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