  *   there is an intermediary node C, which is < N hops away from both
1517  *   nodes A and B, the system is a glueless mesh.
1518  */
1519 static void init_numa_topology_type(void)
1520 {
1521         int a, b, c, n;
1522
1523         n = sched_max_numa_distance;
1524
1525         if (sched_domains_numa_levels <= 2) {
1526                 sched_numa_topology_type = NUMA_DIRECT;
1527                 return;
1528         }
1529
1530         for_each_online_node(a) {
1531                 for_each_online_node(b) {
1532                         /* Find two nodes furthest removed from each other. */
1533                         if (node_distance(a, b) < n)
1534                                 continue;
1535
1536                         /* Is there an intermediary node between a and b? */
1537                         for_each_online_node(c) {
1538                                 if (node_distance(a, c) < n &&
1539                                     node_distance(b, c) < n) {
1540                                         sched_numa_topology_type =
1541                                                         NUMA_GLUELESS_MESH;
1542                                         return;
1543                                 }
1544                         }
1545
1546                         sched_numa_topology_type = NUMA_BACKPLANE;
1547                         return;
1548                 }
1549         }
1550 }
1551
1552 void sched_init_numa(void)
1553 {
1554         int next_distance, curr_distance = node_distance(0, 0);
1555         struct sched_domain_topology_level *tl;
1556         int level = 0;
1557         int i, j, k;
1558
1559         sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1560         if (!sched_domains_numa_distance)
1561                 return;
1562
1563         /* Includes NUMA identity node at level 0. */
1564         sched_domains_numa_distance[level++] = curr_distance;
1565         sched_domains_numa_levels = level;
1566
1567         /*
1568          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1569          * unique distances in the node_distance() table.
1570          *
1571          * Assumes node_distance(0,j) includes all distances in
1572          * node_distance(i,j) in order to avoid cubic time.
1573          */
1574         next_distance = curr_distance;
1575         for (i = 0; i < nr_node_ids; i++) {
1576                 for (j = 0; j < nr_node_ids; j++) {
1577                         for (k = 0; k < nr_node_ids; k++) {
1578                                 int distance = node_distance(i, k);
1579
1580                                 if (distance > curr_distance &&
1581                                     (distance < next_distance ||
1582                                      next_distance == curr_distance))
1583                                         next_distance = distance;
1584
1585                                 /*
1586                                  * While not a strong assumption it would be nice to know
1587                                  * about cases where if node A is connected to B, B is not
1588                                  * equally connected to A.
1589                                  */
1590                                 if (sched_debug() && node_distance(k, i) != distance)
1591                                         sched_numa_warn("Node-distance not symmetric");
1592
1593                                 if (sched_debug() && i && !find_numa_distance(distance))
1594                                         sched_numa_warn("Node-0 not representative");
1595                         }
1596                         if (next_distance != curr_distance) {
1597                                 sched_domains_numa_distance[level++] = next_distance;
1598                                 sched_domains_numa_levels = level;
1599                                 curr_distance = next_distance;
1600                         } else break;
1601                 }
1602
1603                 /*
1604                  * In case of sched_debug() we verify the above assumption.
1605                  */
1606                 if (!sched_debug())
1607                         break;
1608         }
1609
1610         /*
1611          * 'level' contains the number of unique distances
1612          *
1613          * The sched_domains_numa_distance[] array includes the actual distance
1614          * numbers.
1615          */
1616
1617         /*
1618          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1619          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1620          * the array will contain less then 'level' members. This could be
1621          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1622          * in other functions.
1623          *
1624          * We reset it to 'level' at the end of this function.
1625          */
1626         sched_domains_numa_levels = 0;
1627
1628         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1629         if (!sched_domains_numa_masks)
1630                 return;
1631
1632         /*
1633          * Now for each level, construct a mask per node which contains all
1634          * CPUs of nodes that are that many hops away from us.
1635          */
1636         for (i = 0; i < level; i++) {
1637                 sched_domains_numa_masks[i] =
1638                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1639                 if (!sched_domains_numa_masks[i])
1640                         return;
1641
1642                 for (j = 0; j < nr_node_ids; j++) {
1643                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1644                         if (!mask)
1645                                 return;
1646
1647                         sched_domains_numa_masks[i][j] = mask;
1648
1649                         for_each_node(k) {
1650                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1651                                         continue;
1652
1653                                 cpumask_or(mask, mask, cpumask_of_node(k));
1654                         }
1655                 }
1656         }
1657
1658         /* Compute default topology size */
1659         for (i = 0; sched_domain_topology[i].mask; i++);
1660
1661         tl = kzalloc((i + level + 1) *
1662                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1663         if (!tl)
1664                 return;
1665
1666         /*
1667          * Copy the default topology bits..
1668          */
1669         for (i = 0; sched_domain_topology[i].mask; i++)
1670                 tl[i] = sched_domain_topology[i];
1671
1672         /*
1673          * Add the NUMA identity distance, aka single NODE.
1674          */
1675         tl[i++] = (struct sched_domain_topology_level){
1676                 .mask = sd_numa_mask,
1677                 .numa_level = 0,
1678                 SD_INIT_NAME(NODE)
1679         };
1680
1681         /*
1682          * .. and append 'j' levels of NUMA goodness.
1683          */
1684         for (j = 1; j < level; i++, j++) {
1685                 tl[i] = (struct sched_domain_topology_level){
1686                         .mask = sd_numa_mask,
1687                         .sd_flags = cpu_numa_flags,
1688                         .flags = SDTL_OVERLAP,
1689                         .numa_level = j,
1690                         SD_INIT_NAME(NUMA)
1691                 };
1692         }
1693
1694         sched_domain_topology = tl;
1695
1696         sched_domains_numa_levels = level;
1697         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1698
1699         init_numa_topology_type();
1700 }
1701
1702 void sched_domains_numa_masks_set(unsigned int cpu)
1703 {
1704         int node = cpu_to_node(cpu);
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                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
1710                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1711                 }
1712         }
1713 }
1714
1715 void sched_domains_numa_masks_clear(unsigned int cpu)
1716 {
1717         int i, j;
1718
1719         for (i = 0; i < sched_domains_numa_levels; i++) {
1720                 for (j = 0; j < nr_node_ids; j++)
1721                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1722         }
1723 }
1724
1725 /*
1726  * sched_numa_find_closest() - given the NUMA topology, find the cpu
1727  *                             closest to @cpu from @cpumask.
1728  * cpumask: cpumask to find a cpu from
1729  * cpu: cpu to be close to
1730  *
1731  * returns: cpu, or nr_cpu_ids when nothing found.
1732  */
1733 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1734 {
1735         int i, j = cpu_to_node(cpu);
1736
1737         for (i = 0; i < sched_domains_numa_levels; i++) {
1738                 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1739                 if (cpu < nr_cpu_ids)
1740                         return cpu;
1741         }
1742         return nr_cpu_ids;
1743 }
1744
1745 #endif /* CONFIG_NUMA */
1746
1747 static int __sdt_alloc(const struct cpumask *cpu_map)
1748 {
1749         struct sched_domain_topology_level *tl;
1750         int j;
1751
1752         for_each_sd_topology(tl) {
1753                 struct sd_data *sdd = &tl->data;
1754
1755                 sdd->sd = alloc_percpu(struct sched_domain *);
1756                 if (!sdd->sd)
1757                         return -ENOMEM;
1758
1759                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1760                 if (!sdd->sds)
1761                         return -ENOMEM;
1762
1763                 sdd->sg = alloc_percpu(struct sched_group *);
1764                 if (!sdd->sg)
1765                         return -ENOMEM;
1766
1767                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1768                 if (!sdd->sgc)
1769                         return -ENOMEM;
1770
1771                 for_each_cpu(j, cpu_map) {
1772                         struct sched_domain *sd;
1773                         struct sched_domain_shared *sds;
1774                         struct sched_group *sg;
1775                         struct sched_group_capacity *sgc;
1776
1777                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1778                                         GFP_KERNEL, cpu_to_node(j));
1779                         if (!sd)
1780                                 return -ENOMEM;
1781
1782                         *per_cpu_ptr(sdd->sd, j) = sd;
1783
1784                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
1785                                         GFP_KERNEL, cpu_to_node(j));
1786                         if (!sds)
1787                                 return -ENOMEM;
1788
1789                         *per_cpu_ptr(sdd->sds, j) = sds;
1790
1791                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1792                                         GFP_KERNEL, cpu_to_node(j));
1793                         if (!sg)
1794                                 return -ENOMEM;
1795
1796                         sg->next = sg;
1797
1798                         *per_cpu_ptr(sdd->sg, j) = sg;
1799
1800                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1801                                         GFP_KERNEL, cpu_to_node(j));
1802                         if (!sgc)
1803                                 return -ENOMEM;
1804
1805 #ifdef CONFIG_SCHED_DEBUG
1806                         sgc->id = j;
1807 #endif
1808
1809                         *per_cpu_ptr(sdd->sgc, j) = sgc;
1810                 }
1811         }
1812
1813         return 0;
1814 }
1815
1816 static void __sdt_free(const struct cpumask *cpu_map)
1817 {
1818         struct sched_domain_topology_level *tl;
1819         int j;
1820
1821         for_each_sd_topology(tl) {
1822                 struct sd_data *sdd = &tl->data;
1823
1824                 for_each_cpu(j, cpu_map) {
1825                         struct sched_domain *sd;
1826
1827                         if (sdd->sd) {
1828                                 sd = *per_cpu_ptr(sdd->sd, j);
1829                                 if (sd && (sd->flags & SD_OVERLAP))
1830                                         free_sched_groups(sd->groups, 0);
1831                                 kfree(*per_cpu_ptr(sdd->sd, j));
1832                         }
1833
1834                         if (sdd->sds)
1835                                 kfree(*per_cpu_ptr(sdd->sds, j));
1836                         if (sdd->sg)
1837                                 kfree(*per_cpu_ptr(sdd->sg, j));
1838                         if (sdd->sgc)
1839                                 kfree(*per_cpu_ptr(sdd->sgc, j));
1840                 }
1841                 free_percpu(sdd->sd);
1842                 sdd->sd = NULL;
1843                 free_percpu(sdd->sds);
1844                 sdd->sds = NULL;
1845                 free_percpu(sdd->sg);
1846                 sdd->sg = NULL;
1847                 free_percpu(sdd->sgc);
1848                 sdd->sgc = NULL;
1849         }
1850 }
1851
1852 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1853                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1854                 struct sched_domain *child, int dflags, int cpu)
1855 {
1856         struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1857
1858         if (child) {
1859                 sd->level = child->level + 1;
1860                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1861                 child->parent = sd;
1862
1863                 if (!cpumask_subset(sched_domain_span(child),
1864                                     sched_domain_span(sd))) {
1865                         pr_err("BUG: arch topology borken\n");
1866 #ifdef CONFIG_SCHED_DEBUG
1867                         pr_err("     the %s domain not a subset of the %s domain\n",
1868                                         child->name, sd->name);
1869 #endif
1870                         /* Fixup, ensure @sd has at least @child CPUs. */
1871                         cpumask_or(sched_domain_span(sd),
1872                                    sched_domain_span(sd),
1873                                    sched_domain_span(child));
1874                 }
1875
1876         }
1877         set_domain_attribute(sd, attr);
1878
1879         return sd;
1880 }
1881
1882 /*
1883  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
1884  * any two given CPUs at this (non-NUMA) topology level.
1885  */
1886 static bool topology_span_sane(struct sched_domain_topology_level *tl,
1887                               const struct cpumask *cpu_map, int cpu)
1888 {
1889         int i;
1890
1891         /* NUMA levels are allowed to overlap */
1892         if (tl->flags & SDTL_OVERLAP)
1893                 return true;
1894
1895         /*
1896          * Non-NUMA levels cannot partially overlap - they must be either
1897          * completely equal or completely disjoint. Otherwise we can end up
1898          * breaking the sched_group lists - i.e. a later get_group() pass
1899          * breaks the linking done for an earlier span.
1900          */
1901         for_each_cpu(i, cpu_map) {
1902                 if (i == cpu)
1903                         continue;
1904                 /*
1905                  * We should 'and' all those masks with 'cpu_map' to exactly
1906                  * match the topology we're about to build, but that can only
1907                  * remove CPUs, which only lessens our ability to detect
1908                  * overlaps
1909                  */
1910                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
1911                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
1912                         return false;
1913         }
1914
1915         return true;
1916 }
1917
1918 /*
1919  * Find the sched_domain_topology_level where all CPU capacities are visible
1920  * for all CPUs.
1921  */
1922 static struct sched_domain_topology_level
1923 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1924 {
1925         int i, j, asym_level = 0;
1926         bool asym = false;
1927         struct sched_domain_topology_level *tl, *asym_tl = NULL;
1928         unsigned long cap;
1929
1930         /* Is there any asymmetry? */
1931         cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1932
1933         for_each_cpu(i, cpu_map) {
1934                 if (arch_scale_cpu_capacity(i) != cap) {
1935                         asym = true;
1936                         break;
1937                 }
1938         }
1939
1940         if (!asym)
1941                 return NULL;
1942
1943         /*
1944          * Examine topology from all CPU's point of views to detect the lowest
1945          * sched_domain_topology_level where a highest capacity CPU is visible
1946          * to everyone.
1947          */
1948         for_each_cpu(i, cpu_map) {
1949                 unsigned long max_capacity = arch_scale_cpu_capacity(i);
1950                 int tl_id = 0;
1951
1952                 for_each_sd_topology(tl) {
1953                         if (tl_id < asym_level)
1954                                 goto next_level;
1955
1956                         for_each_cpu_and(j, tl->mask(i), cpu_map) {
1957                                 unsigned long capacity;
1958
1959                                 capacity = arch_scale_cpu_capacity(j);
1960
1961                                 if (capacity <= max_capacity)
1962                                         continue;
1963
1964                                 max_capacity = capacity;
1965                                 asym_level = tl_id;
1966                                 asym_tl = tl;
1967                         }
1968 next_level:
1969                         tl_id++;
1970                 }
1971         }
1972
1973         return asym_tl;
1974 }
1975
1976
1977 /*
1978  * Build sched domains for a given set of CPUs and attach the sched domains
1979  * to the individual CPUs
1980  */
1981 static int
1982 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1983 {
1984         enum s_alloc alloc_state = sa_none;
1985         struct sched_domain *sd;
1986         struct s_data d;
1987         struct rq *rq = NULL;
1988         int i, ret = -ENOMEM;
1989         struct sched_domain_topology_level *tl_asym;
1990         bool has_asym = false;
1991
1992         if (WARN_ON(cpumask_empty(cpu_map)))
1993                 goto error;
1994
1995         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1996         if (alloc_state != sa_rootdomain)
1997                 goto error;
1998
1999         tl_asym = asym_cpu_capacity_level(cpu_map);
2000
2001         /* Set up domains for CPUs specified by the cpu_map: */
2002         for_each_cpu(i, cpu_map) {
2003                 struct sched_domain_topology_level *tl;
2004
2005                 sd = NULL;
2006                 for_each_sd_topology(tl) {
2007                         int dflags = 0;
2008
2009                         if (tl == tl_asym) {
2010                                 dflags |= SD_ASYM_CPUCAPACITY;
2011                                 has_asym = true;
2012                         }
2013
2014                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2015                                 goto error;
2016
2017                         sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2018
2019                         if (tl == sched_domain_topology)
2020                                 *per_cpu_ptr(d.sd, i) = sd;
2021                         if (tl->flags & SDTL_OVERLAP)
2022                                 sd->flags |= SD_OVERLAP;
2023                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2024                                 break;
2025                 }
2026         }
2027
2028         /* Build the groups for the domains */
2029         for_each_cpu(i, cpu_map) {
2030                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2031                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2032                         if (sd->flags & SD_OVERLAP) {
2033                                 if (build_overlap_sched_groups(sd, i))
2034                                         goto error;
2035                         } else {
2036                                 if (build_sched_groups(sd, i))
2037                                         goto error;
2038                         }
2039                 }
2040         }
2041
2042         /* Calculate CPU capacity for physical packages and nodes */
2043         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2044                 if (!cpumask_test_cpu(i, cpu_map))
2045                         continue;
2046
2047                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2048                         claim_allocations(i, sd);
2049                         init_sched_groups_capacity(i, sd);
2050                 }
2051         }
2052
2053         /* Attach the domains */
2054         rcu_read_lock();
2055         for_each_cpu(i, cpu_map) {
2056                 rq = cpu_rq(i);
2057                 sd = *per_cpu_ptr(d.sd, i);
2058
2059                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2060                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2061                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2062
2063                 cpu_attach_domain(sd, d.rd, i);
2064         }
2065         rcu_read_unlock();
2066
2067         if (has_asym)
2068                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2069
2070         if (rq && sched_debug_enabled) {
2071                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2072                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2073         }
2074
2075         ret = 0;
2076 error:
2077         __free_domain_allocs(&d, alloc_state, cpu_map);
2078
2079         return ret;
2080 }
2081
2082 /* Current sched domains: */
2083 static cpumask_var_t                    *doms_cur;
2084
2085 /* Number of sched domains in 'doms_cur': */
2086 static int                              ndoms_cur;
2087
2088 /* Attribues of custom domains in 'doms_cur' */
2089 static struct sched_domain_attr         *dattr_cur;
2090
2091 /*
2092  * Special case: If a kmalloc() of a doms_cur partition (array of
2093  * cpumask) fails, then fallback to a single sched domain,
2094  * as determined by the single cpumask fallback_doms.
2095  */
2096 static cpumask_var_t                    fallback_doms;
2097
2098 /*
2099  * arch_update_cpu_topology lets virtualized architectures update the
2100  * CPU core maps. It is supposed to return 1 if the topology changed
2101  * or 0 if it stayed the same.
2102  */
2103 int __weak arch_update_cpu_topology(void)
2104 {
2105         return 0;
2106 }
2107
2108 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2109 {
2110         int i;
2111         cpumask_var_t *doms;
2112
2113         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2114         if (!doms)
2115                 return NULL;
2116         for (i = 0; i < ndoms; i++) {
2117                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2118                         free_sched_domains(doms, i);
2119                         return NULL;
2120                 }
2121         }
2122         return doms;
2123 }
2124
2125 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2126 {
2127         unsigned int i;
2128         for (i = 0; i < ndoms; i++)
2129                 free_cpumask_var(doms[i]);
2130         kfree(doms);
2131 }
2132
2133 /*
2134  * Set up scheduler domains and groups.  For now this just excludes isolated
2135  * CPUs, but could be used to exclude other special cases in the future.
2136  */
2137 int sched_init_domains(const struct cpumask *cpu_map)
2138 {
2139         int err;
2140
2141         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2142         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2143         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2144
2145         arch_update_cpu_topology();
2146         ndoms_cur = 1;
2147         doms_cur = alloc_sched_domains(ndoms_cur);
2148         if (!doms_cur)
2149                 doms_cur = &fallback_doms;
2150         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2151         err = build_sched_domains(doms_cur[0], NULL);
2152         register_sched_domain_sysctl();
2153
2154         return err;
2155 }
2156
2157 /*
2158  * Detach sched domains from a group of CPUs specified in cpu_map
2159  * These CPUs will now be attached to the NULL domain
2160  */
2161 static void detach_destroy_domains(const struct cpumask *cpu_map)
2162 {
2163         unsigned int cpu = cpumask_any(cpu_map);
2164         int i;
2165
2166         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2167                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2168
2169         rcu_read_lock();
2170         for_each_cpu(i, cpu_map)
2171                 cpu_attach_domain(NULL, &def_root_domain, i);
2172         rcu_read_unlock();
2173 }
2174
2175 /* handle null as "default" */
2176 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2177                         struct sched_domain_attr *new, int idx_new)
2178 {
2179         struct sched_domain_attr tmp;
2180
2181         /* Fast path: */
2182         if (!new && !cur)
2183                 return 1;
2184
2185         tmp = SD_ATTR_INIT;
2186
2187         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2188                         new ? (new + idx_new) : &tmp,
2189                         sizeof(struct sched_domain_attr));
2190 }
2191
2192 /*
2193  * Partition sched domains as specified by the 'ndoms_new'
2194  * cpumasks in the array doms_new[] of cpumasks. This compares
2195  * doms_new[] to the current sched domain partitioning, doms_cur[].
2196  * It destroys each deleted domain and builds each new domain.
2197  *
2198  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2199  * The masks don't intersect (don't overlap.) We should setup one
2200  * sched domain for each mask. CPUs not in any of the cpumasks will
2201  * not be load balanced. If the same cpumask appears both in the
2202  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2203  * it as it is.
2204  *
2205  * The passed in 'doms_new' should be allocated using
2206  * alloc_sched_domains.  This routine takes ownership of it and will
2207  * free_sched_domains it when done with it. If the caller failed the
2208  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2209  * and partition_sched_domains() will fallback to the single partition
2210  * 'fallback_doms', it also forces the domains to be rebuilt.
2211  *
2212  * If doms_new == NULL it will be replaced with cpu_online_mask.
2213  * ndoms_new == 0 is a special case for destroying existing domains,
2214  * and it will not create the default domain.
2215  *
2216  * Call with hotplug lock and sched_domains_mutex held
2217  */
2218 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2219                                     struct sched_domain_attr *dattr_new)
2220 {
2221         bool __maybe_unused has_eas = false;
2222         int i, j, n;
2223         int new_topology;
2224
2225         lockdep_assert_held(&sched_domains_mutex);
2226
2227         /* Always unregister in case we don't destroy any domains: */
2228         unregister_sched_domain_sysctl();
2229
2230         /* Let the architecture update CPU core mappings: */
2231         new_topology = arch_update_cpu_topology();
2232
2233         if (!doms_new) {
2234                 WARN_ON_ONCE(dattr_new);
2235                 n = 0;
2236                 doms_new = alloc_sched_domains(1);
2237                 if (doms_new) {
2238                         n = 1;
2239                         cpumask_and(doms_new[0], cpu_active_mask,
2240                                     housekeeping_cpumask(HK_FLAG_DOMAIN));
2241                 }
2242         } else {
2243                 n = ndoms_new;
2244         }
2245
2246         /* Destroy deleted domains: */
2247         for (i = 0; i < ndoms_cur; i++) {
2248                 for (j = 0; j < n && !new_topology; j++) {
2249                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2250                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2251                                 struct root_domain *rd;
2252
2253                                 /*
2254                                  * This domain won't be destroyed and as such
2255                                  * its dl_bw->total_bw needs to be cleared.  It
2256                                  * will be recomputed in function
2257                                  * update_tasks_root_domain().
2258                                  */
2259                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2260                                 dl_clear_root_domain(rd);
2261                                 goto match1;
2262                         }
2263                 }
2264                 /* No match - a current sched domain not in new doms_new[] */
2265                 detach_destroy_domains(doms_cur[i]);
2266 match1:
2267                 ;
2268         }
2269
2270         n = ndoms_cur;
2271         if (!doms_new) {
2272                 n = 0;
2273                 doms_new = &fallback_doms;
2274                 cpumask_and(doms_new[0], cpu_active_mask,
2275                             housekeeping_cpumask(HK_FLAG_DOMAIN));
2276         }
2277
2278         /* Build new domains: */
2279         for (i = 0; i < ndoms_new; i++) {
2280                 for (j = 0; j < n && !new_topology; j++) {
2281                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2282                             dattrs_equal(dattr_new, i, dattr_cur, j))
2283                                 goto match2;
2284                 }
2285                 /* No match - add a new doms_new */
2286                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2287 match2:
2288                 ;
2289         }
2290
2291 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2292         /* Build perf. domains: */
2293         for (i = 0; i < ndoms_new; i++) {
2294                 for (j = 0; j < n && !sched_energy_update; j++) {
2295                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2296                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2297                                 has_eas = true;
2298                                 goto match3;
2299                         }
2300                 }
2301                 /* No match - add perf. domains for a new rd */
2302                 has_eas |= build_perf_domains(doms_new[i]);
2303 match3:
2304                 ;
2305         }
2306         sched_energy_set(has_eas);
2307 #endif
2308
2309         /* Remember the new sched domains: */
2310         if (doms_cur != &fallback_doms)
2311                 free_sched_domains(doms_cur, ndoms_cur);
2312
2313         kfree(dattr_cur);
2314         doms_cur = doms_new;
2315         dattr_cur = dattr_new;
2316         ndoms_cur = ndoms_new;
2317
2318         register_sched_domain_sysctl();
2319 }
2320
2321 /*
2322  * Call with hotplug lock held
2323  */
2324 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2325                              struct sched_domain_attr *dattr_new)
2326 {
2327         mutex_lock(&sched_domains_mutex);
2328         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2329         mutex_unlock(&sched_domains_mutex);
2330 }