kernel/sched/topology.c

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