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