Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
f2cb1360 IM |
2 | /* |
3 | * Scheduler topology setup/handling methods | |
4 | */ | |
f2cb1360 IM |
5 | #include "sched.h" |
6 | ||
7 | DEFINE_MUTEX(sched_domains_mutex); | |
8 | ||
9 | /* Protected by sched_domains_mutex: */ | |
ace80310 | 10 | static cpumask_var_t sched_domains_tmpmask; |
11 | static cpumask_var_t sched_domains_tmpmask2; | |
f2cb1360 IM |
12 | |
13 | #ifdef CONFIG_SCHED_DEBUG | |
14 | ||
f2cb1360 IM |
15 | static int __init sched_debug_setup(char *str) |
16 | { | |
9469eb01 | 17 | sched_debug_enabled = true; |
f2cb1360 IM |
18 | |
19 | return 0; | |
20 | } | |
21 | early_param("sched_debug", sched_debug_setup); | |
22 | ||
23 | static inline bool sched_debug(void) | |
24 | { | |
25 | return sched_debug_enabled; | |
26 | } | |
27 | ||
28 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, | |
29 | struct cpumask *groupmask) | |
30 | { | |
31 | struct sched_group *group = sd->groups; | |
32 | ||
33 | cpumask_clear(groupmask); | |
34 | ||
005f874d | 35 | printk(KERN_DEBUG "%*s domain-%d: ", level, "", level); |
f2cb1360 IM |
36 | |
37 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
38 | printk("does not load-balance\n"); | |
39 | if (sd->parent) | |
97fb7a0a | 40 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent"); |
f2cb1360 IM |
41 | return -1; |
42 | } | |
43 | ||
005f874d | 44 | printk(KERN_CONT "span=%*pbl level=%s\n", |
f2cb1360 IM |
45 | cpumask_pr_args(sched_domain_span(sd)), sd->name); |
46 | ||
47 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
97fb7a0a | 48 | printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu); |
f2cb1360 | 49 | } |
6cd0c583 | 50 | if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) { |
97fb7a0a | 51 | printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu); |
f2cb1360 IM |
52 | } |
53 | ||
54 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); | |
55 | do { | |
56 | if (!group) { | |
57 | printk("\n"); | |
58 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
59 | break; | |
60 | } | |
61 | ||
ae4df9d6 | 62 | if (!cpumask_weight(sched_group_span(group))) { |
f2cb1360 IM |
63 | printk(KERN_CONT "\n"); |
64 | printk(KERN_ERR "ERROR: empty group\n"); | |
65 | break; | |
66 | } | |
67 | ||
68 | if (!(sd->flags & SD_OVERLAP) && | |
ae4df9d6 | 69 | cpumask_intersects(groupmask, sched_group_span(group))) { |
f2cb1360 IM |
70 | printk(KERN_CONT "\n"); |
71 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
72 | break; | |
73 | } | |
74 | ||
ae4df9d6 | 75 | cpumask_or(groupmask, groupmask, sched_group_span(group)); |
f2cb1360 | 76 | |
005f874d PZ |
77 | printk(KERN_CONT " %d:{ span=%*pbl", |
78 | group->sgc->id, | |
ae4df9d6 | 79 | cpumask_pr_args(sched_group_span(group))); |
b0151c25 | 80 | |
af218122 | 81 | if ((sd->flags & SD_OVERLAP) && |
ae4df9d6 | 82 | !cpumask_equal(group_balance_mask(group), sched_group_span(group))) { |
005f874d | 83 | printk(KERN_CONT " mask=%*pbl", |
e5c14b1f | 84 | cpumask_pr_args(group_balance_mask(group))); |
b0151c25 PZ |
85 | } |
86 | ||
005f874d PZ |
87 | if (group->sgc->capacity != SCHED_CAPACITY_SCALE) |
88 | printk(KERN_CONT " cap=%lu", group->sgc->capacity); | |
f2cb1360 | 89 | |
a420b063 PZ |
90 | if (group == sd->groups && sd->child && |
91 | !cpumask_equal(sched_domain_span(sd->child), | |
ae4df9d6 | 92 | sched_group_span(group))) { |
a420b063 PZ |
93 | printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n"); |
94 | } | |
95 | ||
005f874d PZ |
96 | printk(KERN_CONT " }"); |
97 | ||
f2cb1360 | 98 | group = group->next; |
b0151c25 PZ |
99 | |
100 | if (group != sd->groups) | |
101 | printk(KERN_CONT ","); | |
102 | ||
f2cb1360 IM |
103 | } while (group != sd->groups); |
104 | printk(KERN_CONT "\n"); | |
105 | ||
106 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) | |
107 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | |
108 | ||
109 | if (sd->parent && | |
110 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) | |
97fb7a0a | 111 | printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n"); |
f2cb1360 IM |
112 | return 0; |
113 | } | |
114 | ||
115 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | |
116 | { | |
117 | int level = 0; | |
118 | ||
119 | if (!sched_debug_enabled) | |
120 | return; | |
121 | ||
122 | if (!sd) { | |
123 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
124 | return; | |
125 | } | |
126 | ||
005f874d | 127 | printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu); |
f2cb1360 IM |
128 | |
129 | for (;;) { | |
130 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) | |
131 | break; | |
132 | level++; | |
133 | sd = sd->parent; | |
134 | if (!sd) | |
135 | break; | |
136 | } | |
137 | } | |
138 | #else /* !CONFIG_SCHED_DEBUG */ | |
139 | ||
140 | # define sched_debug_enabled 0 | |
141 | # define sched_domain_debug(sd, cpu) do { } while (0) | |
142 | static inline bool sched_debug(void) | |
143 | { | |
144 | return false; | |
145 | } | |
146 | #endif /* CONFIG_SCHED_DEBUG */ | |
147 | ||
148 | static int sd_degenerate(struct sched_domain *sd) | |
149 | { | |
150 | if (cpumask_weight(sched_domain_span(sd)) == 1) | |
151 | return 1; | |
152 | ||
153 | /* Following flags need at least 2 groups */ | |
154 | if (sd->flags & (SD_LOAD_BALANCE | | |
155 | SD_BALANCE_NEWIDLE | | |
156 | SD_BALANCE_FORK | | |
157 | SD_BALANCE_EXEC | | |
158 | SD_SHARE_CPUCAPACITY | | |
159 | SD_ASYM_CPUCAPACITY | | |
160 | SD_SHARE_PKG_RESOURCES | | |
161 | SD_SHARE_POWERDOMAIN)) { | |
162 | if (sd->groups != sd->groups->next) | |
163 | return 0; | |
164 | } | |
165 | ||
166 | /* Following flags don't use groups */ | |
167 | if (sd->flags & (SD_WAKE_AFFINE)) | |
168 | return 0; | |
169 | ||
170 | return 1; | |
171 | } | |
172 | ||
173 | static int | |
174 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
175 | { | |
176 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
177 | ||
178 | if (sd_degenerate(parent)) | |
179 | return 1; | |
180 | ||
181 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) | |
182 | return 0; | |
183 | ||
184 | /* Flags needing groups don't count if only 1 group in parent */ | |
185 | if (parent->groups == parent->groups->next) { | |
186 | pflags &= ~(SD_LOAD_BALANCE | | |
187 | SD_BALANCE_NEWIDLE | | |
188 | SD_BALANCE_FORK | | |
189 | SD_BALANCE_EXEC | | |
190 | SD_ASYM_CPUCAPACITY | | |
191 | SD_SHARE_CPUCAPACITY | | |
192 | SD_SHARE_PKG_RESOURCES | | |
193 | SD_PREFER_SIBLING | | |
194 | SD_SHARE_POWERDOMAIN); | |
195 | if (nr_node_ids == 1) | |
196 | pflags &= ~SD_SERIALIZE; | |
197 | } | |
198 | if (~cflags & pflags) | |
199 | return 0; | |
200 | ||
201 | return 1; | |
202 | } | |
203 | ||
204 | static void free_rootdomain(struct rcu_head *rcu) | |
205 | { | |
206 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu); | |
207 | ||
208 | cpupri_cleanup(&rd->cpupri); | |
209 | cpudl_cleanup(&rd->cpudl); | |
210 | free_cpumask_var(rd->dlo_mask); | |
211 | free_cpumask_var(rd->rto_mask); | |
212 | free_cpumask_var(rd->online); | |
213 | free_cpumask_var(rd->span); | |
214 | kfree(rd); | |
215 | } | |
216 | ||
217 | void rq_attach_root(struct rq *rq, struct root_domain *rd) | |
218 | { | |
219 | struct root_domain *old_rd = NULL; | |
220 | unsigned long flags; | |
221 | ||
222 | raw_spin_lock_irqsave(&rq->lock, flags); | |
223 | ||
224 | if (rq->rd) { | |
225 | old_rd = rq->rd; | |
226 | ||
227 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) | |
228 | set_rq_offline(rq); | |
229 | ||
230 | cpumask_clear_cpu(rq->cpu, old_rd->span); | |
231 | ||
232 | /* | |
233 | * If we dont want to free the old_rd yet then | |
234 | * set old_rd to NULL to skip the freeing later | |
235 | * in this function: | |
236 | */ | |
237 | if (!atomic_dec_and_test(&old_rd->refcount)) | |
238 | old_rd = NULL; | |
239 | } | |
240 | ||
241 | atomic_inc(&rd->refcount); | |
242 | rq->rd = rd; | |
243 | ||
244 | cpumask_set_cpu(rq->cpu, rd->span); | |
245 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) | |
246 | set_rq_online(rq); | |
247 | ||
248 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
249 | ||
250 | if (old_rd) | |
251 | call_rcu_sched(&old_rd->rcu, free_rootdomain); | |
252 | } | |
253 | ||
364f5665 SRV |
254 | void sched_get_rd(struct root_domain *rd) |
255 | { | |
256 | atomic_inc(&rd->refcount); | |
257 | } | |
258 | ||
259 | void sched_put_rd(struct root_domain *rd) | |
260 | { | |
261 | if (!atomic_dec_and_test(&rd->refcount)) | |
262 | return; | |
263 | ||
264 | call_rcu_sched(&rd->rcu, free_rootdomain); | |
265 | } | |
266 | ||
f2cb1360 IM |
267 | static int init_rootdomain(struct root_domain *rd) |
268 | { | |
f2cb1360 IM |
269 | if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL)) |
270 | goto out; | |
271 | if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL)) | |
272 | goto free_span; | |
273 | if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) | |
274 | goto free_online; | |
275 | if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) | |
276 | goto free_dlo_mask; | |
277 | ||
4bdced5c SRRH |
278 | #ifdef HAVE_RT_PUSH_IPI |
279 | rd->rto_cpu = -1; | |
280 | raw_spin_lock_init(&rd->rto_lock); | |
281 | init_irq_work(&rd->rto_push_work, rto_push_irq_work_func); | |
282 | #endif | |
283 | ||
f2cb1360 IM |
284 | init_dl_bw(&rd->dl_bw); |
285 | if (cpudl_init(&rd->cpudl) != 0) | |
286 | goto free_rto_mask; | |
287 | ||
288 | if (cpupri_init(&rd->cpupri) != 0) | |
289 | goto free_cpudl; | |
290 | return 0; | |
291 | ||
292 | free_cpudl: | |
293 | cpudl_cleanup(&rd->cpudl); | |
294 | free_rto_mask: | |
295 | free_cpumask_var(rd->rto_mask); | |
296 | free_dlo_mask: | |
297 | free_cpumask_var(rd->dlo_mask); | |
298 | free_online: | |
299 | free_cpumask_var(rd->online); | |
300 | free_span: | |
301 | free_cpumask_var(rd->span); | |
302 | out: | |
303 | return -ENOMEM; | |
304 | } | |
305 | ||
306 | /* | |
307 | * By default the system creates a single root-domain with all CPUs as | |
308 | * members (mimicking the global state we have today). | |
309 | */ | |
310 | struct root_domain def_root_domain; | |
311 | ||
312 | void init_defrootdomain(void) | |
313 | { | |
314 | init_rootdomain(&def_root_domain); | |
315 | ||
316 | atomic_set(&def_root_domain.refcount, 1); | |
317 | } | |
318 | ||
319 | static struct root_domain *alloc_rootdomain(void) | |
320 | { | |
321 | struct root_domain *rd; | |
322 | ||
4d13a06d | 323 | rd = kzalloc(sizeof(*rd), GFP_KERNEL); |
f2cb1360 IM |
324 | if (!rd) |
325 | return NULL; | |
326 | ||
327 | if (init_rootdomain(rd) != 0) { | |
328 | kfree(rd); | |
329 | return NULL; | |
330 | } | |
331 | ||
332 | return rd; | |
333 | } | |
334 | ||
335 | static void free_sched_groups(struct sched_group *sg, int free_sgc) | |
336 | { | |
337 | struct sched_group *tmp, *first; | |
338 | ||
339 | if (!sg) | |
340 | return; | |
341 | ||
342 | first = sg; | |
343 | do { | |
344 | tmp = sg->next; | |
345 | ||
346 | if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) | |
347 | kfree(sg->sgc); | |
348 | ||
213c5a45 SW |
349 | if (atomic_dec_and_test(&sg->ref)) |
350 | kfree(sg); | |
f2cb1360 IM |
351 | sg = tmp; |
352 | } while (sg != first); | |
353 | } | |
354 | ||
355 | static void destroy_sched_domain(struct sched_domain *sd) | |
356 | { | |
357 | /* | |
a090c4f2 PZ |
358 | * A normal sched domain may have multiple group references, an |
359 | * overlapping domain, having private groups, only one. Iterate, | |
360 | * dropping group/capacity references, freeing where none remain. | |
f2cb1360 | 361 | */ |
213c5a45 SW |
362 | free_sched_groups(sd->groups, 1); |
363 | ||
f2cb1360 IM |
364 | if (sd->shared && atomic_dec_and_test(&sd->shared->ref)) |
365 | kfree(sd->shared); | |
366 | kfree(sd); | |
367 | } | |
368 | ||
369 | static void destroy_sched_domains_rcu(struct rcu_head *rcu) | |
370 | { | |
371 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); | |
372 | ||
373 | while (sd) { | |
374 | struct sched_domain *parent = sd->parent; | |
375 | destroy_sched_domain(sd); | |
376 | sd = parent; | |
377 | } | |
378 | } | |
379 | ||
380 | static void destroy_sched_domains(struct sched_domain *sd) | |
381 | { | |
382 | if (sd) | |
383 | call_rcu(&sd->rcu, destroy_sched_domains_rcu); | |
384 | } | |
385 | ||
386 | /* | |
387 | * Keep a special pointer to the highest sched_domain that has | |
388 | * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this | |
389 | * allows us to avoid some pointer chasing select_idle_sibling(). | |
390 | * | |
391 | * Also keep a unique ID per domain (we use the first CPU number in | |
392 | * the cpumask of the domain), this allows us to quickly tell if | |
393 | * two CPUs are in the same cache domain, see cpus_share_cache(). | |
394 | */ | |
395 | DEFINE_PER_CPU(struct sched_domain *, sd_llc); | |
396 | DEFINE_PER_CPU(int, sd_llc_size); | |
397 | DEFINE_PER_CPU(int, sd_llc_id); | |
398 | DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared); | |
399 | DEFINE_PER_CPU(struct sched_domain *, sd_numa); | |
400 | DEFINE_PER_CPU(struct sched_domain *, sd_asym); | |
df054e84 | 401 | DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity); |
f2cb1360 IM |
402 | |
403 | static void update_top_cache_domain(int cpu) | |
404 | { | |
405 | struct sched_domain_shared *sds = NULL; | |
406 | struct sched_domain *sd; | |
407 | int id = cpu; | |
408 | int size = 1; | |
409 | ||
410 | sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); | |
411 | if (sd) { | |
412 | id = cpumask_first(sched_domain_span(sd)); | |
413 | size = cpumask_weight(sched_domain_span(sd)); | |
414 | sds = sd->shared; | |
415 | } | |
416 | ||
417 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); | |
418 | per_cpu(sd_llc_size, cpu) = size; | |
419 | per_cpu(sd_llc_id, cpu) = id; | |
420 | rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds); | |
421 | ||
422 | sd = lowest_flag_domain(cpu, SD_NUMA); | |
423 | rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); | |
424 | ||
425 | sd = highest_flag_domain(cpu, SD_ASYM_PACKING); | |
426 | rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); | |
427 | } | |
428 | ||
429 | /* | |
430 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | |
431 | * hold the hotplug lock. | |
432 | */ | |
433 | static void | |
434 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
435 | { | |
436 | struct rq *rq = cpu_rq(cpu); | |
437 | struct sched_domain *tmp; | |
438 | ||
439 | /* Remove the sched domains which do not contribute to scheduling. */ | |
440 | for (tmp = sd; tmp; ) { | |
441 | struct sched_domain *parent = tmp->parent; | |
442 | if (!parent) | |
443 | break; | |
444 | ||
445 | if (sd_parent_degenerate(tmp, parent)) { | |
446 | tmp->parent = parent->parent; | |
447 | if (parent->parent) | |
448 | parent->parent->child = tmp; | |
449 | /* | |
450 | * Transfer SD_PREFER_SIBLING down in case of a | |
451 | * degenerate parent; the spans match for this | |
452 | * so the property transfers. | |
453 | */ | |
454 | if (parent->flags & SD_PREFER_SIBLING) | |
455 | tmp->flags |= SD_PREFER_SIBLING; | |
456 | destroy_sched_domain(parent); | |
457 | } else | |
458 | tmp = tmp->parent; | |
459 | } | |
460 | ||
461 | if (sd && sd_degenerate(sd)) { | |
462 | tmp = sd; | |
463 | sd = sd->parent; | |
464 | destroy_sched_domain(tmp); | |
465 | if (sd) | |
466 | sd->child = NULL; | |
467 | } | |
468 | ||
469 | sched_domain_debug(sd, cpu); | |
470 | ||
471 | rq_attach_root(rq, rd); | |
472 | tmp = rq->sd; | |
473 | rcu_assign_pointer(rq->sd, sd); | |
bbdacdfe | 474 | dirty_sched_domain_sysctl(cpu); |
f2cb1360 IM |
475 | destroy_sched_domains(tmp); |
476 | ||
477 | update_top_cache_domain(cpu); | |
478 | } | |
479 | ||
f2cb1360 IM |
480 | struct s_data { |
481 | struct sched_domain ** __percpu sd; | |
482 | struct root_domain *rd; | |
483 | }; | |
484 | ||
485 | enum s_alloc { | |
486 | sa_rootdomain, | |
487 | sa_sd, | |
488 | sa_sd_storage, | |
489 | sa_none, | |
490 | }; | |
491 | ||
35a566e6 PZ |
492 | /* |
493 | * Return the canonical balance CPU for this group, this is the first CPU | |
e5c14b1f | 494 | * of this group that's also in the balance mask. |
35a566e6 | 495 | * |
e5c14b1f PZ |
496 | * The balance mask are all those CPUs that could actually end up at this |
497 | * group. See build_balance_mask(). | |
35a566e6 PZ |
498 | * |
499 | * Also see should_we_balance(). | |
500 | */ | |
501 | int group_balance_cpu(struct sched_group *sg) | |
502 | { | |
e5c14b1f | 503 | return cpumask_first(group_balance_mask(sg)); |
35a566e6 PZ |
504 | } |
505 | ||
506 | ||
507 | /* | |
508 | * NUMA topology (first read the regular topology blurb below) | |
509 | * | |
510 | * Given a node-distance table, for example: | |
511 | * | |
512 | * node 0 1 2 3 | |
513 | * 0: 10 20 30 20 | |
514 | * 1: 20 10 20 30 | |
515 | * 2: 30 20 10 20 | |
516 | * 3: 20 30 20 10 | |
517 | * | |
518 | * which represents a 4 node ring topology like: | |
519 | * | |
520 | * 0 ----- 1 | |
521 | * | | | |
522 | * | | | |
523 | * | | | |
524 | * 3 ----- 2 | |
525 | * | |
526 | * We want to construct domains and groups to represent this. The way we go | |
527 | * about doing this is to build the domains on 'hops'. For each NUMA level we | |
528 | * construct the mask of all nodes reachable in @level hops. | |
529 | * | |
530 | * For the above NUMA topology that gives 3 levels: | |
531 | * | |
532 | * NUMA-2 0-3 0-3 0-3 0-3 | |
533 | * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2} | |
534 | * | |
535 | * NUMA-1 0-1,3 0-2 1-3 0,2-3 | |
536 | * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3} | |
537 | * | |
538 | * NUMA-0 0 1 2 3 | |
539 | * | |
540 | * | |
541 | * As can be seen; things don't nicely line up as with the regular topology. | |
542 | * When we iterate a domain in child domain chunks some nodes can be | |
543 | * represented multiple times -- hence the "overlap" naming for this part of | |
544 | * the topology. | |
545 | * | |
546 | * In order to minimize this overlap, we only build enough groups to cover the | |
547 | * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3. | |
548 | * | |
549 | * Because: | |
550 | * | |
551 | * - the first group of each domain is its child domain; this | |
552 | * gets us the first 0-1,3 | |
553 | * - the only uncovered node is 2, who's child domain is 1-3. | |
554 | * | |
555 | * However, because of the overlap, computing a unique CPU for each group is | |
556 | * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both | |
557 | * groups include the CPUs of Node-0, while those CPUs would not in fact ever | |
558 | * end up at those groups (they would end up in group: 0-1,3). | |
559 | * | |
e5c14b1f | 560 | * To correct this we have to introduce the group balance mask. This mask |
35a566e6 PZ |
561 | * will contain those CPUs in the group that can reach this group given the |
562 | * (child) domain tree. | |
563 | * | |
564 | * With this we can once again compute balance_cpu and sched_group_capacity | |
565 | * relations. | |
566 | * | |
567 | * XXX include words on how balance_cpu is unique and therefore can be | |
568 | * used for sched_group_capacity links. | |
569 | * | |
570 | * | |
571 | * Another 'interesting' topology is: | |
572 | * | |
573 | * node 0 1 2 3 | |
574 | * 0: 10 20 20 30 | |
575 | * 1: 20 10 20 20 | |
576 | * 2: 20 20 10 20 | |
577 | * 3: 30 20 20 10 | |
578 | * | |
579 | * Which looks a little like: | |
580 | * | |
581 | * 0 ----- 1 | |
582 | * | / | | |
583 | * | / | | |
584 | * | / | | |
585 | * 2 ----- 3 | |
586 | * | |
587 | * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3 | |
588 | * are not. | |
589 | * | |
590 | * This leads to a few particularly weird cases where the sched_domain's are | |
97fb7a0a | 591 | * not of the same number for each CPU. Consider: |
35a566e6 PZ |
592 | * |
593 | * NUMA-2 0-3 0-3 | |
594 | * groups: {0-2},{1-3} {1-3},{0-2} | |
595 | * | |
596 | * NUMA-1 0-2 0-3 0-3 1-3 | |
597 | * | |
598 | * NUMA-0 0 1 2 3 | |
599 | * | |
600 | */ | |
601 | ||
602 | ||
f2cb1360 | 603 | /* |
e5c14b1f PZ |
604 | * Build the balance mask; it contains only those CPUs that can arrive at this |
605 | * group and should be considered to continue balancing. | |
35a566e6 PZ |
606 | * |
607 | * We do this during the group creation pass, therefore the group information | |
608 | * isn't complete yet, however since each group represents a (child) domain we | |
609 | * can fully construct this using the sched_domain bits (which are already | |
610 | * complete). | |
f2cb1360 | 611 | */ |
1676330e | 612 | static void |
e5c14b1f | 613 | build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask) |
f2cb1360 | 614 | { |
ae4df9d6 | 615 | const struct cpumask *sg_span = sched_group_span(sg); |
f2cb1360 IM |
616 | struct sd_data *sdd = sd->private; |
617 | struct sched_domain *sibling; | |
618 | int i; | |
619 | ||
1676330e PZ |
620 | cpumask_clear(mask); |
621 | ||
f32d782e | 622 | for_each_cpu(i, sg_span) { |
f2cb1360 | 623 | sibling = *per_cpu_ptr(sdd->sd, i); |
73bb059f PZ |
624 | |
625 | /* | |
626 | * Can happen in the asymmetric case, where these siblings are | |
627 | * unused. The mask will not be empty because those CPUs that | |
628 | * do have the top domain _should_ span the domain. | |
629 | */ | |
630 | if (!sibling->child) | |
631 | continue; | |
632 | ||
633 | /* If we would not end up here, we can't continue from here */ | |
634 | if (!cpumask_equal(sg_span, sched_domain_span(sibling->child))) | |
f2cb1360 IM |
635 | continue; |
636 | ||
1676330e | 637 | cpumask_set_cpu(i, mask); |
f2cb1360 | 638 | } |
73bb059f PZ |
639 | |
640 | /* We must not have empty masks here */ | |
1676330e | 641 | WARN_ON_ONCE(cpumask_empty(mask)); |
f2cb1360 IM |
642 | } |
643 | ||
644 | /* | |
35a566e6 PZ |
645 | * XXX: This creates per-node group entries; since the load-balancer will |
646 | * immediately access remote memory to construct this group's load-balance | |
647 | * statistics having the groups node local is of dubious benefit. | |
f2cb1360 | 648 | */ |
8c033469 LRV |
649 | static struct sched_group * |
650 | build_group_from_child_sched_domain(struct sched_domain *sd, int cpu) | |
651 | { | |
652 | struct sched_group *sg; | |
653 | struct cpumask *sg_span; | |
654 | ||
655 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
656 | GFP_KERNEL, cpu_to_node(cpu)); | |
657 | ||
658 | if (!sg) | |
659 | return NULL; | |
660 | ||
ae4df9d6 | 661 | sg_span = sched_group_span(sg); |
8c033469 LRV |
662 | if (sd->child) |
663 | cpumask_copy(sg_span, sched_domain_span(sd->child)); | |
664 | else | |
665 | cpumask_copy(sg_span, sched_domain_span(sd)); | |
666 | ||
213c5a45 | 667 | atomic_inc(&sg->ref); |
8c033469 LRV |
668 | return sg; |
669 | } | |
670 | ||
671 | static void init_overlap_sched_group(struct sched_domain *sd, | |
1676330e | 672 | struct sched_group *sg) |
8c033469 | 673 | { |
1676330e | 674 | struct cpumask *mask = sched_domains_tmpmask2; |
8c033469 LRV |
675 | struct sd_data *sdd = sd->private; |
676 | struct cpumask *sg_span; | |
1676330e PZ |
677 | int cpu; |
678 | ||
e5c14b1f | 679 | build_balance_mask(sd, sg, mask); |
ae4df9d6 | 680 | cpu = cpumask_first_and(sched_group_span(sg), mask); |
8c033469 LRV |
681 | |
682 | sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); | |
683 | if (atomic_inc_return(&sg->sgc->ref) == 1) | |
e5c14b1f | 684 | cpumask_copy(group_balance_mask(sg), mask); |
35a566e6 | 685 | else |
e5c14b1f | 686 | WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask)); |
8c033469 LRV |
687 | |
688 | /* | |
689 | * Initialize sgc->capacity such that even if we mess up the | |
690 | * domains and no possible iteration will get us here, we won't | |
691 | * die on a /0 trap. | |
692 | */ | |
ae4df9d6 | 693 | sg_span = sched_group_span(sg); |
8c033469 LRV |
694 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); |
695 | sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; | |
e3d6d0cb | 696 | sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; |
8c033469 LRV |
697 | } |
698 | ||
f2cb1360 IM |
699 | static int |
700 | build_overlap_sched_groups(struct sched_domain *sd, int cpu) | |
701 | { | |
91eaed0d | 702 | struct sched_group *first = NULL, *last = NULL, *sg; |
f2cb1360 IM |
703 | const struct cpumask *span = sched_domain_span(sd); |
704 | struct cpumask *covered = sched_domains_tmpmask; | |
705 | struct sd_data *sdd = sd->private; | |
706 | struct sched_domain *sibling; | |
707 | int i; | |
708 | ||
709 | cpumask_clear(covered); | |
710 | ||
0372dd27 | 711 | for_each_cpu_wrap(i, span, cpu) { |
f2cb1360 IM |
712 | struct cpumask *sg_span; |
713 | ||
714 | if (cpumask_test_cpu(i, covered)) | |
715 | continue; | |
716 | ||
717 | sibling = *per_cpu_ptr(sdd->sd, i); | |
718 | ||
c20e1ea4 LRV |
719 | /* |
720 | * Asymmetric node setups can result in situations where the | |
721 | * domain tree is of unequal depth, make sure to skip domains | |
722 | * that already cover the entire range. | |
723 | * | |
724 | * In that case build_sched_domains() will have terminated the | |
725 | * iteration early and our sibling sd spans will be empty. | |
726 | * Domains should always include the CPU they're built on, so | |
727 | * check that. | |
728 | */ | |
f2cb1360 IM |
729 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) |
730 | continue; | |
731 | ||
8c033469 | 732 | sg = build_group_from_child_sched_domain(sibling, cpu); |
f2cb1360 IM |
733 | if (!sg) |
734 | goto fail; | |
735 | ||
ae4df9d6 | 736 | sg_span = sched_group_span(sg); |
f2cb1360 IM |
737 | cpumask_or(covered, covered, sg_span); |
738 | ||
1676330e | 739 | init_overlap_sched_group(sd, sg); |
f2cb1360 | 740 | |
f2cb1360 IM |
741 | if (!first) |
742 | first = sg; | |
743 | if (last) | |
744 | last->next = sg; | |
745 | last = sg; | |
746 | last->next = first; | |
747 | } | |
91eaed0d | 748 | sd->groups = first; |
f2cb1360 IM |
749 | |
750 | return 0; | |
751 | ||
752 | fail: | |
753 | free_sched_groups(first, 0); | |
754 | ||
755 | return -ENOMEM; | |
756 | } | |
757 | ||
35a566e6 PZ |
758 | |
759 | /* | |
760 | * Package topology (also see the load-balance blurb in fair.c) | |
761 | * | |
762 | * The scheduler builds a tree structure to represent a number of important | |
763 | * topology features. By default (default_topology[]) these include: | |
764 | * | |
765 | * - Simultaneous multithreading (SMT) | |
766 | * - Multi-Core Cache (MC) | |
767 | * - Package (DIE) | |
768 | * | |
769 | * Where the last one more or less denotes everything up to a NUMA node. | |
770 | * | |
771 | * The tree consists of 3 primary data structures: | |
772 | * | |
773 | * sched_domain -> sched_group -> sched_group_capacity | |
774 | * ^ ^ ^ ^ | |
775 | * `-' `-' | |
776 | * | |
97fb7a0a | 777 | * The sched_domains are per-CPU and have a two way link (parent & child) and |
35a566e6 PZ |
778 | * denote the ever growing mask of CPUs belonging to that level of topology. |
779 | * | |
780 | * Each sched_domain has a circular (double) linked list of sched_group's, each | |
781 | * denoting the domains of the level below (or individual CPUs in case of the | |
782 | * first domain level). The sched_group linked by a sched_domain includes the | |
783 | * CPU of that sched_domain [*]. | |
784 | * | |
785 | * Take for instance a 2 threaded, 2 core, 2 cache cluster part: | |
786 | * | |
787 | * CPU 0 1 2 3 4 5 6 7 | |
788 | * | |
789 | * DIE [ ] | |
790 | * MC [ ] [ ] | |
791 | * SMT [ ] [ ] [ ] [ ] | |
792 | * | |
793 | * - or - | |
794 | * | |
795 | * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7 | |
796 | * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7 | |
797 | * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7 | |
798 | * | |
799 | * CPU 0 1 2 3 4 5 6 7 | |
800 | * | |
801 | * One way to think about it is: sched_domain moves you up and down among these | |
802 | * topology levels, while sched_group moves you sideways through it, at child | |
803 | * domain granularity. | |
804 | * | |
805 | * sched_group_capacity ensures each unique sched_group has shared storage. | |
806 | * | |
807 | * There are two related construction problems, both require a CPU that | |
808 | * uniquely identify each group (for a given domain): | |
809 | * | |
810 | * - The first is the balance_cpu (see should_we_balance() and the | |
811 | * load-balance blub in fair.c); for each group we only want 1 CPU to | |
812 | * continue balancing at a higher domain. | |
813 | * | |
814 | * - The second is the sched_group_capacity; we want all identical groups | |
815 | * to share a single sched_group_capacity. | |
816 | * | |
817 | * Since these topologies are exclusive by construction. That is, its | |
818 | * impossible for an SMT thread to belong to multiple cores, and cores to | |
819 | * be part of multiple caches. There is a very clear and unique location | |
820 | * for each CPU in the hierarchy. | |
821 | * | |
822 | * Therefore computing a unique CPU for each group is trivial (the iteration | |
823 | * mask is redundant and set all 1s; all CPUs in a group will end up at _that_ | |
824 | * group), we can simply pick the first CPU in each group. | |
825 | * | |
826 | * | |
827 | * [*] in other words, the first group of each domain is its child domain. | |
828 | */ | |
829 | ||
0c0e776a | 830 | static struct sched_group *get_group(int cpu, struct sd_data *sdd) |
f2cb1360 IM |
831 | { |
832 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); | |
833 | struct sched_domain *child = sd->child; | |
0c0e776a | 834 | struct sched_group *sg; |
f2cb1360 IM |
835 | |
836 | if (child) | |
837 | cpu = cpumask_first(sched_domain_span(child)); | |
838 | ||
0c0e776a PZ |
839 | sg = *per_cpu_ptr(sdd->sg, cpu); |
840 | sg->sgc = *per_cpu_ptr(sdd->sgc, cpu); | |
841 | ||
842 | /* For claim_allocations: */ | |
843 | atomic_inc(&sg->ref); | |
844 | atomic_inc(&sg->sgc->ref); | |
f2cb1360 | 845 | |
0c0e776a | 846 | if (child) { |
ae4df9d6 PZ |
847 | cpumask_copy(sched_group_span(sg), sched_domain_span(child)); |
848 | cpumask_copy(group_balance_mask(sg), sched_group_span(sg)); | |
0c0e776a | 849 | } else { |
ae4df9d6 | 850 | cpumask_set_cpu(cpu, sched_group_span(sg)); |
e5c14b1f | 851 | cpumask_set_cpu(cpu, group_balance_mask(sg)); |
f2cb1360 IM |
852 | } |
853 | ||
ae4df9d6 | 854 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg)); |
0c0e776a | 855 | sg->sgc->min_capacity = SCHED_CAPACITY_SCALE; |
e3d6d0cb | 856 | sg->sgc->max_capacity = SCHED_CAPACITY_SCALE; |
0c0e776a PZ |
857 | |
858 | return sg; | |
f2cb1360 IM |
859 | } |
860 | ||
861 | /* | |
862 | * build_sched_groups will build a circular linked list of the groups | |
863 | * covered by the given span, and will set each group's ->cpumask correctly, | |
864 | * and ->cpu_capacity to 0. | |
865 | * | |
866 | * Assumes the sched_domain tree is fully constructed | |
867 | */ | |
868 | static int | |
869 | build_sched_groups(struct sched_domain *sd, int cpu) | |
870 | { | |
871 | struct sched_group *first = NULL, *last = NULL; | |
872 | struct sd_data *sdd = sd->private; | |
873 | const struct cpumask *span = sched_domain_span(sd); | |
874 | struct cpumask *covered; | |
875 | int i; | |
876 | ||
f2cb1360 IM |
877 | lockdep_assert_held(&sched_domains_mutex); |
878 | covered = sched_domains_tmpmask; | |
879 | ||
880 | cpumask_clear(covered); | |
881 | ||
0c0e776a | 882 | for_each_cpu_wrap(i, span, cpu) { |
f2cb1360 | 883 | struct sched_group *sg; |
f2cb1360 IM |
884 | |
885 | if (cpumask_test_cpu(i, covered)) | |
886 | continue; | |
887 | ||
0c0e776a | 888 | sg = get_group(i, sdd); |
f2cb1360 | 889 | |
ae4df9d6 | 890 | cpumask_or(covered, covered, sched_group_span(sg)); |
f2cb1360 IM |
891 | |
892 | if (!first) | |
893 | first = sg; | |
894 | if (last) | |
895 | last->next = sg; | |
896 | last = sg; | |
897 | } | |
898 | last->next = first; | |
0c0e776a | 899 | sd->groups = first; |
f2cb1360 IM |
900 | |
901 | return 0; | |
902 | } | |
903 | ||
904 | /* | |
905 | * Initialize sched groups cpu_capacity. | |
906 | * | |
907 | * cpu_capacity indicates the capacity of sched group, which is used while | |
908 | * distributing the load between different sched groups in a sched domain. | |
909 | * Typically cpu_capacity for all the groups in a sched domain will be same | |
910 | * unless there are asymmetries in the topology. If there are asymmetries, | |
911 | * group having more cpu_capacity will pickup more load compared to the | |
912 | * group having less cpu_capacity. | |
913 | */ | |
914 | static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) | |
915 | { | |
916 | struct sched_group *sg = sd->groups; | |
917 | ||
918 | WARN_ON(!sg); | |
919 | ||
920 | do { | |
921 | int cpu, max_cpu = -1; | |
922 | ||
ae4df9d6 | 923 | sg->group_weight = cpumask_weight(sched_group_span(sg)); |
f2cb1360 IM |
924 | |
925 | if (!(sd->flags & SD_ASYM_PACKING)) | |
926 | goto next; | |
927 | ||
ae4df9d6 | 928 | for_each_cpu(cpu, sched_group_span(sg)) { |
f2cb1360 IM |
929 | if (max_cpu < 0) |
930 | max_cpu = cpu; | |
931 | else if (sched_asym_prefer(cpu, max_cpu)) | |
932 | max_cpu = cpu; | |
933 | } | |
934 | sg->asym_prefer_cpu = max_cpu; | |
935 | ||
936 | next: | |
937 | sg = sg->next; | |
938 | } while (sg != sd->groups); | |
939 | ||
940 | if (cpu != group_balance_cpu(sg)) | |
941 | return; | |
942 | ||
943 | update_group_capacity(sd, cpu); | |
944 | } | |
945 | ||
946 | /* | |
947 | * Initializers for schedule domains | |
948 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
949 | */ | |
950 | ||
951 | static int default_relax_domain_level = -1; | |
952 | int sched_domain_level_max; | |
953 | ||
954 | static int __init setup_relax_domain_level(char *str) | |
955 | { | |
956 | if (kstrtoint(str, 0, &default_relax_domain_level)) | |
957 | pr_warn("Unable to set relax_domain_level\n"); | |
958 | ||
959 | return 1; | |
960 | } | |
961 | __setup("relax_domain_level=", setup_relax_domain_level); | |
962 | ||
963 | static void set_domain_attribute(struct sched_domain *sd, | |
964 | struct sched_domain_attr *attr) | |
965 | { | |
966 | int request; | |
967 | ||
968 | if (!attr || attr->relax_domain_level < 0) { | |
969 | if (default_relax_domain_level < 0) | |
970 | return; | |
971 | else | |
972 | request = default_relax_domain_level; | |
973 | } else | |
974 | request = attr->relax_domain_level; | |
975 | if (request < sd->level) { | |
976 | /* Turn off idle balance on this domain: */ | |
977 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
978 | } else { | |
979 | /* Turn on idle balance on this domain: */ | |
980 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | |
981 | } | |
982 | } | |
983 | ||
984 | static void __sdt_free(const struct cpumask *cpu_map); | |
985 | static int __sdt_alloc(const struct cpumask *cpu_map); | |
986 | ||
987 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, | |
988 | const struct cpumask *cpu_map) | |
989 | { | |
990 | switch (what) { | |
991 | case sa_rootdomain: | |
992 | if (!atomic_read(&d->rd->refcount)) | |
993 | free_rootdomain(&d->rd->rcu); | |
994 | /* Fall through */ | |
995 | case sa_sd: | |
996 | free_percpu(d->sd); | |
997 | /* Fall through */ | |
998 | case sa_sd_storage: | |
999 | __sdt_free(cpu_map); | |
1000 | /* Fall through */ | |
1001 | case sa_none: | |
1002 | break; | |
1003 | } | |
1004 | } | |
1005 | ||
1006 | static enum s_alloc | |
1007 | __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map) | |
1008 | { | |
1009 | memset(d, 0, sizeof(*d)); | |
1010 | ||
1011 | if (__sdt_alloc(cpu_map)) | |
1012 | return sa_sd_storage; | |
1013 | d->sd = alloc_percpu(struct sched_domain *); | |
1014 | if (!d->sd) | |
1015 | return sa_sd_storage; | |
1016 | d->rd = alloc_rootdomain(); | |
1017 | if (!d->rd) | |
1018 | return sa_sd; | |
97fb7a0a | 1019 | |
f2cb1360 IM |
1020 | return sa_rootdomain; |
1021 | } | |
1022 | ||
1023 | /* | |
1024 | * NULL the sd_data elements we've used to build the sched_domain and | |
1025 | * sched_group structure so that the subsequent __free_domain_allocs() | |
1026 | * will not free the data we're using. | |
1027 | */ | |
1028 | static void claim_allocations(int cpu, struct sched_domain *sd) | |
1029 | { | |
1030 | struct sd_data *sdd = sd->private; | |
1031 | ||
1032 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); | |
1033 | *per_cpu_ptr(sdd->sd, cpu) = NULL; | |
1034 | ||
1035 | if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref)) | |
1036 | *per_cpu_ptr(sdd->sds, cpu) = NULL; | |
1037 | ||
1038 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) | |
1039 | *per_cpu_ptr(sdd->sg, cpu) = NULL; | |
1040 | ||
1041 | if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) | |
1042 | *per_cpu_ptr(sdd->sgc, cpu) = NULL; | |
1043 | } | |
1044 | ||
1045 | #ifdef CONFIG_NUMA | |
f2cb1360 | 1046 | enum numa_topology_type sched_numa_topology_type; |
97fb7a0a IM |
1047 | |
1048 | static int sched_domains_numa_levels; | |
1049 | static int sched_domains_curr_level; | |
1050 | ||
1051 | int sched_max_numa_distance; | |
1052 | static int *sched_domains_numa_distance; | |
1053 | static struct cpumask ***sched_domains_numa_masks; | |
f2cb1360 IM |
1054 | #endif |
1055 | ||
1056 | /* | |
1057 | * SD_flags allowed in topology descriptions. | |
1058 | * | |
1059 | * These flags are purely descriptive of the topology and do not prescribe | |
1060 | * behaviour. Behaviour is artificial and mapped in the below sd_init() | |
1061 | * function: | |
1062 | * | |
1063 | * SD_SHARE_CPUCAPACITY - describes SMT topologies | |
1064 | * SD_SHARE_PKG_RESOURCES - describes shared caches | |
1065 | * SD_NUMA - describes NUMA topologies | |
1066 | * SD_SHARE_POWERDOMAIN - describes shared power domain | |
f2cb1360 IM |
1067 | * |
1068 | * Odd one out, which beside describing the topology has a quirk also | |
1069 | * prescribes the desired behaviour that goes along with it: | |
1070 | * | |
1071 | * SD_ASYM_PACKING - describes SMT quirks | |
1072 | */ | |
1073 | #define TOPOLOGY_SD_FLAGS \ | |
97fb7a0a | 1074 | (SD_SHARE_CPUCAPACITY | \ |
f2cb1360 | 1075 | SD_SHARE_PKG_RESOURCES | \ |
97fb7a0a IM |
1076 | SD_NUMA | \ |
1077 | SD_ASYM_PACKING | \ | |
f2cb1360 IM |
1078 | SD_SHARE_POWERDOMAIN) |
1079 | ||
1080 | static struct sched_domain * | |
1081 | sd_init(struct sched_domain_topology_level *tl, | |
1082 | const struct cpumask *cpu_map, | |
05484e09 | 1083 | struct sched_domain *child, int dflags, int cpu) |
f2cb1360 IM |
1084 | { |
1085 | struct sd_data *sdd = &tl->data; | |
1086 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); | |
1087 | int sd_id, sd_weight, sd_flags = 0; | |
1088 | ||
1089 | #ifdef CONFIG_NUMA | |
1090 | /* | |
1091 | * Ugly hack to pass state to sd_numa_mask()... | |
1092 | */ | |
1093 | sched_domains_curr_level = tl->numa_level; | |
1094 | #endif | |
1095 | ||
1096 | sd_weight = cpumask_weight(tl->mask(cpu)); | |
1097 | ||
1098 | if (tl->sd_flags) | |
1099 | sd_flags = (*tl->sd_flags)(); | |
1100 | if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, | |
1101 | "wrong sd_flags in topology description\n")) | |
1102 | sd_flags &= ~TOPOLOGY_SD_FLAGS; | |
1103 | ||
05484e09 MR |
1104 | /* Apply detected topology flags */ |
1105 | sd_flags |= dflags; | |
1106 | ||
f2cb1360 IM |
1107 | *sd = (struct sched_domain){ |
1108 | .min_interval = sd_weight, | |
1109 | .max_interval = 2*sd_weight, | |
1110 | .busy_factor = 32, | |
1111 | .imbalance_pct = 125, | |
1112 | ||
1113 | .cache_nice_tries = 0, | |
1114 | .busy_idx = 0, | |
1115 | .idle_idx = 0, | |
1116 | .newidle_idx = 0, | |
1117 | .wake_idx = 0, | |
1118 | .forkexec_idx = 0, | |
1119 | ||
1120 | .flags = 1*SD_LOAD_BALANCE | |
1121 | | 1*SD_BALANCE_NEWIDLE | |
1122 | | 1*SD_BALANCE_EXEC | |
1123 | | 1*SD_BALANCE_FORK | |
1124 | | 0*SD_BALANCE_WAKE | |
1125 | | 1*SD_WAKE_AFFINE | |
1126 | | 0*SD_SHARE_CPUCAPACITY | |
1127 | | 0*SD_SHARE_PKG_RESOURCES | |
1128 | | 0*SD_SERIALIZE | |
9c63e84d | 1129 | | 1*SD_PREFER_SIBLING |
f2cb1360 IM |
1130 | | 0*SD_NUMA |
1131 | | sd_flags | |
1132 | , | |
1133 | ||
1134 | .last_balance = jiffies, | |
1135 | .balance_interval = sd_weight, | |
1136 | .smt_gain = 0, | |
1137 | .max_newidle_lb_cost = 0, | |
1138 | .next_decay_max_lb_cost = jiffies, | |
1139 | .child = child, | |
1140 | #ifdef CONFIG_SCHED_DEBUG | |
1141 | .name = tl->name, | |
1142 | #endif | |
1143 | }; | |
1144 | ||
1145 | cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); | |
1146 | sd_id = cpumask_first(sched_domain_span(sd)); | |
1147 | ||
1148 | /* | |
1149 | * Convert topological properties into behaviour. | |
1150 | */ | |
1151 | ||
1152 | if (sd->flags & SD_ASYM_CPUCAPACITY) { | |
1153 | struct sched_domain *t = sd; | |
1154 | ||
9c63e84d MR |
1155 | /* |
1156 | * Don't attempt to spread across CPUs of different capacities. | |
1157 | */ | |
1158 | if (sd->child) | |
1159 | sd->child->flags &= ~SD_PREFER_SIBLING; | |
1160 | ||
f2cb1360 IM |
1161 | for_each_lower_domain(t) |
1162 | t->flags |= SD_BALANCE_WAKE; | |
1163 | } | |
1164 | ||
1165 | if (sd->flags & SD_SHARE_CPUCAPACITY) { | |
f2cb1360 IM |
1166 | sd->imbalance_pct = 110; |
1167 | sd->smt_gain = 1178; /* ~15% */ | |
1168 | ||
1169 | } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { | |
1170 | sd->imbalance_pct = 117; | |
1171 | sd->cache_nice_tries = 1; | |
1172 | sd->busy_idx = 2; | |
1173 | ||
1174 | #ifdef CONFIG_NUMA | |
1175 | } else if (sd->flags & SD_NUMA) { | |
1176 | sd->cache_nice_tries = 2; | |
1177 | sd->busy_idx = 3; | |
1178 | sd->idle_idx = 2; | |
1179 | ||
9c63e84d | 1180 | sd->flags &= ~SD_PREFER_SIBLING; |
f2cb1360 IM |
1181 | sd->flags |= SD_SERIALIZE; |
1182 | if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { | |
1183 | sd->flags &= ~(SD_BALANCE_EXEC | | |
1184 | SD_BALANCE_FORK | | |
1185 | SD_WAKE_AFFINE); | |
1186 | } | |
1187 | ||
1188 | #endif | |
1189 | } else { | |
f2cb1360 IM |
1190 | sd->cache_nice_tries = 1; |
1191 | sd->busy_idx = 2; | |
1192 | sd->idle_idx = 1; | |
1193 | } | |
1194 | ||
1195 | /* | |
1196 | * For all levels sharing cache; connect a sched_domain_shared | |
1197 | * instance. | |
1198 | */ | |
1199 | if (sd->flags & SD_SHARE_PKG_RESOURCES) { | |
1200 | sd->shared = *per_cpu_ptr(sdd->sds, sd_id); | |
1201 | atomic_inc(&sd->shared->ref); | |
1202 | atomic_set(&sd->shared->nr_busy_cpus, sd_weight); | |
1203 | } | |
1204 | ||
1205 | sd->private = sdd; | |
1206 | ||
1207 | return sd; | |
1208 | } | |
1209 | ||
1210 | /* | |
1211 | * Topology list, bottom-up. | |
1212 | */ | |
1213 | static struct sched_domain_topology_level default_topology[] = { | |
1214 | #ifdef CONFIG_SCHED_SMT | |
1215 | { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, | |
1216 | #endif | |
1217 | #ifdef CONFIG_SCHED_MC | |
1218 | { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, | |
1219 | #endif | |
1220 | { cpu_cpu_mask, SD_INIT_NAME(DIE) }, | |
1221 | { NULL, }, | |
1222 | }; | |
1223 | ||
1224 | static struct sched_domain_topology_level *sched_domain_topology = | |
1225 | default_topology; | |
1226 | ||
1227 | #define for_each_sd_topology(tl) \ | |
1228 | for (tl = sched_domain_topology; tl->mask; tl++) | |
1229 | ||
1230 | void set_sched_topology(struct sched_domain_topology_level *tl) | |
1231 | { | |
1232 | if (WARN_ON_ONCE(sched_smp_initialized)) | |
1233 | return; | |
1234 | ||
1235 | sched_domain_topology = tl; | |
1236 | } | |
1237 | ||
1238 | #ifdef CONFIG_NUMA | |
1239 | ||
1240 | static const struct cpumask *sd_numa_mask(int cpu) | |
1241 | { | |
1242 | return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; | |
1243 | } | |
1244 | ||
1245 | static void sched_numa_warn(const char *str) | |
1246 | { | |
1247 | static int done = false; | |
1248 | int i,j; | |
1249 | ||
1250 | if (done) | |
1251 | return; | |
1252 | ||
1253 | done = true; | |
1254 | ||
1255 | printk(KERN_WARNING "ERROR: %s\n\n", str); | |
1256 | ||
1257 | for (i = 0; i < nr_node_ids; i++) { | |
1258 | printk(KERN_WARNING " "); | |
1259 | for (j = 0; j < nr_node_ids; j++) | |
1260 | printk(KERN_CONT "%02d ", node_distance(i,j)); | |
1261 | printk(KERN_CONT "\n"); | |
1262 | } | |
1263 | printk(KERN_WARNING "\n"); | |
1264 | } | |
1265 | ||
1266 | bool find_numa_distance(int distance) | |
1267 | { | |
1268 | int i; | |
1269 | ||
1270 | if (distance == node_distance(0, 0)) | |
1271 | return true; | |
1272 | ||
1273 | for (i = 0; i < sched_domains_numa_levels; i++) { | |
1274 | if (sched_domains_numa_distance[i] == distance) | |
1275 | return true; | |
1276 | } | |
1277 | ||
1278 | return false; | |
1279 | } | |
1280 | ||
1281 | /* | |
1282 | * A system can have three types of NUMA topology: | |
1283 | * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system | |
1284 | * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes | |
1285 | * NUMA_BACKPLANE: nodes can reach other nodes through a backplane | |
1286 | * | |
1287 | * The difference between a glueless mesh topology and a backplane | |
1288 | * topology lies in whether communication between not directly | |
1289 | * connected nodes goes through intermediary nodes (where programs | |
1290 | * could run), or through backplane controllers. This affects | |
1291 | * placement of programs. | |
1292 | * | |
1293 | * The type of topology can be discerned with the following tests: | |
1294 | * - If the maximum distance between any nodes is 1 hop, the system | |
1295 | * is directly connected. | |
1296 | * - If for two nodes A and B, located N > 1 hops away from each other, | |
1297 | * there is an intermediary node C, which is < N hops away from both | |
1298 | * nodes A and B, the system is a glueless mesh. | |
1299 | */ | |
1300 | static void init_numa_topology_type(void) | |
1301 | { | |
1302 | int a, b, c, n; | |
1303 | ||
1304 | n = sched_max_numa_distance; | |
1305 | ||
e5e96faf | 1306 | if (sched_domains_numa_levels <= 2) { |
f2cb1360 IM |
1307 | sched_numa_topology_type = NUMA_DIRECT; |
1308 | return; | |
1309 | } | |
1310 | ||
1311 | for_each_online_node(a) { | |
1312 | for_each_online_node(b) { | |
1313 | /* Find two nodes furthest removed from each other. */ | |
1314 | if (node_distance(a, b) < n) | |
1315 | continue; | |
1316 | ||
1317 | /* Is there an intermediary node between a and b? */ | |
1318 | for_each_online_node(c) { | |
1319 | if (node_distance(a, c) < n && | |
1320 | node_distance(b, c) < n) { | |
1321 | sched_numa_topology_type = | |
1322 | NUMA_GLUELESS_MESH; | |
1323 | return; | |
1324 | } | |
1325 | } | |
1326 | ||
1327 | sched_numa_topology_type = NUMA_BACKPLANE; | |
1328 | return; | |
1329 | } | |
1330 | } | |
1331 | } | |
1332 | ||
1333 | void sched_init_numa(void) | |
1334 | { | |
1335 | int next_distance, curr_distance = node_distance(0, 0); | |
1336 | struct sched_domain_topology_level *tl; | |
1337 | int level = 0; | |
1338 | int i, j, k; | |
1339 | ||
1340 | sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); | |
1341 | if (!sched_domains_numa_distance) | |
1342 | return; | |
1343 | ||
051f3ca0 SS |
1344 | /* Includes NUMA identity node at level 0. */ |
1345 | sched_domains_numa_distance[level++] = curr_distance; | |
1346 | sched_domains_numa_levels = level; | |
1347 | ||
f2cb1360 IM |
1348 | /* |
1349 | * O(nr_nodes^2) deduplicating selection sort -- in order to find the | |
1350 | * unique distances in the node_distance() table. | |
1351 | * | |
1352 | * Assumes node_distance(0,j) includes all distances in | |
1353 | * node_distance(i,j) in order to avoid cubic time. | |
1354 | */ | |
1355 | next_distance = curr_distance; | |
1356 | for (i = 0; i < nr_node_ids; i++) { | |
1357 | for (j = 0; j < nr_node_ids; j++) { | |
1358 | for (k = 0; k < nr_node_ids; k++) { | |
1359 | int distance = node_distance(i, k); | |
1360 | ||
1361 | if (distance > curr_distance && | |
1362 | (distance < next_distance || | |
1363 | next_distance == curr_distance)) | |
1364 | next_distance = distance; | |
1365 | ||
1366 | /* | |
1367 | * While not a strong assumption it would be nice to know | |
1368 | * about cases where if node A is connected to B, B is not | |
1369 | * equally connected to A. | |
1370 | */ | |
1371 | if (sched_debug() && node_distance(k, i) != distance) | |
1372 | sched_numa_warn("Node-distance not symmetric"); | |
1373 | ||
1374 | if (sched_debug() && i && !find_numa_distance(distance)) | |
1375 | sched_numa_warn("Node-0 not representative"); | |
1376 | } | |
1377 | if (next_distance != curr_distance) { | |
1378 | sched_domains_numa_distance[level++] = next_distance; | |
1379 | sched_domains_numa_levels = level; | |
1380 | curr_distance = next_distance; | |
1381 | } else break; | |
1382 | } | |
1383 | ||
1384 | /* | |
1385 | * In case of sched_debug() we verify the above assumption. | |
1386 | */ | |
1387 | if (!sched_debug()) | |
1388 | break; | |
1389 | } | |
1390 | ||
f2cb1360 | 1391 | /* |
051f3ca0 | 1392 | * 'level' contains the number of unique distances |
f2cb1360 IM |
1393 | * |
1394 | * The sched_domains_numa_distance[] array includes the actual distance | |
1395 | * numbers. | |
1396 | */ | |
1397 | ||
1398 | /* | |
1399 | * Here, we should temporarily reset sched_domains_numa_levels to 0. | |
1400 | * If it fails to allocate memory for array sched_domains_numa_masks[][], | |
1401 | * the array will contain less then 'level' members. This could be | |
1402 | * dangerous when we use it to iterate array sched_domains_numa_masks[][] | |
1403 | * in other functions. | |
1404 | * | |
1405 | * We reset it to 'level' at the end of this function. | |
1406 | */ | |
1407 | sched_domains_numa_levels = 0; | |
1408 | ||
1409 | sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); | |
1410 | if (!sched_domains_numa_masks) | |
1411 | return; | |
1412 | ||
1413 | /* | |
1414 | * Now for each level, construct a mask per node which contains all | |
1415 | * CPUs of nodes that are that many hops away from us. | |
1416 | */ | |
1417 | for (i = 0; i < level; i++) { | |
1418 | sched_domains_numa_masks[i] = | |
1419 | kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); | |
1420 | if (!sched_domains_numa_masks[i]) | |
1421 | return; | |
1422 | ||
1423 | for (j = 0; j < nr_node_ids; j++) { | |
1424 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); | |
1425 | if (!mask) | |
1426 | return; | |
1427 | ||
1428 | sched_domains_numa_masks[i][j] = mask; | |
1429 | ||
1430 | for_each_node(k) { | |
1431 | if (node_distance(j, k) > sched_domains_numa_distance[i]) | |
1432 | continue; | |
1433 | ||
1434 | cpumask_or(mask, mask, cpumask_of_node(k)); | |
1435 | } | |
1436 | } | |
1437 | } | |
1438 | ||
1439 | /* Compute default topology size */ | |
1440 | for (i = 0; sched_domain_topology[i].mask; i++); | |
1441 | ||
1442 | tl = kzalloc((i + level + 1) * | |
1443 | sizeof(struct sched_domain_topology_level), GFP_KERNEL); | |
1444 | if (!tl) | |
1445 | return; | |
1446 | ||
1447 | /* | |
1448 | * Copy the default topology bits.. | |
1449 | */ | |
1450 | for (i = 0; sched_domain_topology[i].mask; i++) | |
1451 | tl[i] = sched_domain_topology[i]; | |
1452 | ||
051f3ca0 SS |
1453 | /* |
1454 | * Add the NUMA identity distance, aka single NODE. | |
1455 | */ | |
1456 | tl[i++] = (struct sched_domain_topology_level){ | |
1457 | .mask = sd_numa_mask, | |
1458 | .numa_level = 0, | |
1459 | SD_INIT_NAME(NODE) | |
1460 | }; | |
1461 | ||
f2cb1360 IM |
1462 | /* |
1463 | * .. and append 'j' levels of NUMA goodness. | |
1464 | */ | |
051f3ca0 | 1465 | for (j = 1; j < level; i++, j++) { |
f2cb1360 IM |
1466 | tl[i] = (struct sched_domain_topology_level){ |
1467 | .mask = sd_numa_mask, | |
1468 | .sd_flags = cpu_numa_flags, | |
1469 | .flags = SDTL_OVERLAP, | |
1470 | .numa_level = j, | |
1471 | SD_INIT_NAME(NUMA) | |
1472 | }; | |
1473 | } | |
1474 | ||
1475 | sched_domain_topology = tl; | |
1476 | ||
1477 | sched_domains_numa_levels = level; | |
1478 | sched_max_numa_distance = sched_domains_numa_distance[level - 1]; | |
1479 | ||
1480 | init_numa_topology_type(); | |
1481 | } | |
1482 | ||
1483 | void sched_domains_numa_masks_set(unsigned int cpu) | |
1484 | { | |
1485 | int node = cpu_to_node(cpu); | |
1486 | int i, j; | |
1487 | ||
1488 | for (i = 0; i < sched_domains_numa_levels; i++) { | |
1489 | for (j = 0; j < nr_node_ids; j++) { | |
1490 | if (node_distance(j, node) <= sched_domains_numa_distance[i]) | |
1491 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); | |
1492 | } | |
1493 | } | |
1494 | } | |
1495 | ||
1496 | void sched_domains_numa_masks_clear(unsigned int cpu) | |
1497 | { | |
1498 | int i, j; | |
1499 | ||
1500 | for (i = 0; i < sched_domains_numa_levels; i++) { | |
1501 | for (j = 0; j < nr_node_ids; j++) | |
1502 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); | |
1503 | } | |
1504 | } | |
1505 | ||
1506 | #endif /* CONFIG_NUMA */ | |
1507 | ||
1508 | static int __sdt_alloc(const struct cpumask *cpu_map) | |
1509 | { | |
1510 | struct sched_domain_topology_level *tl; | |
1511 | int j; | |
1512 | ||
1513 | for_each_sd_topology(tl) { | |
1514 | struct sd_data *sdd = &tl->data; | |
1515 | ||
1516 | sdd->sd = alloc_percpu(struct sched_domain *); | |
1517 | if (!sdd->sd) | |
1518 | return -ENOMEM; | |
1519 | ||
1520 | sdd->sds = alloc_percpu(struct sched_domain_shared *); | |
1521 | if (!sdd->sds) | |
1522 | return -ENOMEM; | |
1523 | ||
1524 | sdd->sg = alloc_percpu(struct sched_group *); | |
1525 | if (!sdd->sg) | |
1526 | return -ENOMEM; | |
1527 | ||
1528 | sdd->sgc = alloc_percpu(struct sched_group_capacity *); | |
1529 | if (!sdd->sgc) | |
1530 | return -ENOMEM; | |
1531 | ||
1532 | for_each_cpu(j, cpu_map) { | |
1533 | struct sched_domain *sd; | |
1534 | struct sched_domain_shared *sds; | |
1535 | struct sched_group *sg; | |
1536 | struct sched_group_capacity *sgc; | |
1537 | ||
1538 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), | |
1539 | GFP_KERNEL, cpu_to_node(j)); | |
1540 | if (!sd) | |
1541 | return -ENOMEM; | |
1542 | ||
1543 | *per_cpu_ptr(sdd->sd, j) = sd; | |
1544 | ||
1545 | sds = kzalloc_node(sizeof(struct sched_domain_shared), | |
1546 | GFP_KERNEL, cpu_to_node(j)); | |
1547 | if (!sds) | |
1548 | return -ENOMEM; | |
1549 | ||
1550 | *per_cpu_ptr(sdd->sds, j) = sds; | |
1551 | ||
1552 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), | |
1553 | GFP_KERNEL, cpu_to_node(j)); | |
1554 | if (!sg) | |
1555 | return -ENOMEM; | |
1556 | ||
1557 | sg->next = sg; | |
1558 | ||
1559 | *per_cpu_ptr(sdd->sg, j) = sg; | |
1560 | ||
1561 | sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), | |
1562 | GFP_KERNEL, cpu_to_node(j)); | |
1563 | if (!sgc) | |
1564 | return -ENOMEM; | |
1565 | ||
005f874d PZ |
1566 | #ifdef CONFIG_SCHED_DEBUG |
1567 | sgc->id = j; | |
1568 | #endif | |
1569 | ||
f2cb1360 IM |
1570 | *per_cpu_ptr(sdd->sgc, j) = sgc; |
1571 | } | |
1572 | } | |
1573 | ||
1574 | return 0; | |
1575 | } | |
1576 | ||
1577 | static void __sdt_free(const struct cpumask *cpu_map) | |
1578 | { | |
1579 | struct sched_domain_topology_level *tl; | |
1580 | int j; | |
1581 | ||
1582 | for_each_sd_topology(tl) { | |
1583 | struct sd_data *sdd = &tl->data; | |
1584 | ||
1585 | for_each_cpu(j, cpu_map) { | |
1586 | struct sched_domain *sd; | |
1587 | ||
1588 | if (sdd->sd) { | |
1589 | sd = *per_cpu_ptr(sdd->sd, j); | |
1590 | if (sd && (sd->flags & SD_OVERLAP)) | |
1591 | free_sched_groups(sd->groups, 0); | |
1592 | kfree(*per_cpu_ptr(sdd->sd, j)); | |
1593 | } | |
1594 | ||
1595 | if (sdd->sds) | |
1596 | kfree(*per_cpu_ptr(sdd->sds, j)); | |
1597 | if (sdd->sg) | |
1598 | kfree(*per_cpu_ptr(sdd->sg, j)); | |
1599 | if (sdd->sgc) | |
1600 | kfree(*per_cpu_ptr(sdd->sgc, j)); | |
1601 | } | |
1602 | free_percpu(sdd->sd); | |
1603 | sdd->sd = NULL; | |
1604 | free_percpu(sdd->sds); | |
1605 | sdd->sds = NULL; | |
1606 | free_percpu(sdd->sg); | |
1607 | sdd->sg = NULL; | |
1608 | free_percpu(sdd->sgc); | |
1609 | sdd->sgc = NULL; | |
1610 | } | |
1611 | } | |
1612 | ||
181a80d1 | 1613 | static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, |
f2cb1360 | 1614 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
05484e09 | 1615 | struct sched_domain *child, int dflags, int cpu) |
f2cb1360 | 1616 | { |
05484e09 | 1617 | struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu); |
f2cb1360 IM |
1618 | |
1619 | if (child) { | |
1620 | sd->level = child->level + 1; | |
1621 | sched_domain_level_max = max(sched_domain_level_max, sd->level); | |
1622 | child->parent = sd; | |
1623 | ||
1624 | if (!cpumask_subset(sched_domain_span(child), | |
1625 | sched_domain_span(sd))) { | |
1626 | pr_err("BUG: arch topology borken\n"); | |
1627 | #ifdef CONFIG_SCHED_DEBUG | |
1628 | pr_err(" the %s domain not a subset of the %s domain\n", | |
1629 | child->name, sd->name); | |
1630 | #endif | |
97fb7a0a | 1631 | /* Fixup, ensure @sd has at least @child CPUs. */ |
f2cb1360 IM |
1632 | cpumask_or(sched_domain_span(sd), |
1633 | sched_domain_span(sd), | |
1634 | sched_domain_span(child)); | |
1635 | } | |
1636 | ||
1637 | } | |
1638 | set_domain_attribute(sd, attr); | |
1639 | ||
1640 | return sd; | |
1641 | } | |
1642 | ||
05484e09 MR |
1643 | /* |
1644 | * Find the sched_domain_topology_level where all CPU capacities are visible | |
1645 | * for all CPUs. | |
1646 | */ | |
1647 | static struct sched_domain_topology_level | |
1648 | *asym_cpu_capacity_level(const struct cpumask *cpu_map) | |
1649 | { | |
1650 | int i, j, asym_level = 0; | |
1651 | bool asym = false; | |
1652 | struct sched_domain_topology_level *tl, *asym_tl = NULL; | |
1653 | unsigned long cap; | |
1654 | ||
1655 | /* Is there any asymmetry? */ | |
1656 | cap = arch_scale_cpu_capacity(NULL, cpumask_first(cpu_map)); | |
1657 | ||
1658 | for_each_cpu(i, cpu_map) { | |
1659 | if (arch_scale_cpu_capacity(NULL, i) != cap) { | |
1660 | asym = true; | |
1661 | break; | |
1662 | } | |
1663 | } | |
1664 | ||
1665 | if (!asym) | |
1666 | return NULL; | |
1667 | ||
1668 | /* | |
1669 | * Examine topology from all CPU's point of views to detect the lowest | |
1670 | * sched_domain_topology_level where a highest capacity CPU is visible | |
1671 | * to everyone. | |
1672 | */ | |
1673 | for_each_cpu(i, cpu_map) { | |
1674 | unsigned long max_capacity = arch_scale_cpu_capacity(NULL, i); | |
1675 | int tl_id = 0; | |
1676 | ||
1677 | for_each_sd_topology(tl) { | |
1678 | if (tl_id < asym_level) | |
1679 | goto next_level; | |
1680 | ||
1681 | for_each_cpu_and(j, tl->mask(i), cpu_map) { | |
1682 | unsigned long capacity; | |
1683 | ||
1684 | capacity = arch_scale_cpu_capacity(NULL, j); | |
1685 | ||
1686 | if (capacity <= max_capacity) | |
1687 | continue; | |
1688 | ||
1689 | max_capacity = capacity; | |
1690 | asym_level = tl_id; | |
1691 | asym_tl = tl; | |
1692 | } | |
1693 | next_level: | |
1694 | tl_id++; | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | return asym_tl; | |
1699 | } | |
1700 | ||
1701 | ||
f2cb1360 IM |
1702 | /* |
1703 | * Build sched domains for a given set of CPUs and attach the sched domains | |
1704 | * to the individual CPUs | |
1705 | */ | |
1706 | static int | |
1707 | build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr) | |
1708 | { | |
1709 | enum s_alloc alloc_state; | |
1710 | struct sched_domain *sd; | |
1711 | struct s_data d; | |
1712 | struct rq *rq = NULL; | |
1713 | int i, ret = -ENOMEM; | |
05484e09 | 1714 | struct sched_domain_topology_level *tl_asym; |
df054e84 | 1715 | bool has_asym = false; |
f2cb1360 IM |
1716 | |
1717 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); | |
1718 | if (alloc_state != sa_rootdomain) | |
1719 | goto error; | |
1720 | ||
05484e09 MR |
1721 | tl_asym = asym_cpu_capacity_level(cpu_map); |
1722 | ||
f2cb1360 IM |
1723 | /* Set up domains for CPUs specified by the cpu_map: */ |
1724 | for_each_cpu(i, cpu_map) { | |
1725 | struct sched_domain_topology_level *tl; | |
1726 | ||
1727 | sd = NULL; | |
1728 | for_each_sd_topology(tl) { | |
05484e09 MR |
1729 | int dflags = 0; |
1730 | ||
df054e84 | 1731 | if (tl == tl_asym) { |
05484e09 | 1732 | dflags |= SD_ASYM_CPUCAPACITY; |
df054e84 MR |
1733 | has_asym = true; |
1734 | } | |
05484e09 MR |
1735 | |
1736 | sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i); | |
1737 | ||
f2cb1360 IM |
1738 | if (tl == sched_domain_topology) |
1739 | *per_cpu_ptr(d.sd, i) = sd; | |
af85596c | 1740 | if (tl->flags & SDTL_OVERLAP) |
f2cb1360 IM |
1741 | sd->flags |= SD_OVERLAP; |
1742 | if (cpumask_equal(cpu_map, sched_domain_span(sd))) | |
1743 | break; | |
1744 | } | |
1745 | } | |
1746 | ||
1747 | /* Build the groups for the domains */ | |
1748 | for_each_cpu(i, cpu_map) { | |
1749 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { | |
1750 | sd->span_weight = cpumask_weight(sched_domain_span(sd)); | |
1751 | if (sd->flags & SD_OVERLAP) { | |
1752 | if (build_overlap_sched_groups(sd, i)) | |
1753 | goto error; | |
1754 | } else { | |
1755 | if (build_sched_groups(sd, i)) | |
1756 | goto error; | |
1757 | } | |
1758 | } | |
1759 | } | |
1760 | ||
1761 | /* Calculate CPU capacity for physical packages and nodes */ | |
1762 | for (i = nr_cpumask_bits-1; i >= 0; i--) { | |
1763 | if (!cpumask_test_cpu(i, cpu_map)) | |
1764 | continue; | |
1765 | ||
1766 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { | |
1767 | claim_allocations(i, sd); | |
1768 | init_sched_groups_capacity(i, sd); | |
1769 | } | |
1770 | } | |
1771 | ||
1772 | /* Attach the domains */ | |
1773 | rcu_read_lock(); | |
1774 | for_each_cpu(i, cpu_map) { | |
1775 | rq = cpu_rq(i); | |
1776 | sd = *per_cpu_ptr(d.sd, i); | |
1777 | ||
1778 | /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */ | |
1779 | if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity)) | |
1780 | WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig); | |
1781 | ||
1782 | cpu_attach_domain(sd, d.rd, i); | |
1783 | } | |
1784 | rcu_read_unlock(); | |
1785 | ||
df054e84 MR |
1786 | if (has_asym) |
1787 | static_branch_enable_cpuslocked(&sched_asym_cpucapacity); | |
1788 | ||
f2cb1360 | 1789 | if (rq && sched_debug_enabled) { |
bf5015a5 | 1790 | pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n", |
f2cb1360 IM |
1791 | cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity); |
1792 | } | |
1793 | ||
1794 | ret = 0; | |
1795 | error: | |
1796 | __free_domain_allocs(&d, alloc_state, cpu_map); | |
97fb7a0a | 1797 | |
f2cb1360 IM |
1798 | return ret; |
1799 | } | |
1800 | ||
1801 | /* Current sched domains: */ | |
1802 | static cpumask_var_t *doms_cur; | |
1803 | ||
1804 | /* Number of sched domains in 'doms_cur': */ | |
1805 | static int ndoms_cur; | |
1806 | ||
1807 | /* Attribues of custom domains in 'doms_cur' */ | |
1808 | static struct sched_domain_attr *dattr_cur; | |
1809 | ||
1810 | /* | |
1811 | * Special case: If a kmalloc() of a doms_cur partition (array of | |
1812 | * cpumask) fails, then fallback to a single sched domain, | |
1813 | * as determined by the single cpumask fallback_doms. | |
1814 | */ | |
8d5dc512 | 1815 | static cpumask_var_t fallback_doms; |
f2cb1360 IM |
1816 | |
1817 | /* | |
1818 | * arch_update_cpu_topology lets virtualized architectures update the | |
1819 | * CPU core maps. It is supposed to return 1 if the topology changed | |
1820 | * or 0 if it stayed the same. | |
1821 | */ | |
1822 | int __weak arch_update_cpu_topology(void) | |
1823 | { | |
1824 | return 0; | |
1825 | } | |
1826 | ||
1827 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) | |
1828 | { | |
1829 | int i; | |
1830 | cpumask_var_t *doms; | |
1831 | ||
6da2ec56 | 1832 | doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL); |
f2cb1360 IM |
1833 | if (!doms) |
1834 | return NULL; | |
1835 | for (i = 0; i < ndoms; i++) { | |
1836 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { | |
1837 | free_sched_domains(doms, i); | |
1838 | return NULL; | |
1839 | } | |
1840 | } | |
1841 | return doms; | |
1842 | } | |
1843 | ||
1844 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) | |
1845 | { | |
1846 | unsigned int i; | |
1847 | for (i = 0; i < ndoms; i++) | |
1848 | free_cpumask_var(doms[i]); | |
1849 | kfree(doms); | |
1850 | } | |
1851 | ||
1852 | /* | |
1853 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | |
1854 | * For now this just excludes isolated CPUs, but could be used to | |
1855 | * exclude other special cases in the future. | |
1856 | */ | |
8d5dc512 | 1857 | int sched_init_domains(const struct cpumask *cpu_map) |
f2cb1360 IM |
1858 | { |
1859 | int err; | |
1860 | ||
8d5dc512 | 1861 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL); |
1676330e | 1862 | zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL); |
8d5dc512 PZ |
1863 | zalloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
1864 | ||
f2cb1360 IM |
1865 | arch_update_cpu_topology(); |
1866 | ndoms_cur = 1; | |
1867 | doms_cur = alloc_sched_domains(ndoms_cur); | |
1868 | if (!doms_cur) | |
1869 | doms_cur = &fallback_doms; | |
edb93821 | 1870 | cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN)); |
f2cb1360 IM |
1871 | err = build_sched_domains(doms_cur[0], NULL); |
1872 | register_sched_domain_sysctl(); | |
1873 | ||
1874 | return err; | |
1875 | } | |
1876 | ||
1877 | /* | |
1878 | * Detach sched domains from a group of CPUs specified in cpu_map | |
1879 | * These CPUs will now be attached to the NULL domain | |
1880 | */ | |
1881 | static void detach_destroy_domains(const struct cpumask *cpu_map) | |
1882 | { | |
1883 | int i; | |
1884 | ||
1885 | rcu_read_lock(); | |
1886 | for_each_cpu(i, cpu_map) | |
1887 | cpu_attach_domain(NULL, &def_root_domain, i); | |
1888 | rcu_read_unlock(); | |
1889 | } | |
1890 | ||
1891 | /* handle null as "default" */ | |
1892 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
1893 | struct sched_domain_attr *new, int idx_new) | |
1894 | { | |
1895 | struct sched_domain_attr tmp; | |
1896 | ||
1897 | /* Fast path: */ | |
1898 | if (!new && !cur) | |
1899 | return 1; | |
1900 | ||
1901 | tmp = SD_ATTR_INIT; | |
97fb7a0a | 1902 | |
f2cb1360 IM |
1903 | return !memcmp(cur ? (cur + idx_cur) : &tmp, |
1904 | new ? (new + idx_new) : &tmp, | |
1905 | sizeof(struct sched_domain_attr)); | |
1906 | } | |
1907 | ||
1908 | /* | |
1909 | * Partition sched domains as specified by the 'ndoms_new' | |
1910 | * cpumasks in the array doms_new[] of cpumasks. This compares | |
1911 | * doms_new[] to the current sched domain partitioning, doms_cur[]. | |
1912 | * It destroys each deleted domain and builds each new domain. | |
1913 | * | |
1914 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. | |
1915 | * The masks don't intersect (don't overlap.) We should setup one | |
1916 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
1917 | * not be load balanced. If the same cpumask appears both in the | |
1918 | * current 'doms_cur' domains and in the new 'doms_new', we can leave | |
1919 | * it as it is. | |
1920 | * | |
1921 | * The passed in 'doms_new' should be allocated using | |
1922 | * alloc_sched_domains. This routine takes ownership of it and will | |
1923 | * free_sched_domains it when done with it. If the caller failed the | |
1924 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, | |
1925 | * and partition_sched_domains() will fallback to the single partition | |
1926 | * 'fallback_doms', it also forces the domains to be rebuilt. | |
1927 | * | |
1928 | * If doms_new == NULL it will be replaced with cpu_online_mask. | |
1929 | * ndoms_new == 0 is a special case for destroying existing domains, | |
1930 | * and it will not create the default domain. | |
1931 | * | |
1932 | * Call with hotplug lock held | |
1933 | */ | |
1934 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], | |
1935 | struct sched_domain_attr *dattr_new) | |
1936 | { | |
1937 | int i, j, n; | |
1938 | int new_topology; | |
1939 | ||
1940 | mutex_lock(&sched_domains_mutex); | |
1941 | ||
1942 | /* Always unregister in case we don't destroy any domains: */ | |
1943 | unregister_sched_domain_sysctl(); | |
1944 | ||
1945 | /* Let the architecture update CPU core mappings: */ | |
1946 | new_topology = arch_update_cpu_topology(); | |
1947 | ||
09e0dd8e PZ |
1948 | if (!doms_new) { |
1949 | WARN_ON_ONCE(dattr_new); | |
1950 | n = 0; | |
1951 | doms_new = alloc_sched_domains(1); | |
1952 | if (doms_new) { | |
1953 | n = 1; | |
edb93821 FW |
1954 | cpumask_and(doms_new[0], cpu_active_mask, |
1955 | housekeeping_cpumask(HK_FLAG_DOMAIN)); | |
09e0dd8e PZ |
1956 | } |
1957 | } else { | |
1958 | n = ndoms_new; | |
1959 | } | |
f2cb1360 IM |
1960 | |
1961 | /* Destroy deleted domains: */ | |
1962 | for (i = 0; i < ndoms_cur; i++) { | |
1963 | for (j = 0; j < n && !new_topology; j++) { | |
1964 | if (cpumask_equal(doms_cur[i], doms_new[j]) | |
1965 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
1966 | goto match1; | |
1967 | } | |
1968 | /* No match - a current sched domain not in new doms_new[] */ | |
1969 | detach_destroy_domains(doms_cur[i]); | |
1970 | match1: | |
1971 | ; | |
1972 | } | |
1973 | ||
1974 | n = ndoms_cur; | |
09e0dd8e | 1975 | if (!doms_new) { |
f2cb1360 IM |
1976 | n = 0; |
1977 | doms_new = &fallback_doms; | |
edb93821 FW |
1978 | cpumask_and(doms_new[0], cpu_active_mask, |
1979 | housekeeping_cpumask(HK_FLAG_DOMAIN)); | |
f2cb1360 IM |
1980 | } |
1981 | ||
1982 | /* Build new domains: */ | |
1983 | for (i = 0; i < ndoms_new; i++) { | |
1984 | for (j = 0; j < n && !new_topology; j++) { | |
1985 | if (cpumask_equal(doms_new[i], doms_cur[j]) | |
1986 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
1987 | goto match2; | |
1988 | } | |
1989 | /* No match - add a new doms_new */ | |
1990 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); | |
1991 | match2: | |
1992 | ; | |
1993 | } | |
1994 | ||
1995 | /* Remember the new sched domains: */ | |
1996 | if (doms_cur != &fallback_doms) | |
1997 | free_sched_domains(doms_cur, ndoms_cur); | |
1998 | ||
1999 | kfree(dattr_cur); | |
2000 | doms_cur = doms_new; | |
2001 | dattr_cur = dattr_new; | |
2002 | ndoms_cur = ndoms_new; | |
2003 | ||
2004 | register_sched_domain_sysctl(); | |
2005 | ||
2006 | mutex_unlock(&sched_domains_mutex); | |
2007 | } |