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         struct rq_flags rf;
 491
 492         rq_lock_irqsave(rq, &rf);
 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         rq_unlock_irqrestore(rq, &rf);
 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
 723                         if (parent->parent) {
 724                                 parent->parent->child = tmp;
 725                                 if (tmp->flags & SD_SHARE_CPUCAPACITY)
 726                                         parent->parent->groups->flags |= SD_SHARE_CPUCAPACITY;
 727                         }
 728
 729                         /*
 730                          * Transfer SD_PREFER_SIBLING down in case of a
 731                          * degenerate parent; the spans match for this
 732                          * so the property transfers.
 733                          */
 734                         if (parent->flags & SD_PREFER_SIBLING)
 735                                 tmp->flags |= SD_PREFER_SIBLING;
 736                         destroy_sched_domain(parent);
 737                 } else
 738                         tmp = tmp->parent;
 739         }
 740
 741         if (sd && sd_degenerate(sd)) {
 742                 tmp = sd;
 743                 sd = sd->parent;
 744                 destroy_sched_domain(tmp);
 745                 if (sd) {
 746                         struct sched_group *sg = sd->groups;
 747
 748                         /*
 749                          * sched groups hold the flags of the child sched
 750                          * domain for convenience. Clear such flags since
 751                          * the child is being destroyed.
 752                          */
 753                         do {
 754                                 sg->flags = 0;
 755                         } while (sg != sd->groups);
 756
 757                         sd->child = NULL;
 758                 }
 759         }
 760
 761         sched_domain_debug(sd, cpu);
 762
 763         rq_attach_root(rq, rd);
 764         tmp = rq->sd;
 765         rcu_assign_pointer(rq->sd, sd);
 766         dirty_sched_domain_sysctl(cpu);
 767         destroy_sched_domains(tmp);
 768
 769         update_top_cache_domain(cpu);
 770 }
 771
 772 struct s_data {
 773         struct sched_domain * __percpu *sd;
 774         struct root_domain      *rd;
 775 };
 776
 777 enum s_alloc {
 778         sa_rootdomain,
 779         sa_sd,
 780         sa_sd_storage,
 781         sa_none,
 782 };
 783
 784 /*
 785  * Return the canonical balance CPU for this group, this is the first CPU
 786  * of this group that's also in the balance mask.
 787  *
 788  * The balance mask are all those CPUs that could actually end up at this
 789  * group. See build_balance_mask().
 790  *
 791  * Also see should_we_balance().
 792  */
 793 int group_balance_cpu(struct sched_group *sg)
 794 {
 795         return cpumask_first(group_balance_mask(sg));
 796 }
 797
 798
 799 /*
 800  * NUMA topology (first read the regular topology blurb below)
 801  *
 802  * Given a node-distance table, for example:
 803  *
 804  *   node   0   1   2   3
 805  *     0:  10  20  30  20
 806  *     1:  20  10  20  30
 807  *     2:  30  20  10  20
 808  *     3:  20  30  20  10
 809  *
 810  * which represents a 4 node ring topology like:
 811  *
 812  *   0 ----- 1
 813  *   |       |
 814  *   |       |
 815  *   |       |
 816  *   3 ----- 2
 817  *
 818  * We want to construct domains and groups to represent this. The way we go
 819  * about doing this is to build the domains on 'hops'. For each NUMA level we
 820  * construct the mask of all nodes reachable in @level hops.
 821  *
 822  * For the above NUMA topology that gives 3 levels:
 823  *
 824  * NUMA-2       0-3             0-3             0-3             0-3
 825  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 826  *
 827  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 828  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 829  *
 830  * NUMA-0       0               1               2               3
 831  *
 832  *
 833  * As can be seen; things don't nicely line up as with the regular topology.
 834  * When we iterate a domain in child domain chunks some nodes can be
 835  * represented multiple times -- hence the "overlap" naming for this part of
 836  * the topology.
 837  *
 838  * In order to minimize this overlap, we only build enough groups to cover the
 839  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 840  *
 841  * Because:
 842  *
 843  *  - the first group of each domain is its child domain; this
 844  *    gets us the first 0-1,3
 845  *  - the only uncovered node is 2, who's child domain is 1-3.
 846  *
 847  * However, because of the overlap, computing a unique CPU for each group is
 848  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 849  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 850  * end up at those groups (they would end up in group: 0-1,3).
 851  *
 852  * To correct this we have to introduce the group balance mask. This mask
 853  * will contain those CPUs in the group that can reach this group given the
 854  * (child) domain tree.
 855  *
 856  * With this we can once again compute balance_cpu and sched_group_capacity
 857  * relations.
 858  *
 859  * XXX include words on how balance_cpu is unique and therefore can be
 860  * used for sched_group_capacity links.
 861  *
 862  *
 863  * Another 'interesting' topology is:
 864  *
 865  *   node   0   1   2   3
 866  *     0:  10  20  20  30
 867  *     1:  20  10  20  20
 868  *     2:  20  20  10  20
 869  *     3:  30  20  20  10
 870  *
 871  * Which looks a little like:
 872  *
 873  *   0 ----- 1
 874  *   |     / |
 875  *   |   /   |
 876  *   | /     |
 877  *   2 ----- 3
 878  *
 879  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 880  * are not.
 881  *
 882  * This leads to a few particularly weird cases where the sched_domain's are
 883  * not of the same number for each CPU. Consider:
 884  *
 885  * NUMA-2       0-3                                             0-3
 886  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 887  *
 888  * NUMA-1       0-2             0-3             0-3             1-3
 889  *
 890  * NUMA-0       0               1               2               3
 891  *
 892  */
 893
 894
 895 /*
 896  * Build the balance mask; it contains only those CPUs that can arrive at this
 897  * group and should be considered to continue balancing.
 898  *
 899  * We do this during the group creation pass, therefore the group information
 900  * isn't complete yet, however since each group represents a (child) domain we
 901  * can fully construct this using the sched_domain bits (which are already
 902  * complete).
 903  */
 904 static void
 905 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 906 {
 907         const struct cpumask *sg_span = sched_group_span(sg);
 908         struct sd_data *sdd = sd->private;
 909         struct sched_domain *sibling;
 910         int i;
 911
 912         cpumask_clear(mask);
 913
 914         for_each_cpu(i, sg_span) {
 915                 sibling = *per_cpu_ptr(sdd->sd, i);
 916
 917                 /*
 918                  * Can happen in the asymmetric case, where these siblings are
 919                  * unused. The mask will not be empty because those CPUs that
 920                  * do have the top domain _should_ span the domain.
 921                  */
 922                 if (!sibling->child)
 923                         continue;
 924
 925                 /* If we would not end up here, we can't continue from here */
 926                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 927                         continue;
 928
 929                 cpumask_set_cpu(i, mask);
 930         }
 931
 932         /* We must not have empty masks here */
 933         WARN_ON_ONCE(cpumask_empty(mask));
 934 }
 935
 936 /*
 937  * XXX: This creates per-node group entries; since the load-balancer will
 938  * immediately access remote memory to construct this group's load-balance
 939  * statistics having the groups node local is of dubious benefit.
 940  */
 941 static struct sched_group *
 942 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 943 {
 944         struct sched_group *sg;
 945         struct cpumask *sg_span;
 946
 947         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 948                         GFP_KERNEL, cpu_to_node(cpu));
 949
 950         if (!sg)
 951                 return NULL;
 952
 953         sg_span = sched_group_span(sg);
 954         if (sd->child) {
 955                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 956                 sg->flags = sd->child->flags;
 957         } else {
 958                 cpumask_copy(sg_span, sched_domain_span(sd));
 959         }
 960
 961         atomic_inc(&sg->ref);
 962         return sg;
 963 }
 964
 965 static void init_overlap_sched_group(struct sched_domain *sd,
 966                                      struct sched_group *sg)
 967 {
 968         struct cpumask *mask = sched_domains_tmpmask2;
 969         struct sd_data *sdd = sd->private;
 970         struct cpumask *sg_span;
 971         int cpu;
 972
 973         build_balance_mask(sd, sg, mask);
 974         cpu = cpumask_first(mask);
 975
 976         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 977         if (atomic_inc_return(&sg->sgc->ref) == 1)
 978                 cpumask_copy(group_balance_mask(sg), mask);
 979         else
 980                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 981
 982         /*
 983          * Initialize sgc->capacity such that even if we mess up the
 984          * domains and no possible iteration will get us here, we won't
 985          * die on a /0 trap.
 986          */
 987         sg_span = sched_group_span(sg);
 988         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 989         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 990         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 991 }
 992
 993 static struct sched_domain *
 994 find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
 995 {
 996         /*
 997          * The proper descendant would be the one whose child won't span out
 998          * of sd
 999          */
1000         while (sibling->child &&
1001                !cpumask_subset(sched_domain_span(sibling->child),
1002                                sched_domain_span(sd)))
1003                 sibling = sibling->child;
1004
1005         /*
1006          * As we are referencing sgc across different topology level, we need
1007          * to go down to skip those sched_domains which don't contribute to
1008          * scheduling because they will be degenerated in cpu_attach_domain
1009          */
1010         while (sibling->child &&
1011                cpumask_equal(sched_domain_span(sibling->child),
1012                              sched_domain_span(sibling)))
1013                 sibling = sibling->child;
1014
1015         return sibling;
1016 }
1017
1018 static int
1019 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1020 {
1021         struct sched_group *first = NULL, *last = NULL, *sg;
1022         const struct cpumask *span = sched_domain_span(sd);
1023         struct cpumask *covered = sched_domains_tmpmask;
1024         struct sd_data *sdd = sd->private;
1025         struct sched_domain *sibling;
1026         int i;
1027
1028         cpumask_clear(covered);
1029
1030         for_each_cpu_wrap(i, span, cpu) {
1031                 struct cpumask *sg_span;
1032
1033                 if (cpumask_test_cpu(i, covered))
1034                         continue;
1035
1036                 sibling = *per_cpu_ptr(sdd->sd, i);
1037
1038                 /*
1039                  * Asymmetric node setups can result in situations where the
1040                  * domain tree is of unequal depth, make sure to skip domains
1041                  * that already cover the entire range.
1042                  *
1043                  * In that case build_sched_domains() will have terminated the
1044                  * iteration early and our sibling sd spans will be empty.
1045                  * Domains should always include the CPU they're built on, so
1046                  * check that.
1047                  */
1048                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1049                         continue;
1050
1051                 /*
1052                  * Usually we build sched_group by sibling's child sched_domain
1053                  * But for machines whose NUMA diameter are 3 or above, we move
1054                  * to build sched_group by sibling's proper descendant's child
1055                  * domain because sibling's child sched_domain will span out of
1056                  * the sched_domain being built as below.
1057                  *
1058                  * Smallest diameter=3 topology is:
1059                  *
1060                  *   node   0   1   2   3
1061                  *     0:  10  20  30  40
1062                  *     1:  20  10  20  30
1063                  *     2:  30  20  10  20
1064                  *     3:  40  30  20  10
1065                  *
1066                  *   0 --- 1 --- 2 --- 3
1067                  *
1068                  * NUMA-3       0-3             N/A             N/A             0-3
1069                  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
1070                  *
1071                  * NUMA-2       0-2             0-3             0-3             1-3
1072                  *  groups:     {0-1},{1-3}     {0-2},{2-3}     {1-3},{0-1}     {2-3},{0-2}
1073                  *
1074                  * NUMA-1       0-1             0-2             1-3             2-3
1075                  *  groups:     {0},{1}         {1},{2},{0}     {2},{3},{1}     {3},{2}
1076                  *
1077                  * NUMA-0       0               1               2               3
1078                  *
1079                  * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1080                  * group span isn't a subset of the domain span.
1081                  */
1082                 if (sibling->child &&
1083                     !cpumask_subset(sched_domain_span(sibling->child), span))
1084                         sibling = find_descended_sibling(sd, sibling);
1085
1086                 sg = build_group_from_child_sched_domain(sibling, cpu);
1087                 if (!sg)
1088                         goto fail;
1089
1090                 sg_span = sched_group_span(sg);
1091                 cpumask_or(covered, covered, sg_span);
1092
1093                 init_overlap_sched_group(sibling, sg);
1094
1095                 if (!first)
1096                         first = sg;
1097                 if (last)
1098                         last->next = sg;
1099                 last = sg;
1100                 last->next = first;
1101         }
1102         sd->groups = first;
1103
1104         return 0;
1105
1106 fail:
1107         free_sched_groups(first, 0);
1108
1109         return -ENOMEM;
1110 }
1111
1112
1113 /*
1114  * Package topology (also see the load-balance blurb in fair.c)
1115  *
1116  * The scheduler builds a tree structure to represent a number of important
1117  * topology features. By default (default_topology[]) these include:
1118  *
1119  *  - Simultaneous multithreading (SMT)
1120  *  - Multi-Core Cache (MC)
1121  *  - Package (DIE)
1122  *
1123  * Where the last one more or less denotes everything up to a NUMA node.
1124  *
1125  * The tree consists of 3 primary data structures:
1126  *
1127  *      sched_domain -> sched_group -> sched_group_capacity
1128  *          ^ ^             ^ ^
1129  *          `-'             `-'
1130  *
1131  * The sched_domains are per-CPU and have a two way link (parent & child) and
1132  * denote the ever growing mask of CPUs belonging to that level of topology.
1133  *
1134  * Each sched_domain has a circular (double) linked list of sched_group's, each
1135  * denoting the domains of the level below (or individual CPUs in case of the
1136  * first domain level). The sched_group linked by a sched_domain includes the
1137  * CPU of that sched_domain [*].
1138  *
1139  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1140  *
1141  * CPU   0   1   2   3   4   5   6   7
1142  *
1143  * DIE  [                             ]
1144  * MC   [             ] [             ]
1145  * SMT  [     ] [     ] [     ] [     ]
1146  *
1147  *  - or -
1148  *
1149  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1150  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1151  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1152  *
1153  * CPU   0   1   2   3   4   5   6   7
1154  *
1155  * One way to think about it is: sched_domain moves you up and down among these
1156  * topology levels, while sched_group moves you sideways through it, at child
1157  * domain granularity.
1158  *
1159  * sched_group_capacity ensures each unique sched_group has shared storage.
1160  *
1161  * There are two related construction problems, both require a CPU that
1162  * uniquely identify each group (for a given domain):
1163  *
1164  *  - The first is the balance_cpu (see should_we_balance() and the
1165  *    load-balance blub in fair.c); for each group we only want 1 CPU to
1166  *    continue balancing at a higher domain.
1167  *
1168  *  - The second is the sched_group_capacity; we want all identical groups
1169  *    to share a single sched_group_capacity.
1170  *
1171  * Since these topologies are exclusive by construction. That is, its
1172  * impossible for an SMT thread to belong to multiple cores, and cores to
1173  * be part of multiple caches. There is a very clear and unique location
1174  * for each CPU in the hierarchy.
1175  *
1176  * Therefore computing a unique CPU for each group is trivial (the iteration
1177  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1178  * group), we can simply pick the first CPU in each group.
1179  *
1180  *
1181  * [*] in other words, the first group of each domain is its child domain.
1182  */
1183
1184 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1185 {
1186         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1187         struct sched_domain *child = sd->child;
1188         struct sched_group *sg;
1189         bool already_visited;
1190
1191         if (child)
1192                 cpu = cpumask_first(sched_domain_span(child));
1193
1194         sg = *per_cpu_ptr(sdd->sg, cpu);
1195         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1196
1197         /* Increase refcounts for claim_allocations: */
1198         already_visited = atomic_inc_return(&sg->ref) > 1;
1199         /* sgc visits should follow a similar trend as sg */
1200         WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1201
1202         /* If we have already visited that group, it's already initialized. */
1203         if (already_visited)
1204                 return sg;
1205
1206         if (child) {
1207                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1208                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1209                 sg->flags = child->flags;
1210         } else {
1211                 cpumask_set_cpu(cpu, sched_group_span(sg));
1212                 cpumask_set_cpu(cpu, group_balance_mask(sg));
1213         }
1214
1215         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1216         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1217         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1218
1219         return sg;
1220 }
1221
1222 /*
1223  * build_sched_groups will build a circular linked list of the groups
1224  * covered by the given span, will set each group's ->cpumask correctly,
1225  * and will initialize their ->sgc.
1226  *
1227  * Assumes the sched_domain tree is fully constructed
1228  */
1229 static int
1230 build_sched_groups(struct sched_domain *sd, int cpu)
1231 {
1232         struct sched_group *first = NULL, *last = NULL;
1233         struct sd_data *sdd = sd->private;
1234         const struct cpumask *span = sched_domain_span(sd);
1235         struct cpumask *covered;
1236         int i;
1237
1238         lockdep_assert_held(&sched_domains_mutex);
1239         covered = sched_domains_tmpmask;
1240
1241         cpumask_clear(covered);
1242
1243         for_each_cpu_wrap(i, span, cpu) {
1244                 struct sched_group *sg;
1245
1246                 if (cpumask_test_cpu(i, covered))
1247                         continue;
1248
1249                 sg = get_group(i, sdd);
1250
1251                 cpumask_or(covered, covered, sched_group_span(sg));
1252
1253                 if (!first)
1254                         first = sg;
1255                 if (last)
1256                         last->next = sg;
1257                 last = sg;
1258         }
1259         last->next = first;
1260         sd->groups = first;
1261
1262         return 0;
1263 }
1264
1265 /*
1266  * Initialize sched groups cpu_capacity.
1267  *
1268  * cpu_capacity indicates the capacity of sched group, which is used while
1269  * distributing the load between different sched groups in a sched domain.
1270  * Typically cpu_capacity for all the groups in a sched domain will be same
1271  * unless there are asymmetries in the topology. If there are asymmetries,
1272  * group having more cpu_capacity will pickup more load compared to the
1273  * group having less cpu_capacity.
1274  */
1275 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1276 {
1277         struct sched_group *sg = sd->groups;
1278
1279         WARN_ON(!sg);
1280
1281         do {
1282                 int cpu, max_cpu = -1;
1283
1284                 sg->group_weight = cpumask_weight(sched_group_span(sg));
1285
1286                 if (!(sd->flags & SD_ASYM_PACKING))
1287                         goto next;
1288
1289                 for_each_cpu(cpu, sched_group_span(sg)) {
1290                         if (max_cpu < 0)
1291                                 max_cpu = cpu;
1292                         else if (sched_asym_prefer(cpu, max_cpu))
1293                                 max_cpu = cpu;
1294                 }
1295                 sg->asym_prefer_cpu = max_cpu;
1296
1297 next:
1298                 sg = sg->next;
1299         } while (sg != sd->groups);
1300
1301         if (cpu != group_balance_cpu(sg))
1302                 return;
1303
1304         update_group_capacity(sd, cpu);
1305 }
1306
1307 /*
1308  * Asymmetric CPU capacity bits
1309  */
1310 struct asym_cap_data {
1311         struct list_head link;
1312         unsigned long capacity;
1313         unsigned long cpus[];
1314 };
1315
1316 /*
1317  * Set of available CPUs grouped by their corresponding capacities
1318  * Each list entry contains a CPU mask reflecting CPUs that share the same
1319  * capacity.
1320  * The lifespan of data is unlimited.
1321  */
1322 static LIST_HEAD(asym_cap_list);
1323
1324 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
1325
1326 /*
1327  * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1328  * Provides sd_flags reflecting the asymmetry scope.
1329  */
1330 static inline int
1331 asym_cpu_capacity_classify(const struct cpumask *sd_span,
1332                            const struct cpumask *cpu_map)
1333 {
1334         struct asym_cap_data *entry;
1335         int count = 0, miss = 0;
1336
1337         /*
1338          * Count how many unique CPU capacities this domain spans across
1339          * (compare sched_domain CPUs mask with ones representing  available
1340          * CPUs capacities). Take into account CPUs that might be offline:
1341          * skip those.
1342          */
1343         list_for_each_entry(entry, &asym_cap_list, link) {
1344                 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1345                         ++count;
1346                 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1347                         ++miss;
1348         }
1349
1350         WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1351
1352         /* No asymmetry detected */
1353         if (count < 2)
1354                 return 0;
1355         /* Some of the available CPU capacity values have not been detected */
1356         if (miss)
1357                 return SD_ASYM_CPUCAPACITY;
1358
1359         /* Full asymmetry */
1360         return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1361
1362 }
1363
1364 static inline void asym_cpu_capacity_update_data(int cpu)
1365 {
1366         unsigned long capacity = arch_scale_cpu_capacity(cpu);
1367         struct asym_cap_data *entry = NULL;
1368
1369         list_for_each_entry(entry, &asym_cap_list, link) {
1370                 if (capacity == entry->capacity)
1371                         goto done;
1372         }
1373
1374         entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1375         if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1376                 return;
1377         entry->capacity = capacity;
1378         list_add(&entry->link, &asym_cap_list);
1379 done:
1380         __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1381 }
1382
1383 /*
1384  * Build-up/update list of CPUs grouped by their capacities
1385  * An update requires explicit request to rebuild sched domains
1386  * with state indicating CPU topology changes.
1387  */
1388 static void asym_cpu_capacity_scan(void)
1389 {
1390         struct asym_cap_data *entry, *next;
1391         int cpu;
1392
1393         list_for_each_entry(entry, &asym_cap_list, link)
1394                 cpumask_clear(cpu_capacity_span(entry));
1395
1396         for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1397                 asym_cpu_capacity_update_data(cpu);
1398
1399         list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1400                 if (cpumask_empty(cpu_capacity_span(entry))) {
1401                         list_del(&entry->link);
1402                         kfree(entry);
1403                 }
1404         }
1405
1406         /*
1407          * Only one capacity value has been detected i.e. this system is symmetric.
1408          * No need to keep this data around.
1409          */
1410         if (list_is_singular(&asym_cap_list)) {
1411                 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1412                 list_del(&entry->link);
1413                 kfree(entry);
1414         }
1415 }
1416
1417 /*
1418  * Initializers for schedule domains
1419  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1420  */
1421
1422 static int default_relax_domain_level = -1;
1423 int sched_domain_level_max;
1424
1425 static int __init setup_relax_domain_level(char *str)
1426 {
1427         if (kstrtoint(str, 0, &default_relax_domain_level))
1428                 pr_warn("Unable to set relax_domain_level\n");
1429
1430         return 1;
1431 }
1432 __setup("relax_domain_level=", setup_relax_domain_level);
1433
1434 static void set_domain_attribute(struct sched_domain *sd,
1435                                  struct sched_domain_attr *attr)
1436 {
1437         int request;
1438
1439         if (!attr || attr->relax_domain_level < 0) {
1440                 if (default_relax_domain_level < 0)
1441                         return;
1442                 request = default_relax_domain_level;
1443         } else
1444                 request = attr->relax_domain_level;
1445
1446         if (sd->level > request) {
1447                 /* Turn off idle balance on this domain: */
1448                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1449         }
1450 }
1451
1452 static void __sdt_free(const struct cpumask *cpu_map);
1453 static int __sdt_alloc(const struct cpumask *cpu_map);
1454
1455 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1456                                  const struct cpumask *cpu_map)
1457 {
1458         switch (what) {
1459         case sa_rootdomain:
1460                 if (!atomic_read(&d->rd->refcount))
1461                         free_rootdomain(&d->rd->rcu);
1462                 fallthrough;
1463         case sa_sd:
1464                 free_percpu(d->sd);
1465                 fallthrough;
1466         case sa_sd_storage:
1467                 __sdt_free(cpu_map);
1468                 fallthrough;
1469         case sa_none:
1470                 break;
1471         }
1472 }
1473
1474 static enum s_alloc
1475 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1476 {
1477         memset(d, 0, sizeof(*d));
1478
1479         if (__sdt_alloc(cpu_map))
1480                 return sa_sd_storage;
1481         d->sd = alloc_percpu(struct sched_domain *);
1482         if (!d->sd)
1483                 return sa_sd_storage;
1484         d->rd = alloc_rootdomain();
1485         if (!d->rd)
1486                 return sa_sd;
1487
1488         return sa_rootdomain;
1489 }
1490
1491 /*
1492  * NULL the sd_data elements we've used to build the sched_domain and
1493  * sched_group structure so that the subsequent __free_domain_allocs()
1494  * will not free the data we're using.
1495  */
1496 static void claim_allocations(int cpu, struct sched_domain *sd)
1497 {
1498         struct sd_data *sdd = sd->private;
1499
1500         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1501         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1502
1503         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1504                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1505
1506         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1507                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1508
1509         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1510                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1511 }
1512
1513 #ifdef CONFIG_NUMA
1514 enum numa_topology_type sched_numa_topology_type;
1515
1516 static int                      sched_domains_numa_levels;
1517 static int                      sched_domains_curr_level;
1518
1519 int                             sched_max_numa_distance;
1520 static int                      *sched_domains_numa_distance;
1521 static struct cpumask           ***sched_domains_numa_masks;
1522 #endif
1523
1524 /*
1525  * SD_flags allowed in topology descriptions.
1526  *
1527  * These flags are purely descriptive of the topology and do not prescribe
1528  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1529  * function:
1530  *
1531  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1532  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1533  *   SD_NUMA                - describes NUMA topologies
1534  *
1535  * Odd one out, which beside describing the topology has a quirk also
1536  * prescribes the desired behaviour that goes along with it:
1537  *
1538  *   SD_ASYM_PACKING        - describes SMT quirks
1539  */
1540 #define TOPOLOGY_SD_FLAGS               \
1541         (SD_SHARE_CPUCAPACITY   |       \
1542          SD_SHARE_PKG_RESOURCES |       \
1543          SD_NUMA                |       \
1544          SD_ASYM_PACKING)
1545
1546 static struct sched_domain *
1547 sd_init(struct sched_domain_topology_level *tl,
1548         const struct cpumask *cpu_map,
1549         struct sched_domain *child, int cpu)
1550 {
1551         struct sd_data *sdd = &tl->data;
1552         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1553         int sd_id, sd_weight, sd_flags = 0;
1554         struct cpumask *sd_span;
1555
1556 #ifdef CONFIG_NUMA
1557         /*
1558          * Ugly hack to pass state to sd_numa_mask()...
1559          */
1560         sched_domains_curr_level = tl->numa_level;
1561 #endif
1562
1563         sd_weight = cpumask_weight(tl->mask(cpu));
1564
1565         if (tl->sd_flags)
1566                 sd_flags = (*tl->sd_flags)();
1567         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1568                         "wrong sd_flags in topology description\n"))
1569                 sd_flags &= TOPOLOGY_SD_FLAGS;
1570
1571         *sd = (struct sched_domain){
1572                 .min_interval           = sd_weight,
1573                 .max_interval           = 2*sd_weight,
1574                 .busy_factor            = 16,
1575                 .imbalance_pct          = 117,
1576
1577                 .cache_nice_tries       = 0,
1578
1579                 .flags                  = 1*SD_BALANCE_NEWIDLE
1580                                         | 1*SD_BALANCE_EXEC
1581                                         | 1*SD_BALANCE_FORK
1582                                         | 0*SD_BALANCE_WAKE
1583                                         | 1*SD_WAKE_AFFINE
1584                                         | 0*SD_SHARE_CPUCAPACITY
1585                                         | 0*SD_SHARE_PKG_RESOURCES
1586                                         | 0*SD_SERIALIZE
1587                                         | 1*SD_PREFER_SIBLING
1588                                         | 0*SD_NUMA
1589                                         | sd_flags
1590                                         ,
1591
1592                 .last_balance           = jiffies,
1593                 .balance_interval       = sd_weight,
1594                 .max_newidle_lb_cost    = 0,
1595                 .last_decay_max_lb_cost = jiffies,
1596                 .child                  = child,
1597 #ifdef CONFIG_SCHED_DEBUG
1598                 .name                   = tl->name,
1599 #endif
1600         };
1601
1602         sd_span = sched_domain_span(sd);
1603         cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1604         sd_id = cpumask_first(sd_span);
1605
1606         sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1607
1608         WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1609                   (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1610                   "CPU capacity asymmetry not supported on SMT\n");
1611
1612         /*
1613          * Convert topological properties into behaviour.
1614          */
1615         /* Don't attempt to spread across CPUs of different capacities. */
1616         if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1617                 sd->child->flags &= ~SD_PREFER_SIBLING;
1618
1619         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1620                 sd->imbalance_pct = 110;
1621
1622         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1623                 sd->imbalance_pct = 117;
1624                 sd->cache_nice_tries = 1;
1625
1626 #ifdef CONFIG_NUMA
1627         } else if (sd->flags & SD_NUMA) {
1628                 sd->cache_nice_tries = 2;
1629
1630                 sd->flags &= ~SD_PREFER_SIBLING;
1631                 sd->flags |= SD_SERIALIZE;
1632                 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1633                         sd->flags &= ~(SD_BALANCE_EXEC |
1634                                        SD_BALANCE_FORK |
1635                                        SD_WAKE_AFFINE);
1636                 }
1637
1638 #endif
1639         } else {
1640                 sd->cache_nice_tries = 1;
1641         }
1642
1643         /*
1644          * For all levels sharing cache; connect a sched_domain_shared
1645          * instance.
1646          */
1647         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1648                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1649                 atomic_inc(&sd->shared->ref);
1650                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1651         }
1652
1653         sd->private = sdd;
1654
1655         return sd;
1656 }
1657
1658 /*
1659  * Topology list, bottom-up.
1660  */
1661 static struct sched_domain_topology_level default_topology[] = {
1662 #ifdef CONFIG_SCHED_SMT
1663         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1664 #endif
1665
1666 #ifdef CONFIG_SCHED_CLUSTER
1667         { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1668 #endif
1669
1670 #ifdef CONFIG_SCHED_MC
1671         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1672 #endif
1673         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1674         { NULL, },
1675 };
1676
1677 static struct sched_domain_topology_level *sched_domain_topology =
1678         default_topology;
1679 static struct sched_domain_topology_level *sched_domain_topology_saved;
1680
1681 #define for_each_sd_topology(tl)                        \
1682         for (tl = sched_domain_topology; tl->mask; tl++)
1683
1684 void __init set_sched_topology(struct sched_domain_topology_level *tl)
1685 {
1686         if (WARN_ON_ONCE(sched_smp_initialized))
1687                 return;
1688
1689         sched_domain_topology = tl;
1690         sched_domain_topology_saved = NULL;
1691 }
1692
1693 #ifdef CONFIG_NUMA
1694
1695 static const struct cpumask *sd_numa_mask(int cpu)
1696 {
1697         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1698 }
1699
1700 static void sched_numa_warn(const char *str)
1701 {
1702         static int done = false;
1703         int i,j;
1704
1705         if (done)
1706                 return;
1707
1708         done = true;
1709
1710         printk(KERN_WARNING "ERROR: %s\n\n", str);
1711
1712         for (i = 0; i < nr_node_ids; i++) {
1713                 printk(KERN_WARNING "  ");
1714                 for (j = 0; j < nr_node_ids; j++) {
1715                         if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1716                                 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1717                         else
1718                                 printk(KERN_CONT " %02d  ", node_distance(i,j));
1719                 }
1720                 printk(KERN_CONT "\n");
1721         }
1722         printk(KERN_WARNING "\n");
1723 }
1724
1725 bool find_numa_distance(int distance)
1726 {
1727         bool found = false;
1728         int i, *distances;
1729
1730         if (distance == node_distance(0, 0))
1731                 return true;
1732
1733         rcu_read_lock();
1734         distances = rcu_dereference(sched_domains_numa_distance);
1735         if (!distances)
1736                 goto unlock;
1737         for (i = 0; i < sched_domains_numa_levels; i++) {
1738                 if (distances[i] == distance) {
1739                         found = true;
1740                         break;
1741                 }
1742         }
1743 unlock:
1744         rcu_read_unlock();
1745
1746         return found;
1747 }
1748
1749 #define for_each_cpu_node_but(n, nbut)          \
1750         for_each_node_state(n, N_CPU)           \
1751                 if (n == nbut)                  \
1752                         continue;               \
1753                 else
1754
1755 /*
1756  * A system can have three types of NUMA topology:
1757  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1758  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1759  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1760  *
1761  * The difference between a glueless mesh topology and a backplane
1762  * topology lies in whether communication between not directly
1763  * connected nodes goes through intermediary nodes (where programs
1764  * could run), or through backplane controllers. This affects
1765  * placement of programs.
1766  *
1767  * The type of topology can be discerned with the following tests:
1768  * - If the maximum distance between any nodes is 1 hop, the system
1769  *   is directly connected.
1770  * - If for two nodes A and B, located N > 1 hops away from each other,
1771  *   there is an intermediary node C, which is < N hops away from both
1772  *   nodes A and B, the system is a glueless mesh.
1773  */
1774 static void init_numa_topology_type(int offline_node)
1775 {
1776         int a, b, c, n;
1777
1778         n = sched_max_numa_distance;
1779
1780         if (sched_domains_numa_levels <= 2) {
1781                 sched_numa_topology_type = NUMA_DIRECT;
1782                 return;
1783         }
1784
1785         for_each_cpu_node_but(a, offline_node) {
1786                 for_each_cpu_node_but(b, offline_node) {
1787                         /* Find two nodes furthest removed from each other. */
1788                         if (node_distance(a, b) < n)
1789                                 continue;
1790
1791                         /* Is there an intermediary node between a and b? */
1792                         for_each_cpu_node_but(c, offline_node) {
1793                                 if (node_distance(a, c) < n &&
1794                                     node_distance(b, c) < n) {
1795                                         sched_numa_topology_type =
1796                                                         NUMA_GLUELESS_MESH;
1797                                         return;
1798                                 }
1799                         }
1800
1801                         sched_numa_topology_type = NUMA_BACKPLANE;
1802                         return;
1803                 }
1804         }
1805
1806         pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1807         sched_numa_topology_type = NUMA_DIRECT;
1808 }
1809
1810
1811 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1812
1813 void sched_init_numa(int offline_node)
1814 {
1815         struct sched_domain_topology_level *tl;
1816         unsigned long *distance_map;
1817         int nr_levels = 0;
1818         int i, j;
1819         int *distances;
1820         struct cpumask ***masks;
1821
1822         /*
1823          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1824          * unique distances in the node_distance() table.
1825          */
1826         distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1827         if (!distance_map)
1828                 return;
1829
1830         bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1831         for_each_cpu_node_but(i, offline_node) {
1832                 for_each_cpu_node_but(j, offline_node) {
1833                         int distance = node_distance(i, j);
1834
1835                         if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1836                                 sched_numa_warn("Invalid distance value range");
1837                                 bitmap_free(distance_map);
1838                                 return;
1839                         }
1840
1841                         bitmap_set(distance_map, distance, 1);
1842                 }
1843         }
1844         /*
1845          * We can now figure out how many unique distance values there are and
1846          * allocate memory accordingly.
1847          */
1848         nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1849
1850         distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1851         if (!distances) {
1852                 bitmap_free(distance_map);
1853                 return;
1854         }
1855
1856         for (i = 0, j = 0; i < nr_levels; i++, j++) {
1857                 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1858                 distances[i] = j;
1859         }
1860         rcu_assign_pointer(sched_domains_numa_distance, distances);
1861
1862         bitmap_free(distance_map);
1863
1864         /*
1865          * 'nr_levels' contains the number of unique distances
1866          *
1867          * The sched_domains_numa_distance[] array includes the actual distance
1868          * numbers.
1869          */
1870
1871         /*
1872          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1873          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1874          * the array will contain less then 'nr_levels' members. This could be
1875          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1876          * in other functions.
1877          *
1878          * We reset it to 'nr_levels' at the end of this function.
1879          */
1880         sched_domains_numa_levels = 0;
1881
1882         masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1883         if (!masks)
1884                 return;
1885
1886         /*
1887          * Now for each level, construct a mask per node which contains all
1888          * CPUs of nodes that are that many hops away from us.
1889          */
1890         for (i = 0; i < nr_levels; i++) {
1891                 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1892                 if (!masks[i])
1893                         return;
1894
1895                 for_each_cpu_node_but(j, offline_node) {
1896                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1897                         int k;
1898
1899                         if (!mask)
1900                                 return;
1901
1902                         masks[i][j] = mask;
1903
1904                         for_each_cpu_node_but(k, offline_node) {
1905                                 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1906                                         sched_numa_warn("Node-distance not symmetric");
1907
1908                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1909                                         continue;
1910
1911                                 cpumask_or(mask, mask, cpumask_of_node(k));
1912                         }
1913                 }
1914         }
1915         rcu_assign_pointer(sched_domains_numa_masks, masks);
1916
1917         /* Compute default topology size */
1918         for (i = 0; sched_domain_topology[i].mask; i++);
1919
1920         tl = kzalloc((i + nr_levels + 1) *
1921                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1922         if (!tl)
1923                 return;
1924
1925         /*
1926          * Copy the default topology bits..
1927          */
1928         for (i = 0; sched_domain_topology[i].mask; i++)
1929                 tl[i] = sched_domain_topology[i];
1930
1931         /*
1932          * Add the NUMA identity distance, aka single NODE.
1933          */
1934         tl[i++] = (struct sched_domain_topology_level){
1935                 .mask = sd_numa_mask,
1936                 .numa_level = 0,
1937                 SD_INIT_NAME(NODE)
1938         };
1939
1940         /*
1941          * .. and append 'j' levels of NUMA goodness.
1942          */
1943         for (j = 1; j < nr_levels; i++, j++) {
1944                 tl[i] = (struct sched_domain_topology_level){
1945                         .mask = sd_numa_mask,
1946                         .sd_flags = cpu_numa_flags,
1947                         .flags = SDTL_OVERLAP,
1948                         .numa_level = j,
1949                         SD_INIT_NAME(NUMA)
1950                 };
1951         }
1952
1953         sched_domain_topology_saved = sched_domain_topology;
1954         sched_domain_topology = tl;
1955
1956         sched_domains_numa_levels = nr_levels;
1957         WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1958
1959         init_numa_topology_type(offline_node);
1960 }
1961
1962
1963 static void sched_reset_numa(void)
1964 {
1965         int nr_levels, *distances;
1966         struct cpumask ***masks;
1967
1968         nr_levels = sched_domains_numa_levels;
1969         sched_domains_numa_levels = 0;
1970         sched_max_numa_distance = 0;
1971         sched_numa_topology_type = NUMA_DIRECT;
1972         distances = sched_domains_numa_distance;
1973         rcu_assign_pointer(sched_domains_numa_distance, NULL);
1974         masks = sched_domains_numa_masks;
1975         rcu_assign_pointer(sched_domains_numa_masks, NULL);
1976         if (distances || masks) {
1977                 int i, j;
1978
1979                 synchronize_rcu();
1980                 kfree(distances);
1981                 for (i = 0; i < nr_levels && masks; i++) {
1982                         if (!masks[i])
1983                                 continue;
1984                         for_each_node(j)
1985                                 kfree(masks[i][j]);
1986                         kfree(masks[i]);
1987                 }
1988                 kfree(masks);
1989         }
1990         if (sched_domain_topology_saved) {
1991                 kfree(sched_domain_topology);
1992                 sched_domain_topology = sched_domain_topology_saved;
1993                 sched_domain_topology_saved = NULL;
1994         }
1995 }
1996
1997 /*
1998  * Call with hotplug lock held
1999  */
2000 void sched_update_numa(int cpu, bool online)
2001 {
2002         int node;
2003
2004         node = cpu_to_node(cpu);
2005         /*
2006          * Scheduler NUMA topology is updated when the first CPU of a
2007          * node is onlined or the last CPU of a node is offlined.
2008          */
2009         if (cpumask_weight(cpumask_of_node(node)) != 1)
2010                 return;
2011
2012         sched_reset_numa();
2013         sched_init_numa(online ? NUMA_NO_NODE : node);
2014 }
2015
2016 void sched_domains_numa_masks_set(unsigned int cpu)
2017 {
2018         int node = cpu_to_node(cpu);
2019         int i, j;
2020
2021         for (i = 0; i < sched_domains_numa_levels; i++) {
2022                 for (j = 0; j < nr_node_ids; j++) {
2023                         if (!node_state(j, N_CPU))
2024                                 continue;
2025
2026                         /* Set ourselves in the remote node's masks */
2027                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
2028                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2029                 }
2030         }
2031 }
2032
2033 void sched_domains_numa_masks_clear(unsigned int cpu)
2034 {
2035         int i, j;
2036
2037         for (i = 0; i < sched_domains_numa_levels; i++) {
2038                 for (j = 0; j < nr_node_ids; j++) {
2039                         if (sched_domains_numa_masks[i][j])
2040                                 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2041                 }
2042         }
2043 }
2044
2045 /*
2046  * sched_numa_find_closest() - given the NUMA topology, find the cpu
2047  *                             closest to @cpu from @cpumask.
2048  * cpumask: cpumask to find a cpu from
2049  * cpu: cpu to be close to
2050  *
2051  * returns: cpu, or nr_cpu_ids when nothing found.
2052  */
2053 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2054 {
2055         int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2056         struct cpumask ***masks;
2057
2058         rcu_read_lock();
2059         masks = rcu_dereference(sched_domains_numa_masks);
2060         if (!masks)
2061                 goto unlock;
2062         for (i = 0; i < sched_domains_numa_levels; i++) {
2063                 if (!masks[i][j])
2064                         break;
2065                 cpu = cpumask_any_and(cpus, masks[i][j]);
2066                 if (cpu < nr_cpu_ids) {
2067                         found = cpu;
2068                         break;
2069                 }
2070         }
2071 unlock:
2072         rcu_read_unlock();
2073
2074         return found;
2075 }
2076
2077 struct __cmp_key {
2078         const struct cpumask *cpus;
2079         struct cpumask ***masks;
2080         int node;
2081         int cpu;
2082         int w;
2083 };
2084
2085 static int hop_cmp(const void *a, const void *b)
2086 {
2087         struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2088         struct __cmp_key *k = (struct __cmp_key *)a;
2089
2090         if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2091                 return 1;
2092
2093         if (b == k->masks) {
2094                 k->w = 0;
2095                 return 0;
2096         }
2097
2098         prev_hop = *((struct cpumask ***)b - 1);
2099         k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2100         if (k->w <= k->cpu)
2101                 return 0;
2102
2103         return -1;
2104 }
2105
2106 /*
2107  * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth next cpu
2108  *                             closest to @cpu from @cpumask.
2109  * cpumask: cpumask to find a cpu from
2110  * cpu: Nth cpu to find
2111  *
2112  * returns: cpu, or nr_cpu_ids when nothing found.
2113  */
2114 int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2115 {
2116         struct __cmp_key k = { .cpus = cpus, .node = node, .cpu = cpu };
2117         struct cpumask ***hop_masks;
2118         int hop, ret = nr_cpu_ids;
2119
2120         rcu_read_lock();
2121
2122         k.masks = rcu_dereference(sched_domains_numa_masks);
2123         if (!k.masks)
2124                 goto unlock;
2125
2126         hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2127         hop = hop_masks - k.masks;
2128
2129         ret = hop ?
2130                 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2131                 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2132 unlock:
2133         rcu_read_unlock();
2134         return ret;
2135 }
2136 EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2137
2138 /**
2139  * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2140  *                         @node
2141  * @node: The node to count hops from.
2142  * @hops: Include CPUs up to that many hops away. 0 means local node.
2143  *
2144  * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2145  * @node, an error value otherwise.
2146  *
2147  * Requires rcu_lock to be held. Returned cpumask is only valid within that
2148  * read-side section, copy it if required beyond that.
2149  *
2150  * Note that not all hops are equal in distance; see sched_init_numa() for how
2151  * distances and masks are handled.
2152  * Also note that this is a reflection of sched_domains_numa_masks, which may change
2153  * during the lifetime of the system (offline nodes are taken out of the masks).
2154  */
2155 const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2156 {
2157         struct cpumask ***masks;
2158
2159         if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2160                 return ERR_PTR(-EINVAL);
2161
2162         masks = rcu_dereference(sched_domains_numa_masks);
2163         if (!masks)
2164                 return ERR_PTR(-EBUSY);
2165
2166         return masks[hops][node];
2167 }
2168 EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2169
2170 #endif /* CONFIG_NUMA */
2171
2172 static int __sdt_alloc(const struct cpumask *cpu_map)
2173 {
2174         struct sched_domain_topology_level *tl;
2175         int j;
2176
2177         for_each_sd_topology(tl) {
2178                 struct sd_data *sdd = &tl->data;
2179
2180                 sdd->sd = alloc_percpu(struct sched_domain *);
2181                 if (!sdd->sd)
2182                         return -ENOMEM;
2183
2184                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2185                 if (!sdd->sds)
2186                         return -ENOMEM;
2187
2188                 sdd->sg = alloc_percpu(struct sched_group *);
2189                 if (!sdd->sg)
2190                         return -ENOMEM;
2191
2192                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2193                 if (!sdd->sgc)
2194                         return -ENOMEM;
2195
2196                 for_each_cpu(j, cpu_map) {
2197                         struct sched_domain *sd;
2198                         struct sched_domain_shared *sds;
2199                         struct sched_group *sg;
2200                         struct sched_group_capacity *sgc;
2201
2202                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2203                                         GFP_KERNEL, cpu_to_node(j));
2204                         if (!sd)
2205                                 return -ENOMEM;
2206
2207                         *per_cpu_ptr(sdd->sd, j) = sd;
2208
2209                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
2210                                         GFP_KERNEL, cpu_to_node(j));
2211                         if (!sds)
2212                                 return -ENOMEM;
2213
2214                         *per_cpu_ptr(sdd->sds, j) = sds;
2215
2216                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2217                                         GFP_KERNEL, cpu_to_node(j));
2218                         if (!sg)
2219                                 return -ENOMEM;
2220
2221                         sg->next = sg;
2222
2223                         *per_cpu_ptr(sdd->sg, j) = sg;
2224
2225                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2226                                         GFP_KERNEL, cpu_to_node(j));
2227                         if (!sgc)
2228                                 return -ENOMEM;
2229
2230 #ifdef CONFIG_SCHED_DEBUG
2231                         sgc->id = j;
2232 #endif
2233
2234                         *per_cpu_ptr(sdd->sgc, j) = sgc;
2235                 }
2236         }
2237
2238         return 0;
2239 }
2240
2241 static void __sdt_free(const struct cpumask *cpu_map)
2242 {
2243         struct sched_domain_topology_level *tl;
2244         int j;
2245
2246         for_each_sd_topology(tl) {
2247                 struct sd_data *sdd = &tl->data;
2248
2249                 for_each_cpu(j, cpu_map) {
2250                         struct sched_domain *sd;
2251
2252                         if (sdd->sd) {
2253                                 sd = *per_cpu_ptr(sdd->sd, j);
2254                                 if (sd && (sd->flags & SD_OVERLAP))
2255                                         free_sched_groups(sd->groups, 0);
2256                                 kfree(*per_cpu_ptr(sdd->sd, j));
2257                         }
2258
2259                         if (sdd->sds)
2260                                 kfree(*per_cpu_ptr(sdd->sds, j));
2261                         if (sdd->sg)
2262                                 kfree(*per_cpu_ptr(sdd->sg, j));
2263                         if (sdd->sgc)
2264                                 kfree(*per_cpu_ptr(sdd->sgc, j));
2265                 }
2266                 free_percpu(sdd->sd);
2267                 sdd->sd = NULL;
2268                 free_percpu(sdd->sds);
2269                 sdd->sds = NULL;
2270                 free_percpu(sdd->sg);
2271                 sdd->sg = NULL;
2272                 free_percpu(sdd->sgc);
2273                 sdd->sgc = NULL;
2274         }
2275 }
2276
2277 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2278                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2279                 struct sched_domain *child, int cpu)
2280 {
2281         struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2282
2283         if (child) {
2284                 sd->level = child->level + 1;
2285                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2286                 child->parent = sd;
2287
2288                 if (!cpumask_subset(sched_domain_span(child),
2289                                     sched_domain_span(sd))) {
2290                         pr_err("BUG: arch topology borken\n");
2291 #ifdef CONFIG_SCHED_DEBUG
2292                         pr_err("     the %s domain not a subset of the %s domain\n",
2293                                         child->name, sd->name);
2294 #endif
2295                         /* Fixup, ensure @sd has at least @child CPUs. */
2296                         cpumask_or(sched_domain_span(sd),
2297                                    sched_domain_span(sd),
2298                                    sched_domain_span(child));
2299                 }
2300
2301         }
2302         set_domain_attribute(sd, attr);
2303
2304         return sd;
2305 }
2306
2307 /*
2308  * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2309  * any two given CPUs at this (non-NUMA) topology level.
2310  */
2311 static bool topology_span_sane(struct sched_domain_topology_level *tl,
2312                               const struct cpumask *cpu_map, int cpu)
2313 {
2314         int i;
2315
2316         /* NUMA levels are allowed to overlap */
2317         if (tl->flags & SDTL_OVERLAP)
2318                 return true;
2319
2320         /*
2321          * Non-NUMA levels cannot partially overlap - they must be either
2322          * completely equal or completely disjoint. Otherwise we can end up
2323          * breaking the sched_group lists - i.e. a later get_group() pass
2324          * breaks the linking done for an earlier span.
2325          */
2326         for_each_cpu(i, cpu_map) {
2327                 if (i == cpu)
2328                         continue;
2329                 /*
2330                  * We should 'and' all those masks with 'cpu_map' to exactly
2331                  * match the topology we're about to build, but that can only
2332                  * remove CPUs, which only lessens our ability to detect
2333                  * overlaps
2334                  */
2335                 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2336                     cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2337                         return false;
2338         }
2339
2340         return true;
2341 }
2342
2343 /*
2344  * Build sched domains for a given set of CPUs and attach the sched domains
2345  * to the individual CPUs
2346  */
2347 static int
2348 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2349 {
2350         enum s_alloc alloc_state = sa_none;
2351         struct sched_domain *sd;
2352         struct s_data d;
2353         struct rq *rq = NULL;
2354         int i, ret = -ENOMEM;
2355         bool has_asym = false;
2356
2357         if (WARN_ON(cpumask_empty(cpu_map)))
2358                 goto error;
2359
2360         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2361         if (alloc_state != sa_rootdomain)
2362                 goto error;
2363
2364         /* Set up domains for CPUs specified by the cpu_map: */
2365         for_each_cpu(i, cpu_map) {
2366                 struct sched_domain_topology_level *tl;
2367
2368                 sd = NULL;
2369                 for_each_sd_topology(tl) {
2370
2371                         if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2372                                 goto error;
2373
2374                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2375
2376                         has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2377
2378                         if (tl == sched_domain_topology)
2379                                 *per_cpu_ptr(d.sd, i) = sd;
2380                         if (tl->flags & SDTL_OVERLAP)
2381                                 sd->flags |= SD_OVERLAP;
2382                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2383                                 break;
2384                 }
2385         }
2386
2387         /* Build the groups for the domains */
2388         for_each_cpu(i, cpu_map) {
2389                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2390                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
2391                         if (sd->flags & SD_OVERLAP) {
2392                                 if (build_overlap_sched_groups(sd, i))
2393                                         goto error;
2394                         } else {
2395                                 if (build_sched_groups(sd, i))
2396                                         goto error;
2397                         }
2398                 }
2399         }
2400
2401         /*
2402          * Calculate an allowed NUMA imbalance such that LLCs do not get
2403          * imbalanced.
2404          */
2405         for_each_cpu(i, cpu_map) {
2406                 unsigned int imb = 0;
2407                 unsigned int imb_span = 1;
2408
2409                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2410                         struct sched_domain *child = sd->child;
2411
2412                         if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
2413                             (child->flags & SD_SHARE_PKG_RESOURCES)) {
2414                                 struct sched_domain __rcu *top_p;
2415                                 unsigned int nr_llcs;
2416
2417                                 /*
2418                                  * For a single LLC per node, allow an
2419                                  * imbalance up to 12.5% of the node. This is
2420                                  * arbitrary cutoff based two factors -- SMT and
2421                                  * memory channels. For SMT-2, the intent is to
2422                                  * avoid premature sharing of HT resources but
2423                                  * SMT-4 or SMT-8 *may* benefit from a different
2424                                  * cutoff. For memory channels, this is a very
2425                                  * rough estimate of how many channels may be
2426                                  * active and is based on recent CPUs with
2427                                  * many cores.
2428                                  *
2429                                  * For multiple LLCs, allow an imbalance
2430                                  * until multiple tasks would share an LLC
2431                                  * on one node while LLCs on another node
2432                                  * remain idle. This assumes that there are
2433                                  * enough logical CPUs per LLC to avoid SMT
2434                                  * factors and that there is a correlation
2435                                  * between LLCs and memory channels.
2436                                  */
2437                                 nr_llcs = sd->span_weight / child->span_weight;
2438                                 if (nr_llcs == 1)
2439                                         imb = sd->span_weight >> 3;
2440                                 else
2441                                         imb = nr_llcs;
2442                                 imb = max(1U, imb);
2443                                 sd->imb_numa_nr = imb;
2444
2445                                 /* Set span based on the first NUMA domain. */
2446                                 top_p = sd->parent;
2447                                 while (top_p && !(top_p->flags & SD_NUMA)) {
2448                                         top_p = top_p->parent;
2449                                 }
2450                                 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2451                         } else {
2452                                 int factor = max(1U, (sd->span_weight / imb_span));
2453
2454                                 sd->imb_numa_nr = imb * factor;
2455                         }
2456                 }
2457         }
2458
2459         /* Calculate CPU capacity for physical packages and nodes */
2460         for (i = nr_cpumask_bits-1; i >= 0; i--) {
2461                 if (!cpumask_test_cpu(i, cpu_map))
2462                         continue;
2463
2464                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2465                         claim_allocations(i, sd);
2466                         init_sched_groups_capacity(i, sd);
2467                 }
2468         }
2469
2470         /* Attach the domains */
2471         rcu_read_lock();
2472         for_each_cpu(i, cpu_map) {
2473                 rq = cpu_rq(i);
2474                 sd = *per_cpu_ptr(d.sd, i);
2475
2476                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2477                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
2478                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
2479
2480                 cpu_attach_domain(sd, d.rd, i);
2481         }
2482         rcu_read_unlock();
2483
2484         if (has_asym)
2485                 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2486
2487         if (rq && sched_debug_verbose) {
2488                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2489                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
2490         }
2491
2492         ret = 0;
2493 error:
2494         __free_domain_allocs(&d, alloc_state, cpu_map);
2495
2496         return ret;
2497 }
2498
2499 /* Current sched domains: */
2500 static cpumask_var_t                    *doms_cur;
2501
2502 /* Number of sched domains in 'doms_cur': */
2503 static int                              ndoms_cur;
2504
2505 /* Attributes of custom domains in 'doms_cur' */
2506 static struct sched_domain_attr         *dattr_cur;
2507
2508 /*
2509  * Special case: If a kmalloc() of a doms_cur partition (array of
2510  * cpumask) fails, then fallback to a single sched domain,
2511  * as determined by the single cpumask fallback_doms.
2512  */
2513 static cpumask_var_t                    fallback_doms;
2514
2515 /*
2516  * arch_update_cpu_topology lets virtualized architectures update the
2517  * CPU core maps. It is supposed to return 1 if the topology changed
2518  * or 0 if it stayed the same.
2519  */
2520 int __weak arch_update_cpu_topology(void)
2521 {
2522         return 0;
2523 }
2524
2525 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2526 {
2527         int i;
2528         cpumask_var_t *doms;
2529
2530         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2531         if (!doms)
2532                 return NULL;
2533         for (i = 0; i < ndoms; i++) {
2534                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2535                         free_sched_domains(doms, i);
2536                         return NULL;
2537                 }
2538         }
2539         return doms;
2540 }
2541
2542 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2543 {
2544         unsigned int i;
2545         for (i = 0; i < ndoms; i++)
2546                 free_cpumask_var(doms[i]);
2547         kfree(doms);
2548 }
2549
2550 /*
2551  * Set up scheduler domains and groups.  For now this just excludes isolated
2552  * CPUs, but could be used to exclude other special cases in the future.
2553  */
2554 int __init sched_init_domains(const struct cpumask *cpu_map)
2555 {
2556         int err;
2557
2558         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2559         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2560         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2561
2562         arch_update_cpu_topology();
2563         asym_cpu_capacity_scan();
2564         ndoms_cur = 1;
2565         doms_cur = alloc_sched_domains(ndoms_cur);
2566         if (!doms_cur)
2567                 doms_cur = &fallback_doms;
2568         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2569         err = build_sched_domains(doms_cur[0], NULL);
2570
2571         return err;
2572 }
2573
2574 /*
2575  * Detach sched domains from a group of CPUs specified in cpu_map
2576  * These CPUs will now be attached to the NULL domain
2577  */
2578 static void detach_destroy_domains(const struct cpumask *cpu_map)
2579 {
2580         unsigned int cpu = cpumask_any(cpu_map);
2581         int i;
2582
2583         if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2584                 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2585
2586         rcu_read_lock();
2587         for_each_cpu(i, cpu_map)
2588                 cpu_attach_domain(NULL, &def_root_domain, i);
2589         rcu_read_unlock();
2590 }
2591
2592 /* handle null as "default" */
2593 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2594                         struct sched_domain_attr *new, int idx_new)
2595 {
2596         struct sched_domain_attr tmp;
2597
2598         /* Fast path: */
2599         if (!new && !cur)
2600                 return 1;
2601
2602         tmp = SD_ATTR_INIT;
2603
2604         return !memcmp(cur ? (cur + idx_cur) : &tmp,
2605                         new ? (new + idx_new) : &tmp,
2606                         sizeof(struct sched_domain_attr));
2607 }
2608
2609 /*
2610  * Partition sched domains as specified by the 'ndoms_new'
2611  * cpumasks in the array doms_new[] of cpumasks. This compares
2612  * doms_new[] to the current sched domain partitioning, doms_cur[].
2613  * It destroys each deleted domain and builds each new domain.
2614  *
2615  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2616  * The masks don't intersect (don't overlap.) We should setup one
2617  * sched domain for each mask. CPUs not in any of the cpumasks will
2618  * not be load balanced. If the same cpumask appears both in the
2619  * current 'doms_cur' domains and in the new 'doms_new', we can leave
2620  * it as it is.
2621  *
2622  * The passed in 'doms_new' should be allocated using
2623  * alloc_sched_domains.  This routine takes ownership of it and will
2624  * free_sched_domains it when done with it. If the caller failed the
2625  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2626  * and partition_sched_domains() will fallback to the single partition
2627  * 'fallback_doms', it also forces the domains to be rebuilt.
2628  *
2629  * If doms_new == NULL it will be replaced with cpu_online_mask.
2630  * ndoms_new == 0 is a special case for destroying existing domains,
2631  * and it will not create the default domain.
2632  *
2633  * Call with hotplug lock and sched_domains_mutex held
2634  */
2635 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2636                                     struct sched_domain_attr *dattr_new)
2637 {
2638         bool __maybe_unused has_eas = false;
2639         int i, j, n;
2640         int new_topology;
2641
2642         lockdep_assert_held(&sched_domains_mutex);
2643
2644         /* Let the architecture update CPU core mappings: */
2645         new_topology = arch_update_cpu_topology();
2646         /* Trigger rebuilding CPU capacity asymmetry data */
2647         if (new_topology)
2648                 asym_cpu_capacity_scan();
2649
2650         if (!doms_new) {
2651                 WARN_ON_ONCE(dattr_new);
2652                 n = 0;
2653                 doms_new = alloc_sched_domains(1);
2654                 if (doms_new) {
2655                         n = 1;
2656                         cpumask_and(doms_new[0], cpu_active_mask,
2657                                     housekeeping_cpumask(HK_TYPE_DOMAIN));
2658                 }
2659         } else {
2660                 n = ndoms_new;
2661         }
2662
2663         /* Destroy deleted domains: */
2664         for (i = 0; i < ndoms_cur; i++) {
2665                 for (j = 0; j < n && !new_topology; j++) {
2666                         if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2667                             dattrs_equal(dattr_cur, i, dattr_new, j)) {
2668                                 struct root_domain *rd;
2669
2670                                 /*
2671                                  * This domain won't be destroyed and as such
2672                                  * its dl_bw->total_bw needs to be cleared.  It
2673                                  * will be recomputed in function
2674                                  * update_tasks_root_domain().
2675                                  */
2676                                 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2677                                 dl_clear_root_domain(rd);
2678                                 goto match1;
2679                         }
2680                 }
2681                 /* No match - a current sched domain not in new doms_new[] */
2682                 detach_destroy_domains(doms_cur[i]);
2683 match1:
2684                 ;
2685         }
2686
2687         n = ndoms_cur;
2688         if (!doms_new) {
2689                 n = 0;
2690                 doms_new = &fallback_doms;
2691                 cpumask_and(doms_new[0], cpu_active_mask,
2692                             housekeeping_cpumask(HK_TYPE_DOMAIN));
2693         }
2694
2695         /* Build new domains: */
2696         for (i = 0; i < ndoms_new; i++) {
2697                 for (j = 0; j < n && !new_topology; j++) {
2698                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2699                             dattrs_equal(dattr_new, i, dattr_cur, j))
2700                                 goto match2;
2701                 }
2702                 /* No match - add a new doms_new */
2703                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2704 match2:
2705                 ;
2706         }
2707
2708 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2709         /* Build perf. domains: */
2710         for (i = 0; i < ndoms_new; i++) {
2711                 for (j = 0; j < n && !sched_energy_update; j++) {
2712                         if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2713                             cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2714                                 has_eas = true;
2715                                 goto match3;
2716                         }
2717                 }
2718                 /* No match - add perf. domains for a new rd */
2719                 has_eas |= build_perf_domains(doms_new[i]);
2720 match3:
2721                 ;
2722         }
2723         sched_energy_set(has_eas);
2724 #endif
2725
2726         /* Remember the new sched domains: */
2727         if (doms_cur != &fallback_doms)
2728                 free_sched_domains(doms_cur, ndoms_cur);
2729
2730         kfree(dattr_cur);
2731         doms_cur = doms_new;
2732         dattr_cur = dattr_new;
2733         ndoms_cur = ndoms_new;
2734
2735         update_sched_domain_debugfs();
2736 }
2737
2738 /*
2739  * Call with hotplug lock held
2740  */
2741 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2742                              struct sched_domain_attr *dattr_new)
2743 {
2744         mutex_lock(&sched_domains_mutex);
2745         partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2746         mutex_unlock(&sched_domains_mutex);
2747 }