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