kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5 #include "sched.h"
   6
   7 DEFINE_MUTEX(sched_domains_mutex);
   8
   9 /* Protected by sched_domains_mutex: */
  10 static cpumask_var_t sched_domains_tmpmask;
  11 static cpumask_var_t sched_domains_tmpmask2;
  12
  13 #ifdef CONFIG_SCHED_DEBUG
  14
  15 static int __init sched_debug_setup(char *str)
  16 {
  17         sched_debug_enabled = true;
  18
  19         return 0;
  20 }
  21 early_param("sched_debug", sched_debug_setup);
  22
  23 static inline bool sched_debug(void)
  24 {
  25         return sched_debug_enabled;
  26 }
  27
  28 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
  29 const struct sd_flag_debug sd_flag_debug[] = {
  30 #include <linux/sched/sd_flags.h>
  31 };
  32 #undef SD_FLAG
  33
  34 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
  35                                   struct cpumask *groupmask)
  36 {
  37         struct sched_group *group = sd->groups;
  38         unsigned long flags = sd->flags;
  39         unsigned int idx;
  40
  41         cpumask_clear(groupmask);
  42
  43         printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
  44         printk(KERN_CONT "span=%*pbl level=%s\n",
  45                cpumask_pr_args(sched_domain_span(sd)), sd->name);
  46
  47         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
  48                 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
  49         }
  50         if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
  51                 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
  52         }
  53
  54         for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
  55                 unsigned int flag = BIT(idx);
  56                 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
  57
  58                 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
  59                     !(sd->child->flags & flag))
  60                         printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
  61                                sd_flag_debug[idx].name);
  62
  63                 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
  64                     !(sd->parent->flags & flag))
  65                         printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
  66                                sd_flag_debug[idx].name);
  67         }
  68
  69         printk(KERN_DEBUG "%*s groups:", level + 1, "");
  70         do {
  71                 if (!group) {
  72                         printk("\n");
  73                         printk(KERN_ERR "ERROR: group is NULL\n");
  74                         break;
  75                 }
  76
  77                 if (!cpumask_weight(sched_group_span(group))) {
  78                         printk(KERN_CONT "\n");
  79                         printk(KERN_ERR "ERROR: empty group\n");
  80                         break;
  81                 }
  82
  83                 if (!(sd->flags & SD_OVERLAP) &&
  84                     cpumask_intersects(groupmask, sched_group_span(group))) {
  85                         printk(KERN_CONT "\n");
  86                         printk(KERN_ERR "ERROR: repeated CPUs\n");
  87                         break;
  88                 }
  89
  90                 cpumask_or(groupmask, groupmask, sched_group_span(group));
  91
  92                 printk(KERN_CONT " %d:{ span=%*pbl",
  93                                 group->sgc->id,
  94                                 cpumask_pr_args(sched_group_span(group)));
  95
  96                 if ((sd->flags & SD_OVERLAP) &&
  97                     !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
  98                         printk(KERN_CONT " mask=%*pbl",
  99                                 cpumask_pr_args(group_balance_mask(group)));
 100                 }
 101
 102                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
 103                         printk(KERN_CONT " cap=%lu", group->sgc->capacity);
 104
 105                 if (group == sd->groups && sd->child &&
 106                     !cpumask_equal(sched_domain_span(sd->child),
 107                                    sched_group_span(group))) {
 108                         printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
 109                 }
 110
 111                 printk(KERN_CONT " }");
 112
 113                 group = group->next;
 114
 115                 if (group != sd->groups)
 116                         printk(KERN_CONT ",");
 117
 118         } while (group != sd->groups);
 119         printk(KERN_CONT "\n");
 120
 121         if (!cpumask_equal(sched_domain_span(sd), groupmask))
 122                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 123
 124         if (sd->parent &&
 125             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
 126                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
 127         return 0;
 128 }
 129
 130 static void sched_domain_debug(struct sched_domain *sd, int cpu)
 131 {
 132         int level = 0;
 133
 134         if (!sched_debug_enabled)
 135                 return;
 136
 137         if (!sd) {
 138                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
 139                 return;
 140         }
 141
 142         printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
 143
 144         for (;;) {
 145                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
 146                         break;
 147                 level++;
 148                 sd = sd->parent;
 149                 if (!sd)
 150                         break;
 151         }
 152 }
 153 #else /* !CONFIG_SCHED_DEBUG */
 154
 155 # define sched_debug_enabled 0
 156 # define sched_domain_debug(sd, cpu) do { } while (0)
 157 static inline bool sched_debug(void)
 158 {
 159         return false;
 160 }
 161 #endif /* CONFIG_SCHED_DEBUG */
 162
 163 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
 164 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
 165 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
 166 #include <linux/sched/sd_flags.h>
 167 0;
 168 #undef SD_FLAG
 169
 170 static int sd_degenerate(struct sched_domain *sd)
 171 {
 172         if (cpumask_weight(sched_domain_span(sd)) == 1)
 173                 return 1;
 174
 175         /* Following flags need at least 2 groups */
 176         if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
 177             (sd->groups != sd->groups->next))
 178                 return 0;
 179
 180         /* Following flags don't use groups */
 181         if (sd->flags & (SD_WAKE_AFFINE))
 182                 return 0;
 183
 184         return 1;
 185 }
 186
 187 static int
 188 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
 189 {
 190         unsigned long cflags = sd->flags, pflags = parent->flags;
 191
 192         if (sd_degenerate(parent))
 193                 return 1;
 194
 195         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
 196                 return 0;
 197
 198         /* Flags needing groups don't count if only 1 group in parent */
 199         if (parent->groups == parent->groups->next)
 200                 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
 201
 202         if (~cflags & pflags)
 203                 return 0;
 204
 205         return 1;
 206 }
 207
 208 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
 209 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
 210 unsigned int sysctl_sched_energy_aware = 1;
 211 DEFINE_MUTEX(sched_energy_mutex);
 212 bool sched_energy_update;
 213
 214 #ifdef CONFIG_PROC_SYSCTL
 215 int sched_energy_aware_handler(struct ctl_table *table, int write,
 216                 void *buffer, size_t *lenp, loff_t *ppos)
 217 {
 218         int ret, state;
 219
 220         if (write && !capable(CAP_SYS_ADMIN))
 221                 return -EPERM;
 222
 223         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 224         if (!ret && write) {
 225                 state = static_branch_unlikely(&sched_energy_present);
 226                 if (state != sysctl_sched_energy_aware) {
 227                         mutex_lock(&sched_energy_mutex);
 228                         sched_energy_update = 1;
 229                         rebuild_sched_domains();
 230                         sched_energy_update = 0;
 231                         mutex_unlock(&sched_energy_mutex);
 232                 }
 233         }
 234
 235         return ret;
 236 }
 237 #endif
 238
 239 static void free_pd(struct perf_domain *pd)
 240 {
 241         struct perf_domain *tmp;
 242
 243         while (pd) {
 244                 tmp = pd->next;
 245                 kfree(pd);
 246                 pd = tmp;
 247         }
 248 }
 249
 250 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
 251 {
 252         while (pd) {
 253                 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
 254                         return pd;
 255                 pd = pd->next;
 256         }
 257
 258         return NULL;
 259 }
 260
 261 static struct perf_domain *pd_init(int cpu)
 262 {
 263         struct em_perf_domain *obj = em_cpu_get(cpu);
 264         struct perf_domain *pd;
 265
 266         if (!obj) {
 267                 if (sched_debug())
 268                         pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
 269                 return NULL;
 270         }
 271
 272         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 273         if (!pd)
 274                 return NULL;
 275         pd->em_pd = obj;
 276
 277         return pd;
 278 }
 279
 280 static void perf_domain_debug(const struct cpumask *cpu_map,
 281                                                 struct perf_domain *pd)
 282 {
 283         if (!sched_debug() || !pd)
 284                 return;
 285
 286         printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
 287
 288         while (pd) {
 289                 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
 290                                 cpumask_first(perf_domain_span(pd)),
 291                                 cpumask_pr_args(perf_domain_span(pd)),
 292                                 em_pd_nr_perf_states(pd->em_pd));
 293                 pd = pd->next;
 294         }
 295
 296         printk(KERN_CONT "\n");
 297 }
 298
 299 static void destroy_perf_domain_rcu(struct rcu_head *rp)
 300 {
 301         struct perf_domain *pd;
 302
 303         pd = container_of(rp, struct perf_domain, rcu);
 304         free_pd(pd);
 305 }
 306
 307 static void sched_energy_set(bool has_eas)
 308 {
 309         if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
 310                 if (sched_debug())
 311                         pr_info("%s: stopping EAS\n", __func__);
 312                 static_branch_disable_cpuslocked(&sched_energy_present);
 313         } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
 314                 if (sched_debug())
 315                         pr_info("%s: starting EAS\n", __func__);
 316                 static_branch_enable_cpuslocked(&sched_energy_present);
 317         }
 318 }
 319
 320 /*
 321  * EAS can be used on a root domain if it meets all the following conditions:
 322  *    1. an Energy Model (EM) is available;
 323  *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
 324  *    3. no SMT is detected.
 325  *    4. the EM complexity is low enough to keep scheduling overheads low;
 326  *    5. schedutil is driving the frequency of all CPUs of the rd;
 327  *
 328  * The complexity of the Energy Model is defined as:
 329  *
 330  *              C = nr_pd * (nr_cpus + nr_ps)
 331  *
 332  * with parameters defined as:
 333  *  - nr_pd:    the number of performance domains
 334  *  - nr_cpus:  the number of CPUs
 335  *  - nr_ps:    the sum of the number of performance states of all performance
 336  *              domains (for example, on a system with 2 performance domains,
 337  *              with 10 performance states each, nr_ps = 2 * 10 = 20).
 338  *
 339  * It is generally not a good idea to use such a model in the wake-up path on
 340  * very complex platforms because of the associated scheduling overheads. The
 341  * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
 342  * with per-CPU DVFS and less than 8 performance states each, for example.
 343  */
 344 #define EM_MAX_COMPLEXITY 2048
 345
 346 extern struct cpufreq_governor schedutil_gov;
 347 static bool build_perf_domains(const struct cpumask *cpu_map)
 348 {
 349         int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
 350         struct perf_domain *pd = NULL, *tmp;
 351         int cpu = cpumask_first(cpu_map);
 352         struct root_domain *rd = cpu_rq(cpu)->rd;
 353         struct cpufreq_policy *policy;
 354         struct cpufreq_governor *gov;
 355
 356         if (!sysctl_sched_energy_aware)
 357                 goto free;
 358
 359         /* EAS is enabled for asymmetric CPU capacity topologies. */
 360         if (!per_cpu(sd_asym_cpucapacity, cpu)) {
 361                 if (sched_debug()) {
 362                         pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
 363                                         cpumask_pr_args(cpu_map));
 364                 }
 365                 goto free;
 366         }
 367
 368         /* EAS definitely does *not* handle SMT */
 369         if (sched_smt_active()) {
 370                 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
 371                         cpumask_pr_args(cpu_map));
 372                 goto free;
 373         }
 374
 375         for_each_cpu(i, cpu_map) {
 376                 /* Skip already covered CPUs. */
 377                 if (find_pd(pd, i))
 378                         continue;
 379
 380                 /* Do not attempt EAS if schedutil is not being used. */
 381                 policy = cpufreq_cpu_get(i);
 382                 if (!policy)
 383                         goto free;
 384                 gov = policy->governor;
 385                 cpufreq_cpu_put(policy);
 386                 if (gov != &schedutil_gov) {
 387                         if (rd->pd)
 388                                 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
 389                                                 cpumask_pr_args(cpu_map));
 390                         goto free;
 391                 }
 392
 393                 /* Create the new pd and add it to the local list. */
 394                 tmp = pd_init(i);
 395                 if (!tmp)
 396                         goto free;
 397                 tmp->next = pd;
 398                 pd = tmp;
 399
 400                 /*
 401                  * Count performance domains and performance states for the
 402                  * complexity check.
 403                  */
 404                 nr_pd++;
 405                 nr_ps += em_pd_nr_perf_states(pd->em_pd);
 406         }
 407
 408         /* Bail out if the Energy Model complexity is too high. */
 409         if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
 410                 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
 411                                                 cpumask_pr_args(cpu_map));
 412                 goto free;
 413         }
 414
 415         perf_domain_debug(cpu_map, pd);
 416
 417         /* Attach the new list of performance domains to the root domain. */
 418         tmp = rd->pd;
 419         rcu_assign_pointer(rd->pd, pd);
 420         if (tmp)
 421                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 422
 423         return !!pd;
 424
 425 free:
 426         free_pd(pd);
 427         tmp = rd->pd;
 428         rcu_assign_pointer(rd->pd, NULL);
 429         if (tmp)
 430                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 431
 432         return false;
 433 }
 434 #else
 435 static void free_pd(struct perf_domain *pd) { }
 436 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
 437
 438 static void free_rootdomain(struct rcu_head *rcu)
 439 {
 440         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 441
 442         cpupri_cleanup(&rd->cpupri);
 443         cpudl_cleanup(&rd->cpudl);
 444         free_cpumask_var(rd->dlo_mask);
 445         free_cpumask_var(rd->rto_mask);
 446         free_cpumask_var(rd->online);
 447         free_cpumask_var(rd->span);
 448         free_pd(rd->pd);
 449         kfree(rd);
 450 }
 451
 452 void rq_attach_root(struct rq *rq, struct root_domain *rd)
 453 {
 454         struct root_domain *old_rd = NULL;
 455         unsigned long flags;
 456
 457         raw_spin_lock_irqsave(&rq->lock, flags);
 458
 459         if (rq->rd) {
 460                 old_rd = rq->rd;
 461
 462                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
 463                         set_rq_offline(rq);
 464
 465                 cpumask_clear_cpu(rq->cpu, old_rd->span);
 466
 467                 /*
 468                  * If we dont want to free the old_rd yet then
 469                  * set old_rd to NULL to skip the freeing later
 470                  * in this function:
 471                  */
 472                 if (!atomic_dec_and_test(&old_rd->refcount))
 473                         old_rd = NULL;
 474         }
 475
 476         atomic_inc(&rd->refcount);
 477         rq->rd = rd;
 478
 479         cpumask_set_cpu(rq->cpu, rd->span);
 480         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 481                 set_rq_online(rq);
 482
 483         raw_spin_unlock_irqrestore(&rq->lock, flags);
 484
 485         if (old_rd)
 486                 call_rcu(&old_rd->rcu, free_rootdomain);
 487 }
 488
 489 void sched_get_rd(struct root_domain *rd)
 490 {
 491         atomic_inc(&rd->refcount);
 492 }
 493
 494 void sched_put_rd(struct root_domain *rd)
 495 {
 496         if (!atomic_dec_and_test(&rd->refcount))
 497                 return;
 498
 499         call_rcu(&rd->rcu, free_rootdomain);
 500 }
 501
 502 static int init_rootdomain(struct root_domain *rd)
 503 {
 504         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 505                 goto out;
 506         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 507                 goto free_span;
 508         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 509                 goto free_online;
 510         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 511                 goto free_dlo_mask;
 512
 513 #ifdef HAVE_RT_PUSH_IPI
 514         rd->rto_cpu = -1;
 515         raw_spin_lock_init(&rd->rto_lock);
 516         init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
 517 #endif
 518
 519         init_dl_bw(&rd->dl_bw);
 520         if (cpudl_init(&rd->cpudl) != 0)
 521                 goto free_rto_mask;
 522
 523         if (cpupri_init(&rd->cpupri) != 0)
 524                 goto free_cpudl;
 525         return 0;
 526
 527 free_cpudl:
 528         cpudl_cleanup(&rd->cpudl);
 529 free_rto_mask:
 530         free_cpumask_var(rd->rto_mask);
 531 free_dlo_mask:
 532         free_cpumask_var(rd->dlo_mask);
 533 free_online:
 534         free_cpumask_var(rd->online);
 535 free_span:
 536         free_cpumask_var(rd->span);
 537 out:
 538         return -ENOMEM;
 539 }
 540
 541 /*
 542  * By default the system creates a single root-domain with all CPUs as
 543  * members (mimicking the global state we have today).
 544  */
 545 struct root_domain def_root_domain;
 546
 547 void init_defrootdomain(void)
 548 {
 549         init_rootdomain(&def_root_domain);
 550
 551         atomic_set(&def_root_domain.refcount, 1);
 552 }
 553
 554 static struct root_domain *alloc_rootdomain(void)
 555 {
 556         struct root_domain *rd;
 557
 558         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 559         if (!rd)
 560                 return NULL;
 561
 562         if (init_rootdomain(rd) != 0) {
 563                 kfree(rd);
 564                 return NULL;
 565         }
 566
 567         return rd;
 568 }
 569
 570 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 571 {
 572         struct sched_group *tmp, *first;
 573
 574         if (!sg)
 575                 return;
 576
 577         first = sg;
 578         do {
 579                 tmp = sg->next;
 580
 581                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 582                         kfree(sg->sgc);
 583
 584                 if (atomic_dec_and_test(&sg->ref))
 585                         kfree(sg);
 586                 sg = tmp;
 587         } while (sg != first);
 588 }
 589
 590 static void destroy_sched_domain(struct sched_domain *sd)
 591 {
 592         /*
 593          * A normal sched domain may have multiple group references, an
 594          * overlapping domain, having private groups, only one.  Iterate,
 595          * dropping group/capacity references, freeing where none remain.
 596          */
 597         free_sched_groups(sd->groups, 1);
 598
 599         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 600                 kfree(sd->shared);
 601         kfree(sd);
 602 }
 603
 604 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 605 {
 606         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 607
 608         while (sd) {
 609                 struct sched_domain *parent = sd->parent;
 610                 destroy_sched_domain(sd);
 611                 sd = parent;
 612         }
 613 }
 614
 615 static void destroy_sched_domains(struct sched_domain *sd)
 616 {
 617         if (sd)
 618                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 619 }
 620
 621 /*
 622  * Keep a special pointer to the highest sched_domain that has
 623  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
 624  * allows us to avoid some pointer chasing select_idle_sibling().
 625  *
 626  * Also keep a unique ID per domain (we use the first CPU number in
 627  * the cpumask of the domain), this allows us to quickly tell if
 628  * two CPUs are in the same cache domain, see cpus_share_cache().
 629  */
 630 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 631 DEFINE_PER_CPU(int, sd_llc_size);
 632 DEFINE_PER_CPU(int, sd_llc_id);
 633 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 634 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 635 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 636 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 637 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 638
 639 static void update_top_cache_domain(int cpu)
 640 {
 641         struct sched_domain_shared *sds = NULL;
 642         struct sched_domain *sd;
 643         int id = cpu;
 644         int size = 1;
 645
 646         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 647         if (sd) {
 648                 id = cpumask_first(sched_domain_span(sd));
 649                 size = cpumask_weight(sched_domain_span(sd));
 650                 sds = sd->shared;
 651         }
 652
 653         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 654         per_cpu(sd_llc_size, cpu) = size;
 655         per_cpu(sd_llc_id, cpu) = id;
 656         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 657
 658         sd = lowest_flag_domain(cpu, SD_NUMA);
 659         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 660
 661         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 662         rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 663
 664         sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
 665         rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 666 }
 667
 668 /*
 669  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 670  * hold the hotplug lock.
 671  */
 672 static void
 673 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 674 {
 675         struct rq *rq = cpu_rq(cpu);
 676         struct sched_domain *tmp;
 677
 678         /* Remove the sched domains which do not contribute to scheduling. */
 679         for (tmp = sd; tmp; ) {
 680                 struct sched_domain *parent = tmp->parent;
 681                 if (!parent)
 682                         break;
 683
 684                 if (sd_parent_degenerate(tmp, parent)) {
 685                         tmp->parent = parent->parent;
 686                         if (parent->parent)
 687                                 parent->parent->child = tmp;
 688                         /*
 689                          * Transfer SD_PREFER_SIBLING down in case of a
 690                          * degenerate parent; the spans match for this
 691                          * so the property transfers.
 692                          */
 693                         if (parent->flags & SD_PREFER_SIBLING)
 694                                 tmp->flags |= SD_PREFER_SIBLING;
 695                         destroy_sched_domain(parent);
 696                 } else
 697                         tmp = tmp->parent;
 698         }
 699
 700         if (sd && sd_degenerate(sd)) {
 701                 tmp = sd;
 702                 sd = sd->parent;
 703                 destroy_sched_domain(tmp);
 704                 if (sd)
 705                         sd->child = NULL;
 706         }
 707
 708         sched_domain_debug(sd, cpu);
 709
 710         rq_attach_root(rq, rd);
 711         tmp = rq->sd;
 712         rcu_assign_pointer(rq->sd, sd);
 713         dirty_sched_domain_sysctl(cpu);
 714         destroy_sched_domains(tmp);
 715
 716         update_top_cache_domain(cpu);
 717 }
 718
 719 struct s_data {
 720         struct sched_domain * __percpu *sd;
 721         struct root_domain      *rd;
 722 };
 723
 724 enum s_alloc {
 725         sa_rootdomain,
 726         sa_sd,
 727         sa_sd_storage,
 728         sa_none,
 729 };
 730
 731 /*
 732  * Return the canonical balance CPU for this group, this is the first CPU
 733  * of this group that's also in the balance mask.
 734  *
 735  * The balance mask are all those CPUs that could actually end up at this
 736  * group. See build_balance_mask().
 737  *
 738  * Also see should_we_balance().
 739  */
 740 int group_balance_cpu(struct sched_group *sg)
 741 {
 742         return cpumask_first(group_balance_mask(sg));
 743 }
 744
 745
 746 /*
 747  * NUMA topology (first read the regular topology blurb below)
 748  *
 749  * Given a node-distance table, for example:
 750  *
 751  *   node   0   1   2   3
 752  *     0:  10  20  30  20
 753  *     1:  20  10  20  30
 754  *     2:  30  20  10  20
 755  *     3:  20  30  20  10
 756  *
 757  * which represents a 4 node ring topology like:
 758  *
 759  *   0 ----- 1
 760  *   |       |
 761  *   |       |
 762  *   |       |
 763  *   3 ----- 2
 764  *
 765  * We want to construct domains and groups to represent this. The way we go
 766  * about doing this is to build the domains on 'hops'. For each NUMA level we
 767  * construct the mask of all nodes reachable in @level hops.
 768  *
 769  * For the above NUMA topology that gives 3 levels:
 770  *
 771  * NUMA-2       0-3             0-3             0-3             0-3
 772  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 773  *
 774  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 775  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 776  *
 777  * NUMA-0       0               1               2               3
 778  *
 779  *
 780  * As can be seen; things don't nicely line up as with the regular topology.
 781  * When we iterate a domain in child domain chunks some nodes can be
 782  * represented multiple times -- hence the "overlap" naming for this part of
 783  * the topology.
 784  *
 785  * In order to minimize this overlap, we only build enough groups to cover the
 786  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 787  *
 788  * Because:
 789  *
 790  *  - the first group of each domain is its child domain; this
 791  *    gets us the first 0-1,3
 792  *  - the only uncovered node is 2, who's child domain is 1-3.
 793  *
 794  * However, because of the overlap, computing a unique CPU for each group is
 795  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 796  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 797  * end up at those groups (they would end up in group: 0-1,3).
 798  *
 799  * To correct this we have to introduce the group balance mask. This mask
 800  * will contain those CPUs in the group that can reach this group given the
 801  * (child) domain tree.
 802  *
 803  * With this we can once again compute balance_cpu and sched_group_capacity
 804  * relations.
 805  *
 806  * XXX include words on how balance_cpu is unique and therefore can be
 807  * used for sched_group_capacity links.
 808  *
 809  *
 810  * Another 'interesting' topology is:
 811  *
 812  *   node   0   1   2   3
 813  *     0:  10  20  20  30
 814  *     1:  20  10  20  20
 815  *     2:  20  20  10  20
 816  *     3:  30  20  20  10
 817  *
 818  * Which looks a little like:
 819  *
 820  *   0 ----- 1
 821  *   |     / |
 822  *   |   /   |
 823  *   | /     |
 824  *   2 ----- 3
 825  *
 826  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 827  * are not.
 828  *
 829  * This leads to a few particularly weird cases where the sched_domain's are
 830  * not of the same number for each CPU. Consider:
 831  *
 832  * NUMA-2       0-3                                             0-3
 833  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 834  *
 835  * NUMA-1       0-2             0-3             0-3             1-3
 836  *
 837  * NUMA-0       0               1               2               3
 838  *
 839  */
 840
 841
 842 /*
 843  * Build the balance mask; it contains only those CPUs that can arrive at this
 844  * group and should be considered to continue balancing.
 845  *
 846  * We do this during the group creation pass, therefore the group information
 847  * isn't complete yet, however since each group represents a (child) domain we
 848  * can fully construct this using the sched_domain bits (which are already
 849  * complete).
 850  */
 851 static void
 852 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 853 {
 854         const struct cpumask *sg_span = sched_group_span(sg);
 855         struct sd_data *sdd = sd->private;
 856         struct sched_domain *sibling;
 857         int i;
 858
 859         cpumask_clear(mask);
 860
 861         for_each_cpu(i, sg_span) {
 862                 sibling = *per_cpu_ptr(sdd->sd, i);
 863
 864                 /*
 865                  * Can happen in the asymmetric case, where these siblings are
 866                  * unused. The mask will not be empty because those CPUs that
 867                  * do have the top domain _should_ span the domain.
 868                  */
 869                 if (!sibling->child)
 870                         continue;
 871
 872                 /* If we would not end up here, we can't continue from here */
 873                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 874                         continue;
 875
 876                 cpumask_set_cpu(i, mask);
 877         }
 878
 879         /* We must not have empty masks here */
 880         WARN_ON_ONCE(cpumask_empty(mask));
 881 }
 882
 883 /*
 884  * XXX: This creates per-node group entries; since the load-balancer will
 885  * immediately access remote memory to construct this group's load-balance
 886  * statistics having the groups node local is of dubious benefit.
 887  */
 888 static struct sched_group *
 889 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 890 {
 891         struct sched_group *sg;
 892         struct cpumask *sg_span;
 893
 894         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 895                         GFP_KERNEL, cpu_to_node(cpu));
 896
 897         if (!sg)
 898                 return NULL;
 899
 900         sg_span = sched_group_span(sg);
 901         if (sd->child)
 902                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 903         else
 904                 cpumask_copy(sg_span, sched_domain_span(sd));
 905
 906         atomic_inc(&sg->ref);
 907         return sg;
 908 }
 909
 910 static void init_overlap_sched_group(struct sched_domain *sd,
 911                                      struct sched_group *sg)
 912 {
 913         struct cpumask *mask = sched_domains_tmpmask2;
 914         struct sd_data *sdd = sd->private;
 915         struct cpumask *sg_span;
 916         int cpu;
 917
 918         build_balance_mask(sd, sg, mask);
 919         cpu = cpumask_first_and(sched_group_span(sg), mask);
 920
 921         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 922         if (atomic_inc_return(&sg->sgc->ref) == 1)
 923                 cpumask_copy(group_balance_mask(sg), mask);
 924         else
 925                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 926
 927         /*
 928          * Initialize sgc->capacity such that even if we mess up the
 929          * domains and no possible iteration will get us here, we won't
 930          * die on a /0 trap.
 931          */
 932         sg_span = sched_group_span(sg);
 933         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 934         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 935         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 936 }
 937
 938 static int
 939 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
 940 {
 941         struct sched_group *first = NULL, *last = NULL, *sg;
 942         const struct cpumask *span = sched_domain_span(sd);
 943         struct cpumask *covered = sched_domains_tmpmask;
 944         struct sd_data *sdd = sd->private;
 945         struct sched_domain *sibling;
 946         int i;
 947
 948         cpumask_clear(covered);
 949
 950         for_each_cpu_wrap(i, span, cpu) {
 951                 struct cpumask *sg_span;
 952
 953                 if (cpumask_test_cpu(i, covered))
 954                         continue;
 955
 956                 sibling = *per_cpu_ptr(sdd->sd, i);
 957
 958                 /*
 959                  * Asymmetric node setups can result in situations where the
 960                  * domain tree is of unequal depth, make sure to skip domains
 961                  * that already cover the entire range.
 962                  *
 963                  * In that case build_sched_domains() will have terminated the
 964                  * iteration early and our sibling sd spans will be empty.
 965                  * Domains should always include the CPU they're built on, so
 966                  * check that.
 967                  */
 968                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
 969                         continue;
 970
 971                 sg = build_group_from_child_sched_domain(sibling, cpu);
 972                 if (!sg)
 973                         goto fail;
 974
 975                 sg_span = sched_group_span(sg);
 976                 cpumask_or(covered, covered, sg_span);
 977
 978                 init_overlap_sched_group(sd, sg);
 979
 980                 if (!first)
 981                         first = sg;
 982                 if (last)
 983                         last->next = sg;
 984                 last = sg;
 985                 last->next = first;
 986         }
 987         sd->groups = first;
 988
 989         return 0;
 990
 991 fail:
 992         free_sched_groups(first, 0);
 993
 994         return -ENOMEM;
 995 }
 996
 997
 998 /*
 999  * Package topology (also see the load-balance blurb in fair.c)
1000  *
1001  * The scheduler builds a tree structure to represent a number of important
1002  * topology features. By default (default_topology[]) these include:
1003  *
1004  *  - Simultaneous multithreading (SMT)
1005  *  - Multi-Core Cache (MC)
1006  *  - Package (DIE)
1007  *
1008  * Where the last one more or less denotes everything up to a NUMA node.
1009  *
1010  * The tree consists of 3 primary data structures:
1011  *
1012  *      sched_domain -> sched_group -> sched_group_capacity
1013  *          ^ ^             ^ ^
1014  *          `-'             `-'
1015  *
1016  * The sched_domains are per-CPU and have a two way link (parent & child) and
1017  * denote the ever growing mask of CPUs belonging to that level of topology.
1018  *
1019  * Each sched_domain has a circular (double) linked list of sched_group's, each
1020  * denoting the domains of the level below (or individual CPUs in case of the
1021  * first domain level). The sched_group linked by a sched_domain includes the
1022  * CPU of that sched_domain [*].
1023  *
1024  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1025  *
1026  * CPU   0   1   2   3   4   5   6   7
1027  *
1028  * DIE  [                             ]
1029  * MC   [             ] [             ]
1030  * SMT  [     ] [     ] [     ] [     ]
1031  *
1032  *  - or -
1033  *
1034  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1035  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1036  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1037  *
1038  * CPU   0   1   2   3   4   5   6   7
1039  *
1040  * One way to think about it is: sched_domain moves you up and down among these
1041  * topology levels, while sched_group moves you sideways through it, at child
1042  * domain granularity.
1043  *
1044  * sched_group_capacity ensures each unique sched_group has shared storage.
1045  *
1046  * There are two related construction problems, both require a CPU that
1047  * uniquely identify each group (for a given domain):
1048  *
1049  *  - The first is the balance_cpu (see should_we_balance() and the
1050  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1051  *    continue balancing at a higher domain.
1052  *
1053  *  - The second is the sched_group_capacity; we want all identical groups
1054  *    to share a single sched_group_capacity.
1055  *
1056  * Since these topologies are exclusive by construction. That is, its
1057  * impossible for an SMT thread to belong to multiple cores, and cores to
1058  * be part of multiple caches. There is a very clear and unique location
1059  * for each CPU in the hierarchy.
1060  *
1061  * Therefore computing a unique CPU for each group is trivial (the iteration
1062  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1063  * group), we can simply pick the first CPU in each group.
1064  *
1065  *
1066  * [*] in other words, the first group of each domain is its child domain.
1067  */
1068
1069 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1070 {
1071         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1072         struct sched_domain *child = sd->child;
1073         struct sched_group *sg;
1074         bool already_visited;
1075
1076         if (child)
1077                 cpu = cpumask_first(sched_domain_span(child));
1078
1079         sg = *per_cpu_ptr(sdd->sg, cpu);
1080         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1081
1082         /* Increase refcounts for claim_allocations: */
1083         already_visited = atomic_inc_return(&sg->ref) > 1;
1084         /* sgc visits should follow a similar trend as sg */
1085         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1086
1087         /* If we have already visited that group, it's already initialized. */
1088         if (already_visited)
1089                 return sg;
1090
1091         if (child) {
1092                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1093                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1094         } else {
1095                 cpumask_set_cpu(cpu, sched_group_span(sg));
1096                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1097         }
1098
1099         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1100         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1101         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1102
1103         return sg;
1104 }
1105
1106 /*
1107  * build_sched_groups will build a circular linked list of the groups
1108  * covered by the given span, will set each group's ->cpumask correctly,
1109  * and will initialize their ->sgc.
1110  *
1111  * Assumes the sched_domain tree is fully constructed
1112  */
1113 static int
1114 build_sched_groups(struct sched_domain *sd, int cpu)
1115 {
1116         struct sched_group *first = NULL, *last = NULL;
1117         struct sd_data *sdd = sd->private;
1118         const struct cpumask *span = sched_domain_span(sd);
1119         struct cpumask *covered;
1120         int i;
1121
1122         lockdep_assert_held(&sched_domains_mutex);
1123         covered = sched_domains_tmpmask;
1124
1125         cpumask_clear(covered);
1126
1127         for_each_cpu_wrap(i, span, cpu) {
1128                 struct sched_group *sg;
1129
1130                 if (cpumask_test_cpu(i, covered))
1131                         continue;
1132
1133                 sg = get_group(i, sdd);
1134
1135                 cpumask_or(covered, covered, sched_group_span(sg));
1136
1137                 if (!first)
1138                         first = sg;
1139                 if (last)
1140                         last->next = sg;
1141                 last = sg;
1142         }
1143         last->next = first;
1144         sd->groups = first;
1145
1146         return 0;
1147 }
1148
1149 /*
1150  * Initialize sched groups cpu_capacity.
1151  *
1152  * cpu_capacity indicates the capacity of sched group, which is used while
1153  * distributing the load between different sched groups in a sched domain.
1154  * Typically cpu_capacity for all the groups in a sched domain will be same
1155  * unless there are asymmetries in the topology. If there are asymmetries,
1156  * group having more cpu_capacity will pickup more load compared to the
1157  * group having less cpu_capacity.
1158  */
1159 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1160 {
1161         struct sched_group *sg = sd->groups;
1162
1163         WARN_ON(!sg);
1164
1165         do {
1166                 int cpu, max_cpu = -1;
1167
1168                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1169
1170                 if (!(sd->flags & SD_ASYM_PACKING))
1171                         goto next;
1172
1173                 for_each_cpu(cpu, sched_group_span(sg)) {
1174                         if (max_cpu < 0)
1175                                 max_cpu = cpu;
1176                         else if (sched_asym_prefer(cpu, max_cpu))
1177                                 max_cpu = cpu;
1178                 }
1179                 sg->asym_prefer_cpu = max_cpu;
1180
1181 next:
1182                 sg = sg->next;
1183         } while (sg != sd->groups);
1184
1185         if (cpu != group_balance_cpu(sg))
1186                 return;
1187
1188         update_group_capacity(sd, cpu);
1189 }
1190
1191 /*
1192  * Initializers for schedule domains
1193  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1194  */
1195
1196 static int default_relax_domain_level = -1;
1197 int sched_domain_level_max;
1198
1199 static int __init setup_relax_domain_level(char *str)
1200 {
1201         if (kstrtoint(str, 0, &default_relax_domain_level))
1202                 pr_warn("Unable to set relax_domain_level\n");
1203
1204         return 1;
1205 }
1206 __setup("relax_domain_level=", setup_relax_domain_level);
1207
1208 static void set_domain_attribute(struct sched_domain *sd,
1209                                  struct sched_domain_attr *attr)
1210 {
1211         int request;
1212
1213         if (!attr || attr->relax_domain_level < 0) {
1214                 if (default_relax_domain_level < 0)
1215                         return;
1216                 request = default_relax_domain_level;
1217         } else
1218                 request = attr->relax_domain_level;
1219
1220         if (sd->level > request) {
1221                 /* Turn off idle balance on this domain: */
1222                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1223         }
1224 }
1225
1226 static void __sdt_free(const struct cpumask *cpu_map);
1227 static int __sdt_alloc(const struct cpumask *cpu_map);
1228
1229 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1230                                  const struct cpumask *cpu_map)
1231 {
1232         switch (what) {
1233         case sa_rootdomain:
1234                 if (!atomic_read(&d->rd->refcount))
1235                         free_rootdomain(&d->rd->rcu);
1236                 fallthrough;
1237         case sa_sd:
1238                 free_percpu(d->sd);
1239                 fallthrough;
1240         case sa_sd_storage:
1241                 __sdt_free(cpu_map);
1242                 fallthrough;
1243         case sa_none:
1244                 break;
1245         }
1246 }
1247
1248 static enum s_alloc
1249 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1250 {
1251         memset(d, 0, sizeof(*d));
1252
1253         if (__sdt_alloc(cpu_map))
1254                 return sa_sd_storage;
1255         d->sd = alloc_percpu(struct sched_domain *);
1256         if (!d->sd)
1257                 return sa_sd_storage;
1258         d->rd = alloc_rootdomain();
1259         if (!d->rd)
1260                 return sa_sd;
1261
1262         return sa_rootdomain;
1263 }
1264
1265 /*
1266  * NULL the sd_data elements we've used to build the sched_domain and
1267  * sched_group structure so that the subsequent __free_domain_allocs()
1268  * will not free the data we're using.
1269  */
1270 static void claim_allocations(int cpu, struct sched_domain *sd)
1271 {
1272         struct sd_data *sdd = sd->private;
1273
1274         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1275         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1276
1277         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1278                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1279
1280         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1281                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1282
1283         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1284                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1285 }
1286
1287 #ifdef CONFIG_NUMA
1288 enum numa_topology_type sched_numa_topology_type;
1289
1290 static int                      sched_domains_numa_levels;
1291 static int                      sched_domains_curr_level;
1292
1293 int                             sched_max_numa_distance;
1294 static int                      *sched_domains_numa_distance;
1295 static struct cpumask           ***sched_domains_numa_masks;
1296 int __read_mostly               node_reclaim_distance = RECLAIM_DISTANCE;
1297 #endif
1298
1299 /*
1300  * SD_flags allowed in topology descriptions.
1301  *
1302  * These flags are purely descriptive of the topology and do not prescribe
1303  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1304  * function:
1305  *
1306  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1307  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1308  *   SD_NUMA                - describes NUMA topologies
1309  *
1310  * Odd one out, which beside describing the topology has a quirk also
1311  * prescribes the desired behaviour that goes along with it:
1312  *
1313  *   SD_ASYM_PACKING        - describes SMT quirks
1314  */
1315 #define TOPOLOGY_SD_FLAGS               \
1316         (SD_SHARE_CPUCAPACITY   |       \
1317          SD_SHARE_PKG_RESOURCES |       \
1318          SD_NUMA                |       \
1319          SD_ASYM_PACKING)
1320
1321 static struct sched_domain *
1322 sd_init(struct sched_domain_topology_level *tl,
1323         const struct cpumask *cpu_map,
1324         struct sched_domain *child, int dflags, int cpu)
1325 {
1326         struct sd_data *sdd = &tl->data;
1327         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1328         int sd_id, sd_weight, sd_flags = 0;
1329
1330 #ifdef CONFIG_NUMA
1331         /*
1332          * Ugly hack to pass state to sd_numa_mask()...
1333          */
1334         sched_domains_curr_level = tl->numa_level;
1335 #endif
1336
1337         sd_weight = cpumask_weight(tl->mask(cpu));
1338
1339         if (tl->sd_flags)
1340                 sd_flags = (*tl->sd_flags)();
1341         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1342                         "wrong sd_flags in topology description\n"))
1343                 sd_flags &= TOPOLOGY_SD_FLAGS;
1344
1345         /* Apply detected topology flags */
1346         sd_flags |= dflags;
1347
1348         *sd = (struct sched_domain){
1349                 .min_interval           = sd_weight,
1350                 .max_interval           = 2*sd_weight,
1351                 .busy_factor            = 16,
1352                 .imbalance_pct          = 117,
1353
1354                 .cache_nice_tries       = 0,
1355
1356                 .flags                  = 1*SD_BALANCE_NEWIDLE
1357                                         | 1*SD_BALANCE_EXEC
1358                                         | 1*SD_BALANCE_FORK
1359                                         | 0*SD_BALANCE_WAKE
1360                                         | 1*SD_WAKE_AFFINE
1361                                         | 0*SD_SHARE_CPUCAPACITY
1362                                         | 0*SD_SHARE_PKG_RESOURCES
1363                                         | 0*SD_SERIALIZE
1364                                         | 1*SD_PREFER_SIBLING
1365                                         | 0*SD_NUMA
1366                                         | sd_flags
1367                                         ,
1368
1369                 .last_balance           = jiffies,
1370                 .balance_interval       = sd_weight,
1371                 .max_newidle_lb_cost    = 0,
1372                 .next_decay_max_lb_cost = jiffies,
1373                 .child                  = child,
1374 #ifdef CONFIG_SCHED_DEBUG
1375                 .name                   = tl->name,
1376 #endif
1377         };
1378
1379         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1380         sd_id = cpumask_first(sched_domain_span(sd));
1381
1382         /*
1383          * Convert topological properties into behaviour.
1384          */
1385
1386         /* Don't attempt to spread across CPUs of different capacities. */
1387         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1388                 sd->child->flags &= ~SD_PREFER_SIBLING;
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                 int dflags = 0;
2005
2006                 sd = NULL;
2007                 for_each_sd_topology(tl) {
2008                         if (tl == tl_asym) {
2009                                 dflags |= SD_ASYM_CPUCAPACITY;
2010                                 has_asym = true;
2011                         }
2012
2013                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2014                                 goto error;
2015
2016                         sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2017
2018                         if (tl == sched_domain_topology)
2019                                 *per_cpu_ptr(d.sd, i) = sd;
2020                         if (tl->flags & SDTL_OVERLAP)
2021                                 sd->flags |= SD_OVERLAP;
2022                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2023                                 break;
2024                 }
2025         }
2026
2027         /* Build the groups for the domains */
2028         for_each_cpu(i, cpu_map) {
2029                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2030                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2031                         if (sd->flags & SD_OVERLAP) {
2032                                 if (build_overlap_sched_groups(sd, i))
2033                                         goto error;
2034                         } else {
2035                                 if (build_sched_groups(sd, i))
2036                                         goto error;
2037                         }
2038                 }
2039         }
2040
2041         /* Calculate CPU capacity for physical packages and nodes */
2042         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2043                 if (!cpumask_test_cpu(i, cpu_map))
2044                         continue;
2045
2046                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2047                         claim_allocations(i, sd);
2048                         init_sched_groups_capacity(i, sd);
2049                 }
2050         }
2051
2052         /* Attach the domains */
2053         rcu_read_lock();
2054         for_each_cpu(i, cpu_map) {
2055                 rq = cpu_rq(i);
2056                 sd = *per_cpu_ptr(d.sd, i);
2057
2058                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2059                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2060                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2061
2062                 cpu_attach_domain(sd, d.rd, i);
2063         }
2064         rcu_read_unlock();
2065
2066         if (has_asym)
2067                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2068
2069         if (rq && sched_debug_enabled) {
2070                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2071                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2072         }
2073
2074         ret = 0;
2075 error:
2076         __free_domain_allocs(&d, alloc_state, cpu_map);
2077
2078         return ret;
2079 }
2080
2081 /* Current sched domains: */
2082 static cpumask_var_t                    *doms_cur;
2083
2084 /* Number of sched domains in 'doms_cur': */
2085 static int                              ndoms_cur;
2086
2087 /* Attribues of custom domains in 'doms_cur' */
2088 static struct sched_domain_attr         *dattr_cur;
2089
2090 /*
2091  * Special case: If a kmalloc() of a doms_cur partition (array of
2092  * cpumask) fails, then fallback to a single sched domain,
2093  * as determined by the single cpumask fallback_doms.
2094  */
2095 static cpumask_var_t                    fallback_doms;
2096
2097 /*
2098  * arch_update_cpu_topology lets virtualized architectures update the
2099  * CPU core maps. It is supposed to return 1 if the topology changed
2100  * or 0 if it stayed the same.
2101  */
2102 int __weak arch_update_cpu_topology(void)
2103 {
2104         return 0;
2105 }
2106
2107 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2108 {
2109         int i;
2110         cpumask_var_t *doms;
2111
2112         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2113         if (!doms)
2114                 return NULL;
2115         for (i = 0; i < ndoms; i++) {
2116                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2117                         free_sched_domains(doms, i);
2118                         return NULL;
2119                 }
2120         }
2121         return doms;
2122 }
2123
2124 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2125 {
2126         unsigned int i;
2127         for (i = 0; i < ndoms; i++)
2128                 free_cpumask_var(doms[i]);
2129         kfree(doms);
2130 }
2131
2132 /*
2133  * Set up scheduler domains and groups.  For now this just excludes isolated
2134  * CPUs, but could be used to exclude other special cases in the future.
2135  */
2136 int sched_init_domains(const struct cpumask *cpu_map)
2137 {
2138         int err;
2139
2140         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2141         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2142         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2143
2144         arch_update_cpu_topology();
2145         ndoms_cur = 1;
2146         doms_cur = alloc_sched_domains(ndoms_cur);
2147         if (!doms_cur)
2148                 doms_cur = &fallback_doms;
2149         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2150         err = build_sched_domains(doms_cur[0], NULL);
2151         register_sched_domain_sysctl();
2152
2153         return err;
2154 }
2155
2156 /*
2157  * Detach sched domains from a group of CPUs specified in cpu_map
2158  * These CPUs will now be attached to the NULL domain
2159  */
2160 static void detach_destroy_domains(const struct cpumask *cpu_map)
2161 {
2162         unsigned int cpu = cpumask_any(cpu_map);
2163         int i;
2164
2165         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2166                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2167
2168         rcu_read_lock();
2169         for_each_cpu(i, cpu_map)
2170                 cpu_attach_domain(NULL, &def_root_domain, i);
2171         rcu_read_unlock();
2172 }
2173
2174 /* handle null as "default" */
2175 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2176                         struct sched_domain_attr *new, int idx_new)
2177 {
2178         struct sched_domain_attr tmp;
2179
2180         /* Fast path: */
2181         if (!new && !cur)
2182                 return 1;
2183
2184         tmp = SD_ATTR_INIT;
2185
2186         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2187                         new ? (new + idx_new) : &tmp,
2188                         sizeof(struct sched_domain_attr));
2189 }
2190
2191 /*
2192  * Partition sched domains as specified by the 'ndoms_new'
2193  * cpumasks in the array doms_new[] of cpumasks. This compares
2194  * doms_new[] to the current sched domain partitioning, doms_cur[].
2195  * It destroys each deleted domain and builds each new domain.
2196  *
2197  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2198  * The masks don't intersect (don't overlap.) We should setup one
2199  * sched domain for each mask. CPUs not in any of the cpumasks will
2200  * not be load balanced. If the same cpumask appears both in the
2201  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2202  * it as it is.
2203  *
2204  * The passed in 'doms_new' should be allocated using
2205  * alloc_sched_domains.  This routine takes ownership of it and will
2206  * free_sched_domains it when done with it. If the caller failed the
2207  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2208  * and partition_sched_domains() will fallback to the single partition
2209  * 'fallback_doms', it also forces the domains to be rebuilt.
2210  *
2211  * If doms_new == NULL it will be replaced with cpu_online_mask.
2212  * ndoms_new == 0 is a special case for destroying existing domains,
2213  * and it will not create the default domain.
2214  *
2215  * Call with hotplug lock and sched_domains_mutex held
2216  */
2217 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2218                                     struct sched_domain_attr *dattr_new)
2219 {
2220         bool __maybe_unused has_eas = false;
2221         int i, j, n;
2222         int new_topology;
2223
2224         lockdep_assert_held(&sched_domains_mutex);
2225
2226         /* Always unregister in case we don't destroy any domains: */
2227         unregister_sched_domain_sysctl();
2228
2229         /* Let the architecture update CPU core mappings: */
2230         new_topology = arch_update_cpu_topology();
2231
2232         if (!doms_new) {
2233                 WARN_ON_ONCE(dattr_new);
2234                 n = 0;
2235                 doms_new = alloc_sched_domains(1);
2236                 if (doms_new) {
2237                         n = 1;
2238                         cpumask_and(doms_new[0], cpu_active_mask,
2239                                     housekeeping_cpumask(HK_FLAG_DOMAIN));
2240                 }
2241         } else {
2242                 n = ndoms_new;
2243         }
2244
2245         /* Destroy deleted domains: */
2246         for (i = 0; i < ndoms_cur; i++) {
2247                 for (j = 0; j < n && !new_topology; j++) {
2248                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2249                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2250                                 struct root_domain *rd;
2251
2252                                 /*
2253                                  * This domain won't be destroyed and as such
2254                                  * its dl_bw->total_bw needs to be cleared.  It
2255                                  * will be recomputed in function
2256                                  * update_tasks_root_domain().
2257                                  */
2258                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2259                                 dl_clear_root_domain(rd);
2260                                 goto match1;
2261                         }
2262                 }
2263                 /* No match - a current sched domain not in new doms_new[] */
2264                 detach_destroy_domains(doms_cur[i]);
2265 match1:
2266                 ;
2267         }
2268
2269         n = ndoms_cur;
2270         if (!doms_new) {
2271                 n = 0;
2272                 doms_new = &fallback_doms;
2273                 cpumask_and(doms_new[0], cpu_active_mask,
2274                             housekeeping_cpumask(HK_FLAG_DOMAIN));
2275         }
2276
2277         /* Build new domains: */
2278         for (i = 0; i < ndoms_new; i++) {
2279                 for (j = 0; j < n && !new_topology; j++) {
2280                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2281                             dattrs_equal(dattr_new, i, dattr_cur, j))
2282                                 goto match2;
2283                 }
2284                 /* No match - add a new doms_new */
2285                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2286 match2:
2287                 ;
2288         }
2289
2290 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2291         /* Build perf. domains: */
2292         for (i = 0; i < ndoms_new; i++) {
2293                 for (j = 0; j < n && !sched_energy_update; j++) {
2294                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2295                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2296                                 has_eas = true;
2297                                 goto match3;
2298                         }
2299                 }
2300                 /* No match - add perf. domains for a new rd */
2301                 has_eas |= build_perf_domains(doms_new[i]);
2302 match3:
2303                 ;
2304         }
2305         sched_energy_set(has_eas);
2306 #endif
2307
2308         /* Remember the new sched domains: */
2309         if (doms_cur != &fallback_doms)
2310                 free_sched_domains(doms_cur, ndoms_cur);
2311
2312         kfree(dattr_cur);
2313         doms_cur = doms_new;
2314         dattr_cur = dattr_new;
2315         ndoms_cur = ndoms_new;
2316
2317         register_sched_domain_sysctl();
2318 }
2319
2320 /*
2321  * Call with hotplug lock held
2322  */
2323 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2324                              struct sched_domain_attr *dattr_new)
2325 {
2326         mutex_lock(&sched_domains_mutex);
2327         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2328         mutex_unlock(&sched_domains_mutex);
2329 }