kernel/sched/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Scheduler topology setup/handling methods
   4  */
   5 #include "sched.h"
   6
   7 DEFINE_MUTEX(sched_domains_mutex);
   8
   9 /* Protected by sched_domains_mutex: */
  10 static cpumask_var_t sched_domains_tmpmask;
  11 static cpumask_var_t sched_domains_tmpmask2;
  12
  13 #ifdef CONFIG_SCHED_DEBUG
  14
  15 static int __init sched_debug_setup(char *str)
  16 {
  17         sched_debug_enabled = true;
  18
  19         return 0;
  20 }
  21 early_param("sched_debug", sched_debug_setup);
  22
  23 static inline bool sched_debug(void)
  24 {
  25         return sched_debug_enabled;
  26 }
  27
  28 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
  29                                   struct cpumask *groupmask)
  30 {
  31         struct sched_group *group = sd->groups;
  32
  33         cpumask_clear(groupmask);
  34
  35         printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
  36
  37         if (!(sd->flags & SD_LOAD_BALANCE)) {
  38                 printk("does not load-balance\n");
  39                 if (sd->parent)
  40                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
  41                 return -1;
  42         }
  43
  44         printk(KERN_CONT "span=%*pbl level=%s\n",
  45                cpumask_pr_args(sched_domain_span(sd)), sd->name);
  46
  47         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
  48                 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
  49         }
  50         if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
  51                 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
  52         }
  53
  54         printk(KERN_DEBUG "%*s groups:", level + 1, "");
  55         do {
  56                 if (!group) {
  57                         printk("\n");
  58                         printk(KERN_ERR "ERROR: group is NULL\n");
  59                         break;
  60                 }
  61
  62                 if (!cpumask_weight(sched_group_span(group))) {
  63                         printk(KERN_CONT "\n");
  64                         printk(KERN_ERR "ERROR: empty group\n");
  65                         break;
  66                 }
  67
  68                 if (!(sd->flags & SD_OVERLAP) &&
  69                     cpumask_intersects(groupmask, sched_group_span(group))) {
  70                         printk(KERN_CONT "\n");
  71                         printk(KERN_ERR "ERROR: repeated CPUs\n");
  72                         break;
  73                 }
  74
  75                 cpumask_or(groupmask, groupmask, sched_group_span(group));
  76
  77                 printk(KERN_CONT " %d:{ span=%*pbl",
  78                                 group->sgc->id,
  79                                 cpumask_pr_args(sched_group_span(group)));
  80
  81                 if ((sd->flags & SD_OVERLAP) &&
  82                     !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
  83                         printk(KERN_CONT " mask=%*pbl",
  84                                 cpumask_pr_args(group_balance_mask(group)));
  85                 }
  86
  87                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
  88                         printk(KERN_CONT " cap=%lu", group->sgc->capacity);
  89
  90                 if (group == sd->groups && sd->child &&
  91                     !cpumask_equal(sched_domain_span(sd->child),
  92                                    sched_group_span(group))) {
  93                         printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
  94                 }
  95
  96                 printk(KERN_CONT " }");
  97
  98                 group = group->next;
  99
 100                 if (group != sd->groups)
 101                         printk(KERN_CONT ",");
 102
 103         } while (group != sd->groups);
 104         printk(KERN_CONT "\n");
 105
 106         if (!cpumask_equal(sched_domain_span(sd), groupmask))
 107                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 108
 109         if (sd->parent &&
 110             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
 111                 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
 112         return 0;
 113 }
 114
 115 static void sched_domain_debug(struct sched_domain *sd, int cpu)
 116 {
 117         int level = 0;
 118
 119         if (!sched_debug_enabled)
 120                 return;
 121
 122         if (!sd) {
 123                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
 124                 return;
 125         }
 126
 127         printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
 128
 129         for (;;) {
 130                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
 131                         break;
 132                 level++;
 133                 sd = sd->parent;
 134                 if (!sd)
 135                         break;
 136         }
 137 }
 138 #else /* !CONFIG_SCHED_DEBUG */
 139
 140 # define sched_debug_enabled 0
 141 # define sched_domain_debug(sd, cpu) do { } while (0)
 142 static inline bool sched_debug(void)
 143 {
 144         return false;
 145 }
 146 #endif /* CONFIG_SCHED_DEBUG */
 147
 148 static int sd_degenerate(struct sched_domain *sd)
 149 {
 150         if (cpumask_weight(sched_domain_span(sd)) == 1)
 151                 return 1;
 152
 153         /* Following flags need at least 2 groups */
 154         if (sd->flags & (SD_LOAD_BALANCE |
 155                          SD_BALANCE_NEWIDLE |
 156                          SD_BALANCE_FORK |
 157                          SD_BALANCE_EXEC |
 158                          SD_SHARE_CPUCAPACITY |
 159                          SD_ASYM_CPUCAPACITY |
 160                          SD_SHARE_PKG_RESOURCES |
 161                          SD_SHARE_POWERDOMAIN)) {
 162                 if (sd->groups != sd->groups->next)
 163                         return 0;
 164         }
 165
 166         /* Following flags don't use groups */
 167         if (sd->flags & (SD_WAKE_AFFINE))
 168                 return 0;
 169
 170         return 1;
 171 }
 172
 173 static int
 174 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
 175 {
 176         unsigned long cflags = sd->flags, pflags = parent->flags;
 177
 178         if (sd_degenerate(parent))
 179                 return 1;
 180
 181         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
 182                 return 0;
 183
 184         /* Flags needing groups don't count if only 1 group in parent */
 185         if (parent->groups == parent->groups->next) {
 186                 pflags &= ~(SD_LOAD_BALANCE |
 187                                 SD_BALANCE_NEWIDLE |
 188                                 SD_BALANCE_FORK |
 189                                 SD_BALANCE_EXEC |
 190                                 SD_ASYM_CPUCAPACITY |
 191                                 SD_SHARE_CPUCAPACITY |
 192                                 SD_SHARE_PKG_RESOURCES |
 193                                 SD_PREFER_SIBLING |
 194                                 SD_SHARE_POWERDOMAIN);
 195                 if (nr_node_ids == 1)
 196                         pflags &= ~SD_SERIALIZE;
 197         }
 198         if (~cflags & pflags)
 199                 return 0;
 200
 201         return 1;
 202 }
 203
 204 static void free_rootdomain(struct rcu_head *rcu)
 205 {
 206         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
 207
 208         cpupri_cleanup(&rd->cpupri);
 209         cpudl_cleanup(&rd->cpudl);
 210         free_cpumask_var(rd->dlo_mask);
 211         free_cpumask_var(rd->rto_mask);
 212         free_cpumask_var(rd->online);
 213         free_cpumask_var(rd->span);
 214         kfree(rd);
 215 }
 216
 217 void rq_attach_root(struct rq *rq, struct root_domain *rd)
 218 {
 219         struct root_domain *old_rd = NULL;
 220         unsigned long flags;
 221
 222         raw_spin_lock_irqsave(&rq->lock, flags);
 223
 224         if (rq->rd) {
 225                 old_rd = rq->rd;
 226
 227                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
 228                         set_rq_offline(rq);
 229
 230                 cpumask_clear_cpu(rq->cpu, old_rd->span);
 231
 232                 /*
 233                  * If we dont want to free the old_rd yet then
 234                  * set old_rd to NULL to skip the freeing later
 235                  * in this function:
 236                  */
 237                 if (!atomic_dec_and_test(&old_rd->refcount))
 238                         old_rd = NULL;
 239         }
 240
 241         atomic_inc(&rd->refcount);
 242         rq->rd = rd;
 243
 244         cpumask_set_cpu(rq->cpu, rd->span);
 245         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
 246                 set_rq_online(rq);
 247
 248         raw_spin_unlock_irqrestore(&rq->lock, flags);
 249
 250         if (old_rd)
 251                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
 252 }
 253
 254 void sched_get_rd(struct root_domain *rd)
 255 {
 256         atomic_inc(&rd->refcount);
 257 }
 258
 259 void sched_put_rd(struct root_domain *rd)
 260 {
 261         if (!atomic_dec_and_test(&rd->refcount))
 262                 return;
 263
 264         call_rcu_sched(&rd->rcu, free_rootdomain);
 265 }
 266
 267 static int init_rootdomain(struct root_domain *rd)
 268 {
 269         if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
 270                 goto out;
 271         if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
 272                 goto free_span;
 273         if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
 274                 goto free_online;
 275         if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
 276                 goto free_dlo_mask;
 277
 278 #ifdef HAVE_RT_PUSH_IPI
 279         rd->rto_cpu = -1;
 280         raw_spin_lock_init(&rd->rto_lock);
 281         init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
 282 #endif
 283
 284         init_dl_bw(&rd->dl_bw);
 285         if (cpudl_init(&rd->cpudl) != 0)
 286                 goto free_rto_mask;
 287
 288         if (cpupri_init(&rd->cpupri) != 0)
 289                 goto free_cpudl;
 290         return 0;
 291
 292 free_cpudl:
 293         cpudl_cleanup(&rd->cpudl);
 294 free_rto_mask:
 295         free_cpumask_var(rd->rto_mask);
 296 free_dlo_mask:
 297         free_cpumask_var(rd->dlo_mask);
 298 free_online:
 299         free_cpumask_var(rd->online);
 300 free_span:
 301         free_cpumask_var(rd->span);
 302 out:
 303         return -ENOMEM;
 304 }
 305
 306 /*
 307  * By default the system creates a single root-domain with all CPUs as
 308  * members (mimicking the global state we have today).
 309  */
 310 struct root_domain def_root_domain;
 311
 312 void init_defrootdomain(void)
 313 {
 314         init_rootdomain(&def_root_domain);
 315
 316         atomic_set(&def_root_domain.refcount, 1);
 317 }
 318
 319 static struct root_domain *alloc_rootdomain(void)
 320 {
 321         struct root_domain *rd;
 322
 323         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 324         if (!rd)
 325                 return NULL;
 326
 327         if (init_rootdomain(rd) != 0) {
 328                 kfree(rd);
 329                 return NULL;
 330         }
 331
 332         return rd;
 333 }
 334
 335 static void free_sched_groups(struct sched_group *sg, int free_sgc)
 336 {
 337         struct sched_group *tmp, *first;
 338
 339         if (!sg)
 340                 return;
 341
 342         first = sg;
 343         do {
 344                 tmp = sg->next;
 345
 346                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
 347                         kfree(sg->sgc);
 348
 349                 if (atomic_dec_and_test(&sg->ref))
 350                         kfree(sg);
 351                 sg = tmp;
 352         } while (sg != first);
 353 }
 354
 355 static void destroy_sched_domain(struct sched_domain *sd)
 356 {
 357         /*
 358          * A normal sched domain may have multiple group references, an
 359          * overlapping domain, having private groups, only one.  Iterate,
 360          * dropping group/capacity references, freeing where none remain.
 361          */
 362         free_sched_groups(sd->groups, 1);
 363
 364         if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
 365                 kfree(sd->shared);
 366         kfree(sd);
 367 }
 368
 369 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
 370 {
 371         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
 372
 373         while (sd) {
 374                 struct sched_domain *parent = sd->parent;
 375                 destroy_sched_domain(sd);
 376                 sd = parent;
 377         }
 378 }
 379
 380 static void destroy_sched_domains(struct sched_domain *sd)
 381 {
 382         if (sd)
 383                 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
 384 }
 385
 386 /*
 387  * Keep a special pointer to the highest sched_domain that has
 388  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
 389  * allows us to avoid some pointer chasing select_idle_sibling().
 390  *
 391  * Also keep a unique ID per domain (we use the first CPU number in
 392  * the cpumask of the domain), this allows us to quickly tell if
 393  * two CPUs are in the same cache domain, see cpus_share_cache().
 394  */
 395 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
 396 DEFINE_PER_CPU(int, sd_llc_size);
 397 DEFINE_PER_CPU(int, sd_llc_id);
 398 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
 399 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
 400 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
 401 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
 402
 403 static void update_top_cache_domain(int cpu)
 404 {
 405         struct sched_domain_shared *sds = NULL;
 406         struct sched_domain *sd;
 407         int id = cpu;
 408         int size = 1;
 409
 410         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
 411         if (sd) {
 412                 id = cpumask_first(sched_domain_span(sd));
 413                 size = cpumask_weight(sched_domain_span(sd));
 414                 sds = sd->shared;
 415         }
 416
 417         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 418         per_cpu(sd_llc_size, cpu) = size;
 419         per_cpu(sd_llc_id, cpu) = id;
 420         rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 421
 422         sd = lowest_flag_domain(cpu, SD_NUMA);
 423         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 424
 425         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
 426         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
 427 }
 428
 429 /*
 430  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
 431  * hold the hotplug lock.
 432  */
 433 static void
 434 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 435 {
 436         struct rq *rq = cpu_rq(cpu);
 437         struct sched_domain *tmp;
 438
 439         /* Remove the sched domains which do not contribute to scheduling. */
 440         for (tmp = sd; tmp; ) {
 441                 struct sched_domain *parent = tmp->parent;
 442                 if (!parent)
 443                         break;
 444
 445                 if (sd_parent_degenerate(tmp, parent)) {
 446                         tmp->parent = parent->parent;
 447                         if (parent->parent)
 448                                 parent->parent->child = tmp;
 449                         /*
 450                          * Transfer SD_PREFER_SIBLING down in case of a
 451                          * degenerate parent; the spans match for this
 452                          * so the property transfers.
 453                          */
 454                         if (parent->flags & SD_PREFER_SIBLING)
 455                                 tmp->flags |= SD_PREFER_SIBLING;
 456                         destroy_sched_domain(parent);
 457                 } else
 458                         tmp = tmp->parent;
 459         }
 460
 461         if (sd && sd_degenerate(sd)) {
 462                 tmp = sd;
 463                 sd = sd->parent;
 464                 destroy_sched_domain(tmp);
 465                 if (sd)
 466                         sd->child = NULL;
 467         }
 468
 469         sched_domain_debug(sd, cpu);
 470
 471         rq_attach_root(rq, rd);
 472         tmp = rq->sd;
 473         rcu_assign_pointer(rq->sd, sd);
 474         dirty_sched_domain_sysctl(cpu);
 475         destroy_sched_domains(tmp);
 476
 477         update_top_cache_domain(cpu);
 478 }
 479
 480 struct s_data {
 481         struct sched_domain ** __percpu sd;
 482         struct root_domain      *rd;
 483 };
 484
 485 enum s_alloc {
 486         sa_rootdomain,
 487         sa_sd,
 488         sa_sd_storage,
 489         sa_none,
 490 };
 491
 492 /*
 493  * Return the canonical balance CPU for this group, this is the first CPU
 494  * of this group that's also in the balance mask.
 495  *
 496  * The balance mask are all those CPUs that could actually end up at this
 497  * group. See build_balance_mask().
 498  *
 499  * Also see should_we_balance().
 500  */
 501 int group_balance_cpu(struct sched_group *sg)
 502 {
 503         return cpumask_first(group_balance_mask(sg));
 504 }
 505
 506
 507 /*
 508  * NUMA topology (first read the regular topology blurb below)
 509  *
 510  * Given a node-distance table, for example:
 511  *
 512  *   node   0   1   2   3
 513  *     0:  10  20  30  20
 514  *     1:  20  10  20  30
 515  *     2:  30  20  10  20
 516  *     3:  20  30  20  10
 517  *
 518  * which represents a 4 node ring topology like:
 519  *
 520  *   0 ----- 1
 521  *   |       |
 522  *   |       |
 523  *   |       |
 524  *   3 ----- 2
 525  *
 526  * We want to construct domains and groups to represent this. The way we go
 527  * about doing this is to build the domains on 'hops'. For each NUMA level we
 528  * construct the mask of all nodes reachable in @level hops.
 529  *
 530  * For the above NUMA topology that gives 3 levels:
 531  *
 532  * NUMA-2       0-3             0-3             0-3             0-3
 533  *  groups:     {0-1,3},{1-3}   {0-2},{0,2-3}   {1-3},{0-1,3}   {0,2-3},{0-2}
 534  *
 535  * NUMA-1       0-1,3           0-2             1-3             0,2-3
 536  *  groups:     {0},{1},{3}     {0},{1},{2}     {1},{2},{3}     {0},{2},{3}
 537  *
 538  * NUMA-0       0               1               2               3
 539  *
 540  *
 541  * As can be seen; things don't nicely line up as with the regular topology.
 542  * When we iterate a domain in child domain chunks some nodes can be
 543  * represented multiple times -- hence the "overlap" naming for this part of
 544  * the topology.
 545  *
 546  * In order to minimize this overlap, we only build enough groups to cover the
 547  * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
 548  *
 549  * Because:
 550  *
 551  *  - the first group of each domain is its child domain; this
 552  *    gets us the first 0-1,3
 553  *  - the only uncovered node is 2, who's child domain is 1-3.
 554  *
 555  * However, because of the overlap, computing a unique CPU for each group is
 556  * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
 557  * groups include the CPUs of Node-0, while those CPUs would not in fact ever
 558  * end up at those groups (they would end up in group: 0-1,3).
 559  *
 560  * To correct this we have to introduce the group balance mask. This mask
 561  * will contain those CPUs in the group that can reach this group given the
 562  * (child) domain tree.
 563  *
 564  * With this we can once again compute balance_cpu and sched_group_capacity
 565  * relations.
 566  *
 567  * XXX include words on how balance_cpu is unique and therefore can be
 568  * used for sched_group_capacity links.
 569  *
 570  *
 571  * Another 'interesting' topology is:
 572  *
 573  *   node   0   1   2   3
 574  *     0:  10  20  20  30
 575  *     1:  20  10  20  20
 576  *     2:  20  20  10  20
 577  *     3:  30  20  20  10
 578  *
 579  * Which looks a little like:
 580  *
 581  *   0 ----- 1
 582  *   |     / |
 583  *   |   /   |
 584  *   | /     |
 585  *   2 ----- 3
 586  *
 587  * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
 588  * are not.
 589  *
 590  * This leads to a few particularly weird cases where the sched_domain's are
 591  * not of the same number for each CPU. Consider:
 592  *
 593  * NUMA-2       0-3                                             0-3
 594  *  groups:     {0-2},{1-3}                                     {1-3},{0-2}
 595  *
 596  * NUMA-1       0-2             0-3             0-3             1-3
 597  *
 598  * NUMA-0       0               1               2               3
 599  *
 600  */
 601
 602
 603 /*
 604  * Build the balance mask; it contains only those CPUs that can arrive at this
 605  * group and should be considered to continue balancing.
 606  *
 607  * We do this during the group creation pass, therefore the group information
 608  * isn't complete yet, however since each group represents a (child) domain we
 609  * can fully construct this using the sched_domain bits (which are already
 610  * complete).
 611  */
 612 static void
 613 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
 614 {
 615         const struct cpumask *sg_span = sched_group_span(sg);
 616         struct sd_data *sdd = sd->private;
 617         struct sched_domain *sibling;
 618         int i;
 619
 620         cpumask_clear(mask);
 621
 622         for_each_cpu(i, sg_span) {
 623                 sibling = *per_cpu_ptr(sdd->sd, i);
 624
 625                 /*
 626                  * Can happen in the asymmetric case, where these siblings are
 627                  * unused. The mask will not be empty because those CPUs that
 628                  * do have the top domain _should_ span the domain.
 629                  */
 630                 if (!sibling->child)
 631                         continue;
 632
 633                 /* If we would not end up here, we can't continue from here */
 634                 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
 635                         continue;
 636
 637                 cpumask_set_cpu(i, mask);
 638         }
 639
 640         /* We must not have empty masks here */
 641         WARN_ON_ONCE(cpumask_empty(mask));
 642 }
 643
 644 /*
 645  * XXX: This creates per-node group entries; since the load-balancer will
 646  * immediately access remote memory to construct this group's load-balance
 647  * statistics having the groups node local is of dubious benefit.
 648  */
 649 static struct sched_group *
 650 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
 651 {
 652         struct sched_group *sg;
 653         struct cpumask *sg_span;
 654
 655         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
 656                         GFP_KERNEL, cpu_to_node(cpu));
 657
 658         if (!sg)
 659                 return NULL;
 660
 661         sg_span = sched_group_span(sg);
 662         if (sd->child)
 663                 cpumask_copy(sg_span, sched_domain_span(sd->child));
 664         else
 665                 cpumask_copy(sg_span, sched_domain_span(sd));
 666
 667         atomic_inc(&sg->ref);
 668         return sg;
 669 }
 670
 671 static void init_overlap_sched_group(struct sched_domain *sd,
 672                                      struct sched_group *sg)
 673 {
 674         struct cpumask *mask = sched_domains_tmpmask2;
 675         struct sd_data *sdd = sd->private;
 676         struct cpumask *sg_span;
 677         int cpu;
 678
 679         build_balance_mask(sd, sg, mask);
 680         cpu = cpumask_first_and(sched_group_span(sg), mask);
 681
 682         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 683         if (atomic_inc_return(&sg->sgc->ref) == 1)
 684                 cpumask_copy(group_balance_mask(sg), mask);
 685         else
 686                 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
 687
 688         /*
 689          * Initialize sgc->capacity such that even if we mess up the
 690          * domains and no possible iteration will get us here, we won't
 691          * die on a /0 trap.
 692          */
 693         sg_span = sched_group_span(sg);
 694         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
 695         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 696         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 697 }
 698
 699 static int
 700 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
 701 {
 702         struct sched_group *first = NULL, *last = NULL, *sg;
 703         const struct cpumask *span = sched_domain_span(sd);
 704         struct cpumask *covered = sched_domains_tmpmask;
 705         struct sd_data *sdd = sd->private;
 706         struct sched_domain *sibling;
 707         int i;
 708
 709         cpumask_clear(covered);
 710
 711         for_each_cpu_wrap(i, span, cpu) {
 712                 struct cpumask *sg_span;
 713
 714                 if (cpumask_test_cpu(i, covered))
 715                         continue;
 716
 717                 sibling = *per_cpu_ptr(sdd->sd, i);
 718
 719                 /*
 720                  * Asymmetric node setups can result in situations where the
 721                  * domain tree is of unequal depth, make sure to skip domains
 722                  * that already cover the entire range.
 723                  *
 724                  * In that case build_sched_domains() will have terminated the
 725                  * iteration early and our sibling sd spans will be empty.
 726                  * Domains should always include the CPU they're built on, so
 727                  * check that.
 728                  */
 729                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
 730                         continue;
 731
 732                 sg = build_group_from_child_sched_domain(sibling, cpu);
 733                 if (!sg)
 734                         goto fail;
 735
 736                 sg_span = sched_group_span(sg);
 737                 cpumask_or(covered, covered, sg_span);
 738
 739                 init_overlap_sched_group(sd, sg);
 740
 741                 if (!first)
 742                         first = sg;
 743                 if (last)
 744                         last->next = sg;
 745                 last = sg;
 746                 last->next = first;
 747         }
 748         sd->groups = first;
 749
 750         return 0;
 751
 752 fail:
 753         free_sched_groups(first, 0);
 754
 755         return -ENOMEM;
 756 }
 757
 758
 759 /*
 760  * Package topology (also see the load-balance blurb in fair.c)
 761  *
 762  * The scheduler builds a tree structure to represent a number of important
 763  * topology features. By default (default_topology[]) these include:
 764  *
 765  *  - Simultaneous multithreading (SMT)
 766  *  - Multi-Core Cache (MC)
 767  *  - Package (DIE)
 768  *
 769  * Where the last one more or less denotes everything up to a NUMA node.
 770  *
 771  * The tree consists of 3 primary data structures:
 772  *
 773  *      sched_domain -> sched_group -> sched_group_capacity
 774  *          ^ ^             ^ ^
 775  *          `-'             `-'
 776  *
 777  * The sched_domains are per-CPU and have a two way link (parent & child) and
 778  * denote the ever growing mask of CPUs belonging to that level of topology.
 779  *
 780  * Each sched_domain has a circular (double) linked list of sched_group's, each
 781  * denoting the domains of the level below (or individual CPUs in case of the
 782  * first domain level). The sched_group linked by a sched_domain includes the
 783  * CPU of that sched_domain [*].
 784  *
 785  * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
 786  *
 787  * CPU   0   1   2   3   4   5   6   7
 788  *
 789  * DIE  [                             ]
 790  * MC   [             ] [             ]
 791  * SMT  [     ] [     ] [     ] [     ]
 792  *
 793  *  - or -
 794  *
 795  * DIE  0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
 796  * MC   0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
 797  * SMT  0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
 798  *
 799  * CPU   0   1   2   3   4   5   6   7
 800  *
 801  * One way to think about it is: sched_domain moves you up and down among these
 802  * topology levels, while sched_group moves you sideways through it, at child
 803  * domain granularity.
 804  *
 805  * sched_group_capacity ensures each unique sched_group has shared storage.
 806  *
 807  * There are two related construction problems, both require a CPU that
 808  * uniquely identify each group (for a given domain):
 809  *
 810  *  - The first is the balance_cpu (see should_we_balance() and the
 811  *    load-balance blub in fair.c); for each group we only want 1 CPU to
 812  *    continue balancing at a higher domain.
 813  *
 814  *  - The second is the sched_group_capacity; we want all identical groups
 815  *    to share a single sched_group_capacity.
 816  *
 817  * Since these topologies are exclusive by construction. That is, its
 818  * impossible for an SMT thread to belong to multiple cores, and cores to
 819  * be part of multiple caches. There is a very clear and unique location
 820  * for each CPU in the hierarchy.
 821  *
 822  * Therefore computing a unique CPU for each group is trivial (the iteration
 823  * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
 824  * group), we can simply pick the first CPU in each group.
 825  *
 826  *
 827  * [*] in other words, the first group of each domain is its child domain.
 828  */
 829
 830 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
 831 {
 832         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
 833         struct sched_domain *child = sd->child;
 834         struct sched_group *sg;
 835
 836         if (child)
 837                 cpu = cpumask_first(sched_domain_span(child));
 838
 839         sg = *per_cpu_ptr(sdd->sg, cpu);
 840         sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
 841
 842         /* For claim_allocations: */
 843         atomic_inc(&sg->ref);
 844         atomic_inc(&sg->sgc->ref);
 845
 846         if (child) {
 847                 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
 848                 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
 849         } else {
 850                 cpumask_set_cpu(cpu, sched_group_span(sg));
 851                 cpumask_set_cpu(cpu, group_balance_mask(sg));
 852         }
 853
 854         sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
 855         sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
 856         sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
 857
 858         return sg;
 859 }
 860
 861 /*
 862  * build_sched_groups will build a circular linked list of the groups
 863  * covered by the given span, and will set each group's ->cpumask correctly,
 864  * and ->cpu_capacity to 0.
 865  *
 866  * Assumes the sched_domain tree is fully constructed
 867  */
 868 static int
 869 build_sched_groups(struct sched_domain *sd, int cpu)
 870 {
 871         struct sched_group *first = NULL, *last = NULL;
 872         struct sd_data *sdd = sd->private;
 873         const struct cpumask *span = sched_domain_span(sd);
 874         struct cpumask *covered;
 875         int i;
 876
 877         lockdep_assert_held(&sched_domains_mutex);
 878         covered = sched_domains_tmpmask;
 879
 880         cpumask_clear(covered);
 881
 882         for_each_cpu_wrap(i, span, cpu) {
 883                 struct sched_group *sg;
 884
 885                 if (cpumask_test_cpu(i, covered))
 886                         continue;
 887
 888                 sg = get_group(i, sdd);
 889
 890                 cpumask_or(covered, covered, sched_group_span(sg));
 891
 892                 if (!first)
 893                         first = sg;
 894                 if (last)
 895                         last->next = sg;
 896                 last = sg;
 897         }
 898         last->next = first;
 899         sd->groups = first;
 900
 901         return 0;
 902 }
 903
 904 /*
 905  * Initialize sched groups cpu_capacity.
 906  *
 907  * cpu_capacity indicates the capacity of sched group, which is used while
 908  * distributing the load between different sched groups in a sched domain.
 909  * Typically cpu_capacity for all the groups in a sched domain will be same
 910  * unless there are asymmetries in the topology. If there are asymmetries,
 911  * group having more cpu_capacity will pickup more load compared to the
 912  * group having less cpu_capacity.
 913  */
 914 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
 915 {
 916         struct sched_group *sg = sd->groups;
 917
 918         WARN_ON(!sg);
 919
 920         do {
 921                 int cpu, max_cpu = -1;
 922
 923                 sg->group_weight = cpumask_weight(sched_group_span(sg));
 924
 925                 if (!(sd->flags & SD_ASYM_PACKING))
 926                         goto next;
 927
 928                 for_each_cpu(cpu, sched_group_span(sg)) {
 929                         if (max_cpu < 0)
 930                                 max_cpu = cpu;
 931                         else if (sched_asym_prefer(cpu, max_cpu))
 932                                 max_cpu = cpu;
 933                 }
 934                 sg->asym_prefer_cpu = max_cpu;
 935
 936 next:
 937                 sg = sg->next;
 938         } while (sg != sd->groups);
 939
 940         if (cpu != group_balance_cpu(sg))
 941                 return;
 942
 943         update_group_capacity(sd, cpu);
 944 }
 945
 946 /*
 947  * Initializers for schedule domains
 948  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
 949  */
 950
 951 static int default_relax_domain_level = -1;
 952 int sched_domain_level_max;
 953
 954 static int __init setup_relax_domain_level(char *str)
 955 {
 956         if (kstrtoint(str, 0, &default_relax_domain_level))
 957                 pr_warn("Unable to set relax_domain_level\n");
 958
 959         return 1;
 960 }
 961 __setup("relax_domain_level=", setup_relax_domain_level);
 962
 963 static void set_domain_attribute(struct sched_domain *sd,
 964                                  struct sched_domain_attr *attr)
 965 {
 966         int request;
 967
 968         if (!attr || attr->relax_domain_level < 0) {
 969                 if (default_relax_domain_level < 0)
 970                         return;
 971                 else
 972                         request = default_relax_domain_level;
 973         } else
 974                 request = attr->relax_domain_level;
 975         if (request < sd->level) {
 976                 /* Turn off idle balance on this domain: */
 977                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
 978         } else {
 979                 /* Turn on idle balance on this domain: */
 980                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
 981         }
 982 }
 983
 984 static void __sdt_free(const struct cpumask *cpu_map);
 985 static int __sdt_alloc(const struct cpumask *cpu_map);
 986
 987 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
 988                                  const struct cpumask *cpu_map)
 989 {
 990         switch (what) {
 991         case sa_rootdomain:
 992                 if (!atomic_read(&d->rd->refcount))
 993                         free_rootdomain(&d->rd->rcu);
 994                 /* Fall through */
 995         case sa_sd:
 996                 free_percpu(d->sd);
 997                 /* Fall through */
 998         case sa_sd_storage:
 999                 __sdt_free(cpu_map);
1000                 /* Fall through */
1001         case sa_none:
1002                 break;
1003         }
1004 }
1005
1006 static enum s_alloc
1007 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1008 {
1009         memset(d, 0, sizeof(*d));
1010
1011         if (__sdt_alloc(cpu_map))
1012                 return sa_sd_storage;
1013         d->sd = alloc_percpu(struct sched_domain *);
1014         if (!d->sd)
1015                 return sa_sd_storage;
1016         d->rd = alloc_rootdomain();
1017         if (!d->rd)
1018                 return sa_sd;
1019
1020         return sa_rootdomain;
1021 }
1022
1023 /*
1024  * NULL the sd_data elements we've used to build the sched_domain and
1025  * sched_group structure so that the subsequent __free_domain_allocs()
1026  * will not free the data we're using.
1027  */
1028 static void claim_allocations(int cpu, struct sched_domain *sd)
1029 {
1030         struct sd_data *sdd = sd->private;
1031
1032         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1033         *per_cpu_ptr(sdd->sd, cpu) = NULL;
1034
1035         if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1036                 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1037
1038         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1039                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1040
1041         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1042                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1043 }
1044
1045 #ifdef CONFIG_NUMA
1046 enum numa_topology_type sched_numa_topology_type;
1047
1048 static int                      sched_domains_numa_levels;
1049 static int                      sched_domains_curr_level;
1050
1051 int                             sched_max_numa_distance;
1052 static int                      *sched_domains_numa_distance;
1053 static struct cpumask           ***sched_domains_numa_masks;
1054 #endif
1055
1056 /*
1057  * SD_flags allowed in topology descriptions.
1058  *
1059  * These flags are purely descriptive of the topology and do not prescribe
1060  * behaviour. Behaviour is artificial and mapped in the below sd_init()
1061  * function:
1062  *
1063  *   SD_SHARE_CPUCAPACITY   - describes SMT topologies
1064  *   SD_SHARE_PKG_RESOURCES - describes shared caches
1065  *   SD_NUMA                - describes NUMA topologies
1066  *   SD_SHARE_POWERDOMAIN   - describes shared power domain
1067  *
1068  * Odd one out, which beside describing the topology has a quirk also
1069  * prescribes the desired behaviour that goes along with it:
1070  *
1071  *   SD_ASYM_PACKING        - describes SMT quirks
1072  */
1073 #define TOPOLOGY_SD_FLAGS               \
1074         (SD_SHARE_CPUCAPACITY   |       \
1075          SD_SHARE_PKG_RESOURCES |       \
1076          SD_NUMA                |       \
1077          SD_ASYM_PACKING        |       \
1078          SD_SHARE_POWERDOMAIN)
1079
1080 static struct sched_domain *
1081 sd_init(struct sched_domain_topology_level *tl,
1082         const struct cpumask *cpu_map,
1083         struct sched_domain *child, int dflags, int cpu)
1084 {
1085         struct sd_data *sdd = &tl->data;
1086         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1087         int sd_id, sd_weight, sd_flags = 0;
1088
1089 #ifdef CONFIG_NUMA
1090         /*
1091          * Ugly hack to pass state to sd_numa_mask()...
1092          */
1093         sched_domains_curr_level = tl->numa_level;
1094 #endif
1095
1096         sd_weight = cpumask_weight(tl->mask(cpu));
1097
1098         if (tl->sd_flags)
1099                 sd_flags = (*tl->sd_flags)();
1100         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1101                         "wrong sd_flags in topology description\n"))
1102                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1103
1104         /* Apply detected topology flags */
1105         sd_flags |= dflags;
1106
1107         *sd = (struct sched_domain){
1108                 .min_interval           = sd_weight,
1109                 .max_interval           = 2*sd_weight,
1110                 .busy_factor            = 32,
1111                 .imbalance_pct          = 125,
1112
1113                 .cache_nice_tries       = 0,
1114                 .busy_idx               = 0,
1115                 .idle_idx               = 0,
1116                 .newidle_idx            = 0,
1117                 .wake_idx               = 0,
1118                 .forkexec_idx           = 0,
1119
1120                 .flags                  = 1*SD_LOAD_BALANCE
1121                                         | 1*SD_BALANCE_NEWIDLE
1122                                         | 1*SD_BALANCE_EXEC
1123                                         | 1*SD_BALANCE_FORK
1124                                         | 0*SD_BALANCE_WAKE
1125                                         | 1*SD_WAKE_AFFINE
1126                                         | 0*SD_SHARE_CPUCAPACITY
1127                                         | 0*SD_SHARE_PKG_RESOURCES
1128                                         | 0*SD_SERIALIZE
1129                                         | 1*SD_PREFER_SIBLING
1130                                         | 0*SD_NUMA
1131                                         | sd_flags
1132                                         ,
1133
1134                 .last_balance           = jiffies,
1135                 .balance_interval       = sd_weight,
1136                 .max_newidle_lb_cost    = 0,
1137                 .next_decay_max_lb_cost = jiffies,
1138                 .child                  = child,
1139 #ifdef CONFIG_SCHED_DEBUG
1140                 .name                   = tl->name,
1141 #endif
1142         };
1143
1144         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1145         sd_id = cpumask_first(sched_domain_span(sd));
1146
1147         /*
1148          * Convert topological properties into behaviour.
1149          */
1150
1151         if (sd->flags & SD_ASYM_CPUCAPACITY) {
1152                 struct sched_domain *t = sd;
1153
1154                 /*
1155                  * Don't attempt to spread across CPUs of different capacities.
1156                  */
1157                 if (sd->child)
1158                         sd->child->flags &= ~SD_PREFER_SIBLING;
1159
1160                 for_each_lower_domain(t)
1161                         t->flags |= SD_BALANCE_WAKE;
1162         }
1163
1164         if (sd->flags & SD_SHARE_CPUCAPACITY) {
1165                 sd->imbalance_pct = 110;
1166
1167         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1168                 sd->imbalance_pct = 117;
1169                 sd->cache_nice_tries = 1;
1170                 sd->busy_idx = 2;
1171
1172 #ifdef CONFIG_NUMA
1173         } else if (sd->flags & SD_NUMA) {
1174                 sd->cache_nice_tries = 2;
1175                 sd->busy_idx = 3;
1176                 sd->idle_idx = 2;
1177
1178                 sd->flags &= ~SD_PREFER_SIBLING;
1179                 sd->flags |= SD_SERIALIZE;
1180                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1181                         sd->flags &= ~(SD_BALANCE_EXEC |
1182                                        SD_BALANCE_FORK |
1183                                        SD_WAKE_AFFINE);
1184                 }
1185
1186 #endif
1187         } else {
1188                 sd->cache_nice_tries = 1;
1189                 sd->busy_idx = 2;
1190                 sd->idle_idx = 1;
1191         }
1192
1193         /*
1194          * For all levels sharing cache; connect a sched_domain_shared
1195          * instance.
1196          */
1197         if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1198                 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1199                 atomic_inc(&sd->shared->ref);
1200                 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1201         }
1202
1203         sd->private = sdd;
1204
1205         return sd;
1206 }
1207
1208 /*
1209  * Topology list, bottom-up.
1210  */
1211 static struct sched_domain_topology_level default_topology[] = {
1212 #ifdef CONFIG_SCHED_SMT
1213         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1214 #endif
1215 #ifdef CONFIG_SCHED_MC
1216         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1217 #endif
1218         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1219         { NULL, },
1220 };
1221
1222 static struct sched_domain_topology_level *sched_domain_topology =
1223         default_topology;
1224
1225 #define for_each_sd_topology(tl)                        \
1226         for (tl = sched_domain_topology; tl->mask; tl++)
1227
1228 void set_sched_topology(struct sched_domain_topology_level *tl)
1229 {
1230         if (WARN_ON_ONCE(sched_smp_initialized))
1231                 return;
1232
1233         sched_domain_topology = tl;
1234 }
1235
1236 #ifdef CONFIG_NUMA
1237
1238 static const struct cpumask *sd_numa_mask(int cpu)
1239 {
1240         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1241 }
1242
1243 static void sched_numa_warn(const char *str)
1244 {
1245         static int done = false;
1246         int i,j;
1247
1248         if (done)
1249                 return;
1250
1251         done = true;
1252
1253         printk(KERN_WARNING "ERROR: %s\n\n", str);
1254
1255         for (i = 0; i < nr_node_ids; i++) {
1256                 printk(KERN_WARNING "  ");
1257                 for (j = 0; j < nr_node_ids; j++)
1258                         printk(KERN_CONT "%02d ", node_distance(i,j));
1259                 printk(KERN_CONT "\n");
1260         }
1261         printk(KERN_WARNING "\n");
1262 }
1263
1264 bool find_numa_distance(int distance)
1265 {
1266         int i;
1267
1268         if (distance == node_distance(0, 0))
1269                 return true;
1270
1271         for (i = 0; i < sched_domains_numa_levels; i++) {
1272                 if (sched_domains_numa_distance[i] == distance)
1273                         return true;
1274         }
1275
1276         return false;
1277 }
1278
1279 /*
1280  * A system can have three types of NUMA topology:
1281  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1282  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1283  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1284  *
1285  * The difference between a glueless mesh topology and a backplane
1286  * topology lies in whether communication between not directly
1287  * connected nodes goes through intermediary nodes (where programs
1288  * could run), or through backplane controllers. This affects
1289  * placement of programs.
1290  *
1291  * The type of topology can be discerned with the following tests:
1292  * - If the maximum distance between any nodes is 1 hop, the system
1293  *   is directly connected.
1294  * - If for two nodes A and B, located N > 1 hops away from each other,
1295  *   there is an intermediary node C, which is < N hops away from both
1296  *   nodes A and B, the system is a glueless mesh.
1297  */
1298 static void init_numa_topology_type(void)
1299 {
1300         int a, b, c, n;
1301
1302         n = sched_max_numa_distance;
1303
1304         if (sched_domains_numa_levels <= 2) {
1305                 sched_numa_topology_type = NUMA_DIRECT;
1306                 return;
1307         }
1308
1309         for_each_online_node(a) {
1310                 for_each_online_node(b) {
1311                         /* Find two nodes furthest removed from each other. */
1312                         if (node_distance(a, b) < n)
1313                                 continue;
1314
1315                         /* Is there an intermediary node between a and b? */
1316                         for_each_online_node(c) {
1317                                 if (node_distance(a, c) < n &&
1318                                     node_distance(b, c) < n) {
1319                                         sched_numa_topology_type =
1320                                                         NUMA_GLUELESS_MESH;
1321                                         return;
1322                                 }
1323                         }
1324
1325                         sched_numa_topology_type = NUMA_BACKPLANE;
1326                         return;
1327                 }
1328         }
1329 }
1330
1331 void sched_init_numa(void)
1332 {
1333         int next_distance, curr_distance = node_distance(0, 0);
1334         struct sched_domain_topology_level *tl;
1335         int level = 0;
1336         int i, j, k;
1337
1338         sched_domains_numa_distance = kzalloc(sizeof(int) * (nr_node_ids + 1), GFP_KERNEL);
1339         if (!sched_domains_numa_distance)
1340                 return;
1341
1342         /* Includes NUMA identity node at level 0. */
1343         sched_domains_numa_distance[level++] = curr_distance;
1344         sched_domains_numa_levels = level;
1345
1346         /*
1347          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1348          * unique distances in the node_distance() table.
1349          *
1350          * Assumes node_distance(0,j) includes all distances in
1351          * node_distance(i,j) in order to avoid cubic time.
1352          */
1353         next_distance = curr_distance;
1354         for (i = 0; i < nr_node_ids; i++) {
1355                 for (j = 0; j < nr_node_ids; j++) {
1356                         for (k = 0; k < nr_node_ids; k++) {
1357                                 int distance = node_distance(i, k);
1358
1359                                 if (distance > curr_distance &&
1360                                     (distance < next_distance ||
1361                                      next_distance == curr_distance))
1362                                         next_distance = distance;
1363
1364                                 /*
1365                                  * While not a strong assumption it would be nice to know
1366                                  * about cases where if node A is connected to B, B is not
1367                                  * equally connected to A.
1368                                  */
1369                                 if (sched_debug() && node_distance(k, i) != distance)
1370                                         sched_numa_warn("Node-distance not symmetric");
1371
1372                                 if (sched_debug() && i && !find_numa_distance(distance))
1373                                         sched_numa_warn("Node-0 not representative");
1374                         }
1375                         if (next_distance != curr_distance) {
1376                                 sched_domains_numa_distance[level++] = next_distance;
1377                                 sched_domains_numa_levels = level;
1378                                 curr_distance = next_distance;
1379                         } else break;
1380                 }
1381
1382                 /*
1383                  * In case of sched_debug() we verify the above assumption.
1384                  */
1385                 if (!sched_debug())
1386                         break;
1387         }
1388
1389         /*
1390          * 'level' contains the number of unique distances
1391          *
1392          * The sched_domains_numa_distance[] array includes the actual distance
1393          * numbers.
1394          */
1395
1396         /*
1397          * Here, we should temporarily reset sched_domains_numa_levels to 0.
1398          * If it fails to allocate memory for array sched_domains_numa_masks[][],
1399          * the array will contain less then 'level' members. This could be
1400          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1401          * in other functions.
1402          *
1403          * We reset it to 'level' at the end of this function.
1404          */
1405         sched_domains_numa_levels = 0;
1406
1407         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1408         if (!sched_domains_numa_masks)
1409                 return;
1410
1411         /*
1412          * Now for each level, construct a mask per node which contains all
1413          * CPUs of nodes that are that many hops away from us.
1414          */
1415         for (i = 0; i < level; i++) {
1416                 sched_domains_numa_masks[i] =
1417                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1418                 if (!sched_domains_numa_masks[i])
1419                         return;
1420
1421                 for (j = 0; j < nr_node_ids; j++) {
1422                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1423                         if (!mask)
1424                                 return;
1425
1426                         sched_domains_numa_masks[i][j] = mask;
1427
1428                         for_each_node(k) {
1429                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
1430                                         continue;
1431
1432                                 cpumask_or(mask, mask, cpumask_of_node(k));
1433                         }
1434                 }
1435         }
1436
1437         /* Compute default topology size */
1438         for (i = 0; sched_domain_topology[i].mask; i++);
1439
1440         tl = kzalloc((i + level + 1) *
1441                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1442         if (!tl)
1443                 return;
1444
1445         /*
1446          * Copy the default topology bits..
1447          */
1448         for (i = 0; sched_domain_topology[i].mask; i++)
1449                 tl[i] = sched_domain_topology[i];
1450
1451         /*
1452          * Add the NUMA identity distance, aka single NODE.
1453          */
1454         tl[i++] = (struct sched_domain_topology_level){
1455                 .mask = sd_numa_mask,
1456                 .numa_level = 0,
1457                 SD_INIT_NAME(NODE)
1458         };
1459
1460         /*
1461          * .. and append 'j' levels of NUMA goodness.
1462          */
1463         for (j = 1; j < level; i++, j++) {
1464                 tl[i] = (struct sched_domain_topology_level){
1465                         .mask = sd_numa_mask,
1466                         .sd_flags = cpu_numa_flags,
1467                         .flags = SDTL_OVERLAP,
1468                         .numa_level = j,
1469                         SD_INIT_NAME(NUMA)
1470                 };
1471         }
1472
1473         sched_domain_topology = tl;
1474
1475         sched_domains_numa_levels = level;
1476         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1477
1478         init_numa_topology_type();
1479 }
1480
1481 void sched_domains_numa_masks_set(unsigned int cpu)
1482 {
1483         int node = cpu_to_node(cpu);
1484         int i, j;
1485
1486         for (i = 0; i < sched_domains_numa_levels; i++) {
1487                 for (j = 0; j < nr_node_ids; j++) {
1488                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
1489                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1490                 }
1491         }
1492 }
1493
1494 void sched_domains_numa_masks_clear(unsigned int cpu)
1495 {
1496         int i, j;
1497
1498         for (i = 0; i < sched_domains_numa_levels; i++) {
1499                 for (j = 0; j < nr_node_ids; j++)
1500                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1501         }
1502 }
1503
1504 #endif /* CONFIG_NUMA */
1505
1506 static int __sdt_alloc(const struct cpumask *cpu_map)
1507 {
1508         struct sched_domain_topology_level *tl;
1509         int j;
1510
1511         for_each_sd_topology(tl) {
1512                 struct sd_data *sdd = &tl->data;
1513
1514                 sdd->sd = alloc_percpu(struct sched_domain *);
1515                 if (!sdd->sd)
1516                         return -ENOMEM;
1517
1518                 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1519                 if (!sdd->sds)
1520                         return -ENOMEM;
1521
1522                 sdd->sg = alloc_percpu(struct sched_group *);
1523                 if (!sdd->sg)
1524                         return -ENOMEM;
1525
1526                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1527                 if (!sdd->sgc)
1528                         return -ENOMEM;
1529
1530                 for_each_cpu(j, cpu_map) {
1531                         struct sched_domain *sd;
1532                         struct sched_domain_shared *sds;
1533                         struct sched_group *sg;
1534                         struct sched_group_capacity *sgc;
1535
1536                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1537                                         GFP_KERNEL, cpu_to_node(j));
1538                         if (!sd)
1539                                 return -ENOMEM;
1540
1541                         *per_cpu_ptr(sdd->sd, j) = sd;
1542
1543                         sds = kzalloc_node(sizeof(struct sched_domain_shared),
1544                                         GFP_KERNEL, cpu_to_node(j));
1545                         if (!sds)
1546                                 return -ENOMEM;
1547
1548                         *per_cpu_ptr(sdd->sds, j) = sds;
1549
1550                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1551                                         GFP_KERNEL, cpu_to_node(j));
1552                         if (!sg)
1553                                 return -ENOMEM;
1554
1555                         sg->next = sg;
1556
1557                         *per_cpu_ptr(sdd->sg, j) = sg;
1558
1559                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1560                                         GFP_KERNEL, cpu_to_node(j));
1561                         if (!sgc)
1562                                 return -ENOMEM;
1563
1564 #ifdef CONFIG_SCHED_DEBUG
1565                         sgc->id = j;
1566 #endif
1567
1568                         *per_cpu_ptr(sdd->sgc, j) = sgc;
1569                 }
1570         }
1571
1572         return 0;
1573 }
1574
1575 static void __sdt_free(const struct cpumask *cpu_map)
1576 {
1577         struct sched_domain_topology_level *tl;
1578         int j;
1579
1580         for_each_sd_topology(tl) {
1581                 struct sd_data *sdd = &tl->data;
1582
1583                 for_each_cpu(j, cpu_map) {
1584                         struct sched_domain *sd;
1585
1586                         if (sdd->sd) {
1587                                 sd = *per_cpu_ptr(sdd->sd, j);
1588                                 if (sd && (sd->flags & SD_OVERLAP))
1589                                         free_sched_groups(sd->groups, 0);
1590                                 kfree(*per_cpu_ptr(sdd->sd, j));
1591                         }
1592
1593                         if (sdd->sds)
1594                                 kfree(*per_cpu_ptr(sdd->sds, j));
1595                         if (sdd->sg)
1596                                 kfree(*per_cpu_ptr(sdd->sg, j));
1597                         if (sdd->sgc)
1598                                 kfree(*per_cpu_ptr(sdd->sgc, j));
1599                 }
1600                 free_percpu(sdd->sd);
1601                 sdd->sd = NULL;
1602                 free_percpu(sdd->sds);
1603                 sdd->sds = NULL;
1604                 free_percpu(sdd->sg);
1605                 sdd->sg = NULL;
1606                 free_percpu(sdd->sgc);
1607                 sdd->sgc = NULL;
1608         }
1609 }
1610
1611 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1612                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1613                 struct sched_domain *child, int dflags, int cpu)
1614 {
1615         struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1616
1617         if (child) {
1618                 sd->level = child->level + 1;
1619                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1620                 child->parent = sd;
1621
1622                 if (!cpumask_subset(sched_domain_span(child),
1623                                     sched_domain_span(sd))) {
1624                         pr_err("BUG: arch topology borken\n");
1625 #ifdef CONFIG_SCHED_DEBUG
1626                         pr_err("     the %s domain not a subset of the %s domain\n",
1627                                         child->name, sd->name);
1628 #endif
1629                         /* Fixup, ensure @sd has at least @child CPUs. */
1630                         cpumask_or(sched_domain_span(sd),
1631                                    sched_domain_span(sd),
1632                                    sched_domain_span(child));
1633                 }
1634
1635         }
1636         set_domain_attribute(sd, attr);
1637
1638         return sd;
1639 }
1640
1641 /*
1642  * Find the sched_domain_topology_level where all CPU capacities are visible
1643  * for all CPUs.
1644  */
1645 static struct sched_domain_topology_level
1646 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1647 {
1648         int i, j, asym_level = 0;
1649         bool asym = false;
1650         struct sched_domain_topology_level *tl, *asym_tl = NULL;
1651         unsigned long cap;
1652
1653         /* Is there any asymmetry? */
1654         cap = arch_scale_cpu_capacity(NULL, cpumask_first(cpu_map));
1655
1656         for_each_cpu(i, cpu_map) {
1657                 if (arch_scale_cpu_capacity(NULL, i) != cap) {
1658                         asym = true;
1659                         break;
1660                 }
1661         }
1662
1663         if (!asym)
1664                 return NULL;
1665
1666         /*
1667          * Examine topology from all CPU's point of views to detect the lowest
1668          * sched_domain_topology_level where a highest capacity CPU is visible
1669          * to everyone.
1670          */
1671         for_each_cpu(i, cpu_map) {
1672                 unsigned long max_capacity = arch_scale_cpu_capacity(NULL, i);
1673                 int tl_id = 0;
1674
1675                 for_each_sd_topology(tl) {
1676                         if (tl_id < asym_level)
1677                                 goto next_level;
1678
1679                         for_each_cpu_and(j, tl->mask(i), cpu_map) {
1680                                 unsigned long capacity;
1681
1682                                 capacity = arch_scale_cpu_capacity(NULL, j);
1683
1684                                 if (capacity <= max_capacity)
1685                                         continue;
1686
1687                                 max_capacity = capacity;
1688                                 asym_level = tl_id;
1689                                 asym_tl = tl;
1690                         }
1691 next_level:
1692                         tl_id++;
1693                 }
1694         }
1695
1696         return asym_tl;
1697 }
1698
1699
1700 /*
1701  * Build sched domains for a given set of CPUs and attach the sched domains
1702  * to the individual CPUs
1703  */
1704 static int
1705 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1706 {
1707         enum s_alloc alloc_state;
1708         struct sched_domain *sd;
1709         struct s_data d;
1710         struct rq *rq = NULL;
1711         int i, ret = -ENOMEM;
1712         struct sched_domain_topology_level *tl_asym;
1713         bool has_asym = false;
1714
1715         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1716         if (alloc_state != sa_rootdomain)
1717                 goto error;
1718
1719         tl_asym = asym_cpu_capacity_level(cpu_map);
1720
1721         /* Set up domains for CPUs specified by the cpu_map: */
1722         for_each_cpu(i, cpu_map) {
1723                 struct sched_domain_topology_level *tl;
1724
1725                 sd = NULL;
1726                 for_each_sd_topology(tl) {
1727                         int dflags = 0;
1728
1729                         if (tl == tl_asym) {
1730                                 dflags |= SD_ASYM_CPUCAPACITY;
1731                                 has_asym = true;
1732                         }
1733
1734                         sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
1735
1736                         if (tl == sched_domain_topology)
1737                                 *per_cpu_ptr(d.sd, i) = sd;
1738                         if (tl->flags & SDTL_OVERLAP)
1739                                 sd->flags |= SD_OVERLAP;
1740                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1741                                 break;
1742                 }
1743         }
1744
1745         /* Build the groups for the domains */
1746         for_each_cpu(i, cpu_map) {
1747                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1748                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
1749                         if (sd->flags & SD_OVERLAP) {
1750                                 if (build_overlap_sched_groups(sd, i))
1751                                         goto error;
1752                         } else {
1753                                 if (build_sched_groups(sd, i))
1754                                         goto error;
1755                         }
1756                 }
1757         }
1758
1759         /* Calculate CPU capacity for physical packages and nodes */
1760         for (i = nr_cpumask_bits-1; i >= 0; i--) {
1761                 if (!cpumask_test_cpu(i, cpu_map))
1762                         continue;
1763
1764                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1765                         claim_allocations(i, sd);
1766                         init_sched_groups_capacity(i, sd);
1767                 }
1768         }
1769
1770         /* Attach the domains */
1771         rcu_read_lock();
1772         for_each_cpu(i, cpu_map) {
1773                 rq = cpu_rq(i);
1774                 sd = *per_cpu_ptr(d.sd, i);
1775
1776                 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1777                 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1778                         WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1779
1780                 cpu_attach_domain(sd, d.rd, i);
1781         }
1782         rcu_read_unlock();
1783
1784         if (has_asym)
1785                 static_branch_enable_cpuslocked(&sched_asym_cpucapacity);
1786
1787         if (rq && sched_debug_enabled) {
1788                 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
1789                         cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1790         }
1791
1792         ret = 0;
1793 error:
1794         __free_domain_allocs(&d, alloc_state, cpu_map);
1795
1796         return ret;
1797 }
1798
1799 /* Current sched domains: */
1800 static cpumask_var_t                    *doms_cur;
1801
1802 /* Number of sched domains in 'doms_cur': */
1803 static int                              ndoms_cur;
1804
1805 /* Attribues of custom domains in 'doms_cur' */
1806 static struct sched_domain_attr         *dattr_cur;
1807
1808 /*
1809  * Special case: If a kmalloc() of a doms_cur partition (array of
1810  * cpumask) fails, then fallback to a single sched domain,
1811  * as determined by the single cpumask fallback_doms.
1812  */
1813 static cpumask_var_t                    fallback_doms;
1814
1815 /*
1816  * arch_update_cpu_topology lets virtualized architectures update the
1817  * CPU core maps. It is supposed to return 1 if the topology changed
1818  * or 0 if it stayed the same.
1819  */
1820 int __weak arch_update_cpu_topology(void)
1821 {
1822         return 0;
1823 }
1824
1825 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
1826 {
1827         int i;
1828         cpumask_var_t *doms;
1829
1830         doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
1831         if (!doms)
1832                 return NULL;
1833         for (i = 0; i < ndoms; i++) {
1834                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
1835                         free_sched_domains(doms, i);
1836                         return NULL;
1837                 }
1838         }
1839         return doms;
1840 }
1841
1842 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
1843 {
1844         unsigned int i;
1845         for (i = 0; i < ndoms; i++)
1846                 free_cpumask_var(doms[i]);
1847         kfree(doms);
1848 }
1849
1850 /*
1851  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1852  * For now this just excludes isolated CPUs, but could be used to
1853  * exclude other special cases in the future.
1854  */
1855 int sched_init_domains(const struct cpumask *cpu_map)
1856 {
1857         int err;
1858
1859         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
1860         zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
1861         zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
1862
1863         arch_update_cpu_topology();
1864         ndoms_cur = 1;
1865         doms_cur = alloc_sched_domains(ndoms_cur);
1866         if (!doms_cur)
1867                 doms_cur = &fallback_doms;
1868         cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
1869         err = build_sched_domains(doms_cur[0], NULL);
1870         register_sched_domain_sysctl();
1871
1872         return err;
1873 }
1874
1875 /*
1876  * Detach sched domains from a group of CPUs specified in cpu_map
1877  * These CPUs will now be attached to the NULL domain
1878  */
1879 static void detach_destroy_domains(const struct cpumask *cpu_map)
1880 {
1881         int i;
1882
1883         rcu_read_lock();
1884         for_each_cpu(i, cpu_map)
1885                 cpu_attach_domain(NULL, &def_root_domain, i);
1886         rcu_read_unlock();
1887 }
1888
1889 /* handle null as "default" */
1890 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
1891                         struct sched_domain_attr *new, int idx_new)
1892 {
1893         struct sched_domain_attr tmp;
1894
1895         /* Fast path: */
1896         if (!new && !cur)
1897                 return 1;
1898
1899         tmp = SD_ATTR_INIT;
1900
1901         return !memcmp(cur ? (cur + idx_cur) : &tmp,
1902                         new ? (new + idx_new) : &tmp,
1903                         sizeof(struct sched_domain_attr));
1904 }
1905
1906 /*
1907  * Partition sched domains as specified by the 'ndoms_new'
1908  * cpumasks in the array doms_new[] of cpumasks. This compares
1909  * doms_new[] to the current sched domain partitioning, doms_cur[].
1910  * It destroys each deleted domain and builds each new domain.
1911  *
1912  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1913  * The masks don't intersect (don't overlap.) We should setup one
1914  * sched domain for each mask. CPUs not in any of the cpumasks will
1915  * not be load balanced. If the same cpumask appears both in the
1916  * current 'doms_cur' domains and in the new 'doms_new', we can leave
1917  * it as it is.
1918  *
1919  * The passed in 'doms_new' should be allocated using
1920  * alloc_sched_domains.  This routine takes ownership of it and will
1921  * free_sched_domains it when done with it. If the caller failed the
1922  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1923  * and partition_sched_domains() will fallback to the single partition
1924  * 'fallback_doms', it also forces the domains to be rebuilt.
1925  *
1926  * If doms_new == NULL it will be replaced with cpu_online_mask.
1927  * ndoms_new == 0 is a special case for destroying existing domains,
1928  * and it will not create the default domain.
1929  *
1930  * Call with hotplug lock held
1931  */
1932 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1933                              struct sched_domain_attr *dattr_new)
1934 {
1935         int i, j, n;
1936         int new_topology;
1937
1938         mutex_lock(&sched_domains_mutex);
1939
1940         /* Always unregister in case we don't destroy any domains: */
1941         unregister_sched_domain_sysctl();
1942
1943         /* Let the architecture update CPU core mappings: */
1944         new_topology = arch_update_cpu_topology();
1945
1946         if (!doms_new) {
1947                 WARN_ON_ONCE(dattr_new);
1948                 n = 0;
1949                 doms_new = alloc_sched_domains(1);
1950                 if (doms_new) {
1951                         n = 1;
1952                         cpumask_and(doms_new[0], cpu_active_mask,
1953                                     housekeeping_cpumask(HK_FLAG_DOMAIN));
1954                 }
1955         } else {
1956                 n = ndoms_new;
1957         }
1958
1959         /* Destroy deleted domains: */
1960         for (i = 0; i < ndoms_cur; i++) {
1961                 for (j = 0; j < n && !new_topology; j++) {
1962                         if (cpumask_equal(doms_cur[i], doms_new[j])
1963                             && dattrs_equal(dattr_cur, i, dattr_new, j))
1964                                 goto match1;
1965                 }
1966                 /* No match - a current sched domain not in new doms_new[] */
1967                 detach_destroy_domains(doms_cur[i]);
1968 match1:
1969                 ;
1970         }
1971
1972         n = ndoms_cur;
1973         if (!doms_new) {
1974                 n = 0;
1975                 doms_new = &fallback_doms;
1976                 cpumask_and(doms_new[0], cpu_active_mask,
1977                             housekeeping_cpumask(HK_FLAG_DOMAIN));
1978         }
1979
1980         /* Build new domains: */
1981         for (i = 0; i < ndoms_new; i++) {
1982                 for (j = 0; j < n && !new_topology; j++) {
1983                         if (cpumask_equal(doms_new[i], doms_cur[j])
1984                             && dattrs_equal(dattr_new, i, dattr_cur, j))
1985                                 goto match2;
1986                 }
1987                 /* No match - add a new doms_new */
1988                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
1989 match2:
1990                 ;
1991         }
1992
1993         /* Remember the new sched domains: */
1994         if (doms_cur != &fallback_doms)
1995                 free_sched_domains(doms_cur, ndoms_cur);
1996
1997         kfree(dattr_cur);
1998         doms_cur = doms_new;
1999         dattr_cur = dattr_new;
2000         ndoms_cur = ndoms_new;
2001
2002         register_sched_domain_sysctl();
2003
2004         mutex_unlock(&sched_domains_mutex);
2005 }