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