| 1 | |
| 2 | |
| 3 | * Scheduler topology setup/handling methods
|
| 4 |
|
| 5 |
|
| 6 | #include <linux/sched/isolation.h>
|
| 7 | #include <linux/sched/clock.h>
|
| 8 | #include <linux/bsearch.h>
|
| 9 | #include "sched.h"
|
| 10 |
|
| 11 | DEFINE_MUTEX(sched_domains_mutex);
|
| 12 | void sched_domains_mutex_lock(void)
|
| 13 | {
|
| 14 | mutex_lock(&sched_domains_mutex);
|
| 15 | }
|
| 16 | void sched_domains_mutex_unlock(void)
|
| 17 | {
|
| 18 | mutex_unlock(&sched_domains_mutex);
|
| 19 | }
|
| 20 |
|
| 21 |
|
| 22 | static cpumask_var_t sched_domains_llc_id_allocmask;
|
| 23 | static cpumask_var_t sched_domains_tmpmask;
|
| 24 | static cpumask_var_t sched_domains_tmpmask2;
|
| 25 | int max_lid;
|
| 26 |
|
| 27 | static int __init sched_debug_setup(char *str)
|
| 28 | {
|
| 29 | sched_debug_verbose = true;
|
| 30 |
|
| 31 | return 0;
|
| 32 | }
|
| 33 | early_param("sched_verbose", sched_debug_setup);
|
| 34 |
|
| 35 | #ifdef CONFIG_SCHED_SMT
|
| 36 | static bool sched_smt_asym_packing __read_mostly;
|
| 37 |
|
| 38 | static int __init setup_sched_smt_asym_packing(char *str)
|
| 39 | {
|
| 40 | if (strcmp(str, "on"))
|
| 41 | return 0;
|
| 42 |
|
| 43 | sched_smt_asym_packing = true;
|
| 44 | return 1;
|
| 45 | }
|
| 46 | __setup("sched_smt_asym_packing=", setup_sched_smt_asym_packing);
|
| 47 | #endif
|
| 48 |
|
| 49 | static inline bool sched_debug(void)
|
| 50 | {
|
| 51 | return sched_debug_verbose;
|
| 52 | }
|
| 53 |
|
| 54 | #define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
|
| 55 | const struct sd_flag_debug sd_flag_debug[] = {
|
| 56 | #include <linux/sched/sd_flags.h>
|
| 57 | };
|
| 58 | #undef SD_FLAG
|
| 59 |
|
| 60 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
|
| 61 | struct cpumask *groupmask)
|
| 62 | {
|
| 63 | struct sched_group *group = sd->groups;
|
| 64 | unsigned long flags = sd->flags;
|
| 65 | unsigned int idx;
|
| 66 |
|
| 67 | cpumask_clear(groupmask);
|
| 68 |
|
| 69 | printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
|
| 70 | printk(KERN_CONT "span=%*pbl level=%s\n",
|
| 71 | cpumask_pr_args(sched_domain_span(sd)), sd->name);
|
| 72 |
|
| 73 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
|
| 74 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
|
| 75 | }
|
| 76 | if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
|
| 77 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
|
| 78 | }
|
| 79 |
|
| 80 | for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
|
| 81 | unsigned int flag = BIT(idx);
|
| 82 | unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
|
| 83 |
|
| 84 | if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
|
| 85 | !(sd->child->flags & flag))
|
| 86 | printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
|
| 87 | sd_flag_debug[idx].name);
|
| 88 |
|
| 89 | if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
|
| 90 | !(sd->parent->flags & flag))
|
| 91 | printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
|
| 92 | sd_flag_debug[idx].name);
|
| 93 | }
|
| 94 |
|
| 95 | printk(KERN_DEBUG "%*s groups:", level + 1, "");
|
| 96 | do {
|
| 97 | if (!group) {
|
| 98 | printk("\n");
|
| 99 | printk(KERN_ERR "ERROR: group is NULL\n");
|
| 100 | break;
|
| 101 | }
|
| 102 |
|
| 103 | if (cpumask_empty(sched_group_span(group))) {
|
| 104 | printk(KERN_CONT "\n");
|
| 105 | printk(KERN_ERR "ERROR: empty group\n");
|
| 106 | break;
|
| 107 | }
|
| 108 |
|
| 109 | if (!(sd->flags & SD_NUMA) &&
|
| 110 | cpumask_intersects(groupmask, sched_group_span(group))) {
|
| 111 | printk(KERN_CONT "\n");
|
| 112 | printk(KERN_ERR "ERROR: repeated CPUs\n");
|
| 113 | break;
|
| 114 | }
|
| 115 |
|
| 116 | cpumask_or(groupmask, groupmask, sched_group_span(group));
|
| 117 |
|
| 118 | printk(KERN_CONT " %d:{ span=%*pbl",
|
| 119 | group->sgc->id,
|
| 120 | cpumask_pr_args(sched_group_span(group)));
|
| 121 |
|
| 122 | if ((sd->flags & SD_NUMA) &&
|
| 123 | !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
|
| 124 | printk(KERN_CONT " mask=%*pbl",
|
| 125 | cpumask_pr_args(group_balance_mask(group)));
|
| 126 | }
|
| 127 |
|
| 128 | if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
|
| 129 | printk(KERN_CONT " cap=%lu", group->sgc->capacity);
|
| 130 |
|
| 131 | if (group == sd->groups && sd->child &&
|
| 132 | !cpumask_equal(sched_domain_span(sd->child),
|
| 133 | sched_group_span(group))) {
|
| 134 | printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
|
| 135 | }
|
| 136 |
|
| 137 | printk(KERN_CONT " }");
|
| 138 |
|
| 139 | group = group->next;
|
| 140 |
|
| 141 | if (group != sd->groups)
|
| 142 | printk(KERN_CONT ",");
|
| 143 |
|
| 144 | } while (group != sd->groups);
|
| 145 | printk(KERN_CONT "\n");
|
| 146 |
|
| 147 | if (!cpumask_equal(sched_domain_span(sd), groupmask))
|
| 148 | printk(KERN_ERR "ERROR: groups don't span domain->span\n");
|
| 149 |
|
| 150 | if (sd->parent &&
|
| 151 | !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
|
| 152 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
|
| 153 | return 0;
|
| 154 | }
|
| 155 |
|
| 156 | static void sched_domain_debug(struct sched_domain *sd, int cpu)
|
| 157 | {
|
| 158 | int level = 0;
|
| 159 |
|
| 160 | if (!sched_debug_verbose)
|
| 161 | return;
|
| 162 |
|
| 163 | if (!sd) {
|
| 164 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
|
| 165 | return;
|
| 166 | }
|
| 167 |
|
| 168 | printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
|
| 169 |
|
| 170 | for (;;) {
|
| 171 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
|
| 172 | break;
|
| 173 | level++;
|
| 174 | sd = sd->parent;
|
| 175 | if (!sd)
|
| 176 | break;
|
| 177 | }
|
| 178 | }
|
| 179 |
|
| 180 |
|
| 181 | #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
|
| 182 | static const unsigned int SD_DEGENERATE_GROUPS_MASK =
|
| 183 | #include <linux/sched/sd_flags.h>
|
| 184 | 0;
|
| 185 | #undef SD_FLAG
|
| 186 |
|
| 187 | static int sd_degenerate(struct sched_domain *sd)
|
| 188 | {
|
| 189 | if (cpumask_weight(sched_domain_span(sd)) == 1)
|
| 190 | return 1;
|
| 191 |
|
| 192 |
|
| 193 | if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
|
| 194 | (sd->groups != sd->groups->next))
|
| 195 | return 0;
|
| 196 |
|
| 197 |
|
| 198 | if (sd->flags & (SD_WAKE_AFFINE))
|
| 199 | return 0;
|
| 200 |
|
| 201 | return 1;
|
| 202 | }
|
| 203 |
|
| 204 | static int
|
| 205 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
|
| 206 | {
|
| 207 | unsigned long cflags = sd->flags, pflags = parent->flags;
|
| 208 |
|
| 209 | if (sd_degenerate(parent))
|
| 210 | return 1;
|
| 211 |
|
| 212 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
|
| 213 | return 0;
|
| 214 |
|
| 215 |
|
| 216 | if (parent->groups == parent->groups->next)
|
| 217 | pflags &= ~SD_DEGENERATE_GROUPS_MASK;
|
| 218 |
|
| 219 | if (~cflags & pflags)
|
| 220 | return 0;
|
| 221 |
|
| 222 | return 1;
|
| 223 | }
|
| 224 |
|
| 225 | #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
|
| 226 | DEFINE_STATIC_KEY_FALSE(sched_energy_present);
|
| 227 | static unsigned int sysctl_sched_energy_aware = 1;
|
| 228 | static DEFINE_MUTEX(sched_energy_mutex);
|
| 229 | static bool sched_energy_update;
|
| 230 |
|
| 231 | static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
|
| 232 | {
|
| 233 | bool any_asym_capacity = false;
|
| 234 | int i;
|
| 235 |
|
| 236 |
|
| 237 | for_each_cpu(i, cpu_mask) {
|
| 238 | if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
|
| 239 | any_asym_capacity = true;
|
| 240 | break;
|
| 241 | }
|
| 242 | }
|
| 243 | if (!any_asym_capacity) {
|
| 244 | if (sched_debug()) {
|
| 245 | pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
|
| 246 | cpumask_pr_args(cpu_mask));
|
| 247 | }
|
| 248 | return false;
|
| 249 | }
|
| 250 |
|
| 251 |
|
| 252 | if (sched_smt_active()) {
|
| 253 | if (sched_debug()) {
|
| 254 | pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
|
| 255 | cpumask_pr_args(cpu_mask));
|
| 256 | }
|
| 257 | return false;
|
| 258 | }
|
| 259 |
|
| 260 | if (!arch_scale_freq_invariant()) {
|
| 261 | if (sched_debug()) {
|
| 262 | pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
|
| 263 | cpumask_pr_args(cpu_mask));
|
| 264 | }
|
| 265 | return false;
|
| 266 | }
|
| 267 |
|
| 268 | if (!cpufreq_ready_for_eas(cpu_mask)) {
|
| 269 | if (sched_debug()) {
|
| 270 | pr_info("rd %*pbl: Checking EAS: cpufreq is not ready\n",
|
| 271 | cpumask_pr_args(cpu_mask));
|
| 272 | }
|
| 273 | return false;
|
| 274 | }
|
| 275 |
|
| 276 | return true;
|
| 277 | }
|
| 278 |
|
| 279 | void rebuild_sched_domains_energy(void)
|
| 280 | {
|
| 281 | mutex_lock(&sched_energy_mutex);
|
| 282 | sched_energy_update = true;
|
| 283 | rebuild_sched_domains();
|
| 284 | sched_energy_update = false;
|
| 285 | mutex_unlock(&sched_energy_mutex);
|
| 286 | }
|
| 287 |
|
| 288 | #ifdef CONFIG_SYSCTL
|
| 289 | static int sched_energy_aware_handler(const struct ctl_table *table, int write,
|
| 290 | void *buffer, size_t *lenp, loff_t *ppos)
|
| 291 | {
|
| 292 | int ret;
|
| 293 |
|
| 294 | if (write && !capable(CAP_SYS_ADMIN))
|
| 295 | return -EPERM;
|
| 296 |
|
| 297 | if (!sched_is_eas_possible(cpu_active_mask)) {
|
| 298 | if (write) {
|
| 299 | return -EOPNOTSUPP;
|
| 300 | } else {
|
| 301 | *lenp = 0;
|
| 302 | return 0;
|
| 303 | }
|
| 304 | }
|
| 305 |
|
| 306 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
| 307 | if (!ret && write) {
|
| 308 | if (sysctl_sched_energy_aware != sched_energy_enabled())
|
| 309 | rebuild_sched_domains_energy();
|
| 310 | }
|
| 311 |
|
| 312 | return ret;
|
| 313 | }
|
| 314 |
|
| 315 | static const struct ctl_table sched_energy_aware_sysctls[] = {
|
| 316 | {
|
| 317 | .procname = "sched_energy_aware",
|
| 318 | .data = &sysctl_sched_energy_aware,
|
| 319 | .maxlen = sizeof(unsigned int),
|
| 320 | .mode = 0644,
|
| 321 | .proc_handler = sched_energy_aware_handler,
|
| 322 | .extra1 = SYSCTL_ZERO,
|
| 323 | .extra2 = SYSCTL_ONE,
|
| 324 | },
|
| 325 | };
|
| 326 |
|
| 327 | static int __init sched_energy_aware_sysctl_init(void)
|
| 328 | {
|
| 329 | register_sysctl_init("kernel", sched_energy_aware_sysctls);
|
| 330 | return 0;
|
| 331 | }
|
| 332 |
|
| 333 | late_initcall(sched_energy_aware_sysctl_init);
|
| 334 | #endif
|
| 335 |
|
| 336 | static void free_pd(struct perf_domain *pd)
|
| 337 | {
|
| 338 | struct perf_domain *tmp;
|
| 339 |
|
| 340 | while (pd) {
|
| 341 | tmp = pd->next;
|
| 342 | kfree(pd);
|
| 343 | pd = tmp;
|
| 344 | }
|
| 345 | }
|
| 346 |
|
| 347 | static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
|
| 348 | {
|
| 349 | while (pd) {
|
| 350 | if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
|
| 351 | return pd;
|
| 352 | pd = pd->next;
|
| 353 | }
|
| 354 |
|
| 355 | return NULL;
|
| 356 | }
|
| 357 |
|
| 358 | static struct perf_domain *pd_init(int cpu)
|
| 359 | {
|
| 360 | struct em_perf_domain *obj = em_cpu_get(cpu);
|
| 361 | struct perf_domain *pd;
|
| 362 |
|
| 363 | if (!obj) {
|
| 364 | if (sched_debug())
|
| 365 | pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
|
| 366 | return NULL;
|
| 367 | }
|
| 368 |
|
| 369 | pd = kzalloc_obj(*pd);
|
| 370 | if (!pd)
|
| 371 | return NULL;
|
| 372 | pd->em_pd = obj;
|
| 373 |
|
| 374 | return pd;
|
| 375 | }
|
| 376 |
|
| 377 | static void perf_domain_debug(const struct cpumask *cpu_map,
|
| 378 | struct perf_domain *pd)
|
| 379 | {
|
| 380 | if (!sched_debug() || !pd)
|
| 381 | return;
|
| 382 |
|
| 383 | printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
|
| 384 |
|
| 385 | while (pd) {
|
| 386 | printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
|
| 387 | cpumask_first(perf_domain_span(pd)),
|
| 388 | cpumask_pr_args(perf_domain_span(pd)),
|
| 389 | em_pd_nr_perf_states(pd->em_pd));
|
| 390 | pd = pd->next;
|
| 391 | }
|
| 392 |
|
| 393 | printk(KERN_CONT "\n");
|
| 394 | }
|
| 395 |
|
| 396 | static void destroy_perf_domain_rcu(struct rcu_head *rp)
|
| 397 | {
|
| 398 | struct perf_domain *pd;
|
| 399 |
|
| 400 | pd = container_of(rp, struct perf_domain, rcu);
|
| 401 | free_pd(pd);
|
| 402 | }
|
| 403 |
|
| 404 | static void sched_energy_set(bool has_eas)
|
| 405 | {
|
| 406 | if (!has_eas && sched_energy_enabled()) {
|
| 407 | if (sched_debug())
|
| 408 | pr_info("%s: stopping EAS\n", __func__);
|
| 409 | static_branch_disable_cpuslocked(&sched_energy_present);
|
| 410 | } else if (has_eas && !sched_energy_enabled()) {
|
| 411 | if (sched_debug())
|
| 412 | pr_info("%s: starting EAS\n", __func__);
|
| 413 | static_branch_enable_cpuslocked(&sched_energy_present);
|
| 414 | }
|
| 415 | }
|
| 416 |
|
| 417 | |
| 418 | * EAS can be used on a root domain if it meets all the following conditions:
|
| 419 | * 1. an Energy Model (EM) is available;
|
| 420 | * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
|
| 421 | * 3. no SMT is detected.
|
| 422 | * 4. schedutil is driving the frequency of all CPUs of the rd;
|
| 423 | * 5. frequency invariance support is present;
|
| 424 |
|
| 425 | static bool build_perf_domains(const struct cpumask *cpu_map)
|
| 426 | {
|
| 427 | int i;
|
| 428 | struct perf_domain *pd = NULL, *tmp;
|
| 429 | int cpu = cpumask_first(cpu_map);
|
| 430 | struct root_domain *rd = cpu_rq(cpu)->rd;
|
| 431 |
|
| 432 | if (!sysctl_sched_energy_aware)
|
| 433 | goto free;
|
| 434 |
|
| 435 | if (!sched_is_eas_possible(cpu_map))
|
| 436 | goto free;
|
| 437 |
|
| 438 | for_each_cpu(i, cpu_map) {
|
| 439 |
|
| 440 | if (find_pd(pd, i))
|
| 441 | continue;
|
| 442 |
|
| 443 |
|
| 444 | tmp = pd_init(i);
|
| 445 | if (!tmp)
|
| 446 | goto free;
|
| 447 | tmp->next = pd;
|
| 448 | pd = tmp;
|
| 449 | }
|
| 450 |
|
| 451 | perf_domain_debug(cpu_map, pd);
|
| 452 |
|
| 453 |
|
| 454 | tmp = rd->pd;
|
| 455 | rcu_assign_pointer(rd->pd, pd);
|
| 456 | if (tmp)
|
| 457 | call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
|
| 458 |
|
| 459 | return !!pd;
|
| 460 |
|
| 461 | free:
|
| 462 | free_pd(pd);
|
| 463 | tmp = rd->pd;
|
| 464 | rcu_assign_pointer(rd->pd, NULL);
|
| 465 | if (tmp)
|
| 466 | call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
|
| 467 |
|
| 468 | return false;
|
| 469 | }
|
| 470 | #else
|
| 471 | static void free_pd(struct perf_domain *pd) { }
|
| 472 | #endif
|
| 473 |
|
| 474 | static void free_rootdomain(struct rcu_head *rcu)
|
| 475 | {
|
| 476 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
|
| 477 |
|
| 478 | cpupri_cleanup(&rd->cpupri);
|
| 479 | cpudl_cleanup(&rd->cpudl);
|
| 480 | free_cpumask_var(rd->dlo_mask);
|
| 481 | free_cpumask_var(rd->rto_mask);
|
| 482 | free_cpumask_var(rd->online);
|
| 483 | free_cpumask_var(rd->span);
|
| 484 | free_pd(rd->pd);
|
| 485 | kfree(rd);
|
| 486 | }
|
| 487 |
|
| 488 | void rq_attach_root(struct rq *rq, struct root_domain *rd)
|
| 489 | {
|
| 490 | struct root_domain *old_rd = NULL;
|
| 491 | struct rq_flags rf;
|
| 492 |
|
| 493 | rq_lock_irqsave(rq, &rf);
|
| 494 |
|
| 495 | if (rq->rd) {
|
| 496 | old_rd = rq->rd;
|
| 497 |
|
| 498 | if (cpumask_test_cpu(rq->cpu, old_rd->online))
|
| 499 | set_rq_offline(rq);
|
| 500 |
|
| 501 | cpumask_clear_cpu(rq->cpu, old_rd->span);
|
| 502 |
|
| 503 | |
| 504 | * If we don't want to free the old_rd yet then
|
| 505 | * set old_rd to NULL to skip the freeing later
|
| 506 | * in this function:
|
| 507 |
|
| 508 | if (!atomic_dec_and_test(&old_rd->refcount))
|
| 509 | old_rd = NULL;
|
| 510 | }
|
| 511 |
|
| 512 | atomic_inc(&rd->refcount);
|
| 513 | rq->rd = rd;
|
| 514 |
|
| 515 | cpumask_set_cpu(rq->cpu, rd->span);
|
| 516 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
|
| 517 | set_rq_online(rq);
|
| 518 |
|
| 519 | |
| 520 | * Because the rq is not a task, dl_add_task_root_domain() did not
|
| 521 | * move the fair server bw to the rd if it already started.
|
| 522 | * Add it now.
|
| 523 |
|
| 524 | if (rq->fair_server.dl_server)
|
| 525 | __dl_server_attach_root(&rq->fair_server, rq);
|
| 526 |
|
| 527 | #ifdef CONFIG_SCHED_CLASS_EXT
|
| 528 | if (rq->ext_server.dl_server)
|
| 529 | __dl_server_attach_root(&rq->ext_server, rq);
|
| 530 | #endif
|
| 531 |
|
| 532 | rq_unlock_irqrestore(rq, &rf);
|
| 533 |
|
| 534 | if (old_rd)
|
| 535 | call_rcu(&old_rd->rcu, free_rootdomain);
|
| 536 | }
|
| 537 |
|
| 538 | void sched_get_rd(struct root_domain *rd)
|
| 539 | {
|
| 540 | atomic_inc(&rd->refcount);
|
| 541 | }
|
| 542 |
|
| 543 | void sched_put_rd(struct root_domain *rd)
|
| 544 | {
|
| 545 | if (!atomic_dec_and_test(&rd->refcount))
|
| 546 | return;
|
| 547 |
|
| 548 | call_rcu(&rd->rcu, free_rootdomain);
|
| 549 | }
|
| 550 |
|
| 551 | static int init_rootdomain(struct root_domain *rd)
|
| 552 | {
|
| 553 | if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
|
| 554 | goto out;
|
| 555 | if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
|
| 556 | goto free_span;
|
| 557 | if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
|
| 558 | goto free_online;
|
| 559 | if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
|
| 560 | goto free_dlo_mask;
|
| 561 |
|
| 562 | #ifdef HAVE_RT_PUSH_IPI
|
| 563 | rd->rto_cpu = -1;
|
| 564 | raw_spin_lock_init(&rd->rto_lock);
|
| 565 | rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
|
| 566 | #endif
|
| 567 |
|
| 568 | rd->visit_cookie = 0;
|
| 569 | init_dl_bw(&rd->dl_bw);
|
| 570 | if (cpudl_init(&rd->cpudl) != 0)
|
| 571 | goto free_rto_mask;
|
| 572 |
|
| 573 | if (cpupri_init(&rd->cpupri) != 0)
|
| 574 | goto free_cpudl;
|
| 575 | return 0;
|
| 576 |
|
| 577 | free_cpudl:
|
| 578 | cpudl_cleanup(&rd->cpudl);
|
| 579 | free_rto_mask:
|
| 580 | free_cpumask_var(rd->rto_mask);
|
| 581 | free_dlo_mask:
|
| 582 | free_cpumask_var(rd->dlo_mask);
|
| 583 | free_online:
|
| 584 | free_cpumask_var(rd->online);
|
| 585 | free_span:
|
| 586 | free_cpumask_var(rd->span);
|
| 587 | out:
|
| 588 | return -ENOMEM;
|
| 589 | }
|
| 590 |
|
| 591 | |
| 592 | * By default the system creates a single root-domain with all CPUs as
|
| 593 | * members (mimicking the global state we have today).
|
| 594 |
|
| 595 | struct root_domain def_root_domain;
|
| 596 |
|
| 597 | void __init init_defrootdomain(void)
|
| 598 | {
|
| 599 | init_rootdomain(&def_root_domain);
|
| 600 |
|
| 601 | atomic_set(&def_root_domain.refcount, 1);
|
| 602 | }
|
| 603 |
|
| 604 | static struct root_domain *alloc_rootdomain(void)
|
| 605 | {
|
| 606 | struct root_domain *rd;
|
| 607 |
|
| 608 | rd = kzalloc_obj(*rd);
|
| 609 | if (!rd)
|
| 610 | return NULL;
|
| 611 |
|
| 612 | if (init_rootdomain(rd) != 0) {
|
| 613 | kfree(rd);
|
| 614 | return NULL;
|
| 615 | }
|
| 616 |
|
| 617 | return rd;
|
| 618 | }
|
| 619 |
|
| 620 | static void free_sched_groups(struct sched_group *sg, int free_sgc)
|
| 621 | {
|
| 622 | struct sched_group *tmp, *first;
|
| 623 |
|
| 624 | if (!sg)
|
| 625 | return;
|
| 626 |
|
| 627 | first = sg;
|
| 628 | do {
|
| 629 | tmp = sg->next;
|
| 630 |
|
| 631 | if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
|
| 632 | kfree(sg->sgc);
|
| 633 |
|
| 634 | if (atomic_dec_and_test(&sg->ref))
|
| 635 | kfree(sg);
|
| 636 | sg = tmp;
|
| 637 | } while (sg != first);
|
| 638 | }
|
| 639 |
|
| 640 | static void free_sched_domain_shared(struct sched_domain_shared *sds)
|
| 641 | {
|
| 642 | if (sds && atomic_dec_and_test(&sds->ref))
|
| 643 | kfree(sds);
|
| 644 | }
|
| 645 |
|
| 646 | static void destroy_sched_domain(struct sched_domain *sd)
|
| 647 | {
|
| 648 | |
| 649 | * A normal sched domain may have multiple group references, an
|
| 650 | * overlapping domain, having private groups, only one. Iterate,
|
| 651 | * dropping group/capacity references, freeing where none remain.
|
| 652 |
|
| 653 | free_sched_groups(sd->groups, 1);
|
| 654 | free_sched_domain_shared(sd->shared);
|
| 655 |
|
| 656 | #ifdef CONFIG_SCHED_CACHE
|
| 657 |
|
| 658 | kfree(sd->llc_counts);
|
| 659 | #endif
|
| 660 | kfree(sd);
|
| 661 | }
|
| 662 |
|
| 663 | static void destroy_sched_domains_rcu(struct rcu_head *rcu)
|
| 664 | {
|
| 665 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
|
| 666 |
|
| 667 | while (sd) {
|
| 668 | struct sched_domain *parent = sd->parent;
|
| 669 | destroy_sched_domain(sd);
|
| 670 | sd = parent;
|
| 671 | }
|
| 672 | }
|
| 673 |
|
| 674 | static void destroy_sched_domains(struct sched_domain *sd)
|
| 675 | {
|
| 676 | if (sd)
|
| 677 | call_rcu(&sd->rcu, destroy_sched_domains_rcu);
|
| 678 | }
|
| 679 |
|
| 680 | |
| 681 | * Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
|
| 682 | * (Last Level Cache Domain) for this allows us to avoid some pointer chasing
|
| 683 | * select_idle_sibling().
|
| 684 | *
|
| 685 | * Also keep a unique ID per domain (we use the first CPU number in the cpumask
|
| 686 | * of the domain), this allows us to quickly tell if two CPUs are in the same
|
| 687 | * cache domain, see cpus_share_cache().
|
| 688 |
|
| 689 | DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
|
| 690 | DEFINE_PER_CPU(int, sd_llc_size);
|
| 691 | DEFINE_PER_CPU(int, sd_llc_id) = -1;
|
| 692 | DEFINE_PER_CPU(int, sd_share_id);
|
| 693 | DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
|
| 694 | DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_balance_shared);
|
| 695 | DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
|
| 696 | DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
|
| 697 | DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
|
| 698 |
|
| 699 | DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
|
| 700 | DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
|
| 701 |
|
| 702 | static void update_top_cache_domain(int cpu)
|
| 703 | {
|
| 704 | struct sched_domain_shared *sds = NULL;
|
| 705 | struct sched_domain *sd;
|
| 706 | int id = cpu;
|
| 707 | int size = 1;
|
| 708 |
|
| 709 | sd = highest_flag_domain(cpu, SD_SHARE_LLC);
|
| 710 | if (sd) {
|
| 711 | id = cpumask_first(sched_domain_span(sd));
|
| 712 | size = cpumask_weight(sched_domain_span(sd));
|
| 713 |
|
| 714 |
|
| 715 | WARN_ON_ONCE(!sd->shared);
|
| 716 | sds = sd->shared;
|
| 717 | }
|
| 718 |
|
| 719 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
|
| 720 | per_cpu(sd_llc_size, cpu) = size;
|
| 721 | rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
|
| 722 |
|
| 723 | sd = lowest_flag_domain(cpu, SD_CLUSTER);
|
| 724 | if (sd)
|
| 725 | id = cpumask_first(sched_domain_span(sd));
|
| 726 |
|
| 727 | |
| 728 | * This assignment should be placed after the sd_llc_id as
|
| 729 | * we want this id equals to cluster id on cluster machines
|
| 730 | * but equals to LLC id on non-Cluster machines.
|
| 731 |
|
| 732 | per_cpu(sd_share_id, cpu) = id;
|
| 733 |
|
| 734 | sd = lowest_flag_domain(cpu, SD_NUMA);
|
| 735 | rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
|
| 736 |
|
| 737 | sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
|
| 738 | rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
|
| 739 |
|
| 740 | sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
|
| 741 | |
| 742 | * The shared object is attached to sd_asym_cpucapacity only when the
|
| 743 | * asym domain is non-overlapping (i.e., not built from SD_NUMA).
|
| 744 | * On overlapping (NUMA) asym domains we fall back to letting the
|
| 745 | * SD_SHARE_LLC path own the shared object, so sd->shared may be NULL
|
| 746 | * here.
|
| 747 |
|
| 748 | if (sd && sd->shared)
|
| 749 | sds = sd->shared;
|
| 750 |
|
| 751 | rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
|
| 752 | rcu_assign_pointer(per_cpu(sd_balance_shared, cpu), sds);
|
| 753 | }
|
| 754 |
|
| 755 | |
| 756 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
|
| 757 | * hold the hotplug lock.
|
| 758 |
|
| 759 | static void
|
| 760 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
|
| 761 | {
|
| 762 | struct rq *rq = cpu_rq(cpu);
|
| 763 | struct sched_domain *tmp;
|
| 764 |
|
| 765 |
|
| 766 | for (tmp = sd; tmp; ) {
|
| 767 | struct sched_domain *parent = tmp->parent;
|
| 768 | if (!parent)
|
| 769 | break;
|
| 770 |
|
| 771 | if (sd_parent_degenerate(tmp, parent)) {
|
| 772 | tmp->parent = parent->parent;
|
| 773 |
|
| 774 |
|
| 775 | if (parent->shared) {
|
| 776 | |
| 777 | * It is safe to free a sd->shared that
|
| 778 | * has not been published yet. If a
|
| 779 | * sd->shared was published, the refcount
|
| 780 | * will end up being non-zero and it will
|
| 781 | * not be freed here.
|
| 782 |
|
| 783 | free_sched_domain_shared(tmp->shared);
|
| 784 | tmp->shared = parent->shared;
|
| 785 | parent->shared = NULL;
|
| 786 | }
|
| 787 |
|
| 788 | if (parent->parent) {
|
| 789 | parent->parent->child = tmp;
|
| 790 | parent->parent->groups->flags = tmp->flags;
|
| 791 | }
|
| 792 |
|
| 793 | |
| 794 | * Transfer SD_PREFER_SIBLING down in case of a
|
| 795 | * degenerate parent; the spans match for this
|
| 796 | * so the property transfers.
|
| 797 |
|
| 798 | if (parent->flags & SD_PREFER_SIBLING)
|
| 799 | tmp->flags |= SD_PREFER_SIBLING;
|
| 800 | destroy_sched_domain(parent);
|
| 801 | } else
|
| 802 | tmp = tmp->parent;
|
| 803 | }
|
| 804 |
|
| 805 | if (sd && sd_degenerate(sd)) {
|
| 806 | tmp = sd;
|
| 807 | sd = sd->parent;
|
| 808 |
|
| 809 | if (sd) {
|
| 810 | struct sched_group *sg = sd->groups;
|
| 811 |
|
| 812 | #ifdef CONFIG_SCHED_CACHE
|
| 813 |
|
| 814 | sd->llc_counts = tmp->llc_counts;
|
| 815 | sd->llc_max = tmp->llc_max;
|
| 816 | sd->llc_bytes = tmp->llc_bytes;
|
| 817 |
|
| 818 | tmp->llc_counts = NULL;
|
| 819 | tmp->llc_max = 0;
|
| 820 | tmp->llc_bytes = 0;
|
| 821 | #endif
|
| 822 | |
| 823 | * sched groups hold the flags of the child sched
|
| 824 | * domain for convenience. Clear such flags since
|
| 825 | * the child is being destroyed.
|
| 826 |
|
| 827 | do {
|
| 828 | sg->flags = 0;
|
| 829 | } while (sg != sd->groups);
|
| 830 |
|
| 831 | sd->child = NULL;
|
| 832 | }
|
| 833 |
|
| 834 | destroy_sched_domain(tmp);
|
| 835 | }
|
| 836 |
|
| 837 | sched_domain_debug(sd, cpu);
|
| 838 |
|
| 839 | rq_attach_root(rq, rd);
|
| 840 | tmp = rq->sd;
|
| 841 | rcu_assign_pointer(rq->sd, sd);
|
| 842 | dirty_sched_domain_sysctl(cpu);
|
| 843 | destroy_sched_domains(tmp);
|
| 844 |
|
| 845 | update_top_cache_domain(cpu);
|
| 846 | }
|
| 847 |
|
| 848 | struct s_data {
|
| 849 | struct sched_domain_shared * __percpu *sds;
|
| 850 | struct sched_domain * __percpu *sd;
|
| 851 | struct root_domain *rd;
|
| 852 | };
|
| 853 |
|
| 854 | enum s_alloc {
|
| 855 | sa_rootdomain,
|
| 856 | sa_sd,
|
| 857 | sa_sd_shared,
|
| 858 | sa_sd_storage,
|
| 859 | sa_none,
|
| 860 | };
|
| 861 |
|
| 862 | #ifdef CONFIG_SCHED_CACHE
|
| 863 |
|
| 864 | DEFINE_STATIC_KEY_FALSE(sched_cache_present);
|
| 865 | |
| 866 | * Indicator of whether cache aware scheduling
|
| 867 | * is active, used by the scheduler.
|
| 868 |
|
| 869 | DEFINE_STATIC_KEY_FALSE(sched_cache_active);
|
| 870 |
|
| 871 | int sysctl_sched_cache_user = 1;
|
| 872 |
|
| 873 | |
| 874 | * Get the effective LLC size in bytes that @cpu's bottom sched_domain
|
| 875 | * can use. A CPU within a cpuset partition can only use a proportion
|
| 876 | * of the physical LLC, scaled by the ratio of the partition's span
|
| 877 | * weight to the hardware LLC sharing weight. @sd should be the
|
| 878 | * topmost domain with SD_SHARE_LLC.
|
| 879 | *
|
| 880 | * Returns 0 if cacheinfo is not yet populated. This happens during
|
| 881 | * early boot when build_sched_domains() runs before the generic
|
| 882 | * cacheinfo framework has been initialized (cacheinfo_cpu_online()
|
| 883 | * is a device_initcall cpuhp callback). In that case,
|
| 884 | * cacheinfo_cpu_online() will later call sched_update_llc_bytes()
|
| 885 | * to fill in the bottom domain's llc_bytes once the cache attributes
|
| 886 | * are available.
|
| 887 |
|
| 888 | static unsigned long get_effective_llc_bytes(int cpu,
|
| 889 | struct sched_domain *sd)
|
| 890 | {
|
| 891 | struct cacheinfo *ci;
|
| 892 | unsigned int hw_weight;
|
| 893 |
|
| 894 | ci = get_cpu_cacheinfo_llc(cpu);
|
| 895 | if (!ci)
|
| 896 | return 0;
|
| 897 |
|
| 898 | hw_weight = cpumask_weight(&ci->shared_cpu_map);
|
| 899 | if (!hw_weight)
|
| 900 | return 0;
|
| 901 |
|
| 902 | return div_u64((u64)ci->size * sd->span_weight, hw_weight);
|
| 903 | }
|
| 904 |
|
| 905 | static bool alloc_sd_llc(const struct cpumask *cpu_map,
|
| 906 | struct s_data *d)
|
| 907 | {
|
| 908 | struct sched_domain *sd, *top_llc, *parent;
|
| 909 | unsigned int *p;
|
| 910 | int i;
|
| 911 |
|
| 912 | for_each_cpu(i, cpu_map) {
|
| 913 | sd = *per_cpu_ptr(d->sd, i);
|
| 914 | if (!sd)
|
| 915 | goto err;
|
| 916 |
|
| 917 | p = kcalloc_node(max_lid + 1, sizeof(unsigned int),
|
| 918 | GFP_KERNEL, cpu_to_node(i));
|
| 919 | if (!p)
|
| 920 | goto err;
|
| 921 |
|
| 922 | top_llc = sd;
|
| 923 | |
| 924 | * Find the topmost SD_SHARE_LLC domain.
|
| 925 | * Not yet attached to the CPU, so per_cpu(sd_llc, i)
|
| 926 | * can not be used.
|
| 927 |
|
| 928 | while ((parent = rcu_dereference_protected(top_llc->parent, true)) &&
|
| 929 | (parent->flags & SD_SHARE_LLC))
|
| 930 | top_llc = parent;
|
| 931 |
|
| 932 | if (top_llc->flags & SD_SHARE_LLC) {
|
| 933 | sd->llc_max = max_lid + 1;
|
| 934 | sd->llc_counts = p;
|
| 935 | sd->llc_bytes = get_effective_llc_bytes(i, top_llc);
|
| 936 | } else {
|
| 937 |
|
| 938 | kfree(p);
|
| 939 | }
|
| 940 | }
|
| 941 |
|
| 942 | return true;
|
| 943 | err:
|
| 944 | for_each_cpu(i, cpu_map) {
|
| 945 | sd = *per_cpu_ptr(d->sd, i);
|
| 946 | if (sd) {
|
| 947 | kfree(sd->llc_counts);
|
| 948 | sd->llc_counts = NULL;
|
| 949 | sd->llc_max = 0;
|
| 950 | sd->llc_bytes = 0;
|
| 951 | }
|
| 952 | }
|
| 953 |
|
| 954 | return false;
|
| 955 | }
|
| 956 |
|
| 957 | |
| 958 | * Enable/disable cache aware scheduling according to
|
| 959 | * user input and the presence of hardware support.
|
| 960 |
|
| 961 | static void _sched_cache_active_set(void)
|
| 962 | {
|
| 963 | lockdep_assert_cpus_held();
|
| 964 | lockdep_assert_held(&sched_domains_mutex);
|
| 965 |
|
| 966 |
|
| 967 | if (!static_branch_likely(&sched_cache_present)) {
|
| 968 | static_branch_disable_cpuslocked(&sched_cache_active);
|
| 969 | if (sched_debug())
|
| 970 | pr_info("%s: cache aware scheduling not supported on this platform\n", __func__);
|
| 971 | return;
|
| 972 | }
|
| 973 |
|
| 974 | |
| 975 | * user wants it or not ?
|
| 976 | * TBD: read before writing the static key.
|
| 977 | * It is not in the critical path, leave as-is
|
| 978 | * for now.
|
| 979 |
|
| 980 | if (sysctl_sched_cache_user) {
|
| 981 | static_branch_enable_cpuslocked(&sched_cache_active);
|
| 982 | if (sched_debug())
|
| 983 | pr_info("%s: enabling cache aware scheduling\n", __func__);
|
| 984 | } else {
|
| 985 | static_branch_disable_cpuslocked(&sched_cache_active);
|
| 986 | if (sched_debug())
|
| 987 | pr_info("%s: disabling cache aware scheduling\n", __func__);
|
| 988 | }
|
| 989 | }
|
| 990 |
|
| 991 |
|
| 992 | void sched_cache_active_set(void)
|
| 993 | {
|
| 994 | cpus_read_lock();
|
| 995 | sched_domains_mutex_lock();
|
| 996 | _sched_cache_active_set();
|
| 997 | sched_domains_mutex_unlock();
|
| 998 | cpus_read_unlock();
|
| 999 | }
|
| 1000 |
|
| 1001 | |
| 1002 | * Update the bottom sched_domain's llc_bytes for @cpus sharing a physical
|
| 1003 | * LLC. Called from cacheinfo_cpu_online() or
|
| 1004 | * cacheinfo_cpu_pre_down() with cpu hotplug lock held.
|
| 1005 | *
|
| 1006 | * Note: get_effective_llc_bytes() returns 0 on PowerPC.
|
| 1007 | * thus cache aware scheduling is disabled on PowerPC for
|
| 1008 | * now. PowerPC does not use the generic cacheinfo framework --
|
| 1009 | * it has its own cacheinfo with a separate struct cache hierarchy
|
| 1010 | * and does not populates the per-CPU struct cpu_cacheinfo array
|
| 1011 | * that get_cpu_cacheinfo_llc() reads.
|
| 1012 |
|
| 1013 | void sched_update_llc_bytes(const struct cpumask *cpus)
|
| 1014 | {
|
| 1015 | struct sched_domain *sd, *sdp;
|
| 1016 | unsigned int i;
|
| 1017 |
|
| 1018 | sched_domains_mutex_lock();
|
| 1019 |
|
| 1020 | |
| 1021 | * ci->shared_cpu_map is built incrementally as CPUs come
|
| 1022 | * online, so the first CPU in an LLC initially sees
|
| 1023 | * hw_weight == 1 and computes an inflated llc_bytes in
|
| 1024 | * get_effective_llc_bytes(). Re-evaluating every LLC
|
| 1025 | * sibling on each online event corrects this once the full
|
| 1026 | * shared_cpu_map is known.
|
| 1027 | *
|
| 1028 | * The departing CPU's domains have already been detached when
|
| 1029 | * cacheinfo removes it. Use the surviving cache siblings instead.
|
| 1030 | * They may belong to different cpuset partitions, so use each CPU's
|
| 1031 | * own LLC domain to scale its share of the physical cache.
|
| 1032 |
|
| 1033 | for_each_cpu(i, cpus) {
|
| 1034 | sdp = rcu_dereference_sched_domain(per_cpu(sd_llc, i));
|
| 1035 | if (!sdp)
|
| 1036 | continue;
|
| 1037 |
|
| 1038 | sd = rcu_dereference_sched_domain(cpu_rq(i)->sd);
|
| 1039 | if (sd)
|
| 1040 | sd->llc_bytes = get_effective_llc_bytes(i, sdp);
|
| 1041 | }
|
| 1042 |
|
| 1043 | sched_domains_mutex_unlock();
|
| 1044 | }
|
| 1045 |
|
| 1046 | static void sched_cache_set(bool has_multi_llcs)
|
| 1047 | {
|
| 1048 | |
| 1049 | * TBD: check before writing to it. sched domain rebuild
|
| 1050 | * is not in the critical path, leave as-is for now.
|
| 1051 |
|
| 1052 | if (has_multi_llcs)
|
| 1053 | static_branch_enable_cpuslocked(&sched_cache_present);
|
| 1054 | else
|
| 1055 | static_branch_disable_cpuslocked(&sched_cache_present);
|
| 1056 |
|
| 1057 | _sched_cache_active_set();
|
| 1058 | }
|
| 1059 | #else
|
| 1060 | static bool alloc_sd_llc(const struct cpumask *cpu_map,
|
| 1061 | struct s_data *d)
|
| 1062 | {
|
| 1063 | return false;
|
| 1064 | }
|
| 1065 | static inline void sched_cache_set(bool has_multi_llcs) { }
|
| 1066 | #endif
|
| 1067 |
|
| 1068 | |
| 1069 | * Return true if @sd belongs to an LLC group whose enclosing
|
| 1070 | * partition spans more than one LLC. @sd must be the topmost
|
| 1071 | * SD_SHARE_LLC domain.
|
| 1072 | *
|
| 1073 | * Any duplicated parent domains with the same span as @sd are
|
| 1074 | * skipped: before cpu_attach_domain() degeneration these still
|
| 1075 | * exist, after degeneration the loop is a no-op. This makes the
|
| 1076 | * helper usable both during sched domain build and against an
|
| 1077 | * already-attached domain tree.
|
| 1078 | *
|
| 1079 | * Note: For systems with a single LLC per node, cache-aware
|
| 1080 | * scheduling is still enabled when multiple nodes exist.
|
| 1081 | * However, NUMA balancing decisions take precedence over
|
| 1082 | * cache-aware scheduling. Conversely, if there is only one
|
| 1083 | * LLC per partition, cache-aware scheduling should be disabled.
|
| 1084 |
|
| 1085 | static bool sd_in_multi_llcs(struct sched_domain *sd)
|
| 1086 | {
|
| 1087 | struct sched_domain *sdp = sd->parent;
|
| 1088 |
|
| 1089 |
|
| 1090 | if (sd->span_weight == 1)
|
| 1091 | return false;
|
| 1092 |
|
| 1093 | while (sdp && sdp->span_weight == sd->span_weight)
|
| 1094 | sdp = sdp->parent;
|
| 1095 |
|
| 1096 | return !!sdp;
|
| 1097 | }
|
| 1098 |
|
| 1099 | |
| 1100 | * Return the canonical balance CPU for this group, this is the first CPU
|
| 1101 | * of this group that's also in the balance mask.
|
| 1102 | *
|
| 1103 | * The balance mask are all those CPUs that could actually end up at this
|
| 1104 | * group. See build_balance_mask().
|
| 1105 | *
|
| 1106 | * Also see should_we_balance().
|
| 1107 |
|
| 1108 | int group_balance_cpu(struct sched_group *sg)
|
| 1109 | {
|
| 1110 | return cpumask_first(group_balance_mask(sg));
|
| 1111 | }
|
| 1112 |
|
| 1113 |
|
| 1114 | |
| 1115 | * NUMA topology (first read the regular topology blurb below)
|
| 1116 | *
|
| 1117 | * Given a node-distance table, for example:
|
| 1118 | *
|
| 1119 | * node 0 1 2 3
|
| 1120 | * 0: 10 20 30 20
|
| 1121 | * 1: 20 10 20 30
|
| 1122 | * 2: 30 20 10 20
|
| 1123 | * 3: 20 30 20 10
|
| 1124 | *
|
| 1125 | * which represents a 4 node ring topology like:
|
| 1126 | *
|
| 1127 | * 0 ----- 1
|
| 1128 | * | |
|
| 1129 | * | |
|
| 1130 | * | |
|
| 1131 | * 3 ----- 2
|
| 1132 | *
|
| 1133 | * We want to construct domains and groups to represent this. The way we go
|
| 1134 | * about doing this is to build the domains on 'hops'. For each NUMA level we
|
| 1135 | * construct the mask of all nodes reachable in @level hops.
|
| 1136 | *
|
| 1137 | * For the above NUMA topology that gives 3 levels:
|
| 1138 | *
|
| 1139 | * NUMA-2 0-3 0-3 0-3 0-3
|
| 1140 | * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
|
| 1141 | *
|
| 1142 | * NUMA-1 0-1,3 0-2 1-3 0,2-3
|
| 1143 | * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
|
| 1144 | *
|
| 1145 | * NUMA-0 0 1 2 3
|
| 1146 | *
|
| 1147 | *
|
| 1148 | * As can be seen; things don't nicely line up as with the regular topology.
|
| 1149 | * When we iterate a domain in child domain chunks some nodes can be
|
| 1150 | * represented multiple times -- hence the "overlap" naming for this part of
|
| 1151 | * the topology.
|
| 1152 | *
|
| 1153 | * In order to minimize this overlap, we only build enough groups to cover the
|
| 1154 | * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
|
| 1155 | *
|
| 1156 | * Because:
|
| 1157 | *
|
| 1158 | * - the first group of each domain is its child domain; this
|
| 1159 | * gets us the first 0-1,3
|
| 1160 | * - the only uncovered node is 2, who's child domain is 1-3.
|
| 1161 | *
|
| 1162 | * However, because of the overlap, computing a unique CPU for each group is
|
| 1163 | * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
|
| 1164 | * groups include the CPUs of Node-0, while those CPUs would not in fact ever
|
| 1165 | * end up at those groups (they would end up in group: 0-1,3).
|
| 1166 | *
|
| 1167 | * To correct this we have to introduce the group balance mask. This mask
|
| 1168 | * will contain those CPUs in the group that can reach this group given the
|
| 1169 | * (child) domain tree.
|
| 1170 | *
|
| 1171 | * With this we can once again compute balance_cpu and sched_group_capacity
|
| 1172 | * relations.
|
| 1173 | *
|
| 1174 | * XXX include words on how balance_cpu is unique and therefore can be
|
| 1175 | * used for sched_group_capacity links.
|
| 1176 | *
|
| 1177 | *
|
| 1178 | * Another 'interesting' topology is:
|
| 1179 | *
|
| 1180 | * node 0 1 2 3
|
| 1181 | * 0: 10 20 20 30
|
| 1182 | * 1: 20 10 20 20
|
| 1183 | * 2: 20 20 10 20
|
| 1184 | * 3: 30 20 20 10
|
| 1185 | *
|
| 1186 | * Which looks a little like:
|
| 1187 | *
|
| 1188 | * 0 ----- 1
|
| 1189 | * | / |
|
| 1190 | * | / |
|
| 1191 | * | / |
|
| 1192 | * 2 ----- 3
|
| 1193 | *
|
| 1194 | * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
|
| 1195 | * are not.
|
| 1196 | *
|
| 1197 | * This leads to a few particularly weird cases where the sched_domain's are
|
| 1198 | * not of the same number for each CPU. Consider:
|
| 1199 | *
|
| 1200 | * NUMA-2 0-3 0-3
|
| 1201 | * groups: {0-2},{1-3} {1-3},{0-2}
|
| 1202 | *
|
| 1203 | * NUMA-1 0-2 0-3 0-3 1-3
|
| 1204 | *
|
| 1205 | * NUMA-0 0 1 2 3
|
| 1206 | *
|
| 1207 |
|
| 1208 |
|
| 1209 |
|
| 1210 | |
| 1211 | * Build the balance mask; it contains only those CPUs that can arrive at this
|
| 1212 | * group and should be considered to continue balancing.
|
| 1213 | *
|
| 1214 | * We do this during the group creation pass, therefore the group information
|
| 1215 | * isn't complete yet, however since each group represents a (child) domain we
|
| 1216 | * can fully construct this using the sched_domain bits (which are already
|
| 1217 | * complete).
|
| 1218 |
|
| 1219 | static void
|
| 1220 | build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
|
| 1221 | {
|
| 1222 | const struct cpumask *sg_span = sched_group_span(sg);
|
| 1223 | struct sd_data *sdd = sd->private;
|
| 1224 | struct sched_domain *sibling;
|
| 1225 | int i;
|
| 1226 |
|
| 1227 | cpumask_clear(mask);
|
| 1228 |
|
| 1229 | for_each_cpu(i, sg_span) {
|
| 1230 | sibling = *per_cpu_ptr(sdd->sd, i);
|
| 1231 |
|
| 1232 | |
| 1233 | * Can happen in the asymmetric case, where these siblings are
|
| 1234 | * unused. The mask will not be empty because those CPUs that
|
| 1235 | * do have the top domain _should_ span the domain.
|
| 1236 |
|
| 1237 | if (!sibling->child)
|
| 1238 | continue;
|
| 1239 |
|
| 1240 |
|
| 1241 | if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
|
| 1242 | continue;
|
| 1243 |
|
| 1244 | cpumask_set_cpu(i, mask);
|
| 1245 | }
|
| 1246 |
|
| 1247 |
|
| 1248 | WARN_ON_ONCE(cpumask_empty(mask));
|
| 1249 | }
|
| 1250 |
|
| 1251 | |
| 1252 | * XXX: This creates per-node group entries; since the load-balancer will
|
| 1253 | * immediately access remote memory to construct this group's load-balance
|
| 1254 | * statistics having the groups node local is of dubious benefit.
|
| 1255 |
|
| 1256 | static struct sched_group *
|
| 1257 | build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
|
| 1258 | {
|
| 1259 | struct sched_group *sg;
|
| 1260 | struct cpumask *sg_span;
|
| 1261 |
|
| 1262 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
|
| 1263 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1264 |
|
| 1265 | if (!sg)
|
| 1266 | return NULL;
|
| 1267 |
|
| 1268 | sg_span = sched_group_span(sg);
|
| 1269 | if (sd->child) {
|
| 1270 | cpumask_copy(sg_span, sched_domain_span(sd->child));
|
| 1271 | sg->flags = sd->child->flags;
|
| 1272 | } else {
|
| 1273 | cpumask_copy(sg_span, sched_domain_span(sd));
|
| 1274 | }
|
| 1275 |
|
| 1276 | atomic_inc(&sg->ref);
|
| 1277 | return sg;
|
| 1278 | }
|
| 1279 |
|
| 1280 | static void init_overlap_sched_group(struct sched_domain *sd,
|
| 1281 | struct sched_group *sg)
|
| 1282 | {
|
| 1283 | struct cpumask *mask = sched_domains_tmpmask2;
|
| 1284 | struct sd_data *sdd = sd->private;
|
| 1285 | struct cpumask *sg_span;
|
| 1286 | int cpu;
|
| 1287 |
|
| 1288 | build_balance_mask(sd, sg, mask);
|
| 1289 | cpu = cpumask_first(mask);
|
| 1290 |
|
| 1291 | sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
|
| 1292 | if (atomic_inc_return(&sg->sgc->ref) == 1)
|
| 1293 | cpumask_copy(group_balance_mask(sg), mask);
|
| 1294 | else
|
| 1295 | WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
|
| 1296 |
|
| 1297 | |
| 1298 | * Initialize sgc->capacity such that even if we mess up the
|
| 1299 | * domains and no possible iteration will get us here, we won't
|
| 1300 | * die on a /0 trap.
|
| 1301 |
|
| 1302 | sg_span = sched_group_span(sg);
|
| 1303 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
|
| 1304 | sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
|
| 1305 | sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
|
| 1306 | }
|
| 1307 |
|
| 1308 | static struct sched_domain *
|
| 1309 | find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
|
| 1310 | {
|
| 1311 | |
| 1312 | * The proper descendant would be the one whose child won't span out
|
| 1313 | * of sd
|
| 1314 |
|
| 1315 | while (sibling->child &&
|
| 1316 | !cpumask_subset(sched_domain_span(sibling->child),
|
| 1317 | sched_domain_span(sd)))
|
| 1318 | sibling = sibling->child;
|
| 1319 |
|
| 1320 | |
| 1321 | * As we are referencing sgc across different topology level, we need
|
| 1322 | * to go down to skip those sched_domains which don't contribute to
|
| 1323 | * scheduling because they will be degenerated in cpu_attach_domain
|
| 1324 |
|
| 1325 | while (sibling->child &&
|
| 1326 | cpumask_equal(sched_domain_span(sibling->child),
|
| 1327 | sched_domain_span(sibling)))
|
| 1328 | sibling = sibling->child;
|
| 1329 |
|
| 1330 | return sibling;
|
| 1331 | }
|
| 1332 |
|
| 1333 | static int
|
| 1334 | build_overlap_sched_groups(struct sched_domain *sd, int cpu)
|
| 1335 | {
|
| 1336 | struct sched_group *first = NULL, *last = NULL, *sg;
|
| 1337 | const struct cpumask *span = sched_domain_span(sd);
|
| 1338 | struct cpumask *covered = sched_domains_tmpmask;
|
| 1339 | struct sd_data *sdd = sd->private;
|
| 1340 | struct sched_domain *sibling;
|
| 1341 | int i;
|
| 1342 |
|
| 1343 | cpumask_clear(covered);
|
| 1344 |
|
| 1345 | for_each_cpu_wrap(i, span, cpu) {
|
| 1346 | struct cpumask *sg_span;
|
| 1347 |
|
| 1348 | if (cpumask_test_cpu(i, covered))
|
| 1349 | continue;
|
| 1350 |
|
| 1351 | sibling = *per_cpu_ptr(sdd->sd, i);
|
| 1352 |
|
| 1353 | |
| 1354 | * Asymmetric node setups can result in situations where the
|
| 1355 | * domain tree is of unequal depth, make sure to skip domains
|
| 1356 | * that already cover the entire range.
|
| 1357 | *
|
| 1358 | * In that case build_sched_domains() will have terminated the
|
| 1359 | * iteration early and our sibling sd spans will be empty.
|
| 1360 | * Domains should always include the CPU they're built on, so
|
| 1361 | * check that.
|
| 1362 |
|
| 1363 | if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
|
| 1364 | continue;
|
| 1365 |
|
| 1366 | |
| 1367 | * Usually we build sched_group by sibling's child sched_domain
|
| 1368 | * But for machines whose NUMA diameter are 3 or above, we move
|
| 1369 | * to build sched_group by sibling's proper descendant's child
|
| 1370 | * domain because sibling's child sched_domain will span out of
|
| 1371 | * the sched_domain being built as below.
|
| 1372 | *
|
| 1373 | * Smallest diameter=3 topology is:
|
| 1374 | *
|
| 1375 | * node 0 1 2 3
|
| 1376 | * 0: 10 20 30 40
|
| 1377 | * 1: 20 10 20 30
|
| 1378 | * 2: 30 20 10 20
|
| 1379 | * 3: 40 30 20 10
|
| 1380 | *
|
| 1381 | * 0 --- 1 --- 2 --- 3
|
| 1382 | *
|
| 1383 | * NUMA-3 0-3 N/A N/A 0-3
|
| 1384 | * groups: {0-2},{1-3} {1-3},{0-2}
|
| 1385 | *
|
| 1386 | * NUMA-2 0-2 0-3 0-3 1-3
|
| 1387 | * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
|
| 1388 | *
|
| 1389 | * NUMA-1 0-1 0-2 1-3 2-3
|
| 1390 | * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
|
| 1391 | *
|
| 1392 | * NUMA-0 0 1 2 3
|
| 1393 | *
|
| 1394 | * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
|
| 1395 | * group span isn't a subset of the domain span.
|
| 1396 |
|
| 1397 | if (sibling->child &&
|
| 1398 | !cpumask_subset(sched_domain_span(sibling->child), span))
|
| 1399 | sibling = find_descended_sibling(sd, sibling);
|
| 1400 |
|
| 1401 | sg = build_group_from_child_sched_domain(sibling, cpu);
|
| 1402 | if (!sg)
|
| 1403 | goto fail;
|
| 1404 |
|
| 1405 | sg_span = sched_group_span(sg);
|
| 1406 | cpumask_or(covered, covered, sg_span);
|
| 1407 |
|
| 1408 | init_overlap_sched_group(sibling, sg);
|
| 1409 |
|
| 1410 | if (!first)
|
| 1411 | first = sg;
|
| 1412 | if (last)
|
| 1413 | last->next = sg;
|
| 1414 | last = sg;
|
| 1415 | last->next = first;
|
| 1416 | }
|
| 1417 | sd->groups = first;
|
| 1418 |
|
| 1419 | return 0;
|
| 1420 |
|
| 1421 | fail:
|
| 1422 | free_sched_groups(first, 0);
|
| 1423 |
|
| 1424 | return -ENOMEM;
|
| 1425 | }
|
| 1426 |
|
| 1427 |
|
| 1428 | |
| 1429 | * Package topology (also see the load-balance blurb in fair.c)
|
| 1430 | *
|
| 1431 | * The scheduler builds a tree structure to represent a number of important
|
| 1432 | * topology features. By default (default_topology[]) these include:
|
| 1433 | *
|
| 1434 | * - Simultaneous multithreading (SMT)
|
| 1435 | * - Multi-Core Cache (MC)
|
| 1436 | * - Package (PKG)
|
| 1437 | *
|
| 1438 | * Where the last one more or less denotes everything up to a NUMA node.
|
| 1439 | *
|
| 1440 | * The tree consists of 3 primary data structures:
|
| 1441 | *
|
| 1442 | * sched_domain -> sched_group -> sched_group_capacity
|
| 1443 | * ^ ^ ^ ^
|
| 1444 | * `-' `-'
|
| 1445 | *
|
| 1446 | * The sched_domains are per-CPU and have a two way link (parent & child) and
|
| 1447 | * denote the ever growing mask of CPUs belonging to that level of topology.
|
| 1448 | *
|
| 1449 | * Each sched_domain has a circular (double) linked list of sched_group's, each
|
| 1450 | * denoting the domains of the level below (or individual CPUs in case of the
|
| 1451 | * first domain level). The sched_group linked by a sched_domain includes the
|
| 1452 | * CPU of that sched_domain [*].
|
| 1453 | *
|
| 1454 | * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
|
| 1455 | *
|
| 1456 | * CPU 0 1 2 3 4 5 6 7
|
| 1457 | *
|
| 1458 | * PKG [ ]
|
| 1459 | * MC [ ] [ ]
|
| 1460 | * SMT [ ] [ ] [ ] [ ]
|
| 1461 | *
|
| 1462 | * - or -
|
| 1463 | *
|
| 1464 | * PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
|
| 1465 | * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
|
| 1466 | * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
|
| 1467 | *
|
| 1468 | * CPU 0 1 2 3 4 5 6 7
|
| 1469 | *
|
| 1470 | * One way to think about it is: sched_domain moves you up and down among these
|
| 1471 | * topology levels, while sched_group moves you sideways through it, at child
|
| 1472 | * domain granularity.
|
| 1473 | *
|
| 1474 | * sched_group_capacity ensures each unique sched_group has shared storage.
|
| 1475 | *
|
| 1476 | * There are two related construction problems, both require a CPU that
|
| 1477 | * uniquely identify each group (for a given domain):
|
| 1478 | *
|
| 1479 | * - The first is the balance_cpu (see should_we_balance() and the
|
| 1480 | * load-balance blurb in fair.c); for each group we only want 1 CPU to
|
| 1481 | * continue balancing at a higher domain.
|
| 1482 | *
|
| 1483 | * - The second is the sched_group_capacity; we want all identical groups
|
| 1484 | * to share a single sched_group_capacity.
|
| 1485 | *
|
| 1486 | * Since these topologies are exclusive by construction. That is, its
|
| 1487 | * impossible for an SMT thread to belong to multiple cores, and cores to
|
| 1488 | * be part of multiple caches. There is a very clear and unique location
|
| 1489 | * for each CPU in the hierarchy.
|
| 1490 | *
|
| 1491 | * Therefore computing a unique CPU for each group is trivial (the iteration
|
| 1492 | * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
|
| 1493 | * group), we can simply pick the first CPU in each group.
|
| 1494 | *
|
| 1495 | *
|
| 1496 | * [*] in other words, the first group of each domain is its child domain.
|
| 1497 |
|
| 1498 |
|
| 1499 | static struct sched_group *get_group(int cpu, struct sd_data *sdd)
|
| 1500 | {
|
| 1501 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
|
| 1502 | struct sched_domain *child = sd->child;
|
| 1503 | struct sched_group *sg;
|
| 1504 | bool already_visited;
|
| 1505 |
|
| 1506 | if (child)
|
| 1507 | cpu = cpumask_first(sched_domain_span(child));
|
| 1508 |
|
| 1509 | sg = *per_cpu_ptr(sdd->sg, cpu);
|
| 1510 | sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
|
| 1511 |
|
| 1512 |
|
| 1513 | already_visited = atomic_inc_return(&sg->ref) > 1;
|
| 1514 |
|
| 1515 | WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
|
| 1516 |
|
| 1517 |
|
| 1518 | if (already_visited)
|
| 1519 | return sg;
|
| 1520 |
|
| 1521 | if (child) {
|
| 1522 | cpumask_copy(sched_group_span(sg), sched_domain_span(child));
|
| 1523 | cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
|
| 1524 | sg->flags = child->flags;
|
| 1525 | } else {
|
| 1526 | cpumask_set_cpu(cpu, sched_group_span(sg));
|
| 1527 | cpumask_set_cpu(cpu, group_balance_mask(sg));
|
| 1528 | }
|
| 1529 |
|
| 1530 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
|
| 1531 | sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
|
| 1532 | sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
|
| 1533 |
|
| 1534 | return sg;
|
| 1535 | }
|
| 1536 |
|
| 1537 | |
| 1538 | * build_sched_groups will build a circular linked list of the groups
|
| 1539 | * covered by the given span, will set each group's ->cpumask correctly,
|
| 1540 | * and will initialize their ->sgc.
|
| 1541 | *
|
| 1542 | * Assumes the sched_domain tree is fully constructed
|
| 1543 |
|
| 1544 | static int
|
| 1545 | build_sched_groups(struct sched_domain *sd, int cpu)
|
| 1546 | {
|
| 1547 | struct sched_group *first = NULL, *last = NULL;
|
| 1548 | struct sd_data *sdd = sd->private;
|
| 1549 | const struct cpumask *span = sched_domain_span(sd);
|
| 1550 | struct cpumask *covered;
|
| 1551 | int i;
|
| 1552 |
|
| 1553 | lockdep_assert_held(&sched_domains_mutex);
|
| 1554 | covered = sched_domains_tmpmask;
|
| 1555 |
|
| 1556 | cpumask_clear(covered);
|
| 1557 |
|
| 1558 | for_each_cpu_wrap(i, span, cpu) {
|
| 1559 | struct sched_group *sg;
|
| 1560 |
|
| 1561 | if (cpumask_test_cpu(i, covered))
|
| 1562 | continue;
|
| 1563 |
|
| 1564 | sg = get_group(i, sdd);
|
| 1565 |
|
| 1566 | cpumask_or(covered, covered, sched_group_span(sg));
|
| 1567 |
|
| 1568 | if (!first)
|
| 1569 | first = sg;
|
| 1570 | if (last)
|
| 1571 | last->next = sg;
|
| 1572 | last = sg;
|
| 1573 | }
|
| 1574 | last->next = first;
|
| 1575 | sd->groups = first;
|
| 1576 |
|
| 1577 | return 0;
|
| 1578 | }
|
| 1579 |
|
| 1580 | |
| 1581 | * Initialize sched groups cpu_capacity.
|
| 1582 | *
|
| 1583 | * cpu_capacity indicates the capacity of sched group, which is used while
|
| 1584 | * distributing the load between different sched groups in a sched domain.
|
| 1585 | * Typically cpu_capacity for all the groups in a sched domain will be same
|
| 1586 | * unless there are asymmetries in the topology. If there are asymmetries,
|
| 1587 | * group having more cpu_capacity will pickup more load compared to the
|
| 1588 | * group having less cpu_capacity.
|
| 1589 |
|
| 1590 | static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
|
| 1591 | {
|
| 1592 | struct sched_group *sg = sd->groups;
|
| 1593 | struct cpumask *mask = sched_domains_tmpmask2;
|
| 1594 |
|
| 1595 | WARN_ON(!sg);
|
| 1596 |
|
| 1597 | do {
|
| 1598 | int cpu, cores = 0, max_cpu = -1;
|
| 1599 |
|
| 1600 | sg->group_weight = cpumask_weight(sched_group_span(sg));
|
| 1601 |
|
| 1602 | cpumask_copy(mask, sched_group_span(sg));
|
| 1603 | for_each_cpu(cpu, mask) {
|
| 1604 | cores++;
|
| 1605 | cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
|
| 1606 | }
|
| 1607 | sg->cores = cores;
|
| 1608 |
|
| 1609 | if (!(sd->flags & SD_ASYM_PACKING))
|
| 1610 | goto next;
|
| 1611 |
|
| 1612 | for_each_cpu(cpu, sched_group_span(sg)) {
|
| 1613 | if (max_cpu < 0)
|
| 1614 | max_cpu = cpu;
|
| 1615 | else if (sched_asym_prefer(cpu, max_cpu))
|
| 1616 | max_cpu = cpu;
|
| 1617 | }
|
| 1618 | sg->asym_prefer_cpu = max_cpu;
|
| 1619 |
|
| 1620 | next:
|
| 1621 | sg = sg->next;
|
| 1622 | } while (sg != sd->groups);
|
| 1623 |
|
| 1624 | if (cpu != group_balance_cpu(sg))
|
| 1625 | return;
|
| 1626 |
|
| 1627 | update_group_capacity(sd, cpu);
|
| 1628 | }
|
| 1629 |
|
| 1630 |
|
| 1631 | void sched_update_asym_prefer_cpu(int cpu, int old_prio, int new_prio)
|
| 1632 | {
|
| 1633 | int asym_prefer_cpu = cpu;
|
| 1634 | struct sched_domain *sd;
|
| 1635 |
|
| 1636 | guard(rcu)();
|
| 1637 |
|
| 1638 | for_each_domain(cpu, sd) {
|
| 1639 | struct sched_group *sg;
|
| 1640 | int group_cpu;
|
| 1641 |
|
| 1642 | if (!(sd->flags & SD_ASYM_PACKING))
|
| 1643 | continue;
|
| 1644 |
|
| 1645 | |
| 1646 | * Groups of overlapping domain are replicated per NUMA
|
| 1647 | * node and will require updating "asym_prefer_cpu" on
|
| 1648 | * each local copy.
|
| 1649 | *
|
| 1650 | * If you are hitting this warning, consider moving
|
| 1651 | * "sg->asym_prefer_cpu" to "sg->sgc->asym_prefer_cpu"
|
| 1652 | * which is shared by all the overlapping groups.
|
| 1653 |
|
| 1654 | WARN_ON_ONCE(sd->flags & SD_NUMA);
|
| 1655 |
|
| 1656 | sg = sd->groups;
|
| 1657 | if (cpu != sg->asym_prefer_cpu) {
|
| 1658 | |
| 1659 | * Since the parent is a superset of the current group,
|
| 1660 | * if the cpu is not the "asym_prefer_cpu" at the
|
| 1661 | * current level, it cannot be the preferred CPU at a
|
| 1662 | * higher levels either.
|
| 1663 |
|
| 1664 | if (!sched_asym_prefer(cpu, sg->asym_prefer_cpu))
|
| 1665 | return;
|
| 1666 |
|
| 1667 | WRITE_ONCE(sg->asym_prefer_cpu, cpu);
|
| 1668 | continue;
|
| 1669 | }
|
| 1670 |
|
| 1671 |
|
| 1672 | if (new_prio >= old_prio)
|
| 1673 | continue;
|
| 1674 |
|
| 1675 | for_each_cpu(group_cpu, sched_group_span(sg)) {
|
| 1676 | if (sched_asym_prefer(group_cpu, asym_prefer_cpu))
|
| 1677 | asym_prefer_cpu = group_cpu;
|
| 1678 | }
|
| 1679 |
|
| 1680 | WRITE_ONCE(sg->asym_prefer_cpu, asym_prefer_cpu);
|
| 1681 | }
|
| 1682 | }
|
| 1683 |
|
| 1684 | |
| 1685 | * Set of available CPUs grouped by their corresponding capacities
|
| 1686 | * Each list entry contains a CPU mask reflecting CPUs that share the same
|
| 1687 | * capacity.
|
| 1688 | * The lifespan of data is unlimited.
|
| 1689 |
|
| 1690 | LIST_HEAD(asym_cap_list);
|
| 1691 |
|
| 1692 | |
| 1693 | * Verify whether there is any CPU capacity asymmetry in a given sched domain.
|
| 1694 | * Provides sd_flags reflecting the asymmetry scope.
|
| 1695 |
|
| 1696 | static inline int
|
| 1697 | asym_cpu_capacity_classify(const struct cpumask *sd_span,
|
| 1698 | const struct cpumask *cpu_map)
|
| 1699 | {
|
| 1700 | struct asym_cap_data *entry;
|
| 1701 | int count = 0, miss = 0;
|
| 1702 |
|
| 1703 | |
| 1704 | * Count how many unique CPU capacities this domain spans across
|
| 1705 | * (compare sched_domain CPUs mask with ones representing available
|
| 1706 | * CPUs capacities). Take into account CPUs that might be offline:
|
| 1707 | * skip those.
|
| 1708 |
|
| 1709 | list_for_each_entry(entry, &asym_cap_list, link) {
|
| 1710 | if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
|
| 1711 | ++count;
|
| 1712 | else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
|
| 1713 | ++miss;
|
| 1714 | }
|
| 1715 |
|
| 1716 | WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
|
| 1717 |
|
| 1718 |
|
| 1719 | if (count < 2)
|
| 1720 | return 0;
|
| 1721 |
|
| 1722 | if (miss)
|
| 1723 | return SD_ASYM_CPUCAPACITY;
|
| 1724 |
|
| 1725 |
|
| 1726 | return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
|
| 1727 |
|
| 1728 | }
|
| 1729 |
|
| 1730 | static void free_asym_cap_entry(struct rcu_head *head)
|
| 1731 | {
|
| 1732 | struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
|
| 1733 | kfree(entry);
|
| 1734 | }
|
| 1735 |
|
| 1736 | static inline void asym_cpu_capacity_update_data(int cpu)
|
| 1737 | {
|
| 1738 | unsigned long capacity = arch_scale_cpu_capacity(cpu);
|
| 1739 | struct asym_cap_data *insert_entry = NULL;
|
| 1740 | struct asym_cap_data *entry;
|
| 1741 |
|
| 1742 | |
| 1743 | * Search if capacity already exits. If not, track which the entry
|
| 1744 | * where we should insert to keep the list ordered descending.
|
| 1745 |
|
| 1746 | list_for_each_entry(entry, &asym_cap_list, link) {
|
| 1747 | if (capacity == entry->capacity)
|
| 1748 | goto done;
|
| 1749 | else if (!insert_entry && capacity > entry->capacity)
|
| 1750 | insert_entry = list_prev_entry(entry, link);
|
| 1751 | }
|
| 1752 |
|
| 1753 | entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
|
| 1754 | if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
|
| 1755 | return;
|
| 1756 | entry->capacity = capacity;
|
| 1757 |
|
| 1758 |
|
| 1759 | if (!insert_entry)
|
| 1760 | list_add_tail_rcu(&entry->link, &asym_cap_list);
|
| 1761 | else
|
| 1762 | list_add_rcu(&entry->link, &insert_entry->link);
|
| 1763 | done:
|
| 1764 | __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
|
| 1765 | }
|
| 1766 |
|
| 1767 | |
| 1768 | * Build-up/update list of CPUs grouped by their capacities
|
| 1769 | * An update requires explicit request to rebuild sched domains
|
| 1770 | * with state indicating CPU topology changes.
|
| 1771 |
|
| 1772 | static void asym_cpu_capacity_scan(void)
|
| 1773 | {
|
| 1774 | struct asym_cap_data *entry, *next;
|
| 1775 | int cpu;
|
| 1776 |
|
| 1777 | list_for_each_entry(entry, &asym_cap_list, link)
|
| 1778 | cpumask_clear(cpu_capacity_span(entry));
|
| 1779 |
|
| 1780 | for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
|
| 1781 | asym_cpu_capacity_update_data(cpu);
|
| 1782 |
|
| 1783 | list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
|
| 1784 | if (cpumask_empty(cpu_capacity_span(entry))) {
|
| 1785 | list_del_rcu(&entry->link);
|
| 1786 | call_rcu(&entry->rcu, free_asym_cap_entry);
|
| 1787 | }
|
| 1788 | }
|
| 1789 |
|
| 1790 | |
| 1791 | * Only one capacity value has been detected i.e. this system is symmetric.
|
| 1792 | * No need to keep this data around.
|
| 1793 |
|
| 1794 | if (list_is_singular(&asym_cap_list)) {
|
| 1795 | entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
|
| 1796 | list_del_rcu(&entry->link);
|
| 1797 | call_rcu(&entry->rcu, free_asym_cap_entry);
|
| 1798 | }
|
| 1799 | }
|
| 1800 |
|
| 1801 | |
| 1802 | * Initializers for schedule domains
|
| 1803 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
|
| 1804 |
|
| 1805 |
|
| 1806 | static int default_relax_domain_level = -1;
|
| 1807 | int sched_domain_level_max;
|
| 1808 |
|
| 1809 | static int __init setup_relax_domain_level(char *str)
|
| 1810 | {
|
| 1811 | if (kstrtoint(str, 0, &default_relax_domain_level))
|
| 1812 | pr_warn("Unable to set relax_domain_level\n");
|
| 1813 |
|
| 1814 | return 1;
|
| 1815 | }
|
| 1816 | __setup("relax_domain_level=", setup_relax_domain_level);
|
| 1817 |
|
| 1818 | static void set_domain_attribute(struct sched_domain *sd,
|
| 1819 | struct sched_domain_attr *attr)
|
| 1820 | {
|
| 1821 | int request;
|
| 1822 |
|
| 1823 | if (!attr || attr->relax_domain_level < 0) {
|
| 1824 | if (default_relax_domain_level < 0)
|
| 1825 | return;
|
| 1826 | request = default_relax_domain_level;
|
| 1827 | } else
|
| 1828 | request = attr->relax_domain_level;
|
| 1829 |
|
| 1830 | if (sd->level >= request) {
|
| 1831 |
|
| 1832 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
|
| 1833 | }
|
| 1834 | }
|
| 1835 |
|
| 1836 | static void __sdt_free(const struct cpumask *cpu_map);
|
| 1837 | static int __sdt_alloc(const struct cpumask *cpu_map);
|
| 1838 |
|
| 1839 | static void __sds_free(struct s_data *d, const struct cpumask *cpu_map);
|
| 1840 | static int __sds_alloc(struct s_data *d, const struct cpumask *cpu_map);
|
| 1841 |
|
| 1842 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
|
| 1843 | const struct cpumask *cpu_map)
|
| 1844 | {
|
| 1845 | switch (what) {
|
| 1846 | case sa_rootdomain:
|
| 1847 | if (!atomic_read(&d->rd->refcount))
|
| 1848 | free_rootdomain(&d->rd->rcu);
|
| 1849 | fallthrough;
|
| 1850 | case sa_sd:
|
| 1851 | free_percpu(d->sd);
|
| 1852 | fallthrough;
|
| 1853 | case sa_sd_shared:
|
| 1854 | __sds_free(d, cpu_map);
|
| 1855 | fallthrough;
|
| 1856 | case sa_sd_storage:
|
| 1857 | __sdt_free(cpu_map);
|
| 1858 | fallthrough;
|
| 1859 | case sa_none:
|
| 1860 | break;
|
| 1861 | }
|
| 1862 | }
|
| 1863 |
|
| 1864 | static enum s_alloc
|
| 1865 | __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
|
| 1866 | {
|
| 1867 | memset(d, 0, sizeof(*d));
|
| 1868 |
|
| 1869 | if (__sdt_alloc(cpu_map))
|
| 1870 | return sa_sd_storage;
|
| 1871 | if (__sds_alloc(d, cpu_map))
|
| 1872 | return sa_sd_shared;
|
| 1873 | d->sd = alloc_percpu(struct sched_domain *);
|
| 1874 | if (!d->sd)
|
| 1875 | return sa_sd_shared;
|
| 1876 | d->rd = alloc_rootdomain();
|
| 1877 | if (!d->rd)
|
| 1878 | return sa_sd;
|
| 1879 |
|
| 1880 | return sa_rootdomain;
|
| 1881 | }
|
| 1882 |
|
| 1883 | |
| 1884 | * NULL the sd_data elements we've used to build the sched_domain and
|
| 1885 | * sched_group structure so that the subsequent __free_domain_allocs()
|
| 1886 | * will not free the data we're using.
|
| 1887 |
|
| 1888 | static void claim_allocations(int cpu, struct s_data *d)
|
| 1889 | {
|
| 1890 | struct sched_domain *sd;
|
| 1891 |
|
| 1892 | if (atomic_read(&(*per_cpu_ptr(d->sds, cpu))->ref))
|
| 1893 | *per_cpu_ptr(d->sds, cpu) = NULL;
|
| 1894 |
|
| 1895 | for (sd = *per_cpu_ptr(d->sd, cpu); sd; sd = sd->parent) {
|
| 1896 | struct sd_data *sdd = sd->private;
|
| 1897 |
|
| 1898 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
|
| 1899 | *per_cpu_ptr(sdd->sd, cpu) = NULL;
|
| 1900 |
|
| 1901 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
|
| 1902 | *per_cpu_ptr(sdd->sg, cpu) = NULL;
|
| 1903 |
|
| 1904 | if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
|
| 1905 | *per_cpu_ptr(sdd->sgc, cpu) = NULL;
|
| 1906 | }
|
| 1907 | }
|
| 1908 |
|
| 1909 | #ifdef CONFIG_NUMA
|
| 1910 | enum numa_topology_type sched_numa_topology_type;
|
| 1911 |
|
| 1912 | |
| 1913 | * sched_domains_numa_distance is derived from sched_numa_node_distance
|
| 1914 | * and provides a simplified view of NUMA distances used specifically
|
| 1915 | * for building NUMA scheduling domains.
|
| 1916 |
|
| 1917 | static int sched_domains_numa_levels;
|
| 1918 | static int sched_numa_node_levels;
|
| 1919 |
|
| 1920 | int sched_max_numa_distance;
|
| 1921 | static int *sched_domains_numa_distance;
|
| 1922 | static int *sched_numa_node_distance;
|
| 1923 | static struct cpumask ***sched_domains_numa_masks;
|
| 1924 | #endif
|
| 1925 |
|
| 1926 | |
| 1927 | * SD_flags allowed in topology descriptions.
|
| 1928 | *
|
| 1929 | * These flags are purely descriptive of the topology and do not prescribe
|
| 1930 | * behaviour. Behaviour is artificial and mapped in the below sd_init()
|
| 1931 | * function. For details, see include/linux/sched/sd_flags.h.
|
| 1932 | *
|
| 1933 | * SD_SHARE_CPUCAPACITY
|
| 1934 | * SD_SHARE_LLC
|
| 1935 | * SD_CLUSTER
|
| 1936 | * SD_NUMA
|
| 1937 | *
|
| 1938 | * Odd one out, which beside describing the topology has a quirk also
|
| 1939 | * prescribes the desired behaviour that goes along with it:
|
| 1940 | *
|
| 1941 | * SD_ASYM_PACKING - describes SMT quirks
|
| 1942 |
|
| 1943 | #define TOPOLOGY_SD_FLAGS \
|
| 1944 | (SD_SHARE_CPUCAPACITY | \
|
| 1945 | SD_CLUSTER | \
|
| 1946 | SD_SHARE_LLC | \
|
| 1947 | SD_NUMA | \
|
| 1948 | SD_ASYM_PACKING)
|
| 1949 |
|
| 1950 | static struct sched_domain *
|
| 1951 | sd_init(struct sched_domain_topology_level *tl,
|
| 1952 | const struct cpumask *cpu_map,
|
| 1953 | struct sched_domain *child, int cpu)
|
| 1954 | {
|
| 1955 | struct sd_data *sdd = &tl->data;
|
| 1956 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
|
| 1957 | int sd_id, sd_weight, sd_flags = 0;
|
| 1958 | struct cpumask *sd_span;
|
| 1959 | u64 now = sched_clock();
|
| 1960 |
|
| 1961 | sd_span = sched_domain_span(sd);
|
| 1962 | cpumask_and(sd_span, cpu_map, tl->mask(tl, cpu));
|
| 1963 | sd_weight = cpumask_weight(sd_span);
|
| 1964 | sd_id = cpumask_first(sd_span);
|
| 1965 |
|
| 1966 | if (tl->sd_flags)
|
| 1967 | sd_flags = (*tl->sd_flags)();
|
| 1968 | if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
|
| 1969 | "wrong sd_flags in topology description\n"))
|
| 1970 | sd_flags &= TOPOLOGY_SD_FLAGS;
|
| 1971 | #ifdef CONFIG_SCHED_SMT
|
| 1972 | if (sched_smt_asym_packing && (sd_flags & SD_SHARE_CPUCAPACITY))
|
| 1973 | sd_flags |= SD_ASYM_PACKING;
|
| 1974 | #endif
|
| 1975 | sd_flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
|
| 1976 |
|
| 1977 | *sd = (struct sched_domain){
|
| 1978 | .min_interval = sd_weight,
|
| 1979 | .max_interval = 2*sd_weight,
|
| 1980 | .busy_factor = 16,
|
| 1981 | .imbalance_pct = 117,
|
| 1982 |
|
| 1983 | .cache_nice_tries = 0,
|
| 1984 |
|
| 1985 | .flags = 1*SD_BALANCE_NEWIDLE
|
| 1986 | | 1*SD_BALANCE_EXEC
|
| 1987 | | 1*SD_BALANCE_FORK
|
| 1988 | | 0*SD_BALANCE_WAKE
|
| 1989 | | 1*SD_WAKE_AFFINE
|
| 1990 | | 0*SD_SHARE_CPUCAPACITY
|
| 1991 | | 0*SD_SHARE_LLC
|
| 1992 | | 0*SD_SERIALIZE
|
| 1993 | | 1*SD_PREFER_SIBLING
|
| 1994 | | 0*SD_NUMA
|
| 1995 | | sd_flags
|
| 1996 | ,
|
| 1997 |
|
| 1998 | .last_balance = jiffies,
|
| 1999 | .balance_interval = sd_weight,
|
| 2000 |
|
| 2001 |
|
| 2002 | .newidle_call = 512,
|
| 2003 | .newidle_success = 256,
|
| 2004 | .newidle_ratio = 512,
|
| 2005 | .newidle_stamp = now,
|
| 2006 |
|
| 2007 | .max_newidle_lb_cost = 0,
|
| 2008 | .last_decay_max_lb_cost = jiffies,
|
| 2009 | .child = child,
|
| 2010 | .name = tl->name,
|
| 2011 | };
|
| 2012 |
|
| 2013 | WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
|
| 2014 | (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
|
| 2015 | "CPU capacity asymmetry not supported on SMT\n");
|
| 2016 |
|
| 2017 | |
| 2018 | * Convert topological properties into behaviour.
|
| 2019 |
|
| 2020 | if (sd->flags & SD_SHARE_CPUCAPACITY) {
|
| 2021 | sd->imbalance_pct = 110;
|
| 2022 |
|
| 2023 | } else if (sd->flags & SD_SHARE_LLC) {
|
| 2024 | sd->imbalance_pct = 117;
|
| 2025 | sd->cache_nice_tries = 1;
|
| 2026 |
|
| 2027 | #ifdef CONFIG_NUMA
|
| 2028 | } else if (sd->flags & SD_NUMA) {
|
| 2029 | sd->cache_nice_tries = 2;
|
| 2030 |
|
| 2031 | sd->flags &= ~SD_PREFER_SIBLING;
|
| 2032 | sd->flags |= SD_SERIALIZE;
|
| 2033 | if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
|
| 2034 | sd->flags &= ~(SD_BALANCE_EXEC |
|
| 2035 | SD_BALANCE_FORK |
|
| 2036 | SD_WAKE_AFFINE);
|
| 2037 | }
|
| 2038 |
|
| 2039 | #endif
|
| 2040 | } else {
|
| 2041 | sd->cache_nice_tries = 1;
|
| 2042 | }
|
| 2043 |
|
| 2044 | sd->private = sdd;
|
| 2045 |
|
| 2046 | return sd;
|
| 2047 | }
|
| 2048 |
|
| 2049 | #ifdef CONFIG_SCHED_SMT
|
| 2050 | int cpu_smt_flags(void)
|
| 2051 | {
|
| 2052 | return SD_SHARE_CPUCAPACITY | SD_SHARE_LLC;
|
| 2053 | }
|
| 2054 |
|
| 2055 | const struct cpumask *tl_smt_mask(struct sched_domain_topology_level *tl, int cpu)
|
| 2056 | {
|
| 2057 | return cpu_smt_mask(cpu);
|
| 2058 | }
|
| 2059 | #endif
|
| 2060 |
|
| 2061 | #ifdef CONFIG_SCHED_CLUSTER
|
| 2062 | int cpu_cluster_flags(void)
|
| 2063 | {
|
| 2064 | return SD_CLUSTER | SD_SHARE_LLC;
|
| 2065 | }
|
| 2066 |
|
| 2067 | const struct cpumask *tl_cls_mask(struct sched_domain_topology_level *tl, int cpu)
|
| 2068 | {
|
| 2069 | return cpu_clustergroup_mask(cpu);
|
| 2070 | }
|
| 2071 | #endif
|
| 2072 |
|
| 2073 | #ifdef CONFIG_SCHED_MC
|
| 2074 | int cpu_core_flags(void)
|
| 2075 | {
|
| 2076 | return SD_SHARE_LLC;
|
| 2077 | }
|
| 2078 |
|
| 2079 | const struct cpumask *tl_mc_mask(struct sched_domain_topology_level *tl, int cpu)
|
| 2080 | {
|
| 2081 | return cpu_coregroup_mask(cpu);
|
| 2082 | }
|
| 2083 |
|
| 2084 | |
| 2085 | * Majority of architectures have LLC at MC domain level with exception
|
| 2086 | * such as powerpc. Provide a way for arch to specify where its LLC is
|
| 2087 | * if it falls in exception category
|
| 2088 |
|
| 2089 | # ifndef arch_llc_mask
|
| 2090 | #define arch_llc_mask(cpu) cpu_coregroup_mask(cpu)
|
| 2091 | # endif
|
| 2092 |
|
| 2093 | #else
|
| 2094 | #define arch_llc_mask(cpu) cpumask_of(cpu)
|
| 2095 | #endif
|
| 2096 |
|
| 2097 | #define llc_mask(cpu) arch_llc_mask(cpu)
|
| 2098 |
|
| 2099 | const struct cpumask *tl_pkg_mask(struct sched_domain_topology_level *tl, int cpu)
|
| 2100 | {
|
| 2101 | return cpu_node_mask(cpu);
|
| 2102 | }
|
| 2103 |
|
| 2104 | |
| 2105 | * Topology list, bottom-up.
|
| 2106 |
|
| 2107 | static struct sched_domain_topology_level default_topology[] = {
|
| 2108 | #ifdef CONFIG_SCHED_SMT
|
| 2109 | SDTL_INIT(tl_smt_mask, cpu_smt_flags, SMT),
|
| 2110 | #endif
|
| 2111 |
|
| 2112 | #ifdef CONFIG_SCHED_CLUSTER
|
| 2113 | SDTL_INIT(tl_cls_mask, cpu_cluster_flags, CLS),
|
| 2114 | #endif
|
| 2115 |
|
| 2116 | #ifdef CONFIG_SCHED_MC
|
| 2117 | SDTL_INIT(tl_mc_mask, cpu_core_flags, MC),
|
| 2118 | #endif
|
| 2119 | SDTL_INIT(tl_pkg_mask, NULL, PKG),
|
| 2120 | { NULL, },
|
| 2121 | };
|
| 2122 |
|
| 2123 | static struct sched_domain_topology_level *sched_domain_topology =
|
| 2124 | default_topology;
|
| 2125 | static struct sched_domain_topology_level *sched_domain_topology_saved;
|
| 2126 |
|
| 2127 | #define for_each_sd_topology(tl) \
|
| 2128 | for (tl = sched_domain_topology; tl->mask; tl++)
|
| 2129 |
|
| 2130 | void __init set_sched_topology(struct sched_domain_topology_level *tl)
|
| 2131 | {
|
| 2132 | if (WARN_ON_ONCE(sched_smp_initialized))
|
| 2133 | return;
|
| 2134 |
|
| 2135 | sched_domain_topology = tl;
|
| 2136 | sched_domain_topology_saved = NULL;
|
| 2137 | }
|
| 2138 |
|
| 2139 | #ifdef CONFIG_NUMA
|
| 2140 | static int cpu_numa_flags(void)
|
| 2141 | {
|
| 2142 | return SD_NUMA;
|
| 2143 | }
|
| 2144 |
|
| 2145 | static const struct cpumask *sd_numa_mask(struct sched_domain_topology_level *tl, int cpu)
|
| 2146 | {
|
| 2147 | return sched_domains_numa_masks[tl->numa_level][cpu_to_node(cpu)];
|
| 2148 | }
|
| 2149 |
|
| 2150 | static void sched_numa_warn(const char *str)
|
| 2151 | {
|
| 2152 | static int done = false;
|
| 2153 | int i,j;
|
| 2154 |
|
| 2155 | if (done)
|
| 2156 | return;
|
| 2157 |
|
| 2158 | done = true;
|
| 2159 |
|
| 2160 | printk(KERN_WARNING "ERROR: %s\n\n", str);
|
| 2161 |
|
| 2162 | for (i = 0; i < nr_node_ids; i++) {
|
| 2163 | printk(KERN_WARNING " ");
|
| 2164 | for (j = 0; j < nr_node_ids; j++) {
|
| 2165 | if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
|
| 2166 | printk(KERN_CONT "(%02d) ", node_distance(i,j));
|
| 2167 | else
|
| 2168 | printk(KERN_CONT " %02d ", node_distance(i,j));
|
| 2169 | }
|
| 2170 | printk(KERN_CONT "\n");
|
| 2171 | }
|
| 2172 | printk(KERN_WARNING "\n");
|
| 2173 | }
|
| 2174 |
|
| 2175 | bool find_numa_distance(int distance)
|
| 2176 | {
|
| 2177 | bool found = false;
|
| 2178 | int i, *distances;
|
| 2179 |
|
| 2180 | if (distance == node_distance(0, 0))
|
| 2181 | return true;
|
| 2182 |
|
| 2183 | rcu_read_lock();
|
| 2184 | distances = rcu_dereference(sched_numa_node_distance);
|
| 2185 | if (!distances)
|
| 2186 | goto unlock;
|
| 2187 | for (i = 0; i < sched_numa_node_levels; i++) {
|
| 2188 | if (distances[i] == distance) {
|
| 2189 | found = true;
|
| 2190 | break;
|
| 2191 | }
|
| 2192 | }
|
| 2193 | unlock:
|
| 2194 | rcu_read_unlock();
|
| 2195 |
|
| 2196 | return found;
|
| 2197 | }
|
| 2198 |
|
| 2199 | #define for_each_cpu_node_but(n, nbut) \
|
| 2200 | for_each_node_state(n, N_CPU) \
|
| 2201 | if (n == nbut) \
|
| 2202 | continue; \
|
| 2203 | else
|
| 2204 |
|
| 2205 | |
| 2206 | * A system can have three types of NUMA topology:
|
| 2207 | * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
|
| 2208 | * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
|
| 2209 | * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
|
| 2210 | *
|
| 2211 | * The difference between a glueless mesh topology and a backplane
|
| 2212 | * topology lies in whether communication between not directly
|
| 2213 | * connected nodes goes through intermediary nodes (where programs
|
| 2214 | * could run), or through backplane controllers. This affects
|
| 2215 | * placement of programs.
|
| 2216 | *
|
| 2217 | * The type of topology can be discerned with the following tests:
|
| 2218 | * - If the maximum distance between any nodes is 1 hop, the system
|
| 2219 | * is directly connected.
|
| 2220 | * - If for two nodes A and B, located N > 1 hops away from each other,
|
| 2221 | * there is an intermediary node C, which is < N hops away from both
|
| 2222 | * nodes A and B, the system is a glueless mesh.
|
| 2223 |
|
| 2224 | static void init_numa_topology_type(int offline_node)
|
| 2225 | {
|
| 2226 | int a, b, c, n;
|
| 2227 |
|
| 2228 | n = sched_max_numa_distance;
|
| 2229 |
|
| 2230 | if (sched_domains_numa_levels <= 2) {
|
| 2231 | sched_numa_topology_type = NUMA_DIRECT;
|
| 2232 | return;
|
| 2233 | }
|
| 2234 |
|
| 2235 | for_each_cpu_node_but(a, offline_node) {
|
| 2236 | for_each_cpu_node_but(b, offline_node) {
|
| 2237 |
|
| 2238 | if (node_distance(a, b) < n)
|
| 2239 | continue;
|
| 2240 |
|
| 2241 |
|
| 2242 | for_each_cpu_node_but(c, offline_node) {
|
| 2243 | if (node_distance(a, c) < n &&
|
| 2244 | node_distance(b, c) < n) {
|
| 2245 | sched_numa_topology_type =
|
| 2246 | NUMA_GLUELESS_MESH;
|
| 2247 | return;
|
| 2248 | }
|
| 2249 | }
|
| 2250 |
|
| 2251 | sched_numa_topology_type = NUMA_BACKPLANE;
|
| 2252 | return;
|
| 2253 | }
|
| 2254 | }
|
| 2255 |
|
| 2256 | pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
|
| 2257 | sched_numa_topology_type = NUMA_DIRECT;
|
| 2258 | }
|
| 2259 |
|
| 2260 |
|
| 2261 | #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
|
| 2262 |
|
| 2263 | |
| 2264 | * An architecture could modify its NUMA distance, to change
|
| 2265 | * grouping of NUMA nodes and number of NUMA levels when creating
|
| 2266 | * NUMA level sched domains.
|
| 2267 | *
|
| 2268 | * A NUMA level is created for each unique
|
| 2269 | * arch_sched_node_distance.
|
| 2270 |
|
| 2271 | static int numa_node_dist(int i, int j)
|
| 2272 | {
|
| 2273 | return node_distance(i, j);
|
| 2274 | }
|
| 2275 |
|
| 2276 | int arch_sched_node_distance(int from, int to)
|
| 2277 | __weak __alias(numa_node_dist);
|
| 2278 |
|
| 2279 | static bool modified_sched_node_distance(void)
|
| 2280 | {
|
| 2281 | return numa_node_dist != arch_sched_node_distance;
|
| 2282 | }
|
| 2283 |
|
| 2284 | static int sched_record_numa_dist(int offline_node, int (*n_dist)(int, int),
|
| 2285 | int **dist, int *levels)
|
| 2286 | {
|
| 2287 | unsigned long *distance_map __free(bitmap) = NULL;
|
| 2288 | int nr_levels = 0;
|
| 2289 | int i, j;
|
| 2290 | int *distances;
|
| 2291 |
|
| 2292 | |
| 2293 | * O(nr_nodes^2) de-duplicating selection sort -- in order to find the
|
| 2294 | * unique distances in the node_distance() table.
|
| 2295 |
|
| 2296 | distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
|
| 2297 | if (!distance_map)
|
| 2298 | return -ENOMEM;
|
| 2299 |
|
| 2300 | bitmap_zero(distance_map, NR_DISTANCE_VALUES);
|
| 2301 | for_each_cpu_node_but(i, offline_node) {
|
| 2302 | for_each_cpu_node_but(j, offline_node) {
|
| 2303 | int distance = n_dist(i, j);
|
| 2304 |
|
| 2305 | if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
|
| 2306 | sched_numa_warn("Invalid distance value range");
|
| 2307 | return -EINVAL;
|
| 2308 | }
|
| 2309 |
|
| 2310 | bitmap_set(distance_map, distance, 1);
|
| 2311 | }
|
| 2312 | }
|
| 2313 | |
| 2314 | * We can now figure out how many unique distance values there are and
|
| 2315 | * allocate memory accordingly.
|
| 2316 |
|
| 2317 | nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
|
| 2318 |
|
| 2319 | distances = kzalloc_objs(int, nr_levels);
|
| 2320 | if (!distances)
|
| 2321 | return -ENOMEM;
|
| 2322 |
|
| 2323 | for (i = 0, j = 0; i < nr_levels; i++, j++) {
|
| 2324 | j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
|
| 2325 | distances[i] = j;
|
| 2326 | }
|
| 2327 | *dist = distances;
|
| 2328 | *levels = nr_levels;
|
| 2329 |
|
| 2330 | return 0;
|
| 2331 | }
|
| 2332 |
|
| 2333 | void sched_init_numa(int offline_node)
|
| 2334 | {
|
| 2335 | struct sched_domain_topology_level *tl;
|
| 2336 | int nr_levels, nr_node_levels;
|
| 2337 | int i, j;
|
| 2338 | int *distances, *domain_distances;
|
| 2339 | struct cpumask ***masks;
|
| 2340 |
|
| 2341 |
|
| 2342 | if (sched_record_numa_dist(offline_node, numa_node_dist, &distances,
|
| 2343 | &nr_node_levels))
|
| 2344 | return;
|
| 2345 |
|
| 2346 |
|
| 2347 | if (modified_sched_node_distance()) {
|
| 2348 | if (sched_record_numa_dist(offline_node, arch_sched_node_distance,
|
| 2349 | &domain_distances, &nr_levels)) {
|
| 2350 | kfree(distances);
|
| 2351 | return;
|
| 2352 | }
|
| 2353 | } else {
|
| 2354 | domain_distances = distances;
|
| 2355 | nr_levels = nr_node_levels;
|
| 2356 | }
|
| 2357 | rcu_assign_pointer(sched_numa_node_distance, distances);
|
| 2358 | WRITE_ONCE(sched_max_numa_distance, distances[nr_node_levels - 1]);
|
| 2359 | WRITE_ONCE(sched_numa_node_levels, nr_node_levels);
|
| 2360 |
|
| 2361 | |
| 2362 | * 'nr_levels' contains the number of unique distances
|
| 2363 | *
|
| 2364 | * The sched_domains_numa_distance[] array includes the actual distance
|
| 2365 | * numbers.
|
| 2366 |
|
| 2367 |
|
| 2368 | |
| 2369 | * Here, we should temporarily reset sched_domains_numa_levels to 0.
|
| 2370 | * If it fails to allocate memory for array sched_domains_numa_masks[][],
|
| 2371 | * the array will contain less then 'nr_levels' members. This could be
|
| 2372 | * dangerous when we use it to iterate array sched_domains_numa_masks[][]
|
| 2373 | * in other functions.
|
| 2374 | *
|
| 2375 | * We reset it to 'nr_levels' at the end of this function.
|
| 2376 |
|
| 2377 | rcu_assign_pointer(sched_domains_numa_distance, domain_distances);
|
| 2378 |
|
| 2379 | sched_domains_numa_levels = 0;
|
| 2380 |
|
| 2381 | masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
|
| 2382 | if (!masks)
|
| 2383 | return;
|
| 2384 |
|
| 2385 | |
| 2386 | * Now for each level, construct a mask per node which contains all
|
| 2387 | * CPUs of nodes that are that many hops away from us.
|
| 2388 |
|
| 2389 | for (i = 0; i < nr_levels; i++) {
|
| 2390 | masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
|
| 2391 | if (!masks[i])
|
| 2392 | return;
|
| 2393 |
|
| 2394 | for_each_cpu_node_but(j, offline_node) {
|
| 2395 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
|
| 2396 | int k;
|
| 2397 |
|
| 2398 | if (!mask)
|
| 2399 | return;
|
| 2400 |
|
| 2401 | masks[i][j] = mask;
|
| 2402 |
|
| 2403 | for_each_cpu_node_but(k, offline_node) {
|
| 2404 | if (sched_debug() &&
|
| 2405 | (arch_sched_node_distance(j, k) !=
|
| 2406 | arch_sched_node_distance(k, j)))
|
| 2407 | sched_numa_warn("Node-distance not symmetric");
|
| 2408 |
|
| 2409 | if (arch_sched_node_distance(j, k) >
|
| 2410 | sched_domains_numa_distance[i])
|
| 2411 | continue;
|
| 2412 |
|
| 2413 | cpumask_or(mask, mask, cpumask_of_node(k));
|
| 2414 | }
|
| 2415 | }
|
| 2416 | }
|
| 2417 | rcu_assign_pointer(sched_domains_numa_masks, masks);
|
| 2418 |
|
| 2419 |
|
| 2420 | for (i = 0; sched_domain_topology[i].mask; i++);
|
| 2421 |
|
| 2422 | tl = kzalloc((i + nr_levels + 1) *
|
| 2423 | sizeof(struct sched_domain_topology_level), GFP_KERNEL);
|
| 2424 | if (!tl)
|
| 2425 | return;
|
| 2426 |
|
| 2427 | |
| 2428 | * Copy the default topology bits..
|
| 2429 |
|
| 2430 | for (i = 0; sched_domain_topology[i].mask; i++)
|
| 2431 | tl[i] = sched_domain_topology[i];
|
| 2432 |
|
| 2433 | |
| 2434 | * Add the NUMA identity distance, aka single NODE.
|
| 2435 |
|
| 2436 | tl[i++] = SDTL_INIT(sd_numa_mask, NULL, NODE);
|
| 2437 |
|
| 2438 | |
| 2439 | * .. and append 'j' levels of NUMA goodness.
|
| 2440 |
|
| 2441 | for (j = 1; j < nr_levels; i++, j++) {
|
| 2442 | tl[i] = SDTL_INIT(sd_numa_mask, cpu_numa_flags, NUMA);
|
| 2443 | tl[i].numa_level = j;
|
| 2444 | }
|
| 2445 |
|
| 2446 | sched_domain_topology_saved = sched_domain_topology;
|
| 2447 | sched_domain_topology = tl;
|
| 2448 |
|
| 2449 | sched_domains_numa_levels = nr_levels;
|
| 2450 |
|
| 2451 | init_numa_topology_type(offline_node);
|
| 2452 | }
|
| 2453 |
|
| 2454 |
|
| 2455 | static void sched_reset_numa(void)
|
| 2456 | {
|
| 2457 | int nr_levels, *distances, *dom_distances = NULL;
|
| 2458 | struct cpumask ***masks;
|
| 2459 |
|
| 2460 | nr_levels = sched_domains_numa_levels;
|
| 2461 | sched_numa_node_levels = 0;
|
| 2462 | sched_domains_numa_levels = 0;
|
| 2463 | sched_max_numa_distance = 0;
|
| 2464 | sched_numa_topology_type = NUMA_DIRECT;
|
| 2465 | distances = sched_numa_node_distance;
|
| 2466 | if (sched_numa_node_distance != sched_domains_numa_distance)
|
| 2467 | dom_distances = sched_domains_numa_distance;
|
| 2468 | rcu_assign_pointer(sched_numa_node_distance, NULL);
|
| 2469 | rcu_assign_pointer(sched_domains_numa_distance, NULL);
|
| 2470 | masks = sched_domains_numa_masks;
|
| 2471 | rcu_assign_pointer(sched_domains_numa_masks, NULL);
|
| 2472 | if (distances || masks) {
|
| 2473 | int i, j;
|
| 2474 |
|
| 2475 | synchronize_rcu();
|
| 2476 | kfree(distances);
|
| 2477 | kfree(dom_distances);
|
| 2478 | for (i = 0; i < nr_levels && masks; i++) {
|
| 2479 | if (!masks[i])
|
| 2480 | continue;
|
| 2481 | for_each_node(j)
|
| 2482 | kfree(masks[i][j]);
|
| 2483 | kfree(masks[i]);
|
| 2484 | }
|
| 2485 | kfree(masks);
|
| 2486 | }
|
| 2487 | if (sched_domain_topology_saved) {
|
| 2488 | kfree(sched_domain_topology);
|
| 2489 | sched_domain_topology = sched_domain_topology_saved;
|
| 2490 | sched_domain_topology_saved = NULL;
|
| 2491 | }
|
| 2492 | }
|
| 2493 |
|
| 2494 | |
| 2495 | * Call with hotplug lock held
|
| 2496 |
|
| 2497 | void sched_update_numa(int cpu, bool online)
|
| 2498 | {
|
| 2499 | int node;
|
| 2500 |
|
| 2501 | node = cpu_to_node(cpu);
|
| 2502 | |
| 2503 | * Scheduler NUMA topology is updated when the first CPU of a
|
| 2504 | * node is onlined or the last CPU of a node is offlined.
|
| 2505 |
|
| 2506 | if (cpumask_weight(cpumask_of_node(node)) != 1)
|
| 2507 | return;
|
| 2508 |
|
| 2509 | sched_reset_numa();
|
| 2510 | sched_init_numa(online ? NUMA_NO_NODE : node);
|
| 2511 | }
|
| 2512 |
|
| 2513 | void sched_domains_numa_masks_set(unsigned int cpu)
|
| 2514 | {
|
| 2515 | int node = cpu_to_node(cpu);
|
| 2516 | int i, j;
|
| 2517 |
|
| 2518 | for (i = 0; i < sched_domains_numa_levels; i++) {
|
| 2519 | for (j = 0; j < nr_node_ids; j++) {
|
| 2520 | if (!node_state(j, N_CPU))
|
| 2521 | continue;
|
| 2522 |
|
| 2523 |
|
| 2524 | if (arch_sched_node_distance(j, node) <=
|
| 2525 | sched_domains_numa_distance[i])
|
| 2526 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
|
| 2527 | }
|
| 2528 | }
|
| 2529 | }
|
| 2530 |
|
| 2531 | void sched_domains_numa_masks_clear(unsigned int cpu)
|
| 2532 | {
|
| 2533 | int i, j;
|
| 2534 |
|
| 2535 | for (i = 0; i < sched_domains_numa_levels; i++) {
|
| 2536 | for (j = 0; j < nr_node_ids; j++) {
|
| 2537 | if (sched_domains_numa_masks[i][j])
|
| 2538 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
|
| 2539 | }
|
| 2540 | }
|
| 2541 | }
|
| 2542 |
|
| 2543 | |
| 2544 | * sched_numa_find_closest() - given the NUMA topology, find the cpu
|
| 2545 | * closest to @cpu from @cpumask.
|
| 2546 | * cpumask: cpumask to find a cpu from
|
| 2547 | * cpu: cpu to be close to
|
| 2548 | *
|
| 2549 | * returns: cpu, or nr_cpu_ids when nothing found.
|
| 2550 |
|
| 2551 | int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
|
| 2552 | {
|
| 2553 | int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
|
| 2554 | struct cpumask ***masks;
|
| 2555 |
|
| 2556 | rcu_read_lock();
|
| 2557 | masks = rcu_dereference(sched_domains_numa_masks);
|
| 2558 | if (!masks)
|
| 2559 | goto unlock;
|
| 2560 | for (i = 0; i < sched_domains_numa_levels; i++) {
|
| 2561 | if (!masks[i][j])
|
| 2562 | break;
|
| 2563 | cpu = cpumask_any_and_distribute(cpus, masks[i][j]);
|
| 2564 | if (cpu < nr_cpu_ids) {
|
| 2565 | found = cpu;
|
| 2566 | break;
|
| 2567 | }
|
| 2568 | }
|
| 2569 | unlock:
|
| 2570 | rcu_read_unlock();
|
| 2571 |
|
| 2572 | return found;
|
| 2573 | }
|
| 2574 |
|
| 2575 | struct __cmp_key {
|
| 2576 | const struct cpumask *cpus;
|
| 2577 | struct cpumask ***masks;
|
| 2578 | int node;
|
| 2579 | int cpu;
|
| 2580 | int w;
|
| 2581 | };
|
| 2582 |
|
| 2583 | static int hop_cmp(const void *a, const void *b)
|
| 2584 | {
|
| 2585 | struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
|
| 2586 | struct __cmp_key *k = (struct __cmp_key *)a;
|
| 2587 |
|
| 2588 | if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
|
| 2589 | return 1;
|
| 2590 |
|
| 2591 | if (b == k->masks) {
|
| 2592 | k->w = 0;
|
| 2593 | return 0;
|
| 2594 | }
|
| 2595 |
|
| 2596 | prev_hop = *((struct cpumask ***)b - 1);
|
| 2597 | k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
|
| 2598 | if (k->w <= k->cpu)
|
| 2599 | return 0;
|
| 2600 |
|
| 2601 | return -1;
|
| 2602 | }
|
| 2603 |
|
| 2604 | |
| 2605 | * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
|
| 2606 | * from @cpus to @cpu, taking into account distance
|
| 2607 | * from a given @node.
|
| 2608 | * @cpus: cpumask to find a cpu from
|
| 2609 | * @cpu: CPU to start searching
|
| 2610 | * @node: NUMA node to order CPUs by distance
|
| 2611 | *
|
| 2612 | * Return: cpu, or nr_cpu_ids when nothing found.
|
| 2613 |
|
| 2614 | int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
|
| 2615 | {
|
| 2616 | struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
|
| 2617 | struct cpumask ***hop_masks;
|
| 2618 | int hop, ret = nr_cpu_ids;
|
| 2619 |
|
| 2620 | if (node == NUMA_NO_NODE)
|
| 2621 | return cpumask_nth_and(cpu, cpus, cpu_online_mask);
|
| 2622 |
|
| 2623 | rcu_read_lock();
|
| 2624 |
|
| 2625 |
|
| 2626 | node = numa_nearest_node(node, N_CPU);
|
| 2627 | k.node = node;
|
| 2628 |
|
| 2629 | k.masks = rcu_dereference(sched_domains_numa_masks);
|
| 2630 | if (!k.masks)
|
| 2631 | goto unlock;
|
| 2632 |
|
| 2633 | hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
|
| 2634 | if (!hop_masks)
|
| 2635 | goto unlock;
|
| 2636 | hop = hop_masks - k.masks;
|
| 2637 |
|
| 2638 | ret = hop ?
|
| 2639 | cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
|
| 2640 | cpumask_nth_and(cpu, cpus, k.masks[0][node]);
|
| 2641 | unlock:
|
| 2642 | rcu_read_unlock();
|
| 2643 | return ret;
|
| 2644 | }
|
| 2645 | EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
|
| 2646 |
|
| 2647 | |
| 2648 | * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
|
| 2649 | * @node
|
| 2650 | * @node: The node to count hops from.
|
| 2651 | * @hops: Include CPUs up to that many hops away. 0 means local node.
|
| 2652 | *
|
| 2653 | * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
|
| 2654 | * @node, an error value otherwise.
|
| 2655 | *
|
| 2656 | * Requires rcu_lock to be held. Returned cpumask is only valid within that
|
| 2657 | * read-side section, copy it if required beyond that.
|
| 2658 | *
|
| 2659 | * Note that not all hops are equal in distance; see sched_init_numa() for how
|
| 2660 | * distances and masks are handled.
|
| 2661 | * Also note that this is a reflection of sched_domains_numa_masks, which may change
|
| 2662 | * during the lifetime of the system (offline nodes are taken out of the masks).
|
| 2663 |
|
| 2664 | const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
|
| 2665 | {
|
| 2666 | struct cpumask ***masks;
|
| 2667 |
|
| 2668 | if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
|
| 2669 | return ERR_PTR(-EINVAL);
|
| 2670 |
|
| 2671 | masks = rcu_dereference(sched_domains_numa_masks);
|
| 2672 | if (!masks)
|
| 2673 | return ERR_PTR(-EBUSY);
|
| 2674 |
|
| 2675 | return masks[hops][node];
|
| 2676 | }
|
| 2677 | EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
|
| 2678 |
|
| 2679 | #endif
|
| 2680 |
|
| 2681 | static int __sdt_alloc(const struct cpumask *cpu_map)
|
| 2682 | {
|
| 2683 | struct sched_domain_topology_level *tl;
|
| 2684 | int j;
|
| 2685 |
|
| 2686 | for_each_sd_topology(tl) {
|
| 2687 | struct sd_data *sdd = &tl->data;
|
| 2688 |
|
| 2689 | sdd->sd = alloc_percpu(struct sched_domain *);
|
| 2690 | if (!sdd->sd)
|
| 2691 | return -ENOMEM;
|
| 2692 |
|
| 2693 | sdd->sg = alloc_percpu(struct sched_group *);
|
| 2694 | if (!sdd->sg)
|
| 2695 | return -ENOMEM;
|
| 2696 |
|
| 2697 | sdd->sgc = alloc_percpu(struct sched_group_capacity *);
|
| 2698 | if (!sdd->sgc)
|
| 2699 | return -ENOMEM;
|
| 2700 |
|
| 2701 | for_each_cpu(j, cpu_map) {
|
| 2702 | struct sched_domain *sd;
|
| 2703 | struct sched_group *sg;
|
| 2704 | struct sched_group_capacity *sgc;
|
| 2705 |
|
| 2706 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
|
| 2707 | GFP_KERNEL, cpu_to_node(j));
|
| 2708 | if (!sd)
|
| 2709 | return -ENOMEM;
|
| 2710 |
|
| 2711 | *per_cpu_ptr(sdd->sd, j) = sd;
|
| 2712 |
|
| 2713 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
|
| 2714 | GFP_KERNEL, cpu_to_node(j));
|
| 2715 | if (!sg)
|
| 2716 | return -ENOMEM;
|
| 2717 |
|
| 2718 | sg->next = sg;
|
| 2719 |
|
| 2720 | *per_cpu_ptr(sdd->sg, j) = sg;
|
| 2721 |
|
| 2722 | sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
|
| 2723 | GFP_KERNEL, cpu_to_node(j));
|
| 2724 | if (!sgc)
|
| 2725 | return -ENOMEM;
|
| 2726 |
|
| 2727 | sgc->id = j;
|
| 2728 |
|
| 2729 | *per_cpu_ptr(sdd->sgc, j) = sgc;
|
| 2730 | }
|
| 2731 | }
|
| 2732 |
|
| 2733 | return 0;
|
| 2734 | }
|
| 2735 |
|
| 2736 | static void __sdt_free(const struct cpumask *cpu_map)
|
| 2737 | {
|
| 2738 | struct sched_domain_topology_level *tl;
|
| 2739 | int j;
|
| 2740 |
|
| 2741 | for_each_sd_topology(tl) {
|
| 2742 | struct sd_data *sdd = &tl->data;
|
| 2743 |
|
| 2744 | for_each_cpu(j, cpu_map) {
|
| 2745 | struct sched_domain *sd;
|
| 2746 |
|
| 2747 | if (sdd->sd) {
|
| 2748 | sd = *per_cpu_ptr(sdd->sd, j);
|
| 2749 | if (sd && (sd->flags & SD_NUMA))
|
| 2750 | free_sched_groups(sd->groups, 0);
|
| 2751 | kfree(*per_cpu_ptr(sdd->sd, j));
|
| 2752 | }
|
| 2753 |
|
| 2754 | if (sdd->sg)
|
| 2755 | kfree(*per_cpu_ptr(sdd->sg, j));
|
| 2756 | if (sdd->sgc)
|
| 2757 | kfree(*per_cpu_ptr(sdd->sgc, j));
|
| 2758 | }
|
| 2759 | free_percpu(sdd->sd);
|
| 2760 | sdd->sd = NULL;
|
| 2761 | free_percpu(sdd->sg);
|
| 2762 | sdd->sg = NULL;
|
| 2763 | free_percpu(sdd->sgc);
|
| 2764 | sdd->sgc = NULL;
|
| 2765 | }
|
| 2766 | }
|
| 2767 |
|
| 2768 | static int __sds_alloc(struct s_data *d, const struct cpumask *cpu_map)
|
| 2769 | {
|
| 2770 | int j;
|
| 2771 |
|
| 2772 | d->sds = alloc_percpu(struct sched_domain_shared *);
|
| 2773 | if (!d->sds)
|
| 2774 | return -ENOMEM;
|
| 2775 |
|
| 2776 | for_each_cpu(j, cpu_map) {
|
| 2777 | struct sched_domain_shared *sds;
|
| 2778 |
|
| 2779 | sds = kzalloc_node(sizeof(struct sched_domain_shared),
|
| 2780 | GFP_KERNEL, cpu_to_node(j));
|
| 2781 | if (!sds)
|
| 2782 | return -ENOMEM;
|
| 2783 |
|
| 2784 | *per_cpu_ptr(d->sds, j) = sds;
|
| 2785 | }
|
| 2786 |
|
| 2787 | return 0;
|
| 2788 | }
|
| 2789 |
|
| 2790 | static void __sds_free(struct s_data *d, const struct cpumask *cpu_map)
|
| 2791 | {
|
| 2792 | int j;
|
| 2793 |
|
| 2794 | if (!d->sds)
|
| 2795 | return;
|
| 2796 |
|
| 2797 | for_each_cpu(j, cpu_map)
|
| 2798 | kfree(*per_cpu_ptr(d->sds, j));
|
| 2799 |
|
| 2800 | free_percpu(d->sds);
|
| 2801 | d->sds = NULL;
|
| 2802 | }
|
| 2803 |
|
| 2804 | static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
|
| 2805 | const struct cpumask *cpu_map, struct sched_domain_attr *attr,
|
| 2806 | struct sched_domain *child, int cpu)
|
| 2807 | {
|
| 2808 | struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
|
| 2809 |
|
| 2810 | if (child) {
|
| 2811 | sd->level = child->level + 1;
|
| 2812 | sched_domain_level_max = max(sched_domain_level_max, sd->level);
|
| 2813 | child->parent = sd;
|
| 2814 |
|
| 2815 | if (!cpumask_subset(sched_domain_span(child),
|
| 2816 | sched_domain_span(sd))) {
|
| 2817 | pr_err("BUG: arch topology borken\n");
|
| 2818 | pr_err(" the %s domain not a subset of the %s domain\n",
|
| 2819 | child->name, sd->name);
|
| 2820 |
|
| 2821 | cpumask_or(sched_domain_span(sd),
|
| 2822 | sched_domain_span(sd),
|
| 2823 | sched_domain_span(child));
|
| 2824 | }
|
| 2825 |
|
| 2826 | }
|
| 2827 | set_domain_attribute(sd, attr);
|
| 2828 |
|
| 2829 | return sd;
|
| 2830 | }
|
| 2831 |
|
| 2832 | |
| 2833 | * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
|
| 2834 | * any two given CPUs on non-NUMA topology levels.
|
| 2835 |
|
| 2836 | static bool topology_span_sane(const struct cpumask *cpu_map)
|
| 2837 | {
|
| 2838 | struct sched_domain_topology_level *tl;
|
| 2839 | struct cpumask *covered, *id_seen;
|
| 2840 | int cpu;
|
| 2841 |
|
| 2842 | lockdep_assert_held(&sched_domains_mutex);
|
| 2843 | covered = sched_domains_tmpmask;
|
| 2844 | id_seen = sched_domains_tmpmask2;
|
| 2845 |
|
| 2846 | for_each_sd_topology(tl) {
|
| 2847 | int tl_common_flags = 0;
|
| 2848 |
|
| 2849 | if (tl->sd_flags)
|
| 2850 | tl_common_flags = (*tl->sd_flags)();
|
| 2851 |
|
| 2852 |
|
| 2853 | if (tl_common_flags & SD_NUMA)
|
| 2854 | continue;
|
| 2855 |
|
| 2856 | cpumask_clear(covered);
|
| 2857 | cpumask_clear(id_seen);
|
| 2858 |
|
| 2859 | |
| 2860 | * Non-NUMA levels cannot partially overlap - they must be either
|
| 2861 | * completely equal or completely disjoint. Otherwise we can end up
|
| 2862 | * breaking the sched_group lists - i.e. a later get_group() pass
|
| 2863 | * breaks the linking done for an earlier span.
|
| 2864 |
|
| 2865 | for_each_cpu(cpu, cpu_map) {
|
| 2866 | const struct cpumask *tl_cpu_mask = tl->mask(tl, cpu);
|
| 2867 | int id;
|
| 2868 |
|
| 2869 |
|
| 2870 | id = cpumask_first(tl_cpu_mask);
|
| 2871 |
|
| 2872 | if (cpumask_test_cpu(id, id_seen)) {
|
| 2873 |
|
| 2874 | if (!cpumask_equal(tl->mask(tl, id), tl_cpu_mask))
|
| 2875 | return false;
|
| 2876 | } else {
|
| 2877 |
|
| 2878 | if (cpumask_intersects(tl_cpu_mask, covered))
|
| 2879 | return false;
|
| 2880 |
|
| 2881 | cpumask_or(covered, covered, tl_cpu_mask);
|
| 2882 | cpumask_set_cpu(id, id_seen);
|
| 2883 | }
|
| 2884 | }
|
| 2885 | }
|
| 2886 | return true;
|
| 2887 | }
|
| 2888 |
|
| 2889 | |
| 2890 | * Calculate an allowed NUMA imbalance such that LLCs do not get
|
| 2891 | * imbalanced.
|
| 2892 |
|
| 2893 | static void adjust_numa_imbalance(struct sched_domain *sd_llc)
|
| 2894 | {
|
| 2895 | struct sched_domain *parent;
|
| 2896 | unsigned int imb_span = 1;
|
| 2897 | unsigned int imb = 0;
|
| 2898 | unsigned int nr_llcs;
|
| 2899 |
|
| 2900 | WARN_ON(!(sd_llc->flags & SD_SHARE_LLC));
|
| 2901 | WARN_ON(!sd_llc->parent);
|
| 2902 |
|
| 2903 | |
| 2904 | * For a single LLC per node, allow an
|
| 2905 | * imbalance up to 12.5% of the node. This is
|
| 2906 | * arbitrary cutoff based two factors -- SMT and
|
| 2907 | * memory channels. For SMT-2, the intent is to
|
| 2908 | * avoid premature sharing of HT resources but
|
| 2909 | * SMT-4 or SMT-8 *may* benefit from a different
|
| 2910 | * cutoff. For memory channels, this is a very
|
| 2911 | * rough estimate of how many channels may be
|
| 2912 | * active and is based on recent CPUs with
|
| 2913 | * many cores.
|
| 2914 | *
|
| 2915 | * For multiple LLCs, allow an imbalance
|
| 2916 | * until multiple tasks would share an LLC
|
| 2917 | * on one node while LLCs on another node
|
| 2918 | * remain idle. This assumes that there are
|
| 2919 | * enough logical CPUs per LLC to avoid SMT
|
| 2920 | * factors and that there is a correlation
|
| 2921 | * between LLCs and memory channels.
|
| 2922 |
|
| 2923 | nr_llcs = sd_llc->parent->span_weight / sd_llc->span_weight;
|
| 2924 | if (nr_llcs == 1)
|
| 2925 | imb = sd_llc->parent->span_weight >> 3;
|
| 2926 | else
|
| 2927 | imb = nr_llcs;
|
| 2928 |
|
| 2929 | imb = max(1U, imb);
|
| 2930 | sd_llc->parent->imb_numa_nr = imb;
|
| 2931 |
|
| 2932 | |
| 2933 | * Set span based on the first NUMA domain.
|
| 2934 | *
|
| 2935 | * NUMA systems always add a NODE domain before
|
| 2936 | * iterating the NUMA domains. Since this is before
|
| 2937 | * degeneration, start from sd_llc's parent's
|
| 2938 | * parent which is the lowest an SD_NUMA domain can
|
| 2939 | * be relative to sd_llc.
|
| 2940 |
|
| 2941 | parent = sd_llc->parent->parent;
|
| 2942 | while (parent && !(parent->flags & SD_NUMA))
|
| 2943 | parent = parent->parent;
|
| 2944 |
|
| 2945 | imb_span = parent ? parent->span_weight : sd_llc->parent->span_weight;
|
| 2946 |
|
| 2947 |
|
| 2948 | parent = sd_llc->parent;
|
| 2949 | while (parent) {
|
| 2950 | int factor = max(1U, (parent->span_weight / imb_span));
|
| 2951 |
|
| 2952 | parent->imb_numa_nr = imb * factor;
|
| 2953 | parent = parent->parent;
|
| 2954 | }
|
| 2955 | }
|
| 2956 |
|
| 2957 | static void
|
| 2958 | init_sched_domain_shared(struct s_data *d, struct sched_domain *sd, int flags)
|
| 2959 | {
|
| 2960 | struct sched_domain_shared *sds = NULL;
|
| 2961 | int cpu;
|
| 2962 |
|
| 2963 | |
| 2964 | * Multiple domains can try to claim a shared object like
|
| 2965 | * SD_ASYM_CPUCAPACITY and SD_SHARE_LLC which can alias to
|
| 2966 | * same cpumask_first(sched_domain_span(sd)) CPU and can
|
| 2967 | * cause "nr_idle_scan" to be populated incorrectly during
|
| 2968 | * load balancing.
|
| 2969 | *
|
| 2970 | * Find the first CPU in sched_domain_span(sd) with an
|
| 2971 | * unclaimed domain (!alloc_flags) or where the alloc_flag
|
| 2972 | * matches the requested flag (SD_* flag)
|
| 2973 | *
|
| 2974 | * If the domain only has single CPU, allow temporary overlap
|
| 2975 | * in allocation since the domains will be degenerated later.
|
| 2976 |
|
| 2977 | for_each_cpu(cpu, sched_domain_span(sd)) {
|
| 2978 | sds = *per_cpu_ptr(d->sds, cpu);
|
| 2979 |
|
| 2980 | if (!sds->alloc_flags ||
|
| 2981 | sd->span_weight == 1 ||
|
| 2982 | sds->alloc_flags == flags) {
|
| 2983 | sds->alloc_flags = flags;
|
| 2984 | sd->shared = sds;
|
| 2985 | break;
|
| 2986 | }
|
| 2987 | }
|
| 2988 |
|
| 2989 | |
| 2990 | * Use the sd_shared corresponding to the last
|
| 2991 | * CPU in the span if none are avaialable.
|
| 2992 |
|
| 2993 | if (WARN_ON_ONCE(!sd->shared))
|
| 2994 | sd->shared = sds;
|
| 2995 |
|
| 2996 | |
| 2997 | * nr_busy_cpus is consumed only by the NOHZ kick path via
|
| 2998 | * sd_balance_shared; on the asym-capacity path it is initialized but
|
| 2999 | * never read.
|
| 3000 |
|
| 3001 | atomic_set(&sd->shared->nr_busy_cpus, sd->span_weight);
|
| 3002 | atomic_inc(&sd->shared->ref);
|
| 3003 | }
|
| 3004 |
|
| 3005 | |
| 3006 | * For asymmetric CPU capacity, attach sched_domain_shared on the innermost
|
| 3007 | * SD_ASYM_CPUCAPACITY_FULL ancestor of @cpu's base domain when that ancestor is
|
| 3008 | * not an overlapping NUMA-built domain (then LLC should claim shared).
|
| 3009 | *
|
| 3010 | * A CPU may lack any FULL ancestor (e.g., exclusive cpuset symmetric island),
|
| 3011 | * then LLC must claim shared instead.
|
| 3012 | *
|
| 3013 | * Note: SD_ASYM_CPUCAPACITY_FULL is only set when all CPU capacity values
|
| 3014 | * are present in the domain span, so the asym domain we attach to cannot
|
| 3015 | * degenerate into a single-capacity group. The relevant edge cases are instead
|
| 3016 | * covered by the caveats above.
|
| 3017 | *
|
| 3018 | * Return true if this CPU's asym path claimed sd->shared, false otherwise.
|
| 3019 |
|
| 3020 | static bool claim_asym_sched_domain_shared(struct s_data *d, int cpu)
|
| 3021 | {
|
| 3022 | struct sched_domain *sd = *per_cpu_ptr(d->sd, cpu);
|
| 3023 | struct sched_domain *sd_asym;
|
| 3024 |
|
| 3025 | if (!sd)
|
| 3026 | return false;
|
| 3027 |
|
| 3028 | sd_asym = sd;
|
| 3029 | while (sd_asym && !(sd_asym->flags & SD_ASYM_CPUCAPACITY_FULL))
|
| 3030 | sd_asym = sd_asym->parent;
|
| 3031 |
|
| 3032 | if (!sd_asym || (sd_asym->flags & SD_NUMA))
|
| 3033 | return false;
|
| 3034 |
|
| 3035 | init_sched_domain_shared(d, sd_asym, SD_ASYM_CPUCAPACITY);
|
| 3036 | return true;
|
| 3037 | }
|
| 3038 |
|
| 3039 | static int __sched_domains_alloc_llc_id(void)
|
| 3040 | {
|
| 3041 | int lid, max;
|
| 3042 |
|
| 3043 | lockdep_assert_held(&sched_domains_mutex);
|
| 3044 |
|
| 3045 | lid = cpumask_first_zero(sched_domains_llc_id_allocmask);
|
| 3046 | |
| 3047 | * llc_id space should never grow larger than the
|
| 3048 | * possible number of CPUs in the system.
|
| 3049 |
|
| 3050 | if (lid >= nr_cpu_ids)
|
| 3051 | return -1;
|
| 3052 |
|
| 3053 | __cpumask_set_cpu(lid, sched_domains_llc_id_allocmask);
|
| 3054 | max = cpumask_last(sched_domains_llc_id_allocmask);
|
| 3055 | if (max > max_lid)
|
| 3056 | max_lid = max;
|
| 3057 |
|
| 3058 | return lid;
|
| 3059 | }
|
| 3060 |
|
| 3061 | static void __sched_domains_free_llc_id(int cpu)
|
| 3062 | {
|
| 3063 | int i, lid, max;
|
| 3064 |
|
| 3065 | lockdep_assert_held(&sched_domains_mutex);
|
| 3066 |
|
| 3067 | lid = per_cpu(sd_llc_id, cpu);
|
| 3068 | if (lid == -1 || lid >= nr_cpu_ids)
|
| 3069 | return;
|
| 3070 |
|
| 3071 | per_cpu(sd_llc_id, cpu) = -1;
|
| 3072 |
|
| 3073 | for_each_cpu(i, llc_mask(cpu)) {
|
| 3074 |
|
| 3075 | if (per_cpu(sd_llc_id, i) == lid)
|
| 3076 | return;
|
| 3077 | }
|
| 3078 |
|
| 3079 | __cpumask_clear_cpu(lid, sched_domains_llc_id_allocmask);
|
| 3080 |
|
| 3081 | max = cpumask_last(sched_domains_llc_id_allocmask);
|
| 3082 |
|
| 3083 | if (max < max_lid)
|
| 3084 | max_lid = max;
|
| 3085 | }
|
| 3086 |
|
| 3087 | void sched_domains_free_llc_id(int cpu)
|
| 3088 | {
|
| 3089 | sched_domains_mutex_lock();
|
| 3090 | __sched_domains_free_llc_id(cpu);
|
| 3091 | sched_domains_mutex_unlock();
|
| 3092 | }
|
| 3093 |
|
| 3094 | |
| 3095 | * Build sched domains for a given set of CPUs and attach the sched domains
|
| 3096 | * to the individual CPUs
|
| 3097 |
|
| 3098 | static int
|
| 3099 | build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr,
|
| 3100 | bool *multi_llcs)
|
| 3101 | {
|
| 3102 | enum s_alloc alloc_state = sa_none;
|
| 3103 | bool has_multi_llcs = false;
|
| 3104 | struct sched_domain *sd;
|
| 3105 | struct s_data d;
|
| 3106 | struct rq *rq = NULL;
|
| 3107 | int i, ret = -ENOMEM;
|
| 3108 | bool has_asym = false;
|
| 3109 | bool has_cluster = false;
|
| 3110 |
|
| 3111 | if (WARN_ON(cpumask_empty(cpu_map)))
|
| 3112 | goto error;
|
| 3113 |
|
| 3114 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
|
| 3115 | if (alloc_state != sa_rootdomain)
|
| 3116 | goto error;
|
| 3117 |
|
| 3118 |
|
| 3119 | for_each_cpu(i, cpu_map) {
|
| 3120 | struct sched_domain_topology_level *tl;
|
| 3121 | int lid;
|
| 3122 |
|
| 3123 | sd = NULL;
|
| 3124 | for_each_sd_topology(tl) {
|
| 3125 |
|
| 3126 | sd = build_sched_domain(tl, cpu_map, attr, sd, i);
|
| 3127 |
|
| 3128 | has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
|
| 3129 |
|
| 3130 | if (tl == sched_domain_topology)
|
| 3131 | *per_cpu_ptr(d.sd, i) = sd;
|
| 3132 | if (cpumask_equal(cpu_map, sched_domain_span(sd)))
|
| 3133 | break;
|
| 3134 | }
|
| 3135 |
|
| 3136 | lid = per_cpu(sd_llc_id, i);
|
| 3137 | if (lid == -1) {
|
| 3138 |
|
| 3139 | for (int j = cpumask_first(llc_mask(i));
|
| 3140 | j < nr_cpu_ids;
|
| 3141 | j = cpumask_next(j, llc_mask(i))) {
|
| 3142 | if (i == j)
|
| 3143 | continue;
|
| 3144 |
|
| 3145 | lid = per_cpu(sd_llc_id, j);
|
| 3146 |
|
| 3147 | if (lid != -1) {
|
| 3148 | per_cpu(sd_llc_id, i) = lid;
|
| 3149 |
|
| 3150 | break;
|
| 3151 | }
|
| 3152 | }
|
| 3153 |
|
| 3154 |
|
| 3155 | if (lid == -1)
|
| 3156 | per_cpu(sd_llc_id, i) = __sched_domains_alloc_llc_id();
|
| 3157 | }
|
| 3158 | }
|
| 3159 |
|
| 3160 | if (WARN_ON(!topology_span_sane(cpu_map)))
|
| 3161 | goto error;
|
| 3162 |
|
| 3163 |
|
| 3164 | for_each_cpu(i, cpu_map) {
|
| 3165 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
|
| 3166 | sd->span_weight = cpumask_weight(sched_domain_span(sd));
|
| 3167 | if (sd->flags & SD_NUMA) {
|
| 3168 | if (build_overlap_sched_groups(sd, i))
|
| 3169 | goto error;
|
| 3170 | } else {
|
| 3171 | if (build_sched_groups(sd, i))
|
| 3172 | goto error;
|
| 3173 | }
|
| 3174 | }
|
| 3175 | }
|
| 3176 |
|
| 3177 | for_each_cpu(i, cpu_map) {
|
| 3178 | sd = *per_cpu_ptr(d.sd, i);
|
| 3179 | if (!sd)
|
| 3180 | continue;
|
| 3181 |
|
| 3182 | if (has_asym)
|
| 3183 | claim_asym_sched_domain_shared(&d, i);
|
| 3184 |
|
| 3185 |
|
| 3186 | while (sd->parent && (sd->parent->flags & SD_SHARE_LLC))
|
| 3187 | sd = sd->parent;
|
| 3188 |
|
| 3189 | if (sd->flags & SD_SHARE_LLC) {
|
| 3190 | init_sched_domain_shared(&d, sd, SD_SHARE_LLC);
|
| 3191 |
|
| 3192 | |
| 3193 | * In presence of higher domains, adjust the
|
| 3194 | * NUMA imbalance stats for the hierarchy.
|
| 3195 |
|
| 3196 | if (sd->parent) {
|
| 3197 | if (IS_ENABLED(CONFIG_NUMA))
|
| 3198 | adjust_numa_imbalance(sd);
|
| 3199 |
|
| 3200 | if (sd_in_multi_llcs(sd))
|
| 3201 | has_multi_llcs = true;
|
| 3202 | }
|
| 3203 | }
|
| 3204 | }
|
| 3205 |
|
| 3206 |
|
| 3207 | for (i = nr_cpumask_bits-1; i >= 0; i--) {
|
| 3208 | if (!cpumask_test_cpu(i, cpu_map))
|
| 3209 | continue;
|
| 3210 |
|
| 3211 | claim_allocations(i, &d);
|
| 3212 |
|
| 3213 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent)
|
| 3214 | init_sched_groups_capacity(i, sd);
|
| 3215 | }
|
| 3216 |
|
| 3217 | alloc_sd_llc(cpu_map, &d);
|
| 3218 |
|
| 3219 |
|
| 3220 | rcu_read_lock();
|
| 3221 | for_each_cpu(i, cpu_map) {
|
| 3222 | rq = cpu_rq(i);
|
| 3223 | sd = *per_cpu_ptr(d.sd, i);
|
| 3224 |
|
| 3225 | cpu_attach_domain(sd, d.rd, i);
|
| 3226 |
|
| 3227 | if (lowest_flag_domain(i, SD_CLUSTER))
|
| 3228 | has_cluster = true;
|
| 3229 | }
|
| 3230 | rcu_read_unlock();
|
| 3231 |
|
| 3232 | if (has_asym)
|
| 3233 | static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
|
| 3234 |
|
| 3235 | if (has_cluster)
|
| 3236 | static_branch_inc_cpuslocked(&sched_cluster_active);
|
| 3237 |
|
| 3238 | if (rq && sched_debug_verbose)
|
| 3239 | pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
|
| 3240 |
|
| 3241 | ret = 0;
|
| 3242 | error:
|
| 3243 | *multi_llcs = has_multi_llcs;
|
| 3244 | __free_domain_allocs(&d, alloc_state, cpu_map);
|
| 3245 |
|
| 3246 | return ret;
|
| 3247 | }
|
| 3248 |
|
| 3249 |
|
| 3250 | static cpumask_var_t *doms_cur;
|
| 3251 |
|
| 3252 |
|
| 3253 | static int ndoms_cur;
|
| 3254 |
|
| 3255 |
|
| 3256 | static struct sched_domain_attr *dattr_cur;
|
| 3257 |
|
| 3258 | |
| 3259 | * Special case: If a kmalloc() of a doms_cur partition (array of
|
| 3260 | * cpumask) fails, then fallback to a single sched domain,
|
| 3261 | * as determined by the single cpumask fallback_doms.
|
| 3262 |
|
| 3263 | static cpumask_var_t fallback_doms;
|
| 3264 |
|
| 3265 | |
| 3266 | * arch_update_cpu_topology lets virtualized architectures update the
|
| 3267 | * CPU core maps. It is supposed to return 1 if the topology changed
|
| 3268 | * or 0 if it stayed the same.
|
| 3269 |
|
| 3270 | int __weak arch_update_cpu_topology(void)
|
| 3271 | {
|
| 3272 | return 0;
|
| 3273 | }
|
| 3274 |
|
| 3275 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
|
| 3276 | {
|
| 3277 | int i;
|
| 3278 | cpumask_var_t *doms;
|
| 3279 |
|
| 3280 | doms = kmalloc_objs(*doms, ndoms);
|
| 3281 | if (!doms)
|
| 3282 | return NULL;
|
| 3283 | for (i = 0; i < ndoms; i++) {
|
| 3284 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
|
| 3285 | free_sched_domains(doms, i);
|
| 3286 | return NULL;
|
| 3287 | }
|
| 3288 | }
|
| 3289 | return doms;
|
| 3290 | }
|
| 3291 |
|
| 3292 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
|
| 3293 | {
|
| 3294 | unsigned int i;
|
| 3295 | for (i = 0; i < ndoms; i++)
|
| 3296 | free_cpumask_var(doms[i]);
|
| 3297 | kfree(doms);
|
| 3298 | }
|
| 3299 |
|
| 3300 | |
| 3301 | * Set up scheduler domains and groups. For now this just excludes isolated
|
| 3302 | * CPUs, but could be used to exclude other special cases in the future.
|
| 3303 |
|
| 3304 | int __init sched_init_domains(const struct cpumask *cpu_map)
|
| 3305 | {
|
| 3306 | bool multi_llcs;
|
| 3307 | int err;
|
| 3308 |
|
| 3309 | zalloc_cpumask_var(&sched_domains_llc_id_allocmask, GFP_KERNEL);
|
| 3310 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
|
| 3311 | zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
|
| 3312 | zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
|
| 3313 |
|
| 3314 | arch_update_cpu_topology();
|
| 3315 | asym_cpu_capacity_scan();
|
| 3316 | ndoms_cur = 1;
|
| 3317 | doms_cur = alloc_sched_domains(ndoms_cur);
|
| 3318 | if (!doms_cur)
|
| 3319 | doms_cur = &fallback_doms;
|
| 3320 | cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
|
| 3321 | err = build_sched_domains(doms_cur[0], NULL, &multi_llcs);
|
| 3322 | if (!err)
|
| 3323 | sched_cache_set(multi_llcs);
|
| 3324 |
|
| 3325 | return err;
|
| 3326 | }
|
| 3327 |
|
| 3328 | |
| 3329 | * Detach sched domains from a group of CPUs specified in cpu_map
|
| 3330 | * These CPUs will now be attached to the NULL domain
|
| 3331 |
|
| 3332 | static void detach_destroy_domains(const struct cpumask *cpu_map)
|
| 3333 | {
|
| 3334 | unsigned int cpu = cpumask_any(cpu_map);
|
| 3335 | int i;
|
| 3336 |
|
| 3337 | if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
|
| 3338 | static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
|
| 3339 |
|
| 3340 | if (static_branch_unlikely(&sched_cluster_active))
|
| 3341 | static_branch_dec_cpuslocked(&sched_cluster_active);
|
| 3342 |
|
| 3343 | rcu_read_lock();
|
| 3344 | for_each_cpu(i, cpu_map)
|
| 3345 | cpu_attach_domain(NULL, &def_root_domain, i);
|
| 3346 | rcu_read_unlock();
|
| 3347 | }
|
| 3348 |
|
| 3349 |
|
| 3350 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
|
| 3351 | struct sched_domain_attr *new, int idx_new)
|
| 3352 | {
|
| 3353 | struct sched_domain_attr tmp;
|
| 3354 |
|
| 3355 |
|
| 3356 | if (!new && !cur)
|
| 3357 | return 1;
|
| 3358 |
|
| 3359 | tmp = SD_ATTR_INIT;
|
| 3360 |
|
| 3361 | return !memcmp(cur ? (cur + idx_cur) : &tmp,
|
| 3362 | new ? (new + idx_new) : &tmp,
|
| 3363 | sizeof(struct sched_domain_attr));
|
| 3364 | }
|
| 3365 |
|
| 3366 | |
| 3367 | * Partition sched domains as specified by the 'ndoms_new'
|
| 3368 | * cpumasks in the array doms_new[] of cpumasks. This compares
|
| 3369 | * doms_new[] to the current sched domain partitioning, doms_cur[].
|
| 3370 | * It destroys each deleted domain and builds each new domain.
|
| 3371 | *
|
| 3372 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
|
| 3373 | * The masks don't intersect (don't overlap.) We should setup one
|
| 3374 | * sched domain for each mask. CPUs not in any of the cpumasks will
|
| 3375 | * not be load balanced. If the same cpumask appears both in the
|
| 3376 | * current 'doms_cur' domains and in the new 'doms_new', we can leave
|
| 3377 | * it as it is.
|
| 3378 | *
|
| 3379 | * The passed in 'doms_new' should be allocated using
|
| 3380 | * alloc_sched_domains. This routine takes ownership of it and will
|
| 3381 | * free_sched_domains it when done with it. If the caller failed the
|
| 3382 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
|
| 3383 | * and partition_sched_domains() will fallback to the single partition
|
| 3384 | * 'fallback_doms', it also forces the domains to be rebuilt.
|
| 3385 | *
|
| 3386 | * If doms_new == NULL it will be replaced with cpu_online_mask.
|
| 3387 | * ndoms_new == 0 is a special case for destroying existing domains,
|
| 3388 | * and it will not create the default domain.
|
| 3389 | *
|
| 3390 | * Call with hotplug lock and sched_domains_mutex held
|
| 3391 |
|
| 3392 | static void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
|
| 3393 | struct sched_domain_attr *dattr_new)
|
| 3394 | {
|
| 3395 | bool __maybe_unused has_eas = false;
|
| 3396 | bool has_multi_llcs = false, multi_llcs;
|
| 3397 | int i, j, n;
|
| 3398 | int new_topology;
|
| 3399 |
|
| 3400 | lockdep_assert_held(&sched_domains_mutex);
|
| 3401 |
|
| 3402 |
|
| 3403 | new_topology = arch_update_cpu_topology();
|
| 3404 |
|
| 3405 | if (new_topology)
|
| 3406 | asym_cpu_capacity_scan();
|
| 3407 |
|
| 3408 | if (!doms_new) {
|
| 3409 | WARN_ON_ONCE(dattr_new);
|
| 3410 | n = 0;
|
| 3411 | doms_new = alloc_sched_domains(1);
|
| 3412 | if (doms_new) {
|
| 3413 | n = 1;
|
| 3414 | cpumask_and(doms_new[0], cpu_active_mask,
|
| 3415 | housekeeping_cpumask(HK_TYPE_DOMAIN));
|
| 3416 | }
|
| 3417 | } else {
|
| 3418 | n = ndoms_new;
|
| 3419 | }
|
| 3420 |
|
| 3421 |
|
| 3422 | for (i = 0; i < ndoms_cur; i++) {
|
| 3423 | for (j = 0; j < n && !new_topology; j++) {
|
| 3424 | if (cpumask_equal(doms_cur[i], doms_new[j]) &&
|
| 3425 | dattrs_equal(dattr_cur, i, dattr_new, j))
|
| 3426 | goto match1;
|
| 3427 | }
|
| 3428 |
|
| 3429 | detach_destroy_domains(doms_cur[i]);
|
| 3430 | match1:
|
| 3431 | ;
|
| 3432 | }
|
| 3433 |
|
| 3434 | n = ndoms_cur;
|
| 3435 | if (!doms_new) {
|
| 3436 | n = 0;
|
| 3437 | doms_new = &fallback_doms;
|
| 3438 | cpumask_and(doms_new[0], cpu_active_mask,
|
| 3439 | housekeeping_cpumask(HK_TYPE_DOMAIN));
|
| 3440 | }
|
| 3441 |
|
| 3442 |
|
| 3443 | for (i = 0; i < ndoms_new; i++) {
|
| 3444 | for (j = 0; j < n && !new_topology; j++) {
|
| 3445 | if (cpumask_equal(doms_new[i], doms_cur[j]) &&
|
| 3446 | dattrs_equal(dattr_new, i, dattr_cur, j)) {
|
| 3447 | |
| 3448 | * Reused partition has to be taken care
|
| 3449 | * of here, because there could be a corner
|
| 3450 | * case that if the reused partition is skipped
|
| 3451 | * and only new partition is considered, an
|
| 3452 | * incorrect has_multi_llcs would be set. For
|
| 3453 | * example:
|
| 3454 | * If the only multi-LLC partition is reused
|
| 3455 | * and a new single-LLC partition is built,
|
| 3456 | * sched_cache_set(false) disables cache-aware
|
| 3457 | * scheduling globally despite the reused
|
| 3458 | * multi-LLC partition still being active.
|
| 3459 |
|
| 3460 | struct sched_domain *sd;
|
| 3461 | int cpu = cpumask_first(doms_cur[j]);
|
| 3462 |
|
| 3463 | guard(rcu)();
|
| 3464 | sd = rcu_dereference(cpu_rq(cpu)->sd);
|
| 3465 | while (sd && sd->parent && (sd->parent->flags & SD_SHARE_LLC))
|
| 3466 | sd = sd->parent;
|
| 3467 | if (sd && (sd->flags & SD_SHARE_LLC) && sd->parent &&
|
| 3468 | sd_in_multi_llcs(sd))
|
| 3469 | has_multi_llcs = true;
|
| 3470 | goto match2;
|
| 3471 | }
|
| 3472 | }
|
| 3473 |
|
| 3474 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL,
|
| 3475 | &multi_llcs);
|
| 3476 | has_multi_llcs |= multi_llcs;
|
| 3477 | match2:
|
| 3478 | ;
|
| 3479 | }
|
| 3480 | sched_cache_set(has_multi_llcs);
|
| 3481 |
|
| 3482 | #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
|
| 3483 |
|
| 3484 | for (i = 0; i < ndoms_new; i++) {
|
| 3485 | for (j = 0; j < n && !sched_energy_update; j++) {
|
| 3486 | if (cpumask_equal(doms_new[i], doms_cur[j]) &&
|
| 3487 | cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
|
| 3488 | has_eas = true;
|
| 3489 | goto match3;
|
| 3490 | }
|
| 3491 | }
|
| 3492 |
|
| 3493 | has_eas |= build_perf_domains(doms_new[i]);
|
| 3494 | match3:
|
| 3495 | ;
|
| 3496 | }
|
| 3497 | sched_energy_set(has_eas);
|
| 3498 | #endif
|
| 3499 |
|
| 3500 |
|
| 3501 | if (doms_cur != &fallback_doms)
|
| 3502 | free_sched_domains(doms_cur, ndoms_cur);
|
| 3503 |
|
| 3504 | kfree(dattr_cur);
|
| 3505 | doms_cur = doms_new;
|
| 3506 | dattr_cur = dattr_new;
|
| 3507 | ndoms_cur = ndoms_new;
|
| 3508 |
|
| 3509 | update_sched_domain_debugfs();
|
| 3510 | dl_rebuild_rd_accounting();
|
| 3511 | }
|
| 3512 |
|
| 3513 | |
| 3514 | * Call with hotplug lock held
|
| 3515 |
|
| 3516 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
|
| 3517 | struct sched_domain_attr *dattr_new)
|
| 3518 | {
|
| 3519 | sched_domains_mutex_lock();
|
| 3520 | partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
|
| 3521 | sched_domains_mutex_unlock();
|
| 3522 | }
|