kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5
   6 #include <linux/bsearch.h>
   7
   8 DEFINE_MUTEX(sched_domains_mutex);
   9
  10 /* Protected by sched_domains_mutex: */
  11 static cpumask_var_t sched_domains_tmpmask;
  12 static cpumask_var_t sched_domains_tmpmask2;
  13
  14 #ifdef CONFIG_SCHED_DEBUG
  15
  16 static int __init sched_debug_setup(char *str)
  17 {
  18         sched_debug_verbose = true;
  19
  20         return 0;
  21 }
  22 early_param("sched_verbose", sched_debug_setup);
  23
  24 static inline bool sched_debug(void)
  25 {
  26         return sched_debug_verbose;
  27 }
  28
  29 #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
  30 const struct sd_flag_debug sd_flag_debug[] = {
  31 #include <linux/sched/sd_flags.h>
  32 };
  33 #undef SD_FLAG
  34
  35 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
  36                                   struct cpumask *groupmask)
  37 {
  38         struct sched_group *group = sd->groups;
  39         unsigned long flags = sd->flags;
  40         unsigned int idx;
  41
  42         cpumask_clear(groupmask);
  43
  44         printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
  45         printk(KERN_CONT "span=%*pbl level=%s\n",
  46                cpumask_pr_args(sched_domain_span(sd)), sd->name);
  47
  48         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
  49                 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
  50         }
  51         if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
  52                 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
  53         }
  54
  55         for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
  56                 unsigned int flag = BIT(idx);
  57                 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
  58
  59                 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
  60                     !(sd->child->flags & flag))
  61                         printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
  62                                sd_flag_debug[idx].name);
  63
  64                 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
  65                     !(sd->parent->flags & flag))
  66                         printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
  67                                sd_flag_debug[idx].name);
  68         }
  69
  70         printk(KERN_DEBUG "%*s groups:", level + 1, "");
  71         do {
  72                 if (!group) {
  73                         printk("\n");
  74                         printk(KERN_ERR "ERROR: group is NULL\n");
  75                         break;
  76                 }
  77
  78                 if (cpumask_empty(sched_group_span(group))) {
  79                         printk(KERN_CONT "\n");
  80                         printk(KERN_ERR "ERROR: empty group\n");
  81                         break;
  82                 }
  83
  84                 if (!(sd->flags & SD_OVERLAP) &&
  85                     cpumask_intersects(groupmask, sched_group_span(group))) {
  86                         printk(KERN_CONT "\n");
  87                         printk(KERN_ERR "ERROR: repeated CPUs\n");
  88                         break;
  89                 }
  90
  91                 cpumask_or(groupmask, groupmask, sched_group_span(group));
  92
  93                 printk(KERN_CONT " %d:{ span=%*pbl",
  94                                 group->sgc->id,
  95                                 cpumask_pr_args(sched_group_span(group)));
  96
  97                 if ((sd->flags & SD_OVERLAP) &&
  98                     !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
  99                         printk(KERN_CONT " mask=%*pbl",
 100                                 cpumask_pr_args(group_balance_mask(group)));
 101                 }
 102
 103                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
 104                         printk(KERN_CONT " cap=%lu", group->sgc->capacity);
 105
 106                 if (group == sd->groups && sd->child &&
 107                     !cpumask_equal(sched_domain_span(sd->child),
 108                                    sched_group_span(group))) {
 109                         printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
 110                 }
 111
 112                 printk(KERN_CONT " }");
 113
 114                 group = group->next;
 115
 116                 if (group != sd->groups)
 117                         printk(KERN_CONT ",");
 118
 119         } while (group != sd->groups);
 120         printk(KERN_CONT "\n");
 121
 122         if (!cpumask_equal(sched_domain_span(sd), groupmask))
 123                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 124
 125         if (sd->parent &&
 126             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
 127                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
 128         return 0;
 129 }
 130
 131 static void sched_domain_debug(struct sched_domain *sd, int cpu)
 132 {
 133         int level = 0;
 134
 135         if (!sched_debug_verbose)
 136                 return;
 137
 138         if (!sd) {
 139                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
 140                 return;
 141         }
 142
 143         printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
 144
 145         for (;;) {
 146                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
 147                         break;
 148                 level++;
 149                 sd = sd->parent;
 150                 if (!sd)
 151                         break;
 152         }
 153 }
 154 #else /* !CONFIG_SCHED_DEBUG */
 155
 156 # define sched_debug_verbose 0
 157 # define sched_domain_debug(sd, cpu) do { } while (0)
 158 static inline bool sched_debug(void)
 159 {
 160         return false;
 161 }
 162 #endif /* CONFIG_SCHED_DEBUG */
 163
 164 /* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
 165 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
 166 static const unsigned int SD_DEGENERATE_GROUPS_MASK =
 167 #include <linux/sched/sd_flags.h>
 168 0;
 169 #undef SD_FLAG
 170
 171 static int sd_degenerate(struct sched_domain *sd)
 172 {
 173         if (cpumask_weight(sched_domain_span(sd)) == 1)
 174                 return 1;
 175
 176         /* Following flags need at least 2 groups */
 177         if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
 178             (sd->groups != sd->groups->next))
 179                 return 0;
 180
 181         /* Following flags don't use groups */
 182         if (sd->flags & (SD_WAKE_AFFINE))
 183                 return 0;
 184
 185         return 1;
 186 }
 187
 188 static int
 189 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
 190 {
 191         unsigned long cflags = sd->flags, pflags = parent->flags;
 192
 193         if (sd_degenerate(parent))
 194                 return 1;
 195
 196         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
 197                 return 0;
 198
 199         /* Flags needing groups don't count if only 1 group in parent */
 200         if (parent->groups == parent->groups->next)
 201                 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
 202
 203         if (~cflags & pflags)
 204                 return 0;
 205
 206         return 1;
 207 }
 208
 209 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
 210 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
 211 static unsigned int sysctl_sched_energy_aware = 1;
 212 static DEFINE_MUTEX(sched_energy_mutex);
 213 static bool sched_energy_update;
 214
 215 void rebuild_sched_domains_energy(void)
 216 {
 217         mutex_lock(&sched_energy_mutex);
 218         sched_energy_update = true;
 219         rebuild_sched_domains();
 220         sched_energy_update = false;
 221         mutex_unlock(&sched_energy_mutex);
 222 }
 223
 224 #ifdef CONFIG_PROC_SYSCTL
 225 static int sched_energy_aware_handler(struct ctl_table *table, int write,
 226                 void *buffer, size_t *lenp, loff_t *ppos)
 227 {
 228         int ret, state;
 229
 230         if (write && !capable(CAP_SYS_ADMIN))
 231                 return -EPERM;
 232
 233         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 234         if (!ret && write) {
 235                 state = static_branch_unlikely(&sched_energy_present);
 236                 if (state != sysctl_sched_energy_aware)
 237                         rebuild_sched_domains_energy();
 238         }
 239
 240         return ret;
 241 }
 242
 243 static struct ctl_table sched_energy_aware_sysctls[] = {
 244         {
 245                 .procname       = "sched_energy_aware",
 246                 .data           = &sysctl_sched_energy_aware,
 247                 .maxlen         = sizeof(unsigned int),
 248                 .mode           = 0644,
 249                 .proc_handler   = sched_energy_aware_handler,
 250                 .extra1         = SYSCTL_ZERO,
 251                 .extra2         = SYSCTL_ONE,
 252         },
 253         {}
 254 };
 255
 256 static int __init sched_energy_aware_sysctl_init(void)
 257 {
 258         register_sysctl_init("kernel", sched_energy_aware_sysctls);
 259         return 0;
 260 }
 261
 262 late_initcall(sched_energy_aware_sysctl_init);
 263 #endif
 264
 265 static void free_pd(struct perf_domain *pd)
 266 {
 267         struct perf_domain *tmp;
 268
 269         while (pd) {
 270                 tmp = pd->next;
 271                 kfree(pd);
 272                 pd = tmp;
 273         }
 274 }
 275
 276 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
 277 {
 278         while (pd) {
 279                 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
 280                         return pd;
 281                 pd = pd->next;
 282         }
 283
 284         return NULL;
 285 }
 286
 287 static struct perf_domain *pd_init(int cpu)
 288 {
 289         struct em_perf_domain *obj = em_cpu_get(cpu);
 290         struct perf_domain *pd;
 291
 292         if (!obj) {
 293                 if (sched_debug())
 294                         pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
 295                 return NULL;
 296         }
 297
 298         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 299         if (!pd)
 300                 return NULL;
 301         pd->em_pd = obj;
 302
 303         return pd;
 304 }
 305
 306 static void perf_domain_debug(const struct cpumask *cpu_map,
 307                                                 struct perf_domain *pd)
 308 {
 309         if (!sched_debug() || !pd)
 310                 return;
 311
 312         printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
 313
 314         while (pd) {
 315                 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
 316                                 cpumask_first(perf_domain_span(pd)),
 317                                 cpumask_pr_args(perf_domain_span(pd)),
 318                                 em_pd_nr_perf_states(pd->em_pd));
 319                 pd = pd->next;
 320         }
 321
 322         printk(KERN_CONT "\n");
 323 }
 324
 325 static void destroy_perf_domain_rcu(struct rcu_head *rp)
 326 {
 327         struct perf_domain *pd;
 328
 329         pd = container_of(rp, struct perf_domain, rcu);
 330         free_pd(pd);
 331 }
 332
 333 static void sched_energy_set(bool has_eas)
 334 {
 335         if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
 336                 if (sched_debug())
 337                         pr_info("%s: stopping EAS\n", __func__);
 338                 static_branch_disable_cpuslocked(&sched_energy_present);
 339         } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
 340                 if (sched_debug())
 341                         pr_info("%s: starting EAS\n", __func__);
 342                 static_branch_enable_cpuslocked(&sched_energy_present);
 343         }
 344 }
 345
 346 /*
 347  * EAS can be used on a root domain if it meets all the following conditions:
 348  *    1. an Energy Model (EM) is available;
 349  *    2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
 350  *    3. no SMT is detected.
 351  *    4. the EM complexity is low enough to keep scheduling overheads low;
 352  *    5. schedutil is driving the frequency of all CPUs of the rd;
 353  *    6. frequency invariance support is present;
 354  *
 355  * The complexity of the Energy Model is defined as:
 356  *
 357  *              C = nr_pd * (nr_cpus + nr_ps)
 358  *
 359  * with parameters defined as:
 360  *  - nr_pd:    the number of performance domains
 361  *  - nr_cpus:  the number of CPUs
 362  *  - nr_ps:    the sum of the number of performance states of all performance
 363  *              domains (for example, on a system with 2 performance domains,
 364  *              with 10 performance states each, nr_ps = 2 * 10 = 20).
 365  *
 366  * It is generally not a good idea to use such a model in the wake-up path on
 367  * very complex platforms because of the associated scheduling overheads. The
 368  * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
 369  * with per-CPU DVFS and less than 8 performance states each, for example.
 370  */
 371 #define EM_MAX_COMPLEXITY 2048
 372
 373 extern struct cpufreq_governor schedutil_gov;
 374 static bool build_perf_domains(const struct cpumask *cpu_map)
 375 {
 376         int i, nr_pd = 0, nr_ps = 0, nr_cpus = cpumask_weight(cpu_map);
 377         struct perf_domain *pd = NULL, *tmp;
 378         int cpu = cpumask_first(cpu_map);
 379         struct root_domain *rd = cpu_rq(cpu)->rd;
 380         struct cpufreq_policy *policy;
 381         struct cpufreq_governor *gov;
 382
 383         if (!sysctl_sched_energy_aware)
 384                 goto free;
 385
 386         /* EAS is enabled for asymmetric CPU capacity topologies. */
 387         if (!per_cpu(sd_asym_cpucapacity, cpu)) {
 388                 if (sched_debug()) {
 389                         pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
 390                                         cpumask_pr_args(cpu_map));
 391                 }
 392                 goto free;
 393         }
 394
 395         /* EAS definitely does *not* handle SMT */
 396         if (sched_smt_active()) {
 397                 pr_warn("rd %*pbl: Disabling EAS, SMT is not supported\n",
 398                         cpumask_pr_args(cpu_map));
 399                 goto free;
 400         }
 401
 402         if (!arch_scale_freq_invariant()) {
 403                 if (sched_debug()) {
 404                         pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking not yet supported",
 405                                 cpumask_pr_args(cpu_map));
 406                 }
 407                 goto free;
 408         }
 409
 410         for_each_cpu(i, cpu_map) {
 411                 /* Skip already covered CPUs. */
 412                 if (find_pd(pd, i))
 413                         continue;
 414
 415                 /* Do not attempt EAS if schedutil is not being used. */
 416                 policy = cpufreq_cpu_get(i);
 417                 if (!policy)
 418                         goto free;
 419                 gov = policy->governor;
 420                 cpufreq_cpu_put(policy);
 421                 if (gov != &schedutil_gov) {
 422                         if (rd->pd)
 423                                 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
 424                                                 cpumask_pr_args(cpu_map));
 425                         goto free;
 426                 }
 427
 428                 /* Create the new pd and add it to the local list. */
 429                 tmp = pd_init(i);
 430                 if (!tmp)
 431                         goto free;
 432                 tmp->next = pd;
 433                 pd = tmp;
 434
 435                 /*
 436                  * Count performance domains and performance states for the
 437                  * complexity check.
 438                  */
 439                 nr_pd++;
 440                 nr_ps += em_pd_nr_perf_states(pd->em_pd);
 441         }
 442
 443         /* Bail out if the Energy Model complexity is too high. */
 444         if (nr_pd * (nr_ps + nr_cpus) > EM_MAX_COMPLEXITY) {
 445                 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
 446                                                 cpumask_pr_args(cpu_map));
 447                 goto free;
 448         }
 449
 450         perf_domain_debug(cpu_map, pd);
 451
 452         /* Attach the new list of performance domains to the root domain. */
 453         tmp = rd->pd;
 454         rcu_assign_pointer(rd->pd, pd);
 455         if (tmp)
 456                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 457
 458         return !!pd;
 459
 460 free:
 461         free_pd(pd);
 462         tmp = rd->pd;
 463         rcu_assign_pointer(rd->pd, NULL);
 464         if (tmp)
 465                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
 466
 467         return false;
 468 }
 469 #else
 470 static void free_pd(struct perf_domain *pd) { }
 471 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
 472
 473 static void free_rootdomain(struct rcu_head *rcu)
 474 {
 475         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 476
 477         cpupri_cleanup(&rd->cpupri);
 478         cpudl_cleanup(&rd->cpudl);
 479         free_cpumask_var(rd->dlo_mask);
 480         free_cpumask_var(rd->rto_mask);
 481         free_cpumask_var(rd->online);
 482         free_cpumask_var(rd->span);
 483         free_pd(rd->pd);
 484         kfree(rd);
 485 }
 486
 487 void rq_attach_root(struct rq *rq, struct root_domain *rd)
 488 {
 489         struct root_domain *old_rd = NULL;
 490         unsigned long flags;
 491
 492         raw_spin_rq_lock_irqsave(rq, flags);
 493
 494         if (rq->rd) {
 495                 old_rd = rq->rd;
 496
 497                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
 498                         set_rq_offline(rq);
 499
 500                 cpumask_clear_cpu(rq->cpu, old_rd->span);
 501
 502                 /*
 503                  * If we dont want to free the old_rd yet then
 504                  * set old_rd to NULL to skip the freeing later
 505                  * in this function:
 506                  */
 507                 if (!atomic_dec_and_test(&old_rd->refcount))
 508                         old_rd = NULL;
 509         }
 510
 511         atomic_inc(&rd->refcount);
 512         rq->rd = rd;
 513
 514         cpumask_set_cpu(rq->cpu, rd->span);
 515         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 516                 set_rq_online(rq);
 517
 518         raw_spin_rq_unlock_irqrestore(rq, flags);
 519
 520         if (old_rd)
 521                 call_rcu(&old_rd->rcu, free_rootdomain);
 522 }
 523
 524 void sched_get_rd(struct root_domain *rd)
 525 {
 526         atomic_inc(&rd->refcount);
 527 }
 528
 529 void sched_put_rd(struct root_domain *rd)
 530 {
 531         if (!atomic_dec_and_test(&rd->refcount))
 532                 return;
 533
 534         call_rcu(&rd->rcu, free_rootdomain);
 535 }
 536
 537 static int init_rootdomain(struct root_domain *rd)
 538 {
 539         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 540                 goto out;
 541         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 542                 goto free_span;
 543         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 544                 goto free_online;
 545         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 546                 goto free_dlo_mask;
 547
 548 #ifdef HAVE_RT_PUSH_IPI
 549         rd->rto_cpu = -1;
 550         raw_spin_lock_init(&rd->rto_lock);
 551         rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
 552 #endif
 553
 554         rd->visit_gen = 0;
 555         init_dl_bw(&rd->dl_bw);
 556         if (cpudl_init(&rd->cpudl) != 0)
 557                 goto free_rto_mask;
 558
 559         if (cpupri_init(&rd->cpupri) != 0)
 560                 goto free_cpudl;
 561         return 0;
 562
 563 free_cpudl:
 564         cpudl_cleanup(&rd->cpudl);
 565 free_rto_mask:
 566         free_cpumask_var(rd->rto_mask);
 567 free_dlo_mask:
 568         free_cpumask_var(rd->dlo_mask);
 569 free_online:
 570         free_cpumask_var(rd->online);
 571 free_span:
 572         free_cpumask_var(rd->span);
 573 out:
 574         return -ENOMEM;
 575 }
 576
 577 /*
 578  * By default the system creates a single root-domain with all CPUs as
 579  * members (mimicking the global state we have today).
 580  */
 581 struct root_domain def_root_domain;
 582
 583 void __init init_defrootdomain(void)
 584 {
 585         init_rootdomain(&def_root_domain);
 586
 587         atomic_set(&def_root_domain.refcount, 1);
 588 }
 589
 590 static struct root_domain *alloc_rootdomain(void)
 591 {
 592         struct root_domain *rd;
 593
 594         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 595         if (!rd)
 596                 return NULL;
 597
 598         if (init_rootdomain(rd) != 0) {
 599                 kfree(rd);
 600                 return NULL;
 601         }
 602
 603         return rd;
 604 }
 605
 606 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 607 {
 608         struct sched_group *tmp, *first;
 609
 610         if (!sg)
 611                 return;
 612
 613         first = sg;
 614         do {
 615                 tmp = sg->next;
 616
 617                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 618                         kfree(sg->sgc);
 619
 620                 if (atomic_dec_and_test(&sg->ref))
 621                         kfree(sg);
 622                 sg = tmp;
 623         } while (sg != first);
 624 }
 625
 626 static void destroy_sched_domain(struct sched_domain *sd)
 627 {
 628         /*
 629          * A normal sched domain may have multiple group references, an
 630          * overlapping domain, having private groups, only one.  Iterate,
 631          * dropping group/capacity references, freeing where none remain.
 632          */
 633         free_sched_groups(sd->groups, 1);
 634
 635         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 636                 kfree(sd->shared);
 637         kfree(sd);
 638 }
 639
 640 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 641 {
 642         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 643
 644         while (sd) {
 645                 struct sched_domain *parent = sd->parent;
 646                 destroy_sched_domain(sd);
 647                 sd = parent;
 648         }
 649 }
 650
 651 static void destroy_sched_domains(struct sched_domain *sd)
 652 {
 653         if (sd)
 654                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 655 }
 656
 657 /*
 658  * Keep a special pointer to the highest sched_domain that has
 659  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
 660  * allows us to avoid some pointer chasing select_idle_sibling().
 661  *
 662  * Also keep a unique ID per domain (we use the first CPU number in
 663  * the cpumask of the domain), this allows us to quickly tell if
 664  * two CPUs are in the same cache domain, see cpus_share_cache().
 665  */
 666 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 667 DEFINE_PER_CPU(int, sd_llc_size);
 668 DEFINE_PER_CPU(int, sd_llc_id);
 669 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 670 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 671 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 672 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
 673 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 674
 675 static void update_top_cache_domain(int cpu)
 676 {
 677         struct sched_domain_shared *sds = NULL;
 678         struct sched_domain *sd;
 679         int id = cpu;
 680         int size = 1;
 681
 682         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 683         if (sd) {
 684                 id = cpumask_first(sched_domain_span(sd));
 685                 size = cpumask_weight(sched_domain_span(sd));
 686                 sds = sd->shared;
 687         }
 688
 689         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 690         per_cpu(sd_llc_size, cpu) = size;
 691         per_cpu(sd_llc_id, cpu) = id;
 692         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 693
 694         sd = lowest_flag_domain(cpu, SD_NUMA);
 695         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 696
 697         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 698         rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
 699
 700         sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
 701         rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
 702 }
 703
 704 /*
 705  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 706  * hold the hotplug lock.
 707  */
 708 static void
 709 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 710 {
 711         struct rq *rq = cpu_rq(cpu);
 712         struct sched_domain *tmp;
 713
 714         /* Remove the sched domains which do not contribute to scheduling. */
 715         for (tmp = sd; tmp; ) {
 716                 struct sched_domain *parent = tmp->parent;
 717                 if (!parent)
 718                         break;
 719
 720                 if (sd_parent_degenerate(tmp, parent)) {
 721                         tmp->parent = parent->parent;
 722                         if (parent->parent)
 723                                 parent->parent->child = tmp;
 724                         /*
 725                          * Transfer SD_PREFER_SIBLING down in case of a
 726                          * degenerate parent; the spans match for this
 727                          * so the property transfers.
 728                          */
 729                         if (parent->flags & SD_PREFER_SIBLING)
 730                                 tmp->flags |= SD_PREFER_SIBLING;
 731                         destroy_sched_domain(parent);
 732                 } else
 733                         tmp = tmp->parent;
 734         }
 735
 736         if (sd && sd_degenerate(sd)) {
 737                 tmp = sd;
 738                 sd = sd->parent;
 739                 destroy_sched_domain(tmp);
 740                 if (sd) {
 741                         struct sched_group *sg = sd->groups;
 742
 743                         /*
 744                          * sched groups hold the flags of the child sched
 745                          * domain for convenience. Clear such flags since
 746                          * the child is being destroyed.
 747                          */
 748                         do {
 749                                 sg->flags = 0;
 750                         } while (sg != sd->groups);
 751
 752                         sd->child = NULL;
 753                 }
 754         }
 755
 756         sched_domain_debug(sd, cpu);
 757
 758         rq_attach_root(rq, rd);
 759         tmp = rq->sd;
 760         rcu_assign_pointer(rq->sd, sd);
 761         dirty_sched_domain_sysctl(cpu);
 762         destroy_sched_domains(tmp);
 763
 764         update_top_cache_domain(cpu);
 765 }
 766
 767 struct s_data {
 768         struct sched_domain * __percpu *sd;
 769         struct root_domain      *rd;
 770 };
 771
 772 enum s_alloc {
 773         sa_rootdomain,
 774         sa_sd,
 775         sa_sd_storage,
 776         sa_none,
 777 };
 778
 779 /*
 780  * Return the canonical balance CPU for this group, this is the first CPU
 781  * of this group that's also in the balance mask.
 782  *
 783  * The balance mask are all those CPUs that could actually end up at this
 784  * group. See build_balance_mask().
 785  *
 786  * Also see should_we_balance().
 787  */
 788 int group_balance_cpu(struct sched_group *sg)
 789 {
 790         return cpumask_first(group_balance_mask(sg));
 791 }
 792
 793
 794 /*
 795  * NUMA topology (first read the regular topology blurb below)
 796  *
 797  * Given a node-distance table, for example:
 798  *
 799  *   node   0   1   2   3
 800  *     0:  10  20  30  20
 801  *     1:  20  10  20  30
 802  *     2:  30  20  10  20
 803  *     3:  20  30  20  10
 804  *
 805  * which represents a 4 node ring topology like:
 806  *
 807  *   0 ----- 1
 808  *   |       |
 809  *   |       |
 810  *   |       |
 811  *   3 ----- 2
 812  *
 813  * We want to construct domains and groups to represent this. The way we go
 814  * about doing this is to build the domains on 'hops'. For each NUMA level we
 815  * construct the mask of all nodes reachable in @level hops.
 816  *
 817  * For the above NUMA topology that gives 3 levels:
 818  *
 819  * NUMA-2       0-3             0-3             0-3             0-3
 820  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 821  *
 822  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 823  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 824  *
 825  * NUMA-0       0               1               2               3
 826  *
 827  *
 828  * As can be seen; things don't nicely line up as with the regular topology.
 829  * When we iterate a domain in child domain chunks some nodes can be
 830  * represented multiple times -- hence the "overlap" naming for this part of
 831  * the topology.
 832  *
 833  * In order to minimize this overlap, we only build enough groups to cover the
 834  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 835  *
 836  * Because:
 837  *
 838  *  - the first group of each domain is its child domain; this
 839  *    gets us the first 0-1,3
 840  *  - the only uncovered node is 2, who's child domain is 1-3.
 841  *
 842  * However, because of the overlap, computing a unique CPU for each group is
 843  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 844  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 845  * end up at those groups (they would end up in group: 0-1,3).
 846  *
 847  * To correct this we have to introduce the group balance mask. This mask
 848  * will contain those CPUs in the group that can reach this group given the
 849  * (child) domain tree.
 850  *
 851  * With this we can once again compute balance_cpu and sched_group_capacity
 852  * relations.
 853  *
 854  * XXX include words on how balance_cpu is unique and therefore can be
 855  * used for sched_group_capacity links.
 856  *
 857  *
 858  * Another 'interesting' topology is:
 859  *
 860  *   node   0   1   2   3
 861  *     0:  10  20  20  30
 862  *     1:  20  10  20  20
 863  *     2:  20  20  10  20
 864  *     3:  30  20  20  10
 865  *
 866  * Which looks a little like:
 867  *
 868  *   0 ----- 1
 869  *   |     / |
 870  *   |   /   |
 871  *   | /     |
 872  *   2 ----- 3
 873  *
 874  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 875  * are not.
 876  *
 877  * This leads to a few particularly weird cases where the sched_domain's are
 878  * not of the same number for each CPU. Consider:
 879  *
 880  * NUMA-2       0-3                                             0-3
 881  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 882  *
 883  * NUMA-1       0-2             0-3             0-3             1-3
 884  *
 885  * NUMA-0       0               1               2               3
 886  *
 887  */
 888
 889
 890 /*
 891  * Build the balance mask; it contains only those CPUs that can arrive at this
 892  * group and should be considered to continue balancing.
 893  *
 894  * We do this during the group creation pass, therefore the group information
 895  * isn't complete yet, however since each group represents a (child) domain we
 896  * can fully construct this using the sched_domain bits (which are already
 897  * complete).
 898  */
 899 static void
 900 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 901 {
 902         const struct cpumask *sg_span = sched_group_span(sg);
 903         struct sd_data *sdd = sd->private;
 904         struct sched_domain *sibling;
 905         int i;
 906
 907         cpumask_clear(mask);
 908
 909         for_each_cpu(i, sg_span) {
 910                 sibling = *per_cpu_ptr(sdd->sd, i);
 911
 912                 /*
 913                  * Can happen in the asymmetric case, where these siblings are
 914                  * unused. The mask will not be empty because those CPUs that
 915                  * do have the top domain _should_ span the domain.
 916                  */
 917                 if (!sibling->child)
 918                         continue;
 919
 920                 /* If we would not end up here, we can't continue from here */
 921                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 922                         continue;
 923
 924                 cpumask_set_cpu(i, mask);
 925         }
 926
 927         /* We must not have empty masks here */
 928         WARN_ON_ONCE(cpumask_empty(mask));
 929 }
 930
 931 /*
 932  * XXX: This creates per-node group entries; since the load-balancer will
 933  * immediately access remote memory to construct this group's load-balance
 934  * statistics having the groups node local is of dubious benefit.
 935  */
 936 static struct sched_group *
 937 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 938 {
 939         struct sched_group *sg;
 940         struct cpumask *sg_span;
 941
 942         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 943                         GFP_KERNEL, cpu_to_node(cpu));
 944
 945         if (!sg)
 946                 return NULL;
 947
 948         sg_span = sched_group_span(sg);
 949         if (sd->child) {
 950                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 951                 sg->flags = sd->child->flags;
 952         } else {
 953                 cpumask_copy(sg_span, sched_domain_span(sd));
 954         }
 955
 956         atomic_inc(&sg->ref);
 957         return sg;
 958 }
 959
 960 static void init_overlap_sched_group(struct sched_domain *sd,
 961                                      struct sched_group *sg)
 962 {
 963         struct cpumask *mask = sched_domains_tmpmask2;
 964         struct sd_data *sdd = sd->private;
 965         struct cpumask *sg_span;
 966         int cpu;
 967
 968         build_balance_mask(sd, sg, mask);
 969         cpu = cpumask_first(mask);
 970
 971         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 972         if (atomic_inc_return(&sg->sgc->ref) == 1)
 973                 cpumask_copy(group_balance_mask(sg), mask);
 974         else
 975                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 976
 977         /*
 978          * Initialize sgc->capacity such that even if we mess up the
 979          * domains and no possible iteration will get us here, we won't
 980          * die on a /0 trap.
 981          */
 982         sg_span = sched_group_span(sg);
 983         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 984         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 985         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 986 }
 987
 988 static struct sched_domain *
 989 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
 990 {
 991         /*
 992          * The proper descendant would be the one whose child won't span out
 993          * of sd
 994          */
 995         while (sibling->child &&
 996                !cpumask_subset(sched_domain_span(sibling->child),
 997                                sched_domain_span(sd)))
 998                 sibling = sibling->child;
 999
1000         /*
1001          * As we are referencing sgc across different topology level, we need
1002          * to go down to skip those sched_domains which don't contribute to
1003          * scheduling because they will be degenerated in cpu_attach_domain
1004          */
1005         while (sibling->child &&
1006                cpumask_equal(sched_domain_span(sibling->child),
1007                              sched_domain_span(sibling)))
1008                 sibling = sibling->child;
1009
1010         return sibling;
1011 }
1012
1013 static int
1014 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1015 {
1016         struct sched_group *first = NULL, *last = NULL, *sg;
1017         const struct cpumask *span = sched_domain_span(sd);
1018         struct cpumask *covered = sched_domains_tmpmask;
1019         struct sd_data *sdd = sd->private;
1020         struct sched_domain *sibling;
1021         int i;
1022
1023         cpumask_clear(covered);
1024
1025         for_each_cpu_wrap(i, span, cpu) {
1026                 struct cpumask *sg_span;
1027
1028                 if (cpumask_test_cpu(i, covered))
1029                         continue;
1030
1031                 sibling = *per_cpu_ptr(sdd->sd, i);
1032
1033                 /*
1034                  * Asymmetric node setups can result in situations where the
1035                  * domain tree is of unequal depth, make sure to skip domains
1036                  * that already cover the entire range.
1037                  *
1038                  * In that case build_sched_domains() will have terminated the
1039                  * iteration early and our sibling sd spans will be empty.
1040                  * Domains should always include the CPU they're built on, so
1041                  * check that.
1042                  */
1043                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1044                         continue;
1045
1046                 /*
1047                  * Usually we build sched_group by sibling's child sched_domain
1048                  * But for machines whose NUMA diameter are 3 or above, we move
1049                  * to build sched_group by sibling's proper descendant's child
1050                  * domain because sibling's child sched_domain will span out of
1051                  * the sched_domain being built as below.
1052                  *
1053                  * Smallest diameter=3 topology is:
1054                  *
1055                  *   node   0   1   2   3
1056                  *     0:  10  20  30  40
1057                  *     1:  20  10  20  30
1058                  *     2:  30  20  10  20
1059                  *     3:  40  30  20  10
1060                  *
1061                  *   0 --- 1 --- 2 --- 3
1062                  *
1063                  * NUMA-3       0-3             N/A             N/A             0-3
1064                  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1065                  *
1066                  * NUMA-2       0-2             0-3             0-3             1-3
1067                  *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1068                  *
1069                  * NUMA-1       0-1             0-2             1-3             2-3
1070                  *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1071                  *
1072                  * NUMA-0       0               1               2               3
1073                  *
1074                  * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1075                  * group span isn't a subset of the domain span.
1076                  */
1077                 if (sibling->child &&
1078                     !cpumask_subset(sched_domain_span(sibling->child), span))
1079                         sibling = find_descended_sibling(sd, sibling);
1080
1081                 sg = build_group_from_child_sched_domain(sibling, cpu);
1082                 if (!sg)
1083                         goto fail;
1084
1085                 sg_span = sched_group_span(sg);
1086                 cpumask_or(covered, covered, sg_span);
1087
1088                 init_overlap_sched_group(sibling, sg);
1089
1090                 if (!first)
1091                         first = sg;
1092                 if (last)
1093                         last->next = sg;
1094                 last = sg;
1095                 last->next = first;
1096         }
1097         sd->groups = first;
1098
1099         return 0;
1100
1101 fail:
1102         free_sched_groups(first, 0);
1103
1104         return -ENOMEM;
1105 }
1106
1107
1108 /*
1109  * Package topology (also see the load-balance blurb in fair.c)
1110  *
1111  * The scheduler builds a tree structure to represent a number of important
1112  * topology features. By default (default_topology[]) these include:
1113  *
1114  *  - Simultaneous multithreading (SMT)
1115  *  - Multi-Core Cache (MC)
1116  *  - Package (DIE)
1117  *
1118  * Where the last one more or less denotes everything up to a NUMA node.
1119  *
1120  * The tree consists of 3 primary data structures:
1121  *
1122  *      sched_domain -> sched_group -> sched_group_capacity
1123  *          ^ ^             ^ ^
1124  *          `-'             `-'
1125  *
1126  * The sched_domains are per-CPU and have a two way link (parent & child) and
1127  * denote the ever growing mask of CPUs belonging to that level of topology.
1128  *
1129  * Each sched_domain has a circular (double) linked list of sched_group's, each
1130  * denoting the domains of the level below (or individual CPUs in case of the
1131  * first domain level). The sched_group linked by a sched_domain includes the
1132  * CPU of that sched_domain [*].
1133  *
1134  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1135  *
1136  * CPU   0   1   2   3   4   5   6   7
1137  *
1138  * DIE  [                             ]
1139  * MC   [             ] [             ]
1140  * SMT  [     ] [     ] [     ] [     ]
1141  *
1142  *  - or -
1143  *
1144  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1145  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1146  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1147  *
1148  * CPU   0   1   2   3   4   5   6   7
1149  *
1150  * One way to think about it is: sched_domain moves you up and down among these
1151  * topology levels, while sched_group moves you sideways through it, at child
1152  * domain granularity.
1153  *
1154  * sched_group_capacity ensures each unique sched_group has shared storage.
1155  *
1156  * There are two related construction problems, both require a CPU that
1157  * uniquely identify each group (for a given domain):
1158  *
1159  *  - The first is the balance_cpu (see should_we_balance() and the
1160  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1161  *    continue balancing at a higher domain.
1162  *
1163  *  - The second is the sched_group_capacity; we want all identical groups
1164  *    to share a single sched_group_capacity.
1165  *
1166  * Since these topologies are exclusive by construction. That is, its
1167  * impossible for an SMT thread to belong to multiple cores, and cores to
1168  * be part of multiple caches. There is a very clear and unique location
1169  * for each CPU in the hierarchy.
1170  *
1171  * Therefore computing a unique CPU for each group is trivial (the iteration
1172  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1173  * group), we can simply pick the first CPU in each group.
1174  *
1175  *
1176  * [*] in other words, the first group of each domain is its child domain.
1177  */
1178
1179 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1180 {
1181         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1182         struct sched_domain *child = sd->child;
1183         struct sched_group *sg;
1184         bool already_visited;
1185
1186         if (child)
1187                 cpu = cpumask_first(sched_domain_span(child));
1188
1189         sg = *per_cpu_ptr(sdd->sg, cpu);
1190         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1191
1192         /* Increase refcounts for claim_allocations: */
1193         already_visited = atomic_inc_return(&sg->ref) > 1;
1194         /* sgc visits should follow a similar trend as sg */
1195         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1196
1197         /* If we have already visited that group, it's already initialized. */
1198         if (already_visited)
1199                 return sg;
1200
1201         if (child) {
1202                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1203                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1204                 sg->flags = child->flags;
1205         } else {
1206                 cpumask_set_cpu(cpu, sched_group_span(sg));
1207                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1208         }
1209
1210         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1211         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1212         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1213
1214         return sg;
1215 }
1216
1217 /*
1218  * build_sched_groups will build a circular linked list of the groups
1219  * covered by the given span, will set each group's ->cpumask correctly,
1220  * and will initialize their ->sgc.
1221  *
1222  * Assumes the sched_domain tree is fully constructed
1223  */
1224 static int
1225 build_sched_groups(struct sched_domain *sd, int cpu)
1226 {
1227         struct sched_group *first = NULL, *last = NULL;
1228         struct sd_data *sdd = sd->private;
1229         const struct cpumask *span = sched_domain_span(sd);
1230         struct cpumask *covered;
1231         int i;
1232
1233         lockdep_assert_held(&sched_domains_mutex);
1234         covered = sched_domains_tmpmask;
1235
1236         cpumask_clear(covered);
1237
1238         for_each_cpu_wrap(i, span, cpu) {
1239                 struct sched_group *sg;
1240
1241                 if (cpumask_test_cpu(i, covered))
1242                         continue;
1243
1244                 sg = get_group(i, sdd);
1245
1246                 cpumask_or(covered, covered, sched_group_span(sg));
1247
1248                 if (!first)
1249                         first = sg;
1250                 if (last)
1251                         last->next = sg;
1252                 last = sg;
1253         }
1254         last->next = first;
1255         sd->groups = first;
1256
1257         return 0;
1258 }
1259
1260 /*
1261  * Initialize sched groups cpu_capacity.
1262  *
1263  * cpu_capacity indicates the capacity of sched group, which is used while
1264  * distributing the load between different sched groups in a sched domain.
1265  * Typically cpu_capacity for all the groups in a sched domain will be same
1266  * unless there are asymmetries in the topology. If there are asymmetries,
1267  * group having more cpu_capacity will pickup more load compared to the
1268  * group having less cpu_capacity.
1269  */
1270 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1271 {
1272         struct sched_group *sg = sd->groups;
1273
1274         WARN_ON(!sg);
1275
1276         do {
1277                 int cpu, max_cpu = -1;
1278
1279                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1280
1281                 if (!(sd->flags & SD_ASYM_PACKING))
1282                         goto next;
1283
1284                 for_each_cpu(cpu, sched_group_span(sg)) {
1285                         if (max_cpu < 0)
1286                                 max_cpu = cpu;
1287                         else if (sched_asym_prefer(cpu, max_cpu))
1288                                 max_cpu = cpu;
1289                 }
1290                 sg->asym_prefer_cpu = max_cpu;
1291
1292 next:
1293                 sg = sg->next;
1294         } while (sg != sd->groups);
1295
1296         if (cpu != group_balance_cpu(sg))
1297                 return;
1298
1299         update_group_capacity(sd, cpu);
1300 }
1301
1302 /*
1303  * Asymmetric CPU capacity bits
1304  */
1305 struct asym_cap_data {
1306         struct list_head link;
1307         unsigned long capacity;
1308         unsigned long cpus[];
1309 };
1310
1311 /*
1312  * Set of available CPUs grouped by their corresponding capacities
1313  * Each list entry contains a CPU mask reflecting CPUs that share the same
1314  * capacity.
1315  * The lifespan of data is unlimited.
1316  */
1317 static LIST_HEAD(asym_cap_list);
1318
1319 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1320
1321 /*
1322  * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1323  * Provides sd_flags reflecting the asymmetry scope.
1324  */
1325 static inline int
1326 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1327                            const struct cpumask *cpu_map)
1328 {
1329         struct asym_cap_data *entry;
1330         int count = 0, miss = 0;
1331
1332         /*
1333          * Count how many unique CPU capacities this domain spans across
1334          * (compare sched_domain CPUs mask with ones representing  available
1335          * CPUs capacities). Take into account CPUs that might be offline:
1336          * skip those.
1337          */
1338         list_for_each_entry(entry, &asym_cap_list, link) {
1339                 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1340                         ++count;
1341                 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1342                         ++miss;
1343         }
1344
1345         WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1346
1347         /* No asymmetry detected */
1348         if (count < 2)
1349                 return 0;
1350         /* Some of the available CPU capacity values have not been detected */
1351         if (miss)
1352                 return SD_ASYM_CPUCAPACITY;
1353
1354         /* Full asymmetry */
1355         return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1356
1357 }
1358
1359 static inline void asym_cpu_capacity_update_data(int cpu)
1360 {
1361         unsigned long capacity = arch_scale_cpu_capacity(cpu);
1362         struct asym_cap_data *entry = NULL;
1363
1364         list_for_each_entry(entry, &asym_cap_list, link) {
1365                 if (capacity == entry->capacity)
1366                         goto done;
1367         }
1368
1369         entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1370         if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1371                 return;
1372         entry->capacity = capacity;
1373         list_add(&entry->link, &asym_cap_list);
1374 done:
1375         __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1376 }
1377
1378 /*
1379  * Build-up/update list of CPUs grouped by their capacities
1380  * An update requires explicit request to rebuild sched domains
1381  * with state indicating CPU topology changes.
1382  */
1383 static void asym_cpu_capacity_scan(void)
1384 {
1385         struct asym_cap_data *entry, *next;
1386         int cpu;
1387
1388         list_for_each_entry(entry, &asym_cap_list, link)
1389                 cpumask_clear(cpu_capacity_span(entry));
1390
1391         for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1392                 asym_cpu_capacity_update_data(cpu);
1393
1394         list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1395                 if (cpumask_empty(cpu_capacity_span(entry))) {
1396                         list_del(&entry->link);
1397                         kfree(entry);
1398                 }
1399         }
1400
1401         /*
1402          * Only one capacity value has been detected i.e. this system is symmetric.
1403          * No need to keep this data around.
1404          */
1405         if (list_is_singular(&asym_cap_list)) {
1406                 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1407                 list_del(&entry->link);
1408                 kfree(entry);
1409         }
1410 }
1411
1412 /*
1413  * Initializers for schedule domains
1414  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1415  */
1416
1417 static int default_relax_domain_level = -1;
1418 int sched_domain_level_max;
1419
1420 static int __init setup_relax_domain_level(char *str)
1421 {
1422         if (kstrtoint(str, 0, &default_relax_domain_level))
1423                 pr_warn("Unable to set relax_domain_level\n");
1424
1425         return 1;
1426 }
1427 __setup("relax_domain_level=", setup_relax_domain_level);
1428
1429 static void set_domain_attribute(struct sched_domain *sd,
1430                                  struct sched_domain_attr *attr)
1431 {
1432         int request;
1433
1434         if (!attr || attr->relax_domain_level < 0) {
1435                 if (default_relax_domain_level < 0)
1436                         return;
1437                 request = default_relax_domain_level;
1438         } else
1439                 request = attr->relax_domain_level;
1440
1441         if (sd->level > request) {
1442                 /* Turn off idle balance on this domain: */
1443                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1444         }
1445 }
1446
1447 static void __sdt_free(const struct cpumask *cpu_map);
1448 static int __sdt_alloc(const struct cpumask *cpu_map);
1449
1450 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1451                                  const struct cpumask *cpu_map)
1452 {
1453         switch (what) {
1454         case sa_rootdomain:
1455                 if (!atomic_read(&d->rd->refcount))
1456                         free_rootdomain(&d->rd->rcu);
1457                 fallthrough;
1458         case sa_sd:
1459                 free_percpu(d->sd);
1460                 fallthrough;
1461         case sa_sd_storage:
1462                 __sdt_free(cpu_map);
1463                 fallthrough;
1464         case sa_none:
1465                 break;
1466         }
1467 }
1468
1469 static enum s_alloc
1470 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1471 {
1472         memset(d, 0, sizeof(*d));
1473
1474         if (__sdt_alloc(cpu_map))
1475                 return sa_sd_storage;
1476         d->sd = alloc_percpu(struct sched_domain *);
1477         if (!d->sd)
1478                 return sa_sd_storage;
1479         d->rd = alloc_rootdomain();
1480         if (!d->rd)
1481                 return sa_sd;
1482
1483         return sa_rootdomain;
1484 }
1485
1486 /*
1487  * NULL the sd_data elements we've used to build the sched_domain and
1488  * sched_group structure so that the subsequent __free_domain_allocs()
1489  * will not free the data we're using.
1490  */
1491 static void claim_allocations(int cpu, struct sched_domain *sd)
1492 {
1493         struct sd_data *sdd = sd->private;
1494
1495         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1496         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1497
1498         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1499                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1500
1501         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1502                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1503
1504         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1505                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1506 }
1507
1508 #ifdef CONFIG_NUMA
1509 enum numa_topology_type sched_numa_topology_type;
1510
1511 static int                      sched_domains_numa_levels;
1512 static int                      sched_domains_curr_level;
1513
1514 int                             sched_max_numa_distance;
1515 static int                      *sched_domains_numa_distance;
1516 static struct cpumask           ***sched_domains_numa_masks;
1517 #endif
1518
1519 /*
1520  * SD_flags allowed in topology descriptions.
1521  *
1522  * These flags are purely descriptive of the topology and do not prescribe
1523  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1524  * function:
1525  *
1526  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1527  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1528  *   SD_NUMA                - describes NUMA topologies
1529  *
1530  * Odd one out, which beside describing the topology has a quirk also
1531  * prescribes the desired behaviour that goes along with it:
1532  *
1533  *   SD_ASYM_PACKING        - describes SMT quirks
1534  */
1535 #define TOPOLOGY_SD_FLAGS               \
1536         (SD_SHARE_CPUCAPACITY   |       \
1537          SD_SHARE_PKG_RESOURCES |       \
1538          SD_NUMA                |       \
1539          SD_ASYM_PACKING)
1540
1541 static struct sched_domain *
1542 sd_init(struct sched_domain_topology_level *tl,
1543         const struct cpumask *cpu_map,
1544         struct sched_domain *child, int cpu)
1545 {
1546         struct sd_data *sdd = &tl->data;
1547         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1548         int sd_id, sd_weight, sd_flags = 0;
1549         struct cpumask *sd_span;
1550
1551 #ifdef CONFIG_NUMA
1552         /*
1553          * Ugly hack to pass state to sd_numa_mask()...
1554          */
1555         sched_domains_curr_level = tl->numa_level;
1556 #endif
1557
1558         sd_weight = cpumask_weight(tl->mask(cpu));
1559
1560         if (tl->sd_flags)
1561                 sd_flags = (*tl->sd_flags)();
1562         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1563                         "wrong sd_flags in topology description\n"))
1564                 sd_flags &= TOPOLOGY_SD_FLAGS;
1565
1566         *sd = (struct sched_domain){
1567                 .min_interval           = sd_weight,
1568                 .max_interval           = 2*sd_weight,
1569                 .busy_factor            = 16,
1570                 .imbalance_pct          = 117,
1571
1572                 .cache_nice_tries       = 0,
1573
1574                 .flags                  = 1*SD_BALANCE_NEWIDLE
1575                                         | 1*SD_BALANCE_EXEC
1576                                         | 1*SD_BALANCE_FORK
1577                                         | 0*SD_BALANCE_WAKE
1578                                         | 1*SD_WAKE_AFFINE
1579                                         | 0*SD_SHARE_CPUCAPACITY
1580                                         | 0*SD_SHARE_PKG_RESOURCES
1581                                         | 0*SD_SERIALIZE
1582                                         | 1*SD_PREFER_SIBLING
1583                                         | 0*SD_NUMA
1584                                         | sd_flags
1585                                         ,
1586
1587                 .last_balance           = jiffies,
1588                 .balance_interval       = sd_weight,
1589                 .max_newidle_lb_cost    = 0,
1590                 .last_decay_max_lb_cost = jiffies,
1591                 .child                  = child,
1592 #ifdef CONFIG_SCHED_DEBUG
1593                 .name                   = tl->name,
1594 #endif
1595         };
1596
1597         sd_span = sched_domain_span(sd);
1598         cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1599         sd_id = cpumask_first(sd_span);
1600
1601         sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1602
1603         WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1604                   (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1605                   "CPU capacity asymmetry not supported on SMT\n");
1606
1607         /*
1608          * Convert topological properties into behaviour.
1609          */
1610         /* Don't attempt to spread across CPUs of different capacities. */
1611         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1612                 sd->child->flags &= ~SD_PREFER_SIBLING;
1613
1614         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1615                 sd->imbalance_pct = 110;
1616
1617         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1618                 sd->imbalance_pct = 117;
1619                 sd->cache_nice_tries = 1;
1620
1621 #ifdef CONFIG_NUMA
1622         } else if (sd->flags & SD_NUMA) {
1623                 sd->cache_nice_tries = 2;
1624
1625                 sd->flags &= ~SD_PREFER_SIBLING;
1626                 sd->flags |= SD_SERIALIZE;
1627                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1628                         sd->flags &= ~(SD_BALANCE_EXEC |
1629                                        SD_BALANCE_FORK |
1630                                        SD_WAKE_AFFINE);
1631                 }
1632
1633 #endif
1634         } else {
1635                 sd->cache_nice_tries = 1;
1636         }
1637
1638         /*
1639          * For all levels sharing cache; connect a sched_domain_shared
1640          * instance.
1641          */
1642         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1643                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1644                 atomic_inc(&sd->shared->ref);
1645                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1646         }
1647
1648         sd->private = sdd;
1649
1650         return sd;
1651 }
1652
1653 /*
1654  * Topology list, bottom-up.
1655  */
1656 static struct sched_domain_topology_level default_topology[] = {
1657 #ifdef CONFIG_SCHED_SMT
1658         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1659 #endif
1660
1661 #ifdef CONFIG_SCHED_CLUSTER
1662         { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1663 #endif
1664
1665 #ifdef CONFIG_SCHED_MC
1666         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1667 #endif
1668         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1669         { NULL, },
1670 };
1671
1672 static struct sched_domain_topology_level *sched_domain_topology =
1673         default_topology;
1674 static struct sched_domain_topology_level *sched_domain_topology_saved;
1675
1676 #define for_each_sd_topology(tl)                        \
1677         for (tl = sched_domain_topology; tl->mask; tl++)
1678
1679 void set_sched_topology(struct sched_domain_topology_level *tl)
1680 {
1681         if (WARN_ON_ONCE(sched_smp_initialized))
1682                 return;
1683
1684         sched_domain_topology = tl;
1685         sched_domain_topology_saved = NULL;
1686 }
1687
1688 #ifdef CONFIG_NUMA
1689
1690 static const struct cpumask *sd_numa_mask(int cpu)
1691 {
1692         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1693 }
1694
1695 static void sched_numa_warn(const char *str)
1696 {
1697         static int done = false;
1698         int i,j;
1699
1700         if (done)
1701                 return;
1702
1703         done = true;
1704
1705         printk(KERN_WARNING "ERROR: %s\n\n", str);
1706
1707         for (i = 0; i < nr_node_ids; i++) {
1708                 printk(KERN_WARNING "  ");
1709                 for (j = 0; j < nr_node_ids; j++) {
1710                         if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1711                                 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1712                         else
1713                                 printk(KERN_CONT " %02d  ", node_distance(i,j));
1714                 }
1715                 printk(KERN_CONT "\n");
1716         }
1717         printk(KERN_WARNING "\n");
1718 }
1719
1720 bool find_numa_distance(int distance)
1721 {
1722         bool found = false;
1723         int i, *distances;
1724
1725         if (distance == node_distance(0, 0))
1726                 return true;
1727
1728         rcu_read_lock();
1729         distances = rcu_dereference(sched_domains_numa_distance);
1730         if (!distances)
1731                 goto unlock;
1732         for (i = 0; i < sched_domains_numa_levels; i++) {
1733                 if (distances[i] == distance) {
1734                         found = true;
1735                         break;
1736                 }
1737         }
1738 unlock:
1739         rcu_read_unlock();
1740
1741         return found;
1742 }
1743
1744 #define for_each_cpu_node_but(n, nbut)          \
1745         for_each_node_state(n, N_CPU)           \
1746                 if (n == nbut)                  \
1747                         continue;               \
1748                 else
1749
1750 /*
1751  * A system can have three types of NUMA topology:
1752  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1753  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1754  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1755  *
1756  * The difference between a glueless mesh topology and a backplane
1757  * topology lies in whether communication between not directly
1758  * connected nodes goes through intermediary nodes (where programs
1759  * could run), or through backplane controllers. This affects
1760  * placement of programs.
1761  *
1762  * The type of topology can be discerned with the following tests:
1763  * - If the maximum distance between any nodes is 1 hop, the system
1764  *   is directly connected.
1765  * - If for two nodes A and B, located N > 1 hops away from each other,
1766  *   there is an intermediary node C, which is < N hops away from both
1767  *   nodes A and B, the system is a glueless mesh.
1768  */
1769 static void init_numa_topology_type(int offline_node)
1770 {
1771         int a, b, c, n;
1772
1773         n = sched_max_numa_distance;
1774
1775         if (sched_domains_numa_levels <= 2) {
1776                 sched_numa_topology_type = NUMA_DIRECT;
1777                 return;
1778         }
1779
1780         for_each_cpu_node_but(a, offline_node) {
1781                 for_each_cpu_node_but(b, offline_node) {
1782                         /* Find two nodes furthest removed from each other. */
1783                         if (node_distance(a, b) < n)
1784                                 continue;
1785
1786                         /* Is there an intermediary node between a and b? */
1787                         for_each_cpu_node_but(c, offline_node) {
1788                                 if (node_distance(a, c) < n &&
1789                                     node_distance(b, c) < n) {
1790                                         sched_numa_topology_type =
1791                                                         NUMA_GLUELESS_MESH;
1792                                         return;
1793                                 }
1794                         }
1795
1796                         sched_numa_topology_type = NUMA_BACKPLANE;
1797                         return;
1798                 }
1799         }
1800
1801         pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1802         sched_numa_topology_type = NUMA_DIRECT;
1803 }
1804
1805
1806 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1807
1808 void sched_init_numa(int offline_node)
1809 {
1810         struct sched_domain_topology_level *tl;
1811         unsigned long *distance_map;
1812         int nr_levels = 0;
1813         int i, j;
1814         int *distances;
1815         struct cpumask ***masks;
1816
1817         /*
1818          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1819          * unique distances in the node_distance() table.
1820          */
1821         distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1822         if (!distance_map)
1823                 return;
1824
1825         bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1826         for_each_cpu_node_but(i, offline_node) {
1827                 for_each_cpu_node_but(j, offline_node) {
1828                         int distance = node_distance(i, j);
1829
1830                         if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1831                                 sched_numa_warn("Invalid distance value range");
1832                                 bitmap_free(distance_map);
1833                                 return;
1834                         }
1835
1836                         bitmap_set(distance_map, distance, 1);
1837                 }
1838         }
1839         /*
1840          * We can now figure out how many unique distance values there are and
1841          * allocate memory accordingly.
1842          */
1843         nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1844
1845         distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1846         if (!distances) {
1847                 bitmap_free(distance_map);
1848                 return;
1849         }
1850
1851         for (i = 0, j = 0; i < nr_levels; i++, j++) {
1852                 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1853                 distances[i] = j;
1854         }
1855         rcu_assign_pointer(sched_domains_numa_distance, distances);
1856
1857         bitmap_free(distance_map);
1858
1859         /*
1860          * 'nr_levels' contains the number of unique distances
1861          *
1862          * The sched_domains_numa_distance[] array includes the actual distance
1863          * numbers.
1864          */
1865
1866         /*
1867          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1868          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1869          * the array will contain less then 'nr_levels' members. This could be
1870          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1871          * in other functions.
1872          *
1873          * We reset it to 'nr_levels' at the end of this function.
1874          */
1875         sched_domains_numa_levels = 0;
1876
1877         masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1878         if (!masks)
1879                 return;
1880
1881         /*
1882          * Now for each level, construct a mask per node which contains all
1883          * CPUs of nodes that are that many hops away from us.
1884          */
1885         for (i = 0; i < nr_levels; i++) {
1886                 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1887                 if (!masks[i])
1888                         return;
1889
1890                 for_each_cpu_node_but(j, offline_node) {
1891                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1892                         int k;
1893
1894                         if (!mask)
1895                                 return;
1896
1897                         masks[i][j] = mask;
1898
1899                         for_each_cpu_node_but(k, offline_node) {
1900                                 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1901                                         sched_numa_warn("Node-distance not symmetric");
1902
1903                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1904                                         continue;
1905
1906                                 cpumask_or(mask, mask, cpumask_of_node(k));
1907                         }
1908                 }
1909         }
1910         rcu_assign_pointer(sched_domains_numa_masks, masks);
1911
1912         /* Compute default topology size */
1913         for (i = 0; sched_domain_topology[i].mask; i++);
1914
1915         tl = kzalloc((i + nr_levels + 1) *
1916                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1917         if (!tl)
1918                 return;
1919
1920         /*
1921          * Copy the default topology bits..
1922          */
1923         for (i = 0; sched_domain_topology[i].mask; i++)
1924                 tl[i] = sched_domain_topology[i];
1925
1926         /*
1927          * Add the NUMA identity distance, aka single NODE.
1928          */
1929         tl[i++] = (struct sched_domain_topology_level){
1930                 .mask = sd_numa_mask,
1931                 .numa_level = 0,
1932                 SD_INIT_NAME(NODE)
1933         };
1934
1935         /*
1936          * .. and append 'j' levels of NUMA goodness.
1937          */
1938         for (j = 1; j < nr_levels; i++, j++) {
1939                 tl[i] = (struct sched_domain_topology_level){
1940                         .mask = sd_numa_mask,
1941                         .sd_flags = cpu_numa_flags,
1942                         .flags = SDTL_OVERLAP,
1943                         .numa_level = j,
1944                         SD_INIT_NAME(NUMA)
1945                 };
1946         }
1947
1948         sched_domain_topology_saved = sched_domain_topology;
1949         sched_domain_topology = tl;
1950
1951         sched_domains_numa_levels = nr_levels;
1952         WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1953
1954         init_numa_topology_type(offline_node);
1955 }
1956
1957
1958 static void sched_reset_numa(void)
1959 {
1960         int nr_levels, *distances;
1961         struct cpumask ***masks;
1962
1963         nr_levels = sched_domains_numa_levels;
1964         sched_domains_numa_levels = 0;
1965         sched_max_numa_distance = 0;
1966         sched_numa_topology_type = NUMA_DIRECT;
1967         distances = sched_domains_numa_distance;
1968         rcu_assign_pointer(sched_domains_numa_distance, NULL);
1969         masks = sched_domains_numa_masks;
1970         rcu_assign_pointer(sched_domains_numa_masks, NULL);
1971         if (distances || masks) {
1972                 int i, j;
1973
1974                 synchronize_rcu();
1975                 kfree(distances);
1976                 for (i = 0; i < nr_levels && masks; i++) {
1977                         if (!masks[i])
1978                                 continue;
1979                         for_each_node(j)
1980                                 kfree(masks[i][j]);
1981                         kfree(masks[i]);
1982                 }
1983                 kfree(masks);
1984         }
1985         if (sched_domain_topology_saved) {
1986                 kfree(sched_domain_topology);
1987                 sched_domain_topology = sched_domain_topology_saved;
1988                 sched_domain_topology_saved = NULL;
1989         }
1990 }
1991
1992 /*
1993  * Call with hotplug lock held
1994  */
1995 void sched_update_numa(int cpu, bool online)
1996 {
1997         int node;
1998
1999         node = cpu_to_node(cpu);
2000         /*
2001          * Scheduler NUMA topology is updated when the first CPU of a
2002          * node is onlined or the last CPU of a node is offlined.
2003          */
2004         if (cpumask_weight(cpumask_of_node(node)) != 1)
2005                 return;
2006
2007         sched_reset_numa();
2008         sched_init_numa(online ? NUMA_NO_NODE : node);
2009 }
2010
2011 void sched_domains_numa_masks_set(unsigned int cpu)
2012 {
2013         int node = cpu_to_node(cpu);
2014         int i, j;
2015
2016         for (i = 0; i < sched_domains_numa_levels; i++) {
2017                 for (j = 0; j < nr_node_ids; j++) {
2018                         if (!node_state(j, N_CPU))
2019                                 continue;
2020
2021                         /* Set ourselves in the remote node's masks */
2022                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
2023                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2024                 }
2025         }
2026 }
2027
2028 void sched_domains_numa_masks_clear(unsigned int cpu)
2029 {
2030         int i, j;
2031
2032         for (i = 0; i < sched_domains_numa_levels; i++) {
2033                 for (j = 0; j < nr_node_ids; j++) {
2034                         if (sched_domains_numa_masks[i][j])
2035                                 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2036                 }
2037         }
2038 }
2039
2040 /*
2041  * sched_numa_find_closest() - given the NUMA topology, find the cpu
2042  *                             closest to @cpu from @cpumask.
2043  * cpumask: cpumask to find a cpu from
2044  * cpu: cpu to be close to
2045  *
2046  * returns: cpu, or nr_cpu_ids when nothing found.
2047  */
2048 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2049 {
2050         int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2051         struct cpumask ***masks;
2052
2053         rcu_read_lock();
2054         masks = rcu_dereference(sched_domains_numa_masks);
2055         if (!masks)
2056                 goto unlock;
2057         for (i = 0; i < sched_domains_numa_levels; i++) {
2058                 if (!masks[i][j])
2059                         break;
2060                 cpu = cpumask_any_and(cpus, masks[i][j]);
2061                 if (cpu < nr_cpu_ids) {
2062                         found = cpu;
2063                         break;
2064                 }
2065         }
2066 unlock:
2067         rcu_read_unlock();
2068
2069         return found;
2070 }
2071
2072 struct __cmp_key {
2073         const struct cpumask *cpus;
2074         struct cpumask ***masks;
2075         int node;
2076         int cpu;
2077         int w;
2078 };
2079
2080 static int hop_cmp(const void *a, const void *b)
2081 {
2082         struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2083         struct __cmp_key *k = (struct __cmp_key *)a;
2084
2085         if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2086                 return 1;
2087
2088         if (b == k->masks) {
2089                 k->w = 0;
2090                 return 0;
2091         }
2092
2093         prev_hop = *((struct cpumask ***)b - 1);
2094         k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2095         if (k->w <= k->cpu)
2096                 return 0;
2097
2098         return -1;
2099 }
2100
2101 /*
2102  * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth next cpu
2103  *                             closest to @cpu from @cpumask.
2104  * cpumask: cpumask to find a cpu from
2105  * cpu: Nth cpu to find
2106  *
2107  * returns: cpu, or nr_cpu_ids when nothing found.
2108  */
2109 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2110 {
2111         struct __cmp_key k = { .cpus = cpus, .node = node, .cpu = cpu };
2112         struct cpumask ***hop_masks;
2113         int hop, ret = nr_cpu_ids;
2114
2115         rcu_read_lock();
2116
2117         k.masks = rcu_dereference(sched_domains_numa_masks);
2118         if (!k.masks)
2119                 goto unlock;
2120
2121         hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2122         hop = hop_masks - k.masks;
2123
2124         ret = hop ?
2125                 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2126                 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2127 unlock:
2128         rcu_read_unlock();
2129         return ret;
2130 }
2131 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2132
2133 /**
2134  * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2135  *                         @node
2136  * @node: The node to count hops from.
2137  * @hops: Include CPUs up to that many hops away. 0 means local node.
2138  *
2139  * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2140  * @node, an error value otherwise.
2141  *
2142  * Requires rcu_lock to be held. Returned cpumask is only valid within that
2143  * read-side section, copy it if required beyond that.
2144  *
2145  * Note that not all hops are equal in distance; see sched_init_numa() for how
2146  * distances and masks are handled.
2147  * Also note that this is a reflection of sched_domains_numa_masks, which may change
2148  * during the lifetime of the system (offline nodes are taken out of the masks).
2149  */
2150 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2151 {
2152         struct cpumask ***masks;
2153
2154         if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2155                 return ERR_PTR(-EINVAL);
2156
2157         masks = rcu_dereference(sched_domains_numa_masks);
2158         if (!masks)
2159                 return ERR_PTR(-EBUSY);
2160
2161         return masks[hops][node];
2162 }
2163 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2164
2165 #endif /* CONFIG_NUMA */
2166
2167 static int __sdt_alloc(const struct cpumask *cpu_map)
2168 {
2169         struct sched_domain_topology_level *tl;
2170         int j;
2171
2172         for_each_sd_topology(tl) {
2173                 struct sd_data *sdd = &tl->data;
2174
2175                 sdd->sd = alloc_percpu(struct sched_domain *);
2176                 if (!sdd->sd)
2177                         return -ENOMEM;
2178
2179                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2180                 if (!sdd->sds)
2181                         return -ENOMEM;
2182
2183                 sdd->sg = alloc_percpu(struct sched_group *);
2184                 if (!sdd->sg)
2185                         return -ENOMEM;
2186
2187                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2188                 if (!sdd->sgc)
2189                         return -ENOMEM;
2190
2191                 for_each_cpu(j, cpu_map) {
2192                         struct sched_domain *sd;
2193                         struct sched_domain_shared *sds;
2194                         struct sched_group *sg;
2195                         struct sched_group_capacity *sgc;
2196
2197                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2198                                         GFP_KERNEL, cpu_to_node(j));
2199                         if (!sd)
2200                                 return -ENOMEM;
2201
2202                         *per_cpu_ptr(sdd->sd, j) = sd;
2203
2204                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
2205                                         GFP_KERNEL, cpu_to_node(j));
2206                         if (!sds)
2207                                 return -ENOMEM;
2208
2209                         *per_cpu_ptr(sdd->sds, j) = sds;
2210
2211                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2212                                         GFP_KERNEL, cpu_to_node(j));
2213                         if (!sg)
2214                                 return -ENOMEM;
2215
2216                         sg->next = sg;
2217
2218                         *per_cpu_ptr(sdd->sg, j) = sg;
2219
2220                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2221                                         GFP_KERNEL, cpu_to_node(j));
2222                         if (!sgc)
2223                                 return -ENOMEM;
2224
2225 #ifdef CONFIG_SCHED_DEBUG
2226                         sgc->id = j;
2227 #endif
2228
2229                         *per_cpu_ptr(sdd->sgc, j) = sgc;
2230                 }
2231         }
2232
2233         return 0;
2234 }
2235
2236 static void __sdt_free(const struct cpumask *cpu_map)
2237 {
2238         struct sched_domain_topology_level *tl;
2239         int j;
2240
2241         for_each_sd_topology(tl) {
2242                 struct sd_data *sdd = &tl->data;
2243
2244                 for_each_cpu(j, cpu_map) {
2245                         struct sched_domain *sd;
2246
2247                         if (sdd->sd) {
2248                                 sd = *per_cpu_ptr(sdd->sd, j);
2249                                 if (sd && (sd->flags & SD_OVERLAP))
2250                                         free_sched_groups(sd->groups, 0);
2251                                 kfree(*per_cpu_ptr(sdd->sd, j));
2252                         }
2253
2254                         if (sdd->sds)
2255                                 kfree(*per_cpu_ptr(sdd->sds, j));
2256                         if (sdd->sg)
2257                                 kfree(*per_cpu_ptr(sdd->sg, j));
2258                         if (sdd->sgc)
2259                                 kfree(*per_cpu_ptr(sdd->sgc, j));
2260                 }
2261                 free_percpu(sdd->sd);
2262                 sdd->sd = NULL;
2263                 free_percpu(sdd->sds);
2264                 sdd->sds = NULL;
2265                 free_percpu(sdd->sg);
2266                 sdd->sg = NULL;
2267                 free_percpu(sdd->sgc);
2268                 sdd->sgc = NULL;
2269         }
2270 }
2271
2272 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2273                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2274                 struct sched_domain *child, int cpu)
2275 {
2276         struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2277
2278         if (child) {
2279                 sd->level = child->level + 1;
2280                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2281                 child->parent = sd;
2282
2283                 if (!cpumask_subset(sched_domain_span(child),
2284                                     sched_domain_span(sd))) {
2285                         pr_err("BUG: arch topology borken\n");
2286 #ifdef CONFIG_SCHED_DEBUG
2287                         pr_err("     the %s domain not a subset of the %s domain\n",
2288                                         child->name, sd->name);
2289 #endif
2290                         /* Fixup, ensure @sd has at least @child CPUs. */
2291                         cpumask_or(sched_domain_span(sd),
2292                                    sched_domain_span(sd),
2293                                    sched_domain_span(child));
2294                 }
2295
2296         }
2297         set_domain_attribute(sd, attr);
2298
2299         return sd;
2300 }
2301
2302 /*
2303  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2304  * any two given CPUs at this (non-NUMA) topology level.
2305  */
2306 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2307                               const struct cpumask *cpu_map, int cpu)
2308 {
2309         int i;
2310
2311         /* NUMA levels are allowed to overlap */
2312         if (tl->flags & SDTL_OVERLAP)
2313                 return true;
2314
2315         /*
2316          * Non-NUMA levels cannot partially overlap - they must be either
2317          * completely equal or completely disjoint. Otherwise we can end up
2318          * breaking the sched_group lists - i.e. a later get_group() pass
2319          * breaks the linking done for an earlier span.
2320          */
2321         for_each_cpu(i, cpu_map) {
2322                 if (i == cpu)
2323                         continue;
2324                 /*
2325                  * We should 'and' all those masks with 'cpu_map' to exactly
2326                  * match the topology we're about to build, but that can only
2327                  * remove CPUs, which only lessens our ability to detect
2328                  * overlaps
2329                  */
2330                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2331                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2332                         return false;
2333         }
2334
2335         return true;
2336 }
2337
2338 /*
2339  * Build sched domains for a given set of CPUs and attach the sched domains
2340  * to the individual CPUs
2341  */
2342 static int
2343 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2344 {
2345         enum s_alloc alloc_state = sa_none;
2346         struct sched_domain *sd;
2347         struct s_data d;
2348         struct rq *rq = NULL;
2349         int i, ret = -ENOMEM;
2350         bool has_asym = false;
2351
2352         if (WARN_ON(cpumask_empty(cpu_map)))
2353                 goto error;
2354
2355         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2356         if (alloc_state != sa_rootdomain)
2357                 goto error;
2358
2359         /* Set up domains for CPUs specified by the cpu_map: */
2360         for_each_cpu(i, cpu_map) {
2361                 struct sched_domain_topology_level *tl;
2362
2363                 sd = NULL;
2364                 for_each_sd_topology(tl) {
2365
2366                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2367                                 goto error;
2368
2369                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2370
2371                         has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2372
2373                         if (tl == sched_domain_topology)
2374                                 *per_cpu_ptr(d.sd, i) = sd;
2375                         if (tl->flags & SDTL_OVERLAP)
2376                                 sd->flags |= SD_OVERLAP;
2377                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2378                                 break;
2379                 }
2380         }
2381
2382         /* Build the groups for the domains */
2383         for_each_cpu(i, cpu_map) {
2384                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2385                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2386                         if (sd->flags & SD_OVERLAP) {
2387                                 if (build_overlap_sched_groups(sd, i))
2388                                         goto error;
2389                         } else {
2390                                 if (build_sched_groups(sd, i))
2391                                         goto error;
2392                         }
2393                 }
2394         }
2395
2396         /*
2397          * Calculate an allowed NUMA imbalance such that LLCs do not get
2398          * imbalanced.
2399          */
2400         for_each_cpu(i, cpu_map) {
2401                 unsigned int imb = 0;
2402                 unsigned int imb_span = 1;
2403
2404                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2405                         struct sched_domain *child = sd->child;
2406
2407                         if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2408                             (child->flags & SD_SHARE_PKG_RESOURCES)) {
2409                                 struct sched_domain __rcu *top_p;
2410                                 unsigned int nr_llcs;
2411
2412                                 /*
2413                                  * For a single LLC per node, allow an
2414                                  * imbalance up to 12.5% of the node. This is
2415                                  * arbitrary cutoff based two factors -- SMT and
2416                                  * memory channels. For SMT-2, the intent is to
2417                                  * avoid premature sharing of HT resources but
2418                                  * SMT-4 or SMT-8 *may* benefit from a different
2419                                  * cutoff. For memory channels, this is a very
2420                                  * rough estimate of how many channels may be
2421                                  * active and is based on recent CPUs with
2422                                  * many cores.
2423                                  *
2424                                  * For multiple LLCs, allow an imbalance
2425                                  * until multiple tasks would share an LLC
2426                                  * on one node while LLCs on another node
2427                                  * remain idle. This assumes that there are
2428                                  * enough logical CPUs per LLC to avoid SMT
2429                                  * factors and that there is a correlation
2430                                  * between LLCs and memory channels.
2431                                  */
2432                                 nr_llcs = sd->span_weight / child->span_weight;
2433                                 if (nr_llcs == 1)
2434                                         imb = sd->span_weight >> 3;
2435                                 else
2436                                         imb = nr_llcs;
2437                                 imb = max(1U, imb);
2438                                 sd->imb_numa_nr = imb;
2439
2440                                 /* Set span based on the first NUMA domain. */
2441                                 top_p = sd->parent;
2442                                 while (top_p && !(top_p->flags & SD_NUMA)) {
2443                                         top_p = top_p->parent;
2444                                 }
2445                                 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2446                         } else {
2447                                 int factor = max(1U, (sd->span_weight / imb_span));
2448
2449                                 sd->imb_numa_nr = imb * factor;
2450                         }
2451                 }
2452         }
2453
2454         /* Calculate CPU capacity for physical packages and nodes */
2455         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2456                 if (!cpumask_test_cpu(i, cpu_map))
2457                         continue;
2458
2459                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2460                         claim_allocations(i, sd);
2461                         init_sched_groups_capacity(i, sd);
2462                 }
2463         }
2464
2465         /* Attach the domains */
2466         rcu_read_lock();
2467         for_each_cpu(i, cpu_map) {
2468                 rq = cpu_rq(i);
2469                 sd = *per_cpu_ptr(d.sd, i);
2470
2471                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2472                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2473                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2474
2475                 cpu_attach_domain(sd, d.rd, i);
2476         }
2477         rcu_read_unlock();
2478
2479         if (has_asym)
2480                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2481
2482         if (rq && sched_debug_verbose) {
2483                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2484                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2485         }
2486
2487         ret = 0;
2488 error:
2489         __free_domain_allocs(&d, alloc_state, cpu_map);
2490
2491         return ret;
2492 }
2493
2494 /* Current sched domains: */
2495 static cpumask_var_t                    *doms_cur;
2496
2497 /* Number of sched domains in 'doms_cur': */
2498 static int                              ndoms_cur;
2499
2500 /* Attributes of custom domains in 'doms_cur' */
2501 static struct sched_domain_attr         *dattr_cur;
2502
2503 /*
2504  * Special case: If a kmalloc() of a doms_cur partition (array of
2505  * cpumask) fails, then fallback to a single sched domain,
2506  * as determined by the single cpumask fallback_doms.
2507  */
2508 static cpumask_var_t                    fallback_doms;
2509
2510 /*
2511  * arch_update_cpu_topology lets virtualized architectures update the
2512  * CPU core maps. It is supposed to return 1 if the topology changed
2513  * or 0 if it stayed the same.
2514  */
2515 int __weak arch_update_cpu_topology(void)
2516 {
2517         return 0;
2518 }
2519
2520 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2521 {
2522         int i;
2523         cpumask_var_t *doms;
2524
2525         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2526         if (!doms)
2527                 return NULL;
2528         for (i = 0; i < ndoms; i++) {
2529                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2530                         free_sched_domains(doms, i);
2531                         return NULL;
2532                 }
2533         }
2534         return doms;
2535 }
2536
2537 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2538 {
2539         unsigned int i;
2540         for (i = 0; i < ndoms; i++)
2541                 free_cpumask_var(doms[i]);
2542         kfree(doms);
2543 }
2544
2545 /*
2546  * Set up scheduler domains and groups.  For now this just excludes isolated
2547  * CPUs, but could be used to exclude other special cases in the future.
2548  */
2549 int __init sched_init_domains(const struct cpumask *cpu_map)
2550 {
2551         int err;
2552
2553         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2554         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2555         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2556
2557         arch_update_cpu_topology();
2558         asym_cpu_capacity_scan();
2559         ndoms_cur = 1;
2560         doms_cur = alloc_sched_domains(ndoms_cur);
2561         if (!doms_cur)
2562                 doms_cur = &fallback_doms;
2563         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2564         err = build_sched_domains(doms_cur[0], NULL);
2565
2566         return err;
2567 }
2568
2569 /*
2570  * Detach sched domains from a group of CPUs specified in cpu_map
2571  * These CPUs will now be attached to the NULL domain
2572  */
2573 static void detach_destroy_domains(const struct cpumask *cpu_map)
2574 {
2575         unsigned int cpu = cpumask_any(cpu_map);
2576         int i;
2577
2578         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2579                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2580
2581         rcu_read_lock();
2582         for_each_cpu(i, cpu_map)
2583                 cpu_attach_domain(NULL, &def_root_domain, i);
2584         rcu_read_unlock();
2585 }
2586
2587 /* handle null as "default" */
2588 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2589                         struct sched_domain_attr *new, int idx_new)
2590 {
2591         struct sched_domain_attr tmp;
2592
2593         /* Fast path: */
2594         if (!new && !cur)
2595                 return 1;
2596
2597         tmp = SD_ATTR_INIT;
2598
2599         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2600                         new ? (new + idx_new) : &tmp,
2601                         sizeof(struct sched_domain_attr));
2602 }
2603
2604 /*
2605  * Partition sched domains as specified by the 'ndoms_new'
2606  * cpumasks in the array doms_new[] of cpumasks. This compares
2607  * doms_new[] to the current sched domain partitioning, doms_cur[].
2608  * It destroys each deleted domain and builds each new domain.
2609  *
2610  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2611  * The masks don't intersect (don't overlap.) We should setup one
2612  * sched domain for each mask. CPUs not in any of the cpumasks will
2613  * not be load balanced. If the same cpumask appears both in the
2614  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2615  * it as it is.
2616  *
2617  * The passed in 'doms_new' should be allocated using
2618  * alloc_sched_domains.  This routine takes ownership of it and will
2619  * free_sched_domains it when done with it. If the caller failed the
2620  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2621  * and partition_sched_domains() will fallback to the single partition
2622  * 'fallback_doms', it also forces the domains to be rebuilt.
2623  *
2624  * If doms_new == NULL it will be replaced with cpu_online_mask.
2625  * ndoms_new == 0 is a special case for destroying existing domains,
2626  * and it will not create the default domain.
2627  *
2628  * Call with hotplug lock and sched_domains_mutex held
2629  */
2630 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2631                                     struct sched_domain_attr *dattr_new)
2632 {
2633         bool __maybe_unused has_eas = false;
2634         int i, j, n;
2635         int new_topology;
2636
2637         lockdep_assert_held(&sched_domains_mutex);
2638
2639         /* Let the architecture update CPU core mappings: */
2640         new_topology = arch_update_cpu_topology();
2641         /* Trigger rebuilding CPU capacity asymmetry data */
2642         if (new_topology)
2643                 asym_cpu_capacity_scan();
2644
2645         if (!doms_new) {
2646                 WARN_ON_ONCE(dattr_new);
2647                 n = 0;
2648                 doms_new = alloc_sched_domains(1);
2649                 if (doms_new) {
2650                         n = 1;
2651                         cpumask_and(doms_new[0], cpu_active_mask,
2652                                     housekeeping_cpumask(HK_TYPE_DOMAIN));
2653                 }
2654         } else {
2655                 n = ndoms_new;
2656         }
2657
2658         /* Destroy deleted domains: */
2659         for (i = 0; i < ndoms_cur; i++) {
2660                 for (j = 0; j < n && !new_topology; j++) {
2661                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2662                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2663                                 struct root_domain *rd;
2664
2665                                 /*
2666                                  * This domain won't be destroyed and as such
2667                                  * its dl_bw->total_bw needs to be cleared.  It
2668                                  * will be recomputed in function
2669                                  * update_tasks_root_domain().
2670                                  */
2671                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2672                                 dl_clear_root_domain(rd);
2673                                 goto match1;
2674                         }
2675                 }
2676                 /* No match - a current sched domain not in new doms_new[] */
2677                 detach_destroy_domains(doms_cur[i]);
2678 match1:
2679                 ;
2680         }
2681
2682         n = ndoms_cur;
2683         if (!doms_new) {
2684                 n = 0;
2685                 doms_new = &fallback_doms;
2686                 cpumask_and(doms_new[0], cpu_active_mask,
2687                             housekeeping_cpumask(HK_TYPE_DOMAIN));
2688         }
2689
2690         /* Build new domains: */
2691         for (i = 0; i < ndoms_new; i++) {
2692                 for (j = 0; j < n && !new_topology; j++) {
2693                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2694                             dattrs_equal(dattr_new, i, dattr_cur, j))
2695                                 goto match2;
2696                 }
2697                 /* No match - add a new doms_new */
2698                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2699 match2:
2700                 ;
2701         }
2702
2703 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2704         /* Build perf. domains: */
2705         for (i = 0; i < ndoms_new; i++) {
2706                 for (j = 0; j < n && !sched_energy_update; j++) {
2707                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2708                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2709                                 has_eas = true;
2710                                 goto match3;
2711                         }
2712                 }
2713                 /* No match - add perf. domains for a new rd */
2714                 has_eas |= build_perf_domains(doms_new[i]);
2715 match3:
2716                 ;
2717         }
2718         sched_energy_set(has_eas);
2719 #endif
2720
2721         /* Remember the new sched domains: */
2722         if (doms_cur != &fallback_doms)
2723                 free_sched_domains(doms_cur, ndoms_cur);
2724
2725         kfree(dattr_cur);
2726         doms_cur = doms_new;
2727         dattr_cur = dattr_new;
2728         ndoms_cur = ndoms_new;
2729
2730         update_sched_domain_debugfs();
2731 }
2732
2733 /*
2734  * Call with hotplug lock held
2735  */
2736 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2737                              struct sched_domain_attr *dattr_new)
2738 {
2739         mutex_lock(&sched_domains_mutex);
2740         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2741         mutex_unlock(&sched_domains_mutex);
2742 }